blob: 4950de4add3bf267d02a7a3d8b593199ac9f8484 [file] [edit]
; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 6
; RUN: llc -mtriple=amdgpu11.00--amdpal -verify-misched < %s | FileCheck --check-prefixes=GFX11-PAL %s
; RUN: llc -mtriple=amdgpu11.00--amdpal -amdgpu-use-amdgpu-trackers=1 -verify-misched < %s | FileCheck --check-prefixes=GFX11-PAL-GCNTRACKERS %s
; RUN: llc -mtriple=amdgpu8.02 -amdgpu-scalarize-global-loads=false -verify-misched < %s | FileCheck --check-prefixes=TONGA %s
; RUN: llc -mtriple=amdgpu8.02 -amdgpu-scalarize-global-loads=false -amdgpu-use-amdgpu-trackers=1 -verify-misched < %s | FileCheck --check-prefixes=TONGA-GCNTRACKERS %s
; RUN: llc -mtriple=amdgpu9.08 -verify-misched < %s | FileCheck --check-prefixes=GFX908 %s
; RUN: llc -mtriple=amdgpu9.08 -amdgpu-use-amdgpu-trackers=1 -verify-misched < %s | FileCheck --check-prefixes=GFX908-GCNTRACKERS %s
; RUN: llc -mtriple=amdgpu6.00 -verify-misched < %s | FileCheck --check-prefixes=GENERIC %s
; RUN: llc -mtriple=amdgpu6.00 -amdgpu-use-amdgpu-trackers=1 -verify-misched < %s | FileCheck --check-prefixes=GENERIC-GCNTRACKERS %s
; GCN Trackers are sensitive to minor changes in RP, and will avoid scheduling certain instructions, which, if scheduled,
; allow scheduling of other instructions which reduce RP
; CHECK-LABEL: {{^}}return_72xi32:
; GFX11-PAL: NumSgprs: 33
; GFX11-PAL-GCNTRACKERS: NumSgprs: 33
; GFX11-PAL: NumVgprs: 64
; GFX11-PAL-GCNTRACKERS: NumVgprs: 64
; GFX11-PAL: ScratchSize: 220
; GFX11-PAL-GCNTRACKERS: ScratchSize: 248
; CHECK-LABEL: {{^}}call_72xi32:
; GFX11-PAL: NumSgprs: 40
; GFX11-PAL-GCNTRACKERS: NumSgprs: 37
; GFX11-PAL: NumVgprs: 64
; GFX11-PAL-GCNTRACKERS: NumVgprs: 64
; GFX11-PAL: ScratchSize: 2780
; GFX11-PAL-GCNTRACKERS: ScratchSize: 2808
define amdgpu_gfx <72 x i32> @return_72xi32(<72 x i32> %val) #1 {
; GFX11-PAL-LABEL: return_72xi32:
; GFX11-PAL: ; %bb.0:
; GFX11-PAL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-PAL-NEXT: s_clause 0xc ; 52-byte Folded Spill
; GFX11-PAL-NEXT: scratch_store_b32 off, v40, s32 offset:212
; GFX11-PAL-NEXT: ; meta instruction
; GFX11-PAL-NEXT: scratch_store_b32 off, v41, s32 offset:208
; GFX11-PAL-NEXT: ; meta instruction
; GFX11-PAL-NEXT: scratch_store_b32 off, v42, s32 offset:204
; GFX11-PAL-NEXT: ; meta instruction
; GFX11-PAL-NEXT: scratch_store_b32 off, v43, s32 offset:200
; GFX11-PAL-NEXT: ; meta instruction
; GFX11-PAL-NEXT: scratch_store_b32 off, v44, s32 offset:196
; GFX11-PAL-NEXT: ; meta instruction
; GFX11-PAL-NEXT: scratch_store_b32 off, v56, s32 offset:192
; GFX11-PAL-NEXT: ; meta instruction
; GFX11-PAL-NEXT: scratch_store_b32 off, v57, s32 offset:188
; GFX11-PAL-NEXT: ; meta instruction
; GFX11-PAL-NEXT: scratch_store_b32 off, v58, s32 offset:184
; GFX11-PAL-NEXT: ; meta instruction
; GFX11-PAL-NEXT: scratch_store_b32 off, v59, s32 offset:180
; GFX11-PAL-NEXT: ; meta instruction
; GFX11-PAL-NEXT: scratch_store_b32 off, v60, s32 offset:176
; GFX11-PAL-NEXT: ; meta instruction
; GFX11-PAL-NEXT: scratch_store_b32 off, v61, s32 offset:172
; GFX11-PAL-NEXT: ; meta instruction
; GFX11-PAL-NEXT: scratch_store_b32 off, v62, s32 offset:168
; GFX11-PAL-NEXT: ; meta instruction
; GFX11-PAL-NEXT: scratch_store_b32 off, v63, s32 offset:164
; GFX11-PAL-NEXT: s_clause 0x11
; GFX11-PAL-NEXT: scratch_load_b32 v36, off, s32 offset:16
; GFX11-PAL-NEXT: scratch_load_b32 v35, off, s32 offset:12
; GFX11-PAL-NEXT: scratch_load_b32 v34, off, s32 offset:8
; GFX11-PAL-NEXT: scratch_load_b32 v51, off, s32 offset:32
; GFX11-PAL-NEXT: scratch_load_b32 v50, off, s32 offset:28
; GFX11-PAL-NEXT: scratch_load_b32 v49, off, s32 offset:24
; GFX11-PAL-NEXT: scratch_load_b32 v55, off, s32 offset:48
; GFX11-PAL-NEXT: scratch_load_b32 v54, off, s32 offset:44
; GFX11-PAL-NEXT: scratch_load_b32 v53, off, s32 offset:40
; GFX11-PAL-NEXT: scratch_load_b32 v40, off, s32 offset:64
; GFX11-PAL-NEXT: scratch_load_b32 v39, off, s32 offset:60
; GFX11-PAL-NEXT: scratch_load_b32 v38, off, s32 offset:56
; GFX11-PAL-NEXT: scratch_load_b32 v44, off, s32 offset:80
; GFX11-PAL-NEXT: scratch_load_b32 v43, off, s32 offset:76
; GFX11-PAL-NEXT: scratch_load_b32 v42, off, s32 offset:72
; GFX11-PAL-NEXT: scratch_load_b32 v59, off, s32 offset:96
; GFX11-PAL-NEXT: scratch_load_b32 v58, off, s32 offset:92
; GFX11-PAL-NEXT: scratch_load_b32 v57, off, s32 offset:88
; GFX11-PAL-NEXT: s_clause 0x1
; GFX11-PAL-NEXT: scratch_store_b128 v0, v[21:24], off offset:80
; GFX11-PAL-NEXT: scratch_store_b128 v0, v[17:20], off offset:64
; GFX11-PAL-NEXT: s_clause 0x5
; GFX11-PAL-NEXT: scratch_load_b32 v23, off, s32 offset:112
; GFX11-PAL-NEXT: scratch_load_b32 v22, off, s32 offset:108
; GFX11-PAL-NEXT: scratch_load_b32 v21, off, s32 offset:104
; GFX11-PAL-NEXT: scratch_load_b32 v19, off, s32 offset:128
; GFX11-PAL-NEXT: scratch_load_b32 v18, off, s32 offset:124
; GFX11-PAL-NEXT: scratch_load_b32 v17, off, s32 offset:120
; GFX11-PAL-NEXT: s_clause 0x1
; GFX11-PAL-NEXT: scratch_store_b128 v0, v[13:16], off offset:48
; GFX11-PAL-NEXT: scratch_store_b128 v0, v[9:12], off offset:32
; GFX11-PAL-NEXT: s_clause 0x10
; GFX11-PAL-NEXT: scratch_load_b32 v15, off, s32 offset:144
; GFX11-PAL-NEXT: scratch_load_b32 v14, off, s32 offset:140
; GFX11-PAL-NEXT: scratch_load_b32 v13, off, s32 offset:136
; GFX11-PAL-NEXT: scratch_load_b32 v63, off, s32 offset:160
; GFX11-PAL-NEXT: scratch_load_b32 v62, off, s32 offset:156
; GFX11-PAL-NEXT: scratch_load_b32 v61, off, s32 offset:152
; GFX11-PAL-NEXT: scratch_load_b32 v60, off, s32 offset:148
; GFX11-PAL-NEXT: scratch_load_b32 v12, off, s32 offset:132
; GFX11-PAL-NEXT: scratch_load_b32 v16, off, s32 offset:116
; GFX11-PAL-NEXT: scratch_load_b32 v20, off, s32 offset:100
; GFX11-PAL-NEXT: scratch_load_b32 v56, off, s32 offset:84
; GFX11-PAL-NEXT: scratch_load_b32 v41, off, s32 offset:68
; GFX11-PAL-NEXT: scratch_load_b32 v37, off, s32 offset:52
; GFX11-PAL-NEXT: scratch_load_b32 v52, off, s32 offset:36
; GFX11-PAL-NEXT: scratch_load_b32 v48, off, s32 offset:20
; GFX11-PAL-NEXT: scratch_load_b32 v33, off, s32 offset:4
; GFX11-PAL-NEXT: scratch_load_b32 v32, off, s32
; GFX11-PAL-NEXT: s_waitcnt vmcnt(10)
; GFX11-PAL-NEXT: scratch_store_b128 v0, v[60:63], off offset:272
; GFX11-PAL-NEXT: s_waitcnt vmcnt(9)
; GFX11-PAL-NEXT: scratch_store_b128 v0, v[12:15], off offset:256
; GFX11-PAL-NEXT: s_waitcnt vmcnt(8)
; GFX11-PAL-NEXT: scratch_store_b128 v0, v[16:19], off offset:240
; GFX11-PAL-NEXT: s_waitcnt vmcnt(7)
; GFX11-PAL-NEXT: scratch_store_b128 v0, v[20:23], off offset:224
; GFX11-PAL-NEXT: s_waitcnt vmcnt(6)
; GFX11-PAL-NEXT: scratch_store_b128 v0, v[56:59], off offset:208
; GFX11-PAL-NEXT: s_waitcnt vmcnt(5)
; GFX11-PAL-NEXT: scratch_store_b128 v0, v[41:44], off offset:192
; GFX11-PAL-NEXT: s_waitcnt vmcnt(4)
; GFX11-PAL-NEXT: scratch_store_b128 v0, v[37:40], off offset:176
; GFX11-PAL-NEXT: s_waitcnt vmcnt(3)
; GFX11-PAL-NEXT: scratch_store_b128 v0, v[52:55], off offset:160
; GFX11-PAL-NEXT: s_waitcnt vmcnt(2)
; GFX11-PAL-NEXT: scratch_store_b128 v0, v[48:51], off offset:144
; GFX11-PAL-NEXT: s_waitcnt vmcnt(1)
; GFX11-PAL-NEXT: scratch_store_b128 v0, v[33:36], off offset:128
; GFX11-PAL-NEXT: s_waitcnt vmcnt(0)
; GFX11-PAL-NEXT: s_clause 0x3
; GFX11-PAL-NEXT: scratch_store_b128 v0, v[29:32], off offset:112
; GFX11-PAL-NEXT: scratch_store_b128 v0, v[25:28], off offset:96
; GFX11-PAL-NEXT: scratch_store_b128 v0, v[5:8], off offset:16
; GFX11-PAL-NEXT: scratch_store_b128 v0, v[1:4], off
; GFX11-PAL-NEXT: s_clause 0xc ; 52-byte Folded Reload
; GFX11-PAL-NEXT: scratch_load_b32 v63, off, s32 offset:164
; GFX11-PAL-NEXT: scratch_load_b32 v62, off, s32 offset:168
; GFX11-PAL-NEXT: scratch_load_b32 v61, off, s32 offset:172
; GFX11-PAL-NEXT: scratch_load_b32 v60, off, s32 offset:176
; GFX11-PAL-NEXT: scratch_load_b32 v59, off, s32 offset:180
; GFX11-PAL-NEXT: scratch_load_b32 v58, off, s32 offset:184
; GFX11-PAL-NEXT: scratch_load_b32 v57, off, s32 offset:188
; GFX11-PAL-NEXT: scratch_load_b32 v56, off, s32 offset:192
; GFX11-PAL-NEXT: scratch_load_b32 v44, off, s32 offset:196
; GFX11-PAL-NEXT: scratch_load_b32 v43, off, s32 offset:200
; GFX11-PAL-NEXT: scratch_load_b32 v42, off, s32 offset:204
; GFX11-PAL-NEXT: scratch_load_b32 v41, off, s32 offset:208
; GFX11-PAL-NEXT: scratch_load_b32 v40, off, s32 offset:212
; GFX11-PAL-NEXT: s_waitcnt vmcnt(0)
; GFX11-PAL-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-PAL-GCNTRACKERS-LABEL: return_72xi32:
; GFX11-PAL-GCNTRACKERS: ; %bb.0:
; GFX11-PAL-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-PAL-GCNTRACKERS-NEXT: s_clause 0xf ; 64-byte Folded Spill
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_store_b32 off, v40, s32 offset:224
; GFX11-PAL-GCNTRACKERS-NEXT: ; meta instruction
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_store_b32 off, v41, s32 offset:220
; GFX11-PAL-GCNTRACKERS-NEXT: ; meta instruction
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_store_b32 off, v42, s32 offset:216
; GFX11-PAL-GCNTRACKERS-NEXT: ; meta instruction
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_store_b32 off, v43, s32 offset:212
; GFX11-PAL-GCNTRACKERS-NEXT: ; meta instruction
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_store_b32 off, v44, s32 offset:208
; GFX11-PAL-GCNTRACKERS-NEXT: ; meta instruction
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_store_b32 off, v45, s32 offset:204
; GFX11-PAL-GCNTRACKERS-NEXT: ; meta instruction
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_store_b32 off, v46, s32 offset:200
; GFX11-PAL-GCNTRACKERS-NEXT: ; meta instruction
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_store_b32 off, v47, s32 offset:196
; GFX11-PAL-GCNTRACKERS-NEXT: ; meta instruction
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_store_b32 off, v56, s32 offset:192
; GFX11-PAL-GCNTRACKERS-NEXT: ; meta instruction
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_store_b32 off, v57, s32 offset:188
; GFX11-PAL-GCNTRACKERS-NEXT: ; meta instruction
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_store_b32 off, v58, s32 offset:184
; GFX11-PAL-GCNTRACKERS-NEXT: ; meta instruction
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_store_b32 off, v59, s32 offset:180
; GFX11-PAL-GCNTRACKERS-NEXT: ; meta instruction
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_store_b32 off, v60, s32 offset:176
; GFX11-PAL-GCNTRACKERS-NEXT: ; meta instruction
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_store_b32 off, v61, s32 offset:172
; GFX11-PAL-GCNTRACKERS-NEXT: ; meta instruction
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_store_b32 off, v62, s32 offset:168
; GFX11-PAL-GCNTRACKERS-NEXT: ; meta instruction
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_store_b32 off, v63, s32 offset:164
; GFX11-PAL-GCNTRACKERS-NEXT: v_dual_mov_b32 v47, v31 :: v_dual_mov_b32 v46, v30
; GFX11-PAL-GCNTRACKERS-NEXT: v_mov_b32_e32 v45, v29
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_store_b128 off, v[25:28], s32 offset:228 ; 16-byte Folded Spill
; GFX11-PAL-GCNTRACKERS-NEXT: v_dual_mov_b32 v28, v8 :: v_dual_mov_b32 v27, v7
; GFX11-PAL-GCNTRACKERS-NEXT: v_dual_mov_b32 v26, v6 :: v_dual_mov_b32 v25, v5
; GFX11-PAL-GCNTRACKERS-NEXT: v_dual_mov_b32 v32, v4 :: v_dual_mov_b32 v31, v3
; GFX11-PAL-GCNTRACKERS-NEXT: v_dual_mov_b32 v30, v2 :: v_dual_mov_b32 v29, v1
; GFX11-PAL-GCNTRACKERS-NEXT: s_clause 0x1b
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b32 v36, off, s32 offset:16
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b32 v35, off, s32 offset:12
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b32 v34, off, s32 offset:8
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b32 v33, off, s32 offset:4
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b32 v51, off, s32 offset:32
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b32 v50, off, s32 offset:28
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b32 v49, off, s32 offset:24
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b32 v55, off, s32 offset:48
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b32 v54, off, s32 offset:44
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b32 v53, off, s32 offset:40
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b32 v40, off, s32 offset:64
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b32 v39, off, s32 offset:60
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b32 v38, off, s32 offset:56
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b32 v44, off, s32 offset:80
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b32 v43, off, s32 offset:76
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b32 v42, off, s32 offset:72
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b32 v59, off, s32 offset:96
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b32 v58, off, s32 offset:92
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b32 v57, off, s32 offset:88
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b32 v56, off, s32 offset:84
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b32 v63, off, s32 offset:112
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b32 v62, off, s32 offset:108
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b32 v61, off, s32 offset:104
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b32 v8, off, s32 offset:128
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b32 v7, off, s32 offset:124
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b32 v6, off, s32 offset:120
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b32 v4, off, s32 offset:144
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b32 v3, off, s32 offset:140
; GFX11-PAL-GCNTRACKERS-NEXT: s_clause 0x1
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_store_b128 v0, v[21:24], off offset:80
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_store_b128 v0, v[17:20], off offset:64
; GFX11-PAL-GCNTRACKERS-NEXT: s_clause 0xb
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b32 v2, off, s32 offset:136
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b32 v23, off, s32 offset:160
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b32 v22, off, s32 offset:156
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b32 v21, off, s32 offset:152
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b32 v20, off, s32 offset:148
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b32 v1, off, s32 offset:132
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b32 v5, off, s32 offset:116
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b32 v60, off, s32 offset:100
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b32 v41, off, s32 offset:68
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b32 v37, off, s32 offset:52
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b32 v52, off, s32 offset:36
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b32 v48, off, s32 offset:20
; GFX11-PAL-GCNTRACKERS-NEXT: s_clause 0x1
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_store_b128 v0, v[13:16], off offset:48
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_store_b128 v0, v[9:12], off offset:32
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b32 v12, off, s32
; GFX11-PAL-GCNTRACKERS-NEXT: v_dual_mov_b32 v9, v45 :: v_dual_mov_b32 v10, v46
; GFX11-PAL-GCNTRACKERS-NEXT: v_mov_b32_e32 v11, v47
; GFX11-PAL-GCNTRACKERS-NEXT: s_waitcnt vmcnt(8)
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_store_b128 v0, v[20:23], off offset:272
; GFX11-PAL-GCNTRACKERS-NEXT: s_waitcnt vmcnt(7)
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_store_b128 v0, v[1:4], off offset:256
; GFX11-PAL-GCNTRACKERS-NEXT: s_waitcnt vmcnt(6)
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_store_b128 v0, v[5:8], off offset:240
; GFX11-PAL-GCNTRACKERS-NEXT: s_waitcnt vmcnt(5)
; GFX11-PAL-GCNTRACKERS-NEXT: s_clause 0x1
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_store_b128 v0, v[60:63], off offset:224
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_store_b128 v0, v[56:59], off offset:208
; GFX11-PAL-GCNTRACKERS-NEXT: s_waitcnt vmcnt(4)
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_store_b128 v0, v[41:44], off offset:192
; GFX11-PAL-GCNTRACKERS-NEXT: s_waitcnt vmcnt(3)
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_store_b128 v0, v[37:40], off offset:176
; GFX11-PAL-GCNTRACKERS-NEXT: s_waitcnt vmcnt(2)
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_store_b128 v0, v[52:55], off offset:160
; GFX11-PAL-GCNTRACKERS-NEXT: s_waitcnt vmcnt(1)
; GFX11-PAL-GCNTRACKERS-NEXT: s_clause 0x1
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_store_b128 v0, v[48:51], off offset:144
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_store_b128 v0, v[33:36], off offset:128
; GFX11-PAL-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_store_b128 v0, v[9:12], off offset:112
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b128 v[1:4], off, s32 offset:228 ; 16-byte Folded Reload
; GFX11-PAL-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; GFX11-PAL-GCNTRACKERS-NEXT: s_clause 0x2
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_store_b128 v0, v[1:4], off offset:96
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_store_b128 v0, v[25:28], off offset:16
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_store_b128 v0, v[29:32], off
; GFX11-PAL-GCNTRACKERS-NEXT: s_clause 0xf ; 64-byte Folded Reload
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b32 v63, off, s32 offset:164
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b32 v62, off, s32 offset:168
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b32 v61, off, s32 offset:172
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b32 v60, off, s32 offset:176
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b32 v59, off, s32 offset:180
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b32 v58, off, s32 offset:184
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b32 v57, off, s32 offset:188
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b32 v56, off, s32 offset:192
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b32 v47, off, s32 offset:196
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b32 v46, off, s32 offset:200
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b32 v45, off, s32 offset:204
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b32 v44, off, s32 offset:208
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b32 v43, off, s32 offset:212
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b32 v42, off, s32 offset:216
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b32 v41, off, s32 offset:220
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b32 v40, off, s32 offset:224
; GFX11-PAL-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; GFX11-PAL-GCNTRACKERS-NEXT: s_setpc_b64 s[30:31]
;
; TONGA-LABEL: return_72xi32:
; TONGA: ; %bb.0:
; TONGA-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; TONGA-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:160
; TONGA-NEXT: s_waitcnt vmcnt(0)
; TONGA-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:284
; TONGA-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:156
; TONGA-NEXT: s_waitcnt vmcnt(0)
; TONGA-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:280
; TONGA-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:152
; TONGA-NEXT: s_waitcnt vmcnt(0)
; TONGA-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:276
; TONGA-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:148
; TONGA-NEXT: s_waitcnt vmcnt(0)
; TONGA-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:272
; TONGA-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:144
; TONGA-NEXT: s_waitcnt vmcnt(0)
; TONGA-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:268
; TONGA-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:140
; TONGA-NEXT: s_waitcnt vmcnt(0)
; TONGA-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:264
; TONGA-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:136
; TONGA-NEXT: s_waitcnt vmcnt(0)
; TONGA-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:260
; TONGA-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:132
; TONGA-NEXT: s_waitcnt vmcnt(0)
; TONGA-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:256
; TONGA-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:128
; TONGA-NEXT: s_waitcnt vmcnt(0)
; TONGA-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:252
; TONGA-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:124
; TONGA-NEXT: s_waitcnt vmcnt(0)
; TONGA-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:248
; TONGA-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:120
; TONGA-NEXT: s_waitcnt vmcnt(0)
; TONGA-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:244
; TONGA-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:116
; TONGA-NEXT: s_waitcnt vmcnt(0)
; TONGA-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:240
; TONGA-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:112
; TONGA-NEXT: s_waitcnt vmcnt(0)
; TONGA-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:236
; TONGA-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:108
; TONGA-NEXT: s_waitcnt vmcnt(0)
; TONGA-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:232
; TONGA-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:104
; TONGA-NEXT: s_waitcnt vmcnt(0)
; TONGA-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:228
; TONGA-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:100
; TONGA-NEXT: s_waitcnt vmcnt(0)
; TONGA-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:224
; TONGA-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:96
; TONGA-NEXT: s_waitcnt vmcnt(0)
; TONGA-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:220
; TONGA-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:92
; TONGA-NEXT: s_waitcnt vmcnt(0)
; TONGA-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:216
; TONGA-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:88
; TONGA-NEXT: s_waitcnt vmcnt(0)
; TONGA-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:212
; TONGA-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:84
; TONGA-NEXT: s_waitcnt vmcnt(0)
; TONGA-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:208
; TONGA-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:80
; TONGA-NEXT: s_waitcnt vmcnt(0)
; TONGA-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:204
; TONGA-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:76
; TONGA-NEXT: s_waitcnt vmcnt(0)
; TONGA-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:200
; TONGA-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:72
; TONGA-NEXT: s_waitcnt vmcnt(0)
; TONGA-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:196
; TONGA-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:68
; TONGA-NEXT: s_waitcnt vmcnt(0)
; TONGA-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:192
; TONGA-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:64
; TONGA-NEXT: s_waitcnt vmcnt(0)
; TONGA-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:188
; TONGA-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:60
; TONGA-NEXT: s_waitcnt vmcnt(0)
; TONGA-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:184
; TONGA-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:56
; TONGA-NEXT: s_waitcnt vmcnt(0)
; TONGA-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:180
; TONGA-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:52
; TONGA-NEXT: s_waitcnt vmcnt(0)
; TONGA-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:176
; TONGA-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:48
; TONGA-NEXT: s_waitcnt vmcnt(0)
; TONGA-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:172
; TONGA-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:44
; TONGA-NEXT: s_waitcnt vmcnt(0)
; TONGA-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:168
; TONGA-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:40
; TONGA-NEXT: s_waitcnt vmcnt(0)
; TONGA-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:164
; TONGA-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:36
; TONGA-NEXT: s_waitcnt vmcnt(0)
; TONGA-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:160
; TONGA-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:32
; TONGA-NEXT: s_waitcnt vmcnt(0)
; TONGA-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:156
; TONGA-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:28
; TONGA-NEXT: s_waitcnt vmcnt(0)
; TONGA-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:152
; TONGA-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:24
; TONGA-NEXT: s_waitcnt vmcnt(0)
; TONGA-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:148
; TONGA-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:20
; TONGA-NEXT: s_waitcnt vmcnt(0)
; TONGA-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:144
; TONGA-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:16
; TONGA-NEXT: s_waitcnt vmcnt(0)
; TONGA-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:140
; TONGA-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:12
; TONGA-NEXT: s_waitcnt vmcnt(0)
; TONGA-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:136
; TONGA-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:8
; TONGA-NEXT: s_waitcnt vmcnt(0)
; TONGA-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:132
; TONGA-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:4
; TONGA-NEXT: s_waitcnt vmcnt(0)
; TONGA-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:128
; TONGA-NEXT: buffer_load_dword v32, off, s[0:3], s32
; TONGA-NEXT: s_waitcnt vmcnt(0)
; TONGA-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:124
; TONGA-NEXT: buffer_store_dword v31, v0, s[0:3], 0 offen offset:120
; TONGA-NEXT: buffer_store_dword v30, v0, s[0:3], 0 offen offset:116
; TONGA-NEXT: buffer_store_dword v29, v0, s[0:3], 0 offen offset:112
; TONGA-NEXT: buffer_store_dword v28, v0, s[0:3], 0 offen offset:108
; TONGA-NEXT: buffer_store_dword v27, v0, s[0:3], 0 offen offset:104
; TONGA-NEXT: buffer_store_dword v26, v0, s[0:3], 0 offen offset:100
; TONGA-NEXT: buffer_store_dword v25, v0, s[0:3], 0 offen offset:96
; TONGA-NEXT: buffer_store_dword v24, v0, s[0:3], 0 offen offset:92
; TONGA-NEXT: buffer_store_dword v23, v0, s[0:3], 0 offen offset:88
; TONGA-NEXT: buffer_store_dword v22, v0, s[0:3], 0 offen offset:84
; TONGA-NEXT: buffer_store_dword v21, v0, s[0:3], 0 offen offset:80
; TONGA-NEXT: buffer_store_dword v20, v0, s[0:3], 0 offen offset:76
; TONGA-NEXT: buffer_store_dword v19, v0, s[0:3], 0 offen offset:72
; TONGA-NEXT: buffer_store_dword v18, v0, s[0:3], 0 offen offset:68
; TONGA-NEXT: buffer_store_dword v17, v0, s[0:3], 0 offen offset:64
; TONGA-NEXT: buffer_store_dword v16, v0, s[0:3], 0 offen offset:60
; TONGA-NEXT: buffer_store_dword v15, v0, s[0:3], 0 offen offset:56
; TONGA-NEXT: buffer_store_dword v14, v0, s[0:3], 0 offen offset:52
; TONGA-NEXT: buffer_store_dword v13, v0, s[0:3], 0 offen offset:48
; TONGA-NEXT: buffer_store_dword v12, v0, s[0:3], 0 offen offset:44
; TONGA-NEXT: buffer_store_dword v11, v0, s[0:3], 0 offen offset:40
; TONGA-NEXT: buffer_store_dword v10, v0, s[0:3], 0 offen offset:36
; TONGA-NEXT: buffer_store_dword v9, v0, s[0:3], 0 offen offset:32
; TONGA-NEXT: buffer_store_dword v8, v0, s[0:3], 0 offen offset:28
; TONGA-NEXT: buffer_store_dword v7, v0, s[0:3], 0 offen offset:24
; TONGA-NEXT: buffer_store_dword v6, v0, s[0:3], 0 offen offset:20
; TONGA-NEXT: buffer_store_dword v5, v0, s[0:3], 0 offen offset:16
; TONGA-NEXT: buffer_store_dword v4, v0, s[0:3], 0 offen offset:12
; TONGA-NEXT: buffer_store_dword v3, v0, s[0:3], 0 offen offset:8
; TONGA-NEXT: buffer_store_dword v2, v0, s[0:3], 0 offen offset:4
; TONGA-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
; TONGA-NEXT: s_waitcnt vmcnt(0)
; TONGA-NEXT: s_setpc_b64 s[30:31]
;
; TONGA-GCNTRACKERS-LABEL: return_72xi32:
; TONGA-GCNTRACKERS: ; %bb.0:
; TONGA-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:160
; TONGA-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:284
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:156
; TONGA-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:280
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:152
; TONGA-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:276
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:148
; TONGA-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:272
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:144
; TONGA-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:268
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:140
; TONGA-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:264
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:136
; TONGA-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:260
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:132
; TONGA-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:256
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:128
; TONGA-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:252
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:124
; TONGA-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:248
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:120
; TONGA-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:244
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:116
; TONGA-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:240
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:112
; TONGA-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:236
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:108
; TONGA-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:232
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:104
; TONGA-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:228
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:100
; TONGA-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:224
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:96
; TONGA-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:220
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:92
; TONGA-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:216
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:88
; TONGA-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:212
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:84
; TONGA-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:208
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:80
; TONGA-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:204
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:76
; TONGA-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:200
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:72
; TONGA-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:196
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:68
; TONGA-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:192
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:64
; TONGA-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:188
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:60
; TONGA-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:184
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:56
; TONGA-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:180
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:52
; TONGA-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:176
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:48
; TONGA-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:172
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:44
; TONGA-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:168
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:40
; TONGA-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:164
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:36
; TONGA-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:160
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:32
; TONGA-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:156
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:28
; TONGA-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:152
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:24
; TONGA-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:148
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:20
; TONGA-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:144
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:16
; TONGA-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:140
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:12
; TONGA-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:136
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:8
; TONGA-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:132
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:4
; TONGA-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:128
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32
; TONGA-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:124
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v31, v0, s[0:3], 0 offen offset:120
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v30, v0, s[0:3], 0 offen offset:116
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v29, v0, s[0:3], 0 offen offset:112
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v28, v0, s[0:3], 0 offen offset:108
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v27, v0, s[0:3], 0 offen offset:104
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v26, v0, s[0:3], 0 offen offset:100
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v25, v0, s[0:3], 0 offen offset:96
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v24, v0, s[0:3], 0 offen offset:92
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v23, v0, s[0:3], 0 offen offset:88
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v22, v0, s[0:3], 0 offen offset:84
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v21, v0, s[0:3], 0 offen offset:80
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v20, v0, s[0:3], 0 offen offset:76
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v19, v0, s[0:3], 0 offen offset:72
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v18, v0, s[0:3], 0 offen offset:68
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v17, v0, s[0:3], 0 offen offset:64
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v16, v0, s[0:3], 0 offen offset:60
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v15, v0, s[0:3], 0 offen offset:56
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v14, v0, s[0:3], 0 offen offset:52
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v13, v0, s[0:3], 0 offen offset:48
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v12, v0, s[0:3], 0 offen offset:44
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v11, v0, s[0:3], 0 offen offset:40
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v10, v0, s[0:3], 0 offen offset:36
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v9, v0, s[0:3], 0 offen offset:32
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v8, v0, s[0:3], 0 offen offset:28
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v7, v0, s[0:3], 0 offen offset:24
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v6, v0, s[0:3], 0 offen offset:20
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v5, v0, s[0:3], 0 offen offset:16
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v4, v0, s[0:3], 0 offen offset:12
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v3, v0, s[0:3], 0 offen offset:8
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v2, v0, s[0:3], 0 offen offset:4
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
; TONGA-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; TONGA-GCNTRACKERS-NEXT: s_setpc_b64 s[30:31]
;
; GFX908-LABEL: return_72xi32:
; GFX908: ; %bb.0:
; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX908-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:160
; GFX908-NEXT: s_waitcnt vmcnt(0)
; GFX908-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:284
; GFX908-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:156
; GFX908-NEXT: s_waitcnt vmcnt(0)
; GFX908-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:280
; GFX908-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:152
; GFX908-NEXT: s_waitcnt vmcnt(0)
; GFX908-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:276
; GFX908-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:148
; GFX908-NEXT: s_waitcnt vmcnt(0)
; GFX908-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:272
; GFX908-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:144
; GFX908-NEXT: s_waitcnt vmcnt(0)
; GFX908-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:268
; GFX908-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:140
; GFX908-NEXT: s_waitcnt vmcnt(0)
; GFX908-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:264
; GFX908-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:136
; GFX908-NEXT: s_waitcnt vmcnt(0)
; GFX908-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:260
; GFX908-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:132
; GFX908-NEXT: s_waitcnt vmcnt(0)
; GFX908-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:256
; GFX908-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:128
; GFX908-NEXT: s_waitcnt vmcnt(0)
; GFX908-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:252
; GFX908-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:124
; GFX908-NEXT: s_waitcnt vmcnt(0)
; GFX908-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:248
; GFX908-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:120
; GFX908-NEXT: s_waitcnt vmcnt(0)
; GFX908-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:244
; GFX908-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:116
; GFX908-NEXT: s_waitcnt vmcnt(0)
; GFX908-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:240
; GFX908-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:112
; GFX908-NEXT: s_waitcnt vmcnt(0)
; GFX908-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:236
; GFX908-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:108
; GFX908-NEXT: s_waitcnt vmcnt(0)
; GFX908-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:232
; GFX908-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:104
; GFX908-NEXT: s_waitcnt vmcnt(0)
; GFX908-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:228
; GFX908-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:100
; GFX908-NEXT: s_waitcnt vmcnt(0)
; GFX908-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:224
; GFX908-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:96
; GFX908-NEXT: s_waitcnt vmcnt(0)
; GFX908-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:220
; GFX908-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:92
; GFX908-NEXT: s_waitcnt vmcnt(0)
; GFX908-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:216
; GFX908-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:88
; GFX908-NEXT: s_waitcnt vmcnt(0)
; GFX908-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:212
; GFX908-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:84
; GFX908-NEXT: s_waitcnt vmcnt(0)
; GFX908-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:208
; GFX908-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:80
; GFX908-NEXT: s_waitcnt vmcnt(0)
; GFX908-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:204
; GFX908-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:76
; GFX908-NEXT: s_waitcnt vmcnt(0)
; GFX908-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:200
; GFX908-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:72
; GFX908-NEXT: s_waitcnt vmcnt(0)
; GFX908-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:196
; GFX908-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:68
; GFX908-NEXT: s_waitcnt vmcnt(0)
; GFX908-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:192
; GFX908-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:64
; GFX908-NEXT: s_waitcnt vmcnt(0)
; GFX908-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:188
; GFX908-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:60
; GFX908-NEXT: s_waitcnt vmcnt(0)
; GFX908-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:184
; GFX908-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:56
; GFX908-NEXT: s_waitcnt vmcnt(0)
; GFX908-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:180
; GFX908-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:52
; GFX908-NEXT: s_waitcnt vmcnt(0)
; GFX908-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:176
; GFX908-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:48
; GFX908-NEXT: s_waitcnt vmcnt(0)
; GFX908-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:172
; GFX908-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:44
; GFX908-NEXT: s_waitcnt vmcnt(0)
; GFX908-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:168
; GFX908-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:40
; GFX908-NEXT: s_waitcnt vmcnt(0)
; GFX908-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:164
; GFX908-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:36
; GFX908-NEXT: s_waitcnt vmcnt(0)
; GFX908-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:160
; GFX908-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:32
; GFX908-NEXT: s_waitcnt vmcnt(0)
; GFX908-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:156
; GFX908-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:28
; GFX908-NEXT: s_waitcnt vmcnt(0)
; GFX908-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:152
; GFX908-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:24
; GFX908-NEXT: s_waitcnt vmcnt(0)
; GFX908-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:148
; GFX908-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:20
; GFX908-NEXT: s_waitcnt vmcnt(0)
; GFX908-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:144
; GFX908-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:16
; GFX908-NEXT: s_waitcnt vmcnt(0)
; GFX908-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:140
; GFX908-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:12
; GFX908-NEXT: s_waitcnt vmcnt(0)
; GFX908-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:136
; GFX908-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:8
; GFX908-NEXT: s_waitcnt vmcnt(0)
; GFX908-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:132
; GFX908-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:4
; GFX908-NEXT: s_waitcnt vmcnt(0)
; GFX908-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:128
; GFX908-NEXT: buffer_load_dword v32, off, s[0:3], s32
; GFX908-NEXT: s_waitcnt vmcnt(0)
; GFX908-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:124
; GFX908-NEXT: buffer_store_dword v31, v0, s[0:3], 0 offen offset:120
; GFX908-NEXT: buffer_store_dword v30, v0, s[0:3], 0 offen offset:116
; GFX908-NEXT: buffer_store_dword v29, v0, s[0:3], 0 offen offset:112
; GFX908-NEXT: buffer_store_dword v28, v0, s[0:3], 0 offen offset:108
; GFX908-NEXT: buffer_store_dword v27, v0, s[0:3], 0 offen offset:104
; GFX908-NEXT: buffer_store_dword v26, v0, s[0:3], 0 offen offset:100
; GFX908-NEXT: buffer_store_dword v25, v0, s[0:3], 0 offen offset:96
; GFX908-NEXT: buffer_store_dword v24, v0, s[0:3], 0 offen offset:92
; GFX908-NEXT: buffer_store_dword v23, v0, s[0:3], 0 offen offset:88
; GFX908-NEXT: buffer_store_dword v22, v0, s[0:3], 0 offen offset:84
; GFX908-NEXT: buffer_store_dword v21, v0, s[0:3], 0 offen offset:80
; GFX908-NEXT: buffer_store_dword v20, v0, s[0:3], 0 offen offset:76
; GFX908-NEXT: buffer_store_dword v19, v0, s[0:3], 0 offen offset:72
; GFX908-NEXT: buffer_store_dword v18, v0, s[0:3], 0 offen offset:68
; GFX908-NEXT: buffer_store_dword v17, v0, s[0:3], 0 offen offset:64
; GFX908-NEXT: buffer_store_dword v16, v0, s[0:3], 0 offen offset:60
; GFX908-NEXT: buffer_store_dword v15, v0, s[0:3], 0 offen offset:56
; GFX908-NEXT: buffer_store_dword v14, v0, s[0:3], 0 offen offset:52
; GFX908-NEXT: buffer_store_dword v13, v0, s[0:3], 0 offen offset:48
; GFX908-NEXT: buffer_store_dword v12, v0, s[0:3], 0 offen offset:44
; GFX908-NEXT: buffer_store_dword v11, v0, s[0:3], 0 offen offset:40
; GFX908-NEXT: buffer_store_dword v10, v0, s[0:3], 0 offen offset:36
; GFX908-NEXT: buffer_store_dword v9, v0, s[0:3], 0 offen offset:32
; GFX908-NEXT: buffer_store_dword v8, v0, s[0:3], 0 offen offset:28
; GFX908-NEXT: buffer_store_dword v7, v0, s[0:3], 0 offen offset:24
; GFX908-NEXT: buffer_store_dword v6, v0, s[0:3], 0 offen offset:20
; GFX908-NEXT: buffer_store_dword v5, v0, s[0:3], 0 offen offset:16
; GFX908-NEXT: buffer_store_dword v4, v0, s[0:3], 0 offen offset:12
; GFX908-NEXT: buffer_store_dword v3, v0, s[0:3], 0 offen offset:8
; GFX908-NEXT: buffer_store_dword v2, v0, s[0:3], 0 offen offset:4
; GFX908-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
; GFX908-NEXT: s_waitcnt vmcnt(0)
; GFX908-NEXT: s_setpc_b64 s[30:31]
;
; GFX908-GCNTRACKERS-LABEL: return_72xi32:
; GFX908-GCNTRACKERS: ; %bb.0:
; GFX908-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX908-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:160
; GFX908-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:284
; GFX908-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:156
; GFX908-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:280
; GFX908-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:152
; GFX908-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:276
; GFX908-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:148
; GFX908-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:272
; GFX908-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:144
; GFX908-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:268
; GFX908-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:140
; GFX908-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:264
; GFX908-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:136
; GFX908-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:260
; GFX908-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:132
; GFX908-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:256
; GFX908-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:128
; GFX908-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:252
; GFX908-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:124
; GFX908-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:248
; GFX908-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:120
; GFX908-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:244
; GFX908-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:116
; GFX908-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:240
; GFX908-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:112
; GFX908-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:236
; GFX908-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:108
; GFX908-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:232
; GFX908-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:104
; GFX908-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:228
; GFX908-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:100
; GFX908-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:224
; GFX908-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:96
; GFX908-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:220
; GFX908-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:92
; GFX908-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:216
; GFX908-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:88
; GFX908-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:212
; GFX908-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:84
; GFX908-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:208
; GFX908-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:80
; GFX908-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:204
; GFX908-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:76
; GFX908-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:200
; GFX908-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:72
; GFX908-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:196
; GFX908-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:68
; GFX908-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:192
; GFX908-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:64
; GFX908-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:188
; GFX908-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:60
; GFX908-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:184
; GFX908-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:56
; GFX908-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:180
; GFX908-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:52
; GFX908-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:176
; GFX908-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:48
; GFX908-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:172
; GFX908-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:44
; GFX908-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:168
; GFX908-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:40
; GFX908-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:164
; GFX908-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:36
; GFX908-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:160
; GFX908-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:32
; GFX908-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:156
; GFX908-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:28
; GFX908-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:152
; GFX908-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:24
; GFX908-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:148
; GFX908-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:20
; GFX908-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:144
; GFX908-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:16
; GFX908-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:140
; GFX908-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:12
; GFX908-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:136
; GFX908-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:8
; GFX908-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:132
; GFX908-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:4
; GFX908-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:128
; GFX908-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32
; GFX908-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:124
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v31, v0, s[0:3], 0 offen offset:120
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v30, v0, s[0:3], 0 offen offset:116
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v29, v0, s[0:3], 0 offen offset:112
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v28, v0, s[0:3], 0 offen offset:108
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v27, v0, s[0:3], 0 offen offset:104
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v26, v0, s[0:3], 0 offen offset:100
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v25, v0, s[0:3], 0 offen offset:96
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v24, v0, s[0:3], 0 offen offset:92
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v23, v0, s[0:3], 0 offen offset:88
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v22, v0, s[0:3], 0 offen offset:84
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v21, v0, s[0:3], 0 offen offset:80
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v20, v0, s[0:3], 0 offen offset:76
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v19, v0, s[0:3], 0 offen offset:72
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v18, v0, s[0:3], 0 offen offset:68
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v17, v0, s[0:3], 0 offen offset:64
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v16, v0, s[0:3], 0 offen offset:60
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v15, v0, s[0:3], 0 offen offset:56
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v14, v0, s[0:3], 0 offen offset:52
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v13, v0, s[0:3], 0 offen offset:48
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v12, v0, s[0:3], 0 offen offset:44
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v11, v0, s[0:3], 0 offen offset:40
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v10, v0, s[0:3], 0 offen offset:36
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v9, v0, s[0:3], 0 offen offset:32
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v8, v0, s[0:3], 0 offen offset:28
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v7, v0, s[0:3], 0 offen offset:24
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v6, v0, s[0:3], 0 offen offset:20
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v5, v0, s[0:3], 0 offen offset:16
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v4, v0, s[0:3], 0 offen offset:12
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v3, v0, s[0:3], 0 offen offset:8
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v2, v0, s[0:3], 0 offen offset:4
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
; GFX908-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; GFX908-GCNTRACKERS-NEXT: s_setpc_b64 s[30:31]
;
; GENERIC-LABEL: return_72xi32:
; GENERIC: ; %bb.0:
; GENERIC-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GENERIC-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:160
; GENERIC-NEXT: s_waitcnt vmcnt(0)
; GENERIC-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:284
; GENERIC-NEXT: s_waitcnt expcnt(0)
; GENERIC-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:156
; GENERIC-NEXT: s_waitcnt vmcnt(0)
; GENERIC-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:280
; GENERIC-NEXT: s_waitcnt expcnt(0)
; GENERIC-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:152
; GENERIC-NEXT: s_waitcnt vmcnt(0)
; GENERIC-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:276
; GENERIC-NEXT: s_waitcnt expcnt(0)
; GENERIC-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:148
; GENERIC-NEXT: s_waitcnt vmcnt(0)
; GENERIC-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:272
; GENERIC-NEXT: s_waitcnt expcnt(0)
; GENERIC-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:144
; GENERIC-NEXT: s_waitcnt vmcnt(0)
; GENERIC-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:268
; GENERIC-NEXT: s_waitcnt expcnt(0)
; GENERIC-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:140
; GENERIC-NEXT: s_waitcnt vmcnt(0)
; GENERIC-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:264
; GENERIC-NEXT: s_waitcnt expcnt(0)
; GENERIC-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:136
; GENERIC-NEXT: s_waitcnt vmcnt(0)
; GENERIC-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:260
; GENERIC-NEXT: s_waitcnt expcnt(0)
; GENERIC-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:132
; GENERIC-NEXT: s_waitcnt vmcnt(0)
; GENERIC-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:256
; GENERIC-NEXT: s_waitcnt expcnt(0)
; GENERIC-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:128
; GENERIC-NEXT: s_waitcnt vmcnt(0)
; GENERIC-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:252
; GENERIC-NEXT: s_waitcnt expcnt(0)
; GENERIC-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:124
; GENERIC-NEXT: s_waitcnt vmcnt(0)
; GENERIC-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:248
; GENERIC-NEXT: s_waitcnt expcnt(0)
; GENERIC-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:120
; GENERIC-NEXT: s_waitcnt vmcnt(0)
; GENERIC-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:244
; GENERIC-NEXT: s_waitcnt expcnt(0)
; GENERIC-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:116
; GENERIC-NEXT: s_waitcnt vmcnt(0)
; GENERIC-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:240
; GENERIC-NEXT: s_waitcnt expcnt(0)
; GENERIC-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:112
; GENERIC-NEXT: s_waitcnt vmcnt(0)
; GENERIC-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:236
; GENERIC-NEXT: s_waitcnt expcnt(0)
; GENERIC-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:108
; GENERIC-NEXT: s_waitcnt vmcnt(0)
; GENERIC-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:232
; GENERIC-NEXT: s_waitcnt expcnt(0)
; GENERIC-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:104
; GENERIC-NEXT: s_waitcnt vmcnt(0)
; GENERIC-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:228
; GENERIC-NEXT: s_waitcnt expcnt(0)
; GENERIC-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:100
; GENERIC-NEXT: s_waitcnt vmcnt(0)
; GENERIC-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:224
; GENERIC-NEXT: s_waitcnt expcnt(0)
; GENERIC-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:96
; GENERIC-NEXT: s_waitcnt vmcnt(0)
; GENERIC-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:220
; GENERIC-NEXT: s_waitcnt expcnt(0)
; GENERIC-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:92
; GENERIC-NEXT: s_waitcnt vmcnt(0)
; GENERIC-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:216
; GENERIC-NEXT: s_waitcnt expcnt(0)
; GENERIC-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:88
; GENERIC-NEXT: s_waitcnt vmcnt(0)
; GENERIC-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:212
; GENERIC-NEXT: s_waitcnt expcnt(0)
; GENERIC-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:84
; GENERIC-NEXT: s_waitcnt vmcnt(0)
; GENERIC-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:208
; GENERIC-NEXT: s_waitcnt expcnt(0)
; GENERIC-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:80
; GENERIC-NEXT: s_waitcnt vmcnt(0)
; GENERIC-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:204
; GENERIC-NEXT: s_waitcnt expcnt(0)
; GENERIC-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:76
; GENERIC-NEXT: s_waitcnt vmcnt(0)
; GENERIC-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:200
; GENERIC-NEXT: s_waitcnt expcnt(0)
; GENERIC-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:72
; GENERIC-NEXT: s_waitcnt vmcnt(0)
; GENERIC-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:196
; GENERIC-NEXT: s_waitcnt expcnt(0)
; GENERIC-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:68
; GENERIC-NEXT: s_waitcnt vmcnt(0)
; GENERIC-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:192
; GENERIC-NEXT: s_waitcnt expcnt(0)
; GENERIC-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:64
; GENERIC-NEXT: s_waitcnt vmcnt(0)
; GENERIC-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:188
; GENERIC-NEXT: s_waitcnt expcnt(0)
; GENERIC-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:60
; GENERIC-NEXT: s_waitcnt vmcnt(0)
; GENERIC-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:184
; GENERIC-NEXT: s_waitcnt expcnt(0)
; GENERIC-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:56
; GENERIC-NEXT: s_waitcnt vmcnt(0)
; GENERIC-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:180
; GENERIC-NEXT: s_waitcnt expcnt(0)
; GENERIC-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:52
; GENERIC-NEXT: s_waitcnt vmcnt(0)
; GENERIC-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:176
; GENERIC-NEXT: s_waitcnt expcnt(0)
; GENERIC-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:48
; GENERIC-NEXT: s_waitcnt vmcnt(0)
; GENERIC-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:172
; GENERIC-NEXT: s_waitcnt expcnt(0)
; GENERIC-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:44
; GENERIC-NEXT: s_waitcnt vmcnt(0)
; GENERIC-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:168
; GENERIC-NEXT: s_waitcnt expcnt(0)
; GENERIC-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:40
; GENERIC-NEXT: s_waitcnt vmcnt(0)
; GENERIC-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:164
; GENERIC-NEXT: s_waitcnt expcnt(0)
; GENERIC-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:36
; GENERIC-NEXT: s_waitcnt vmcnt(0)
; GENERIC-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:160
; GENERIC-NEXT: s_waitcnt expcnt(0)
; GENERIC-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:32
; GENERIC-NEXT: s_waitcnt vmcnt(0)
; GENERIC-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:156
; GENERIC-NEXT: s_waitcnt expcnt(0)
; GENERIC-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:28
; GENERIC-NEXT: s_waitcnt vmcnt(0)
; GENERIC-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:152
; GENERIC-NEXT: s_waitcnt expcnt(0)
; GENERIC-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:24
; GENERIC-NEXT: s_waitcnt vmcnt(0)
; GENERIC-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:148
; GENERIC-NEXT: s_waitcnt expcnt(0)
; GENERIC-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:20
; GENERIC-NEXT: s_waitcnt vmcnt(0)
; GENERIC-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:144
; GENERIC-NEXT: s_waitcnt expcnt(0)
; GENERIC-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:16
; GENERIC-NEXT: s_waitcnt vmcnt(0)
; GENERIC-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:140
; GENERIC-NEXT: s_waitcnt expcnt(0)
; GENERIC-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:12
; GENERIC-NEXT: s_waitcnt vmcnt(0)
; GENERIC-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:136
; GENERIC-NEXT: s_waitcnt expcnt(0)
; GENERIC-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:8
; GENERIC-NEXT: s_waitcnt vmcnt(0)
; GENERIC-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:132
; GENERIC-NEXT: s_waitcnt expcnt(0)
; GENERIC-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:4
; GENERIC-NEXT: s_waitcnt vmcnt(0)
; GENERIC-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:128
; GENERIC-NEXT: s_waitcnt expcnt(0)
; GENERIC-NEXT: buffer_load_dword v32, off, s[0:3], s32
; GENERIC-NEXT: s_waitcnt vmcnt(0)
; GENERIC-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:124
; GENERIC-NEXT: buffer_store_dword v31, v0, s[0:3], 0 offen offset:120
; GENERIC-NEXT: buffer_store_dword v30, v0, s[0:3], 0 offen offset:116
; GENERIC-NEXT: buffer_store_dword v29, v0, s[0:3], 0 offen offset:112
; GENERIC-NEXT: buffer_store_dword v28, v0, s[0:3], 0 offen offset:108
; GENERIC-NEXT: buffer_store_dword v27, v0, s[0:3], 0 offen offset:104
; GENERIC-NEXT: buffer_store_dword v26, v0, s[0:3], 0 offen offset:100
; GENERIC-NEXT: buffer_store_dword v25, v0, s[0:3], 0 offen offset:96
; GENERIC-NEXT: buffer_store_dword v24, v0, s[0:3], 0 offen offset:92
; GENERIC-NEXT: buffer_store_dword v23, v0, s[0:3], 0 offen offset:88
; GENERIC-NEXT: buffer_store_dword v22, v0, s[0:3], 0 offen offset:84
; GENERIC-NEXT: buffer_store_dword v21, v0, s[0:3], 0 offen offset:80
; GENERIC-NEXT: buffer_store_dword v20, v0, s[0:3], 0 offen offset:76
; GENERIC-NEXT: buffer_store_dword v19, v0, s[0:3], 0 offen offset:72
; GENERIC-NEXT: buffer_store_dword v18, v0, s[0:3], 0 offen offset:68
; GENERIC-NEXT: buffer_store_dword v17, v0, s[0:3], 0 offen offset:64
; GENERIC-NEXT: buffer_store_dword v16, v0, s[0:3], 0 offen offset:60
; GENERIC-NEXT: buffer_store_dword v15, v0, s[0:3], 0 offen offset:56
; GENERIC-NEXT: buffer_store_dword v14, v0, s[0:3], 0 offen offset:52
; GENERIC-NEXT: buffer_store_dword v13, v0, s[0:3], 0 offen offset:48
; GENERIC-NEXT: buffer_store_dword v12, v0, s[0:3], 0 offen offset:44
; GENERIC-NEXT: buffer_store_dword v11, v0, s[0:3], 0 offen offset:40
; GENERIC-NEXT: buffer_store_dword v10, v0, s[0:3], 0 offen offset:36
; GENERIC-NEXT: buffer_store_dword v9, v0, s[0:3], 0 offen offset:32
; GENERIC-NEXT: buffer_store_dword v8, v0, s[0:3], 0 offen offset:28
; GENERIC-NEXT: buffer_store_dword v7, v0, s[0:3], 0 offen offset:24
; GENERIC-NEXT: buffer_store_dword v6, v0, s[0:3], 0 offen offset:20
; GENERIC-NEXT: buffer_store_dword v5, v0, s[0:3], 0 offen offset:16
; GENERIC-NEXT: buffer_store_dword v4, v0, s[0:3], 0 offen offset:12
; GENERIC-NEXT: buffer_store_dword v3, v0, s[0:3], 0 offen offset:8
; GENERIC-NEXT: buffer_store_dword v2, v0, s[0:3], 0 offen offset:4
; GENERIC-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
; GENERIC-NEXT: s_waitcnt vmcnt(0) expcnt(0)
; GENERIC-NEXT: s_setpc_b64 s[30:31]
;
; GENERIC-GCNTRACKERS-LABEL: return_72xi32:
; GENERIC-GCNTRACKERS: ; %bb.0:
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:160
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:284
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt expcnt(0)
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:156
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:280
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt expcnt(0)
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:152
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:276
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt expcnt(0)
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:148
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:272
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt expcnt(0)
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:144
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:268
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt expcnt(0)
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:140
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:264
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt expcnt(0)
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:136
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:260
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt expcnt(0)
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:132
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:256
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt expcnt(0)
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:128
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:252
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt expcnt(0)
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:124
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:248
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt expcnt(0)
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:120
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:244
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt expcnt(0)
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:116
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:240
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt expcnt(0)
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:112
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:236
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt expcnt(0)
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:108
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:232
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt expcnt(0)
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:104
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:228
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt expcnt(0)
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:100
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:224
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt expcnt(0)
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:96
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:220
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt expcnt(0)
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:92
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:216
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt expcnt(0)
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:88
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:212
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt expcnt(0)
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:84
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:208
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt expcnt(0)
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:80
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:204
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt expcnt(0)
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:76
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:200
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt expcnt(0)
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:72
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:196
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt expcnt(0)
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:68
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:192
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt expcnt(0)
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:64
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:188
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt expcnt(0)
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:60
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:184
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt expcnt(0)
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:56
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:180
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt expcnt(0)
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:52
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:176
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt expcnt(0)
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:48
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:172
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt expcnt(0)
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:44
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:168
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt expcnt(0)
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:40
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:164
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt expcnt(0)
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:36
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:160
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt expcnt(0)
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:32
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:156
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt expcnt(0)
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:28
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:152
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt expcnt(0)
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:24
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:148
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt expcnt(0)
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:20
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:144
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt expcnt(0)
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:16
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:140
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt expcnt(0)
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:12
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:136
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt expcnt(0)
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:8
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:132
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt expcnt(0)
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:4
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:128
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt expcnt(0)
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s32
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:124
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v31, v0, s[0:3], 0 offen offset:120
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v30, v0, s[0:3], 0 offen offset:116
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v29, v0, s[0:3], 0 offen offset:112
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v28, v0, s[0:3], 0 offen offset:108
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v27, v0, s[0:3], 0 offen offset:104
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v26, v0, s[0:3], 0 offen offset:100
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v25, v0, s[0:3], 0 offen offset:96
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v24, v0, s[0:3], 0 offen offset:92
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v23, v0, s[0:3], 0 offen offset:88
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v22, v0, s[0:3], 0 offen offset:84
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v21, v0, s[0:3], 0 offen offset:80
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v20, v0, s[0:3], 0 offen offset:76
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v19, v0, s[0:3], 0 offen offset:72
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v18, v0, s[0:3], 0 offen offset:68
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v17, v0, s[0:3], 0 offen offset:64
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v16, v0, s[0:3], 0 offen offset:60
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v15, v0, s[0:3], 0 offen offset:56
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v14, v0, s[0:3], 0 offen offset:52
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v13, v0, s[0:3], 0 offen offset:48
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v12, v0, s[0:3], 0 offen offset:44
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v11, v0, s[0:3], 0 offen offset:40
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v10, v0, s[0:3], 0 offen offset:36
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v9, v0, s[0:3], 0 offen offset:32
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v8, v0, s[0:3], 0 offen offset:28
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v7, v0, s[0:3], 0 offen offset:24
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v6, v0, s[0:3], 0 offen offset:20
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v5, v0, s[0:3], 0 offen offset:16
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v4, v0, s[0:3], 0 offen offset:12
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v3, v0, s[0:3], 0 offen offset:8
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v2, v0, s[0:3], 0 offen offset:4
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0) expcnt(0)
; GENERIC-GCNTRACKERS-NEXT: s_setpc_b64 s[30:31]
ret <72 x i32> %val
}
define amdgpu_gfx void @call_72xi32() #1 {
; GFX11-PAL-LABEL: call_72xi32:
; GFX11-PAL: ; %bb.0: ; %entry
; GFX11-PAL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-PAL-NEXT: s_mov_b32 s38, s33
; GFX11-PAL-NEXT: s_add_i32 s33, s32, 0x1ff
; GFX11-PAL-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-PAL-NEXT: s_and_b32 s33, s33, 0xfffffe00
; GFX11-PAL-NEXT: s_or_saveexec_b32 s0, -1
; GFX11-PAL-NEXT: scratch_store_b32 off, v62, s33 offset:1600 ; 4-byte Folded Spill
; GFX11-PAL-NEXT: s_mov_b32 exec_lo, s0
; GFX11-PAL-NEXT: s_mov_b32 s39, s34
; GFX11-PAL-NEXT: s_mov_b32 s34, s32
; GFX11-PAL-NEXT: s_addk_i32 s32, 0xa00
; GFX11-PAL-NEXT: s_clause 0xd ; 56-byte Folded Spill
; GFX11-PAL-NEXT: scratch_store_b32 off, v40, s33 offset:52
; GFX11-PAL-NEXT: ; meta instruction
; GFX11-PAL-NEXT: scratch_store_b32 off, v41, s33 offset:48
; GFX11-PAL-NEXT: ; meta instruction
; GFX11-PAL-NEXT: scratch_store_b32 off, v42, s33 offset:44
; GFX11-PAL-NEXT: ; meta instruction
; GFX11-PAL-NEXT: scratch_store_b32 off, v43, s33 offset:40
; GFX11-PAL-NEXT: ; meta instruction
; GFX11-PAL-NEXT: scratch_store_b32 off, v44, s33 offset:36
; GFX11-PAL-NEXT: ; meta instruction
; GFX11-PAL-NEXT: scratch_store_b32 off, v45, s33 offset:32
; GFX11-PAL-NEXT: ; meta instruction
; GFX11-PAL-NEXT: scratch_store_b32 off, v46, s33 offset:28
; GFX11-PAL-NEXT: ; meta instruction
; GFX11-PAL-NEXT: scratch_store_b32 off, v47, s33 offset:24
; GFX11-PAL-NEXT: ; meta instruction
; GFX11-PAL-NEXT: scratch_store_b32 off, v56, s33 offset:20
; GFX11-PAL-NEXT: ; meta instruction
; GFX11-PAL-NEXT: scratch_store_b32 off, v57, s33 offset:16
; GFX11-PAL-NEXT: ; meta instruction
; GFX11-PAL-NEXT: scratch_store_b32 off, v58, s33 offset:12
; GFX11-PAL-NEXT: ; meta instruction
; GFX11-PAL-NEXT: scratch_store_b32 off, v59, s33 offset:8
; GFX11-PAL-NEXT: ; meta instruction
; GFX11-PAL-NEXT: scratch_store_b32 off, v60, s33 offset:4
; GFX11-PAL-NEXT: ; meta instruction
; GFX11-PAL-NEXT: scratch_store_b32 off, v61, s33
; GFX11-PAL-NEXT: v_writelane_b32 v62, s30, 0
; GFX11-PAL-NEXT: v_writelane_b32 v62, s31, 1
; GFX11-PAL-NEXT: s_mov_b32 s0, 0
; GFX11-PAL-NEXT: v_mov_b32_e32 v4, 0
; GFX11-PAL-NEXT: s_mov_b32 s1, s0
; GFX11-PAL-NEXT: s_mov_b32 s2, s0
; GFX11-PAL-NEXT: s_mov_b32 s3, s0
; GFX11-PAL-NEXT: v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-PAL-NEXT: v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
; GFX11-PAL-NEXT: s_add_i32 s0, s32, 0xa0
; GFX11-PAL-NEXT: s_add_i32 s1, s32, 0x90
; GFX11-PAL-NEXT: scratch_store_b128 off, v[0:3], s32
; GFX11-PAL-NEXT: scratch_store_b32 off, v4, s0
; GFX11-PAL-NEXT: scratch_store_b128 off, v[0:3], s1
; GFX11-PAL-NEXT: s_add_i32 s0, s32, 0x80
; GFX11-PAL-NEXT: s_add_i32 s1, s32, 0x70
; GFX11-PAL-NEXT: scratch_store_b128 off, v[0:3], s0
; GFX11-PAL-NEXT: scratch_store_b128 off, v[0:3], s1
; GFX11-PAL-NEXT: s_add_i32 s0, s32, 0x60
; GFX11-PAL-NEXT: s_add_i32 s1, s32, 0x50
; GFX11-PAL-NEXT: scratch_store_b128 off, v[0:3], s0
; GFX11-PAL-NEXT: scratch_store_b128 off, v[0:3], s1
; GFX11-PAL-NEXT: s_add_i32 s0, s32, 64
; GFX11-PAL-NEXT: s_add_i32 s1, s32, 48
; GFX11-PAL-NEXT: scratch_store_b128 off, v[0:3], s0
; GFX11-PAL-NEXT: scratch_store_b128 off, v[0:3], s1
; GFX11-PAL-NEXT: s_add_i32 s0, s32, 32
; GFX11-PAL-NEXT: s_add_i32 s1, s32, 16
; GFX11-PAL-NEXT: s_add_i32 s2, s33, 0x200
; GFX11-PAL-NEXT: scratch_store_b128 off, v[0:3], s0
; GFX11-PAL-NEXT: scratch_store_b128 off, v[0:3], s1
; GFX11-PAL-NEXT: v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v3, 0
; GFX11-PAL-NEXT: v_dual_mov_b32 v1, 0 :: v_dual_mov_b32 v2, 0
; GFX11-PAL-NEXT: v_dual_mov_b32 v5, 0 :: v_dual_mov_b32 v4, 0
; GFX11-PAL-NEXT: v_dual_mov_b32 v7, 0 :: v_dual_mov_b32 v6, 0
; GFX11-PAL-NEXT: v_dual_mov_b32 v9, 0 :: v_dual_mov_b32 v8, 0
; GFX11-PAL-NEXT: v_dual_mov_b32 v11, 0 :: v_dual_mov_b32 v10, 0
; GFX11-PAL-NEXT: v_dual_mov_b32 v13, 0 :: v_dual_mov_b32 v12, 0
; GFX11-PAL-NEXT: v_dual_mov_b32 v15, 0 :: v_dual_mov_b32 v14, 0
; GFX11-PAL-NEXT: v_dual_mov_b32 v17, 0 :: v_dual_mov_b32 v16, 0
; GFX11-PAL-NEXT: v_dual_mov_b32 v19, 0 :: v_dual_mov_b32 v18, 0
; GFX11-PAL-NEXT: v_dual_mov_b32 v21, 0 :: v_dual_mov_b32 v20, 0
; GFX11-PAL-NEXT: v_dual_mov_b32 v23, 0 :: v_dual_mov_b32 v22, 0
; GFX11-PAL-NEXT: v_dual_mov_b32 v25, 0 :: v_dual_mov_b32 v24, 0
; GFX11-PAL-NEXT: v_dual_mov_b32 v27, 0 :: v_dual_mov_b32 v26, 0
; GFX11-PAL-NEXT: v_dual_mov_b32 v29, 0 :: v_dual_mov_b32 v28, 0
; GFX11-PAL-NEXT: v_dual_mov_b32 v31, 0 :: v_dual_mov_b32 v30, 0
; GFX11-PAL-NEXT: s_mov_b32 s1, return_72xi32@abs32@hi
; GFX11-PAL-NEXT: s_mov_b32 s0, return_72xi32@abs32@lo
; GFX11-PAL-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-PAL-NEXT: s_swappc_b64 s[30:31], s[0:1]
; GFX11-PAL-NEXT: s_clause 0xb
; GFX11-PAL-NEXT: scratch_load_b128 v[43:46], off, s33 offset:624
; GFX11-PAL-NEXT: scratch_load_b128 v[47:50], off, s33 offset:640
; GFX11-PAL-NEXT: scratch_load_b128 v[16:19], off, s33 offset:784
; GFX11-PAL-NEXT: scratch_load_b128 v[12:15], off, s33 offset:768
; GFX11-PAL-NEXT: scratch_load_b128 v[8:11], off, s33 offset:752
; GFX11-PAL-NEXT: scratch_load_b128 v[4:7], off, s33 offset:736
; GFX11-PAL-NEXT: scratch_load_b128 v[0:3], off, s33 offset:720
; GFX11-PAL-NEXT: scratch_load_b128 v[24:27], off, s33 offset:656
; GFX11-PAL-NEXT: scratch_load_b128 v[36:39], off, s33 offset:704
; GFX11-PAL-NEXT: scratch_load_b128 v[32:35], off, s33 offset:688
; GFX11-PAL-NEXT: scratch_load_b128 v[28:31], off, s33 offset:672
; GFX11-PAL-NEXT: scratch_load_b128 v[20:23], off, s33 offset:512
; GFX11-PAL-NEXT: s_waitcnt vmcnt(6)
; GFX11-PAL-NEXT: v_mov_b32_e32 v6, 24
; GFX11-PAL-NEXT: s_add_i32 s2, s32, 0xa0
; GFX11-PAL-NEXT: s_add_i32 s3, s32, 0x90
; GFX11-PAL-NEXT: s_add_i32 s35, s32, 0x80
; GFX11-PAL-NEXT: s_add_i32 s36, s32, 0x70
; GFX11-PAL-NEXT: s_add_i32 s37, s32, 0x6c
; GFX11-PAL-NEXT: s_waitcnt vmcnt(4)
; GFX11-PAL-NEXT: v_dual_mov_b32 v40, v0 :: v_dual_mov_b32 v53, v26
; GFX11-PAL-NEXT: v_dual_mov_b32 v41, v1 :: v_dual_mov_b32 v42, v2
; GFX11-PAL-NEXT: v_mov_b32_e32 v51, v24
; GFX11-PAL-NEXT: s_waitcnt vmcnt(0)
; GFX11-PAL-NEXT: scratch_store_b128 off, v[20:23], s33 offset:1584 ; 16-byte Folded Spill
; GFX11-PAL-NEXT: scratch_load_b128 v[20:23], off, s33 offset:528
; GFX11-PAL-NEXT: v_mov_b32_e32 v52, v25
; GFX11-PAL-NEXT: s_waitcnt vmcnt(0)
; GFX11-PAL-NEXT: scratch_store_b128 off, v[20:23], s33 offset:1536 ; 16-byte Folded Spill
; GFX11-PAL-NEXT: scratch_load_b128 v[20:23], off, s33 offset:544
; GFX11-PAL-NEXT: s_waitcnt vmcnt(0)
; GFX11-PAL-NEXT: scratch_store_b128 off, v[20:23], s33 offset:1568 ; 16-byte Folded Spill
; GFX11-PAL-NEXT: scratch_load_b128 v[20:23], off, s33 offset:560
; GFX11-PAL-NEXT: s_waitcnt vmcnt(0)
; GFX11-PAL-NEXT: scratch_store_b128 off, v[20:23], s33 offset:1552 ; 16-byte Folded Spill
; GFX11-PAL-NEXT: s_clause 0x2
; GFX11-PAL-NEXT: scratch_load_b128 v[20:23], off, s33 offset:576
; GFX11-PAL-NEXT: scratch_load_b128 v[58:61], off, s33 offset:592
; GFX11-PAL-NEXT: scratch_load_b128 v[54:57], off, s33 offset:608
; GFX11-PAL-NEXT: scratch_store_b128 off, v[46:49], s32
; GFX11-PAL-NEXT: scratch_store_b32 off, v19, s2
; GFX11-PAL-NEXT: scratch_store_b128 off, v[15:18], s3
; GFX11-PAL-NEXT: scratch_store_b128 off, v[11:14], s35
; GFX11-PAL-NEXT: scratch_store_b128 off, v[7:10], s36
; GFX11-PAL-NEXT: s_add_i32 s2, s32, 0x60
; GFX11-PAL-NEXT: scratch_store_b32 off, v6, s37
; GFX11-PAL-NEXT: scratch_store_b96 off, v[3:5], s2
; GFX11-PAL-NEXT: s_add_i32 s2, s32, 0x50
; GFX11-PAL-NEXT: s_add_i32 s3, s32, 64
; GFX11-PAL-NEXT: s_add_i32 s35, s32, 48
; GFX11-PAL-NEXT: s_add_i32 s36, s32, 32
; GFX11-PAL-NEXT: scratch_store_b128 off, v[39:42], s2
; GFX11-PAL-NEXT: scratch_store_b128 off, v[35:38], s3
; GFX11-PAL-NEXT: scratch_store_b128 off, v[31:34], s35
; GFX11-PAL-NEXT: scratch_store_b128 off, v[27:30], s36
; GFX11-PAL-NEXT: s_add_i32 s2, s32, 16
; GFX11-PAL-NEXT: v_mov_b32_e32 v29, v43
; GFX11-PAL-NEXT: scratch_store_b128 off, v[50:53], s2
; GFX11-PAL-NEXT: s_clause 0x3 ; 64-byte Folded Reload
; GFX11-PAL-NEXT: scratch_load_b128 v[1:4], off, s33 offset:1584
; GFX11-PAL-NEXT: scratch_load_b128 v[5:8], off, s33 offset:1536
; GFX11-PAL-NEXT: scratch_load_b128 v[9:12], off, s33 offset:1568
; GFX11-PAL-NEXT: scratch_load_b128 v[13:16], off, s33 offset:1552
; GFX11-PAL-NEXT: s_add_i32 s2, s33, 0x400
; GFX11-PAL-NEXT: v_dual_mov_b32 v30, v44 :: v_dual_mov_b32 v31, v45
; GFX11-PAL-NEXT: s_waitcnt vmcnt(3)
; GFX11-PAL-NEXT: v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, 42
; GFX11-PAL-NEXT: v_dual_mov_b32 v17, v20 :: v_dual_mov_b32 v18, v21
; GFX11-PAL-NEXT: v_dual_mov_b32 v19, v22 :: v_dual_mov_b32 v20, v23
; GFX11-PAL-NEXT: v_dual_mov_b32 v21, v58 :: v_dual_mov_b32 v22, v59
; GFX11-PAL-NEXT: v_dual_mov_b32 v23, v60 :: v_dual_mov_b32 v24, v61
; GFX11-PAL-NEXT: v_dual_mov_b32 v25, v54 :: v_dual_mov_b32 v26, v55
; GFX11-PAL-NEXT: v_dual_mov_b32 v27, v56 :: v_dual_mov_b32 v28, v57
; GFX11-PAL-NEXT: s_swappc_b64 s[30:31], s[0:1]
; GFX11-PAL-NEXT: s_clause 0xd ; 56-byte Folded Reload
; GFX11-PAL-NEXT: scratch_load_b32 v61, off, s33
; GFX11-PAL-NEXT: scratch_load_b32 v60, off, s33 offset:4
; GFX11-PAL-NEXT: scratch_load_b32 v59, off, s33 offset:8
; GFX11-PAL-NEXT: scratch_load_b32 v58, off, s33 offset:12
; GFX11-PAL-NEXT: scratch_load_b32 v57, off, s33 offset:16
; GFX11-PAL-NEXT: scratch_load_b32 v56, off, s33 offset:20
; GFX11-PAL-NEXT: scratch_load_b32 v47, off, s33 offset:24
; GFX11-PAL-NEXT: scratch_load_b32 v46, off, s33 offset:28
; GFX11-PAL-NEXT: scratch_load_b32 v45, off, s33 offset:32
; GFX11-PAL-NEXT: scratch_load_b32 v44, off, s33 offset:36
; GFX11-PAL-NEXT: scratch_load_b32 v43, off, s33 offset:40
; GFX11-PAL-NEXT: scratch_load_b32 v42, off, s33 offset:44
; GFX11-PAL-NEXT: scratch_load_b32 v41, off, s33 offset:48
; GFX11-PAL-NEXT: scratch_load_b32 v40, off, s33 offset:52
; GFX11-PAL-NEXT: v_readlane_b32 s30, v62, 0
; GFX11-PAL-NEXT: v_readlane_b32 s31, v62, 1
; GFX11-PAL-NEXT: s_mov_b32 s32, s34
; GFX11-PAL-NEXT: s_mov_b32 s34, s39
; GFX11-PAL-NEXT: s_or_saveexec_b32 s0, -1
; GFX11-PAL-NEXT: scratch_load_b32 v62, off, s33 offset:1600 ; 4-byte Folded Reload
; GFX11-PAL-NEXT: s_mov_b32 exec_lo, s0
; GFX11-PAL-NEXT: s_mov_b32 s33, s38
; GFX11-PAL-NEXT: s_waitcnt vmcnt(0)
; GFX11-PAL-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-PAL-GCNTRACKERS-LABEL: call_72xi32:
; GFX11-PAL-GCNTRACKERS: ; %bb.0: ; %entry
; GFX11-PAL-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-PAL-GCNTRACKERS-NEXT: s_mov_b32 s35, s33
; GFX11-PAL-GCNTRACKERS-NEXT: s_add_i32 s33, s32, 0x1ff
; GFX11-PAL-GCNTRACKERS-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-PAL-GCNTRACKERS-NEXT: s_and_b32 s33, s33, 0xfffffe00
; GFX11-PAL-GCNTRACKERS-NEXT: s_or_saveexec_b32 s0, -1
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_store_b32 off, v61, s33 offset:1916 ; 4-byte Folded Spill
; GFX11-PAL-GCNTRACKERS-NEXT: s_mov_b32 exec_lo, s0
; GFX11-PAL-GCNTRACKERS-NEXT: s_mov_b32 s36, s34
; GFX11-PAL-GCNTRACKERS-NEXT: s_mov_b32 s34, s32
; GFX11-PAL-GCNTRACKERS-NEXT: s_addk_i32 s32, 0xa00
; GFX11-PAL-GCNTRACKERS-NEXT: s_clause 0xc ; 52-byte Folded Spill
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_store_b32 off, v40, s33 offset:48
; GFX11-PAL-GCNTRACKERS-NEXT: ; meta instruction
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_store_b32 off, v41, s33 offset:44
; GFX11-PAL-GCNTRACKERS-NEXT: ; meta instruction
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_store_b32 off, v42, s33 offset:40
; GFX11-PAL-GCNTRACKERS-NEXT: ; meta instruction
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_store_b32 off, v43, s33 offset:36
; GFX11-PAL-GCNTRACKERS-NEXT: ; meta instruction
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_store_b32 off, v44, s33 offset:32
; GFX11-PAL-GCNTRACKERS-NEXT: ; meta instruction
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_store_b32 off, v45, s33 offset:28
; GFX11-PAL-GCNTRACKERS-NEXT: ; meta instruction
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_store_b32 off, v46, s33 offset:24
; GFX11-PAL-GCNTRACKERS-NEXT: ; meta instruction
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_store_b32 off, v47, s33 offset:20
; GFX11-PAL-GCNTRACKERS-NEXT: ; meta instruction
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_store_b32 off, v56, s33 offset:16
; GFX11-PAL-GCNTRACKERS-NEXT: ; meta instruction
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_store_b32 off, v57, s33 offset:12
; GFX11-PAL-GCNTRACKERS-NEXT: ; meta instruction
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_store_b32 off, v58, s33 offset:8
; GFX11-PAL-GCNTRACKERS-NEXT: ; meta instruction
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_store_b32 off, v59, s33 offset:4
; GFX11-PAL-GCNTRACKERS-NEXT: ; meta instruction
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_store_b32 off, v60, s33
; GFX11-PAL-GCNTRACKERS-NEXT: v_writelane_b32 v61, s30, 0
; GFX11-PAL-GCNTRACKERS-NEXT: v_writelane_b32 v61, s31, 1
; GFX11-PAL-GCNTRACKERS-NEXT: s_mov_b32 s0, 0
; GFX11-PAL-GCNTRACKERS-NEXT: v_mov_b32_e32 v4, 0
; GFX11-PAL-GCNTRACKERS-NEXT: s_mov_b32 s1, s0
; GFX11-PAL-GCNTRACKERS-NEXT: s_mov_b32 s2, s0
; GFX11-PAL-GCNTRACKERS-NEXT: s_mov_b32 s3, s0
; GFX11-PAL-GCNTRACKERS-NEXT: v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-PAL-GCNTRACKERS-NEXT: v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
; GFX11-PAL-GCNTRACKERS-NEXT: s_add_i32 s0, s32, 0xa0
; GFX11-PAL-GCNTRACKERS-NEXT: s_add_i32 s1, s32, 0x90
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_store_b128 off, v[0:3], s32
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_store_b32 off, v4, s0
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_store_b128 off, v[0:3], s1
; GFX11-PAL-GCNTRACKERS-NEXT: s_add_i32 s0, s32, 0x80
; GFX11-PAL-GCNTRACKERS-NEXT: s_add_i32 s1, s32, 0x70
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_store_b128 off, v[0:3], s0
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_store_b128 off, v[0:3], s1
; GFX11-PAL-GCNTRACKERS-NEXT: s_add_i32 s0, s32, 0x60
; GFX11-PAL-GCNTRACKERS-NEXT: s_add_i32 s1, s32, 0x50
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_store_b128 off, v[0:3], s0
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_store_b128 off, v[0:3], s1
; GFX11-PAL-GCNTRACKERS-NEXT: s_add_i32 s0, s32, 64
; GFX11-PAL-GCNTRACKERS-NEXT: s_add_i32 s1, s32, 48
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_store_b128 off, v[0:3], s0
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_store_b128 off, v[0:3], s1
; GFX11-PAL-GCNTRACKERS-NEXT: s_add_i32 s0, s32, 32
; GFX11-PAL-GCNTRACKERS-NEXT: s_add_i32 s1, s32, 16
; GFX11-PAL-GCNTRACKERS-NEXT: s_add_i32 s2, s33, 0x200
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_store_b128 off, v[0:3], s0
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_store_b128 off, v[0:3], s1
; GFX11-PAL-GCNTRACKERS-NEXT: v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v3, 0
; GFX11-PAL-GCNTRACKERS-NEXT: v_dual_mov_b32 v1, 0 :: v_dual_mov_b32 v2, 0
; GFX11-PAL-GCNTRACKERS-NEXT: v_dual_mov_b32 v5, 0 :: v_dual_mov_b32 v4, 0
; GFX11-PAL-GCNTRACKERS-NEXT: v_dual_mov_b32 v7, 0 :: v_dual_mov_b32 v6, 0
; GFX11-PAL-GCNTRACKERS-NEXT: v_dual_mov_b32 v9, 0 :: v_dual_mov_b32 v8, 0
; GFX11-PAL-GCNTRACKERS-NEXT: v_dual_mov_b32 v11, 0 :: v_dual_mov_b32 v10, 0
; GFX11-PAL-GCNTRACKERS-NEXT: v_dual_mov_b32 v13, 0 :: v_dual_mov_b32 v12, 0
; GFX11-PAL-GCNTRACKERS-NEXT: v_dual_mov_b32 v15, 0 :: v_dual_mov_b32 v14, 0
; GFX11-PAL-GCNTRACKERS-NEXT: v_dual_mov_b32 v17, 0 :: v_dual_mov_b32 v16, 0
; GFX11-PAL-GCNTRACKERS-NEXT: v_dual_mov_b32 v19, 0 :: v_dual_mov_b32 v18, 0
; GFX11-PAL-GCNTRACKERS-NEXT: v_dual_mov_b32 v21, 0 :: v_dual_mov_b32 v20, 0
; GFX11-PAL-GCNTRACKERS-NEXT: v_dual_mov_b32 v23, 0 :: v_dual_mov_b32 v22, 0
; GFX11-PAL-GCNTRACKERS-NEXT: v_dual_mov_b32 v25, 0 :: v_dual_mov_b32 v24, 0
; GFX11-PAL-GCNTRACKERS-NEXT: v_dual_mov_b32 v27, 0 :: v_dual_mov_b32 v26, 0
; GFX11-PAL-GCNTRACKERS-NEXT: v_dual_mov_b32 v29, 0 :: v_dual_mov_b32 v28, 0
; GFX11-PAL-GCNTRACKERS-NEXT: v_dual_mov_b32 v31, 0 :: v_dual_mov_b32 v30, 0
; GFX11-PAL-GCNTRACKERS-NEXT: s_mov_b32 s1, return_72xi32@abs32@hi
; GFX11-PAL-GCNTRACKERS-NEXT: s_mov_b32 s0, return_72xi32@abs32@lo
; GFX11-PAL-GCNTRACKERS-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-PAL-GCNTRACKERS-NEXT: s_swappc_b64 s[30:31], s[0:1]
; GFX11-PAL-GCNTRACKERS-NEXT: s_clause 0x1
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b128 v[0:3], off, s33 offset:624
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b128 v[4:7], off, s33 offset:640
; GFX11-PAL-GCNTRACKERS-NEXT: s_add_i32 s2, s32, 0xa0
; GFX11-PAL-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; GFX11-PAL-GCNTRACKERS-NEXT: v_dual_mov_b32 v32, v7 :: v_dual_mov_b32 v31, v6
; GFX11-PAL-GCNTRACKERS-NEXT: v_dual_mov_b32 v30, v5 :: v_dual_mov_b32 v29, v4
; GFX11-PAL-GCNTRACKERS-NEXT: v_dual_mov_b32 v28, v3 :: v_dual_mov_b32 v27, v2
; GFX11-PAL-GCNTRACKERS-NEXT: v_dual_mov_b32 v26, v1 :: v_dual_mov_b32 v25, v0
; GFX11-PAL-GCNTRACKERS-NEXT: s_clause 0x5
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b128 v[9:12], off, s33 offset:736
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b128 v[1:4], off, s33 offset:512
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b128 v[21:24], off, s33 offset:784
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b128 v[17:20], off, s33 offset:768
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b128 v[13:16], off, s33 offset:752
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b128 v[5:8], off, s33 offset:720
; GFX11-PAL-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; GFX11-PAL-GCNTRACKERS-NEXT: s_clause 0x7 ; 128-byte Folded Spill
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_store_b128 off, v[5:8], s33 offset:1536
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_store_b128 off, v[9:12], s33 offset:1552
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_store_b128 off, v[13:16], s33 offset:1568
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_store_b128 off, v[17:20], s33 offset:1584
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_store_b128 off, v[21:24], s33 offset:1600
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_store_b128 off, v[25:28], s33 offset:1616
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_store_b128 off, v[29:32], s33 offset:1632
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_store_b128 off, v[33:36], s33 offset:1648
; GFX11-PAL-GCNTRACKERS-NEXT: s_clause 0x1
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b128 v[12:15], off, s33 offset:704
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b128 v[8:11], off, s33 offset:688
; GFX11-PAL-GCNTRACKERS-NEXT: v_dual_mov_b32 v0, 24 :: v_dual_mov_b32 v1, 42
; GFX11-PAL-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; GFX11-PAL-GCNTRACKERS-NEXT: v_dual_mov_b32 v16, v15 :: v_dual_mov_b32 v15, v14
; GFX11-PAL-GCNTRACKERS-NEXT: v_mov_b32_e32 v14, v13
; GFX11-PAL-GCNTRACKERS-NEXT: v_mov_b32_e32 v13, v12
; GFX11-PAL-GCNTRACKERS-NEXT: v_mov_b32_e32 v12, v11
; GFX11-PAL-GCNTRACKERS-NEXT: v_mov_b32_e32 v11, v10
; GFX11-PAL-GCNTRACKERS-NEXT: v_mov_b32_e32 v10, v9
; GFX11-PAL-GCNTRACKERS-NEXT: v_mov_b32_e32 v9, v8
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b128 v[5:8], off, s33 offset:672
; GFX11-PAL-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; GFX11-PAL-GCNTRACKERS-NEXT: v_dual_mov_b32 v20, v16 :: v_dual_mov_b32 v19, v15
; GFX11-PAL-GCNTRACKERS-NEXT: v_dual_mov_b32 v18, v14 :: v_dual_mov_b32 v17, v13
; GFX11-PAL-GCNTRACKERS-NEXT: v_mov_b32_e32 v16, v12
; GFX11-PAL-GCNTRACKERS-NEXT: v_mov_b32_e32 v15, v11
; GFX11-PAL-GCNTRACKERS-NEXT: v_mov_b32_e32 v14, v10
; GFX11-PAL-GCNTRACKERS-NEXT: v_mov_b32_e32 v13, v9
; GFX11-PAL-GCNTRACKERS-NEXT: v_mov_b32_e32 v12, v8
; GFX11-PAL-GCNTRACKERS-NEXT: v_mov_b32_e32 v11, v7
; GFX11-PAL-GCNTRACKERS-NEXT: v_mov_b32_e32 v10, v6
; GFX11-PAL-GCNTRACKERS-NEXT: v_mov_b32_e32 v9, v5
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b128 v[5:8], off, s33 offset:656
; GFX11-PAL-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; GFX11-PAL-GCNTRACKERS-NEXT: s_clause 0x7 ; 128-byte Folded Spill
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_store_b128 off, v[5:8], s33 offset:1744
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_store_b128 off, v[9:12], s33 offset:1760
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_store_b128 off, v[13:16], s33 offset:1776
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_store_b128 off, v[17:20], s33 offset:1792
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_store_b128 off, v[21:24], s33 offset:1808
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_store_b128 off, v[25:28], s33 offset:1824
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_store_b128 off, v[29:32], s33 offset:1840
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_store_b128 off, v[33:36], s33 offset:1856
; GFX11-PAL-GCNTRACKERS-NEXT: s_clause 0x1
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b128 v[5:8], off, s33 offset:528
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b128 v[57:60], off, s33 offset:608
; GFX11-PAL-GCNTRACKERS-NEXT: s_waitcnt vmcnt(1)
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_store_b128 off, v[5:8], s33 offset:1728 ; 16-byte Folded Spill
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b128 v[5:8], off, s33 offset:544
; GFX11-PAL-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_store_b128 off, v[5:8], s33 offset:1712 ; 16-byte Folded Spill
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b128 v[5:8], off, s33 offset:560
; GFX11-PAL-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_store_b128 off, v[5:8], s33 offset:1696 ; 16-byte Folded Spill
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b128 v[5:8], off, s33 offset:576
; GFX11-PAL-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_store_b128 off, v[5:8], s33 offset:1680 ; 16-byte Folded Spill
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b128 v[5:8], off, s33 offset:592
; GFX11-PAL-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_store_b128 off, v[5:8], s33 offset:1664 ; 16-byte Folded Spill
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_store_b128 off, v[28:31], s32
; GFX11-PAL-GCNTRACKERS-NEXT: s_clause 0x2 ; 44-byte Folded Spill
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_store_b128 off, v[25:28], s33 offset:1872
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_store_b128 off, v[29:32], s33 offset:1888
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_store_b96 off, v[33:35], s33 offset:1904
; GFX11-PAL-GCNTRACKERS-NEXT: s_clause 0x7 ; 128-byte Folded Reload
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b128 v[5:8], off, s33 offset:1536
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b128 v[9:12], off, s33 offset:1552
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b128 v[13:16], off, s33 offset:1568
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b128 v[17:20], off, s33 offset:1584
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b128 v[21:24], off, s33 offset:1600
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b128 v[25:28], off, s33 offset:1616
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b128 v[29:32], off, s33 offset:1632
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b128 v[33:36], off, s33 offset:1648
; GFX11-PAL-GCNTRACKERS-NEXT: s_waitcnt vmcnt(3)
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_store_b32 off, v24, s2
; GFX11-PAL-GCNTRACKERS-NEXT: s_clause 0x7 ; 128-byte Folded Reload
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b128 v[5:8], off, s33 offset:1536
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b128 v[9:12], off, s33 offset:1552
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b128 v[13:16], off, s33 offset:1568
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b128 v[17:20], off, s33 offset:1584
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b128 v[21:24], off, s33 offset:1600
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b128 v[25:28], off, s33 offset:1616
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b128 v[29:32], off, s33 offset:1632
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b128 v[33:36], off, s33 offset:1648
; GFX11-PAL-GCNTRACKERS-NEXT: s_add_i32 s2, s32, 0x90
; GFX11-PAL-GCNTRACKERS-NEXT: s_waitcnt vmcnt(3)
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_store_b128 off, v[20:23], s2
; GFX11-PAL-GCNTRACKERS-NEXT: s_clause 0x7 ; 128-byte Folded Reload
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b128 v[5:8], off, s33 offset:1536
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b128 v[9:12], off, s33 offset:1552
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b128 v[13:16], off, s33 offset:1568
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b128 v[17:20], off, s33 offset:1584
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b128 v[21:24], off, s33 offset:1600
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b128 v[25:28], off, s33 offset:1616
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b128 v[29:32], off, s33 offset:1632
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b128 v[33:36], off, s33 offset:1648
; GFX11-PAL-GCNTRACKERS-NEXT: s_add_i32 s2, s32, 0x80
; GFX11-PAL-GCNTRACKERS-NEXT: s_waitcnt vmcnt(4)
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_store_b128 off, v[16:19], s2
; GFX11-PAL-GCNTRACKERS-NEXT: s_clause 0x7 ; 128-byte Folded Reload
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b128 v[5:8], off, s33 offset:1536
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b128 v[9:12], off, s33 offset:1552
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b128 v[13:16], off, s33 offset:1568
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b128 v[17:20], off, s33 offset:1584
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b128 v[21:24], off, s33 offset:1600
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b128 v[25:28], off, s33 offset:1616
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b128 v[29:32], off, s33 offset:1632
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b128 v[33:36], off, s33 offset:1648
; GFX11-PAL-GCNTRACKERS-NEXT: s_add_i32 s2, s32, 0x70
; GFX11-PAL-GCNTRACKERS-NEXT: s_waitcnt vmcnt(5)
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_store_b128 off, v[12:15], s2
; GFX11-PAL-GCNTRACKERS-NEXT: s_clause 0x7 ; 128-byte Folded Reload
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b128 v[5:8], off, s33 offset:1536
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b128 v[9:12], off, s33 offset:1552
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b128 v[13:16], off, s33 offset:1568
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b128 v[17:20], off, s33 offset:1584
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b128 v[21:24], off, s33 offset:1600
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b128 v[25:28], off, s33 offset:1616
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b128 v[29:32], off, s33 offset:1632
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b128 v[33:36], off, s33 offset:1648
; GFX11-PAL-GCNTRACKERS-NEXT: s_add_i32 s2, s32, 0x6c
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_store_b32 off, v0, s2
; GFX11-PAL-GCNTRACKERS-NEXT: s_add_i32 s2, s32, 0x60
; GFX11-PAL-GCNTRACKERS-NEXT: s_waitcnt vmcnt(6)
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_store_b96 off, v[8:10], s2
; GFX11-PAL-GCNTRACKERS-NEXT: s_clause 0x17 ; 384-byte Folded Reload
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b128 v[7:10], off, s33 offset:1744
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b128 v[11:14], off, s33 offset:1760
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b128 v[15:18], off, s33 offset:1776
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b128 v[19:22], off, s33 offset:1792
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b128 v[23:26], off, s33 offset:1808
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b128 v[27:30], off, s33 offset:1824
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b128 v[31:34], off, s33 offset:1840
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b128 v[35:38], off, s33 offset:1856
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b128 v[23:26], off, s33 offset:1536
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b128 v[27:30], off, s33 offset:1552
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b128 v[31:34], off, s33 offset:1568
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b128 v[35:38], off, s33 offset:1584
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b128 v[39:42], off, s33 offset:1600
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b128 v[43:46], off, s33 offset:1616
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b128 v[47:50], off, s33 offset:1632
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b128 v[51:54], off, s33 offset:1648
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b128 v[24:27], off, s33 offset:1536
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b128 v[28:31], off, s33 offset:1552
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b128 v[32:35], off, s33 offset:1568
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b128 v[36:39], off, s33 offset:1584
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b128 v[40:43], off, s33 offset:1600
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b128 v[44:47], off, s33 offset:1616
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b128 v[48:51], off, s33 offset:1632
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b128 v[52:55], off, s33 offset:1648
; GFX11-PAL-GCNTRACKERS-NEXT: s_add_i32 s2, s32, 0x50
; GFX11-PAL-GCNTRACKERS-NEXT: s_waitcnt vmcnt(7)
; GFX11-PAL-GCNTRACKERS-NEXT: v_mov_b32_e32 v24, v25
; GFX11-PAL-GCNTRACKERS-NEXT: s_clause 0x7 ; 128-byte Folded Reload
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b128 v[25:28], off, s33 offset:1536
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b128 v[29:32], off, s33 offset:1552
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b128 v[33:36], off, s33 offset:1568
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b128 v[37:40], off, s33 offset:1584
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b128 v[41:44], off, s33 offset:1600
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b128 v[45:48], off, s33 offset:1616
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b128 v[49:52], off, s33 offset:1632
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b128 v[53:56], off, s33 offset:1648
; GFX11-PAL-GCNTRACKERS-NEXT: v_dual_mov_b32 v5, v7 :: v_dual_mov_b32 v6, v8
; GFX11-PAL-GCNTRACKERS-NEXT: s_waitcnt vmcnt(7)
; GFX11-PAL-GCNTRACKERS-NEXT: v_dual_mov_b32 v7, v9 :: v_dual_mov_b32 v26, v58
; GFX11-PAL-GCNTRACKERS-NEXT: v_mov_b32_e32 v8, v10
; GFX11-PAL-GCNTRACKERS-NEXT: v_dual_mov_b32 v9, v11 :: v_dual_mov_b32 v28, v60
; GFX11-PAL-GCNTRACKERS-NEXT: v_mov_b32_e32 v10, v12
; GFX11-PAL-GCNTRACKERS-NEXT: v_mov_b32_e32 v11, v13
; GFX11-PAL-GCNTRACKERS-NEXT: s_waitcnt vmcnt(4)
; GFX11-PAL-GCNTRACKERS-NEXT: v_dual_mov_b32 v12, v14 :: v_dual_mov_b32 v37, v5
; GFX11-PAL-GCNTRACKERS-NEXT: v_dual_mov_b32 v38, v6 :: v_dual_mov_b32 v39, v7
; GFX11-PAL-GCNTRACKERS-NEXT: v_mov_b32_e32 v25, v27
; GFX11-PAL-GCNTRACKERS-NEXT: v_mov_b32_e32 v27, v59
; GFX11-PAL-GCNTRACKERS-NEXT: v_mov_b32_e32 v13, v15
; GFX11-PAL-GCNTRACKERS-NEXT: v_mov_b32_e32 v14, v16
; GFX11-PAL-GCNTRACKERS-NEXT: v_mov_b32_e32 v15, v17
; GFX11-PAL-GCNTRACKERS-NEXT: v_mov_b32_e32 v16, v18
; GFX11-PAL-GCNTRACKERS-NEXT: v_mov_b32_e32 v17, v19
; GFX11-PAL-GCNTRACKERS-NEXT: v_mov_b32_e32 v18, v20
; GFX11-PAL-GCNTRACKERS-NEXT: v_mov_b32_e32 v19, v21
; GFX11-PAL-GCNTRACKERS-NEXT: v_dual_mov_b32 v20, v22 :: v_dual_mov_b32 v21, v23
; GFX11-PAL-GCNTRACKERS-NEXT: v_dual_mov_b32 v22, v24 :: v_dual_mov_b32 v23, v25
; GFX11-PAL-GCNTRACKERS-NEXT: v_mov_b32_e32 v25, v57
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_store_b128 off, v[20:23], s2
; GFX11-PAL-GCNTRACKERS-NEXT: s_add_i32 s2, s32, 64
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_store_b128 off, v[16:19], s2
; GFX11-PAL-GCNTRACKERS-NEXT: s_add_i32 s2, s32, 48
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_store_b128 off, v[12:15], s2
; GFX11-PAL-GCNTRACKERS-NEXT: s_add_i32 s2, s32, 32
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_store_b128 off, v[8:11], s2
; GFX11-PAL-GCNTRACKERS-NEXT: s_clause 0x2 ; 44-byte Folded Reload
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b128 v[8:11], off, s33 offset:1872
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b128 v[12:15], off, s33 offset:1888
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b96 v[16:18], off, s33 offset:1904
; GFX11-PAL-GCNTRACKERS-NEXT: s_add_i32 s2, s32, 16
; GFX11-PAL-GCNTRACKERS-NEXT: s_waitcnt vmcnt(1)
; GFX11-PAL-GCNTRACKERS-NEXT: v_dual_mov_b32 v31, v10 :: v_dual_mov_b32 v36, v15
; GFX11-PAL-GCNTRACKERS-NEXT: v_dual_mov_b32 v30, v9 :: v_dual_mov_b32 v29, v8
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_store_b128 off, v[36:39], s2
; GFX11-PAL-GCNTRACKERS-NEXT: s_clause 0x4 ; 80-byte Folded Reload
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b128 v[5:8], off, s33 offset:1728
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b128 v[9:12], off, s33 offset:1712
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b128 v[13:16], off, s33 offset:1696
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b128 v[17:20], off, s33 offset:1680
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b128 v[21:24], off, s33 offset:1664
; GFX11-PAL-GCNTRACKERS-NEXT: s_add_i32 s2, s33, 0x400
; GFX11-PAL-GCNTRACKERS-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-PAL-GCNTRACKERS-NEXT: v_mov_b32_e32 v0, s2
; GFX11-PAL-GCNTRACKERS-NEXT: s_swappc_b64 s[30:31], s[0:1]
; GFX11-PAL-GCNTRACKERS-NEXT: s_clause 0xc ; 52-byte Folded Reload
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b32 v60, off, s33
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b32 v59, off, s33 offset:4
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b32 v58, off, s33 offset:8
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b32 v57, off, s33 offset:12
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b32 v56, off, s33 offset:16
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b32 v47, off, s33 offset:20
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b32 v46, off, s33 offset:24
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b32 v45, off, s33 offset:28
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b32 v44, off, s33 offset:32
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b32 v43, off, s33 offset:36
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b32 v42, off, s33 offset:40
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b32 v41, off, s33 offset:44
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b32 v40, off, s33 offset:48
; GFX11-PAL-GCNTRACKERS-NEXT: v_readlane_b32 s30, v61, 0
; GFX11-PAL-GCNTRACKERS-NEXT: v_readlane_b32 s31, v61, 1
; GFX11-PAL-GCNTRACKERS-NEXT: s_mov_b32 s32, s34
; GFX11-PAL-GCNTRACKERS-NEXT: s_mov_b32 s34, s36
; GFX11-PAL-GCNTRACKERS-NEXT: s_or_saveexec_b32 s0, -1
; GFX11-PAL-GCNTRACKERS-NEXT: scratch_load_b32 v61, off, s33 offset:1916 ; 4-byte Folded Reload
; GFX11-PAL-GCNTRACKERS-NEXT: s_mov_b32 exec_lo, s0
; GFX11-PAL-GCNTRACKERS-NEXT: s_mov_b32 s33, s35
; GFX11-PAL-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; GFX11-PAL-GCNTRACKERS-NEXT: s_setpc_b64 s[30:31]
;
; TONGA-LABEL: call_72xi32:
; TONGA: ; %bb.0: ; %entry
; TONGA-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; TONGA-NEXT: s_mov_b32 s35, s33
; TONGA-NEXT: s_add_i32 s33, s32, 0x7fc0
; TONGA-NEXT: s_and_b32 s33, s33, 0xffff8000
; TONGA-NEXT: s_or_saveexec_b64 s[36:37], -1
; TONGA-NEXT: buffer_store_dword v63, off, s[0:3], s33 offset:1564 ; 4-byte Folded Spill
; TONGA-NEXT: s_mov_b64 exec, s[36:37]
; TONGA-NEXT: s_mov_b32 s38, s34
; TONGA-NEXT: s_mov_b32 s34, s32
; TONGA-NEXT: s_add_i32 s32, s32, 0x28000
; TONGA-NEXT: buffer_store_dword v40, off, s[0:3], s33 offset:56 ; 4-byte Folded Spill
; TONGA-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:52 ; 4-byte Folded Spill
; TONGA-NEXT: buffer_store_dword v42, off, s[0:3], s33 offset:48 ; 4-byte Folded Spill
; TONGA-NEXT: buffer_store_dword v43, off, s[0:3], s33 offset:44 ; 4-byte Folded Spill
; TONGA-NEXT: buffer_store_dword v44, off, s[0:3], s33 offset:40 ; 4-byte Folded Spill
; TONGA-NEXT: buffer_store_dword v45, off, s[0:3], s33 offset:36 ; 4-byte Folded Spill
; TONGA-NEXT: buffer_store_dword v46, off, s[0:3], s33 offset:32 ; 4-byte Folded Spill
; TONGA-NEXT: buffer_store_dword v47, off, s[0:3], s33 offset:28 ; 4-byte Folded Spill
; TONGA-NEXT: buffer_store_dword v56, off, s[0:3], s33 offset:24 ; 4-byte Folded Spill
; TONGA-NEXT: buffer_store_dword v57, off, s[0:3], s33 offset:20 ; 4-byte Folded Spill
; TONGA-NEXT: buffer_store_dword v58, off, s[0:3], s33 offset:16 ; 4-byte Folded Spill
; TONGA-NEXT: buffer_store_dword v59, off, s[0:3], s33 offset:12 ; 4-byte Folded Spill
; TONGA-NEXT: buffer_store_dword v60, off, s[0:3], s33 offset:8 ; 4-byte Folded Spill
; TONGA-NEXT: buffer_store_dword v61, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; TONGA-NEXT: buffer_store_dword v62, off, s[0:3], s33 ; 4-byte Folded Spill
; TONGA-NEXT: v_writelane_b32 v63, s30, 0
; TONGA-NEXT: v_writelane_b32 v63, s31, 1
; TONGA-NEXT: s_getpc_b64 s[36:37]
; TONGA-NEXT: s_add_u32 s36, s36, return_72xi32@gotpcrel32@lo+4
; TONGA-NEXT: s_addc_u32 s37, s37, return_72xi32@gotpcrel32@hi+12
; TONGA-NEXT: s_load_dwordx2 s[36:37], s[36:37], 0x0
; TONGA-NEXT: v_mov_b32_e32 v0, 0
; TONGA-NEXT: buffer_store_dword v0, off, s[0:3], s32
; TONGA-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:4
; TONGA-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:8
; TONGA-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:12
; TONGA-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:16
; TONGA-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:20
; TONGA-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:24
; TONGA-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:28
; TONGA-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:32
; TONGA-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:36
; TONGA-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:40
; TONGA-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:44
; TONGA-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:48
; TONGA-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:52
; TONGA-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:56
; TONGA-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:60
; TONGA-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:64
; TONGA-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:68
; TONGA-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:72
; TONGA-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:76
; TONGA-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:80
; TONGA-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:84
; TONGA-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:88
; TONGA-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:92
; TONGA-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:96
; TONGA-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:100
; TONGA-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:104
; TONGA-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:108
; TONGA-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:112
; TONGA-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:116
; TONGA-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:120
; TONGA-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:124
; TONGA-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:128
; TONGA-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:132
; TONGA-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:136
; TONGA-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:140
; TONGA-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:144
; TONGA-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:148
; TONGA-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:152
; TONGA-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:156
; TONGA-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:160
; TONGA-NEXT: v_lshrrev_b32_e64 v0, 6, s33
; TONGA-NEXT: s_movk_i32 vcc_lo, 0x200
; TONGA-NEXT: v_add_u32_e32 v0, vcc, vcc_lo, v0
; TONGA-NEXT: v_mov_b32_e32 v1, 0
; TONGA-NEXT: v_mov_b32_e32 v2, 0
; TONGA-NEXT: v_mov_b32_e32 v3, 0
; TONGA-NEXT: v_mov_b32_e32 v4, 0
; TONGA-NEXT: v_mov_b32_e32 v5, 0
; TONGA-NEXT: v_mov_b32_e32 v6, 0
; TONGA-NEXT: v_mov_b32_e32 v7, 0
; TONGA-NEXT: v_mov_b32_e32 v8, 0
; TONGA-NEXT: v_mov_b32_e32 v9, 0
; TONGA-NEXT: v_mov_b32_e32 v10, 0
; TONGA-NEXT: v_mov_b32_e32 v11, 0
; TONGA-NEXT: v_mov_b32_e32 v12, 0
; TONGA-NEXT: v_mov_b32_e32 v13, 0
; TONGA-NEXT: v_mov_b32_e32 v14, 0
; TONGA-NEXT: v_mov_b32_e32 v15, 0
; TONGA-NEXT: v_mov_b32_e32 v16, 0
; TONGA-NEXT: v_mov_b32_e32 v17, 0
; TONGA-NEXT: v_mov_b32_e32 v18, 0
; TONGA-NEXT: v_mov_b32_e32 v19, 0
; TONGA-NEXT: v_mov_b32_e32 v20, 0
; TONGA-NEXT: v_mov_b32_e32 v21, 0
; TONGA-NEXT: v_mov_b32_e32 v22, 0
; TONGA-NEXT: v_mov_b32_e32 v23, 0
; TONGA-NEXT: v_mov_b32_e32 v24, 0
; TONGA-NEXT: v_mov_b32_e32 v25, 0
; TONGA-NEXT: v_mov_b32_e32 v26, 0
; TONGA-NEXT: v_mov_b32_e32 v27, 0
; TONGA-NEXT: v_mov_b32_e32 v28, 0
; TONGA-NEXT: v_mov_b32_e32 v29, 0
; TONGA-NEXT: v_mov_b32_e32 v30, 0
; TONGA-NEXT: v_mov_b32_e32 v31, 0
; TONGA-NEXT: s_waitcnt lgkmcnt(0)
; TONGA-NEXT: s_swappc_b64 s[30:31], s[36:37]
; TONGA-NEXT: buffer_load_dword v0, off, s[0:3], s33 offset:636
; TONGA-NEXT: buffer_load_dword v1, off, s[0:3], s33 offset:640
; TONGA-NEXT: buffer_load_dword v32, off, s[0:3], s33 offset:644
; TONGA-NEXT: buffer_load_dword v33, off, s[0:3], s33 offset:648
; TONGA-NEXT: buffer_load_dword v34, off, s[0:3], s33 offset:652
; TONGA-NEXT: buffer_load_dword v35, off, s[0:3], s33 offset:656
; TONGA-NEXT: buffer_load_dword v36, off, s[0:3], s33 offset:660
; TONGA-NEXT: buffer_load_dword v37, off, s[0:3], s33 offset:664
; TONGA-NEXT: buffer_load_dword v38, off, s[0:3], s33 offset:668
; TONGA-NEXT: buffer_load_dword v39, off, s[0:3], s33 offset:672
; TONGA-NEXT: buffer_load_dword v48, off, s[0:3], s33 offset:676
; TONGA-NEXT: buffer_load_dword v49, off, s[0:3], s33 offset:680
; TONGA-NEXT: buffer_load_dword v50, off, s[0:3], s33 offset:684
; TONGA-NEXT: buffer_load_dword v51, off, s[0:3], s33 offset:688
; TONGA-NEXT: buffer_load_dword v52, off, s[0:3], s33 offset:692
; TONGA-NEXT: buffer_load_dword v53, off, s[0:3], s33 offset:696
; TONGA-NEXT: buffer_load_dword v54, off, s[0:3], s33 offset:700
; TONGA-NEXT: buffer_load_dword v55, off, s[0:3], s33 offset:704
; TONGA-NEXT: buffer_load_dword v40, off, s[0:3], s33 offset:708
; TONGA-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:712
; TONGA-NEXT: buffer_load_dword v42, off, s[0:3], s33 offset:716
; TONGA-NEXT: buffer_load_dword v43, off, s[0:3], s33 offset:720
; TONGA-NEXT: buffer_load_dword v44, off, s[0:3], s33 offset:724
; TONGA-NEXT: buffer_load_dword v45, off, s[0:3], s33 offset:728
; TONGA-NEXT: buffer_load_dword v46, off, s[0:3], s33 offset:732
; TONGA-NEXT: buffer_load_dword v47, off, s[0:3], s33 offset:736
; TONGA-NEXT: buffer_load_dword v56, off, s[0:3], s33 offset:740
; TONGA-NEXT: buffer_load_dword v57, off, s[0:3], s33 offset:748
; TONGA-NEXT: buffer_load_dword v58, off, s[0:3], s33 offset:752
; TONGA-NEXT: buffer_load_dword v59, off, s[0:3], s33 offset:756
; TONGA-NEXT: buffer_load_dword v60, off, s[0:3], s33 offset:760
; TONGA-NEXT: buffer_load_dword v61, off, s[0:3], s33 offset:764
; TONGA-NEXT: buffer_load_dword v62, off, s[0:3], s33 offset:768
; TONGA-NEXT: buffer_load_dword v2, off, s[0:3], s33 offset:772
; TONGA-NEXT: buffer_load_dword v3, off, s[0:3], s33 offset:776
; TONGA-NEXT: buffer_load_dword v4, off, s[0:3], s33 offset:780
; TONGA-NEXT: buffer_load_dword v5, off, s[0:3], s33 offset:784
; TONGA-NEXT: buffer_load_dword v6, off, s[0:3], s33 offset:788
; TONGA-NEXT: buffer_load_dword v7, off, s[0:3], s33 offset:792
; TONGA-NEXT: buffer_load_dword v8, off, s[0:3], s33 offset:796
; TONGA-NEXT: buffer_load_dword v9, off, s[0:3], s33 offset:516
; TONGA-NEXT: s_movk_i32 vcc_lo, 0x400
; TONGA-NEXT: s_waitcnt vmcnt(0)
; TONGA-NEXT: buffer_store_dword v9, off, s[0:3], s33 offset:1536 ; 4-byte Folded Spill
; TONGA-NEXT: buffer_load_dword v9, off, s[0:3], s33 offset:520
; TONGA-NEXT: s_waitcnt vmcnt(0)
; TONGA-NEXT: buffer_store_dword v9, off, s[0:3], s33 offset:1540 ; 4-byte Folded Spill
; TONGA-NEXT: buffer_load_dword v9, off, s[0:3], s33 offset:524
; TONGA-NEXT: s_waitcnt vmcnt(0)
; TONGA-NEXT: buffer_store_dword v9, off, s[0:3], s33 offset:1544 ; 4-byte Folded Spill
; TONGA-NEXT: buffer_load_dword v9, off, s[0:3], s33 offset:528
; TONGA-NEXT: s_waitcnt vmcnt(0)
; TONGA-NEXT: buffer_store_dword v9, off, s[0:3], s33 offset:1548 ; 4-byte Folded Spill
; TONGA-NEXT: buffer_load_dword v9, off, s[0:3], s33 offset:532
; TONGA-NEXT: s_waitcnt vmcnt(0)
; TONGA-NEXT: buffer_store_dword v9, off, s[0:3], s33 offset:1552 ; 4-byte Folded Spill
; TONGA-NEXT: buffer_load_dword v9, off, s[0:3], s33 offset:536
; TONGA-NEXT: s_waitcnt vmcnt(0)
; TONGA-NEXT: buffer_store_dword v9, off, s[0:3], s33 offset:1556 ; 4-byte Folded Spill
; TONGA-NEXT: buffer_load_dword v9, off, s[0:3], s33 offset:540
; TONGA-NEXT: s_waitcnt vmcnt(0)
; TONGA-NEXT: buffer_store_dword v9, off, s[0:3], s33 offset:1560 ; 4-byte Folded Spill
; TONGA-NEXT: buffer_load_dword v9, off, s[0:3], s33 offset:544
; TONGA-NEXT: buffer_load_dword v10, off, s[0:3], s33 offset:548
; TONGA-NEXT: buffer_load_dword v11, off, s[0:3], s33 offset:552
; TONGA-NEXT: buffer_load_dword v12, off, s[0:3], s33 offset:556
; TONGA-NEXT: buffer_load_dword v13, off, s[0:3], s33 offset:560
; TONGA-NEXT: buffer_load_dword v14, off, s[0:3], s33 offset:564
; TONGA-NEXT: buffer_load_dword v15, off, s[0:3], s33 offset:568
; TONGA-NEXT: buffer_load_dword v16, off, s[0:3], s33 offset:572
; TONGA-NEXT: buffer_load_dword v17, off, s[0:3], s33 offset:576
; TONGA-NEXT: buffer_load_dword v18, off, s[0:3], s33 offset:580
; TONGA-NEXT: buffer_load_dword v19, off, s[0:3], s33 offset:584
; TONGA-NEXT: buffer_load_dword v20, off, s[0:3], s33 offset:588
; TONGA-NEXT: buffer_load_dword v21, off, s[0:3], s33 offset:592
; TONGA-NEXT: buffer_load_dword v22, off, s[0:3], s33 offset:596
; TONGA-NEXT: buffer_load_dword v23, off, s[0:3], s33 offset:600
; TONGA-NEXT: buffer_load_dword v24, off, s[0:3], s33 offset:604
; TONGA-NEXT: buffer_load_dword v25, off, s[0:3], s33 offset:608
; TONGA-NEXT: buffer_load_dword v26, off, s[0:3], s33 offset:612
; TONGA-NEXT: buffer_load_dword v27, off, s[0:3], s33 offset:616
; TONGA-NEXT: buffer_load_dword v28, off, s[0:3], s33 offset:620
; TONGA-NEXT: buffer_load_dword v29, off, s[0:3], s33 offset:624
; TONGA-NEXT: buffer_load_dword v30, off, s[0:3], s33 offset:628
; TONGA-NEXT: buffer_load_dword v31, off, s[0:3], s33 offset:632
; TONGA-NEXT: buffer_store_dword v0, off, s[0:3], s32
; TONGA-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:4
; TONGA-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:8
; TONGA-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:12
; TONGA-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:16
; TONGA-NEXT: buffer_store_dword v35, off, s[0:3], s32 offset:20
; TONGA-NEXT: buffer_store_dword v36, off, s[0:3], s32 offset:24
; TONGA-NEXT: buffer_store_dword v37, off, s[0:3], s32 offset:28
; TONGA-NEXT: buffer_store_dword v38, off, s[0:3], s32 offset:32
; TONGA-NEXT: buffer_store_dword v39, off, s[0:3], s32 offset:36
; TONGA-NEXT: buffer_store_dword v48, off, s[0:3], s32 offset:40
; TONGA-NEXT: buffer_store_dword v49, off, s[0:3], s32 offset:44
; TONGA-NEXT: buffer_store_dword v50, off, s[0:3], s32 offset:48
; TONGA-NEXT: buffer_store_dword v51, off, s[0:3], s32 offset:52
; TONGA-NEXT: buffer_store_dword v52, off, s[0:3], s32 offset:56
; TONGA-NEXT: buffer_store_dword v53, off, s[0:3], s32 offset:60
; TONGA-NEXT: buffer_store_dword v54, off, s[0:3], s32 offset:64
; TONGA-NEXT: buffer_store_dword v55, off, s[0:3], s32 offset:68
; TONGA-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:72
; TONGA-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:76
; TONGA-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:80
; TONGA-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:84
; TONGA-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:88
; TONGA-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:92
; TONGA-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:96
; TONGA-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:100
; TONGA-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:104
; TONGA-NEXT: v_mov_b32_e32 v0, 24
; TONGA-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:108
; TONGA-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:112
; TONGA-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:116
; TONGA-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:120
; TONGA-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:124
; TONGA-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:128
; TONGA-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:132
; TONGA-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:136
; TONGA-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:140
; TONGA-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:144
; TONGA-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:148
; TONGA-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:152
; TONGA-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:156
; TONGA-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:160
; TONGA-NEXT: buffer_load_dword v2, off, s[0:3], s33 offset:1536 ; 4-byte Folded Reload
; TONGA-NEXT: buffer_load_dword v3, off, s[0:3], s33 offset:1540 ; 4-byte Folded Reload
; TONGA-NEXT: buffer_load_dword v4, off, s[0:3], s33 offset:1544 ; 4-byte Folded Reload
; TONGA-NEXT: buffer_load_dword v5, off, s[0:3], s33 offset:1548 ; 4-byte Folded Reload
; TONGA-NEXT: buffer_load_dword v6, off, s[0:3], s33 offset:1552 ; 4-byte Folded Reload
; TONGA-NEXT: buffer_load_dword v7, off, s[0:3], s33 offset:1556 ; 4-byte Folded Reload
; TONGA-NEXT: buffer_load_dword v8, off, s[0:3], s33 offset:1560 ; 4-byte Folded Reload
; TONGA-NEXT: v_lshrrev_b32_e64 v0, 6, s33
; TONGA-NEXT: v_add_u32_e32 v0, vcc, vcc_lo, v0
; TONGA-NEXT: v_mov_b32_e32 v1, 42
; TONGA-NEXT: s_swappc_b64 s[30:31], s[36:37]
; TONGA-NEXT: buffer_load_dword v62, off, s[0:3], s33 ; 4-byte Folded Reload
; TONGA-NEXT: buffer_load_dword v61, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; TONGA-NEXT: buffer_load_dword v60, off, s[0:3], s33 offset:8 ; 4-byte Folded Reload
; TONGA-NEXT: buffer_load_dword v59, off, s[0:3], s33 offset:12 ; 4-byte Folded Reload
; TONGA-NEXT: buffer_load_dword v58, off, s[0:3], s33 offset:16 ; 4-byte Folded Reload
; TONGA-NEXT: buffer_load_dword v57, off, s[0:3], s33 offset:20 ; 4-byte Folded Reload
; TONGA-NEXT: buffer_load_dword v56, off, s[0:3], s33 offset:24 ; 4-byte Folded Reload
; TONGA-NEXT: buffer_load_dword v47, off, s[0:3], s33 offset:28 ; 4-byte Folded Reload
; TONGA-NEXT: buffer_load_dword v46, off, s[0:3], s33 offset:32 ; 4-byte Folded Reload
; TONGA-NEXT: buffer_load_dword v45, off, s[0:3], s33 offset:36 ; 4-byte Folded Reload
; TONGA-NEXT: buffer_load_dword v44, off, s[0:3], s33 offset:40 ; 4-byte Folded Reload
; TONGA-NEXT: buffer_load_dword v43, off, s[0:3], s33 offset:44 ; 4-byte Folded Reload
; TONGA-NEXT: buffer_load_dword v42, off, s[0:3], s33 offset:48 ; 4-byte Folded Reload
; TONGA-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:52 ; 4-byte Folded Reload
; TONGA-NEXT: buffer_load_dword v40, off, s[0:3], s33 offset:56 ; 4-byte Folded Reload
; TONGA-NEXT: v_readlane_b32 s30, v63, 0
; TONGA-NEXT: v_readlane_b32 s31, v63, 1
; TONGA-NEXT: s_mov_b32 s32, s34
; TONGA-NEXT: s_mov_b32 s34, s38
; TONGA-NEXT: s_or_saveexec_b64 s[36:37], -1
; TONGA-NEXT: buffer_load_dword v63, off, s[0:3], s33 offset:1564 ; 4-byte Folded Reload
; TONGA-NEXT: s_mov_b64 exec, s[36:37]
; TONGA-NEXT: s_mov_b32 s33, s35
; TONGA-NEXT: s_waitcnt vmcnt(0)
; TONGA-NEXT: s_setpc_b64 s[30:31]
;
; TONGA-GCNTRACKERS-LABEL: call_72xi32:
; TONGA-GCNTRACKERS: ; %bb.0: ; %entry
; TONGA-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; TONGA-GCNTRACKERS-NEXT: s_mov_b32 s35, s33
; TONGA-GCNTRACKERS-NEXT: s_add_i32 s33, s32, 0x7fc0
; TONGA-GCNTRACKERS-NEXT: s_and_b32 s33, s33, 0xffff8000
; TONGA-GCNTRACKERS-NEXT: s_or_saveexec_b64 s[36:37], -1
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v63, off, s[0:3], s33 offset:1564 ; 4-byte Folded Spill
; TONGA-GCNTRACKERS-NEXT: s_mov_b64 exec, s[36:37]
; TONGA-GCNTRACKERS-NEXT: s_mov_b32 s38, s34
; TONGA-GCNTRACKERS-NEXT: s_mov_b32 s34, s32
; TONGA-GCNTRACKERS-NEXT: s_add_i32 s32, s32, 0x28000
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v40, off, s[0:3], s33 offset:56 ; 4-byte Folded Spill
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:52 ; 4-byte Folded Spill
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v42, off, s[0:3], s33 offset:48 ; 4-byte Folded Spill
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v43, off, s[0:3], s33 offset:44 ; 4-byte Folded Spill
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v44, off, s[0:3], s33 offset:40 ; 4-byte Folded Spill
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v45, off, s[0:3], s33 offset:36 ; 4-byte Folded Spill
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v46, off, s[0:3], s33 offset:32 ; 4-byte Folded Spill
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v47, off, s[0:3], s33 offset:28 ; 4-byte Folded Spill
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v56, off, s[0:3], s33 offset:24 ; 4-byte Folded Spill
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v57, off, s[0:3], s33 offset:20 ; 4-byte Folded Spill
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v58, off, s[0:3], s33 offset:16 ; 4-byte Folded Spill
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v59, off, s[0:3], s33 offset:12 ; 4-byte Folded Spill
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v60, off, s[0:3], s33 offset:8 ; 4-byte Folded Spill
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v61, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v62, off, s[0:3], s33 ; 4-byte Folded Spill
; TONGA-GCNTRACKERS-NEXT: v_writelane_b32 v63, s30, 0
; TONGA-GCNTRACKERS-NEXT: v_writelane_b32 v63, s31, 1
; TONGA-GCNTRACKERS-NEXT: s_getpc_b64 s[36:37]
; TONGA-GCNTRACKERS-NEXT: s_add_u32 s36, s36, return_72xi32@gotpcrel32@lo+4
; TONGA-GCNTRACKERS-NEXT: s_addc_u32 s37, s37, return_72xi32@gotpcrel32@hi+12
; TONGA-GCNTRACKERS-NEXT: s_load_dwordx2 s[36:37], s[36:37], 0x0
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v0, 0
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:4
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:8
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:12
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:16
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:20
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:24
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:28
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:32
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:36
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:40
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:44
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:48
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:52
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:56
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:60
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:64
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:68
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:72
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:76
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:80
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:84
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:88
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:92
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:96
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:100
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:104
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:108
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:112
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:116
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:120
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:124
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:128
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:132
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:136
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:140
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:144
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:148
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:152
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:156
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:160
; TONGA-GCNTRACKERS-NEXT: v_lshrrev_b32_e64 v0, 6, s33
; TONGA-GCNTRACKERS-NEXT: s_movk_i32 vcc_lo, 0x200
; TONGA-GCNTRACKERS-NEXT: v_add_u32_e32 v0, vcc, vcc_lo, v0
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v1, 0
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v2, 0
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v3, 0
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v4, 0
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v5, 0
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v6, 0
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v7, 0
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v8, 0
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v9, 0
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v10, 0
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v11, 0
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v12, 0
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v13, 0
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v14, 0
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v15, 0
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v16, 0
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v17, 0
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v18, 0
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v19, 0
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v20, 0
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v21, 0
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v22, 0
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v23, 0
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v24, 0
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v25, 0
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v26, 0
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v27, 0
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v28, 0
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v29, 0
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v30, 0
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v31, 0
; TONGA-GCNTRACKERS-NEXT: s_waitcnt lgkmcnt(0)
; TONGA-GCNTRACKERS-NEXT: s_swappc_b64 s[30:31], s[36:37]
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v0, off, s[0:3], s33 offset:636
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v1, off, s[0:3], s33 offset:640
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s33 offset:644
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v33, off, s[0:3], s33 offset:648
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v34, off, s[0:3], s33 offset:652
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v35, off, s[0:3], s33 offset:656
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v36, off, s[0:3], s33 offset:660
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v37, off, s[0:3], s33 offset:664
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v38, off, s[0:3], s33 offset:668
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v39, off, s[0:3], s33 offset:672
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v48, off, s[0:3], s33 offset:676
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v49, off, s[0:3], s33 offset:680
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v50, off, s[0:3], s33 offset:684
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v51, off, s[0:3], s33 offset:688
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v52, off, s[0:3], s33 offset:692
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v53, off, s[0:3], s33 offset:696
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v54, off, s[0:3], s33 offset:700
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v55, off, s[0:3], s33 offset:704
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v40, off, s[0:3], s33 offset:708
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:712
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v42, off, s[0:3], s33 offset:716
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v43, off, s[0:3], s33 offset:720
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v44, off, s[0:3], s33 offset:724
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v45, off, s[0:3], s33 offset:728
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v46, off, s[0:3], s33 offset:732
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v47, off, s[0:3], s33 offset:736
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v56, off, s[0:3], s33 offset:740
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v57, off, s[0:3], s33 offset:748
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v58, off, s[0:3], s33 offset:752
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v59, off, s[0:3], s33 offset:756
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v60, off, s[0:3], s33 offset:760
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v61, off, s[0:3], s33 offset:764
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v62, off, s[0:3], s33 offset:768
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v2, off, s[0:3], s33 offset:772
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v3, off, s[0:3], s33 offset:776
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v4, off, s[0:3], s33 offset:780
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v5, off, s[0:3], s33 offset:784
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v6, off, s[0:3], s33 offset:788
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v7, off, s[0:3], s33 offset:792
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v8, off, s[0:3], s33 offset:796
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v12, off, s[0:3], s33 offset:516
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v13, off, s[0:3], s33 offset:520
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v14, off, s[0:3], s33 offset:524
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v15, off, s[0:3], s33 offset:528
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v16, off, s[0:3], s33 offset:532
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v17, off, s[0:3], s33 offset:536
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v18, off, s[0:3], s33 offset:540
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v9, off, s[0:3], s33 offset:544
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v10, off, s[0:3], s33 offset:548
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v11, off, s[0:3], s33 offset:552
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v19, off, s[0:3], s33 offset:556
; TONGA-GCNTRACKERS-NEXT: s_movk_i32 vcc_lo, 0x400
; TONGA-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v19, off, s[0:3], s33 offset:1540 ; 4-byte Folded Spill
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v19, off, s[0:3], s33 offset:560
; TONGA-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v19, off, s[0:3], s33 offset:1544 ; 4-byte Folded Spill
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v19, off, s[0:3], s33 offset:564
; TONGA-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v19, off, s[0:3], s33 offset:1548 ; 4-byte Folded Spill
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v19, off, s[0:3], s33 offset:568
; TONGA-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v19, off, s[0:3], s33 offset:1552 ; 4-byte Folded Spill
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v19, off, s[0:3], s33 offset:572
; TONGA-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v19, off, s[0:3], s33 offset:1556 ; 4-byte Folded Spill
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v19, off, s[0:3], s33 offset:576
; TONGA-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v19, off, s[0:3], s33 offset:1560 ; 4-byte Folded Spill
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v19, off, s[0:3], s33 offset:580
; TONGA-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v19, off, s[0:3], s33 offset:1536 ; 4-byte Folded Spill
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v19, off, s[0:3], s33 offset:584
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v20, off, s[0:3], s33 offset:588
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v21, off, s[0:3], s33 offset:592
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v22, off, s[0:3], s33 offset:596
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v23, off, s[0:3], s33 offset:600
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v24, off, s[0:3], s33 offset:604
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v25, off, s[0:3], s33 offset:608
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v26, off, s[0:3], s33 offset:612
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v27, off, s[0:3], s33 offset:616
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v28, off, s[0:3], s33 offset:620
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v29, off, s[0:3], s33 offset:624
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v30, off, s[0:3], s33 offset:628
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v31, off, s[0:3], s33 offset:632
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:4
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:8
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:12
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:16
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v35, off, s[0:3], s32 offset:20
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v36, off, s[0:3], s32 offset:24
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v37, off, s[0:3], s32 offset:28
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v38, off, s[0:3], s32 offset:32
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v39, off, s[0:3], s32 offset:36
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v48, off, s[0:3], s32 offset:40
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v49, off, s[0:3], s32 offset:44
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v50, off, s[0:3], s32 offset:48
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v51, off, s[0:3], s32 offset:52
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v52, off, s[0:3], s32 offset:56
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v53, off, s[0:3], s32 offset:60
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v54, off, s[0:3], s32 offset:64
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v55, off, s[0:3], s32 offset:68
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:72
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:76
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:80
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:84
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:88
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:92
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:96
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:100
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:104
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v0, 24
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:108
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:112
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:116
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:120
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:124
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:128
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:132
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:136
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:140
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:144
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:148
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:152
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:156
; TONGA-GCNTRACKERS-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:160
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v2, v12
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v3, v13
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v4, v14
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v5, v15
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v6, v16
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v7, v17
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v8, v18
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v12, off, s[0:3], s33 offset:1540 ; 4-byte Folded Reload
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v13, off, s[0:3], s33 offset:1544 ; 4-byte Folded Reload
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v14, off, s[0:3], s33 offset:1548 ; 4-byte Folded Reload
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v15, off, s[0:3], s33 offset:1552 ; 4-byte Folded Reload
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v16, off, s[0:3], s33 offset:1556 ; 4-byte Folded Reload
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v17, off, s[0:3], s33 offset:1560 ; 4-byte Folded Reload
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v18, off, s[0:3], s33 offset:1536 ; 4-byte Folded Reload
; TONGA-GCNTRACKERS-NEXT: v_lshrrev_b32_e64 v0, 6, s33
; TONGA-GCNTRACKERS-NEXT: v_add_u32_e32 v0, vcc, vcc_lo, v0
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v1, 42
; TONGA-GCNTRACKERS-NEXT: s_swappc_b64 s[30:31], s[36:37]
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v62, off, s[0:3], s33 ; 4-byte Folded Reload
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v61, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v60, off, s[0:3], s33 offset:8 ; 4-byte Folded Reload
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v59, off, s[0:3], s33 offset:12 ; 4-byte Folded Reload
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v58, off, s[0:3], s33 offset:16 ; 4-byte Folded Reload
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v57, off, s[0:3], s33 offset:20 ; 4-byte Folded Reload
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v56, off, s[0:3], s33 offset:24 ; 4-byte Folded Reload
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v47, off, s[0:3], s33 offset:28 ; 4-byte Folded Reload
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v46, off, s[0:3], s33 offset:32 ; 4-byte Folded Reload
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v45, off, s[0:3], s33 offset:36 ; 4-byte Folded Reload
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v44, off, s[0:3], s33 offset:40 ; 4-byte Folded Reload
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v43, off, s[0:3], s33 offset:44 ; 4-byte Folded Reload
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v42, off, s[0:3], s33 offset:48 ; 4-byte Folded Reload
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:52 ; 4-byte Folded Reload
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v40, off, s[0:3], s33 offset:56 ; 4-byte Folded Reload
; TONGA-GCNTRACKERS-NEXT: v_readlane_b32 s30, v63, 0
; TONGA-GCNTRACKERS-NEXT: v_readlane_b32 s31, v63, 1
; TONGA-GCNTRACKERS-NEXT: s_mov_b32 s32, s34
; TONGA-GCNTRACKERS-NEXT: s_mov_b32 s34, s38
; TONGA-GCNTRACKERS-NEXT: s_or_saveexec_b64 s[36:37], -1
; TONGA-GCNTRACKERS-NEXT: buffer_load_dword v63, off, s[0:3], s33 offset:1564 ; 4-byte Folded Reload
; TONGA-GCNTRACKERS-NEXT: s_mov_b64 exec, s[36:37]
; TONGA-GCNTRACKERS-NEXT: s_mov_b32 s33, s35
; TONGA-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; TONGA-GCNTRACKERS-NEXT: s_setpc_b64 s[30:31]
;
; GFX908-LABEL: call_72xi32:
; GFX908: ; %bb.0: ; %entry
; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX908-NEXT: s_mov_b32 s35, s33
; GFX908-NEXT: s_add_i32 s33, s32, 0x7fc0
; GFX908-NEXT: s_and_b32 s33, s33, 0xffff8000
; GFX908-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX908-NEXT: buffer_store_dword v62, off, s[0:3], s33 offset:1024 ; 4-byte Folded Spill
; GFX908-NEXT: s_mov_b64 exec, s[36:37]
; GFX908-NEXT: s_mov_b32 s38, s34
; GFX908-NEXT: s_mov_b32 s34, s32
; GFX908-NEXT: s_add_i32 s32, s32, 0x20000
; GFX908-NEXT: v_accvgpr_write_b32 a8, v40 ; Reload Reuse
; GFX908-NEXT: v_accvgpr_write_b32 a9, v41 ; Reload Reuse
; GFX908-NEXT: v_accvgpr_write_b32 a10, v42 ; Reload Reuse
; GFX908-NEXT: v_accvgpr_write_b32 a11, v43 ; Reload Reuse
; GFX908-NEXT: v_accvgpr_write_b32 a12, v44 ; Reload Reuse
; GFX908-NEXT: v_accvgpr_write_b32 a13, v45 ; Reload Reuse
; GFX908-NEXT: v_accvgpr_write_b32 a14, v46 ; Reload Reuse
; GFX908-NEXT: v_accvgpr_write_b32 a15, v47 ; Reload Reuse
; GFX908-NEXT: v_accvgpr_write_b32 a16, v56 ; Reload Reuse
; GFX908-NEXT: v_accvgpr_write_b32 a17, v57 ; Reload Reuse
; GFX908-NEXT: v_accvgpr_write_b32 a18, v58 ; Reload Reuse
; GFX908-NEXT: v_accvgpr_write_b32 a19, v59 ; Reload Reuse
; GFX908-NEXT: v_accvgpr_write_b32 a20, v60 ; Reload Reuse
; GFX908-NEXT: v_accvgpr_write_b32 a21, v61 ; Reload Reuse
; GFX908-NEXT: v_writelane_b32 v62, s30, 0
; GFX908-NEXT: v_writelane_b32 v62, s31, 1
; GFX908-NEXT: s_getpc_b64 s[36:37]
; GFX908-NEXT: s_add_u32 s36, s36, return_72xi32@gotpcrel32@lo+4
; GFX908-NEXT: s_addc_u32 s37, s37, return_72xi32@gotpcrel32@hi+12
; GFX908-NEXT: s_load_dwordx2 s[36:37], s[36:37], 0x0
; GFX908-NEXT: v_mov_b32_e32 v0, 0
; GFX908-NEXT: buffer_store_dword v0, off, s[0:3], s32
; GFX908-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:4
; GFX908-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:8
; GFX908-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:12
; GFX908-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:16
; GFX908-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:20
; GFX908-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:24
; GFX908-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:28
; GFX908-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:32
; GFX908-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:36
; GFX908-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:40
; GFX908-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:44
; GFX908-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:48
; GFX908-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:52
; GFX908-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:56
; GFX908-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:60
; GFX908-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:64
; GFX908-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:68
; GFX908-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:72
; GFX908-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:76
; GFX908-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:80
; GFX908-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:84
; GFX908-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:88
; GFX908-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:92
; GFX908-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:96
; GFX908-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:100
; GFX908-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:104
; GFX908-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:108
; GFX908-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:112
; GFX908-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:116
; GFX908-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:120
; GFX908-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:124
; GFX908-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:128
; GFX908-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:132
; GFX908-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:136
; GFX908-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:140
; GFX908-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:144
; GFX908-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:148
; GFX908-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:152
; GFX908-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:156
; GFX908-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:160
; GFX908-NEXT: v_lshrrev_b32_e64 v0, 6, s33
; GFX908-NEXT: v_mov_b32_e32 v1, 0
; GFX908-NEXT: v_mov_b32_e32 v2, 0
; GFX908-NEXT: v_mov_b32_e32 v3, 0
; GFX908-NEXT: v_mov_b32_e32 v4, 0
; GFX908-NEXT: v_mov_b32_e32 v5, 0
; GFX908-NEXT: v_mov_b32_e32 v6, 0
; GFX908-NEXT: v_mov_b32_e32 v7, 0
; GFX908-NEXT: v_mov_b32_e32 v8, 0
; GFX908-NEXT: v_mov_b32_e32 v9, 0
; GFX908-NEXT: v_mov_b32_e32 v10, 0
; GFX908-NEXT: v_mov_b32_e32 v11, 0
; GFX908-NEXT: v_mov_b32_e32 v12, 0
; GFX908-NEXT: v_mov_b32_e32 v13, 0
; GFX908-NEXT: v_mov_b32_e32 v14, 0
; GFX908-NEXT: v_mov_b32_e32 v15, 0
; GFX908-NEXT: v_mov_b32_e32 v16, 0
; GFX908-NEXT: v_mov_b32_e32 v17, 0
; GFX908-NEXT: v_mov_b32_e32 v18, 0
; GFX908-NEXT: v_mov_b32_e32 v19, 0
; GFX908-NEXT: v_mov_b32_e32 v20, 0
; GFX908-NEXT: v_mov_b32_e32 v21, 0
; GFX908-NEXT: v_mov_b32_e32 v22, 0
; GFX908-NEXT: v_mov_b32_e32 v23, 0
; GFX908-NEXT: v_mov_b32_e32 v24, 0
; GFX908-NEXT: v_mov_b32_e32 v25, 0
; GFX908-NEXT: v_mov_b32_e32 v26, 0
; GFX908-NEXT: v_mov_b32_e32 v27, 0
; GFX908-NEXT: v_mov_b32_e32 v28, 0
; GFX908-NEXT: v_mov_b32_e32 v29, 0
; GFX908-NEXT: v_mov_b32_e32 v30, 0
; GFX908-NEXT: v_mov_b32_e32 v31, 0
; GFX908-NEXT: s_waitcnt lgkmcnt(0)
; GFX908-NEXT: s_swappc_b64 s[30:31], s[36:37]
; GFX908-NEXT: buffer_load_dword v0, off, s[0:3], s33 offset:124
; GFX908-NEXT: buffer_load_dword v1, off, s[0:3], s33 offset:128
; GFX908-NEXT: buffer_load_dword v32, off, s[0:3], s33 offset:132
; GFX908-NEXT: buffer_load_dword v33, off, s[0:3], s33 offset:136
; GFX908-NEXT: buffer_load_dword v34, off, s[0:3], s33 offset:140
; GFX908-NEXT: buffer_load_dword v35, off, s[0:3], s33 offset:144
; GFX908-NEXT: buffer_load_dword v36, off, s[0:3], s33 offset:148
; GFX908-NEXT: buffer_load_dword v37, off, s[0:3], s33 offset:152
; GFX908-NEXT: buffer_load_dword v38, off, s[0:3], s33 offset:156
; GFX908-NEXT: buffer_load_dword v39, off, s[0:3], s33 offset:160
; GFX908-NEXT: buffer_load_dword v48, off, s[0:3], s33 offset:164
; GFX908-NEXT: buffer_load_dword v49, off, s[0:3], s33 offset:168
; GFX908-NEXT: buffer_load_dword v50, off, s[0:3], s33 offset:172
; GFX908-NEXT: buffer_load_dword v51, off, s[0:3], s33 offset:176
; GFX908-NEXT: buffer_load_dword v52, off, s[0:3], s33 offset:180
; GFX908-NEXT: buffer_load_dword v53, off, s[0:3], s33 offset:184
; GFX908-NEXT: buffer_load_dword v54, off, s[0:3], s33 offset:188
; GFX908-NEXT: buffer_load_dword v55, off, s[0:3], s33 offset:192
; GFX908-NEXT: buffer_load_dword v40, off, s[0:3], s33 offset:196
; GFX908-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:200
; GFX908-NEXT: buffer_load_dword v42, off, s[0:3], s33 offset:204
; GFX908-NEXT: buffer_load_dword v43, off, s[0:3], s33 offset:208
; GFX908-NEXT: buffer_load_dword v44, off, s[0:3], s33 offset:212
; GFX908-NEXT: buffer_load_dword v45, off, s[0:3], s33 offset:216
; GFX908-NEXT: buffer_load_dword v46, off, s[0:3], s33 offset:220
; GFX908-NEXT: buffer_load_dword v47, off, s[0:3], s33 offset:224
; GFX908-NEXT: buffer_load_dword v56, off, s[0:3], s33 offset:228
; GFX908-NEXT: buffer_load_dword v57, off, s[0:3], s33 offset:236
; GFX908-NEXT: buffer_load_dword v58, off, s[0:3], s33 offset:240
; GFX908-NEXT: buffer_load_dword v59, off, s[0:3], s33 offset:244
; GFX908-NEXT: buffer_load_dword v60, off, s[0:3], s33 offset:248
; GFX908-NEXT: buffer_load_dword v61, off, s[0:3], s33 offset:252
; GFX908-NEXT: buffer_load_dword v2, off, s[0:3], s33 offset:256
; GFX908-NEXT: buffer_load_dword v3, off, s[0:3], s33 offset:260
; GFX908-NEXT: buffer_load_dword v4, off, s[0:3], s33 offset:264
; GFX908-NEXT: buffer_load_dword v5, off, s[0:3], s33 offset:268
; GFX908-NEXT: buffer_load_dword v6, off, s[0:3], s33 offset:272
; GFX908-NEXT: buffer_load_dword v7, off, s[0:3], s33 offset:276
; GFX908-NEXT: buffer_load_dword v8, off, s[0:3], s33 offset:280
; GFX908-NEXT: buffer_load_dword v9, off, s[0:3], s33 offset:284
; GFX908-NEXT: buffer_load_dword v10, off, s[0:3], s33 offset:4
; GFX908-NEXT: s_waitcnt vmcnt(0)
; GFX908-NEXT: v_accvgpr_write_b32 a0, v10
; GFX908-NEXT: buffer_load_dword v10, off, s[0:3], s33 offset:8
; GFX908-NEXT: s_waitcnt vmcnt(0)
; GFX908-NEXT: v_accvgpr_write_b32 a1, v10
; GFX908-NEXT: buffer_load_dword v10, off, s[0:3], s33 offset:12
; GFX908-NEXT: s_waitcnt vmcnt(0)
; GFX908-NEXT: v_accvgpr_write_b32 a2, v10
; GFX908-NEXT: buffer_load_dword v10, off, s[0:3], s33 offset:16
; GFX908-NEXT: s_waitcnt vmcnt(0)
; GFX908-NEXT: v_accvgpr_write_b32 a3, v10
; GFX908-NEXT: buffer_load_dword v10, off, s[0:3], s33 offset:20
; GFX908-NEXT: s_waitcnt vmcnt(0)
; GFX908-NEXT: v_accvgpr_write_b32 a4, v10
; GFX908-NEXT: buffer_load_dword v10, off, s[0:3], s33 offset:24
; GFX908-NEXT: s_waitcnt vmcnt(0)
; GFX908-NEXT: v_accvgpr_write_b32 a5, v10
; GFX908-NEXT: buffer_load_dword v10, off, s[0:3], s33 offset:28
; GFX908-NEXT: s_waitcnt vmcnt(0)
; GFX908-NEXT: v_accvgpr_write_b32 a6, v10
; GFX908-NEXT: buffer_load_dword v10, off, s[0:3], s33 offset:32
; GFX908-NEXT: s_waitcnt vmcnt(0)
; GFX908-NEXT: v_accvgpr_write_b32 a7, v10
; GFX908-NEXT: buffer_load_dword v10, off, s[0:3], s33 offset:36
; GFX908-NEXT: buffer_load_dword v11, off, s[0:3], s33 offset:40
; GFX908-NEXT: buffer_load_dword v12, off, s[0:3], s33 offset:44
; GFX908-NEXT: buffer_load_dword v13, off, s[0:3], s33 offset:48
; GFX908-NEXT: buffer_load_dword v14, off, s[0:3], s33 offset:52
; GFX908-NEXT: buffer_load_dword v15, off, s[0:3], s33 offset:56
; GFX908-NEXT: buffer_load_dword v16, off, s[0:3], s33 offset:60
; GFX908-NEXT: buffer_load_dword v17, off, s[0:3], s33 offset:64
; GFX908-NEXT: buffer_load_dword v18, off, s[0:3], s33 offset:68
; GFX908-NEXT: buffer_load_dword v19, off, s[0:3], s33 offset:72
; GFX908-NEXT: buffer_load_dword v20, off, s[0:3], s33 offset:76
; GFX908-NEXT: buffer_load_dword v21, off, s[0:3], s33 offset:80
; GFX908-NEXT: buffer_load_dword v22, off, s[0:3], s33 offset:84
; GFX908-NEXT: buffer_load_dword v23, off, s[0:3], s33 offset:88
; GFX908-NEXT: buffer_load_dword v24, off, s[0:3], s33 offset:92
; GFX908-NEXT: buffer_load_dword v25, off, s[0:3], s33 offset:96
; GFX908-NEXT: buffer_load_dword v26, off, s[0:3], s33 offset:100
; GFX908-NEXT: buffer_load_dword v27, off, s[0:3], s33 offset:104
; GFX908-NEXT: buffer_load_dword v28, off, s[0:3], s33 offset:108
; GFX908-NEXT: buffer_load_dword v29, off, s[0:3], s33 offset:112
; GFX908-NEXT: buffer_load_dword v30, off, s[0:3], s33 offset:116
; GFX908-NEXT: buffer_load_dword v31, off, s[0:3], s33 offset:120
; GFX908-NEXT: s_nop 0
; GFX908-NEXT: buffer_store_dword v0, off, s[0:3], s32
; GFX908-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:4
; GFX908-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:8
; GFX908-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:12
; GFX908-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:16
; GFX908-NEXT: buffer_store_dword v35, off, s[0:3], s32 offset:20
; GFX908-NEXT: buffer_store_dword v36, off, s[0:3], s32 offset:24
; GFX908-NEXT: buffer_store_dword v37, off, s[0:3], s32 offset:28
; GFX908-NEXT: buffer_store_dword v38, off, s[0:3], s32 offset:32
; GFX908-NEXT: buffer_store_dword v39, off, s[0:3], s32 offset:36
; GFX908-NEXT: buffer_store_dword v48, off, s[0:3], s32 offset:40
; GFX908-NEXT: buffer_store_dword v49, off, s[0:3], s32 offset:44
; GFX908-NEXT: buffer_store_dword v50, off, s[0:3], s32 offset:48
; GFX908-NEXT: buffer_store_dword v51, off, s[0:3], s32 offset:52
; GFX908-NEXT: buffer_store_dword v52, off, s[0:3], s32 offset:56
; GFX908-NEXT: buffer_store_dword v53, off, s[0:3], s32 offset:60
; GFX908-NEXT: buffer_store_dword v54, off, s[0:3], s32 offset:64
; GFX908-NEXT: buffer_store_dword v55, off, s[0:3], s32 offset:68
; GFX908-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:72
; GFX908-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:76
; GFX908-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:80
; GFX908-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:84
; GFX908-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:88
; GFX908-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:92
; GFX908-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:96
; GFX908-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:100
; GFX908-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:104
; GFX908-NEXT: v_mov_b32_e32 v0, 24
; GFX908-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:108
; GFX908-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:112
; GFX908-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:116
; GFX908-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:120
; GFX908-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:124
; GFX908-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:128
; GFX908-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:132
; GFX908-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:136
; GFX908-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:140
; GFX908-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:144
; GFX908-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:148
; GFX908-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:152
; GFX908-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:156
; GFX908-NEXT: buffer_store_dword v9, off, s[0:3], s32 offset:160
; GFX908-NEXT: v_lshrrev_b32_e64 v0, 6, s33
; GFX908-NEXT: v_accvgpr_read_b32 v2, a0
; GFX908-NEXT: v_accvgpr_read_b32 v3, a1
; GFX908-NEXT: v_accvgpr_read_b32 v4, a2
; GFX908-NEXT: v_accvgpr_read_b32 v5, a3
; GFX908-NEXT: v_accvgpr_read_b32 v6, a4
; GFX908-NEXT: v_accvgpr_read_b32 v7, a5
; GFX908-NEXT: v_accvgpr_read_b32 v8, a6
; GFX908-NEXT: v_accvgpr_read_b32 v9, a7
; GFX908-NEXT: v_add_u32_e32 v0, 0x200, v0
; GFX908-NEXT: v_mov_b32_e32 v1, 42
; GFX908-NEXT: s_swappc_b64 s[30:31], s[36:37]
; GFX908-NEXT: v_readlane_b32 s30, v62, 0
; GFX908-NEXT: v_readlane_b32 s31, v62, 1
; GFX908-NEXT: v_accvgpr_read_b32 v61, a21 ; Reload Reuse
; GFX908-NEXT: v_accvgpr_read_b32 v60, a20 ; Reload Reuse
; GFX908-NEXT: v_accvgpr_read_b32 v59, a19 ; Reload Reuse
; GFX908-NEXT: v_accvgpr_read_b32 v58, a18 ; Reload Reuse
; GFX908-NEXT: v_accvgpr_read_b32 v57, a17 ; Reload Reuse
; GFX908-NEXT: v_accvgpr_read_b32 v56, a16 ; Reload Reuse
; GFX908-NEXT: v_accvgpr_read_b32 v47, a15 ; Reload Reuse
; GFX908-NEXT: v_accvgpr_read_b32 v46, a14 ; Reload Reuse
; GFX908-NEXT: v_accvgpr_read_b32 v45, a13 ; Reload Reuse
; GFX908-NEXT: v_accvgpr_read_b32 v44, a12 ; Reload Reuse
; GFX908-NEXT: v_accvgpr_read_b32 v43, a11 ; Reload Reuse
; GFX908-NEXT: v_accvgpr_read_b32 v42, a10 ; Reload Reuse
; GFX908-NEXT: v_accvgpr_read_b32 v41, a9 ; Reload Reuse
; GFX908-NEXT: v_accvgpr_read_b32 v40, a8 ; Reload Reuse
; GFX908-NEXT: s_mov_b32 s32, s34
; GFX908-NEXT: s_mov_b32 s34, s38
; GFX908-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX908-NEXT: buffer_load_dword v62, off, s[0:3], s33 offset:1024 ; 4-byte Folded Reload
; GFX908-NEXT: s_mov_b64 exec, s[36:37]
; GFX908-NEXT: s_mov_b32 s33, s35
; GFX908-NEXT: s_waitcnt vmcnt(0)
; GFX908-NEXT: s_setpc_b64 s[30:31]
;
; GFX908-GCNTRACKERS-LABEL: call_72xi32:
; GFX908-GCNTRACKERS: ; %bb.0: ; %entry
; GFX908-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX908-GCNTRACKERS-NEXT: s_mov_b32 s35, s33
; GFX908-GCNTRACKERS-NEXT: s_add_i32 s33, s32, 0x7fc0
; GFX908-GCNTRACKERS-NEXT: s_and_b32 s33, s33, 0xffff8000
; GFX908-GCNTRACKERS-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v62, off, s[0:3], s33 offset:1024 ; 4-byte Folded Spill
; GFX908-GCNTRACKERS-NEXT: s_mov_b64 exec, s[36:37]
; GFX908-GCNTRACKERS-NEXT: s_mov_b32 s38, s34
; GFX908-GCNTRACKERS-NEXT: s_mov_b32 s34, s32
; GFX908-GCNTRACKERS-NEXT: s_add_i32 s32, s32, 0x20000
; GFX908-GCNTRACKERS-NEXT: v_accvgpr_write_b32 a8, v40 ; Reload Reuse
; GFX908-GCNTRACKERS-NEXT: v_accvgpr_write_b32 a9, v41 ; Reload Reuse
; GFX908-GCNTRACKERS-NEXT: v_accvgpr_write_b32 a10, v42 ; Reload Reuse
; GFX908-GCNTRACKERS-NEXT: v_accvgpr_write_b32 a11, v43 ; Reload Reuse
; GFX908-GCNTRACKERS-NEXT: v_accvgpr_write_b32 a12, v44 ; Reload Reuse
; GFX908-GCNTRACKERS-NEXT: v_accvgpr_write_b32 a13, v45 ; Reload Reuse
; GFX908-GCNTRACKERS-NEXT: v_accvgpr_write_b32 a14, v46 ; Reload Reuse
; GFX908-GCNTRACKERS-NEXT: v_accvgpr_write_b32 a15, v47 ; Reload Reuse
; GFX908-GCNTRACKERS-NEXT: v_accvgpr_write_b32 a16, v56 ; Reload Reuse
; GFX908-GCNTRACKERS-NEXT: v_accvgpr_write_b32 a17, v57 ; Reload Reuse
; GFX908-GCNTRACKERS-NEXT: v_accvgpr_write_b32 a18, v58 ; Reload Reuse
; GFX908-GCNTRACKERS-NEXT: v_accvgpr_write_b32 a19, v59 ; Reload Reuse
; GFX908-GCNTRACKERS-NEXT: v_accvgpr_write_b32 a20, v60 ; Reload Reuse
; GFX908-GCNTRACKERS-NEXT: v_accvgpr_write_b32 a21, v61 ; Reload Reuse
; GFX908-GCNTRACKERS-NEXT: v_writelane_b32 v62, s30, 0
; GFX908-GCNTRACKERS-NEXT: v_writelane_b32 v62, s31, 1
; GFX908-GCNTRACKERS-NEXT: s_getpc_b64 s[36:37]
; GFX908-GCNTRACKERS-NEXT: s_add_u32 s36, s36, return_72xi32@gotpcrel32@lo+4
; GFX908-GCNTRACKERS-NEXT: s_addc_u32 s37, s37, return_72xi32@gotpcrel32@hi+12
; GFX908-GCNTRACKERS-NEXT: s_load_dwordx2 s[36:37], s[36:37], 0x0
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v0, 0
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:4
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:8
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:12
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:16
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:20
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:24
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:28
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:32
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:36
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:40
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:44
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:48
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:52
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:56
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:60
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:64
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:68
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:72
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:76
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:80
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:84
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:88
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:92
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:96
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:100
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:104
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:108
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:112
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:116
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:120
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:124
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:128
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:132
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:136
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:140
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:144
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:148
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:152
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:156
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:160
; GFX908-GCNTRACKERS-NEXT: v_lshrrev_b32_e64 v0, 6, s33
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v1, 0
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v2, 0
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v3, 0
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v4, 0
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v5, 0
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v6, 0
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v7, 0
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v8, 0
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v9, 0
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v10, 0
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v11, 0
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v12, 0
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v13, 0
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v14, 0
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v15, 0
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v16, 0
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v17, 0
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v18, 0
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v19, 0
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v20, 0
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v21, 0
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v22, 0
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v23, 0
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v24, 0
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v25, 0
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v26, 0
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v27, 0
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v28, 0
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v29, 0
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v30, 0
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v31, 0
; GFX908-GCNTRACKERS-NEXT: s_waitcnt lgkmcnt(0)
; GFX908-GCNTRACKERS-NEXT: s_swappc_b64 s[30:31], s[36:37]
; GFX908-GCNTRACKERS-NEXT: buffer_load_dword v0, off, s[0:3], s33 offset:124
; GFX908-GCNTRACKERS-NEXT: buffer_load_dword v1, off, s[0:3], s33 offset:128
; GFX908-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s33 offset:132
; GFX908-GCNTRACKERS-NEXT: buffer_load_dword v33, off, s[0:3], s33 offset:136
; GFX908-GCNTRACKERS-NEXT: buffer_load_dword v34, off, s[0:3], s33 offset:140
; GFX908-GCNTRACKERS-NEXT: buffer_load_dword v35, off, s[0:3], s33 offset:144
; GFX908-GCNTRACKERS-NEXT: buffer_load_dword v36, off, s[0:3], s33 offset:148
; GFX908-GCNTRACKERS-NEXT: buffer_load_dword v37, off, s[0:3], s33 offset:152
; GFX908-GCNTRACKERS-NEXT: buffer_load_dword v38, off, s[0:3], s33 offset:156
; GFX908-GCNTRACKERS-NEXT: buffer_load_dword v39, off, s[0:3], s33 offset:160
; GFX908-GCNTRACKERS-NEXT: buffer_load_dword v48, off, s[0:3], s33 offset:164
; GFX908-GCNTRACKERS-NEXT: buffer_load_dword v49, off, s[0:3], s33 offset:168
; GFX908-GCNTRACKERS-NEXT: buffer_load_dword v50, off, s[0:3], s33 offset:172
; GFX908-GCNTRACKERS-NEXT: buffer_load_dword v51, off, s[0:3], s33 offset:176
; GFX908-GCNTRACKERS-NEXT: buffer_load_dword v52, off, s[0:3], s33 offset:180
; GFX908-GCNTRACKERS-NEXT: buffer_load_dword v53, off, s[0:3], s33 offset:184
; GFX908-GCNTRACKERS-NEXT: buffer_load_dword v54, off, s[0:3], s33 offset:188
; GFX908-GCNTRACKERS-NEXT: buffer_load_dword v55, off, s[0:3], s33 offset:192
; GFX908-GCNTRACKERS-NEXT: buffer_load_dword v40, off, s[0:3], s33 offset:196
; GFX908-GCNTRACKERS-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:200
; GFX908-GCNTRACKERS-NEXT: buffer_load_dword v42, off, s[0:3], s33 offset:204
; GFX908-GCNTRACKERS-NEXT: buffer_load_dword v43, off, s[0:3], s33 offset:208
; GFX908-GCNTRACKERS-NEXT: buffer_load_dword v44, off, s[0:3], s33 offset:212
; GFX908-GCNTRACKERS-NEXT: buffer_load_dword v45, off, s[0:3], s33 offset:216
; GFX908-GCNTRACKERS-NEXT: buffer_load_dword v46, off, s[0:3], s33 offset:220
; GFX908-GCNTRACKERS-NEXT: buffer_load_dword v47, off, s[0:3], s33 offset:224
; GFX908-GCNTRACKERS-NEXT: buffer_load_dword v56, off, s[0:3], s33 offset:228
; GFX908-GCNTRACKERS-NEXT: buffer_load_dword v57, off, s[0:3], s33 offset:236
; GFX908-GCNTRACKERS-NEXT: buffer_load_dword v58, off, s[0:3], s33 offset:240
; GFX908-GCNTRACKERS-NEXT: buffer_load_dword v59, off, s[0:3], s33 offset:244
; GFX908-GCNTRACKERS-NEXT: buffer_load_dword v60, off, s[0:3], s33 offset:248
; GFX908-GCNTRACKERS-NEXT: buffer_load_dword v61, off, s[0:3], s33 offset:252
; GFX908-GCNTRACKERS-NEXT: buffer_load_dword v2, off, s[0:3], s33 offset:256
; GFX908-GCNTRACKERS-NEXT: buffer_load_dword v3, off, s[0:3], s33 offset:260
; GFX908-GCNTRACKERS-NEXT: buffer_load_dword v4, off, s[0:3], s33 offset:264
; GFX908-GCNTRACKERS-NEXT: buffer_load_dword v5, off, s[0:3], s33 offset:268
; GFX908-GCNTRACKERS-NEXT: buffer_load_dword v6, off, s[0:3], s33 offset:272
; GFX908-GCNTRACKERS-NEXT: buffer_load_dword v7, off, s[0:3], s33 offset:276
; GFX908-GCNTRACKERS-NEXT: buffer_load_dword v8, off, s[0:3], s33 offset:280
; GFX908-GCNTRACKERS-NEXT: buffer_load_dword v9, off, s[0:3], s33 offset:284
; GFX908-GCNTRACKERS-NEXT: buffer_load_dword v12, off, s[0:3], s33 offset:4
; GFX908-GCNTRACKERS-NEXT: buffer_load_dword v13, off, s[0:3], s33 offset:8
; GFX908-GCNTRACKERS-NEXT: buffer_load_dword v14, off, s[0:3], s33 offset:12
; GFX908-GCNTRACKERS-NEXT: buffer_load_dword v15, off, s[0:3], s33 offset:16
; GFX908-GCNTRACKERS-NEXT: buffer_load_dword v16, off, s[0:3], s33 offset:20
; GFX908-GCNTRACKERS-NEXT: buffer_load_dword v17, off, s[0:3], s33 offset:24
; GFX908-GCNTRACKERS-NEXT: buffer_load_dword v18, off, s[0:3], s33 offset:28
; GFX908-GCNTRACKERS-NEXT: buffer_load_dword v19, off, s[0:3], s33 offset:32
; GFX908-GCNTRACKERS-NEXT: buffer_load_dword v10, off, s[0:3], s33 offset:36
; GFX908-GCNTRACKERS-NEXT: buffer_load_dword v11, off, s[0:3], s33 offset:40
; GFX908-GCNTRACKERS-NEXT: buffer_load_dword v20, off, s[0:3], s33 offset:44
; GFX908-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; GFX908-GCNTRACKERS-NEXT: v_accvgpr_write_b32 a2, v20
; GFX908-GCNTRACKERS-NEXT: buffer_load_dword v20, off, s[0:3], s33 offset:48
; GFX908-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; GFX908-GCNTRACKERS-NEXT: v_accvgpr_write_b32 a3, v20
; GFX908-GCNTRACKERS-NEXT: buffer_load_dword v20, off, s[0:3], s33 offset:52
; GFX908-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; GFX908-GCNTRACKERS-NEXT: v_accvgpr_write_b32 a4, v20
; GFX908-GCNTRACKERS-NEXT: buffer_load_dword v20, off, s[0:3], s33 offset:56
; GFX908-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; GFX908-GCNTRACKERS-NEXT: v_accvgpr_write_b32 a5, v20
; GFX908-GCNTRACKERS-NEXT: buffer_load_dword v20, off, s[0:3], s33 offset:60
; GFX908-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; GFX908-GCNTRACKERS-NEXT: v_accvgpr_write_b32 a6, v20
; GFX908-GCNTRACKERS-NEXT: buffer_load_dword v20, off, s[0:3], s33 offset:64
; GFX908-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; GFX908-GCNTRACKERS-NEXT: v_accvgpr_write_b32 a7, v20
; GFX908-GCNTRACKERS-NEXT: buffer_load_dword v20, off, s[0:3], s33 offset:68
; GFX908-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; GFX908-GCNTRACKERS-NEXT: v_accvgpr_write_b32 a1, v20
; GFX908-GCNTRACKERS-NEXT: buffer_load_dword v20, off, s[0:3], s33 offset:72
; GFX908-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; GFX908-GCNTRACKERS-NEXT: v_accvgpr_write_b32 a0, v20
; GFX908-GCNTRACKERS-NEXT: buffer_load_dword v20, off, s[0:3], s33 offset:76
; GFX908-GCNTRACKERS-NEXT: buffer_load_dword v21, off, s[0:3], s33 offset:80
; GFX908-GCNTRACKERS-NEXT: buffer_load_dword v22, off, s[0:3], s33 offset:84
; GFX908-GCNTRACKERS-NEXT: buffer_load_dword v23, off, s[0:3], s33 offset:88
; GFX908-GCNTRACKERS-NEXT: buffer_load_dword v24, off, s[0:3], s33 offset:92
; GFX908-GCNTRACKERS-NEXT: buffer_load_dword v25, off, s[0:3], s33 offset:96
; GFX908-GCNTRACKERS-NEXT: buffer_load_dword v26, off, s[0:3], s33 offset:100
; GFX908-GCNTRACKERS-NEXT: buffer_load_dword v27, off, s[0:3], s33 offset:104
; GFX908-GCNTRACKERS-NEXT: buffer_load_dword v28, off, s[0:3], s33 offset:108
; GFX908-GCNTRACKERS-NEXT: buffer_load_dword v29, off, s[0:3], s33 offset:112
; GFX908-GCNTRACKERS-NEXT: buffer_load_dword v30, off, s[0:3], s33 offset:116
; GFX908-GCNTRACKERS-NEXT: buffer_load_dword v31, off, s[0:3], s33 offset:120
; GFX908-GCNTRACKERS-NEXT: s_nop 0
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:4
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:8
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:12
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:16
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v35, off, s[0:3], s32 offset:20
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v36, off, s[0:3], s32 offset:24
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v37, off, s[0:3], s32 offset:28
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v38, off, s[0:3], s32 offset:32
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v39, off, s[0:3], s32 offset:36
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v48, off, s[0:3], s32 offset:40
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v49, off, s[0:3], s32 offset:44
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v50, off, s[0:3], s32 offset:48
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v51, off, s[0:3], s32 offset:52
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v52, off, s[0:3], s32 offset:56
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v53, off, s[0:3], s32 offset:60
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v54, off, s[0:3], s32 offset:64
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v55, off, s[0:3], s32 offset:68
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:72
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:76
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:80
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:84
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:88
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:92
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:96
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:100
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:104
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v0, 24
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:108
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:112
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:116
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:120
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:124
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:128
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:132
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:136
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:140
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:144
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:148
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:152
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:156
; GFX908-GCNTRACKERS-NEXT: buffer_store_dword v9, off, s[0:3], s32 offset:160
; GFX908-GCNTRACKERS-NEXT: v_lshrrev_b32_e64 v0, 6, s33
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v2, v12
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v3, v13
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v4, v14
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v5, v15
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v6, v16
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v7, v17
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v8, v18
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v9, v19
; GFX908-GCNTRACKERS-NEXT: v_add_u32_e32 v0, 0x200, v0
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v1, 42
; GFX908-GCNTRACKERS-NEXT: v_accvgpr_read_b32 v12, a2
; GFX908-GCNTRACKERS-NEXT: v_accvgpr_read_b32 v13, a3
; GFX908-GCNTRACKERS-NEXT: v_accvgpr_read_b32 v14, a4
; GFX908-GCNTRACKERS-NEXT: v_accvgpr_read_b32 v15, a5
; GFX908-GCNTRACKERS-NEXT: v_accvgpr_read_b32 v16, a6
; GFX908-GCNTRACKERS-NEXT: v_accvgpr_read_b32 v17, a7
; GFX908-GCNTRACKERS-NEXT: v_accvgpr_read_b32 v18, a1
; GFX908-GCNTRACKERS-NEXT: v_accvgpr_read_b32 v19, a0
; GFX908-GCNTRACKERS-NEXT: s_swappc_b64 s[30:31], s[36:37]
; GFX908-GCNTRACKERS-NEXT: v_readlane_b32 s30, v62, 0
; GFX908-GCNTRACKERS-NEXT: v_readlane_b32 s31, v62, 1
; GFX908-GCNTRACKERS-NEXT: v_accvgpr_read_b32 v61, a21 ; Reload Reuse
; GFX908-GCNTRACKERS-NEXT: v_accvgpr_read_b32 v60, a20 ; Reload Reuse
; GFX908-GCNTRACKERS-NEXT: v_accvgpr_read_b32 v59, a19 ; Reload Reuse
; GFX908-GCNTRACKERS-NEXT: v_accvgpr_read_b32 v58, a18 ; Reload Reuse
; GFX908-GCNTRACKERS-NEXT: v_accvgpr_read_b32 v57, a17 ; Reload Reuse
; GFX908-GCNTRACKERS-NEXT: v_accvgpr_read_b32 v56, a16 ; Reload Reuse
; GFX908-GCNTRACKERS-NEXT: v_accvgpr_read_b32 v47, a15 ; Reload Reuse
; GFX908-GCNTRACKERS-NEXT: v_accvgpr_read_b32 v46, a14 ; Reload Reuse
; GFX908-GCNTRACKERS-NEXT: v_accvgpr_read_b32 v45, a13 ; Reload Reuse
; GFX908-GCNTRACKERS-NEXT: v_accvgpr_read_b32 v44, a12 ; Reload Reuse
; GFX908-GCNTRACKERS-NEXT: v_accvgpr_read_b32 v43, a11 ; Reload Reuse
; GFX908-GCNTRACKERS-NEXT: v_accvgpr_read_b32 v42, a10 ; Reload Reuse
; GFX908-GCNTRACKERS-NEXT: v_accvgpr_read_b32 v41, a9 ; Reload Reuse
; GFX908-GCNTRACKERS-NEXT: v_accvgpr_read_b32 v40, a8 ; Reload Reuse
; GFX908-GCNTRACKERS-NEXT: s_mov_b32 s32, s34
; GFX908-GCNTRACKERS-NEXT: s_mov_b32 s34, s38
; GFX908-GCNTRACKERS-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX908-GCNTRACKERS-NEXT: buffer_load_dword v62, off, s[0:3], s33 offset:1024 ; 4-byte Folded Reload
; GFX908-GCNTRACKERS-NEXT: s_mov_b64 exec, s[36:37]
; GFX908-GCNTRACKERS-NEXT: s_mov_b32 s33, s35
; GFX908-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; GFX908-GCNTRACKERS-NEXT: s_setpc_b64 s[30:31]
;
; GENERIC-LABEL: call_72xi32:
; GENERIC: ; %bb.0: ; %entry
; GENERIC-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GENERIC-NEXT: s_mov_b32 s35, s33
; GENERIC-NEXT: s_add_i32 s33, s32, 0x7fc0
; GENERIC-NEXT: s_and_b32 s33, s33, 0xffff8000
; GENERIC-NEXT: s_or_saveexec_b64 s[36:37], -1
; GENERIC-NEXT: buffer_store_dword v63, off, s[0:3], s33 offset:1564 ; 4-byte Folded Spill
; GENERIC-NEXT: s_mov_b64 exec, s[36:37]
; GENERIC-NEXT: s_mov_b32 s38, s34
; GENERIC-NEXT: s_mov_b32 s34, s32
; GENERIC-NEXT: s_add_i32 s32, s32, 0x28000
; GENERIC-NEXT: buffer_store_dword v40, off, s[0:3], s33 offset:56 ; 4-byte Folded Spill
; GENERIC-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:52 ; 4-byte Folded Spill
; GENERIC-NEXT: buffer_store_dword v42, off, s[0:3], s33 offset:48 ; 4-byte Folded Spill
; GENERIC-NEXT: buffer_store_dword v43, off, s[0:3], s33 offset:44 ; 4-byte Folded Spill
; GENERIC-NEXT: buffer_store_dword v44, off, s[0:3], s33 offset:40 ; 4-byte Folded Spill
; GENERIC-NEXT: buffer_store_dword v45, off, s[0:3], s33 offset:36 ; 4-byte Folded Spill
; GENERIC-NEXT: buffer_store_dword v46, off, s[0:3], s33 offset:32 ; 4-byte Folded Spill
; GENERIC-NEXT: buffer_store_dword v47, off, s[0:3], s33 offset:28 ; 4-byte Folded Spill
; GENERIC-NEXT: buffer_store_dword v56, off, s[0:3], s33 offset:24 ; 4-byte Folded Spill
; GENERIC-NEXT: buffer_store_dword v57, off, s[0:3], s33 offset:20 ; 4-byte Folded Spill
; GENERIC-NEXT: buffer_store_dword v58, off, s[0:3], s33 offset:16 ; 4-byte Folded Spill
; GENERIC-NEXT: buffer_store_dword v59, off, s[0:3], s33 offset:12 ; 4-byte Folded Spill
; GENERIC-NEXT: buffer_store_dword v60, off, s[0:3], s33 offset:8 ; 4-byte Folded Spill
; GENERIC-NEXT: buffer_store_dword v61, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GENERIC-NEXT: buffer_store_dword v62, off, s[0:3], s33 ; 4-byte Folded Spill
; GENERIC-NEXT: v_writelane_b32 v63, s30, 0
; GENERIC-NEXT: v_writelane_b32 v63, s31, 1
; GENERIC-NEXT: s_getpc_b64 s[36:37]
; GENERIC-NEXT: s_add_u32 s36, s36, return_72xi32@gotpcrel32@lo+4
; GENERIC-NEXT: s_addc_u32 s37, s37, return_72xi32@gotpcrel32@hi+12
; GENERIC-NEXT: s_load_dwordx2 s[36:37], s[36:37], 0x0
; GENERIC-NEXT: v_mov_b32_e32 v0, 0
; GENERIC-NEXT: buffer_store_dword v0, off, s[0:3], s32
; GENERIC-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:4
; GENERIC-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:8
; GENERIC-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:12
; GENERIC-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:16
; GENERIC-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:20
; GENERIC-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:24
; GENERIC-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:28
; GENERIC-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:32
; GENERIC-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:36
; GENERIC-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:40
; GENERIC-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:44
; GENERIC-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:48
; GENERIC-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:52
; GENERIC-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:56
; GENERIC-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:60
; GENERIC-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:64
; GENERIC-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:68
; GENERIC-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:72
; GENERIC-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:76
; GENERIC-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:80
; GENERIC-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:84
; GENERIC-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:88
; GENERIC-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:92
; GENERIC-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:96
; GENERIC-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:100
; GENERIC-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:104
; GENERIC-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:108
; GENERIC-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:112
; GENERIC-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:116
; GENERIC-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:120
; GENERIC-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:124
; GENERIC-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:128
; GENERIC-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:132
; GENERIC-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:136
; GENERIC-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:140
; GENERIC-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:144
; GENERIC-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:148
; GENERIC-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:152
; GENERIC-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:156
; GENERIC-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:160
; GENERIC-NEXT: s_waitcnt expcnt(0)
; GENERIC-NEXT: v_lshr_b32_e64 v0, s33, 6
; GENERIC-NEXT: s_movk_i32 vcc_lo, 0x200
; GENERIC-NEXT: v_add_i32_e32 v0, vcc, vcc_lo, v0
; GENERIC-NEXT: v_mov_b32_e32 v1, 0
; GENERIC-NEXT: v_mov_b32_e32 v2, 0
; GENERIC-NEXT: v_mov_b32_e32 v3, 0
; GENERIC-NEXT: v_mov_b32_e32 v4, 0
; GENERIC-NEXT: v_mov_b32_e32 v5, 0
; GENERIC-NEXT: v_mov_b32_e32 v6, 0
; GENERIC-NEXT: v_mov_b32_e32 v7, 0
; GENERIC-NEXT: v_mov_b32_e32 v8, 0
; GENERIC-NEXT: v_mov_b32_e32 v9, 0
; GENERIC-NEXT: v_mov_b32_e32 v10, 0
; GENERIC-NEXT: v_mov_b32_e32 v11, 0
; GENERIC-NEXT: v_mov_b32_e32 v12, 0
; GENERIC-NEXT: v_mov_b32_e32 v13, 0
; GENERIC-NEXT: v_mov_b32_e32 v14, 0
; GENERIC-NEXT: v_mov_b32_e32 v15, 0
; GENERIC-NEXT: v_mov_b32_e32 v16, 0
; GENERIC-NEXT: v_mov_b32_e32 v17, 0
; GENERIC-NEXT: v_mov_b32_e32 v18, 0
; GENERIC-NEXT: v_mov_b32_e32 v19, 0
; GENERIC-NEXT: v_mov_b32_e32 v20, 0
; GENERIC-NEXT: v_mov_b32_e32 v21, 0
; GENERIC-NEXT: v_mov_b32_e32 v22, 0
; GENERIC-NEXT: v_mov_b32_e32 v23, 0
; GENERIC-NEXT: v_mov_b32_e32 v24, 0
; GENERIC-NEXT: v_mov_b32_e32 v25, 0
; GENERIC-NEXT: v_mov_b32_e32 v26, 0
; GENERIC-NEXT: v_mov_b32_e32 v27, 0
; GENERIC-NEXT: v_mov_b32_e32 v28, 0
; GENERIC-NEXT: v_mov_b32_e32 v29, 0
; GENERIC-NEXT: v_mov_b32_e32 v30, 0
; GENERIC-NEXT: v_mov_b32_e32 v31, 0
; GENERIC-NEXT: s_waitcnt lgkmcnt(0)
; GENERIC-NEXT: s_swappc_b64 s[30:31], s[36:37]
; GENERIC-NEXT: buffer_load_dword v0, off, s[0:3], s33 offset:636
; GENERIC-NEXT: buffer_load_dword v1, off, s[0:3], s33 offset:640
; GENERIC-NEXT: buffer_load_dword v32, off, s[0:3], s33 offset:644
; GENERIC-NEXT: buffer_load_dword v33, off, s[0:3], s33 offset:648
; GENERIC-NEXT: buffer_load_dword v34, off, s[0:3], s33 offset:652
; GENERIC-NEXT: buffer_load_dword v35, off, s[0:3], s33 offset:656
; GENERIC-NEXT: buffer_load_dword v36, off, s[0:3], s33 offset:660
; GENERIC-NEXT: buffer_load_dword v37, off, s[0:3], s33 offset:664
; GENERIC-NEXT: buffer_load_dword v38, off, s[0:3], s33 offset:668
; GENERIC-NEXT: buffer_load_dword v39, off, s[0:3], s33 offset:672
; GENERIC-NEXT: buffer_load_dword v48, off, s[0:3], s33 offset:676
; GENERIC-NEXT: buffer_load_dword v49, off, s[0:3], s33 offset:680
; GENERIC-NEXT: buffer_load_dword v50, off, s[0:3], s33 offset:684
; GENERIC-NEXT: buffer_load_dword v51, off, s[0:3], s33 offset:688
; GENERIC-NEXT: buffer_load_dword v52, off, s[0:3], s33 offset:692
; GENERIC-NEXT: buffer_load_dword v53, off, s[0:3], s33 offset:696
; GENERIC-NEXT: buffer_load_dword v54, off, s[0:3], s33 offset:700
; GENERIC-NEXT: buffer_load_dword v55, off, s[0:3], s33 offset:704
; GENERIC-NEXT: buffer_load_dword v40, off, s[0:3], s33 offset:708
; GENERIC-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:712
; GENERIC-NEXT: buffer_load_dword v42, off, s[0:3], s33 offset:716
; GENERIC-NEXT: buffer_load_dword v43, off, s[0:3], s33 offset:720
; GENERIC-NEXT: buffer_load_dword v44, off, s[0:3], s33 offset:724
; GENERIC-NEXT: buffer_load_dword v45, off, s[0:3], s33 offset:728
; GENERIC-NEXT: buffer_load_dword v46, off, s[0:3], s33 offset:732
; GENERIC-NEXT: buffer_load_dword v47, off, s[0:3], s33 offset:736
; GENERIC-NEXT: buffer_load_dword v56, off, s[0:3], s33 offset:740
; GENERIC-NEXT: buffer_load_dword v57, off, s[0:3], s33 offset:748
; GENERIC-NEXT: buffer_load_dword v58, off, s[0:3], s33 offset:752
; GENERIC-NEXT: buffer_load_dword v59, off, s[0:3], s33 offset:756
; GENERIC-NEXT: buffer_load_dword v60, off, s[0:3], s33 offset:760
; GENERIC-NEXT: buffer_load_dword v61, off, s[0:3], s33 offset:764
; GENERIC-NEXT: buffer_load_dword v62, off, s[0:3], s33 offset:768
; GENERIC-NEXT: buffer_load_dword v2, off, s[0:3], s33 offset:772
; GENERIC-NEXT: buffer_load_dword v3, off, s[0:3], s33 offset:776
; GENERIC-NEXT: buffer_load_dword v4, off, s[0:3], s33 offset:780
; GENERIC-NEXT: buffer_load_dword v5, off, s[0:3], s33 offset:784
; GENERIC-NEXT: buffer_load_dword v6, off, s[0:3], s33 offset:788
; GENERIC-NEXT: buffer_load_dword v7, off, s[0:3], s33 offset:792
; GENERIC-NEXT: buffer_load_dword v8, off, s[0:3], s33 offset:796
; GENERIC-NEXT: buffer_load_dword v9, off, s[0:3], s33 offset:516
; GENERIC-NEXT: s_movk_i32 vcc_lo, 0x400
; GENERIC-NEXT: s_waitcnt vmcnt(0)
; GENERIC-NEXT: buffer_store_dword v9, off, s[0:3], s33 offset:1536 ; 4-byte Folded Spill
; GENERIC-NEXT: s_waitcnt expcnt(0)
; GENERIC-NEXT: buffer_load_dword v9, off, s[0:3], s33 offset:520
; GENERIC-NEXT: s_waitcnt vmcnt(0)
; GENERIC-NEXT: buffer_store_dword v9, off, s[0:3], s33 offset:1540 ; 4-byte Folded Spill
; GENERIC-NEXT: s_waitcnt expcnt(0)
; GENERIC-NEXT: buffer_load_dword v9, off, s[0:3], s33 offset:524
; GENERIC-NEXT: s_waitcnt vmcnt(0)
; GENERIC-NEXT: buffer_store_dword v9, off, s[0:3], s33 offset:1544 ; 4-byte Folded Spill
; GENERIC-NEXT: s_waitcnt expcnt(0)
; GENERIC-NEXT: buffer_load_dword v9, off, s[0:3], s33 offset:528
; GENERIC-NEXT: s_waitcnt vmcnt(0)
; GENERIC-NEXT: buffer_store_dword v9, off, s[0:3], s33 offset:1548 ; 4-byte Folded Spill
; GENERIC-NEXT: s_waitcnt expcnt(0)
; GENERIC-NEXT: buffer_load_dword v9, off, s[0:3], s33 offset:532
; GENERIC-NEXT: s_waitcnt vmcnt(0)
; GENERIC-NEXT: buffer_store_dword v9, off, s[0:3], s33 offset:1552 ; 4-byte Folded Spill
; GENERIC-NEXT: s_waitcnt expcnt(0)
; GENERIC-NEXT: buffer_load_dword v9, off, s[0:3], s33 offset:536
; GENERIC-NEXT: s_waitcnt vmcnt(0)
; GENERIC-NEXT: buffer_store_dword v9, off, s[0:3], s33 offset:1556 ; 4-byte Folded Spill
; GENERIC-NEXT: s_waitcnt expcnt(0)
; GENERIC-NEXT: buffer_load_dword v9, off, s[0:3], s33 offset:540
; GENERIC-NEXT: s_waitcnt vmcnt(0)
; GENERIC-NEXT: buffer_store_dword v9, off, s[0:3], s33 offset:1560 ; 4-byte Folded Spill
; GENERIC-NEXT: s_waitcnt expcnt(0)
; GENERIC-NEXT: buffer_load_dword v9, off, s[0:3], s33 offset:544
; GENERIC-NEXT: buffer_load_dword v10, off, s[0:3], s33 offset:548
; GENERIC-NEXT: buffer_load_dword v11, off, s[0:3], s33 offset:552
; GENERIC-NEXT: buffer_load_dword v12, off, s[0:3], s33 offset:556
; GENERIC-NEXT: buffer_load_dword v13, off, s[0:3], s33 offset:560
; GENERIC-NEXT: buffer_load_dword v14, off, s[0:3], s33 offset:564
; GENERIC-NEXT: buffer_load_dword v15, off, s[0:3], s33 offset:568
; GENERIC-NEXT: buffer_load_dword v16, off, s[0:3], s33 offset:572
; GENERIC-NEXT: buffer_load_dword v17, off, s[0:3], s33 offset:576
; GENERIC-NEXT: buffer_load_dword v18, off, s[0:3], s33 offset:580
; GENERIC-NEXT: buffer_load_dword v19, off, s[0:3], s33 offset:584
; GENERIC-NEXT: buffer_load_dword v20, off, s[0:3], s33 offset:588
; GENERIC-NEXT: buffer_load_dword v21, off, s[0:3], s33 offset:592
; GENERIC-NEXT: buffer_load_dword v22, off, s[0:3], s33 offset:596
; GENERIC-NEXT: buffer_load_dword v23, off, s[0:3], s33 offset:600
; GENERIC-NEXT: buffer_load_dword v24, off, s[0:3], s33 offset:604
; GENERIC-NEXT: buffer_load_dword v25, off, s[0:3], s33 offset:608
; GENERIC-NEXT: buffer_load_dword v26, off, s[0:3], s33 offset:612
; GENERIC-NEXT: buffer_load_dword v27, off, s[0:3], s33 offset:616
; GENERIC-NEXT: buffer_load_dword v28, off, s[0:3], s33 offset:620
; GENERIC-NEXT: buffer_load_dword v29, off, s[0:3], s33 offset:624
; GENERIC-NEXT: buffer_load_dword v30, off, s[0:3], s33 offset:628
; GENERIC-NEXT: buffer_load_dword v31, off, s[0:3], s33 offset:632
; GENERIC-NEXT: buffer_store_dword v0, off, s[0:3], s32
; GENERIC-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:4
; GENERIC-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:8
; GENERIC-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:12
; GENERIC-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:16
; GENERIC-NEXT: buffer_store_dword v35, off, s[0:3], s32 offset:20
; GENERIC-NEXT: buffer_store_dword v36, off, s[0:3], s32 offset:24
; GENERIC-NEXT: buffer_store_dword v37, off, s[0:3], s32 offset:28
; GENERIC-NEXT: buffer_store_dword v38, off, s[0:3], s32 offset:32
; GENERIC-NEXT: buffer_store_dword v39, off, s[0:3], s32 offset:36
; GENERIC-NEXT: buffer_store_dword v48, off, s[0:3], s32 offset:40
; GENERIC-NEXT: buffer_store_dword v49, off, s[0:3], s32 offset:44
; GENERIC-NEXT: buffer_store_dword v50, off, s[0:3], s32 offset:48
; GENERIC-NEXT: buffer_store_dword v51, off, s[0:3], s32 offset:52
; GENERIC-NEXT: buffer_store_dword v52, off, s[0:3], s32 offset:56
; GENERIC-NEXT: buffer_store_dword v53, off, s[0:3], s32 offset:60
; GENERIC-NEXT: buffer_store_dword v54, off, s[0:3], s32 offset:64
; GENERIC-NEXT: buffer_store_dword v55, off, s[0:3], s32 offset:68
; GENERIC-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:72
; GENERIC-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:76
; GENERIC-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:80
; GENERIC-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:84
; GENERIC-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:88
; GENERIC-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:92
; GENERIC-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:96
; GENERIC-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:100
; GENERIC-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:104
; GENERIC-NEXT: v_mov_b32_e32 v0, 24
; GENERIC-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:108
; GENERIC-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:112
; GENERIC-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:116
; GENERIC-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:120
; GENERIC-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:124
; GENERIC-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:128
; GENERIC-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:132
; GENERIC-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:136
; GENERIC-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:140
; GENERIC-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:144
; GENERIC-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:148
; GENERIC-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:152
; GENERIC-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:156
; GENERIC-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:160
; GENERIC-NEXT: s_waitcnt expcnt(6)
; GENERIC-NEXT: buffer_load_dword v2, off, s[0:3], s33 offset:1536 ; 4-byte Folded Reload
; GENERIC-NEXT: s_waitcnt expcnt(5)
; GENERIC-NEXT: buffer_load_dword v3, off, s[0:3], s33 offset:1540 ; 4-byte Folded Reload
; GENERIC-NEXT: s_waitcnt expcnt(4)
; GENERIC-NEXT: buffer_load_dword v4, off, s[0:3], s33 offset:1544 ; 4-byte Folded Reload
; GENERIC-NEXT: s_waitcnt expcnt(3)
; GENERIC-NEXT: buffer_load_dword v5, off, s[0:3], s33 offset:1548 ; 4-byte Folded Reload
; GENERIC-NEXT: s_waitcnt expcnt(2)
; GENERIC-NEXT: buffer_load_dword v6, off, s[0:3], s33 offset:1552 ; 4-byte Folded Reload
; GENERIC-NEXT: s_waitcnt expcnt(1)
; GENERIC-NEXT: buffer_load_dword v7, off, s[0:3], s33 offset:1556 ; 4-byte Folded Reload
; GENERIC-NEXT: s_waitcnt expcnt(0)
; GENERIC-NEXT: buffer_load_dword v8, off, s[0:3], s33 offset:1560 ; 4-byte Folded Reload
; GENERIC-NEXT: v_lshr_b32_e64 v0, s33, 6
; GENERIC-NEXT: v_add_i32_e32 v0, vcc, vcc_lo, v0
; GENERIC-NEXT: v_mov_b32_e32 v1, 42
; GENERIC-NEXT: s_swappc_b64 s[30:31], s[36:37]
; GENERIC-NEXT: buffer_load_dword v62, off, s[0:3], s33 ; 4-byte Folded Reload
; GENERIC-NEXT: buffer_load_dword v61, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; GENERIC-NEXT: buffer_load_dword v60, off, s[0:3], s33 offset:8 ; 4-byte Folded Reload
; GENERIC-NEXT: buffer_load_dword v59, off, s[0:3], s33 offset:12 ; 4-byte Folded Reload
; GENERIC-NEXT: buffer_load_dword v58, off, s[0:3], s33 offset:16 ; 4-byte Folded Reload
; GENERIC-NEXT: buffer_load_dword v57, off, s[0:3], s33 offset:20 ; 4-byte Folded Reload
; GENERIC-NEXT: buffer_load_dword v56, off, s[0:3], s33 offset:24 ; 4-byte Folded Reload
; GENERIC-NEXT: buffer_load_dword v47, off, s[0:3], s33 offset:28 ; 4-byte Folded Reload
; GENERIC-NEXT: buffer_load_dword v46, off, s[0:3], s33 offset:32 ; 4-byte Folded Reload
; GENERIC-NEXT: buffer_load_dword v45, off, s[0:3], s33 offset:36 ; 4-byte Folded Reload
; GENERIC-NEXT: buffer_load_dword v44, off, s[0:3], s33 offset:40 ; 4-byte Folded Reload
; GENERIC-NEXT: buffer_load_dword v43, off, s[0:3], s33 offset:44 ; 4-byte Folded Reload
; GENERIC-NEXT: buffer_load_dword v42, off, s[0:3], s33 offset:48 ; 4-byte Folded Reload
; GENERIC-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:52 ; 4-byte Folded Reload
; GENERIC-NEXT: buffer_load_dword v40, off, s[0:3], s33 offset:56 ; 4-byte Folded Reload
; GENERIC-NEXT: v_readlane_b32 s30, v63, 0
; GENERIC-NEXT: v_readlane_b32 s31, v63, 1
; GENERIC-NEXT: s_mov_b32 s32, s34
; GENERIC-NEXT: s_mov_b32 s34, s38
; GENERIC-NEXT: s_or_saveexec_b64 s[36:37], -1
; GENERIC-NEXT: buffer_load_dword v63, off, s[0:3], s33 offset:1564 ; 4-byte Folded Reload
; GENERIC-NEXT: s_mov_b64 exec, s[36:37]
; GENERIC-NEXT: s_mov_b32 s33, s35
; GENERIC-NEXT: s_waitcnt vmcnt(0)
; GENERIC-NEXT: s_setpc_b64 s[30:31]
;
; GENERIC-GCNTRACKERS-LABEL: call_72xi32:
; GENERIC-GCNTRACKERS: ; %bb.0: ; %entry
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GENERIC-GCNTRACKERS-NEXT: s_mov_b32 s35, s33
; GENERIC-GCNTRACKERS-NEXT: s_add_i32 s33, s32, 0x7fc0
; GENERIC-GCNTRACKERS-NEXT: s_and_b32 s33, s33, 0xffff8000
; GENERIC-GCNTRACKERS-NEXT: s_or_saveexec_b64 s[36:37], -1
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v63, off, s[0:3], s33 offset:1564 ; 4-byte Folded Spill
; GENERIC-GCNTRACKERS-NEXT: s_mov_b64 exec, s[36:37]
; GENERIC-GCNTRACKERS-NEXT: s_mov_b32 s38, s34
; GENERIC-GCNTRACKERS-NEXT: s_mov_b32 s34, s32
; GENERIC-GCNTRACKERS-NEXT: s_add_i32 s32, s32, 0x28000
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v40, off, s[0:3], s33 offset:56 ; 4-byte Folded Spill
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:52 ; 4-byte Folded Spill
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v42, off, s[0:3], s33 offset:48 ; 4-byte Folded Spill
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v43, off, s[0:3], s33 offset:44 ; 4-byte Folded Spill
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v44, off, s[0:3], s33 offset:40 ; 4-byte Folded Spill
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v45, off, s[0:3], s33 offset:36 ; 4-byte Folded Spill
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v46, off, s[0:3], s33 offset:32 ; 4-byte Folded Spill
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v47, off, s[0:3], s33 offset:28 ; 4-byte Folded Spill
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v56, off, s[0:3], s33 offset:24 ; 4-byte Folded Spill
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v57, off, s[0:3], s33 offset:20 ; 4-byte Folded Spill
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v58, off, s[0:3], s33 offset:16 ; 4-byte Folded Spill
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v59, off, s[0:3], s33 offset:12 ; 4-byte Folded Spill
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v60, off, s[0:3], s33 offset:8 ; 4-byte Folded Spill
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v61, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v62, off, s[0:3], s33 ; 4-byte Folded Spill
; GENERIC-GCNTRACKERS-NEXT: v_writelane_b32 v63, s30, 0
; GENERIC-GCNTRACKERS-NEXT: v_writelane_b32 v63, s31, 1
; GENERIC-GCNTRACKERS-NEXT: s_getpc_b64 s[36:37]
; GENERIC-GCNTRACKERS-NEXT: s_add_u32 s36, s36, return_72xi32@gotpcrel32@lo+4
; GENERIC-GCNTRACKERS-NEXT: s_addc_u32 s37, s37, return_72xi32@gotpcrel32@hi+12
; GENERIC-GCNTRACKERS-NEXT: s_load_dwordx2 s[36:37], s[36:37], 0x0
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v0, 0
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:4
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:8
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:12
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:16
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:20
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:24
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:28
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:32
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:36
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:40
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:44
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:48
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:52
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:56
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:60
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:64
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:68
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:72
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:76
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:80
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:84
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:88
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:92
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:96
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:100
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:104
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:108
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:112
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:116
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:120
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:124
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:128
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:132
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:136
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:140
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:144
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:148
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:152
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:156
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:160
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt expcnt(0)
; GENERIC-GCNTRACKERS-NEXT: v_lshr_b32_e64 v0, s33, 6
; GENERIC-GCNTRACKERS-NEXT: s_movk_i32 vcc_lo, 0x200
; GENERIC-GCNTRACKERS-NEXT: v_add_i32_e32 v0, vcc, vcc_lo, v0
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v1, 0
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v2, 0
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v3, 0
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v4, 0
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v5, 0
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v6, 0
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v7, 0
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v8, 0
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v9, 0
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v10, 0
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v11, 0
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v12, 0
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v13, 0
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v14, 0
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v15, 0
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v16, 0
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v17, 0
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v18, 0
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v19, 0
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v20, 0
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v21, 0
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v22, 0
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v23, 0
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v24, 0
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v25, 0
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v26, 0
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v27, 0
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v28, 0
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v29, 0
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v30, 0
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v31, 0
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt lgkmcnt(0)
; GENERIC-GCNTRACKERS-NEXT: s_swappc_b64 s[30:31], s[36:37]
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v0, off, s[0:3], s33 offset:636
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v1, off, s[0:3], s33 offset:640
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v32, off, s[0:3], s33 offset:644
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v33, off, s[0:3], s33 offset:648
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v34, off, s[0:3], s33 offset:652
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v35, off, s[0:3], s33 offset:656
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v36, off, s[0:3], s33 offset:660
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v37, off, s[0:3], s33 offset:664
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v38, off, s[0:3], s33 offset:668
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v39, off, s[0:3], s33 offset:672
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v48, off, s[0:3], s33 offset:676
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v49, off, s[0:3], s33 offset:680
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v50, off, s[0:3], s33 offset:684
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v51, off, s[0:3], s33 offset:688
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v52, off, s[0:3], s33 offset:692
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v53, off, s[0:3], s33 offset:696
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v54, off, s[0:3], s33 offset:700
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v55, off, s[0:3], s33 offset:704
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v40, off, s[0:3], s33 offset:708
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:712
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v42, off, s[0:3], s33 offset:716
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v43, off, s[0:3], s33 offset:720
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v44, off, s[0:3], s33 offset:724
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v45, off, s[0:3], s33 offset:728
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v46, off, s[0:3], s33 offset:732
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v47, off, s[0:3], s33 offset:736
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v56, off, s[0:3], s33 offset:740
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v57, off, s[0:3], s33 offset:748
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v58, off, s[0:3], s33 offset:752
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v59, off, s[0:3], s33 offset:756
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v60, off, s[0:3], s33 offset:760
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v61, off, s[0:3], s33 offset:764
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v62, off, s[0:3], s33 offset:768
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v2, off, s[0:3], s33 offset:772
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v3, off, s[0:3], s33 offset:776
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v4, off, s[0:3], s33 offset:780
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v5, off, s[0:3], s33 offset:784
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v6, off, s[0:3], s33 offset:788
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v7, off, s[0:3], s33 offset:792
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v8, off, s[0:3], s33 offset:796
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v12, off, s[0:3], s33 offset:516
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v13, off, s[0:3], s33 offset:520
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v14, off, s[0:3], s33 offset:524
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v15, off, s[0:3], s33 offset:528
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v16, off, s[0:3], s33 offset:532
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v17, off, s[0:3], s33 offset:536
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v18, off, s[0:3], s33 offset:540
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v9, off, s[0:3], s33 offset:544
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v10, off, s[0:3], s33 offset:548
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v11, off, s[0:3], s33 offset:552
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v19, off, s[0:3], s33 offset:556
; GENERIC-GCNTRACKERS-NEXT: s_movk_i32 vcc_lo, 0x400
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v19, off, s[0:3], s33 offset:1540 ; 4-byte Folded Spill
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt expcnt(0)
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v19, off, s[0:3], s33 offset:560
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v19, off, s[0:3], s33 offset:1544 ; 4-byte Folded Spill
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt expcnt(0)
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v19, off, s[0:3], s33 offset:564
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v19, off, s[0:3], s33 offset:1548 ; 4-byte Folded Spill
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt expcnt(0)
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v19, off, s[0:3], s33 offset:568
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v19, off, s[0:3], s33 offset:1552 ; 4-byte Folded Spill
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt expcnt(0)
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v19, off, s[0:3], s33 offset:572
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v19, off, s[0:3], s33 offset:1556 ; 4-byte Folded Spill
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt expcnt(0)
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v19, off, s[0:3], s33 offset:576
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v19, off, s[0:3], s33 offset:1560 ; 4-byte Folded Spill
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt expcnt(0)
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v19, off, s[0:3], s33 offset:580
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v19, off, s[0:3], s33 offset:1536 ; 4-byte Folded Spill
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt expcnt(0)
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v19, off, s[0:3], s33 offset:584
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v20, off, s[0:3], s33 offset:588
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v21, off, s[0:3], s33 offset:592
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v22, off, s[0:3], s33 offset:596
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v23, off, s[0:3], s33 offset:600
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v24, off, s[0:3], s33 offset:604
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v25, off, s[0:3], s33 offset:608
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v26, off, s[0:3], s33 offset:612
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v27, off, s[0:3], s33 offset:616
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v28, off, s[0:3], s33 offset:620
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v29, off, s[0:3], s33 offset:624
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v30, off, s[0:3], s33 offset:628
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v31, off, s[0:3], s33 offset:632
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:4
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:8
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:12
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:16
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v35, off, s[0:3], s32 offset:20
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v36, off, s[0:3], s32 offset:24
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v37, off, s[0:3], s32 offset:28
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v38, off, s[0:3], s32 offset:32
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v39, off, s[0:3], s32 offset:36
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v48, off, s[0:3], s32 offset:40
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v49, off, s[0:3], s32 offset:44
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v50, off, s[0:3], s32 offset:48
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v51, off, s[0:3], s32 offset:52
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v52, off, s[0:3], s32 offset:56
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v53, off, s[0:3], s32 offset:60
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v54, off, s[0:3], s32 offset:64
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v55, off, s[0:3], s32 offset:68
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:72
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:76
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:80
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:84
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:88
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:92
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:96
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:100
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:104
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v0, 24
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:108
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:112
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:116
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:120
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:124
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:128
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:132
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:136
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:140
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:144
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:148
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:152
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:156
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:160
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt expcnt(6)
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v2, v12
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt expcnt(5)
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v3, v13
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt expcnt(4)
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v4, v14
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt expcnt(3)
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v5, v15
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt expcnt(2)
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v6, v16
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt expcnt(1)
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v7, v17
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt expcnt(0)
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v8, v18
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v12, off, s[0:3], s33 offset:1540 ; 4-byte Folded Reload
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v13, off, s[0:3], s33 offset:1544 ; 4-byte Folded Reload
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v14, off, s[0:3], s33 offset:1548 ; 4-byte Folded Reload
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v15, off, s[0:3], s33 offset:1552 ; 4-byte Folded Reload
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v16, off, s[0:3], s33 offset:1556 ; 4-byte Folded Reload
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v17, off, s[0:3], s33 offset:1560 ; 4-byte Folded Reload
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v18, off, s[0:3], s33 offset:1536 ; 4-byte Folded Reload
; GENERIC-GCNTRACKERS-NEXT: v_lshr_b32_e64 v0, s33, 6
; GENERIC-GCNTRACKERS-NEXT: v_add_i32_e32 v0, vcc, vcc_lo, v0
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v1, 42
; GENERIC-GCNTRACKERS-NEXT: s_swappc_b64 s[30:31], s[36:37]
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v62, off, s[0:3], s33 ; 4-byte Folded Reload
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v61, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v60, off, s[0:3], s33 offset:8 ; 4-byte Folded Reload
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v59, off, s[0:3], s33 offset:12 ; 4-byte Folded Reload
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v58, off, s[0:3], s33 offset:16 ; 4-byte Folded Reload
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v57, off, s[0:3], s33 offset:20 ; 4-byte Folded Reload
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v56, off, s[0:3], s33 offset:24 ; 4-byte Folded Reload
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v47, off, s[0:3], s33 offset:28 ; 4-byte Folded Reload
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v46, off, s[0:3], s33 offset:32 ; 4-byte Folded Reload
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v45, off, s[0:3], s33 offset:36 ; 4-byte Folded Reload
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v44, off, s[0:3], s33 offset:40 ; 4-byte Folded Reload
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v43, off, s[0:3], s33 offset:44 ; 4-byte Folded Reload
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v42, off, s[0:3], s33 offset:48 ; 4-byte Folded Reload
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:52 ; 4-byte Folded Reload
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v40, off, s[0:3], s33 offset:56 ; 4-byte Folded Reload
; GENERIC-GCNTRACKERS-NEXT: v_readlane_b32 s30, v63, 0
; GENERIC-GCNTRACKERS-NEXT: v_readlane_b32 s31, v63, 1
; GENERIC-GCNTRACKERS-NEXT: s_mov_b32 s32, s34
; GENERIC-GCNTRACKERS-NEXT: s_mov_b32 s34, s38
; GENERIC-GCNTRACKERS-NEXT: s_or_saveexec_b64 s[36:37], -1
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v63, off, s[0:3], s33 offset:1564 ; 4-byte Folded Reload
; GENERIC-GCNTRACKERS-NEXT: s_mov_b64 exec, s[36:37]
; GENERIC-GCNTRACKERS-NEXT: s_mov_b32 s33, s35
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; GENERIC-GCNTRACKERS-NEXT: s_setpc_b64 s[30:31]
entry:
%ret.0 = call amdgpu_gfx <72 x i32> @return_72xi32(<72 x i32> zeroinitializer)
%val.0 = insertelement <72 x i32> %ret.0, i32 42, i32 0
%val.1 = insertelement <72 x i32> %val.0, i32 24, i32 58
%ret.1 = call amdgpu_gfx <72 x i32> @return_72xi32(<72 x i32> %val.1)
ret void
}
; CHECK-LABEL: {{^}}global_extload_v16f16_to_v16f64:
; TONGA: NumSgprs: 96
; TONGA-GCNTRACKERS: NumSgprs: 96
; TONGA: NumVgprs: 21
; TONGA-GCNTRACKERS: NumVgprs: 23
; TONGA: Occupancy: 8
; TONGA-GCNTRACKERS: Occupancy: 8
define amdgpu_kernel void @global_extload_v16f16_to_v16f64(ptr addrspace(1) %out, ptr addrspace(1) %in) {
; GFX11-PAL-LABEL: global_extload_v16f16_to_v16f64:
; GFX11-PAL: ; %bb.0:
; GFX11-PAL-NEXT: s_load_b128 s[0:3], s[4:5], 0x0
; GFX11-PAL-NEXT: s_waitcnt lgkmcnt(0)
; GFX11-PAL-NEXT: s_load_b256 s[4:11], s[2:3], 0x0
; GFX11-PAL-NEXT: s_waitcnt lgkmcnt(0)
; GFX11-PAL-NEXT: v_cvt_f32_f16_e32 v14, s9
; GFX11-PAL-NEXT: s_lshr_b32 s9, s9, 16
; GFX11-PAL-NEXT: v_cvt_f32_f16_e32 v11, s8
; GFX11-PAL-NEXT: s_lshr_b32 s8, s8, 16
; GFX11-PAL-NEXT: v_cvt_f32_f16_e32 v2, s5
; GFX11-PAL-NEXT: s_lshr_b32 s3, s5, 16
; GFX11-PAL-NEXT: v_cvt_f32_f16_e32 v6, s7
; GFX11-PAL-NEXT: s_lshr_b32 s5, s7, 16
; GFX11-PAL-NEXT: s_lshr_b32 s7, s11, 16
; GFX11-PAL-NEXT: v_cvt_f32_f16_e32 v23, s9
; GFX11-PAL-NEXT: v_cvt_f32_f16_e32 v0, s4
; GFX11-PAL-NEXT: s_lshr_b32 s2, s4, 16
; GFX11-PAL-NEXT: v_cvt_f32_f16_e32 v3, s6
; GFX11-PAL-NEXT: s_lshr_b32 s4, s6, 16
; GFX11-PAL-NEXT: s_lshr_b32 s6, s10, 16
; GFX11-PAL-NEXT: v_cvt_f32_f16_e32 v22, s8
; GFX11-PAL-NEXT: v_cvt_f32_f16_e32 v10, s11
; GFX11-PAL-NEXT: v_cvt_f32_f16_e32 v19, s7
; GFX11-PAL-NEXT: v_cvt_f32_f16_e32 v7, s10
; GFX11-PAL-NEXT: v_cvt_f32_f16_e32 v18, s6
; GFX11-PAL-NEXT: v_cvt_f32_f16_e32 v15, s5
; GFX11-PAL-NEXT: v_cvt_f32_f16_e32 v34, s4
; GFX11-PAL-NEXT: v_cvt_f64_f32_e32 v[28:29], v14
; GFX11-PAL-NEXT: v_cvt_f64_f32_e32 v[30:31], v23
; GFX11-PAL-NEXT: v_cvt_f32_f16_e32 v33, s3
; GFX11-PAL-NEXT: v_cvt_f64_f32_e32 v[24:25], v11
; GFX11-PAL-NEXT: v_cvt_f64_f32_e32 v[26:27], v22
; GFX11-PAL-NEXT: v_cvt_f32_f16_e32 v32, s2
; GFX11-PAL-NEXT: v_cvt_f64_f32_e32 v[20:21], v10
; GFX11-PAL-NEXT: v_cvt_f64_f32_e32 v[22:23], v19
; GFX11-PAL-NEXT: v_cvt_f64_f32_e32 v[16:17], v7
; GFX11-PAL-NEXT: v_cvt_f64_f32_e32 v[18:19], v18
; GFX11-PAL-NEXT: v_cvt_f64_f32_e32 v[12:13], v6
; GFX11-PAL-NEXT: v_cvt_f64_f32_e32 v[14:15], v15
; GFX11-PAL-NEXT: v_cvt_f64_f32_e32 v[8:9], v3
; GFX11-PAL-NEXT: v_cvt_f64_f32_e32 v[10:11], v34
; GFX11-PAL-NEXT: v_cvt_f64_f32_e32 v[4:5], v2
; GFX11-PAL-NEXT: v_cvt_f64_f32_e32 v[6:7], v33
; GFX11-PAL-NEXT: v_cvt_f64_f32_e32 v[0:1], v0
; GFX11-PAL-NEXT: v_cvt_f64_f32_e32 v[2:3], v32
; GFX11-PAL-NEXT: v_mov_b32_e32 v32, 0
; GFX11-PAL-NEXT: s_clause 0x7
; GFX11-PAL-NEXT: global_store_b128 v32, v[28:31], s[0:1] offset:80
; GFX11-PAL-NEXT: global_store_b128 v32, v[24:27], s[0:1] offset:64
; GFX11-PAL-NEXT: global_store_b128 v32, v[20:23], s[0:1] offset:112
; GFX11-PAL-NEXT: global_store_b128 v32, v[16:19], s[0:1] offset:96
; GFX11-PAL-NEXT: global_store_b128 v32, v[12:15], s[0:1] offset:48
; GFX11-PAL-NEXT: global_store_b128 v32, v[8:11], s[0:1] offset:32
; GFX11-PAL-NEXT: global_store_b128 v32, v[4:7], s[0:1] offset:16
; GFX11-PAL-NEXT: global_store_b128 v32, v[0:3], s[0:1]
; GFX11-PAL-NEXT: s_endpgm
;
; GFX11-PAL-GCNTRACKERS-LABEL: global_extload_v16f16_to_v16f64:
; GFX11-PAL-GCNTRACKERS: ; %bb.0:
; GFX11-PAL-GCNTRACKERS-NEXT: s_load_b128 s[0:3], s[4:5], 0x0
; GFX11-PAL-GCNTRACKERS-NEXT: s_waitcnt lgkmcnt(0)
; GFX11-PAL-GCNTRACKERS-NEXT: s_load_b256 s[4:11], s[2:3], 0x0
; GFX11-PAL-GCNTRACKERS-NEXT: s_waitcnt lgkmcnt(0)
; GFX11-PAL-GCNTRACKERS-NEXT: v_cvt_f32_f16_e32 v14, s9
; GFX11-PAL-GCNTRACKERS-NEXT: s_lshr_b32 s9, s9, 16
; GFX11-PAL-GCNTRACKERS-NEXT: v_cvt_f32_f16_e32 v11, s8
; GFX11-PAL-GCNTRACKERS-NEXT: s_lshr_b32 s8, s8, 16
; GFX11-PAL-GCNTRACKERS-NEXT: v_cvt_f32_f16_e32 v2, s5
; GFX11-PAL-GCNTRACKERS-NEXT: s_lshr_b32 s3, s5, 16
; GFX11-PAL-GCNTRACKERS-NEXT: v_cvt_f32_f16_e32 v6, s7
; GFX11-PAL-GCNTRACKERS-NEXT: s_lshr_b32 s5, s7, 16
; GFX11-PAL-GCNTRACKERS-NEXT: s_lshr_b32 s7, s11, 16
; GFX11-PAL-GCNTRACKERS-NEXT: v_cvt_f32_f16_e32 v23, s9
; GFX11-PAL-GCNTRACKERS-NEXT: v_cvt_f32_f16_e32 v0, s4
; GFX11-PAL-GCNTRACKERS-NEXT: s_lshr_b32 s2, s4, 16
; GFX11-PAL-GCNTRACKERS-NEXT: v_cvt_f32_f16_e32 v3, s6
; GFX11-PAL-GCNTRACKERS-NEXT: s_lshr_b32 s4, s6, 16
; GFX11-PAL-GCNTRACKERS-NEXT: s_lshr_b32 s6, s10, 16
; GFX11-PAL-GCNTRACKERS-NEXT: v_cvt_f32_f16_e32 v22, s8
; GFX11-PAL-GCNTRACKERS-NEXT: v_cvt_f32_f16_e32 v10, s11
; GFX11-PAL-GCNTRACKERS-NEXT: v_cvt_f32_f16_e32 v19, s7
; GFX11-PAL-GCNTRACKERS-NEXT: v_cvt_f32_f16_e32 v7, s10
; GFX11-PAL-GCNTRACKERS-NEXT: v_cvt_f32_f16_e32 v18, s6
; GFX11-PAL-GCNTRACKERS-NEXT: v_cvt_f32_f16_e32 v15, s5
; GFX11-PAL-GCNTRACKERS-NEXT: v_cvt_f32_f16_e32 v34, s4
; GFX11-PAL-GCNTRACKERS-NEXT: v_cvt_f64_f32_e32 v[28:29], v14
; GFX11-PAL-GCNTRACKERS-NEXT: v_cvt_f64_f32_e32 v[30:31], v23
; GFX11-PAL-GCNTRACKERS-NEXT: v_cvt_f32_f16_e32 v33, s3
; GFX11-PAL-GCNTRACKERS-NEXT: v_cvt_f64_f32_e32 v[24:25], v11
; GFX11-PAL-GCNTRACKERS-NEXT: v_cvt_f64_f32_e32 v[26:27], v22
; GFX11-PAL-GCNTRACKERS-NEXT: v_cvt_f32_f16_e32 v32, s2
; GFX11-PAL-GCNTRACKERS-NEXT: v_cvt_f64_f32_e32 v[20:21], v10
; GFX11-PAL-GCNTRACKERS-NEXT: v_cvt_f64_f32_e32 v[22:23], v19
; GFX11-PAL-GCNTRACKERS-NEXT: v_cvt_f64_f32_e32 v[16:17], v7
; GFX11-PAL-GCNTRACKERS-NEXT: v_cvt_f64_f32_e32 v[18:19], v18
; GFX11-PAL-GCNTRACKERS-NEXT: v_cvt_f64_f32_e32 v[12:13], v6
; GFX11-PAL-GCNTRACKERS-NEXT: v_cvt_f64_f32_e32 v[14:15], v15
; GFX11-PAL-GCNTRACKERS-NEXT: v_cvt_f64_f32_e32 v[8:9], v3
; GFX11-PAL-GCNTRACKERS-NEXT: v_cvt_f64_f32_e32 v[10:11], v34
; GFX11-PAL-GCNTRACKERS-NEXT: v_cvt_f64_f32_e32 v[4:5], v2
; GFX11-PAL-GCNTRACKERS-NEXT: v_cvt_f64_f32_e32 v[6:7], v33
; GFX11-PAL-GCNTRACKERS-NEXT: v_cvt_f64_f32_e32 v[0:1], v0
; GFX11-PAL-GCNTRACKERS-NEXT: v_cvt_f64_f32_e32 v[2:3], v32
; GFX11-PAL-GCNTRACKERS-NEXT: v_mov_b32_e32 v32, 0
; GFX11-PAL-GCNTRACKERS-NEXT: s_clause 0x7
; GFX11-PAL-GCNTRACKERS-NEXT: global_store_b128 v32, v[28:31], s[0:1] offset:80
; GFX11-PAL-GCNTRACKERS-NEXT: global_store_b128 v32, v[24:27], s[0:1] offset:64
; GFX11-PAL-GCNTRACKERS-NEXT: global_store_b128 v32, v[20:23], s[0:1] offset:112
; GFX11-PAL-GCNTRACKERS-NEXT: global_store_b128 v32, v[16:19], s[0:1] offset:96
; GFX11-PAL-GCNTRACKERS-NEXT: global_store_b128 v32, v[12:15], s[0:1] offset:48
; GFX11-PAL-GCNTRACKERS-NEXT: global_store_b128 v32, v[8:11], s[0:1] offset:32
; GFX11-PAL-GCNTRACKERS-NEXT: global_store_b128 v32, v[4:7], s[0:1] offset:16
; GFX11-PAL-GCNTRACKERS-NEXT: global_store_b128 v32, v[0:3], s[0:1]
; GFX11-PAL-GCNTRACKERS-NEXT: s_endpgm
;
; TONGA-LABEL: global_extload_v16f16_to_v16f64:
; TONGA: ; %bb.0:
; TONGA-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24
; TONGA-NEXT: s_waitcnt lgkmcnt(0)
; TONGA-NEXT: v_mov_b32_e32 v0, s2
; TONGA-NEXT: v_mov_b32_e32 v1, s3
; TONGA-NEXT: flat_load_dwordx4 v[4:7], v[0:1]
; TONGA-NEXT: s_add_u32 s2, s2, 16
; TONGA-NEXT: s_addc_u32 s3, s3, 0
; TONGA-NEXT: v_mov_b32_e32 v0, s2
; TONGA-NEXT: v_mov_b32_e32 v1, s3
; TONGA-NEXT: flat_load_dwordx4 v[0:3], v[0:1]
; TONGA-NEXT: s_add_u32 s2, s0, 48
; TONGA-NEXT: s_addc_u32 s3, s1, 0
; TONGA-NEXT: v_mov_b32_e32 v14, s3
; TONGA-NEXT: v_mov_b32_e32 v13, s2
; TONGA-NEXT: s_add_u32 s2, s0, 32
; TONGA-NEXT: s_addc_u32 s3, s1, 0
; TONGA-NEXT: v_mov_b32_e32 v16, s3
; TONGA-NEXT: v_mov_b32_e32 v15, s2
; TONGA-NEXT: s_add_u32 s2, s0, 16
; TONGA-NEXT: s_addc_u32 s3, s1, 0
; TONGA-NEXT: v_mov_b32_e32 v18, s3
; TONGA-NEXT: v_mov_b32_e32 v17, s2
; TONGA-NEXT: v_mov_b32_e32 v12, s1
; TONGA-NEXT: v_mov_b32_e32 v11, s0
; TONGA-NEXT: s_add_u32 s2, s0, 0x50
; TONGA-NEXT: s_addc_u32 s3, s1, 0
; TONGA-NEXT: s_waitcnt vmcnt(1)
; TONGA-NEXT: v_cvt_f32_f16_e32 v8, v7
; TONGA-NEXT: v_cvt_f32_f16_sdwa v9, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1
; TONGA-NEXT: v_cvt_f64_f32_e32 v[7:8], v8
; TONGA-NEXT: v_cvt_f64_f32_e32 v[9:10], v9
; TONGA-NEXT: flat_store_dwordx4 v[13:14], v[7:10]
; TONGA-NEXT: v_mov_b32_e32 v14, s3
; TONGA-NEXT: v_cvt_f32_f16_e32 v7, v6
; TONGA-NEXT: v_cvt_f32_f16_sdwa v8, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1
; TONGA-NEXT: s_waitcnt vmcnt(1)
; TONGA-NEXT: v_cvt_f32_f16_e32 v10, v2
; TONGA-NEXT: v_mov_b32_e32 v13, s2
; TONGA-NEXT: v_cvt_f64_f32_e32 v[6:7], v7
; TONGA-NEXT: v_cvt_f64_f32_e32 v[8:9], v8
; TONGA-NEXT: s_add_u32 s2, s0, 64
; TONGA-NEXT: s_addc_u32 s3, s1, 0
; TONGA-NEXT: flat_store_dwordx4 v[15:16], v[6:9]
; TONGA-NEXT: v_mov_b32_e32 v16, s3
; TONGA-NEXT: v_cvt_f32_f16_e32 v6, v5
; TONGA-NEXT: v_cvt_f32_f16_sdwa v7, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1
; TONGA-NEXT: v_cvt_f32_f16_e32 v8, v4
; TONGA-NEXT: v_cvt_f32_f16_sdwa v9, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1
; TONGA-NEXT: v_cvt_f64_f32_e32 v[4:5], v6
; TONGA-NEXT: v_cvt_f64_f32_e32 v[6:7], v7
; TONGA-NEXT: v_mov_b32_e32 v15, s2
; TONGA-NEXT: s_add_u32 s2, s0, 0x70
; TONGA-NEXT: s_addc_u32 s3, s1, 0
; TONGA-NEXT: flat_store_dwordx4 v[17:18], v[4:7]
; TONGA-NEXT: v_cvt_f32_f16_sdwa v17, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1
; TONGA-NEXT: v_cvt_f64_f32_e32 v[4:5], v8
; TONGA-NEXT: v_cvt_f64_f32_e32 v[6:7], v9
; TONGA-NEXT: v_cvt_f32_f16_sdwa v9, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1
; TONGA-NEXT: v_cvt_f32_f16_sdwa v8, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1
; TONGA-NEXT: v_cvt_f32_f16_e32 v2, v1
; TONGA-NEXT: flat_store_dwordx4 v[11:12], v[4:7]
; TONGA-NEXT: v_cvt_f32_f16_sdwa v11, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1
; TONGA-NEXT: v_cvt_f32_f16_e32 v7, v3
; TONGA-NEXT: v_cvt_f64_f32_e32 v[3:4], v9
; TONGA-NEXT: v_cvt_f32_f16_e32 v9, v0
; TONGA-NEXT: v_cvt_f64_f32_e32 v[1:2], v2
; TONGA-NEXT: v_cvt_f64_f32_e32 v[5:6], v10
; TONGA-NEXT: v_cvt_f64_f32_e32 v[11:12], v11
; TONGA-NEXT: v_cvt_f64_f32_e32 v[9:10], v9
; TONGA-NEXT: flat_store_dwordx4 v[13:14], v[1:4]
; TONGA-NEXT: s_add_u32 s0, s0, 0x60
; TONGA-NEXT: v_cvt_f64_f32_e32 v[0:1], v7
; TONGA-NEXT: v_cvt_f64_f32_e32 v[2:3], v17
; TONGA-NEXT: v_cvt_f64_f32_e32 v[7:8], v8
; TONGA-NEXT: s_addc_u32 s1, s1, 0
; TONGA-NEXT: v_mov_b32_e32 v20, s3
; TONGA-NEXT: v_mov_b32_e32 v19, s2
; TONGA-NEXT: v_mov_b32_e32 v14, s1
; TONGA-NEXT: v_mov_b32_e32 v13, s0
; TONGA-NEXT: flat_store_dwordx4 v[15:16], v[9:12]
; TONGA-NEXT: flat_store_dwordx4 v[19:20], v[0:3]
; TONGA-NEXT: flat_store_dwordx4 v[13:14], v[5:8]
; TONGA-NEXT: s_endpgm
;
; TONGA-GCNTRACKERS-LABEL: global_extload_v16f16_to_v16f64:
; TONGA-GCNTRACKERS: ; %bb.0:
; TONGA-GCNTRACKERS-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24
; TONGA-GCNTRACKERS-NEXT: s_waitcnt lgkmcnt(0)
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v0, s2
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v1, s3
; TONGA-GCNTRACKERS-NEXT: flat_load_dwordx4 v[4:7], v[0:1]
; TONGA-GCNTRACKERS-NEXT: s_add_u32 s2, s2, 16
; TONGA-GCNTRACKERS-NEXT: s_addc_u32 s3, s3, 0
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v0, s2
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v1, s3
; TONGA-GCNTRACKERS-NEXT: flat_load_dwordx4 v[0:3], v[0:1]
; TONGA-GCNTRACKERS-NEXT: s_add_u32 s2, s0, 48
; TONGA-GCNTRACKERS-NEXT: s_addc_u32 s3, s1, 0
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v14, s3
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v13, s2
; TONGA-GCNTRACKERS-NEXT: s_add_u32 s2, s0, 32
; TONGA-GCNTRACKERS-NEXT: s_addc_u32 s3, s1, 0
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v16, s3
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v15, s2
; TONGA-GCNTRACKERS-NEXT: s_add_u32 s2, s0, 16
; TONGA-GCNTRACKERS-NEXT: s_addc_u32 s3, s1, 0
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v18, s3
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v17, s2
; TONGA-GCNTRACKERS-NEXT: s_add_u32 s2, s0, 0x50
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v12, s1
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v11, s0
; TONGA-GCNTRACKERS-NEXT: s_addc_u32 s3, s1, 0
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v20, s3
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v19, s2
; TONGA-GCNTRACKERS-NEXT: s_add_u32 s2, s0, 64
; TONGA-GCNTRACKERS-NEXT: s_addc_u32 s3, s1, 0
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v22, s3
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v21, s2
; TONGA-GCNTRACKERS-NEXT: s_add_u32 s2, s0, 0x70
; TONGA-GCNTRACKERS-NEXT: s_addc_u32 s3, s1, 0
; TONGA-GCNTRACKERS-NEXT: s_add_u32 s0, s0, 0x60
; TONGA-GCNTRACKERS-NEXT: s_addc_u32 s1, s1, 0
; TONGA-GCNTRACKERS-NEXT: s_waitcnt vmcnt(1)
; TONGA-GCNTRACKERS-NEXT: v_cvt_f32_f16_e32 v8, v7
; TONGA-GCNTRACKERS-NEXT: v_cvt_f32_f16_sdwa v9, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1
; TONGA-GCNTRACKERS-NEXT: v_cvt_f64_f32_e32 v[7:8], v8
; TONGA-GCNTRACKERS-NEXT: v_cvt_f64_f32_e32 v[9:10], v9
; TONGA-GCNTRACKERS-NEXT: flat_store_dwordx4 v[13:14], v[7:10]
; TONGA-GCNTRACKERS-NEXT: s_waitcnt vmcnt(1)
; TONGA-GCNTRACKERS-NEXT: v_cvt_f32_f16_e32 v13, v1
; TONGA-GCNTRACKERS-NEXT: v_cvt_f32_f16_e32 v7, v6
; TONGA-GCNTRACKERS-NEXT: v_cvt_f32_f16_sdwa v8, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1
; TONGA-GCNTRACKERS-NEXT: v_cvt_f32_f16_e32 v10, v0
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v14, s2
; TONGA-GCNTRACKERS-NEXT: v_cvt_f64_f32_e32 v[6:7], v7
; TONGA-GCNTRACKERS-NEXT: v_cvt_f64_f32_e32 v[8:9], v8
; TONGA-GCNTRACKERS-NEXT: flat_store_dwordx4 v[15:16], v[6:9]
; TONGA-GCNTRACKERS-NEXT: v_cvt_f32_f16_sdwa v16, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1
; TONGA-GCNTRACKERS-NEXT: v_cvt_f32_f16_e32 v6, v5
; TONGA-GCNTRACKERS-NEXT: v_cvt_f32_f16_sdwa v7, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1
; TONGA-GCNTRACKERS-NEXT: v_cvt_f32_f16_e32 v9, v4
; TONGA-GCNTRACKERS-NEXT: v_cvt_f32_f16_sdwa v4, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1
; TONGA-GCNTRACKERS-NEXT: v_cvt_f64_f32_e32 v[5:6], v6
; TONGA-GCNTRACKERS-NEXT: v_cvt_f64_f32_e32 v[7:8], v7
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v15, s3
; TONGA-GCNTRACKERS-NEXT: flat_store_dwordx4 v[17:18], v[5:8]
; TONGA-GCNTRACKERS-NEXT: v_cvt_f32_f16_sdwa v17, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1
; TONGA-GCNTRACKERS-NEXT: v_cvt_f32_f16_e32 v6, v2
; TONGA-GCNTRACKERS-NEXT: v_cvt_f32_f16_e32 v7, v3
; TONGA-GCNTRACKERS-NEXT: v_cvt_f32_f16_sdwa v8, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1
; TONGA-GCNTRACKERS-NEXT: v_cvt_f64_f32_e32 v[2:3], v9
; TONGA-GCNTRACKERS-NEXT: v_cvt_f64_f32_e32 v[4:5], v4
; TONGA-GCNTRACKERS-NEXT: v_cvt_f32_f16_sdwa v9, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1
; TONGA-GCNTRACKERS-NEXT: v_cvt_f64_f32_e32 v[0:1], v6
; TONGA-GCNTRACKERS-NEXT: v_cvt_f64_f32_e32 v[6:7], v7
; TONGA-GCNTRACKERS-NEXT: flat_store_dwordx4 v[11:12], v[2:5]
; TONGA-GCNTRACKERS-NEXT: v_cvt_f64_f32_e32 v[10:11], v10
; TONGA-GCNTRACKERS-NEXT: v_cvt_f64_f32_e32 v[2:3], v13
; TONGA-GCNTRACKERS-NEXT: v_cvt_f64_f32_e32 v[4:5], v17
; TONGA-GCNTRACKERS-NEXT: v_cvt_f64_f32_e32 v[12:13], v9
; TONGA-GCNTRACKERS-NEXT: v_cvt_f64_f32_e32 v[8:9], v8
; TONGA-GCNTRACKERS-NEXT: flat_store_dwordx4 v[19:20], v[2:5]
; TONGA-GCNTRACKERS-NEXT: s_nop 0
; TONGA-GCNTRACKERS-NEXT: v_cvt_f64_f32_e32 v[2:3], v16
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v5, s1
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v4, s0
; TONGA-GCNTRACKERS-NEXT: flat_store_dwordx4 v[21:22], v[10:13]
; TONGA-GCNTRACKERS-NEXT: flat_store_dwordx4 v[14:15], v[6:9]
; TONGA-GCNTRACKERS-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; TONGA-GCNTRACKERS-NEXT: s_endpgm
;
; GFX908-LABEL: global_extload_v16f16_to_v16f64:
; GFX908: ; %bb.0:
; GFX908-NEXT: s_load_dwordx4 s[8:11], s[4:5], 0x24
; GFX908-NEXT: v_mov_b32_e32 v16, 0
; GFX908-NEXT: s_waitcnt lgkmcnt(0)
; GFX908-NEXT: s_load_dwordx8 s[0:7], s[10:11], 0x0
; GFX908-NEXT: s_waitcnt lgkmcnt(0)
; GFX908-NEXT: v_cvt_f32_f16_e32 v6, s5
; GFX908-NEXT: s_lshr_b32 s5, s5, 16
; GFX908-NEXT: v_cvt_f32_f16_e32 v2, s1
; GFX908-NEXT: v_cvt_f32_f16_e32 v8, s5
; GFX908-NEXT: s_lshr_b32 s1, s1, 16
; GFX908-NEXT: v_cvt_f32_f16_e32 v18, s1
; GFX908-NEXT: s_lshr_b32 s1, s4, 16
; GFX908-NEXT: v_cvt_f64_f32_e32 v[4:5], v2
; GFX908-NEXT: v_cvt_f32_f16_e32 v2, s4
; GFX908-NEXT: v_cvt_f64_f32_e32 v[6:7], v6
; GFX908-NEXT: v_cvt_f64_f32_e32 v[8:9], v8
; GFX908-NEXT: v_cvt_f32_f16_e32 v10, s1
; GFX908-NEXT: s_lshr_b32 s1, s7, 16
; GFX908-NEXT: v_cvt_f32_f16_e32 v11, s7
; GFX908-NEXT: global_store_dwordx4 v16, v[6:9], s[8:9] offset:80
; GFX908-NEXT: v_cvt_f32_f16_e32 v0, s0
; GFX908-NEXT: v_cvt_f64_f32_e32 v[6:7], v2
; GFX908-NEXT: v_cvt_f64_f32_e32 v[8:9], v10
; GFX908-NEXT: v_cvt_f32_f16_e32 v2, s1
; GFX908-NEXT: s_lshr_b32 s1, s6, 16
; GFX908-NEXT: v_cvt_f32_f16_e32 v10, s6
; GFX908-NEXT: global_store_dwordx4 v16, v[6:9], s[8:9] offset:64
; GFX908-NEXT: s_lshr_b32 s0, s0, 16
; GFX908-NEXT: v_cvt_f64_f32_e32 v[6:7], v11
; GFX908-NEXT: v_cvt_f64_f32_e32 v[8:9], v2
; GFX908-NEXT: v_cvt_f32_f16_e32 v2, s1
; GFX908-NEXT: s_lshr_b32 s1, s3, 16
; GFX908-NEXT: v_cvt_f32_f16_e32 v17, s0
; GFX908-NEXT: s_lshr_b32 s0, s2, 16
; GFX908-NEXT: v_cvt_f32_f16_e32 v11, s3
; GFX908-NEXT: global_store_dwordx4 v16, v[6:9], s[8:9] offset:112
; GFX908-NEXT: v_cvt_f32_f16_e32 v3, s2
; GFX908-NEXT: v_cvt_f64_f32_e32 v[6:7], v10
; GFX908-NEXT: v_cvt_f64_f32_e32 v[8:9], v2
; GFX908-NEXT: v_cvt_f32_f16_e32 v2, s1
; GFX908-NEXT: v_cvt_f32_f16_e32 v14, s0
; GFX908-NEXT: v_cvt_f64_f32_e32 v[12:13], v3
; GFX908-NEXT: global_store_dwordx4 v16, v[6:9], s[8:9] offset:96
; GFX908-NEXT: v_cvt_f64_f32_e32 v[0:1], v0
; GFX908-NEXT: v_cvt_f64_f32_e32 v[8:9], v11
; GFX908-NEXT: v_cvt_f64_f32_e32 v[10:11], v2
; GFX908-NEXT: v_cvt_f64_f32_e32 v[14:15], v14
; GFX908-NEXT: v_cvt_f64_f32_e32 v[6:7], v18
; GFX908-NEXT: v_cvt_f64_f32_e32 v[2:3], v17
; GFX908-NEXT: global_store_dwordx4 v16, v[8:11], s[8:9] offset:48
; GFX908-NEXT: global_store_dwordx4 v16, v[12:15], s[8:9] offset:32
; GFX908-NEXT: global_store_dwordx4 v16, v[4:7], s[8:9] offset:16
; GFX908-NEXT: global_store_dwordx4 v16, v[0:3], s[8:9]
; GFX908-NEXT: s_endpgm
;
; GFX908-GCNTRACKERS-LABEL: global_extload_v16f16_to_v16f64:
; GFX908-GCNTRACKERS: ; %bb.0:
; GFX908-GCNTRACKERS-NEXT: s_load_dwordx4 s[8:11], s[4:5], 0x24
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v20, 0
; GFX908-GCNTRACKERS-NEXT: s_waitcnt lgkmcnt(0)
; GFX908-GCNTRACKERS-NEXT: s_load_dwordx8 s[0:7], s[10:11], 0x0
; GFX908-GCNTRACKERS-NEXT: s_waitcnt lgkmcnt(0)
; GFX908-GCNTRACKERS-NEXT: v_cvt_f32_f16_e32 v2, s1
; GFX908-GCNTRACKERS-NEXT: s_lshr_b32 s1, s1, 16
; GFX908-GCNTRACKERS-NEXT: v_cvt_f32_f16_e32 v6, s1
; GFX908-GCNTRACKERS-NEXT: s_lshr_b32 s1, s5, 16
; GFX908-GCNTRACKERS-NEXT: v_cvt_f32_f16_e32 v10, s5
; GFX908-GCNTRACKERS-NEXT: v_cvt_f32_f16_e32 v12, s1
; GFX908-GCNTRACKERS-NEXT: s_lshr_b32 s1, s4, 16
; GFX908-GCNTRACKERS-NEXT: v_cvt_f32_f16_e32 v15, s4
; GFX908-GCNTRACKERS-NEXT: v_cvt_f64_f32_e32 v[10:11], v10
; GFX908-GCNTRACKERS-NEXT: v_cvt_f64_f32_e32 v[12:13], v12
; GFX908-GCNTRACKERS-NEXT: v_cvt_f32_f16_e32 v16, s1
; GFX908-GCNTRACKERS-NEXT: s_lshr_b32 s1, s7, 16
; GFX908-GCNTRACKERS-NEXT: v_cvt_f32_f16_e32 v3, s2
; GFX908-GCNTRACKERS-NEXT: v_cvt_f32_f16_e32 v14, s7
; GFX908-GCNTRACKERS-NEXT: global_store_dwordx4 v20, v[10:13], s[8:9] offset:80
; GFX908-GCNTRACKERS-NEXT: v_cvt_f32_f16_e32 v0, s0
; GFX908-GCNTRACKERS-NEXT: v_cvt_f64_f32_e32 v[10:11], v15
; GFX908-GCNTRACKERS-NEXT: v_cvt_f64_f32_e32 v[12:13], v16
; GFX908-GCNTRACKERS-NEXT: v_cvt_f32_f16_e32 v16, s1
; GFX908-GCNTRACKERS-NEXT: s_lshr_b32 s0, s0, 16
; GFX908-GCNTRACKERS-NEXT: v_cvt_f32_f16_e32 v21, s0
; GFX908-GCNTRACKERS-NEXT: s_lshr_b32 s0, s3, 16
; GFX908-GCNTRACKERS-NEXT: s_lshr_b32 s1, s6, 16
; GFX908-GCNTRACKERS-NEXT: s_lshr_b32 s2, s2, 16
; GFX908-GCNTRACKERS-NEXT: v_cvt_f64_f32_e32 v[4:5], v2
; GFX908-GCNTRACKERS-NEXT: v_cvt_f64_f32_e32 v[8:9], v3
; GFX908-GCNTRACKERS-NEXT: v_cvt_f32_f16_e32 v2, s3
; GFX908-GCNTRACKERS-NEXT: v_cvt_f32_f16_e32 v3, s6
; GFX908-GCNTRACKERS-NEXT: v_cvt_f64_f32_e32 v[14:15], v14
; GFX908-GCNTRACKERS-NEXT: global_store_dwordx4 v20, v[10:13], s[8:9] offset:64
; GFX908-GCNTRACKERS-NEXT: v_cvt_f64_f32_e32 v[16:17], v16
; GFX908-GCNTRACKERS-NEXT: v_cvt_f32_f16_e32 v10, s1
; GFX908-GCNTRACKERS-NEXT: v_cvt_f32_f16_e32 v11, s0
; GFX908-GCNTRACKERS-NEXT: v_cvt_f32_f16_e32 v22, s2
; GFX908-GCNTRACKERS-NEXT: v_cvt_f64_f32_e32 v[12:13], v3
; GFX908-GCNTRACKERS-NEXT: global_store_dwordx4 v20, v[14:17], s[8:9] offset:112
; GFX908-GCNTRACKERS-NEXT: v_cvt_f64_f32_e32 v[18:19], v11
; GFX908-GCNTRACKERS-NEXT: v_cvt_f64_f32_e32 v[14:15], v10
; GFX908-GCNTRACKERS-NEXT: v_cvt_f64_f32_e32 v[16:17], v2
; GFX908-GCNTRACKERS-NEXT: v_cvt_f64_f32_e32 v[10:11], v22
; GFX908-GCNTRACKERS-NEXT: v_cvt_f64_f32_e32 v[6:7], v6
; GFX908-GCNTRACKERS-NEXT: v_cvt_f64_f32_e32 v[0:1], v0
; GFX908-GCNTRACKERS-NEXT: v_cvt_f64_f32_e32 v[2:3], v21
; GFX908-GCNTRACKERS-NEXT: global_store_dwordx4 v20, v[12:15], s[8:9] offset:96
; GFX908-GCNTRACKERS-NEXT: global_store_dwordx4 v20, v[16:19], s[8:9] offset:48
; GFX908-GCNTRACKERS-NEXT: global_store_dwordx4 v20, v[8:11], s[8:9] offset:32
; GFX908-GCNTRACKERS-NEXT: global_store_dwordx4 v20, v[4:7], s[8:9] offset:16
; GFX908-GCNTRACKERS-NEXT: global_store_dwordx4 v20, v[0:3], s[8:9]
; GFX908-GCNTRACKERS-NEXT: s_endpgm
;
; GENERIC-LABEL: global_extload_v16f16_to_v16f64:
; GENERIC: ; %bb.0:
; GENERIC-NEXT: s_load_dwordx4 s[8:11], s[4:5], 0x9
; GENERIC-NEXT: s_waitcnt lgkmcnt(0)
; GENERIC-NEXT: s_load_dwordx8 s[0:7], s[10:11], 0x0
; GENERIC-NEXT: s_mov_b32 s11, 0xf000
; GENERIC-NEXT: s_mov_b32 s10, -1
; GENERIC-NEXT: s_waitcnt lgkmcnt(0)
; GENERIC-NEXT: s_lshr_b32 s12, s7, 16
; GENERIC-NEXT: v_cvt_f32_f16_e32 v8, s7
; GENERIC-NEXT: v_cvt_f32_f16_e32 v10, s12
; GENERIC-NEXT: s_lshr_b32 s13, s6, 16
; GENERIC-NEXT: v_cvt_f32_f16_e32 v13, s6
; GENERIC-NEXT: v_cvt_f64_f32_e32 v[8:9], v8
; GENERIC-NEXT: v_cvt_f64_f32_e32 v[10:11], v10
; GENERIC-NEXT: v_cvt_f32_f16_e32 v14, s13
; GENERIC-NEXT: s_lshr_b32 s14, s5, 16
; GENERIC-NEXT: v_cvt_f32_f16_e32 v12, s5
; GENERIC-NEXT: v_cvt_f32_f16_e32 v15, s14
; GENERIC-NEXT: buffer_store_dwordx4 v[8:11], off, s[8:11], 0 offset:112
; GENERIC-NEXT: s_waitcnt expcnt(0)
; GENERIC-NEXT: v_cvt_f64_f32_e32 v[8:9], v13
; GENERIC-NEXT: v_cvt_f64_f32_e32 v[10:11], v14
; GENERIC-NEXT: s_lshr_b32 s15, s4, 16
; GENERIC-NEXT: s_lshr_b32 s16, s3, 16
; GENERIC-NEXT: v_cvt_f64_f32_e32 v[12:13], v12
; GENERIC-NEXT: v_cvt_f64_f32_e32 v[14:15], v15
; GENERIC-NEXT: buffer_store_dwordx4 v[8:11], off, s[8:11], 0 offset:96
; GENERIC-NEXT: s_waitcnt expcnt(0)
; GENERIC-NEXT: v_cvt_f32_f16_e32 v10, s4
; GENERIC-NEXT: v_cvt_f32_f16_e32 v11, s15
; GENERIC-NEXT: s_lshr_b32 s17, s0, 16
; GENERIC-NEXT: v_cvt_f32_f16_e32 v0, s0
; GENERIC-NEXT: s_lshr_b32 s0, s2, 16
; GENERIC-NEXT: v_cvt_f32_f16_e32 v17, s3
; GENERIC-NEXT: v_cvt_f32_f16_e32 v19, s16
; GENERIC-NEXT: v_cvt_f32_f16_e32 v2, s17
; GENERIC-NEXT: s_lshr_b32 s17, s1, 16
; GENERIC-NEXT: v_cvt_f32_f16_e32 v4, s2
; GENERIC-NEXT: v_cvt_f32_f16_e32 v6, s0
; GENERIC-NEXT: v_cvt_f32_f16_e32 v16, s1
; GENERIC-NEXT: v_cvt_f32_f16_e32 v18, s17
; GENERIC-NEXT: buffer_store_dwordx4 v[12:15], off, s[8:11], 0 offset:80
; GENERIC-NEXT: s_waitcnt expcnt(0)
; GENERIC-NEXT: v_cvt_f64_f32_e32 v[12:13], v10
; GENERIC-NEXT: v_cvt_f64_f32_e32 v[14:15], v11
; GENERIC-NEXT: v_cvt_f64_f32_e32 v[8:9], v17
; GENERIC-NEXT: v_cvt_f64_f32_e32 v[10:11], v19
; GENERIC-NEXT: v_cvt_f64_f32_e32 v[4:5], v4
; GENERIC-NEXT: v_cvt_f64_f32_e32 v[6:7], v6
; GENERIC-NEXT: v_cvt_f64_f32_e32 v[16:17], v16
; GENERIC-NEXT: v_cvt_f64_f32_e32 v[18:19], v18
; GENERIC-NEXT: v_cvt_f64_f32_e32 v[0:1], v0
; GENERIC-NEXT: v_cvt_f64_f32_e32 v[2:3], v2
; GENERIC-NEXT: buffer_store_dwordx4 v[12:15], off, s[8:11], 0 offset:64
; GENERIC-NEXT: buffer_store_dwordx4 v[8:11], off, s[8:11], 0 offset:48
; GENERIC-NEXT: buffer_store_dwordx4 v[4:7], off, s[8:11], 0 offset:32
; GENERIC-NEXT: buffer_store_dwordx4 v[16:19], off, s[8:11], 0 offset:16
; GENERIC-NEXT: buffer_store_dwordx4 v[0:3], off, s[8:11], 0
; GENERIC-NEXT: s_endpgm
;
; GENERIC-GCNTRACKERS-LABEL: global_extload_v16f16_to_v16f64:
; GENERIC-GCNTRACKERS: ; %bb.0:
; GENERIC-GCNTRACKERS-NEXT: s_load_dwordx4 s[8:11], s[4:5], 0x9
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt lgkmcnt(0)
; GENERIC-GCNTRACKERS-NEXT: s_load_dwordx8 s[0:7], s[10:11], 0x0
; GENERIC-GCNTRACKERS-NEXT: s_mov_b32 s11, 0xf000
; GENERIC-GCNTRACKERS-NEXT: s_mov_b32 s10, -1
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt lgkmcnt(0)
; GENERIC-GCNTRACKERS-NEXT: v_cvt_f32_f16_e32 v6, s2
; GENERIC-GCNTRACKERS-NEXT: s_lshr_b32 s12, s7, 16
; GENERIC-GCNTRACKERS-NEXT: v_cvt_f32_f16_e32 v11, s7
; GENERIC-GCNTRACKERS-NEXT: v_cvt_f32_f16_e32 v16, s12
; GENERIC-GCNTRACKERS-NEXT: v_cvt_f64_f32_e32 v[8:9], v6
; GENERIC-GCNTRACKERS-NEXT: v_cvt_f32_f16_e32 v6, s3
; GENERIC-GCNTRACKERS-NEXT: s_lshr_b32 s13, s6, 16
; GENERIC-GCNTRACKERS-NEXT: s_lshr_b32 s14, s5, 16
; GENERIC-GCNTRACKERS-NEXT: v_cvt_f64_f32_e32 v[14:15], v11
; GENERIC-GCNTRACKERS-NEXT: v_cvt_f64_f32_e32 v[12:13], v6
; GENERIC-GCNTRACKERS-NEXT: v_cvt_f32_f16_e32 v6, s6
; GENERIC-GCNTRACKERS-NEXT: v_cvt_f64_f32_e32 v[16:17], v16
; GENERIC-GCNTRACKERS-NEXT: v_cvt_f32_f16_e32 v11, s13
; GENERIC-GCNTRACKERS-NEXT: v_cvt_f32_f16_e32 v22, s5
; GENERIC-GCNTRACKERS-NEXT: v_cvt_f64_f32_e32 v[18:19], v6
; GENERIC-GCNTRACKERS-NEXT: v_cvt_f32_f16_e32 v6, s14
; GENERIC-GCNTRACKERS-NEXT: s_lshr_b32 s15, s4, 16
; GENERIC-GCNTRACKERS-NEXT: s_lshr_b32 s16, s3, 16
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dwordx4 v[14:17], off, s[8:11], 0 offset:112
; GENERIC-GCNTRACKERS-NEXT: v_cvt_f64_f32_e32 v[20:21], v11
; GENERIC-GCNTRACKERS-NEXT: v_cvt_f32_f16_e32 v11, s4
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt expcnt(0)
; GENERIC-GCNTRACKERS-NEXT: v_cvt_f64_f32_e32 v[16:17], v6
; GENERIC-GCNTRACKERS-NEXT: v_cvt_f32_f16_e32 v6, s15
; GENERIC-GCNTRACKERS-NEXT: s_lshr_b32 s17, s2, 16
; GENERIC-GCNTRACKERS-NEXT: v_cvt_f64_f32_e32 v[14:15], v22
; GENERIC-GCNTRACKERS-NEXT: v_cvt_f32_f16_e32 v22, s16
; GENERIC-GCNTRACKERS-NEXT: s_lshr_b32 s18, s1, 16
; GENERIC-GCNTRACKERS-NEXT: v_cvt_f32_f16_e32 v10, s17
; GENERIC-GCNTRACKERS-NEXT: v_cvt_f32_f16_e32 v0, s0
; GENERIC-GCNTRACKERS-NEXT: s_lshr_b32 s0, s0, 16
; GENERIC-GCNTRACKERS-NEXT: v_cvt_f32_f16_e32 v4, s1
; GENERIC-GCNTRACKERS-NEXT: v_cvt_f32_f16_e32 v7, s18
; GENERIC-GCNTRACKERS-NEXT: v_cvt_f32_f16_e32 v2, s0
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dwordx4 v[18:21], off, s[8:11], 0 offset:96
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt expcnt(0)
; GENERIC-GCNTRACKERS-NEXT: v_cvt_f64_f32_e32 v[18:19], v11
; GENERIC-GCNTRACKERS-NEXT: v_cvt_f64_f32_e32 v[20:21], v6
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dwordx4 v[14:17], off, s[8:11], 0 offset:80
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt expcnt(0)
; GENERIC-GCNTRACKERS-NEXT: v_cvt_f64_f32_e32 v[14:15], v22
; GENERIC-GCNTRACKERS-NEXT: v_cvt_f64_f32_e32 v[10:11], v10
; GENERIC-GCNTRACKERS-NEXT: v_cvt_f64_f32_e32 v[4:5], v4
; GENERIC-GCNTRACKERS-NEXT: v_cvt_f64_f32_e32 v[6:7], v7
; GENERIC-GCNTRACKERS-NEXT: v_cvt_f64_f32_e32 v[0:1], v0
; GENERIC-GCNTRACKERS-NEXT: v_cvt_f64_f32_e32 v[2:3], v2
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dwordx4 v[18:21], off, s[8:11], 0 offset:64
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dwordx4 v[12:15], off, s[8:11], 0 offset:48
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dwordx4 v[8:11], off, s[8:11], 0 offset:32
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dwordx4 v[4:7], off, s[8:11], 0 offset:16
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dwordx4 v[0:3], off, s[8:11], 0
; GENERIC-GCNTRACKERS-NEXT: s_endpgm
%val = load <16 x half>, ptr addrspace(1) %in
%cvt = fpext <16 x half> %val to <16 x double>
store <16 x double> %cvt, ptr addrspace(1) %out
ret void
}
; CHECK-LABEL: {{^}}constant_zextload_v64i16_to_v64i32:
; GENERIC: NumSgprs: 69
; GENERIC-GCNTRACKERS: NumSgprs: 45
; GENERIC: NumVgprs: 20
; GENERIC-GCNTRACKERS: NumVgprs: 20
; GENERIC: Occupancy: 7
; GENERIC-GCNTRACKERS: Occupancy: 10
define amdgpu_kernel void @constant_zextload_v64i16_to_v64i32(ptr addrspace(1) %out, ptr addrspace(4) %in) {
; GFX11-PAL-LABEL: constant_zextload_v64i16_to_v64i32:
; GFX11-PAL: ; %bb.0:
; GFX11-PAL-NEXT: s_load_b128 s[36:39], s[4:5], 0x0
; GFX11-PAL-NEXT: s_waitcnt lgkmcnt(0)
; GFX11-PAL-NEXT: s_clause 0x1
; GFX11-PAL-NEXT: s_load_b512 s[16:31], s[38:39], 0x0
; GFX11-PAL-NEXT: s_load_b512 s[0:15], s[38:39], 0x40
; GFX11-PAL-NEXT: s_waitcnt lgkmcnt(0)
; GFX11-PAL-NEXT: s_lshr_b32 s49, s31, 16
; GFX11-PAL-NEXT: s_lshr_b32 s65, s15, 16
; GFX11-PAL-NEXT: s_lshr_b32 s66, s14, 16
; GFX11-PAL-NEXT: s_and_b32 s14, s14, 0xffff
; GFX11-PAL-NEXT: s_and_b32 s15, s15, 0xffff
; GFX11-PAL-NEXT: s_lshr_b32 s63, s13, 16
; GFX11-PAL-NEXT: s_lshr_b32 s64, s12, 16
; GFX11-PAL-NEXT: s_and_b32 s13, s13, 0xffff
; GFX11-PAL-NEXT: s_and_b32 s12, s12, 0xffff
; GFX11-PAL-NEXT: v_dual_mov_b32 v24, 0 :: v_dual_mov_b32 v1, s66
; GFX11-PAL-NEXT: s_lshr_b32 s61, s11, 16
; GFX11-PAL-NEXT: s_lshr_b32 s62, s10, 16
; GFX11-PAL-NEXT: s_and_b32 s11, s11, 0xffff
; GFX11-PAL-NEXT: s_and_b32 s10, s10, 0xffff
; GFX11-PAL-NEXT: v_dual_mov_b32 v0, s14 :: v_dual_mov_b32 v3, s65
; GFX11-PAL-NEXT: v_dual_mov_b32 v2, s15 :: v_dual_mov_b32 v5, s64
; GFX11-PAL-NEXT: s_lshr_b32 s59, s9, 16
; GFX11-PAL-NEXT: s_lshr_b32 s60, s8, 16
; GFX11-PAL-NEXT: s_and_b32 s9, s9, 0xffff
; GFX11-PAL-NEXT: s_and_b32 s8, s8, 0xffff
; GFX11-PAL-NEXT: v_dual_mov_b32 v4, s12 :: v_dual_mov_b32 v7, s63
; GFX11-PAL-NEXT: v_dual_mov_b32 v6, s13 :: v_dual_mov_b32 v9, s62
; GFX11-PAL-NEXT: v_dual_mov_b32 v8, s10 :: v_dual_mov_b32 v11, s61
; GFX11-PAL-NEXT: v_dual_mov_b32 v10, s11 :: v_dual_mov_b32 v13, s60
; GFX11-PAL-NEXT: s_lshr_b32 s57, s7, 16
; GFX11-PAL-NEXT: s_lshr_b32 s58, s6, 16
; GFX11-PAL-NEXT: s_and_b32 s7, s7, 0xffff
; GFX11-PAL-NEXT: v_dual_mov_b32 v12, s8 :: v_dual_mov_b32 v15, s59
; GFX11-PAL-NEXT: v_mov_b32_e32 v14, s9
; GFX11-PAL-NEXT: s_and_b32 s6, s6, 0xffff
; GFX11-PAL-NEXT: s_lshr_b32 s55, s5, 16
; GFX11-PAL-NEXT: s_lshr_b32 s56, s4, 16
; GFX11-PAL-NEXT: s_and_b32 s5, s5, 0xffff
; GFX11-PAL-NEXT: s_and_b32 s4, s4, 0xffff
; GFX11-PAL-NEXT: s_lshr_b32 s53, s3, 16
; GFX11-PAL-NEXT: s_lshr_b32 s54, s2, 16
; GFX11-PAL-NEXT: s_and_b32 s3, s3, 0xffff
; GFX11-PAL-NEXT: s_and_b32 s2, s2, 0xffff
; GFX11-PAL-NEXT: s_clause 0x3
; GFX11-PAL-NEXT: global_store_b128 v24, v[0:3], s[36:37] offset:240
; GFX11-PAL-NEXT: global_store_b128 v24, v[4:7], s[36:37] offset:224
; GFX11-PAL-NEXT: global_store_b128 v24, v[8:11], s[36:37] offset:208
; GFX11-PAL-NEXT: global_store_b128 v24, v[12:15], s[36:37] offset:192
; GFX11-PAL-NEXT: v_dual_mov_b32 v0, s6 :: v_dual_mov_b32 v3, s57
; GFX11-PAL-NEXT: v_dual_mov_b32 v1, s58 :: v_dual_mov_b32 v2, s7
; GFX11-PAL-NEXT: v_mov_b32_e32 v5, s56
; GFX11-PAL-NEXT: s_lshr_b32 s51, s1, 16
; GFX11-PAL-NEXT: s_lshr_b32 s52, s0, 16
; GFX11-PAL-NEXT: s_and_b32 s1, s1, 0xffff
; GFX11-PAL-NEXT: s_and_b32 s0, s0, 0xffff
; GFX11-PAL-NEXT: v_dual_mov_b32 v4, s4 :: v_dual_mov_b32 v7, s55
; GFX11-PAL-NEXT: v_dual_mov_b32 v6, s5 :: v_dual_mov_b32 v9, s54
; GFX11-PAL-NEXT: s_lshr_b32 s50, s30, 16
; GFX11-PAL-NEXT: s_and_b32 s31, s31, 0xffff
; GFX11-PAL-NEXT: s_and_b32 s30, s30, 0xffff
; GFX11-PAL-NEXT: v_dual_mov_b32 v8, s2 :: v_dual_mov_b32 v11, s53
; GFX11-PAL-NEXT: v_dual_mov_b32 v10, s3 :: v_dual_mov_b32 v13, s52
; GFX11-PAL-NEXT: s_lshr_b32 s45, s27, 16
; GFX11-PAL-NEXT: s_lshr_b32 s46, s26, 16
; GFX11-PAL-NEXT: s_lshr_b32 s47, s29, 16
; GFX11-PAL-NEXT: s_lshr_b32 s48, s28, 16
; GFX11-PAL-NEXT: s_and_b32 s27, s27, 0xffff
; GFX11-PAL-NEXT: s_and_b32 s26, s26, 0xffff
; GFX11-PAL-NEXT: s_and_b32 s29, s29, 0xffff
; GFX11-PAL-NEXT: s_and_b32 s28, s28, 0xffff
; GFX11-PAL-NEXT: v_dual_mov_b32 v12, s0 :: v_dual_mov_b32 v15, s51
; GFX11-PAL-NEXT: v_dual_mov_b32 v14, s1 :: v_dual_mov_b32 v17, s50
; GFX11-PAL-NEXT: s_lshr_b32 s43, s25, 16
; GFX11-PAL-NEXT: s_lshr_b32 s44, s24, 16
; GFX11-PAL-NEXT: s_and_b32 s25, s25, 0xffff
; GFX11-PAL-NEXT: s_and_b32 s24, s24, 0xffff
; GFX11-PAL-NEXT: v_dual_mov_b32 v16, s30 :: v_dual_mov_b32 v19, s49
; GFX11-PAL-NEXT: v_dual_mov_b32 v18, s31 :: v_dual_mov_b32 v21, s48
; GFX11-PAL-NEXT: s_lshr_b32 s41, s23, 16
; GFX11-PAL-NEXT: s_lshr_b32 s42, s22, 16
; GFX11-PAL-NEXT: s_and_b32 s23, s23, 0xffff
; GFX11-PAL-NEXT: s_and_b32 s22, s22, 0xffff
; GFX11-PAL-NEXT: v_dual_mov_b32 v20, s28 :: v_dual_mov_b32 v23, s47
; GFX11-PAL-NEXT: v_mov_b32_e32 v22, s29
; GFX11-PAL-NEXT: s_clause 0x5
; GFX11-PAL-NEXT: global_store_b128 v24, v[0:3], s[36:37] offset:176
; GFX11-PAL-NEXT: global_store_b128 v24, v[4:7], s[36:37] offset:160
; GFX11-PAL-NEXT: global_store_b128 v24, v[8:11], s[36:37] offset:144
; GFX11-PAL-NEXT: global_store_b128 v24, v[12:15], s[36:37] offset:128
; GFX11-PAL-NEXT: global_store_b128 v24, v[16:19], s[36:37] offset:112
; GFX11-PAL-NEXT: global_store_b128 v24, v[20:23], s[36:37] offset:96
; GFX11-PAL-NEXT: v_dual_mov_b32 v0, s26 :: v_dual_mov_b32 v3, s45
; GFX11-PAL-NEXT: v_dual_mov_b32 v1, s46 :: v_dual_mov_b32 v2, s27
; GFX11-PAL-NEXT: v_mov_b32_e32 v5, s44
; GFX11-PAL-NEXT: s_lshr_b32 s39, s21, 16
; GFX11-PAL-NEXT: s_lshr_b32 s40, s20, 16
; GFX11-PAL-NEXT: s_and_b32 s21, s21, 0xffff
; GFX11-PAL-NEXT: s_and_b32 s20, s20, 0xffff
; GFX11-PAL-NEXT: v_dual_mov_b32 v4, s24 :: v_dual_mov_b32 v7, s43
; GFX11-PAL-NEXT: v_dual_mov_b32 v6, s25 :: v_dual_mov_b32 v9, s42
; GFX11-PAL-NEXT: s_lshr_b32 s35, s19, 16
; GFX11-PAL-NEXT: s_lshr_b32 s38, s18, 16
; GFX11-PAL-NEXT: s_and_b32 s19, s19, 0xffff
; GFX11-PAL-NEXT: s_and_b32 s18, s18, 0xffff
; GFX11-PAL-NEXT: v_dual_mov_b32 v8, s22 :: v_dual_mov_b32 v11, s41
; GFX11-PAL-NEXT: v_dual_mov_b32 v10, s23 :: v_dual_mov_b32 v13, s40
; GFX11-PAL-NEXT: s_lshr_b32 s33, s17, 16
; GFX11-PAL-NEXT: s_lshr_b32 s34, s16, 16
; GFX11-PAL-NEXT: s_and_b32 s17, s17, 0xffff
; GFX11-PAL-NEXT: s_and_b32 s16, s16, 0xffff
; GFX11-PAL-NEXT: v_dual_mov_b32 v12, s20 :: v_dual_mov_b32 v15, s39
; GFX11-PAL-NEXT: v_dual_mov_b32 v14, s21 :: v_dual_mov_b32 v17, s38
; GFX11-PAL-NEXT: v_dual_mov_b32 v16, s18 :: v_dual_mov_b32 v19, s35
; GFX11-PAL-NEXT: v_dual_mov_b32 v18, s19 :: v_dual_mov_b32 v21, s34
; GFX11-PAL-NEXT: v_dual_mov_b32 v20, s16 :: v_dual_mov_b32 v23, s33
; GFX11-PAL-NEXT: v_mov_b32_e32 v22, s17
; GFX11-PAL-NEXT: s_clause 0x5
; GFX11-PAL-NEXT: global_store_b128 v24, v[0:3], s[36:37] offset:80
; GFX11-PAL-NEXT: global_store_b128 v24, v[4:7], s[36:37] offset:64
; GFX11-PAL-NEXT: global_store_b128 v24, v[8:11], s[36:37] offset:48
; GFX11-PAL-NEXT: global_store_b128 v24, v[12:15], s[36:37] offset:32
; GFX11-PAL-NEXT: global_store_b128 v24, v[16:19], s[36:37] offset:16
; GFX11-PAL-NEXT: global_store_b128 v24, v[20:23], s[36:37]
; GFX11-PAL-NEXT: s_endpgm
;
; GFX11-PAL-GCNTRACKERS-LABEL: constant_zextload_v64i16_to_v64i32:
; GFX11-PAL-GCNTRACKERS: ; %bb.0:
; GFX11-PAL-GCNTRACKERS-NEXT: s_load_b128 s[36:39], s[4:5], 0x0
; GFX11-PAL-GCNTRACKERS-NEXT: s_waitcnt lgkmcnt(0)
; GFX11-PAL-GCNTRACKERS-NEXT: s_clause 0x1
; GFX11-PAL-GCNTRACKERS-NEXT: s_load_b512 s[16:31], s[38:39], 0x0
; GFX11-PAL-GCNTRACKERS-NEXT: s_load_b512 s[0:15], s[38:39], 0x40
; GFX11-PAL-GCNTRACKERS-NEXT: s_waitcnt lgkmcnt(0)
; GFX11-PAL-GCNTRACKERS-NEXT: s_lshr_b32 s49, s31, 16
; GFX11-PAL-GCNTRACKERS-NEXT: s_lshr_b32 s65, s15, 16
; GFX11-PAL-GCNTRACKERS-NEXT: s_lshr_b32 s66, s14, 16
; GFX11-PAL-GCNTRACKERS-NEXT: s_and_b32 s14, s14, 0xffff
; GFX11-PAL-GCNTRACKERS-NEXT: s_and_b32 s15, s15, 0xffff
; GFX11-PAL-GCNTRACKERS-NEXT: s_lshr_b32 s63, s13, 16
; GFX11-PAL-GCNTRACKERS-NEXT: s_lshr_b32 s64, s12, 16
; GFX11-PAL-GCNTRACKERS-NEXT: s_and_b32 s13, s13, 0xffff
; GFX11-PAL-GCNTRACKERS-NEXT: s_and_b32 s12, s12, 0xffff
; GFX11-PAL-GCNTRACKERS-NEXT: v_dual_mov_b32 v24, 0 :: v_dual_mov_b32 v1, s66
; GFX11-PAL-GCNTRACKERS-NEXT: s_lshr_b32 s61, s11, 16
; GFX11-PAL-GCNTRACKERS-NEXT: s_lshr_b32 s62, s10, 16
; GFX11-PAL-GCNTRACKERS-NEXT: s_and_b32 s11, s11, 0xffff
; GFX11-PAL-GCNTRACKERS-NEXT: s_and_b32 s10, s10, 0xffff
; GFX11-PAL-GCNTRACKERS-NEXT: v_dual_mov_b32 v0, s14 :: v_dual_mov_b32 v3, s65
; GFX11-PAL-GCNTRACKERS-NEXT: v_dual_mov_b32 v2, s15 :: v_dual_mov_b32 v5, s64
; GFX11-PAL-GCNTRACKERS-NEXT: s_lshr_b32 s59, s9, 16
; GFX11-PAL-GCNTRACKERS-NEXT: s_lshr_b32 s60, s8, 16
; GFX11-PAL-GCNTRACKERS-NEXT: s_and_b32 s9, s9, 0xffff
; GFX11-PAL-GCNTRACKERS-NEXT: s_and_b32 s8, s8, 0xffff
; GFX11-PAL-GCNTRACKERS-NEXT: v_dual_mov_b32 v4, s12 :: v_dual_mov_b32 v7, s63
; GFX11-PAL-GCNTRACKERS-NEXT: v_dual_mov_b32 v6, s13 :: v_dual_mov_b32 v9, s62
; GFX11-PAL-GCNTRACKERS-NEXT: v_dual_mov_b32 v8, s10 :: v_dual_mov_b32 v11, s61
; GFX11-PAL-GCNTRACKERS-NEXT: v_dual_mov_b32 v10, s11 :: v_dual_mov_b32 v13, s60
; GFX11-PAL-GCNTRACKERS-NEXT: s_lshr_b32 s57, s7, 16
; GFX11-PAL-GCNTRACKERS-NEXT: s_lshr_b32 s58, s6, 16
; GFX11-PAL-GCNTRACKERS-NEXT: s_and_b32 s7, s7, 0xffff
; GFX11-PAL-GCNTRACKERS-NEXT: v_dual_mov_b32 v12, s8 :: v_dual_mov_b32 v15, s59
; GFX11-PAL-GCNTRACKERS-NEXT: v_mov_b32_e32 v14, s9
; GFX11-PAL-GCNTRACKERS-NEXT: s_and_b32 s6, s6, 0xffff
; GFX11-PAL-GCNTRACKERS-NEXT: s_lshr_b32 s55, s5, 16
; GFX11-PAL-GCNTRACKERS-NEXT: s_lshr_b32 s56, s4, 16
; GFX11-PAL-GCNTRACKERS-NEXT: s_and_b32 s5, s5, 0xffff
; GFX11-PAL-GCNTRACKERS-NEXT: s_and_b32 s4, s4, 0xffff
; GFX11-PAL-GCNTRACKERS-NEXT: s_lshr_b32 s53, s3, 16
; GFX11-PAL-GCNTRACKERS-NEXT: s_lshr_b32 s54, s2, 16
; GFX11-PAL-GCNTRACKERS-NEXT: s_and_b32 s3, s3, 0xffff
; GFX11-PAL-GCNTRACKERS-NEXT: s_and_b32 s2, s2, 0xffff
; GFX11-PAL-GCNTRACKERS-NEXT: s_clause 0x3
; GFX11-PAL-GCNTRACKERS-NEXT: global_store_b128 v24, v[0:3], s[36:37] offset:240
; GFX11-PAL-GCNTRACKERS-NEXT: global_store_b128 v24, v[4:7], s[36:37] offset:224
; GFX11-PAL-GCNTRACKERS-NEXT: global_store_b128 v24, v[8:11], s[36:37] offset:208
; GFX11-PAL-GCNTRACKERS-NEXT: global_store_b128 v24, v[12:15], s[36:37] offset:192
; GFX11-PAL-GCNTRACKERS-NEXT: v_dual_mov_b32 v0, s6 :: v_dual_mov_b32 v3, s57
; GFX11-PAL-GCNTRACKERS-NEXT: v_dual_mov_b32 v1, s58 :: v_dual_mov_b32 v2, s7
; GFX11-PAL-GCNTRACKERS-NEXT: v_mov_b32_e32 v5, s56
; GFX11-PAL-GCNTRACKERS-NEXT: s_lshr_b32 s51, s1, 16
; GFX11-PAL-GCNTRACKERS-NEXT: s_lshr_b32 s52, s0, 16
; GFX11-PAL-GCNTRACKERS-NEXT: s_and_b32 s1, s1, 0xffff
; GFX11-PAL-GCNTRACKERS-NEXT: s_and_b32 s0, s0, 0xffff
; GFX11-PAL-GCNTRACKERS-NEXT: v_dual_mov_b32 v4, s4 :: v_dual_mov_b32 v7, s55
; GFX11-PAL-GCNTRACKERS-NEXT: v_dual_mov_b32 v6, s5 :: v_dual_mov_b32 v9, s54
; GFX11-PAL-GCNTRACKERS-NEXT: s_lshr_b32 s50, s30, 16
; GFX11-PAL-GCNTRACKERS-NEXT: s_and_b32 s31, s31, 0xffff
; GFX11-PAL-GCNTRACKERS-NEXT: s_and_b32 s30, s30, 0xffff
; GFX11-PAL-GCNTRACKERS-NEXT: v_dual_mov_b32 v8, s2 :: v_dual_mov_b32 v11, s53
; GFX11-PAL-GCNTRACKERS-NEXT: v_dual_mov_b32 v10, s3 :: v_dual_mov_b32 v13, s52
; GFX11-PAL-GCNTRACKERS-NEXT: s_lshr_b32 s45, s27, 16
; GFX11-PAL-GCNTRACKERS-NEXT: s_lshr_b32 s46, s26, 16
; GFX11-PAL-GCNTRACKERS-NEXT: s_lshr_b32 s47, s29, 16
; GFX11-PAL-GCNTRACKERS-NEXT: s_lshr_b32 s48, s28, 16
; GFX11-PAL-GCNTRACKERS-NEXT: s_and_b32 s27, s27, 0xffff
; GFX11-PAL-GCNTRACKERS-NEXT: s_and_b32 s26, s26, 0xffff
; GFX11-PAL-GCNTRACKERS-NEXT: s_and_b32 s29, s29, 0xffff
; GFX11-PAL-GCNTRACKERS-NEXT: s_and_b32 s28, s28, 0xffff
; GFX11-PAL-GCNTRACKERS-NEXT: v_dual_mov_b32 v12, s0 :: v_dual_mov_b32 v15, s51
; GFX11-PAL-GCNTRACKERS-NEXT: v_dual_mov_b32 v14, s1 :: v_dual_mov_b32 v17, s50
; GFX11-PAL-GCNTRACKERS-NEXT: s_lshr_b32 s43, s25, 16
; GFX11-PAL-GCNTRACKERS-NEXT: s_lshr_b32 s44, s24, 16
; GFX11-PAL-GCNTRACKERS-NEXT: s_and_b32 s25, s25, 0xffff
; GFX11-PAL-GCNTRACKERS-NEXT: s_and_b32 s24, s24, 0xffff
; GFX11-PAL-GCNTRACKERS-NEXT: v_dual_mov_b32 v16, s30 :: v_dual_mov_b32 v19, s49
; GFX11-PAL-GCNTRACKERS-NEXT: v_dual_mov_b32 v18, s31 :: v_dual_mov_b32 v21, s48
; GFX11-PAL-GCNTRACKERS-NEXT: s_lshr_b32 s41, s23, 16
; GFX11-PAL-GCNTRACKERS-NEXT: s_lshr_b32 s42, s22, 16
; GFX11-PAL-GCNTRACKERS-NEXT: s_and_b32 s23, s23, 0xffff
; GFX11-PAL-GCNTRACKERS-NEXT: s_and_b32 s22, s22, 0xffff
; GFX11-PAL-GCNTRACKERS-NEXT: v_dual_mov_b32 v20, s28 :: v_dual_mov_b32 v23, s47
; GFX11-PAL-GCNTRACKERS-NEXT: v_mov_b32_e32 v22, s29
; GFX11-PAL-GCNTRACKERS-NEXT: s_clause 0x5
; GFX11-PAL-GCNTRACKERS-NEXT: global_store_b128 v24, v[0:3], s[36:37] offset:176
; GFX11-PAL-GCNTRACKERS-NEXT: global_store_b128 v24, v[4:7], s[36:37] offset:160
; GFX11-PAL-GCNTRACKERS-NEXT: global_store_b128 v24, v[8:11], s[36:37] offset:144
; GFX11-PAL-GCNTRACKERS-NEXT: global_store_b128 v24, v[12:15], s[36:37] offset:128
; GFX11-PAL-GCNTRACKERS-NEXT: global_store_b128 v24, v[16:19], s[36:37] offset:112
; GFX11-PAL-GCNTRACKERS-NEXT: global_store_b128 v24, v[20:23], s[36:37] offset:96
; GFX11-PAL-GCNTRACKERS-NEXT: v_dual_mov_b32 v0, s26 :: v_dual_mov_b32 v3, s45
; GFX11-PAL-GCNTRACKERS-NEXT: v_dual_mov_b32 v1, s46 :: v_dual_mov_b32 v2, s27
; GFX11-PAL-GCNTRACKERS-NEXT: v_mov_b32_e32 v5, s44
; GFX11-PAL-GCNTRACKERS-NEXT: s_lshr_b32 s39, s21, 16
; GFX11-PAL-GCNTRACKERS-NEXT: s_lshr_b32 s40, s20, 16
; GFX11-PAL-GCNTRACKERS-NEXT: s_and_b32 s21, s21, 0xffff
; GFX11-PAL-GCNTRACKERS-NEXT: s_and_b32 s20, s20, 0xffff
; GFX11-PAL-GCNTRACKERS-NEXT: v_dual_mov_b32 v4, s24 :: v_dual_mov_b32 v7, s43
; GFX11-PAL-GCNTRACKERS-NEXT: v_dual_mov_b32 v6, s25 :: v_dual_mov_b32 v9, s42
; GFX11-PAL-GCNTRACKERS-NEXT: s_lshr_b32 s35, s19, 16
; GFX11-PAL-GCNTRACKERS-NEXT: s_lshr_b32 s38, s18, 16
; GFX11-PAL-GCNTRACKERS-NEXT: s_and_b32 s19, s19, 0xffff
; GFX11-PAL-GCNTRACKERS-NEXT: s_and_b32 s18, s18, 0xffff
; GFX11-PAL-GCNTRACKERS-NEXT: v_dual_mov_b32 v8, s22 :: v_dual_mov_b32 v11, s41
; GFX11-PAL-GCNTRACKERS-NEXT: v_dual_mov_b32 v10, s23 :: v_dual_mov_b32 v13, s40
; GFX11-PAL-GCNTRACKERS-NEXT: s_lshr_b32 s33, s17, 16
; GFX11-PAL-GCNTRACKERS-NEXT: s_lshr_b32 s34, s16, 16
; GFX11-PAL-GCNTRACKERS-NEXT: s_and_b32 s17, s17, 0xffff
; GFX11-PAL-GCNTRACKERS-NEXT: s_and_b32 s16, s16, 0xffff
; GFX11-PAL-GCNTRACKERS-NEXT: v_dual_mov_b32 v12, s20 :: v_dual_mov_b32 v15, s39
; GFX11-PAL-GCNTRACKERS-NEXT: v_dual_mov_b32 v14, s21 :: v_dual_mov_b32 v17, s38
; GFX11-PAL-GCNTRACKERS-NEXT: v_dual_mov_b32 v16, s18 :: v_dual_mov_b32 v19, s35
; GFX11-PAL-GCNTRACKERS-NEXT: v_dual_mov_b32 v18, s19 :: v_dual_mov_b32 v21, s34
; GFX11-PAL-GCNTRACKERS-NEXT: v_dual_mov_b32 v20, s16 :: v_dual_mov_b32 v23, s33
; GFX11-PAL-GCNTRACKERS-NEXT: v_mov_b32_e32 v22, s17
; GFX11-PAL-GCNTRACKERS-NEXT: s_clause 0x5
; GFX11-PAL-GCNTRACKERS-NEXT: global_store_b128 v24, v[0:3], s[36:37] offset:80
; GFX11-PAL-GCNTRACKERS-NEXT: global_store_b128 v24, v[4:7], s[36:37] offset:64
; GFX11-PAL-GCNTRACKERS-NEXT: global_store_b128 v24, v[8:11], s[36:37] offset:48
; GFX11-PAL-GCNTRACKERS-NEXT: global_store_b128 v24, v[12:15], s[36:37] offset:32
; GFX11-PAL-GCNTRACKERS-NEXT: global_store_b128 v24, v[16:19], s[36:37] offset:16
; GFX11-PAL-GCNTRACKERS-NEXT: global_store_b128 v24, v[20:23], s[36:37]
; GFX11-PAL-GCNTRACKERS-NEXT: s_endpgm
;
; TONGA-LABEL: constant_zextload_v64i16_to_v64i32:
; TONGA: ; %bb.0:
; TONGA-NEXT: s_load_dwordx4 s[36:39], s[4:5], 0x24
; TONGA-NEXT: s_waitcnt lgkmcnt(0)
; TONGA-NEXT: s_load_dwordx16 s[16:31], s[38:39], 0x0
; TONGA-NEXT: s_load_dwordx16 s[0:15], s[38:39], 0x40
; TONGA-NEXT: s_waitcnt lgkmcnt(0)
; TONGA-NEXT: s_lshr_b32 s33, s17, 16
; TONGA-NEXT: s_lshr_b32 s34, s16, 16
; TONGA-NEXT: s_lshr_b32 s35, s19, 16
; TONGA-NEXT: s_lshr_b32 s40, s18, 16
; TONGA-NEXT: s_lshr_b32 s41, s21, 16
; TONGA-NEXT: s_lshr_b32 s42, s20, 16
; TONGA-NEXT: s_lshr_b32 s43, s23, 16
; TONGA-NEXT: s_lshr_b32 s44, s22, 16
; TONGA-NEXT: s_lshr_b32 s45, s25, 16
; TONGA-NEXT: s_lshr_b32 s46, s24, 16
; TONGA-NEXT: s_lshr_b32 s47, s27, 16
; TONGA-NEXT: s_lshr_b32 s48, s26, 16
; TONGA-NEXT: s_lshr_b32 s38, s29, 16
; TONGA-NEXT: s_lshr_b32 s39, s28, 16
; TONGA-NEXT: s_lshr_b32 s49, s31, 16
; TONGA-NEXT: s_lshr_b32 s50, s30, 16
; TONGA-NEXT: s_lshr_b32 s51, s1, 16
; TONGA-NEXT: s_lshr_b32 s52, s0, 16
; TONGA-NEXT: s_lshr_b32 s53, s3, 16
; TONGA-NEXT: s_lshr_b32 s54, s2, 16
; TONGA-NEXT: s_lshr_b32 s55, s5, 16
; TONGA-NEXT: s_lshr_b32 s56, s4, 16
; TONGA-NEXT: s_lshr_b32 s57, s7, 16
; TONGA-NEXT: s_lshr_b32 s58, s6, 16
; TONGA-NEXT: s_lshr_b32 s59, s9, 16
; TONGA-NEXT: s_and_b32 s17, s17, 0xffff
; TONGA-NEXT: s_and_b32 s16, s16, 0xffff
; TONGA-NEXT: s_and_b32 s19, s19, 0xffff
; TONGA-NEXT: s_and_b32 s18, s18, 0xffff
; TONGA-NEXT: s_and_b32 s21, s21, 0xffff
; TONGA-NEXT: s_and_b32 s20, s20, 0xffff
; TONGA-NEXT: s_and_b32 s23, s23, 0xffff
; TONGA-NEXT: s_and_b32 s22, s22, 0xffff
; TONGA-NEXT: s_and_b32 s25, s25, 0xffff
; TONGA-NEXT: s_and_b32 s24, s24, 0xffff
; TONGA-NEXT: s_and_b32 s27, s27, 0xffff
; TONGA-NEXT: s_and_b32 s26, s26, 0xffff
; TONGA-NEXT: s_and_b32 s29, s29, 0xffff
; TONGA-NEXT: s_and_b32 s28, s28, 0xffff
; TONGA-NEXT: s_and_b32 s31, s31, 0xffff
; TONGA-NEXT: s_and_b32 s30, s30, 0xffff
; TONGA-NEXT: s_lshr_b32 s60, s8, 16
; TONGA-NEXT: s_lshr_b32 s61, s11, 16
; TONGA-NEXT: s_lshr_b32 s62, s10, 16
; TONGA-NEXT: s_lshr_b32 s63, s13, 16
; TONGA-NEXT: s_lshr_b32 s64, s12, 16
; TONGA-NEXT: s_lshr_b32 s65, s15, 16
; TONGA-NEXT: s_lshr_b32 s66, s14, 16
; TONGA-NEXT: s_and_b32 s67, s1, 0xffff
; TONGA-NEXT: s_and_b32 s68, s0, 0xffff
; TONGA-NEXT: s_and_b32 s3, s3, 0xffff
; TONGA-NEXT: s_and_b32 s2, s2, 0xffff
; TONGA-NEXT: s_and_b32 s5, s5, 0xffff
; TONGA-NEXT: s_and_b32 s4, s4, 0xffff
; TONGA-NEXT: s_and_b32 s7, s7, 0xffff
; TONGA-NEXT: s_and_b32 s6, s6, 0xffff
; TONGA-NEXT: s_and_b32 s9, s9, 0xffff
; TONGA-NEXT: s_and_b32 s8, s8, 0xffff
; TONGA-NEXT: s_and_b32 s11, s11, 0xffff
; TONGA-NEXT: s_and_b32 s10, s10, 0xffff
; TONGA-NEXT: s_and_b32 s13, s13, 0xffff
; TONGA-NEXT: s_and_b32 s12, s12, 0xffff
; TONGA-NEXT: s_and_b32 s0, s15, 0xffff
; TONGA-NEXT: s_and_b32 s1, s14, 0xffff
; TONGA-NEXT: v_mov_b32_e32 v2, s0
; TONGA-NEXT: s_add_u32 s0, s36, 0xf0
; TONGA-NEXT: v_mov_b32_e32 v0, s1
; TONGA-NEXT: s_addc_u32 s1, s37, 0
; TONGA-NEXT: v_mov_b32_e32 v5, s1
; TONGA-NEXT: v_mov_b32_e32 v4, s0
; TONGA-NEXT: s_add_u32 s0, s36, 0xe0
; TONGA-NEXT: v_mov_b32_e32 v1, s66
; TONGA-NEXT: v_mov_b32_e32 v3, s65
; TONGA-NEXT: s_addc_u32 s1, s37, 0
; TONGA-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; TONGA-NEXT: v_mov_b32_e32 v5, s1
; TONGA-NEXT: v_mov_b32_e32 v4, s0
; TONGA-NEXT: s_add_u32 s0, s36, 0xd0
; TONGA-NEXT: v_mov_b32_e32 v0, s12
; TONGA-NEXT: v_mov_b32_e32 v1, s64
; TONGA-NEXT: v_mov_b32_e32 v2, s13
; TONGA-NEXT: v_mov_b32_e32 v3, s63
; TONGA-NEXT: s_addc_u32 s1, s37, 0
; TONGA-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; TONGA-NEXT: v_mov_b32_e32 v5, s1
; TONGA-NEXT: v_mov_b32_e32 v4, s0
; TONGA-NEXT: s_add_u32 s0, s36, 0xc0
; TONGA-NEXT: v_mov_b32_e32 v0, s10
; TONGA-NEXT: v_mov_b32_e32 v1, s62
; TONGA-NEXT: v_mov_b32_e32 v2, s11
; TONGA-NEXT: v_mov_b32_e32 v3, s61
; TONGA-NEXT: s_addc_u32 s1, s37, 0
; TONGA-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; TONGA-NEXT: v_mov_b32_e32 v5, s1
; TONGA-NEXT: v_mov_b32_e32 v4, s0
; TONGA-NEXT: s_add_u32 s0, s36, 0xb0
; TONGA-NEXT: v_mov_b32_e32 v0, s8
; TONGA-NEXT: v_mov_b32_e32 v1, s60
; TONGA-NEXT: v_mov_b32_e32 v2, s9
; TONGA-NEXT: v_mov_b32_e32 v3, s59
; TONGA-NEXT: s_addc_u32 s1, s37, 0
; TONGA-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; TONGA-NEXT: v_mov_b32_e32 v5, s1
; TONGA-NEXT: v_mov_b32_e32 v4, s0
; TONGA-NEXT: s_add_u32 s0, s36, 0xa0
; TONGA-NEXT: v_mov_b32_e32 v0, s6
; TONGA-NEXT: v_mov_b32_e32 v1, s58
; TONGA-NEXT: v_mov_b32_e32 v2, s7
; TONGA-NEXT: v_mov_b32_e32 v3, s57
; TONGA-NEXT: s_addc_u32 s1, s37, 0
; TONGA-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; TONGA-NEXT: v_mov_b32_e32 v5, s1
; TONGA-NEXT: v_mov_b32_e32 v4, s0
; TONGA-NEXT: s_add_u32 s0, s36, 0x90
; TONGA-NEXT: v_mov_b32_e32 v0, s4
; TONGA-NEXT: v_mov_b32_e32 v1, s56
; TONGA-NEXT: v_mov_b32_e32 v2, s5
; TONGA-NEXT: v_mov_b32_e32 v3, s55
; TONGA-NEXT: s_addc_u32 s1, s37, 0
; TONGA-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; TONGA-NEXT: v_mov_b32_e32 v5, s1
; TONGA-NEXT: v_mov_b32_e32 v4, s0
; TONGA-NEXT: s_add_u32 s0, s36, 0x80
; TONGA-NEXT: v_mov_b32_e32 v0, s2
; TONGA-NEXT: v_mov_b32_e32 v1, s54
; TONGA-NEXT: v_mov_b32_e32 v2, s3
; TONGA-NEXT: v_mov_b32_e32 v3, s53
; TONGA-NEXT: s_addc_u32 s1, s37, 0
; TONGA-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; TONGA-NEXT: v_mov_b32_e32 v5, s1
; TONGA-NEXT: v_mov_b32_e32 v4, s0
; TONGA-NEXT: s_add_u32 s0, s36, 0x70
; TONGA-NEXT: v_mov_b32_e32 v0, s68
; TONGA-NEXT: v_mov_b32_e32 v1, s52
; TONGA-NEXT: v_mov_b32_e32 v2, s67
; TONGA-NEXT: v_mov_b32_e32 v3, s51
; TONGA-NEXT: s_addc_u32 s1, s37, 0
; TONGA-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; TONGA-NEXT: v_mov_b32_e32 v5, s1
; TONGA-NEXT: v_mov_b32_e32 v4, s0
; TONGA-NEXT: s_add_u32 s0, s36, 0x60
; TONGA-NEXT: v_mov_b32_e32 v0, s30
; TONGA-NEXT: v_mov_b32_e32 v1, s50
; TONGA-NEXT: v_mov_b32_e32 v2, s31
; TONGA-NEXT: v_mov_b32_e32 v3, s49
; TONGA-NEXT: s_addc_u32 s1, s37, 0
; TONGA-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; TONGA-NEXT: v_mov_b32_e32 v5, s1
; TONGA-NEXT: v_mov_b32_e32 v4, s0
; TONGA-NEXT: s_add_u32 s0, s36, 0x50
; TONGA-NEXT: v_mov_b32_e32 v0, s28
; TONGA-NEXT: v_mov_b32_e32 v1, s39
; TONGA-NEXT: v_mov_b32_e32 v2, s29
; TONGA-NEXT: v_mov_b32_e32 v3, s38
; TONGA-NEXT: s_addc_u32 s1, s37, 0
; TONGA-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; TONGA-NEXT: v_mov_b32_e32 v5, s1
; TONGA-NEXT: v_mov_b32_e32 v4, s0
; TONGA-NEXT: s_add_u32 s0, s36, 64
; TONGA-NEXT: v_mov_b32_e32 v0, s26
; TONGA-NEXT: v_mov_b32_e32 v1, s48
; TONGA-NEXT: v_mov_b32_e32 v2, s27
; TONGA-NEXT: v_mov_b32_e32 v3, s47
; TONGA-NEXT: s_addc_u32 s1, s37, 0
; TONGA-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; TONGA-NEXT: v_mov_b32_e32 v5, s1
; TONGA-NEXT: v_mov_b32_e32 v4, s0
; TONGA-NEXT: s_add_u32 s0, s36, 48
; TONGA-NEXT: v_mov_b32_e32 v0, s24
; TONGA-NEXT: v_mov_b32_e32 v1, s46
; TONGA-NEXT: v_mov_b32_e32 v2, s25
; TONGA-NEXT: v_mov_b32_e32 v3, s45
; TONGA-NEXT: s_addc_u32 s1, s37, 0
; TONGA-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; TONGA-NEXT: v_mov_b32_e32 v5, s1
; TONGA-NEXT: v_mov_b32_e32 v4, s0
; TONGA-NEXT: s_add_u32 s0, s36, 32
; TONGA-NEXT: v_mov_b32_e32 v0, s22
; TONGA-NEXT: v_mov_b32_e32 v1, s44
; TONGA-NEXT: v_mov_b32_e32 v2, s23
; TONGA-NEXT: v_mov_b32_e32 v3, s43
; TONGA-NEXT: s_addc_u32 s1, s37, 0
; TONGA-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; TONGA-NEXT: v_mov_b32_e32 v5, s1
; TONGA-NEXT: v_mov_b32_e32 v4, s0
; TONGA-NEXT: s_add_u32 s0, s36, 16
; TONGA-NEXT: v_mov_b32_e32 v0, s20
; TONGA-NEXT: v_mov_b32_e32 v1, s42
; TONGA-NEXT: v_mov_b32_e32 v2, s21
; TONGA-NEXT: v_mov_b32_e32 v3, s41
; TONGA-NEXT: s_addc_u32 s1, s37, 0
; TONGA-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; TONGA-NEXT: v_mov_b32_e32 v5, s1
; TONGA-NEXT: v_mov_b32_e32 v0, s18
; TONGA-NEXT: v_mov_b32_e32 v1, s40
; TONGA-NEXT: v_mov_b32_e32 v2, s19
; TONGA-NEXT: v_mov_b32_e32 v3, s35
; TONGA-NEXT: v_mov_b32_e32 v4, s0
; TONGA-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; TONGA-NEXT: v_mov_b32_e32 v4, s36
; TONGA-NEXT: v_mov_b32_e32 v0, s16
; TONGA-NEXT: v_mov_b32_e32 v1, s34
; TONGA-NEXT: v_mov_b32_e32 v2, s17
; TONGA-NEXT: v_mov_b32_e32 v3, s33
; TONGA-NEXT: v_mov_b32_e32 v5, s37
; TONGA-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; TONGA-NEXT: s_endpgm
;
; TONGA-GCNTRACKERS-LABEL: constant_zextload_v64i16_to_v64i32:
; TONGA-GCNTRACKERS: ; %bb.0:
; TONGA-GCNTRACKERS-NEXT: s_load_dwordx4 s[36:39], s[4:5], 0x24
; TONGA-GCNTRACKERS-NEXT: s_waitcnt lgkmcnt(0)
; TONGA-GCNTRACKERS-NEXT: s_load_dwordx16 s[0:15], s[38:39], 0x0
; TONGA-GCNTRACKERS-NEXT: s_load_dwordx16 s[16:31], s[38:39], 0x40
; TONGA-GCNTRACKERS-NEXT: s_waitcnt lgkmcnt(0)
; TONGA-GCNTRACKERS-NEXT: s_lshr_b32 s33, s1, 16
; TONGA-GCNTRACKERS-NEXT: s_lshr_b32 s34, s0, 16
; TONGA-GCNTRACKERS-NEXT: s_lshr_b32 s35, s3, 16
; TONGA-GCNTRACKERS-NEXT: s_lshr_b32 s40, s2, 16
; TONGA-GCNTRACKERS-NEXT: s_lshr_b32 s41, s5, 16
; TONGA-GCNTRACKERS-NEXT: s_lshr_b32 s42, s4, 16
; TONGA-GCNTRACKERS-NEXT: s_lshr_b32 s43, s7, 16
; TONGA-GCNTRACKERS-NEXT: s_lshr_b32 s44, s6, 16
; TONGA-GCNTRACKERS-NEXT: s_lshr_b32 s45, s9, 16
; TONGA-GCNTRACKERS-NEXT: s_lshr_b32 s46, s8, 16
; TONGA-GCNTRACKERS-NEXT: s_lshr_b32 s47, s11, 16
; TONGA-GCNTRACKERS-NEXT: s_lshr_b32 s48, s10, 16
; TONGA-GCNTRACKERS-NEXT: s_lshr_b32 s38, s13, 16
; TONGA-GCNTRACKERS-NEXT: s_lshr_b32 s39, s12, 16
; TONGA-GCNTRACKERS-NEXT: s_lshr_b32 s49, s15, 16
; TONGA-GCNTRACKERS-NEXT: s_lshr_b32 s50, s14, 16
; TONGA-GCNTRACKERS-NEXT: s_lshr_b32 s51, s17, 16
; TONGA-GCNTRACKERS-NEXT: s_lshr_b32 s52, s16, 16
; TONGA-GCNTRACKERS-NEXT: s_lshr_b32 s53, s19, 16
; TONGA-GCNTRACKERS-NEXT: s_lshr_b32 s54, s18, 16
; TONGA-GCNTRACKERS-NEXT: s_lshr_b32 s55, s21, 16
; TONGA-GCNTRACKERS-NEXT: s_lshr_b32 s56, s20, 16
; TONGA-GCNTRACKERS-NEXT: s_lshr_b32 s57, s23, 16
; TONGA-GCNTRACKERS-NEXT: s_lshr_b32 s58, s22, 16
; TONGA-GCNTRACKERS-NEXT: s_lshr_b32 s59, s25, 16
; TONGA-GCNTRACKERS-NEXT: s_lshr_b32 s60, s24, 16
; TONGA-GCNTRACKERS-NEXT: s_lshr_b32 s61, s27, 16
; TONGA-GCNTRACKERS-NEXT: s_lshr_b32 s62, s26, 16
; TONGA-GCNTRACKERS-NEXT: s_lshr_b32 s63, s29, 16
; TONGA-GCNTRACKERS-NEXT: s_lshr_b32 s64, s28, 16
; TONGA-GCNTRACKERS-NEXT: s_lshr_b32 s65, s31, 16
; TONGA-GCNTRACKERS-NEXT: s_lshr_b32 s66, s30, 16
; TONGA-GCNTRACKERS-NEXT: s_and_b32 s1, s1, 0xffff
; TONGA-GCNTRACKERS-NEXT: s_and_b32 s0, s0, 0xffff
; TONGA-GCNTRACKERS-NEXT: s_and_b32 s3, s3, 0xffff
; TONGA-GCNTRACKERS-NEXT: s_and_b32 s2, s2, 0xffff
; TONGA-GCNTRACKERS-NEXT: s_and_b32 s5, s5, 0xffff
; TONGA-GCNTRACKERS-NEXT: s_and_b32 s4, s4, 0xffff
; TONGA-GCNTRACKERS-NEXT: s_and_b32 s7, s7, 0xffff
; TONGA-GCNTRACKERS-NEXT: s_and_b32 s6, s6, 0xffff
; TONGA-GCNTRACKERS-NEXT: s_and_b32 s9, s9, 0xffff
; TONGA-GCNTRACKERS-NEXT: s_and_b32 s8, s8, 0xffff
; TONGA-GCNTRACKERS-NEXT: s_and_b32 s67, s11, 0xffff
; TONGA-GCNTRACKERS-NEXT: s_and_b32 s68, s10, 0xffff
; TONGA-GCNTRACKERS-NEXT: s_and_b32 s13, s13, 0xffff
; TONGA-GCNTRACKERS-NEXT: s_and_b32 s12, s12, 0xffff
; TONGA-GCNTRACKERS-NEXT: s_and_b32 s15, s15, 0xffff
; TONGA-GCNTRACKERS-NEXT: s_and_b32 s14, s14, 0xffff
; TONGA-GCNTRACKERS-NEXT: s_and_b32 s17, s17, 0xffff
; TONGA-GCNTRACKERS-NEXT: s_and_b32 s16, s16, 0xffff
; TONGA-GCNTRACKERS-NEXT: s_and_b32 s19, s19, 0xffff
; TONGA-GCNTRACKERS-NEXT: s_and_b32 s18, s18, 0xffff
; TONGA-GCNTRACKERS-NEXT: s_and_b32 s21, s21, 0xffff
; TONGA-GCNTRACKERS-NEXT: s_and_b32 s20, s20, 0xffff
; TONGA-GCNTRACKERS-NEXT: s_and_b32 s23, s23, 0xffff
; TONGA-GCNTRACKERS-NEXT: s_and_b32 s22, s22, 0xffff
; TONGA-GCNTRACKERS-NEXT: s_and_b32 s25, s25, 0xffff
; TONGA-GCNTRACKERS-NEXT: s_and_b32 s24, s24, 0xffff
; TONGA-GCNTRACKERS-NEXT: s_and_b32 s27, s27, 0xffff
; TONGA-GCNTRACKERS-NEXT: s_and_b32 s26, s26, 0xffff
; TONGA-GCNTRACKERS-NEXT: s_and_b32 s29, s29, 0xffff
; TONGA-GCNTRACKERS-NEXT: s_and_b32 s28, s28, 0xffff
; TONGA-GCNTRACKERS-NEXT: s_and_b32 s10, s31, 0xffff
; TONGA-GCNTRACKERS-NEXT: s_and_b32 s11, s30, 0xffff
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v2, s10
; TONGA-GCNTRACKERS-NEXT: s_add_u32 s10, s36, 0xf0
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v0, s11
; TONGA-GCNTRACKERS-NEXT: s_addc_u32 s11, s37, 0
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v4, s10
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v5, s11
; TONGA-GCNTRACKERS-NEXT: s_add_u32 s10, s36, 0xe0
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v1, s66
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v3, s65
; TONGA-GCNTRACKERS-NEXT: s_addc_u32 s11, s37, 0
; TONGA-GCNTRACKERS-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v4, s10
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v5, s11
; TONGA-GCNTRACKERS-NEXT: s_add_u32 s10, s36, 0xd0
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v0, s28
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v1, s64
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v2, s29
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v3, s63
; TONGA-GCNTRACKERS-NEXT: s_addc_u32 s11, s37, 0
; TONGA-GCNTRACKERS-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v4, s10
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v5, s11
; TONGA-GCNTRACKERS-NEXT: s_add_u32 s10, s36, 0xc0
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v0, s26
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v1, s62
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v2, s27
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v3, s61
; TONGA-GCNTRACKERS-NEXT: s_addc_u32 s11, s37, 0
; TONGA-GCNTRACKERS-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v4, s10
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v5, s11
; TONGA-GCNTRACKERS-NEXT: s_add_u32 s10, s36, 0xb0
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v0, s24
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v1, s60
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v2, s25
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v3, s59
; TONGA-GCNTRACKERS-NEXT: s_addc_u32 s11, s37, 0
; TONGA-GCNTRACKERS-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v4, s10
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v5, s11
; TONGA-GCNTRACKERS-NEXT: s_add_u32 s10, s36, 0xa0
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v0, s22
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v1, s58
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v2, s23
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v3, s57
; TONGA-GCNTRACKERS-NEXT: s_addc_u32 s11, s37, 0
; TONGA-GCNTRACKERS-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v4, s10
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v5, s11
; TONGA-GCNTRACKERS-NEXT: s_add_u32 s10, s36, 0x90
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v0, s20
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v1, s56
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v2, s21
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v3, s55
; TONGA-GCNTRACKERS-NEXT: s_addc_u32 s11, s37, 0
; TONGA-GCNTRACKERS-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v4, s10
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v5, s11
; TONGA-GCNTRACKERS-NEXT: s_add_u32 s10, s36, 0x80
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v0, s18
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v1, s54
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v2, s19
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v3, s53
; TONGA-GCNTRACKERS-NEXT: s_addc_u32 s11, s37, 0
; TONGA-GCNTRACKERS-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v4, s10
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v5, s11
; TONGA-GCNTRACKERS-NEXT: s_add_u32 s10, s36, 0x70
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v0, s16
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v1, s52
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v2, s17
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v3, s51
; TONGA-GCNTRACKERS-NEXT: s_addc_u32 s11, s37, 0
; TONGA-GCNTRACKERS-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v4, s10
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v5, s11
; TONGA-GCNTRACKERS-NEXT: s_add_u32 s10, s36, 0x60
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v0, s14
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v1, s50
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v2, s15
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v3, s49
; TONGA-GCNTRACKERS-NEXT: s_addc_u32 s11, s37, 0
; TONGA-GCNTRACKERS-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v4, s10
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v5, s11
; TONGA-GCNTRACKERS-NEXT: s_add_u32 s10, s36, 0x50
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v0, s12
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v1, s39
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v2, s13
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v3, s38
; TONGA-GCNTRACKERS-NEXT: s_addc_u32 s11, s37, 0
; TONGA-GCNTRACKERS-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v4, s10
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v0, s68
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v1, s48
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v2, s67
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v3, s47
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v5, s11
; TONGA-GCNTRACKERS-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; TONGA-GCNTRACKERS-NEXT: s_nop 0
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v0, s8
; TONGA-GCNTRACKERS-NEXT: s_add_u32 s8, s36, 64
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v2, s9
; TONGA-GCNTRACKERS-NEXT: s_addc_u32 s9, s37, 0
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v1, s46
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v3, s45
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v4, s8
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v5, s9
; TONGA-GCNTRACKERS-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; TONGA-GCNTRACKERS-NEXT: s_nop 0
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v0, s6
; TONGA-GCNTRACKERS-NEXT: s_add_u32 s6, s36, 48
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v2, s7
; TONGA-GCNTRACKERS-NEXT: s_addc_u32 s7, s37, 0
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v1, s44
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v3, s43
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v4, s6
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v5, s7
; TONGA-GCNTRACKERS-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; TONGA-GCNTRACKERS-NEXT: s_nop 0
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v0, s4
; TONGA-GCNTRACKERS-NEXT: s_add_u32 s4, s36, 32
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v2, s5
; TONGA-GCNTRACKERS-NEXT: s_addc_u32 s5, s37, 0
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v1, s42
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v3, s41
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v4, s4
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v5, s5
; TONGA-GCNTRACKERS-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; TONGA-GCNTRACKERS-NEXT: s_nop 0
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v0, s2
; TONGA-GCNTRACKERS-NEXT: s_add_u32 s2, s36, 16
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v2, s3
; TONGA-GCNTRACKERS-NEXT: s_addc_u32 s3, s37, 0
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v1, s40
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v3, s35
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v5, s3
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v4, s2
; TONGA-GCNTRACKERS-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v4, s36
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v0, s0
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v1, s34
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v2, s1
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v3, s33
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v5, s37
; TONGA-GCNTRACKERS-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; TONGA-GCNTRACKERS-NEXT: s_endpgm
;
; GFX908-LABEL: constant_zextload_v64i16_to_v64i32:
; GFX908: ; %bb.0:
; GFX908-NEXT: s_load_dwordx4 s[36:39], s[4:5], 0x24
; GFX908-NEXT: v_mov_b32_e32 v0, 0
; GFX908-NEXT: s_waitcnt lgkmcnt(0)
; GFX908-NEXT: s_load_dwordx16 s[0:15], s[38:39], 0x40
; GFX908-NEXT: s_load_dwordx16 s[16:31], s[38:39], 0x0
; GFX908-NEXT: s_waitcnt lgkmcnt(0)
; GFX908-NEXT: s_lshr_b32 s65, s15, 16
; GFX908-NEXT: s_lshr_b32 s66, s14, 16
; GFX908-NEXT: s_and_b32 s15, s15, 0xffff
; GFX908-NEXT: s_and_b32 s14, s14, 0xffff
; GFX908-NEXT: s_lshr_b32 s63, s13, 16
; GFX908-NEXT: s_lshr_b32 s64, s12, 16
; GFX908-NEXT: s_and_b32 s13, s13, 0xffff
; GFX908-NEXT: s_and_b32 s12, s12, 0xffff
; GFX908-NEXT: v_mov_b32_e32 v1, s14
; GFX908-NEXT: v_mov_b32_e32 v2, s66
; GFX908-NEXT: v_mov_b32_e32 v3, s15
; GFX908-NEXT: v_mov_b32_e32 v4, s65
; GFX908-NEXT: s_lshr_b32 s61, s11, 16
; GFX908-NEXT: s_lshr_b32 s62, s10, 16
; GFX908-NEXT: s_and_b32 s11, s11, 0xffff
; GFX908-NEXT: s_and_b32 s10, s10, 0xffff
; GFX908-NEXT: global_store_dwordx4 v0, v[1:4], s[36:37] offset:240
; GFX908-NEXT: s_lshr_b32 s59, s9, 16
; GFX908-NEXT: v_mov_b32_e32 v1, s12
; GFX908-NEXT: v_mov_b32_e32 v2, s64
; GFX908-NEXT: v_mov_b32_e32 v3, s13
; GFX908-NEXT: v_mov_b32_e32 v4, s63
; GFX908-NEXT: s_lshr_b32 s60, s8, 16
; GFX908-NEXT: s_and_b32 s9, s9, 0xffff
; GFX908-NEXT: s_and_b32 s8, s8, 0xffff
; GFX908-NEXT: global_store_dwordx4 v0, v[1:4], s[36:37] offset:224
; GFX908-NEXT: s_lshr_b32 s57, s7, 16
; GFX908-NEXT: v_mov_b32_e32 v1, s10
; GFX908-NEXT: v_mov_b32_e32 v2, s62
; GFX908-NEXT: v_mov_b32_e32 v3, s11
; GFX908-NEXT: v_mov_b32_e32 v4, s61
; GFX908-NEXT: s_lshr_b32 s58, s6, 16
; GFX908-NEXT: s_and_b32 s7, s7, 0xffff
; GFX908-NEXT: s_and_b32 s6, s6, 0xffff
; GFX908-NEXT: global_store_dwordx4 v0, v[1:4], s[36:37] offset:208
; GFX908-NEXT: s_lshr_b32 s55, s5, 16
; GFX908-NEXT: v_mov_b32_e32 v1, s8
; GFX908-NEXT: v_mov_b32_e32 v2, s60
; GFX908-NEXT: v_mov_b32_e32 v3, s9
; GFX908-NEXT: v_mov_b32_e32 v4, s59
; GFX908-NEXT: s_lshr_b32 s56, s4, 16
; GFX908-NEXT: s_and_b32 s5, s5, 0xffff
; GFX908-NEXT: s_and_b32 s4, s4, 0xffff
; GFX908-NEXT: global_store_dwordx4 v0, v[1:4], s[36:37] offset:192
; GFX908-NEXT: s_lshr_b32 s53, s3, 16
; GFX908-NEXT: v_mov_b32_e32 v1, s6
; GFX908-NEXT: v_mov_b32_e32 v2, s58
; GFX908-NEXT: v_mov_b32_e32 v3, s7
; GFX908-NEXT: v_mov_b32_e32 v4, s57
; GFX908-NEXT: s_lshr_b32 s54, s2, 16
; GFX908-NEXT: s_and_b32 s3, s3, 0xffff
; GFX908-NEXT: s_and_b32 s2, s2, 0xffff
; GFX908-NEXT: global_store_dwordx4 v0, v[1:4], s[36:37] offset:176
; GFX908-NEXT: s_lshr_b32 s51, s1, 16
; GFX908-NEXT: v_mov_b32_e32 v1, s4
; GFX908-NEXT: v_mov_b32_e32 v2, s56
; GFX908-NEXT: v_mov_b32_e32 v3, s5
; GFX908-NEXT: v_mov_b32_e32 v4, s55
; GFX908-NEXT: s_lshr_b32 s52, s0, 16
; GFX908-NEXT: s_and_b32 s1, s1, 0xffff
; GFX908-NEXT: s_and_b32 s0, s0, 0xffff
; GFX908-NEXT: global_store_dwordx4 v0, v[1:4], s[36:37] offset:160
; GFX908-NEXT: s_lshr_b32 s49, s31, 16
; GFX908-NEXT: v_mov_b32_e32 v1, s2
; GFX908-NEXT: v_mov_b32_e32 v2, s54
; GFX908-NEXT: v_mov_b32_e32 v3, s3
; GFX908-NEXT: v_mov_b32_e32 v4, s53
; GFX908-NEXT: s_lshr_b32 s50, s30, 16
; GFX908-NEXT: s_and_b32 s31, s31, 0xffff
; GFX908-NEXT: s_and_b32 s30, s30, 0xffff
; GFX908-NEXT: global_store_dwordx4 v0, v[1:4], s[36:37] offset:144
; GFX908-NEXT: s_lshr_b32 s38, s29, 16
; GFX908-NEXT: v_mov_b32_e32 v1, s0
; GFX908-NEXT: v_mov_b32_e32 v2, s52
; GFX908-NEXT: v_mov_b32_e32 v3, s1
; GFX908-NEXT: v_mov_b32_e32 v4, s51
; GFX908-NEXT: s_lshr_b32 s39, s28, 16
; GFX908-NEXT: s_and_b32 s29, s29, 0xffff
; GFX908-NEXT: s_and_b32 s28, s28, 0xffff
; GFX908-NEXT: global_store_dwordx4 v0, v[1:4], s[36:37] offset:128
; GFX908-NEXT: s_lshr_b32 s47, s27, 16
; GFX908-NEXT: v_mov_b32_e32 v1, s30
; GFX908-NEXT: v_mov_b32_e32 v2, s50
; GFX908-NEXT: v_mov_b32_e32 v3, s31
; GFX908-NEXT: v_mov_b32_e32 v4, s49
; GFX908-NEXT: s_lshr_b32 s48, s26, 16
; GFX908-NEXT: s_and_b32 s27, s27, 0xffff
; GFX908-NEXT: s_and_b32 s26, s26, 0xffff
; GFX908-NEXT: global_store_dwordx4 v0, v[1:4], s[36:37] offset:112
; GFX908-NEXT: s_lshr_b32 s45, s25, 16
; GFX908-NEXT: v_mov_b32_e32 v1, s28
; GFX908-NEXT: v_mov_b32_e32 v2, s39
; GFX908-NEXT: v_mov_b32_e32 v3, s29
; GFX908-NEXT: v_mov_b32_e32 v4, s38
; GFX908-NEXT: s_lshr_b32 s46, s24, 16
; GFX908-NEXT: s_and_b32 s25, s25, 0xffff
; GFX908-NEXT: s_and_b32 s24, s24, 0xffff
; GFX908-NEXT: global_store_dwordx4 v0, v[1:4], s[36:37] offset:96
; GFX908-NEXT: s_lshr_b32 s43, s23, 16
; GFX908-NEXT: v_mov_b32_e32 v1, s26
; GFX908-NEXT: v_mov_b32_e32 v2, s48
; GFX908-NEXT: v_mov_b32_e32 v3, s27
; GFX908-NEXT: v_mov_b32_e32 v4, s47
; GFX908-NEXT: s_lshr_b32 s44, s22, 16
; GFX908-NEXT: s_and_b32 s23, s23, 0xffff
; GFX908-NEXT: s_and_b32 s22, s22, 0xffff
; GFX908-NEXT: global_store_dwordx4 v0, v[1:4], s[36:37] offset:80
; GFX908-NEXT: s_lshr_b32 s41, s21, 16
; GFX908-NEXT: v_mov_b32_e32 v1, s24
; GFX908-NEXT: v_mov_b32_e32 v2, s46
; GFX908-NEXT: v_mov_b32_e32 v3, s25
; GFX908-NEXT: v_mov_b32_e32 v4, s45
; GFX908-NEXT: s_lshr_b32 s42, s20, 16
; GFX908-NEXT: s_and_b32 s21, s21, 0xffff
; GFX908-NEXT: s_and_b32 s20, s20, 0xffff
; GFX908-NEXT: global_store_dwordx4 v0, v[1:4], s[36:37] offset:64
; GFX908-NEXT: s_lshr_b32 s35, s19, 16
; GFX908-NEXT: v_mov_b32_e32 v1, s22
; GFX908-NEXT: v_mov_b32_e32 v2, s44
; GFX908-NEXT: v_mov_b32_e32 v3, s23
; GFX908-NEXT: v_mov_b32_e32 v4, s43
; GFX908-NEXT: s_lshr_b32 s40, s18, 16
; GFX908-NEXT: s_and_b32 s19, s19, 0xffff
; GFX908-NEXT: s_and_b32 s18, s18, 0xffff
; GFX908-NEXT: global_store_dwordx4 v0, v[1:4], s[36:37] offset:48
; GFX908-NEXT: s_lshr_b32 s33, s17, 16
; GFX908-NEXT: v_mov_b32_e32 v1, s20
; GFX908-NEXT: v_mov_b32_e32 v2, s42
; GFX908-NEXT: v_mov_b32_e32 v3, s21
; GFX908-NEXT: v_mov_b32_e32 v4, s41
; GFX908-NEXT: s_lshr_b32 s34, s16, 16
; GFX908-NEXT: s_and_b32 s17, s17, 0xffff
; GFX908-NEXT: s_and_b32 s16, s16, 0xffff
; GFX908-NEXT: global_store_dwordx4 v0, v[1:4], s[36:37] offset:32
; GFX908-NEXT: s_nop 0
; GFX908-NEXT: v_mov_b32_e32 v1, s18
; GFX908-NEXT: v_mov_b32_e32 v2, s40
; GFX908-NEXT: v_mov_b32_e32 v3, s19
; GFX908-NEXT: v_mov_b32_e32 v4, s35
; GFX908-NEXT: global_store_dwordx4 v0, v[1:4], s[36:37] offset:16
; GFX908-NEXT: s_nop 0
; GFX908-NEXT: v_mov_b32_e32 v1, s16
; GFX908-NEXT: v_mov_b32_e32 v2, s34
; GFX908-NEXT: v_mov_b32_e32 v3, s17
; GFX908-NEXT: v_mov_b32_e32 v4, s33
; GFX908-NEXT: global_store_dwordx4 v0, v[1:4], s[36:37]
; GFX908-NEXT: s_endpgm
;
; GFX908-GCNTRACKERS-LABEL: constant_zextload_v64i16_to_v64i32:
; GFX908-GCNTRACKERS: ; %bb.0:
; GFX908-GCNTRACKERS-NEXT: s_load_dwordx4 s[36:39], s[4:5], 0x24
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v0, 0
; GFX908-GCNTRACKERS-NEXT: s_waitcnt lgkmcnt(0)
; GFX908-GCNTRACKERS-NEXT: s_load_dwordx16 s[16:31], s[38:39], 0x40
; GFX908-GCNTRACKERS-NEXT: s_load_dwordx16 s[0:15], s[38:39], 0x0
; GFX908-GCNTRACKERS-NEXT: s_waitcnt lgkmcnt(0)
; GFX908-GCNTRACKERS-NEXT: s_lshr_b32 s65, s31, 16
; GFX908-GCNTRACKERS-NEXT: s_lshr_b32 s66, s30, 16
; GFX908-GCNTRACKERS-NEXT: s_and_b32 s31, s31, 0xffff
; GFX908-GCNTRACKERS-NEXT: s_and_b32 s30, s30, 0xffff
; GFX908-GCNTRACKERS-NEXT: s_lshr_b32 s63, s29, 16
; GFX908-GCNTRACKERS-NEXT: s_lshr_b32 s64, s28, 16
; GFX908-GCNTRACKERS-NEXT: s_and_b32 s29, s29, 0xffff
; GFX908-GCNTRACKERS-NEXT: s_and_b32 s28, s28, 0xffff
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v1, s30
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v2, s66
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v3, s31
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v4, s65
; GFX908-GCNTRACKERS-NEXT: s_lshr_b32 s61, s27, 16
; GFX908-GCNTRACKERS-NEXT: s_lshr_b32 s62, s26, 16
; GFX908-GCNTRACKERS-NEXT: s_and_b32 s27, s27, 0xffff
; GFX908-GCNTRACKERS-NEXT: s_and_b32 s26, s26, 0xffff
; GFX908-GCNTRACKERS-NEXT: global_store_dwordx4 v0, v[1:4], s[36:37] offset:240
; GFX908-GCNTRACKERS-NEXT: s_lshr_b32 s59, s25, 16
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v1, s28
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v2, s64
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v3, s29
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v4, s63
; GFX908-GCNTRACKERS-NEXT: s_lshr_b32 s60, s24, 16
; GFX908-GCNTRACKERS-NEXT: s_and_b32 s25, s25, 0xffff
; GFX908-GCNTRACKERS-NEXT: s_and_b32 s24, s24, 0xffff
; GFX908-GCNTRACKERS-NEXT: global_store_dwordx4 v0, v[1:4], s[36:37] offset:224
; GFX908-GCNTRACKERS-NEXT: s_lshr_b32 s57, s23, 16
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v1, s26
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v2, s62
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v3, s27
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v4, s61
; GFX908-GCNTRACKERS-NEXT: s_lshr_b32 s58, s22, 16
; GFX908-GCNTRACKERS-NEXT: s_and_b32 s23, s23, 0xffff
; GFX908-GCNTRACKERS-NEXT: s_and_b32 s22, s22, 0xffff
; GFX908-GCNTRACKERS-NEXT: global_store_dwordx4 v0, v[1:4], s[36:37] offset:208
; GFX908-GCNTRACKERS-NEXT: s_lshr_b32 s55, s21, 16
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v1, s24
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v2, s60
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v3, s25
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v4, s59
; GFX908-GCNTRACKERS-NEXT: s_lshr_b32 s56, s20, 16
; GFX908-GCNTRACKERS-NEXT: s_and_b32 s21, s21, 0xffff
; GFX908-GCNTRACKERS-NEXT: s_and_b32 s20, s20, 0xffff
; GFX908-GCNTRACKERS-NEXT: global_store_dwordx4 v0, v[1:4], s[36:37] offset:192
; GFX908-GCNTRACKERS-NEXT: s_lshr_b32 s53, s19, 16
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v1, s22
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v2, s58
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v3, s23
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v4, s57
; GFX908-GCNTRACKERS-NEXT: s_lshr_b32 s54, s18, 16
; GFX908-GCNTRACKERS-NEXT: s_and_b32 s19, s19, 0xffff
; GFX908-GCNTRACKERS-NEXT: s_and_b32 s18, s18, 0xffff
; GFX908-GCNTRACKERS-NEXT: global_store_dwordx4 v0, v[1:4], s[36:37] offset:176
; GFX908-GCNTRACKERS-NEXT: s_lshr_b32 s51, s17, 16
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v1, s20
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v2, s56
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v3, s21
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v4, s55
; GFX908-GCNTRACKERS-NEXT: s_lshr_b32 s52, s16, 16
; GFX908-GCNTRACKERS-NEXT: s_and_b32 s17, s17, 0xffff
; GFX908-GCNTRACKERS-NEXT: s_and_b32 s16, s16, 0xffff
; GFX908-GCNTRACKERS-NEXT: global_store_dwordx4 v0, v[1:4], s[36:37] offset:160
; GFX908-GCNTRACKERS-NEXT: s_lshr_b32 s49, s15, 16
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v1, s18
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v2, s54
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v3, s19
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v4, s53
; GFX908-GCNTRACKERS-NEXT: s_lshr_b32 s50, s14, 16
; GFX908-GCNTRACKERS-NEXT: s_and_b32 s15, s15, 0xffff
; GFX908-GCNTRACKERS-NEXT: s_and_b32 s14, s14, 0xffff
; GFX908-GCNTRACKERS-NEXT: global_store_dwordx4 v0, v[1:4], s[36:37] offset:144
; GFX908-GCNTRACKERS-NEXT: s_lshr_b32 s38, s13, 16
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v1, s16
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v2, s52
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v3, s17
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v4, s51
; GFX908-GCNTRACKERS-NEXT: s_lshr_b32 s39, s12, 16
; GFX908-GCNTRACKERS-NEXT: s_and_b32 s13, s13, 0xffff
; GFX908-GCNTRACKERS-NEXT: s_and_b32 s12, s12, 0xffff
; GFX908-GCNTRACKERS-NEXT: global_store_dwordx4 v0, v[1:4], s[36:37] offset:128
; GFX908-GCNTRACKERS-NEXT: s_lshr_b32 s47, s11, 16
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v1, s14
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v2, s50
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v3, s15
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v4, s49
; GFX908-GCNTRACKERS-NEXT: s_lshr_b32 s48, s10, 16
; GFX908-GCNTRACKERS-NEXT: s_and_b32 s11, s11, 0xffff
; GFX908-GCNTRACKERS-NEXT: s_and_b32 s10, s10, 0xffff
; GFX908-GCNTRACKERS-NEXT: global_store_dwordx4 v0, v[1:4], s[36:37] offset:112
; GFX908-GCNTRACKERS-NEXT: s_lshr_b32 s45, s9, 16
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v1, s12
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v2, s39
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v3, s13
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v4, s38
; GFX908-GCNTRACKERS-NEXT: s_lshr_b32 s46, s8, 16
; GFX908-GCNTRACKERS-NEXT: s_and_b32 s9, s9, 0xffff
; GFX908-GCNTRACKERS-NEXT: s_and_b32 s8, s8, 0xffff
; GFX908-GCNTRACKERS-NEXT: global_store_dwordx4 v0, v[1:4], s[36:37] offset:96
; GFX908-GCNTRACKERS-NEXT: s_lshr_b32 s43, s7, 16
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v1, s10
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v2, s48
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v3, s11
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v4, s47
; GFX908-GCNTRACKERS-NEXT: s_lshr_b32 s44, s6, 16
; GFX908-GCNTRACKERS-NEXT: s_and_b32 s7, s7, 0xffff
; GFX908-GCNTRACKERS-NEXT: s_and_b32 s6, s6, 0xffff
; GFX908-GCNTRACKERS-NEXT: global_store_dwordx4 v0, v[1:4], s[36:37] offset:80
; GFX908-GCNTRACKERS-NEXT: s_lshr_b32 s41, s5, 16
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v1, s8
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v2, s46
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v3, s9
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v4, s45
; GFX908-GCNTRACKERS-NEXT: s_lshr_b32 s42, s4, 16
; GFX908-GCNTRACKERS-NEXT: s_and_b32 s5, s5, 0xffff
; GFX908-GCNTRACKERS-NEXT: s_and_b32 s4, s4, 0xffff
; GFX908-GCNTRACKERS-NEXT: global_store_dwordx4 v0, v[1:4], s[36:37] offset:64
; GFX908-GCNTRACKERS-NEXT: s_lshr_b32 s35, s3, 16
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v1, s6
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v2, s44
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v3, s7
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v4, s43
; GFX908-GCNTRACKERS-NEXT: s_lshr_b32 s40, s2, 16
; GFX908-GCNTRACKERS-NEXT: s_and_b32 s3, s3, 0xffff
; GFX908-GCNTRACKERS-NEXT: s_and_b32 s2, s2, 0xffff
; GFX908-GCNTRACKERS-NEXT: global_store_dwordx4 v0, v[1:4], s[36:37] offset:48
; GFX908-GCNTRACKERS-NEXT: s_lshr_b32 s33, s1, 16
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v1, s4
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v2, s42
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v3, s5
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v4, s41
; GFX908-GCNTRACKERS-NEXT: s_lshr_b32 s34, s0, 16
; GFX908-GCNTRACKERS-NEXT: s_and_b32 s1, s1, 0xffff
; GFX908-GCNTRACKERS-NEXT: s_and_b32 s0, s0, 0xffff
; GFX908-GCNTRACKERS-NEXT: global_store_dwordx4 v0, v[1:4], s[36:37] offset:32
; GFX908-GCNTRACKERS-NEXT: s_nop 0
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v1, s2
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v2, s40
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v3, s3
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v4, s35
; GFX908-GCNTRACKERS-NEXT: global_store_dwordx4 v0, v[1:4], s[36:37] offset:16
; GFX908-GCNTRACKERS-NEXT: s_nop 0
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v1, s0
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v2, s34
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v3, s1
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v4, s33
; GFX908-GCNTRACKERS-NEXT: global_store_dwordx4 v0, v[1:4], s[36:37]
; GFX908-GCNTRACKERS-NEXT: s_endpgm
;
; GENERIC-LABEL: constant_zextload_v64i16_to_v64i32:
; GENERIC: ; %bb.0:
; GENERIC-NEXT: s_load_dwordx4 s[36:39], s[4:5], 0x9
; GENERIC-NEXT: s_waitcnt lgkmcnt(0)
; GENERIC-NEXT: s_load_dwordx16 s[0:15], s[38:39], 0x0
; GENERIC-NEXT: s_load_dwordx16 s[16:31], s[38:39], 0x10
; GENERIC-NEXT: s_waitcnt lgkmcnt(0)
; GENERIC-NEXT: s_lshr_b32 s35, s3, 16
; GENERIC-NEXT: s_lshr_b32 s40, s2, 16
; GENERIC-NEXT: s_and_b32 s55, s3, 0xffff
; GENERIC-NEXT: s_and_b32 s56, s2, 0xffff
; GENERIC-NEXT: s_lshr_b32 s2, s22, 16
; GENERIC-NEXT: s_lshr_b32 s3, s25, 16
; GENERIC-NEXT: s_lshr_b32 s67, s31, 16
; GENERIC-NEXT: s_lshr_b32 s68, s30, 16
; GENERIC-NEXT: s_and_b32 s31, s31, 0xffff
; GENERIC-NEXT: s_and_b32 s30, s30, 0xffff
; GENERIC-NEXT: s_lshr_b32 s33, s1, 16
; GENERIC-NEXT: s_lshr_b32 s34, s0, 16
; GENERIC-NEXT: s_and_b32 s44, s1, 0xffff
; GENERIC-NEXT: s_and_b32 s46, s0, 0xffff
; GENERIC-NEXT: s_lshr_b32 s59, s21, 16
; GENERIC-NEXT: s_lshr_b32 s60, s20, 16
; GENERIC-NEXT: s_lshr_b32 s65, s29, 16
; GENERIC-NEXT: s_lshr_b32 s66, s28, 16
; GENERIC-NEXT: s_and_b32 s21, s21, 0xffff
; GENERIC-NEXT: s_and_b32 s20, s20, 0xffff
; GENERIC-NEXT: s_and_b32 s29, s29, 0xffff
; GENERIC-NEXT: s_and_b32 s28, s28, 0xffff
; GENERIC-NEXT: s_mov_b32 s0, s36
; GENERIC-NEXT: s_mov_b32 s1, s37
; GENERIC-NEXT: v_mov_b32_e32 v0, s30
; GENERIC-NEXT: v_mov_b32_e32 v1, s68
; GENERIC-NEXT: v_mov_b32_e32 v2, s31
; GENERIC-NEXT: v_mov_b32_e32 v3, s67
; GENERIC-NEXT: v_mov_b32_e32 v15, s3
; GENERIC-NEXT: v_mov_b32_e32 v17, s2
; GENERIC-NEXT: s_mov_b32 s3, 0xf000
; GENERIC-NEXT: s_mov_b32 s2, -1
; GENERIC-NEXT: s_lshr_b32 s57, s19, 16
; GENERIC-NEXT: s_lshr_b32 s58, s18, 16
; GENERIC-NEXT: s_lshr_b32 s61, s23, 16
; GENERIC-NEXT: s_lshr_b32 s62, s24, 16
; GENERIC-NEXT: s_lshr_b32 s63, s27, 16
; GENERIC-NEXT: s_lshr_b32 s64, s26, 16
; GENERIC-NEXT: s_and_b32 s19, s19, 0xffff
; GENERIC-NEXT: s_and_b32 s18, s18, 0xffff
; GENERIC-NEXT: s_and_b32 s23, s23, 0xffff
; GENERIC-NEXT: s_and_b32 s22, s22, 0xffff
; GENERIC-NEXT: s_and_b32 s25, s25, 0xffff
; GENERIC-NEXT: s_and_b32 s24, s24, 0xffff
; GENERIC-NEXT: s_and_b32 s27, s27, 0xffff
; GENERIC-NEXT: s_and_b32 s26, s26, 0xffff
; GENERIC-NEXT: v_mov_b32_e32 v4, s28
; GENERIC-NEXT: v_mov_b32_e32 v5, s66
; GENERIC-NEXT: v_mov_b32_e32 v6, s29
; GENERIC-NEXT: v_mov_b32_e32 v7, s65
; GENERIC-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:240
; GENERIC-NEXT: s_waitcnt expcnt(0)
; GENERIC-NEXT: v_mov_b32_e32 v0, s20
; GENERIC-NEXT: v_mov_b32_e32 v1, s60
; GENERIC-NEXT: v_mov_b32_e32 v2, s21
; GENERIC-NEXT: v_mov_b32_e32 v3, s59
; GENERIC-NEXT: s_lshr_b32 s38, s17, 16
; GENERIC-NEXT: s_lshr_b32 s39, s16, 16
; GENERIC-NEXT: s_and_b32 s17, s17, 0xffff
; GENERIC-NEXT: s_and_b32 s16, s16, 0xffff
; GENERIC-NEXT: v_mov_b32_e32 v8, s26
; GENERIC-NEXT: v_mov_b32_e32 v9, s64
; GENERIC-NEXT: v_mov_b32_e32 v10, s27
; GENERIC-NEXT: v_mov_b32_e32 v11, s63
; GENERIC-NEXT: v_mov_b32_e32 v12, s24
; GENERIC-NEXT: v_mov_b32_e32 v13, s62
; GENERIC-NEXT: v_mov_b32_e32 v14, s25
; GENERIC-NEXT: v_mov_b32_e32 v16, s22
; GENERIC-NEXT: v_mov_b32_e32 v18, s23
; GENERIC-NEXT: v_mov_b32_e32 v19, s61
; GENERIC-NEXT: buffer_store_dwordx4 v[4:7], off, s[0:3], 0 offset:224
; GENERIC-NEXT: buffer_store_dwordx4 v[8:11], off, s[0:3], 0 offset:208
; GENERIC-NEXT: buffer_store_dwordx4 v[12:15], off, s[0:3], 0 offset:192
; GENERIC-NEXT: buffer_store_dwordx4 v[16:19], off, s[0:3], 0 offset:176
; GENERIC-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:160
; GENERIC-NEXT: s_waitcnt expcnt(0)
; GENERIC-NEXT: v_mov_b32_e32 v0, s18
; GENERIC-NEXT: v_mov_b32_e32 v1, s58
; GENERIC-NEXT: v_mov_b32_e32 v2, s19
; GENERIC-NEXT: v_mov_b32_e32 v3, s57
; GENERIC-NEXT: s_lshr_b32 s53, s15, 16
; GENERIC-NEXT: s_lshr_b32 s54, s14, 16
; GENERIC-NEXT: s_and_b32 s15, s15, 0xffff
; GENERIC-NEXT: s_and_b32 s14, s14, 0xffff
; GENERIC-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:144
; GENERIC-NEXT: s_waitcnt expcnt(0)
; GENERIC-NEXT: v_mov_b32_e32 v0, s16
; GENERIC-NEXT: v_mov_b32_e32 v1, s39
; GENERIC-NEXT: v_mov_b32_e32 v2, s17
; GENERIC-NEXT: v_mov_b32_e32 v3, s38
; GENERIC-NEXT: s_lshr_b32 s51, s13, 16
; GENERIC-NEXT: s_lshr_b32 s52, s12, 16
; GENERIC-NEXT: s_and_b32 s13, s13, 0xffff
; GENERIC-NEXT: s_and_b32 s12, s12, 0xffff
; GENERIC-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:128
; GENERIC-NEXT: s_waitcnt expcnt(0)
; GENERIC-NEXT: v_mov_b32_e32 v0, s14
; GENERIC-NEXT: v_mov_b32_e32 v1, s54
; GENERIC-NEXT: v_mov_b32_e32 v2, s15
; GENERIC-NEXT: v_mov_b32_e32 v3, s53
; GENERIC-NEXT: s_lshr_b32 s49, s11, 16
; GENERIC-NEXT: s_lshr_b32 s50, s10, 16
; GENERIC-NEXT: s_and_b32 s11, s11, 0xffff
; GENERIC-NEXT: s_and_b32 s10, s10, 0xffff
; GENERIC-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:112
; GENERIC-NEXT: s_waitcnt expcnt(0)
; GENERIC-NEXT: v_mov_b32_e32 v0, s12
; GENERIC-NEXT: v_mov_b32_e32 v1, s52
; GENERIC-NEXT: v_mov_b32_e32 v2, s13
; GENERIC-NEXT: v_mov_b32_e32 v3, s51
; GENERIC-NEXT: s_lshr_b32 s47, s9, 16
; GENERIC-NEXT: s_lshr_b32 s48, s8, 16
; GENERIC-NEXT: s_and_b32 s9, s9, 0xffff
; GENERIC-NEXT: s_and_b32 s8, s8, 0xffff
; GENERIC-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:96
; GENERIC-NEXT: s_waitcnt expcnt(0)
; GENERIC-NEXT: v_mov_b32_e32 v0, s10
; GENERIC-NEXT: v_mov_b32_e32 v1, s50
; GENERIC-NEXT: v_mov_b32_e32 v2, s11
; GENERIC-NEXT: v_mov_b32_e32 v3, s49
; GENERIC-NEXT: s_lshr_b32 s43, s7, 16
; GENERIC-NEXT: s_lshr_b32 s45, s6, 16
; GENERIC-NEXT: s_and_b32 s7, s7, 0xffff
; GENERIC-NEXT: s_and_b32 s6, s6, 0xffff
; GENERIC-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:80
; GENERIC-NEXT: s_waitcnt expcnt(0)
; GENERIC-NEXT: v_mov_b32_e32 v0, s8
; GENERIC-NEXT: v_mov_b32_e32 v1, s48
; GENERIC-NEXT: v_mov_b32_e32 v2, s9
; GENERIC-NEXT: v_mov_b32_e32 v3, s47
; GENERIC-NEXT: s_lshr_b32 s41, s5, 16
; GENERIC-NEXT: s_lshr_b32 s42, s4, 16
; GENERIC-NEXT: s_and_b32 s5, s5, 0xffff
; GENERIC-NEXT: s_and_b32 s4, s4, 0xffff
; GENERIC-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:64
; GENERIC-NEXT: s_waitcnt expcnt(0)
; GENERIC-NEXT: v_mov_b32_e32 v0, s6
; GENERIC-NEXT: v_mov_b32_e32 v1, s45
; GENERIC-NEXT: v_mov_b32_e32 v2, s7
; GENERIC-NEXT: v_mov_b32_e32 v3, s43
; GENERIC-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:48
; GENERIC-NEXT: s_waitcnt expcnt(0)
; GENERIC-NEXT: v_mov_b32_e32 v0, s4
; GENERIC-NEXT: v_mov_b32_e32 v1, s42
; GENERIC-NEXT: v_mov_b32_e32 v2, s5
; GENERIC-NEXT: v_mov_b32_e32 v3, s41
; GENERIC-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:32
; GENERIC-NEXT: s_waitcnt expcnt(0)
; GENERIC-NEXT: v_mov_b32_e32 v0, s56
; GENERIC-NEXT: v_mov_b32_e32 v1, s40
; GENERIC-NEXT: v_mov_b32_e32 v2, s55
; GENERIC-NEXT: v_mov_b32_e32 v3, s35
; GENERIC-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:16
; GENERIC-NEXT: s_waitcnt expcnt(0)
; GENERIC-NEXT: v_mov_b32_e32 v0, s46
; GENERIC-NEXT: v_mov_b32_e32 v1, s34
; GENERIC-NEXT: v_mov_b32_e32 v2, s44
; GENERIC-NEXT: v_mov_b32_e32 v3, s33
; GENERIC-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0
; GENERIC-NEXT: s_endpgm
;
; GENERIC-GCNTRACKERS-LABEL: constant_zextload_v64i16_to_v64i32:
; GENERIC-GCNTRACKERS: ; %bb.0:
; GENERIC-GCNTRACKERS-NEXT: s_load_dwordx4 s[16:19], s[4:5], 0x9
; GENERIC-GCNTRACKERS-NEXT: s_mov_b32 s39, 0xf000
; GENERIC-GCNTRACKERS-NEXT: s_mov_b32 s38, -1
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt lgkmcnt(0)
; GENERIC-GCNTRACKERS-NEXT: s_load_dwordx16 s[0:15], s[18:19], 0x0
; GENERIC-GCNTRACKERS-NEXT: s_mov_b32 s36, s16
; GENERIC-GCNTRACKERS-NEXT: s_mov_b32 s37, s17
; GENERIC-GCNTRACKERS-NEXT: s_load_dwordx16 s[16:31], s[18:19], 0x10
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt lgkmcnt(0)
; GENERIC-GCNTRACKERS-NEXT: s_lshr_b32 s43, s7, 16
; GENERIC-GCNTRACKERS-NEXT: s_and_b32 s7, s7, 0xffff
; GENERIC-GCNTRACKERS-NEXT: s_lshr_b32 s41, s5, 16
; GENERIC-GCNTRACKERS-NEXT: s_and_b32 s44, s30, 0xffff
; GENERIC-GCNTRACKERS-NEXT: s_lshr_b32 s30, s30, 16
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v1, s30
; GENERIC-GCNTRACKERS-NEXT: s_and_b32 s30, s31, 0xffff
; GENERIC-GCNTRACKERS-NEXT: s_lshr_b32 s31, s31, 16
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v3, s31
; GENERIC-GCNTRACKERS-NEXT: s_and_b32 s31, s28, 0xffff
; GENERIC-GCNTRACKERS-NEXT: s_lshr_b32 s28, s28, 16
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v5, s28
; GENERIC-GCNTRACKERS-NEXT: s_and_b32 s28, s29, 0xffff
; GENERIC-GCNTRACKERS-NEXT: s_lshr_b32 s29, s29, 16
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v7, s29
; GENERIC-GCNTRACKERS-NEXT: s_and_b32 s29, s26, 0xffff
; GENERIC-GCNTRACKERS-NEXT: s_lshr_b32 s26, s26, 16
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v9, s26
; GENERIC-GCNTRACKERS-NEXT: s_and_b32 s26, s27, 0xffff
; GENERIC-GCNTRACKERS-NEXT: s_lshr_b32 s27, s27, 16
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v11, s27
; GENERIC-GCNTRACKERS-NEXT: s_and_b32 s27, s24, 0xffff
; GENERIC-GCNTRACKERS-NEXT: s_lshr_b32 s24, s24, 16
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v13, s24
; GENERIC-GCNTRACKERS-NEXT: s_and_b32 s24, s25, 0xffff
; GENERIC-GCNTRACKERS-NEXT: s_lshr_b32 s25, s25, 16
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v15, s25
; GENERIC-GCNTRACKERS-NEXT: s_and_b32 s25, s22, 0xffff
; GENERIC-GCNTRACKERS-NEXT: s_lshr_b32 s22, s22, 16
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v17, s22
; GENERIC-GCNTRACKERS-NEXT: s_and_b32 s22, s23, 0xffff
; GENERIC-GCNTRACKERS-NEXT: s_lshr_b32 s23, s23, 16
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v0, s44
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v2, s30
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v19, s23
; GENERIC-GCNTRACKERS-NEXT: s_and_b32 s23, s20, 0xffff
; GENERIC-GCNTRACKERS-NEXT: s_lshr_b32 s20, s20, 16
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dwordx4 v[0:3], off, s[36:39], 0 offset:240
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt expcnt(0)
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v1, s20
; GENERIC-GCNTRACKERS-NEXT: s_and_b32 s20, s21, 0xffff
; GENERIC-GCNTRACKERS-NEXT: s_lshr_b32 s21, s21, 16
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v4, s31
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v6, s28
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v0, s23
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v2, s20
; GENERIC-GCNTRACKERS-NEXT: s_lshr_b32 s20, s19, 16
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v3, s21
; GENERIC-GCNTRACKERS-NEXT: s_lshr_b32 s21, s18, 16
; GENERIC-GCNTRACKERS-NEXT: s_and_b32 s19, s19, 0xffff
; GENERIC-GCNTRACKERS-NEXT: s_and_b32 s18, s18, 0xffff
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v8, s29
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v10, s26
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v12, s27
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v14, s24
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v16, s25
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v18, s22
; GENERIC-GCNTRACKERS-NEXT: s_lshr_b32 s22, s17, 16
; GENERIC-GCNTRACKERS-NEXT: s_lshr_b32 s23, s16, 16
; GENERIC-GCNTRACKERS-NEXT: s_and_b32 s17, s17, 0xffff
; GENERIC-GCNTRACKERS-NEXT: s_and_b32 s16, s16, 0xffff
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dwordx4 v[4:7], off, s[36:39], 0 offset:224
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dwordx4 v[8:11], off, s[36:39], 0 offset:208
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dwordx4 v[12:15], off, s[36:39], 0 offset:192
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dwordx4 v[16:19], off, s[36:39], 0 offset:176
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dwordx4 v[0:3], off, s[36:39], 0 offset:160
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt expcnt(0)
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v0, s18
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v1, s21
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v2, s19
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v3, s20
; GENERIC-GCNTRACKERS-NEXT: s_lshr_b32 s24, s15, 16
; GENERIC-GCNTRACKERS-NEXT: s_lshr_b32 s25, s14, 16
; GENERIC-GCNTRACKERS-NEXT: s_and_b32 s15, s15, 0xffff
; GENERIC-GCNTRACKERS-NEXT: s_and_b32 s14, s14, 0xffff
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dwordx4 v[0:3], off, s[36:39], 0 offset:144
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt expcnt(0)
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v0, s16
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v1, s23
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v2, s17
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v3, s22
; GENERIC-GCNTRACKERS-NEXT: s_lshr_b32 s26, s13, 16
; GENERIC-GCNTRACKERS-NEXT: s_lshr_b32 s27, s12, 16
; GENERIC-GCNTRACKERS-NEXT: s_and_b32 s13, s13, 0xffff
; GENERIC-GCNTRACKERS-NEXT: s_and_b32 s12, s12, 0xffff
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dwordx4 v[0:3], off, s[36:39], 0 offset:128
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt expcnt(0)
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v0, s14
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v1, s25
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v2, s15
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v3, s24
; GENERIC-GCNTRACKERS-NEXT: s_lshr_b32 s28, s11, 16
; GENERIC-GCNTRACKERS-NEXT: s_lshr_b32 s29, s10, 16
; GENERIC-GCNTRACKERS-NEXT: s_and_b32 s11, s11, 0xffff
; GENERIC-GCNTRACKERS-NEXT: s_and_b32 s10, s10, 0xffff
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dwordx4 v[0:3], off, s[36:39], 0 offset:112
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt expcnt(0)
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v0, s12
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v1, s27
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v2, s13
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v3, s26
; GENERIC-GCNTRACKERS-NEXT: s_lshr_b32 s30, s9, 16
; GENERIC-GCNTRACKERS-NEXT: s_lshr_b32 s31, s8, 16
; GENERIC-GCNTRACKERS-NEXT: s_and_b32 s9, s9, 0xffff
; GENERIC-GCNTRACKERS-NEXT: s_and_b32 s8, s8, 0xffff
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dwordx4 v[0:3], off, s[36:39], 0 offset:96
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt expcnt(0)
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v0, s10
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v1, s29
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v2, s11
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v3, s28
; GENERIC-GCNTRACKERS-NEXT: s_lshr_b32 s44, s6, 16
; GENERIC-GCNTRACKERS-NEXT: s_and_b32 s6, s6, 0xffff
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dwordx4 v[0:3], off, s[36:39], 0 offset:80
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt expcnt(0)
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v0, s8
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v1, s31
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v2, s9
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v3, s30
; GENERIC-GCNTRACKERS-NEXT: s_lshr_b32 s42, s4, 16
; GENERIC-GCNTRACKERS-NEXT: s_and_b32 s5, s5, 0xffff
; GENERIC-GCNTRACKERS-NEXT: s_and_b32 s4, s4, 0xffff
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dwordx4 v[0:3], off, s[36:39], 0 offset:64
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt expcnt(0)
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v0, s6
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v1, s44
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v2, s7
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v3, s43
; GENERIC-GCNTRACKERS-NEXT: s_lshr_b32 s35, s3, 16
; GENERIC-GCNTRACKERS-NEXT: s_lshr_b32 s40, s2, 16
; GENERIC-GCNTRACKERS-NEXT: s_and_b32 s3, s3, 0xffff
; GENERIC-GCNTRACKERS-NEXT: s_and_b32 s2, s2, 0xffff
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dwordx4 v[0:3], off, s[36:39], 0 offset:48
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt expcnt(0)
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v0, s4
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v1, s42
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v2, s5
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v3, s41
; GENERIC-GCNTRACKERS-NEXT: s_lshr_b32 s33, s1, 16
; GENERIC-GCNTRACKERS-NEXT: s_lshr_b32 s34, s0, 16
; GENERIC-GCNTRACKERS-NEXT: s_and_b32 s1, s1, 0xffff
; GENERIC-GCNTRACKERS-NEXT: s_and_b32 s0, s0, 0xffff
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dwordx4 v[0:3], off, s[36:39], 0 offset:32
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt expcnt(0)
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v0, s2
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v1, s40
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v2, s3
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v3, s35
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dwordx4 v[0:3], off, s[36:39], 0 offset:16
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt expcnt(0)
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v0, s0
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v1, s34
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v2, s1
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v3, s33
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dwordx4 v[0:3], off, s[36:39], 0
; GENERIC-GCNTRACKERS-NEXT: s_endpgm
%load = load <64 x i16>, ptr addrspace(4) %in
%ext = zext <64 x i16> %load to <64 x i32>
store <64 x i32> %ext, ptr addrspace(1) %out
ret void
}
; CHECK-LABEL: {{^}}excess_soft_clause_reg_pressure:
; GFX908: NumSgprs: 64
; GFX908-GCNTRACKERS: NumSgprs: 64
; GFX908: NumVgprs: 41
; GFX908-GCNTRACKERS: NumVgprs: 39
; GFX908: Occupancy: 5
; GFX908-GCNTRACKERS: Occupancy: 6
define protected amdgpu_kernel void @excess_soft_clause_reg_pressure(ptr addrspace(4) %wei_ptr, ptr addrspace(1) %out_ptr, ptr addrspace(1) %in) {
; GFX11-PAL-LABEL: excess_soft_clause_reg_pressure:
; GFX11-PAL: ; %bb.0: ; %entry
; GFX11-PAL-NEXT: s_load_b256 s[0:7], s[4:5], 0x0
; GFX11-PAL-NEXT: v_dual_mov_b32 v11, 0 :: v_dual_and_b32 v34, 0x3ff, v0
; GFX11-PAL-NEXT: v_dual_mov_b32 v1, 0 :: v_dual_mov_b32 v6, 0
; GFX11-PAL-NEXT: s_delay_alu instid0(VALU_DEP_2)
; GFX11-PAL-NEXT: v_dual_mov_b32 v21, 0 :: v_dual_lshlrev_b32 v2, 2, v34
; GFX11-PAL-NEXT: v_dual_mov_b32 v7, 0 :: v_dual_mov_b32 v8, 0
; GFX11-PAL-NEXT: v_dual_mov_b32 v9, 0 :: v_dual_mov_b32 v10, 0
; GFX11-PAL-NEXT: v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v17, 0
; GFX11-PAL-NEXT: v_dual_mov_b32 v16, 0 :: v_dual_mov_b32 v19, 0
; GFX11-PAL-NEXT: v_dual_mov_b32 v18, 0 :: v_dual_mov_b32 v23, 0
; GFX11-PAL-NEXT: v_dual_mov_b32 v20, 0 :: v_dual_mov_b32 v25, 0
; GFX11-PAL-NEXT: v_dual_mov_b32 v22, 0 :: v_dual_mov_b32 v27, 0
; GFX11-PAL-NEXT: s_waitcnt lgkmcnt(0)
; GFX11-PAL-NEXT: v_add_co_u32 v2, s4, s4, v2
; GFX11-PAL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-PAL-NEXT: v_add_co_ci_u32_e64 v3, null, s5, 0, s4
; GFX11-PAL-NEXT: v_dual_mov_b32 v24, 0 :: v_dual_mov_b32 v29, 0
; GFX11-PAL-NEXT: v_add_co_u32 v4, vcc_lo, 0x188, v2
; GFX11-PAL-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-PAL-NEXT: v_add_co_ci_u32_e64 v5, null, 0, v3, vcc_lo
; GFX11-PAL-NEXT: v_dual_mov_b32 v26, 0 :: v_dual_mov_b32 v3, 0
; GFX11-PAL-NEXT: v_dual_mov_b32 v28, 0 :: v_dual_mov_b32 v31, 0
; GFX11-PAL-NEXT: v_dual_mov_b32 v2, 0 :: v_dual_mov_b32 v33, 0
; GFX11-PAL-NEXT: v_dual_mov_b32 v30, 0 :: v_dual_mov_b32 v13, 0
; GFX11-PAL-NEXT: v_dual_mov_b32 v32, 0 :: v_dual_mov_b32 v15, 0
; GFX11-PAL-NEXT: v_mov_b32_e32 v12, 0
; GFX11-PAL-NEXT: v_mov_b32_e32 v14, 0
; GFX11-PAL-NEXT: s_movk_i32 s4, 0x200
; GFX11-PAL-NEXT: .LBB4_1: ; %for.cond28.preheader
; GFX11-PAL-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-PAL-NEXT: s_clause 0x3
; GFX11-PAL-NEXT: global_load_b32 v35, v[4:5], off offset:-392
; GFX11-PAL-NEXT: global_load_b32 v36, v[4:5], off offset:-196
; GFX11-PAL-NEXT: global_load_b32 v37, v[4:5], off
; GFX11-PAL-NEXT: global_load_b32 v38, v[4:5], off offset:196
; GFX11-PAL-NEXT: s_clause 0x4
; GFX11-PAL-NEXT: s_load_b512 s[84:99], s[0:1], 0x0
; GFX11-PAL-NEXT: s_load_b512 s[52:67], s[0:1], 0x1000
; GFX11-PAL-NEXT: s_load_b512 s[68:83], s[0:1], 0x40
; GFX11-PAL-NEXT: s_load_b512 s[36:51], s[0:1], 0x2000
; GFX11-PAL-NEXT: s_load_b512 s[16:31], s[0:1], 0x3000
; GFX11-PAL-NEXT: v_add_co_u32 v4, vcc_lo, 0x310, v4
; GFX11-PAL-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-PAL-NEXT: v_add_co_ci_u32_e64 v5, null, 0, v5, vcc_lo
; GFX11-PAL-NEXT: s_waitcnt vmcnt(3) lgkmcnt(0)
; GFX11-PAL-NEXT: v_fmac_f32_e32 v32, s85, v35
; GFX11-PAL-NEXT: v_fmac_f32_e32 v33, s84, v35
; GFX11-PAL-NEXT: v_fmac_f32_e32 v30, s87, v35
; GFX11-PAL-NEXT: v_fmac_f32_e32 v31, s86, v35
; GFX11-PAL-NEXT: v_fmac_f32_e32 v3, s89, v35
; GFX11-PAL-NEXT: v_fmac_f32_e32 v2, s88, v35
; GFX11-PAL-NEXT: v_fmac_f32_e32 v28, s92, v35
; GFX11-PAL-NEXT: v_fmac_f32_e32 v1, s90, v35
; GFX11-PAL-NEXT: v_fmac_f32_e32 v26, s94, v35
; GFX11-PAL-NEXT: v_fmac_f32_e32 v29, s91, v35
; GFX11-PAL-NEXT: v_fmac_f32_e32 v24, s96, v35
; GFX11-PAL-NEXT: v_fmac_f32_e32 v27, s93, v35
; GFX11-PAL-NEXT: v_fmac_f32_e32 v22, s98, v35
; GFX11-PAL-NEXT: v_fmac_f32_e32 v25, s95, v35
; GFX11-PAL-NEXT: v_fmac_f32_e32 v20, s68, v35
; GFX11-PAL-NEXT: v_fmac_f32_e32 v23, s97, v35
; GFX11-PAL-NEXT: v_fmac_f32_e32 v18, s70, v35
; GFX11-PAL-NEXT: v_fmac_f32_e32 v21, s99, v35
; GFX11-PAL-NEXT: s_load_b512 s[84:99], s[0:1], 0x1040
; GFX11-PAL-NEXT: v_fmac_f32_e32 v16, s72, v35
; GFX11-PAL-NEXT: v_fmac_f32_e32 v19, s69, v35
; GFX11-PAL-NEXT: v_fmac_f32_e32 v0, s74, v35
; GFX11-PAL-NEXT: v_fmac_f32_e32 v17, s71, v35
; GFX11-PAL-NEXT: v_fmac_f32_e32 v10, s75, v35
; GFX11-PAL-NEXT: v_fmac_f32_e32 v11, s73, v35
; GFX11-PAL-NEXT: v_fmac_f32_e32 v8, s77, v35
; GFX11-PAL-NEXT: v_fmac_f32_e32 v9, s76, v35
; GFX11-PAL-NEXT: v_fmac_f32_e32 v6, s79, v35
; GFX11-PAL-NEXT: v_fmac_f32_e32 v7, s78, v35
; GFX11-PAL-NEXT: v_fmac_f32_e32 v13, s81, v35
; GFX11-PAL-NEXT: v_fmac_f32_e32 v12, s80, v35
; GFX11-PAL-NEXT: v_fmac_f32_e32 v15, s83, v35
; GFX11-PAL-NEXT: v_fmac_f32_e32 v14, s82, v35
; GFX11-PAL-NEXT: s_load_b512 s[68:83], s[0:1], 0x2040
; GFX11-PAL-NEXT: s_waitcnt vmcnt(2)
; GFX11-PAL-NEXT: v_fmac_f32_e32 v32, s53, v36
; GFX11-PAL-NEXT: v_fmac_f32_e32 v33, s52, v36
; GFX11-PAL-NEXT: v_fmac_f32_e32 v30, s55, v36
; GFX11-PAL-NEXT: v_fmac_f32_e32 v31, s54, v36
; GFX11-PAL-NEXT: v_fmac_f32_e32 v3, s57, v36
; GFX11-PAL-NEXT: v_fmac_f32_e32 v2, s56, v36
; GFX11-PAL-NEXT: v_fmac_f32_e32 v28, s60, v36
; GFX11-PAL-NEXT: v_fmac_f32_e32 v1, s58, v36
; GFX11-PAL-NEXT: v_fmac_f32_e32 v26, s62, v36
; GFX11-PAL-NEXT: v_fmac_f32_e32 v29, s59, v36
; GFX11-PAL-NEXT: v_fmac_f32_e32 v24, s64, v36
; GFX11-PAL-NEXT: v_fmac_f32_e32 v27, s61, v36
; GFX11-PAL-NEXT: v_fmac_f32_e32 v22, s66, v36
; GFX11-PAL-NEXT: v_fmac_f32_e32 v25, s63, v36
; GFX11-PAL-NEXT: s_waitcnt lgkmcnt(0)
; GFX11-PAL-NEXT: v_fmac_f32_e32 v20, s84, v36
; GFX11-PAL-NEXT: v_fmac_f32_e32 v23, s65, v36
; GFX11-PAL-NEXT: v_fmac_f32_e32 v18, s86, v36
; GFX11-PAL-NEXT: v_fmac_f32_e32 v21, s67, v36
; GFX11-PAL-NEXT: s_load_b512 s[52:67], s[0:1], 0x3040
; GFX11-PAL-NEXT: v_fmac_f32_e32 v16, s88, v36
; GFX11-PAL-NEXT: v_fmac_f32_e32 v19, s85, v36
; GFX11-PAL-NEXT: v_fmac_f32_e32 v0, s90, v36
; GFX11-PAL-NEXT: v_fmac_f32_e32 v17, s87, v36
; GFX11-PAL-NEXT: v_fmac_f32_e32 v10, s91, v36
; GFX11-PAL-NEXT: v_fmac_f32_e32 v11, s89, v36
; GFX11-PAL-NEXT: v_fmac_f32_e32 v8, s93, v36
; GFX11-PAL-NEXT: v_fmac_f32_e32 v9, s92, v36
; GFX11-PAL-NEXT: v_fmac_f32_e32 v6, s95, v36
; GFX11-PAL-NEXT: v_fmac_f32_e32 v7, s94, v36
; GFX11-PAL-NEXT: v_fmac_f32_e32 v13, s97, v36
; GFX11-PAL-NEXT: v_fmac_f32_e32 v12, s96, v36
; GFX11-PAL-NEXT: v_fmac_f32_e32 v15, s99, v36
; GFX11-PAL-NEXT: v_fmac_f32_e32 v14, s98, v36
; GFX11-PAL-NEXT: s_waitcnt vmcnt(1)
; GFX11-PAL-NEXT: v_fmac_f32_e32 v32, s37, v37
; GFX11-PAL-NEXT: v_fmac_f32_e32 v33, s36, v37
; GFX11-PAL-NEXT: v_fmac_f32_e32 v30, s39, v37
; GFX11-PAL-NEXT: v_fmac_f32_e32 v31, s38, v37
; GFX11-PAL-NEXT: v_fmac_f32_e32 v3, s41, v37
; GFX11-PAL-NEXT: v_fmac_f32_e32 v2, s40, v37
; GFX11-PAL-NEXT: v_fmac_f32_e32 v28, s44, v37
; GFX11-PAL-NEXT: v_fmac_f32_e32 v1, s42, v37
; GFX11-PAL-NEXT: v_fmac_f32_e32 v26, s46, v37
; GFX11-PAL-NEXT: v_fmac_f32_e32 v29, s43, v37
; GFX11-PAL-NEXT: v_fmac_f32_e32 v24, s48, v37
; GFX11-PAL-NEXT: v_fmac_f32_e32 v27, s45, v37
; GFX11-PAL-NEXT: v_fmac_f32_e32 v22, s50, v37
; GFX11-PAL-NEXT: v_fmac_f32_e32 v25, s47, v37
; GFX11-PAL-NEXT: v_fmac_f32_e32 v20, s68, v37
; GFX11-PAL-NEXT: v_fmac_f32_e32 v23, s49, v37
; GFX11-PAL-NEXT: v_fmac_f32_e32 v18, s70, v37
; GFX11-PAL-NEXT: v_fmac_f32_e32 v21, s51, v37
; GFX11-PAL-NEXT: v_fmac_f32_e32 v16, s72, v37
; GFX11-PAL-NEXT: v_fmac_f32_e32 v19, s69, v37
; GFX11-PAL-NEXT: v_fmac_f32_e32 v0, s74, v37
; GFX11-PAL-NEXT: v_fmac_f32_e32 v17, s71, v37
; GFX11-PAL-NEXT: v_fmac_f32_e32 v10, s75, v37
; GFX11-PAL-NEXT: v_fmac_f32_e32 v11, s73, v37
; GFX11-PAL-NEXT: v_fmac_f32_e32 v8, s77, v37
; GFX11-PAL-NEXT: v_fmac_f32_e32 v9, s76, v37
; GFX11-PAL-NEXT: v_fmac_f32_e32 v6, s79, v37
; GFX11-PAL-NEXT: v_fmac_f32_e32 v7, s78, v37
; GFX11-PAL-NEXT: v_fmac_f32_e32 v13, s81, v37
; GFX11-PAL-NEXT: v_fmac_f32_e32 v12, s80, v37
; GFX11-PAL-NEXT: v_fmac_f32_e32 v15, s83, v37
; GFX11-PAL-NEXT: v_fmac_f32_e32 v14, s82, v37
; GFX11-PAL-NEXT: s_add_u32 s0, s0, 0x4000
; GFX11-PAL-NEXT: s_waitcnt vmcnt(0)
; GFX11-PAL-NEXT: v_fmac_f32_e32 v32, s17, v38
; GFX11-PAL-NEXT: v_fmac_f32_e32 v33, s16, v38
; GFX11-PAL-NEXT: v_fmac_f32_e32 v30, s19, v38
; GFX11-PAL-NEXT: v_fmac_f32_e32 v31, s18, v38
; GFX11-PAL-NEXT: v_fmac_f32_e32 v3, s21, v38
; GFX11-PAL-NEXT: v_fmac_f32_e32 v2, s20, v38
; GFX11-PAL-NEXT: v_fmac_f32_e32 v28, s24, v38
; GFX11-PAL-NEXT: v_fmac_f32_e32 v1, s22, v38
; GFX11-PAL-NEXT: v_fmac_f32_e32 v26, s26, v38
; GFX11-PAL-NEXT: v_fmac_f32_e32 v29, s23, v38
; GFX11-PAL-NEXT: v_fmac_f32_e32 v24, s28, v38
; GFX11-PAL-NEXT: v_fmac_f32_e32 v27, s25, v38
; GFX11-PAL-NEXT: v_fmac_f32_e32 v22, s30, v38
; GFX11-PAL-NEXT: v_fmac_f32_e32 v25, s27, v38
; GFX11-PAL-NEXT: s_waitcnt lgkmcnt(0)
; GFX11-PAL-NEXT: v_fmac_f32_e32 v20, s52, v38
; GFX11-PAL-NEXT: v_fmac_f32_e32 v23, s29, v38
; GFX11-PAL-NEXT: v_fmac_f32_e32 v18, s54, v38
; GFX11-PAL-NEXT: v_fmac_f32_e32 v21, s31, v38
; GFX11-PAL-NEXT: v_fmac_f32_e32 v16, s56, v38
; GFX11-PAL-NEXT: v_fmac_f32_e32 v19, s53, v38
; GFX11-PAL-NEXT: v_fmac_f32_e32 v0, s58, v38
; GFX11-PAL-NEXT: v_fmac_f32_e32 v17, s55, v38
; GFX11-PAL-NEXT: v_fmac_f32_e32 v10, s59, v38
; GFX11-PAL-NEXT: v_fmac_f32_e32 v11, s57, v38
; GFX11-PAL-NEXT: v_fmac_f32_e32 v8, s61, v38
; GFX11-PAL-NEXT: v_fmac_f32_e32 v9, s60, v38
; GFX11-PAL-NEXT: v_fmac_f32_e32 v6, s63, v38
; GFX11-PAL-NEXT: v_fmac_f32_e32 v7, s62, v38
; GFX11-PAL-NEXT: v_fmac_f32_e32 v13, s65, v38
; GFX11-PAL-NEXT: v_fmac_f32_e32 v12, s64, v38
; GFX11-PAL-NEXT: v_fmac_f32_e32 v15, s67, v38
; GFX11-PAL-NEXT: v_fmac_f32_e32 v14, s66, v38
; GFX11-PAL-NEXT: s_addc_u32 s1, s1, 0
; GFX11-PAL-NEXT: s_add_i32 s4, s4, -1
; GFX11-PAL-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-PAL-NEXT: s_cmp_lg_u32 s4, 0
; GFX11-PAL-NEXT: s_cbranch_scc1 .LBB4_1
; GFX11-PAL-NEXT: ; %bb.2: ; %for.cond.cleanup26
; GFX11-PAL-NEXT: s_lshl_b32 s0, s13, 8
; GFX11-PAL-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_3) | instid1(SALU_CYCLE_1)
; GFX11-PAL-NEXT: v_add3_u32 v4, s6, s0, v34
; GFX11-PAL-NEXT: s_mov_b32 s0, 0
; GFX11-PAL-NEXT: v_mov_b32_e32 v5, 0
; GFX11-PAL-NEXT: s_mov_b32 s1, s0
; GFX11-PAL-NEXT: v_dual_mov_b32 v36, s0 :: v_dual_mov_b32 v37, s1
; GFX11-PAL-NEXT: v_mul_hi_u32 v4, 0x5397829d, v4
; GFX11-PAL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-PAL-NEXT: v_lshrrev_b32_e32 v4, 4, v4
; GFX11-PAL-NEXT: v_mul_lo_u32 v4, 0x31000, v4
; GFX11-PAL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-PAL-NEXT: v_lshlrev_b64 v[34:35], 2, v[4:5]
; GFX11-PAL-NEXT: v_mov_b32_e32 v4, s0
; GFX11-PAL-NEXT: v_add_co_u32 v34, vcc_lo, s2, v34
; GFX11-PAL-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-PAL-NEXT: v_add_co_ci_u32_e64 v35, null, s3, v35, vcc_lo
; GFX11-PAL-NEXT: s_clause 0x5
; GFX11-PAL-NEXT: global_store_b32 v[34:35], v33, off
; GFX11-PAL-NEXT: global_store_b32 v[34:35], v32, off offset:784
; GFX11-PAL-NEXT: global_store_b64 v[34:35], v[36:37], off offset:840
; GFX11-PAL-NEXT: global_store_b32 v[34:35], v31, off offset:1568
; GFX11-PAL-NEXT: global_store_b32 v[34:35], v30, off offset:2352
; GFX11-PAL-NEXT: global_store_b32 v[34:35], v2, off offset:3136
; GFX11-PAL-NEXT: v_add_co_u32 v38, vcc_lo, 0x310, v34
; GFX11-PAL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-PAL-NEXT: v_add_co_ci_u32_e64 v39, null, 0, v35, vcc_lo
; GFX11-PAL-NEXT: v_add_co_u32 v40, vcc_lo, 0x620, v34
; GFX11-PAL-NEXT: v_add_co_ci_u32_e64 v41, null, 0, v35, vcc_lo
; GFX11-PAL-NEXT: v_add_co_u32 v42, vcc_lo, 0x930, v34
; GFX11-PAL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-PAL-NEXT: v_add_co_ci_u32_e64 v43, null, 0, v35, vcc_lo
; GFX11-PAL-NEXT: v_add_co_u32 v44, vcc_lo, 0xc40, v34
; GFX11-PAL-NEXT: v_add_co_ci_u32_e64 v45, null, 0, v35, vcc_lo
; GFX11-PAL-NEXT: v_add_co_u32 v30, vcc_lo, 0xf50, v34
; GFX11-PAL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-PAL-NEXT: v_add_co_ci_u32_e64 v31, null, 0, v35, vcc_lo
; GFX11-PAL-NEXT: v_add_co_u32 v32, vcc_lo, 0xf50, v38
; GFX11-PAL-NEXT: v_add_co_ci_u32_e64 v33, null, 0, v39, vcc_lo
; GFX11-PAL-NEXT: v_mov_b32_e32 v2, s0
; GFX11-PAL-NEXT: v_add_co_u32 v46, vcc_lo, 0xf50, v40
; GFX11-PAL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-PAL-NEXT: v_add_co_ci_u32_e64 v47, null, 0, v41, vcc_lo
; GFX11-PAL-NEXT: v_add_co_u32 v48, vcc_lo, 0xf50, v42
; GFX11-PAL-NEXT: v_add_co_ci_u32_e64 v49, null, 0, v43, vcc_lo
; GFX11-PAL-NEXT: s_clause 0x3
; GFX11-PAL-NEXT: global_store_b64 v[34:35], v[3:4], off offset:3920
; GFX11-PAL-NEXT: global_store_b64 v[38:39], v[1:2], off offset:3920
; GFX11-PAL-NEXT: global_store_b32 v[40:41], v29, off offset:3920
; GFX11-PAL-NEXT: global_store_b32 v[42:43], v28, off offset:3920
; GFX11-PAL-NEXT: v_add_co_u32 v1, vcc_lo, 0xf50, v44
; GFX11-PAL-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-PAL-NEXT: v_add_co_ci_u32_e64 v2, null, 0, v45, vcc_lo
; GFX11-PAL-NEXT: v_add_co_u32 v3, vcc_lo, 0xf50, v30
; GFX11-PAL-NEXT: s_clause 0x1
; GFX11-PAL-NEXT: global_store_b64 v[42:43], v[36:37], off offset:3976
; GFX11-PAL-NEXT: global_store_b32 v[44:45], v27, off offset:3920
; GFX11-PAL-NEXT: v_add_co_ci_u32_e64 v4, null, 0, v31, vcc_lo
; GFX11-PAL-NEXT: global_store_b32 v[30:31], v26, off offset:3920
; GFX11-PAL-NEXT: v_add_co_u32 v26, vcc_lo, 0xf50, v32
; GFX11-PAL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-PAL-NEXT: v_add_co_ci_u32_e64 v27, null, 0, v33, vcc_lo
; GFX11-PAL-NEXT: v_add_co_u32 v28, vcc_lo, 0xf50, v46
; GFX11-PAL-NEXT: v_add_co_ci_u32_e64 v29, null, 0, v47, vcc_lo
; GFX11-PAL-NEXT: global_store_b32 v[46:47], v24, off offset:3920
; GFX11-PAL-NEXT: v_add_co_u32 v24, vcc_lo, 0xf50, v48
; GFX11-PAL-NEXT: global_store_b32 v[32:33], v25, off offset:3920
; GFX11-PAL-NEXT: v_add_co_ci_u32_e64 v25, null, 0, v49, vcc_lo
; GFX11-PAL-NEXT: v_add_co_u32 v30, vcc_lo, 0xf50, v1
; GFX11-PAL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1)
; GFX11-PAL-NEXT: v_add_co_ci_u32_e64 v31, null, 0, v2, vcc_lo
; GFX11-PAL-NEXT: global_store_b32 v[1:2], v22, off offset:3920
; GFX11-PAL-NEXT: v_add_co_u32 v1, vcc_lo, 0xf50, v3
; GFX11-PAL-NEXT: v_add_co_ci_u32_e64 v2, null, 0, v4, vcc_lo
; GFX11-PAL-NEXT: global_store_b32 v[3:4], v21, off offset:3920
; GFX11-PAL-NEXT: v_add_co_u32 v3, vcc_lo, 0xf50, v26
; GFX11-PAL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1)
; GFX11-PAL-NEXT: v_add_co_ci_u32_e64 v4, null, 0, v27, vcc_lo
; GFX11-PAL-NEXT: global_store_b32 v[26:27], v20, off offset:3920
; GFX11-PAL-NEXT: v_add_co_u32 v20, vcc_lo, 0xf50, v28
; GFX11-PAL-NEXT: v_add_co_ci_u32_e64 v21, null, 0, v29, vcc_lo
; GFX11-PAL-NEXT: v_add_co_u32 v22, vcc_lo, 0xf50, v24
; GFX11-PAL-NEXT: global_store_b32 v[48:49], v23, off offset:3920
; GFX11-PAL-NEXT: v_add_co_ci_u32_e64 v23, null, 0, v25, vcc_lo
; GFX11-PAL-NEXT: global_store_b32 v[24:25], v18, off offset:3920
; GFX11-PAL-NEXT: v_add_co_u32 v18, vcc_lo, 0xf50, v30
; GFX11-PAL-NEXT: global_store_b32 v[28:29], v19, off offset:3920
; GFX11-PAL-NEXT: v_add_co_ci_u32_e64 v19, null, 0, v31, vcc_lo
; GFX11-PAL-NEXT: v_add_co_u32 v24, vcc_lo, 0xf50, v1
; GFX11-PAL-NEXT: global_store_b32 v[30:31], v17, off offset:3920
; GFX11-PAL-NEXT: v_add_co_ci_u32_e64 v25, null, 0, v2, vcc_lo
; GFX11-PAL-NEXT: global_store_b32 v[1:2], v16, off offset:3920
; GFX11-PAL-NEXT: v_add_co_u32 v1, vcc_lo, 0xf50, v3
; GFX11-PAL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1)
; GFX11-PAL-NEXT: v_add_co_ci_u32_e64 v2, null, 0, v4, vcc_lo
; GFX11-PAL-NEXT: global_store_b32 v[3:4], v11, off offset:3920
; GFX11-PAL-NEXT: v_add_co_u32 v3, vcc_lo, 0xf50, v20
; GFX11-PAL-NEXT: v_add_co_ci_u32_e64 v4, null, 0, v21, vcc_lo
; GFX11-PAL-NEXT: v_add_co_u32 v16, vcc_lo, 0xf50, v22
; GFX11-PAL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1)
; GFX11-PAL-NEXT: v_add_co_ci_u32_e64 v17, null, 0, v23, vcc_lo
; GFX11-PAL-NEXT: global_store_b32 v[22:23], v10, off offset:3920
; GFX11-PAL-NEXT: v_add_co_u32 v10, vcc_lo, 0xf50, v18
; GFX11-PAL-NEXT: v_add_co_ci_u32_e64 v11, null, 0, v19, vcc_lo
; GFX11-PAL-NEXT: global_store_b32 v[18:19], v9, off offset:3920
; GFX11-PAL-NEXT: v_add_co_u32 v18, vcc_lo, 0xf50, v24
; GFX11-PAL-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-PAL-NEXT: v_add_co_ci_u32_e64 v19, null, 0, v25, vcc_lo
; GFX11-PAL-NEXT: global_store_b32 v[24:25], v8, off offset:3920
; GFX11-PAL-NEXT: v_add_co_u32 v8, vcc_lo, 0xf50, v1
; GFX11-PAL-NEXT: global_store_b32 v[20:21], v0, off offset:3920
; GFX11-PAL-NEXT: v_add_co_ci_u32_e64 v9, null, 0, v2, vcc_lo
; GFX11-PAL-NEXT: global_store_b32 v[1:2], v7, off offset:3920
; GFX11-PAL-NEXT: global_store_b32 v[3:4], v6, off offset:3920
; GFX11-PAL-NEXT: global_store_b32 v[16:17], v12, off offset:3920
; GFX11-PAL-NEXT: global_store_b32 v[10:11], v13, off offset:3920
; GFX11-PAL-NEXT: global_store_b32 v[18:19], v14, off offset:3920
; GFX11-PAL-NEXT: global_store_b32 v[8:9], v15, off offset:3920
; GFX11-PAL-NEXT: s_endpgm
;
; GFX11-PAL-GCNTRACKERS-LABEL: excess_soft_clause_reg_pressure:
; GFX11-PAL-GCNTRACKERS: ; %bb.0: ; %entry
; GFX11-PAL-GCNTRACKERS-NEXT: s_load_b256 s[16:23], s[4:5], 0x0
; GFX11-PAL-GCNTRACKERS-NEXT: v_dual_mov_b32 v11, 0 :: v_dual_and_b32 v34, 0x3ff, v0
; GFX11-PAL-GCNTRACKERS-NEXT: v_dual_mov_b32 v1, 0 :: v_dual_mov_b32 v6, 0
; GFX11-PAL-GCNTRACKERS-NEXT: s_delay_alu instid0(VALU_DEP_2)
; GFX11-PAL-GCNTRACKERS-NEXT: v_dual_mov_b32 v21, 0 :: v_dual_lshlrev_b32 v2, 2, v34
; GFX11-PAL-GCNTRACKERS-NEXT: v_dual_mov_b32 v7, 0 :: v_dual_mov_b32 v8, 0
; GFX11-PAL-GCNTRACKERS-NEXT: v_dual_mov_b32 v9, 0 :: v_dual_mov_b32 v10, 0
; GFX11-PAL-GCNTRACKERS-NEXT: v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v17, 0
; GFX11-PAL-GCNTRACKERS-NEXT: v_dual_mov_b32 v16, 0 :: v_dual_mov_b32 v19, 0
; GFX11-PAL-GCNTRACKERS-NEXT: v_dual_mov_b32 v18, 0 :: v_dual_mov_b32 v23, 0
; GFX11-PAL-GCNTRACKERS-NEXT: v_dual_mov_b32 v20, 0 :: v_dual_mov_b32 v25, 0
; GFX11-PAL-GCNTRACKERS-NEXT: v_dual_mov_b32 v22, 0 :: v_dual_mov_b32 v27, 0
; GFX11-PAL-GCNTRACKERS-NEXT: s_waitcnt lgkmcnt(0)
; GFX11-PAL-GCNTRACKERS-NEXT: v_add_co_u32 v2, s0, s20, v2
; GFX11-PAL-GCNTRACKERS-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-PAL-GCNTRACKERS-NEXT: v_add_co_ci_u32_e64 v3, null, s21, 0, s0
; GFX11-PAL-GCNTRACKERS-NEXT: v_dual_mov_b32 v24, 0 :: v_dual_mov_b32 v29, 0
; GFX11-PAL-GCNTRACKERS-NEXT: v_add_co_u32 v4, vcc_lo, 0x188, v2
; GFX11-PAL-GCNTRACKERS-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-PAL-GCNTRACKERS-NEXT: v_add_co_ci_u32_e64 v5, null, 0, v3, vcc_lo
; GFX11-PAL-GCNTRACKERS-NEXT: v_dual_mov_b32 v26, 0 :: v_dual_mov_b32 v3, 0
; GFX11-PAL-GCNTRACKERS-NEXT: v_dual_mov_b32 v28, 0 :: v_dual_mov_b32 v31, 0
; GFX11-PAL-GCNTRACKERS-NEXT: v_dual_mov_b32 v2, 0 :: v_dual_mov_b32 v33, 0
; GFX11-PAL-GCNTRACKERS-NEXT: v_dual_mov_b32 v30, 0 :: v_dual_mov_b32 v13, 0
; GFX11-PAL-GCNTRACKERS-NEXT: v_dual_mov_b32 v32, 0 :: v_dual_mov_b32 v15, 0
; GFX11-PAL-GCNTRACKERS-NEXT: v_mov_b32_e32 v12, 0
; GFX11-PAL-GCNTRACKERS-NEXT: v_mov_b32_e32 v14, 0
; GFX11-PAL-GCNTRACKERS-NEXT: s_mov_b32 s33, s13
; GFX11-PAL-GCNTRACKERS-NEXT: s_movk_i32 s34, 0x200
; GFX11-PAL-GCNTRACKERS-NEXT: ; implicit-def: $vgpr50 : SGPR spill to VGPR lane
; GFX11-PAL-GCNTRACKERS-NEXT: .LBB4_1: ; %for.cond28.preheader
; GFX11-PAL-GCNTRACKERS-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-PAL-GCNTRACKERS-NEXT: s_clause 0x3
; GFX11-PAL-GCNTRACKERS-NEXT: global_load_b32 v37, v[4:5], off offset:-392
; GFX11-PAL-GCNTRACKERS-NEXT: global_load_b32 v38, v[4:5], off offset:-196
; GFX11-PAL-GCNTRACKERS-NEXT: global_load_b32 v36, v[4:5], off
; GFX11-PAL-GCNTRACKERS-NEXT: global_load_b32 v35, v[4:5], off offset:196
; GFX11-PAL-GCNTRACKERS-NEXT: s_clause 0x4
; GFX11-PAL-GCNTRACKERS-NEXT: s_load_b512 s[84:99], s[16:17], 0x0
; GFX11-PAL-GCNTRACKERS-NEXT: s_load_b512 s[52:67], s[16:17], 0x1000
; GFX11-PAL-GCNTRACKERS-NEXT: s_load_b512 s[36:51], s[16:17], 0x2000
; GFX11-PAL-GCNTRACKERS-NEXT: s_load_b512 s[0:15], s[16:17], 0x3000
; GFX11-PAL-GCNTRACKERS-NEXT: s_load_b512 s[68:83], s[16:17], 0x40
; GFX11-PAL-GCNTRACKERS-NEXT: v_add_co_u32 v4, vcc_lo, 0x310, v4
; GFX11-PAL-GCNTRACKERS-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-PAL-GCNTRACKERS-NEXT: v_add_co_ci_u32_e64 v5, null, 0, v5, vcc_lo
; GFX11-PAL-GCNTRACKERS-NEXT: s_waitcnt lgkmcnt(0)
; GFX11-PAL-GCNTRACKERS-NEXT: v_writelane_b32 v50, s0, 0
; GFX11-PAL-GCNTRACKERS-NEXT: v_writelane_b32 v50, s1, 1
; GFX11-PAL-GCNTRACKERS-NEXT: v_writelane_b32 v50, s2, 2
; GFX11-PAL-GCNTRACKERS-NEXT: v_writelane_b32 v50, s3, 3
; GFX11-PAL-GCNTRACKERS-NEXT: v_writelane_b32 v50, s4, 4
; GFX11-PAL-GCNTRACKERS-NEXT: v_writelane_b32 v50, s5, 5
; GFX11-PAL-GCNTRACKERS-NEXT: v_writelane_b32 v50, s6, 6
; GFX11-PAL-GCNTRACKERS-NEXT: v_writelane_b32 v50, s7, 7
; GFX11-PAL-GCNTRACKERS-NEXT: v_writelane_b32 v50, s8, 8
; GFX11-PAL-GCNTRACKERS-NEXT: v_writelane_b32 v50, s9, 9
; GFX11-PAL-GCNTRACKERS-NEXT: v_writelane_b32 v50, s10, 10
; GFX11-PAL-GCNTRACKERS-NEXT: v_writelane_b32 v50, s11, 11
; GFX11-PAL-GCNTRACKERS-NEXT: v_writelane_b32 v50, s12, 12
; GFX11-PAL-GCNTRACKERS-NEXT: v_writelane_b32 v50, s13, 13
; GFX11-PAL-GCNTRACKERS-NEXT: v_writelane_b32 v50, s14, 14
; GFX11-PAL-GCNTRACKERS-NEXT: v_writelane_b32 v50, s15, 15
; GFX11-PAL-GCNTRACKERS-NEXT: s_load_b512 s[0:15], s[16:17], 0x1040
; GFX11-PAL-GCNTRACKERS-NEXT: s_waitcnt vmcnt(3)
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v32, s85, v37
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v33, s84, v37
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v30, s87, v37
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v31, s86, v37
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v3, s89, v37
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v2, s88, v37
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v28, s92, v37
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v1, s90, v37
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v26, s94, v37
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v29, s91, v37
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v24, s96, v37
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v27, s93, v37
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v22, s98, v37
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v25, s95, v37
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v20, s68, v37
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v23, s97, v37
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v18, s70, v37
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v21, s99, v37
; GFX11-PAL-GCNTRACKERS-NEXT: s_load_b512 s[84:99], s[16:17], 0x2040
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v16, s72, v37
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v19, s69, v37
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v0, s74, v37
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v17, s71, v37
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v10, s75, v37
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v11, s73, v37
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v8, s77, v37
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v9, s76, v37
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v6, s79, v37
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v7, s78, v37
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v13, s81, v37
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v12, s80, v37
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v15, s83, v37
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v14, s82, v37
; GFX11-PAL-GCNTRACKERS-NEXT: s_load_b512 s[68:83], s[16:17], 0x3040
; GFX11-PAL-GCNTRACKERS-NEXT: s_waitcnt vmcnt(2)
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v32, s53, v38
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v33, s52, v38
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v30, s55, v38
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v31, s54, v38
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v3, s57, v38
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v2, s56, v38
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v28, s60, v38
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v1, s58, v38
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v26, s62, v38
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v29, s59, v38
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v24, s64, v38
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v27, s61, v38
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v22, s66, v38
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v25, s63, v38
; GFX11-PAL-GCNTRACKERS-NEXT: s_waitcnt lgkmcnt(0)
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v20, s0, v38
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v23, s65, v38
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v18, s2, v38
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v21, s67, v38
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v16, s4, v38
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v19, s1, v38
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v0, s6, v38
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v17, s3, v38
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v10, s7, v38
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v11, s5, v38
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v8, s9, v38
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v9, s8, v38
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v6, s11, v38
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v7, s10, v38
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v13, s13, v38
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v12, s12, v38
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v15, s15, v38
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v14, s14, v38
; GFX11-PAL-GCNTRACKERS-NEXT: v_readlane_b32 s0, v50, 0
; GFX11-PAL-GCNTRACKERS-NEXT: s_waitcnt vmcnt(1)
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v32, s37, v36
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v16, s88, v36
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v19, s85, v36
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v0, s90, v36
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v17, s87, v36
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v10, s91, v36
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v11, s89, v36
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v8, s93, v36
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v9, s92, v36
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v6, s95, v36
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v7, s94, v36
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v13, s97, v36
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v12, s96, v36
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v15, s99, v36
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v14, s98, v36
; GFX11-PAL-GCNTRACKERS-NEXT: v_readlane_b32 s1, v50, 1
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v33, s36, v36
; GFX11-PAL-GCNTRACKERS-NEXT: v_readlane_b32 s2, v50, 2
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v30, s39, v36
; GFX11-PAL-GCNTRACKERS-NEXT: v_readlane_b32 s3, v50, 3
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v31, s38, v36
; GFX11-PAL-GCNTRACKERS-NEXT: v_readlane_b32 s4, v50, 4
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v3, s41, v36
; GFX11-PAL-GCNTRACKERS-NEXT: v_readlane_b32 s5, v50, 5
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v2, s40, v36
; GFX11-PAL-GCNTRACKERS-NEXT: v_readlane_b32 s6, v50, 6
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v28, s44, v36
; GFX11-PAL-GCNTRACKERS-NEXT: v_readlane_b32 s7, v50, 7
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v26, s46, v36
; GFX11-PAL-GCNTRACKERS-NEXT: v_readlane_b32 s8, v50, 8
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v1, s42, v36
; GFX11-PAL-GCNTRACKERS-NEXT: v_readlane_b32 s9, v50, 9
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v24, s48, v36
; GFX11-PAL-GCNTRACKERS-NEXT: v_readlane_b32 s10, v50, 10
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v29, s43, v36
; GFX11-PAL-GCNTRACKERS-NEXT: v_readlane_b32 s11, v50, 11
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v22, s50, v36
; GFX11-PAL-GCNTRACKERS-NEXT: v_readlane_b32 s12, v50, 12
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v27, s45, v36
; GFX11-PAL-GCNTRACKERS-NEXT: v_readlane_b32 s13, v50, 13
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v20, s84, v36
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v23, s49, v36
; GFX11-PAL-GCNTRACKERS-NEXT: v_readlane_b32 s14, v50, 14
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v25, s47, v36
; GFX11-PAL-GCNTRACKERS-NEXT: v_readlane_b32 s15, v50, 15
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v18, s86, v36
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v21, s51, v36
; GFX11-PAL-GCNTRACKERS-NEXT: s_add_u32 s16, s16, 0x4000
; GFX11-PAL-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v32, s1, v35
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v33, s0, v35
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v30, s3, v35
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v31, s2, v35
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v3, s5, v35
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v2, s4, v35
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v28, s8, v35
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v1, s6, v35
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v26, s10, v35
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v29, s7, v35
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v24, s12, v35
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v27, s9, v35
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v22, s14, v35
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v25, s11, v35
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v20, s68, v35
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v23, s13, v35
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v18, s70, v35
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v21, s15, v35
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v16, s72, v35
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v19, s69, v35
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v0, s74, v35
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v17, s71, v35
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v10, s75, v35
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v11, s73, v35
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v8, s77, v35
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v9, s76, v35
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v6, s79, v35
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v7, s78, v35
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v13, s81, v35
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v12, s80, v35
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v15, s83, v35
; GFX11-PAL-GCNTRACKERS-NEXT: v_fmac_f32_e32 v14, s82, v35
; GFX11-PAL-GCNTRACKERS-NEXT: s_addc_u32 s17, s17, 0
; GFX11-PAL-GCNTRACKERS-NEXT: s_add_i32 s34, s34, -1
; GFX11-PAL-GCNTRACKERS-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-PAL-GCNTRACKERS-NEXT: s_cmp_lg_u32 s34, 0
; GFX11-PAL-GCNTRACKERS-NEXT: s_cbranch_scc1 .LBB4_1
; GFX11-PAL-GCNTRACKERS-NEXT: ; %bb.2: ; %for.cond.cleanup26
; GFX11-PAL-GCNTRACKERS-NEXT: s_lshl_b32 s0, s33, 8
; GFX11-PAL-GCNTRACKERS-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_3) | instid1(SALU_CYCLE_1)
; GFX11-PAL-GCNTRACKERS-NEXT: v_add3_u32 v4, s22, s0, v34
; GFX11-PAL-GCNTRACKERS-NEXT: s_mov_b32 s0, 0
; GFX11-PAL-GCNTRACKERS-NEXT: v_mov_b32_e32 v5, 0
; GFX11-PAL-GCNTRACKERS-NEXT: s_mov_b32 s1, s0
; GFX11-PAL-GCNTRACKERS-NEXT: v_dual_mov_b32 v36, s0 :: v_dual_mov_b32 v37, s1
; GFX11-PAL-GCNTRACKERS-NEXT: v_mul_hi_u32 v4, 0x5397829d, v4
; GFX11-PAL-GCNTRACKERS-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-PAL-GCNTRACKERS-NEXT: v_lshrrev_b32_e32 v4, 4, v4
; GFX11-PAL-GCNTRACKERS-NEXT: v_mul_lo_u32 v4, 0x31000, v4
; GFX11-PAL-GCNTRACKERS-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-PAL-GCNTRACKERS-NEXT: v_lshlrev_b64 v[34:35], 2, v[4:5]
; GFX11-PAL-GCNTRACKERS-NEXT: v_mov_b32_e32 v4, s0
; GFX11-PAL-GCNTRACKERS-NEXT: v_add_co_u32 v34, vcc_lo, s18, v34
; GFX11-PAL-GCNTRACKERS-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-PAL-GCNTRACKERS-NEXT: v_add_co_ci_u32_e64 v35, null, s19, v35, vcc_lo
; GFX11-PAL-GCNTRACKERS-NEXT: s_clause 0x5
; GFX11-PAL-GCNTRACKERS-NEXT: global_store_b32 v[34:35], v33, off
; GFX11-PAL-GCNTRACKERS-NEXT: global_store_b32 v[34:35], v32, off offset:784
; GFX11-PAL-GCNTRACKERS-NEXT: global_store_b64 v[34:35], v[36:37], off offset:840
; GFX11-PAL-GCNTRACKERS-NEXT: global_store_b32 v[34:35], v31, off offset:1568
; GFX11-PAL-GCNTRACKERS-NEXT: global_store_b32 v[34:35], v30, off offset:2352
; GFX11-PAL-GCNTRACKERS-NEXT: global_store_b32 v[34:35], v2, off offset:3136
; GFX11-PAL-GCNTRACKERS-NEXT: v_add_co_u32 v38, vcc_lo, 0x310, v34
; GFX11-PAL-GCNTRACKERS-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-PAL-GCNTRACKERS-NEXT: v_add_co_ci_u32_e64 v39, null, 0, v35, vcc_lo
; GFX11-PAL-GCNTRACKERS-NEXT: v_add_co_u32 v40, vcc_lo, 0x620, v34
; GFX11-PAL-GCNTRACKERS-NEXT: v_add_co_ci_u32_e64 v41, null, 0, v35, vcc_lo
; GFX11-PAL-GCNTRACKERS-NEXT: v_add_co_u32 v42, vcc_lo, 0x930, v34
; GFX11-PAL-GCNTRACKERS-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-PAL-GCNTRACKERS-NEXT: v_add_co_ci_u32_e64 v43, null, 0, v35, vcc_lo
; GFX11-PAL-GCNTRACKERS-NEXT: v_add_co_u32 v44, vcc_lo, 0xc40, v34
; GFX11-PAL-GCNTRACKERS-NEXT: v_add_co_ci_u32_e64 v45, null, 0, v35, vcc_lo
; GFX11-PAL-GCNTRACKERS-NEXT: v_add_co_u32 v30, vcc_lo, 0xf50, v34
; GFX11-PAL-GCNTRACKERS-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-PAL-GCNTRACKERS-NEXT: v_add_co_ci_u32_e64 v31, null, 0, v35, vcc_lo
; GFX11-PAL-GCNTRACKERS-NEXT: v_add_co_u32 v32, vcc_lo, 0xf50, v38
; GFX11-PAL-GCNTRACKERS-NEXT: v_add_co_ci_u32_e64 v33, null, 0, v39, vcc_lo
; GFX11-PAL-GCNTRACKERS-NEXT: v_mov_b32_e32 v2, s0
; GFX11-PAL-GCNTRACKERS-NEXT: v_add_co_u32 v46, vcc_lo, 0xf50, v40
; GFX11-PAL-GCNTRACKERS-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-PAL-GCNTRACKERS-NEXT: v_add_co_ci_u32_e64 v47, null, 0, v41, vcc_lo
; GFX11-PAL-GCNTRACKERS-NEXT: v_add_co_u32 v48, vcc_lo, 0xf50, v42
; GFX11-PAL-GCNTRACKERS-NEXT: v_add_co_ci_u32_e64 v49, null, 0, v43, vcc_lo
; GFX11-PAL-GCNTRACKERS-NEXT: s_clause 0x3
; GFX11-PAL-GCNTRACKERS-NEXT: global_store_b64 v[34:35], v[3:4], off offset:3920
; GFX11-PAL-GCNTRACKERS-NEXT: global_store_b64 v[38:39], v[1:2], off offset:3920
; GFX11-PAL-GCNTRACKERS-NEXT: global_store_b32 v[40:41], v29, off offset:3920
; GFX11-PAL-GCNTRACKERS-NEXT: global_store_b32 v[42:43], v28, off offset:3920
; GFX11-PAL-GCNTRACKERS-NEXT: v_add_co_u32 v1, vcc_lo, 0xf50, v44
; GFX11-PAL-GCNTRACKERS-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-PAL-GCNTRACKERS-NEXT: v_add_co_ci_u32_e64 v2, null, 0, v45, vcc_lo
; GFX11-PAL-GCNTRACKERS-NEXT: v_add_co_u32 v3, vcc_lo, 0xf50, v30
; GFX11-PAL-GCNTRACKERS-NEXT: s_clause 0x1
; GFX11-PAL-GCNTRACKERS-NEXT: global_store_b64 v[42:43], v[36:37], off offset:3976
; GFX11-PAL-GCNTRACKERS-NEXT: global_store_b32 v[44:45], v27, off offset:3920
; GFX11-PAL-GCNTRACKERS-NEXT: v_add_co_ci_u32_e64 v4, null, 0, v31, vcc_lo
; GFX11-PAL-GCNTRACKERS-NEXT: global_store_b32 v[30:31], v26, off offset:3920
; GFX11-PAL-GCNTRACKERS-NEXT: v_add_co_u32 v26, vcc_lo, 0xf50, v32
; GFX11-PAL-GCNTRACKERS-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-PAL-GCNTRACKERS-NEXT: v_add_co_ci_u32_e64 v27, null, 0, v33, vcc_lo
; GFX11-PAL-GCNTRACKERS-NEXT: v_add_co_u32 v28, vcc_lo, 0xf50, v46
; GFX11-PAL-GCNTRACKERS-NEXT: v_add_co_ci_u32_e64 v29, null, 0, v47, vcc_lo
; GFX11-PAL-GCNTRACKERS-NEXT: global_store_b32 v[46:47], v24, off offset:3920
; GFX11-PAL-GCNTRACKERS-NEXT: v_add_co_u32 v24, vcc_lo, 0xf50, v48
; GFX11-PAL-GCNTRACKERS-NEXT: global_store_b32 v[32:33], v25, off offset:3920
; GFX11-PAL-GCNTRACKERS-NEXT: v_add_co_ci_u32_e64 v25, null, 0, v49, vcc_lo
; GFX11-PAL-GCNTRACKERS-NEXT: v_add_co_u32 v30, vcc_lo, 0xf50, v1
; GFX11-PAL-GCNTRACKERS-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1)
; GFX11-PAL-GCNTRACKERS-NEXT: v_add_co_ci_u32_e64 v31, null, 0, v2, vcc_lo
; GFX11-PAL-GCNTRACKERS-NEXT: global_store_b32 v[1:2], v22, off offset:3920
; GFX11-PAL-GCNTRACKERS-NEXT: v_add_co_u32 v1, vcc_lo, 0xf50, v3
; GFX11-PAL-GCNTRACKERS-NEXT: v_add_co_ci_u32_e64 v2, null, 0, v4, vcc_lo
; GFX11-PAL-GCNTRACKERS-NEXT: global_store_b32 v[3:4], v21, off offset:3920
; GFX11-PAL-GCNTRACKERS-NEXT: v_add_co_u32 v3, vcc_lo, 0xf50, v26
; GFX11-PAL-GCNTRACKERS-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1)
; GFX11-PAL-GCNTRACKERS-NEXT: v_add_co_ci_u32_e64 v4, null, 0, v27, vcc_lo
; GFX11-PAL-GCNTRACKERS-NEXT: global_store_b32 v[26:27], v20, off offset:3920
; GFX11-PAL-GCNTRACKERS-NEXT: v_add_co_u32 v20, vcc_lo, 0xf50, v28
; GFX11-PAL-GCNTRACKERS-NEXT: v_add_co_ci_u32_e64 v21, null, 0, v29, vcc_lo
; GFX11-PAL-GCNTRACKERS-NEXT: v_add_co_u32 v22, vcc_lo, 0xf50, v24
; GFX11-PAL-GCNTRACKERS-NEXT: global_store_b32 v[48:49], v23, off offset:3920
; GFX11-PAL-GCNTRACKERS-NEXT: v_add_co_ci_u32_e64 v23, null, 0, v25, vcc_lo
; GFX11-PAL-GCNTRACKERS-NEXT: global_store_b32 v[24:25], v18, off offset:3920
; GFX11-PAL-GCNTRACKERS-NEXT: v_add_co_u32 v18, vcc_lo, 0xf50, v30
; GFX11-PAL-GCNTRACKERS-NEXT: global_store_b32 v[28:29], v19, off offset:3920
; GFX11-PAL-GCNTRACKERS-NEXT: v_add_co_ci_u32_e64 v19, null, 0, v31, vcc_lo
; GFX11-PAL-GCNTRACKERS-NEXT: v_add_co_u32 v24, vcc_lo, 0xf50, v1
; GFX11-PAL-GCNTRACKERS-NEXT: global_store_b32 v[30:31], v17, off offset:3920
; GFX11-PAL-GCNTRACKERS-NEXT: v_add_co_ci_u32_e64 v25, null, 0, v2, vcc_lo
; GFX11-PAL-GCNTRACKERS-NEXT: global_store_b32 v[1:2], v16, off offset:3920
; GFX11-PAL-GCNTRACKERS-NEXT: v_add_co_u32 v1, vcc_lo, 0xf50, v3
; GFX11-PAL-GCNTRACKERS-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1)
; GFX11-PAL-GCNTRACKERS-NEXT: v_add_co_ci_u32_e64 v2, null, 0, v4, vcc_lo
; GFX11-PAL-GCNTRACKERS-NEXT: global_store_b32 v[3:4], v11, off offset:3920
; GFX11-PAL-GCNTRACKERS-NEXT: v_add_co_u32 v3, vcc_lo, 0xf50, v20
; GFX11-PAL-GCNTRACKERS-NEXT: v_add_co_ci_u32_e64 v4, null, 0, v21, vcc_lo
; GFX11-PAL-GCNTRACKERS-NEXT: v_add_co_u32 v16, vcc_lo, 0xf50, v22
; GFX11-PAL-GCNTRACKERS-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1)
; GFX11-PAL-GCNTRACKERS-NEXT: v_add_co_ci_u32_e64 v17, null, 0, v23, vcc_lo
; GFX11-PAL-GCNTRACKERS-NEXT: global_store_b32 v[22:23], v10, off offset:3920
; GFX11-PAL-GCNTRACKERS-NEXT: v_add_co_u32 v10, vcc_lo, 0xf50, v18
; GFX11-PAL-GCNTRACKERS-NEXT: v_add_co_ci_u32_e64 v11, null, 0, v19, vcc_lo
; GFX11-PAL-GCNTRACKERS-NEXT: global_store_b32 v[18:19], v9, off offset:3920
; GFX11-PAL-GCNTRACKERS-NEXT: v_add_co_u32 v18, vcc_lo, 0xf50, v24
; GFX11-PAL-GCNTRACKERS-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-PAL-GCNTRACKERS-NEXT: v_add_co_ci_u32_e64 v19, null, 0, v25, vcc_lo
; GFX11-PAL-GCNTRACKERS-NEXT: global_store_b32 v[24:25], v8, off offset:3920
; GFX11-PAL-GCNTRACKERS-NEXT: v_add_co_u32 v8, vcc_lo, 0xf50, v1
; GFX11-PAL-GCNTRACKERS-NEXT: global_store_b32 v[20:21], v0, off offset:3920
; GFX11-PAL-GCNTRACKERS-NEXT: v_add_co_ci_u32_e64 v9, null, 0, v2, vcc_lo
; GFX11-PAL-GCNTRACKERS-NEXT: global_store_b32 v[1:2], v7, off offset:3920
; GFX11-PAL-GCNTRACKERS-NEXT: global_store_b32 v[3:4], v6, off offset:3920
; GFX11-PAL-GCNTRACKERS-NEXT: global_store_b32 v[16:17], v12, off offset:3920
; GFX11-PAL-GCNTRACKERS-NEXT: global_store_b32 v[10:11], v13, off offset:3920
; GFX11-PAL-GCNTRACKERS-NEXT: global_store_b32 v[18:19], v14, off offset:3920
; GFX11-PAL-GCNTRACKERS-NEXT: global_store_b32 v[8:9], v15, off offset:3920
; GFX11-PAL-GCNTRACKERS-NEXT: s_endpgm
;
; TONGA-LABEL: excess_soft_clause_reg_pressure:
; TONGA: ; %bb.0: ; %entry
; TONGA-NEXT: s_load_dwordx8 s[0:7], s[4:5], 0x24
; TONGA-NEXT: v_lshlrev_b32_e32 v1, 2, v0
; TONGA-NEXT: s_waitcnt lgkmcnt(0)
; TONGA-NEXT: s_movk_i32 s7, 0x188
; TONGA-NEXT: s_movk_i32 s9, 0x24c
; TONGA-NEXT: s_movk_i32 s10, 0x310
; TONGA-NEXT: v_mov_b32_e32 v2, s5
; TONGA-NEXT: v_add_u32_e32 v1, vcc, s4, v1
; TONGA-NEXT: v_addc_u32_e32 v2, vcc, 0, v2, vcc
; TONGA-NEXT: s_movk_i32 s4, 0x200
; TONGA-NEXT: s_movk_i32 s5, 0xc4
; TONGA-NEXT: v_mov_b32_e32 v11, 0
; TONGA-NEXT: v_mov_b32_e32 v13, 0
; TONGA-NEXT: v_mov_b32_e32 v15, 0
; TONGA-NEXT: v_mov_b32_e32 v18, 0
; TONGA-NEXT: v_mov_b32_e32 v19, 0
; TONGA-NEXT: v_mov_b32_e32 v20, 0
; TONGA-NEXT: v_mov_b32_e32 v21, 0
; TONGA-NEXT: v_mov_b32_e32 v22, 0
; TONGA-NEXT: v_mov_b32_e32 v23, 0
; TONGA-NEXT: v_mov_b32_e32 v24, 0
; TONGA-NEXT: v_mov_b32_e32 v25, 0
; TONGA-NEXT: v_mov_b32_e32 v26, 0
; TONGA-NEXT: v_mov_b32_e32 v27, 0
; TONGA-NEXT: v_mov_b32_e32 v28, 0
; TONGA-NEXT: v_mov_b32_e32 v29, 0
; TONGA-NEXT: v_mov_b32_e32 v30, 0
; TONGA-NEXT: v_mov_b32_e32 v31, 0
; TONGA-NEXT: v_mov_b32_e32 v32, 0
; TONGA-NEXT: v_mov_b32_e32 v33, 0
; TONGA-NEXT: v_mov_b32_e32 v34, 0
; TONGA-NEXT: v_mov_b32_e32 v35, 0
; TONGA-NEXT: v_mov_b32_e32 v36, 0
; TONGA-NEXT: v_mov_b32_e32 v37, 0
; TONGA-NEXT: v_mov_b32_e32 v38, 0
; TONGA-NEXT: v_mov_b32_e32 v39, 0
; TONGA-NEXT: v_mov_b32_e32 v40, 0
; TONGA-NEXT: v_mov_b32_e32 v17, 0
; TONGA-NEXT: v_mov_b32_e32 v16, 0
; TONGA-NEXT: v_mov_b32_e32 v14, 0
; TONGA-NEXT: v_mov_b32_e32 v12, 0
; TONGA-NEXT: v_mov_b32_e32 v3, 0
; TONGA-NEXT: v_mov_b32_e32 v4, 0
; TONGA-NEXT: .LBB4_1: ; %for.cond28.preheader
; TONGA-NEXT: ; =>This Inner Loop Header: Depth=1
; TONGA-NEXT: flat_load_dword v41, v[1:2]
; TONGA-NEXT: v_add_u32_e32 v5, vcc, s5, v1
; TONGA-NEXT: v_addc_u32_e32 v6, vcc, 0, v2, vcc
; TONGA-NEXT: v_add_u32_e32 v7, vcc, s7, v1
; TONGA-NEXT: v_addc_u32_e32 v8, vcc, 0, v2, vcc
; TONGA-NEXT: v_add_u32_e32 v9, vcc, s9, v1
; TONGA-NEXT: v_addc_u32_e32 v10, vcc, 0, v2, vcc
; TONGA-NEXT: flat_load_dword v42, v[5:6]
; TONGA-NEXT: flat_load_dword v6, v[7:8]
; TONGA-NEXT: flat_load_dword v5, v[9:10]
; TONGA-NEXT: s_load_dwordx16 s[12:27], s[0:1], 0x0
; TONGA-NEXT: s_load_dwordx16 s[36:51], s[0:1], 0x40
; TONGA-NEXT: s_load_dwordx16 s[52:67], s[0:1], 0x1000
; TONGA-NEXT: s_load_dwordx16 s[68:83], s[0:1], 0x1040
; TONGA-NEXT: v_add_u32_e32 v1, vcc, s10, v1
; TONGA-NEXT: v_addc_u32_e32 v2, vcc, 0, v2, vcc
; TONGA-NEXT: s_waitcnt vmcnt(3) lgkmcnt(0)
; TONGA-NEXT: v_mul_f32_e32 v7, s17, v41
; TONGA-NEXT: v_mul_f32_e32 v8, s18, v41
; TONGA-NEXT: v_mul_f32_e32 v9, s12, v41
; TONGA-NEXT: v_add_f32_e32 v43, v7, v4
; TONGA-NEXT: v_mul_f32_e32 v4, s14, v41
; TONGA-NEXT: v_mul_f32_e32 v7, s15, v41
; TONGA-NEXT: v_add_f32_e32 v44, v8, v3
; TONGA-NEXT: v_mul_f32_e32 v3, s16, v41
; TONGA-NEXT: v_mul_f32_e32 v10, s13, v41
; TONGA-NEXT: v_mul_f32_e32 v8, s19, v41
; TONGA-NEXT: v_add_f32_e32 v48, v9, v40
; TONGA-NEXT: v_mul_f32_e32 v9, s20, v41
; TONGA-NEXT: v_add_f32_e32 v50, v4, v38
; TONGA-NEXT: v_mul_f32_e32 v4, s22, v41
; TONGA-NEXT: v_add_f32_e32 v51, v7, v37
; TONGA-NEXT: v_mul_f32_e32 v7, s23, v41
; TONGA-NEXT: v_add_f32_e32 v52, v3, v36
; TONGA-NEXT: v_mul_f32_e32 v3, s24, v41
; TONGA-NEXT: v_add_f32_e32 v49, v10, v39
; TONGA-NEXT: v_mul_f32_e32 v10, s21, v41
; TONGA-NEXT: v_add_f32_e32 v35, v8, v35
; TONGA-NEXT: v_mul_f32_e32 v8, s25, v41
; TONGA-NEXT: v_add_f32_e32 v34, v9, v34
; TONGA-NEXT: v_mul_f32_e32 v9, s26, v41
; TONGA-NEXT: v_add_f32_e32 v32, v4, v32
; TONGA-NEXT: v_mul_f32_e32 v4, s36, v41
; TONGA-NEXT: v_add_f32_e32 v31, v7, v31
; TONGA-NEXT: v_mul_f32_e32 v7, s37, v41
; TONGA-NEXT: v_add_f32_e32 v30, v3, v30
; TONGA-NEXT: v_mul_f32_e32 v3, s38, v41
; TONGA-NEXT: v_add_f32_e32 v33, v10, v33
; TONGA-NEXT: v_mul_f32_e32 v10, s27, v41
; TONGA-NEXT: s_load_dwordx16 s[12:27], s[0:1], 0x2000
; TONGA-NEXT: v_add_f32_e32 v29, v8, v29
; TONGA-NEXT: v_mul_f32_e32 v8, s39, v41
; TONGA-NEXT: v_add_f32_e32 v28, v9, v28
; TONGA-NEXT: v_mul_f32_e32 v9, s40, v41
; TONGA-NEXT: v_add_f32_e32 v37, v4, v26
; TONGA-NEXT: v_mul_f32_e32 v4, s42, v41
; TONGA-NEXT: v_add_f32_e32 v38, v7, v25
; TONGA-NEXT: v_mul_f32_e32 v7, s43, v41
; TONGA-NEXT: v_add_f32_e32 v39, v3, v24
; TONGA-NEXT: v_mul_f32_e32 v3, s44, v41
; TONGA-NEXT: v_add_f32_e32 v36, v10, v27
; TONGA-NEXT: v_mul_f32_e32 v10, s41, v41
; TONGA-NEXT: v_add_f32_e32 v40, v8, v23
; TONGA-NEXT: v_mul_f32_e32 v8, s45, v41
; TONGA-NEXT: v_add_f32_e32 v22, v9, v22
; TONGA-NEXT: v_mul_f32_e32 v9, s46, v41
; TONGA-NEXT: v_add_f32_e32 v46, v4, v20
; TONGA-NEXT: v_mul_f32_e32 v4, s48, v41
; TONGA-NEXT: v_add_f32_e32 v47, v7, v19
; TONGA-NEXT: v_mul_f32_e32 v7, s49, v41
; TONGA-NEXT: v_add_f32_e32 v3, v3, v18
; TONGA-NEXT: v_mul_f32_e32 v18, s50, v41
; TONGA-NEXT: v_mul_f32_e32 v19, s51, v41
; TONGA-NEXT: s_waitcnt vmcnt(2)
; TONGA-NEXT: v_mul_f32_e32 v23, s61, v42
; TONGA-NEXT: v_add_f32_e32 v45, v10, v21
; TONGA-NEXT: v_mul_f32_e32 v10, s47, v41
; TONGA-NEXT: s_load_dwordx16 s[36:51], s[0:1], 0x2040
; TONGA-NEXT: v_add_f32_e32 v21, v8, v15
; TONGA-NEXT: v_add_f32_e32 v13, v9, v13
; TONGA-NEXT: v_add_f32_e32 v9, v4, v17
; TONGA-NEXT: v_add_f32_e32 v8, v7, v16
; TONGA-NEXT: v_add_f32_e32 v7, v18, v14
; TONGA-NEXT: v_add_f32_e32 v4, v19, v12
; TONGA-NEXT: v_mul_f32_e32 v17, s57, v42
; TONGA-NEXT: v_mul_f32_e32 v18, s58, v42
; TONGA-NEXT: v_mul_f32_e32 v19, s59, v42
; TONGA-NEXT: v_mul_f32_e32 v20, s60, v42
; TONGA-NEXT: v_mul_f32_e32 v24, s62, v42
; TONGA-NEXT: v_mul_f32_e32 v26, s64, v42
; TONGA-NEXT: v_add_f32_e32 v23, v23, v33
; TONGA-NEXT: v_mul_f32_e32 v33, s68, v42
; TONGA-NEXT: v_add_f32_e32 v10, v10, v11
; TONGA-NEXT: v_mul_f32_e32 v11, s52, v42
; TONGA-NEXT: v_mul_f32_e32 v12, s53, v42
; TONGA-NEXT: v_mul_f32_e32 v14, s54, v42
; TONGA-NEXT: v_mul_f32_e32 v15, s55, v42
; TONGA-NEXT: v_mul_f32_e32 v16, s56, v42
; TONGA-NEXT: v_mul_f32_e32 v25, s63, v42
; TONGA-NEXT: v_add_f32_e32 v17, v17, v43
; TONGA-NEXT: v_add_f32_e32 v18, v18, v44
; TONGA-NEXT: v_mul_f32_e32 v27, s65, v42
; TONGA-NEXT: v_add_f32_e32 v19, v19, v35
; TONGA-NEXT: v_mul_f32_e32 v35, s66, v42
; TONGA-NEXT: v_add_f32_e32 v20, v20, v34
; TONGA-NEXT: v_mul_f32_e32 v34, s67, v42
; TONGA-NEXT: s_load_dwordx16 s[52:67], s[0:1], 0x3000
; TONGA-NEXT: v_add_f32_e32 v24, v24, v32
; TONGA-NEXT: v_mul_f32_e32 v32, s69, v42
; TONGA-NEXT: v_mul_f32_e32 v41, s70, v42
; TONGA-NEXT: v_add_f32_e32 v26, v26, v30
; TONGA-NEXT: v_mul_f32_e32 v43, s71, v42
; TONGA-NEXT: v_mul_f32_e32 v44, s72, v42
; TONGA-NEXT: v_add_f32_e32 v30, v33, v37
; TONGA-NEXT: v_mul_f32_e32 v37, s75, v42
; TONGA-NEXT: v_add_f32_e32 v25, v25, v31
; TONGA-NEXT: v_add_f32_e32 v27, v27, v29
; TONGA-NEXT: v_add_f32_e32 v29, v34, v36
; TONGA-NEXT: v_add_f32_e32 v31, v32, v38
; TONGA-NEXT: v_mul_f32_e32 v38, s76, v42
; TONGA-NEXT: v_add_f32_e32 v32, v41, v39
; TONGA-NEXT: v_mul_f32_e32 v39, s77, v42
; TONGA-NEXT: v_add_f32_e32 v33, v43, v40
; TONGA-NEXT: v_mul_f32_e32 v40, s78, v42
; TONGA-NEXT: v_add_f32_e32 v34, v44, v22
; TONGA-NEXT: v_mul_f32_e32 v41, s79, v42
; TONGA-NEXT: v_mul_f32_e32 v43, s80, v42
; TONGA-NEXT: v_mul_f32_e32 v44, s81, v42
; TONGA-NEXT: v_add_f32_e32 v22, v37, v47
; TONGA-NEXT: v_mul_f32_e32 v37, s82, v42
; TONGA-NEXT: v_add_f32_e32 v11, v11, v48
; TONGA-NEXT: v_add_f32_e32 v12, v12, v49
; TONGA-NEXT: v_add_f32_e32 v14, v14, v50
; TONGA-NEXT: v_add_f32_e32 v15, v15, v51
; TONGA-NEXT: v_add_f32_e32 v16, v16, v52
; TONGA-NEXT: v_add_f32_e32 v28, v35, v28
; TONGA-NEXT: v_mul_f32_e32 v35, s73, v42
; TONGA-NEXT: v_mul_f32_e32 v36, s74, v42
; TONGA-NEXT: v_mul_f32_e32 v42, s83, v42
; TONGA-NEXT: s_load_dwordx16 s[68:83], s[0:1], 0x3040
; TONGA-NEXT: v_add_f32_e32 v3, v38, v3
; TONGA-NEXT: s_waitcnt vmcnt(1) lgkmcnt(0)
; TONGA-NEXT: v_mul_f32_e32 v38, s12, v6
; TONGA-NEXT: v_add_f32_e32 v21, v39, v21
; TONGA-NEXT: v_mul_f32_e32 v39, s13, v6
; TONGA-NEXT: v_add_f32_e32 v13, v40, v13
; TONGA-NEXT: v_mul_f32_e32 v40, s14, v6
; TONGA-NEXT: v_add_f32_e32 v10, v41, v10
; TONGA-NEXT: v_mul_f32_e32 v41, s15, v6
; TONGA-NEXT: v_add_f32_e32 v9, v43, v9
; TONGA-NEXT: v_mul_f32_e32 v43, s16, v6
; TONGA-NEXT: v_add_f32_e32 v8, v44, v8
; TONGA-NEXT: v_mul_f32_e32 v44, s17, v6
; TONGA-NEXT: v_add_f32_e32 v7, v37, v7
; TONGA-NEXT: v_mul_f32_e32 v37, s18, v6
; TONGA-NEXT: v_add_f32_e32 v35, v35, v45
; TONGA-NEXT: v_add_f32_e32 v4, v42, v4
; TONGA-NEXT: v_mul_f32_e32 v42, s19, v6
; TONGA-NEXT: v_add_f32_e32 v11, v38, v11
; TONGA-NEXT: v_mul_f32_e32 v38, s20, v6
; TONGA-NEXT: v_add_f32_e32 v12, v39, v12
; TONGA-NEXT: v_mul_f32_e32 v39, s21, v6
; TONGA-NEXT: v_add_f32_e32 v14, v40, v14
; TONGA-NEXT: v_mul_f32_e32 v40, s22, v6
; TONGA-NEXT: v_add_f32_e32 v15, v41, v15
; TONGA-NEXT: v_mul_f32_e32 v41, s23, v6
; TONGA-NEXT: v_add_f32_e32 v16, v43, v16
; TONGA-NEXT: v_mul_f32_e32 v43, s24, v6
; TONGA-NEXT: v_add_f32_e32 v44, v44, v17
; TONGA-NEXT: v_mul_f32_e32 v17, s25, v6
; TONGA-NEXT: v_add_f32_e32 v45, v37, v18
; TONGA-NEXT: v_mul_f32_e32 v18, s26, v6
; TONGA-NEXT: v_mul_f32_e32 v37, s27, v6
; TONGA-NEXT: v_add_f32_e32 v36, v36, v46
; TONGA-NEXT: v_add_f32_e32 v19, v42, v19
; TONGA-NEXT: v_add_f32_e32 v20, v38, v20
; TONGA-NEXT: v_mul_f32_e32 v38, s36, v6
; TONGA-NEXT: v_add_f32_e32 v23, v39, v23
; TONGA-NEXT: v_mul_f32_e32 v39, s37, v6
; TONGA-NEXT: v_add_f32_e32 v24, v40, v24
; TONGA-NEXT: v_mul_f32_e32 v40, s38, v6
; TONGA-NEXT: v_add_f32_e32 v25, v41, v25
; TONGA-NEXT: v_mul_f32_e32 v41, s39, v6
; TONGA-NEXT: v_add_f32_e32 v26, v43, v26
; TONGA-NEXT: v_mul_f32_e32 v42, s40, v6
; TONGA-NEXT: v_add_f32_e32 v17, v17, v27
; TONGA-NEXT: v_mul_f32_e32 v27, s41, v6
; TONGA-NEXT: v_add_f32_e32 v18, v18, v28
; TONGA-NEXT: v_mul_f32_e32 v28, s42, v6
; TONGA-NEXT: v_add_f32_e32 v43, v37, v29
; TONGA-NEXT: v_mul_f32_e32 v29, s43, v6
; TONGA-NEXT: v_add_f32_e32 v46, v38, v30
; TONGA-NEXT: v_mul_f32_e32 v30, s44, v6
; TONGA-NEXT: v_add_f32_e32 v47, v39, v31
; TONGA-NEXT: v_mul_f32_e32 v31, s45, v6
; TONGA-NEXT: v_add_f32_e32 v48, v40, v32
; TONGA-NEXT: v_mul_f32_e32 v32, s46, v6
; TONGA-NEXT: v_add_f32_e32 v41, v41, v33
; TONGA-NEXT: v_mul_f32_e32 v33, s47, v6
; TONGA-NEXT: v_add_f32_e32 v42, v42, v34
; TONGA-NEXT: v_mul_f32_e32 v34, s48, v6
; TONGA-NEXT: v_add_f32_e32 v49, v27, v35
; TONGA-NEXT: v_mul_f32_e32 v27, s49, v6
; TONGA-NEXT: v_add_f32_e32 v50, v28, v36
; TONGA-NEXT: v_mul_f32_e32 v28, s50, v6
; TONGA-NEXT: v_mul_f32_e32 v6, s51, v6
; TONGA-NEXT: v_add_f32_e32 v51, v29, v22
; TONGA-NEXT: s_waitcnt vmcnt(0)
; TONGA-NEXT: v_mul_f32_e32 v22, s52, v5
; TONGA-NEXT: v_mul_f32_e32 v29, s53, v5
; TONGA-NEXT: v_add_f32_e32 v3, v30, v3
; TONGA-NEXT: v_add_f32_e32 v52, v31, v21
; TONGA-NEXT: v_mul_f32_e32 v21, s54, v5
; TONGA-NEXT: v_add_f32_e32 v13, v32, v13
; TONGA-NEXT: v_mul_f32_e32 v30, s55, v5
; TONGA-NEXT: v_mul_f32_e32 v31, s56, v5
; TONGA-NEXT: v_mul_f32_e32 v32, s59, v5
; TONGA-NEXT: v_add_f32_e32 v8, v27, v8
; TONGA-NEXT: v_mul_f32_e32 v27, s60, v5
; TONGA-NEXT: v_add_f32_e32 v4, v6, v4
; TONGA-NEXT: v_mul_f32_e32 v6, s62, v5
; TONGA-NEXT: v_add_f32_e32 v40, v22, v11
; TONGA-NEXT: v_mul_f32_e32 v11, s63, v5
; TONGA-NEXT: v_add_f32_e32 v39, v29, v12
; TONGA-NEXT: v_mul_f32_e32 v12, s64, v5
; TONGA-NEXT: s_add_u32 s0, s0, 0x4000
; TONGA-NEXT: v_add_f32_e32 v9, v34, v9
; TONGA-NEXT: v_add_f32_e32 v7, v28, v7
; TONGA-NEXT: v_mul_f32_e32 v28, s61, v5
; TONGA-NEXT: v_add_f32_e32 v38, v21, v14
; TONGA-NEXT: v_mul_f32_e32 v14, s65, v5
; TONGA-NEXT: v_add_f32_e32 v37, v30, v15
; TONGA-NEXT: v_mul_f32_e32 v15, s66, v5
; TONGA-NEXT: v_add_f32_e32 v36, v31, v16
; TONGA-NEXT: v_mul_f32_e32 v16, s67, v5
; TONGA-NEXT: v_add_f32_e32 v35, v32, v19
; TONGA-NEXT: v_mul_f32_e32 v19, s68, v5
; TONGA-NEXT: v_add_f32_e32 v34, v27, v20
; TONGA-NEXT: v_mul_f32_e32 v20, s69, v5
; TONGA-NEXT: v_mul_f32_e32 v21, s70, v5
; TONGA-NEXT: v_add_f32_e32 v32, v6, v24
; TONGA-NEXT: v_mul_f32_e32 v6, s71, v5
; TONGA-NEXT: v_add_f32_e32 v31, v11, v25
; TONGA-NEXT: v_mul_f32_e32 v11, s72, v5
; TONGA-NEXT: v_add_f32_e32 v30, v12, v26
; TONGA-NEXT: v_mul_f32_e32 v12, s73, v5
; TONGA-NEXT: s_addc_u32 s1, s1, 0
; TONGA-NEXT: s_add_i32 s4, s4, -1
; TONGA-NEXT: v_add_f32_e32 v10, v33, v10
; TONGA-NEXT: v_add_f32_e32 v33, v28, v23
; TONGA-NEXT: v_add_f32_e32 v29, v14, v17
; TONGA-NEXT: v_mul_f32_e32 v14, s74, v5
; TONGA-NEXT: v_add_f32_e32 v28, v15, v18
; TONGA-NEXT: v_mul_f32_e32 v15, s75, v5
; TONGA-NEXT: v_add_f32_e32 v27, v16, v43
; TONGA-NEXT: v_mul_f32_e32 v16, s76, v5
; TONGA-NEXT: v_add_f32_e32 v26, v19, v46
; TONGA-NEXT: v_mul_f32_e32 v17, s77, v5
; TONGA-NEXT: v_add_f32_e32 v25, v20, v47
; TONGA-NEXT: v_mul_f32_e32 v43, s78, v5
; TONGA-NEXT: v_add_f32_e32 v24, v21, v48
; TONGA-NEXT: v_mul_f32_e32 v46, s79, v5
; TONGA-NEXT: v_add_f32_e32 v23, v6, v41
; TONGA-NEXT: v_mul_f32_e32 v6, s80, v5
; TONGA-NEXT: v_add_f32_e32 v22, v11, v42
; TONGA-NEXT: v_mul_f32_e32 v41, s81, v5
; TONGA-NEXT: v_add_f32_e32 v21, v12, v49
; TONGA-NEXT: v_mul_f32_e32 v12, s82, v5
; TONGA-NEXT: v_mul_f32_e32 v42, s83, v5
; TONGA-NEXT: v_mul_f32_e32 v47, s57, v5
; TONGA-NEXT: v_mul_f32_e32 v5, s58, v5
; TONGA-NEXT: s_cmp_lg_u32 s4, 0
; TONGA-NEXT: v_add_f32_e32 v20, v14, v50
; TONGA-NEXT: v_add_f32_e32 v19, v15, v51
; TONGA-NEXT: v_add_f32_e32 v18, v16, v3
; TONGA-NEXT: v_add_f32_e32 v15, v17, v52
; TONGA-NEXT: v_add_f32_e32 v13, v43, v13
; TONGA-NEXT: v_add_f32_e32 v11, v46, v10
; TONGA-NEXT: v_add_f32_e32 v17, v6, v9
; TONGA-NEXT: v_add_f32_e32 v16, v41, v8
; TONGA-NEXT: v_add_f32_e32 v14, v12, v7
; TONGA-NEXT: v_add_f32_e32 v12, v42, v4
; TONGA-NEXT: v_add_f32_e32 v4, v47, v44
; TONGA-NEXT: v_add_f32_e32 v3, v5, v45
; TONGA-NEXT: s_cbranch_scc1 .LBB4_1
; TONGA-NEXT: ; %bb.2: ; %for.cond.cleanup26
; TONGA-NEXT: s_lshl_b32 s0, s8, 8
; TONGA-NEXT: s_add_i32 s0, s6, s0
; TONGA-NEXT: v_add_u32_e32 v0, vcc, s0, v0
; TONGA-NEXT: s_mov_b32 s0, 0x5397829d
; TONGA-NEXT: v_mul_hi_u32 v0, v0, s0
; TONGA-NEXT: s_mov_b32 s0, 0x31000
; TONGA-NEXT: v_mov_b32_e32 v1, 0
; TONGA-NEXT: v_mov_b32_e32 v2, s3
; TONGA-NEXT: v_lshrrev_b32_e32 v0, 4, v0
; TONGA-NEXT: v_mul_lo_u32 v0, v0, s0
; TONGA-NEXT: s_movk_i32 s0, 0x310
; TONGA-NEXT: v_lshlrev_b64 v[0:1], 2, v[0:1]
; TONGA-NEXT: v_add_u32_e32 v0, vcc, s2, v0
; TONGA-NEXT: v_addc_u32_e32 v1, vcc, v2, v1, vcc
; TONGA-NEXT: v_add_u32_e32 v5, vcc, s0, v0
; TONGA-NEXT: v_addc_u32_e32 v6, vcc, 0, v1, vcc
; TONGA-NEXT: s_movk_i32 s0, 0x348
; TONGA-NEXT: flat_store_dword v[5:6], v39
; TONGA-NEXT: v_add_u32_e32 v5, vcc, s0, v0
; TONGA-NEXT: s_mov_b32 s0, 0
; TONGA-NEXT: s_mov_b32 s1, s0
; TONGA-NEXT: v_addc_u32_e32 v6, vcc, 0, v1, vcc
; TONGA-NEXT: v_mov_b32_e32 v8, s1
; TONGA-NEXT: v_mov_b32_e32 v7, s0
; TONGA-NEXT: s_movk_i32 s1, 0x620
; TONGA-NEXT: flat_store_dwordx2 v[5:6], v[7:8]
; TONGA-NEXT: v_add_u32_e32 v5, vcc, s1, v0
; TONGA-NEXT: v_addc_u32_e32 v6, vcc, 0, v1, vcc
; TONGA-NEXT: s_movk_i32 s1, 0x930
; TONGA-NEXT: flat_store_dword v[5:6], v38
; TONGA-NEXT: v_add_u32_e32 v5, vcc, s1, v0
; TONGA-NEXT: v_addc_u32_e32 v6, vcc, 0, v1, vcc
; TONGA-NEXT: s_movk_i32 s1, 0xc40
; TONGA-NEXT: flat_store_dword v[5:6], v37
; TONGA-NEXT: v_add_u32_e32 v5, vcc, s1, v0
; TONGA-NEXT: v_addc_u32_e32 v6, vcc, 0, v1, vcc
; TONGA-NEXT: s_movk_i32 s1, 0xf50
; TONGA-NEXT: v_add_u32_e32 v9, vcc, s1, v0
; TONGA-NEXT: flat_store_dword v[5:6], v36
; TONGA-NEXT: v_addc_u32_e32 v10, vcc, 0, v1, vcc
; TONGA-NEXT: v_mov_b32_e32 v5, s0
; TONGA-NEXT: s_movk_i32 s1, 0x1260
; TONGA-NEXT: flat_store_dwordx2 v[9:10], v[4:5]
; TONGA-NEXT: v_add_u32_e32 v5, vcc, s1, v0
; TONGA-NEXT: v_addc_u32_e32 v6, vcc, 0, v1, vcc
; TONGA-NEXT: v_mov_b32_e32 v4, s0
; TONGA-NEXT: s_movk_i32 s0, 0x1570
; TONGA-NEXT: v_add_u32_e32 v2, vcc, s0, v0
; TONGA-NEXT: flat_store_dwordx2 v[5:6], v[3:4]
; TONGA-NEXT: v_addc_u32_e32 v3, vcc, 0, v1, vcc
; TONGA-NEXT: s_movk_i32 s0, 0x1880
; TONGA-NEXT: flat_store_dword v[2:3], v35
; TONGA-NEXT: v_add_u32_e32 v2, vcc, s0, v0
; TONGA-NEXT: v_addc_u32_e32 v3, vcc, 0, v1, vcc
; TONGA-NEXT: s_movk_i32 s0, 0x18b8
; TONGA-NEXT: flat_store_dword v[2:3], v34
; TONGA-NEXT: v_add_u32_e32 v2, vcc, s0, v0
; TONGA-NEXT: v_addc_u32_e32 v3, vcc, 0, v1, vcc
; TONGA-NEXT: s_movk_i32 s0, 0x1b90
; TONGA-NEXT: flat_store_dwordx2 v[2:3], v[7:8]
; TONGA-NEXT: v_add_u32_e32 v2, vcc, s0, v0
; TONGA-NEXT: v_addc_u32_e32 v3, vcc, 0, v1, vcc
; TONGA-NEXT: s_movk_i32 s0, 0x1ea0
; TONGA-NEXT: flat_store_dword v[2:3], v33
; TONGA-NEXT: v_add_u32_e32 v2, vcc, s0, v0
; TONGA-NEXT: v_addc_u32_e32 v3, vcc, 0, v1, vcc
; TONGA-NEXT: s_movk_i32 s0, 0x21b0
; TONGA-NEXT: flat_store_dword v[0:1], v40
; TONGA-NEXT: flat_store_dword v[2:3], v32
; TONGA-NEXT: v_add_u32_e32 v2, vcc, s0, v0
; TONGA-NEXT: v_addc_u32_e32 v3, vcc, 0, v1, vcc
; TONGA-NEXT: s_movk_i32 s0, 0x24c0
; TONGA-NEXT: flat_store_dword v[2:3], v31
; TONGA-NEXT: v_add_u32_e32 v2, vcc, s0, v0
; TONGA-NEXT: v_addc_u32_e32 v3, vcc, 0, v1, vcc
; TONGA-NEXT: s_movk_i32 s0, 0x27d0
; TONGA-NEXT: flat_store_dword v[2:3], v30
; TONGA-NEXT: v_add_u32_e32 v2, vcc, s0, v0
; TONGA-NEXT: v_addc_u32_e32 v3, vcc, 0, v1, vcc
; TONGA-NEXT: s_movk_i32 s0, 0x2ae0
; TONGA-NEXT: flat_store_dword v[2:3], v29
; TONGA-NEXT: v_add_u32_e32 v2, vcc, s0, v0
; TONGA-NEXT: v_addc_u32_e32 v3, vcc, 0, v1, vcc
; TONGA-NEXT: s_movk_i32 s0, 0x2df0
; TONGA-NEXT: flat_store_dword v[2:3], v28
; TONGA-NEXT: v_add_u32_e32 v2, vcc, s0, v0
; TONGA-NEXT: v_addc_u32_e32 v3, vcc, 0, v1, vcc
; TONGA-NEXT: s_movk_i32 s0, 0x3100
; TONGA-NEXT: flat_store_dword v[2:3], v27
; TONGA-NEXT: v_add_u32_e32 v2, vcc, s0, v0
; TONGA-NEXT: v_addc_u32_e32 v3, vcc, 0, v1, vcc
; TONGA-NEXT: s_movk_i32 s0, 0x3410
; TONGA-NEXT: flat_store_dword v[2:3], v26
; TONGA-NEXT: v_add_u32_e32 v2, vcc, s0, v0
; TONGA-NEXT: v_addc_u32_e32 v3, vcc, 0, v1, vcc
; TONGA-NEXT: s_movk_i32 s0, 0x3720
; TONGA-NEXT: flat_store_dword v[2:3], v25
; TONGA-NEXT: v_add_u32_e32 v2, vcc, s0, v0
; TONGA-NEXT: v_addc_u32_e32 v3, vcc, 0, v1, vcc
; TONGA-NEXT: s_movk_i32 s0, 0x3a30
; TONGA-NEXT: flat_store_dword v[2:3], v24
; TONGA-NEXT: v_add_u32_e32 v2, vcc, s0, v0
; TONGA-NEXT: v_addc_u32_e32 v3, vcc, 0, v1, vcc
; TONGA-NEXT: s_movk_i32 s0, 0x3d40
; TONGA-NEXT: flat_store_dword v[2:3], v23
; TONGA-NEXT: v_add_u32_e32 v2, vcc, s0, v0
; TONGA-NEXT: v_addc_u32_e32 v3, vcc, 0, v1, vcc
; TONGA-NEXT: s_movk_i32 s0, 0x4050
; TONGA-NEXT: flat_store_dword v[2:3], v22
; TONGA-NEXT: v_add_u32_e32 v2, vcc, s0, v0
; TONGA-NEXT: v_addc_u32_e32 v3, vcc, 0, v1, vcc
; TONGA-NEXT: s_movk_i32 s0, 0x4360
; TONGA-NEXT: flat_store_dword v[2:3], v21
; TONGA-NEXT: v_add_u32_e32 v2, vcc, s0, v0
; TONGA-NEXT: v_addc_u32_e32 v3, vcc, 0, v1, vcc
; TONGA-NEXT: s_movk_i32 s0, 0x4670
; TONGA-NEXT: flat_store_dword v[2:3], v20
; TONGA-NEXT: v_add_u32_e32 v2, vcc, s0, v0
; TONGA-NEXT: v_addc_u32_e32 v3, vcc, 0, v1, vcc
; TONGA-NEXT: s_movk_i32 s0, 0x4980
; TONGA-NEXT: flat_store_dword v[2:3], v19
; TONGA-NEXT: v_add_u32_e32 v2, vcc, s0, v0
; TONGA-NEXT: v_addc_u32_e32 v3, vcc, 0, v1, vcc
; TONGA-NEXT: s_movk_i32 s0, 0x4c90
; TONGA-NEXT: flat_store_dword v[2:3], v18
; TONGA-NEXT: v_add_u32_e32 v2, vcc, s0, v0
; TONGA-NEXT: v_addc_u32_e32 v3, vcc, 0, v1, vcc
; TONGA-NEXT: s_movk_i32 s0, 0x4fa0
; TONGA-NEXT: flat_store_dword v[2:3], v15
; TONGA-NEXT: v_add_u32_e32 v2, vcc, s0, v0
; TONGA-NEXT: v_addc_u32_e32 v3, vcc, 0, v1, vcc
; TONGA-NEXT: s_movk_i32 s0, 0x52b0
; TONGA-NEXT: flat_store_dword v[2:3], v13
; TONGA-NEXT: v_add_u32_e32 v2, vcc, s0, v0
; TONGA-NEXT: v_addc_u32_e32 v3, vcc, 0, v1, vcc
; TONGA-NEXT: s_movk_i32 s0, 0x55c0
; TONGA-NEXT: flat_store_dword v[2:3], v11
; TONGA-NEXT: v_add_u32_e32 v2, vcc, s0, v0
; TONGA-NEXT: v_addc_u32_e32 v3, vcc, 0, v1, vcc
; TONGA-NEXT: s_movk_i32 s0, 0x58d0
; TONGA-NEXT: flat_store_dword v[2:3], v17
; TONGA-NEXT: v_add_u32_e32 v2, vcc, s0, v0
; TONGA-NEXT: v_addc_u32_e32 v3, vcc, 0, v1, vcc
; TONGA-NEXT: flat_store_dword v[2:3], v16
; TONGA-NEXT: v_add_u32_e32 v2, vcc, 0x5be0, v0
; TONGA-NEXT: v_addc_u32_e32 v3, vcc, 0, v1, vcc
; TONGA-NEXT: v_add_u32_e32 v0, vcc, 0x5ef0, v0
; TONGA-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
; TONGA-NEXT: flat_store_dword v[2:3], v14
; TONGA-NEXT: flat_store_dword v[0:1], v12
; TONGA-NEXT: s_endpgm
;
; TONGA-GCNTRACKERS-LABEL: excess_soft_clause_reg_pressure:
; TONGA-GCNTRACKERS: ; %bb.0: ; %entry
; TONGA-GCNTRACKERS-NEXT: s_load_dwordx8 s[0:7], s[4:5], 0x24
; TONGA-GCNTRACKERS-NEXT: ; implicit-def: $vgpr47 : SGPR spill to VGPR lane
; TONGA-GCNTRACKERS-NEXT: v_lshlrev_b32_e32 v1, 2, v0
; TONGA-GCNTRACKERS-NEXT: s_mov_b32 s33, s8
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v3, 0
; TONGA-GCNTRACKERS-NEXT: s_movk_i32 s34, 0x200
; TONGA-GCNTRACKERS-NEXT: s_waitcnt lgkmcnt(0)
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v2, s5
; TONGA-GCNTRACKERS-NEXT: v_writelane_b32 v47, s0, 0
; TONGA-GCNTRACKERS-NEXT: v_writelane_b32 v47, s1, 1
; TONGA-GCNTRACKERS-NEXT: v_writelane_b32 v47, s2, 2
; TONGA-GCNTRACKERS-NEXT: v_writelane_b32 v47, s3, 3
; TONGA-GCNTRACKERS-NEXT: v_writelane_b32 v47, s4, 4
; TONGA-GCNTRACKERS-NEXT: v_writelane_b32 v47, s5, 5
; TONGA-GCNTRACKERS-NEXT: v_add_u32_e32 v1, vcc, s4, v1
; TONGA-GCNTRACKERS-NEXT: v_writelane_b32 v47, s6, 6
; TONGA-GCNTRACKERS-NEXT: v_addc_u32_e32 v2, vcc, 0, v2, vcc
; TONGA-GCNTRACKERS-NEXT: s_movk_i32 s35, 0xc4
; TONGA-GCNTRACKERS-NEXT: s_movk_i32 s84, 0x188
; TONGA-GCNTRACKERS-NEXT: s_movk_i32 s85, 0x24c
; TONGA-GCNTRACKERS-NEXT: s_movk_i32 s86, 0x310
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v11, 0
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v13, 0
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v16, 0
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v18, 0
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v19, 0
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v20, 0
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v21, 0
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v22, 0
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v23, 0
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v24, 0
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v25, 0
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v26, 0
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v27, 0
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v28, 0
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v29, 0
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v30, 0
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v31, 0
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v32, 0
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v33, 0
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v34, 0
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v35, 0
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v4, 0
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v36, 0
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v37, 0
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v38, 0
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v39, 0
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v40, 0
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v17, 0
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v15, 0
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v14, 0
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v12, 0
; TONGA-GCNTRACKERS-NEXT: v_writelane_b32 v47, s7, 7
; TONGA-GCNTRACKERS-NEXT: .LBB4_1: ; %for.cond28.preheader
; TONGA-GCNTRACKERS-NEXT: ; =>This Inner Loop Header: Depth=1
; TONGA-GCNTRACKERS-NEXT: flat_load_dword v41, v[1:2]
; TONGA-GCNTRACKERS-NEXT: v_add_u32_e32 v5, vcc, s35, v1
; TONGA-GCNTRACKERS-NEXT: v_addc_u32_e32 v6, vcc, 0, v2, vcc
; TONGA-GCNTRACKERS-NEXT: v_add_u32_e32 v7, vcc, s84, v1
; TONGA-GCNTRACKERS-NEXT: v_addc_u32_e32 v8, vcc, 0, v2, vcc
; TONGA-GCNTRACKERS-NEXT: v_add_u32_e32 v9, vcc, s85, v1
; TONGA-GCNTRACKERS-NEXT: v_addc_u32_e32 v10, vcc, 0, v2, vcc
; TONGA-GCNTRACKERS-NEXT: flat_load_dword v42, v[5:6]
; TONGA-GCNTRACKERS-NEXT: flat_load_dword v6, v[7:8]
; TONGA-GCNTRACKERS-NEXT: flat_load_dword v5, v[9:10]
; TONGA-GCNTRACKERS-NEXT: v_readlane_b32 s68, v47, 0
; TONGA-GCNTRACKERS-NEXT: v_readlane_b32 s69, v47, 1
; TONGA-GCNTRACKERS-NEXT: s_load_dwordx16 s[12:27], s[68:69], 0x0
; TONGA-GCNTRACKERS-NEXT: s_load_dwordx16 s[36:51], s[68:69], 0x40
; TONGA-GCNTRACKERS-NEXT: v_readlane_b32 s70, v47, 2
; TONGA-GCNTRACKERS-NEXT: v_readlane_b32 s71, v47, 3
; TONGA-GCNTRACKERS-NEXT: v_readlane_b32 s72, v47, 4
; TONGA-GCNTRACKERS-NEXT: v_readlane_b32 s73, v47, 5
; TONGA-GCNTRACKERS-NEXT: v_readlane_b32 s74, v47, 6
; TONGA-GCNTRACKERS-NEXT: v_readlane_b32 s75, v47, 7
; TONGA-GCNTRACKERS-NEXT: s_load_dwordx16 s[52:67], s[68:69], 0x2040
; TONGA-GCNTRACKERS-NEXT: v_add_u32_e32 v1, vcc, s86, v1
; TONGA-GCNTRACKERS-NEXT: v_addc_u32_e32 v2, vcc, 0, v2, vcc
; TONGA-GCNTRACKERS-NEXT: s_waitcnt vmcnt(3) lgkmcnt(0)
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v7, s12, v41
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v8, s13, v41
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v9, s14, v41
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v10, s15, v41
; TONGA-GCNTRACKERS-NEXT: s_load_dwordx16 s[0:15], s[68:69], 0x2000
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v7, v7, v40
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v40, s17, v41
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v43, s16, v41
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v8, v8, v39
; TONGA-GCNTRACKERS-NEXT: s_waitcnt lgkmcnt(0)
; TONGA-GCNTRACKERS-NEXT: v_writelane_b32 v47, s0, 8
; TONGA-GCNTRACKERS-NEXT: v_writelane_b32 v47, s1, 9
; TONGA-GCNTRACKERS-NEXT: v_writelane_b32 v47, s2, 10
; TONGA-GCNTRACKERS-NEXT: v_writelane_b32 v47, s3, 11
; TONGA-GCNTRACKERS-NEXT: v_writelane_b32 v47, s4, 12
; TONGA-GCNTRACKERS-NEXT: v_writelane_b32 v47, s5, 13
; TONGA-GCNTRACKERS-NEXT: v_writelane_b32 v47, s6, 14
; TONGA-GCNTRACKERS-NEXT: v_writelane_b32 v47, s7, 15
; TONGA-GCNTRACKERS-NEXT: v_writelane_b32 v47, s8, 16
; TONGA-GCNTRACKERS-NEXT: v_writelane_b32 v47, s9, 17
; TONGA-GCNTRACKERS-NEXT: v_writelane_b32 v47, s10, 18
; TONGA-GCNTRACKERS-NEXT: v_writelane_b32 v47, s11, 19
; TONGA-GCNTRACKERS-NEXT: v_writelane_b32 v47, s12, 20
; TONGA-GCNTRACKERS-NEXT: v_writelane_b32 v47, s13, 21
; TONGA-GCNTRACKERS-NEXT: v_writelane_b32 v47, s14, 22
; TONGA-GCNTRACKERS-NEXT: v_writelane_b32 v47, s15, 23
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v39, s18, v41
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v9, v9, v38
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v38, s19, v41
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v10, v10, v37
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v37, s20, v41
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v4, v40, v4
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v40, s22, v41
; TONGA-GCNTRACKERS-NEXT: s_load_dwordx16 s[0:15], s[68:69], 0x1000
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v36, v43, v36
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v43, s21, v41
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v3, v39, v3
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v39, s23, v41
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v35, v38, v35
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v38, s24, v41
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v34, v37, v34
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v37, s25, v41
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v32, v40, v32
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v40, s27, v41
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v33, v43, v33
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v43, s26, v41
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v31, v39, v31
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v39, s36, v41
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v30, v38, v30
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v38, s37, v41
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v29, v37, v29
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v37, s38, v41
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v27, v40, v27
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v40, s40, v41
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v28, v43, v28
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v43, s39, v41
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v26, v39, v26
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v39, s41, v41
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v25, v38, v25
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v38, s42, v41
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v24, v37, v24
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v37, s43, v41
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v22, v40, v22
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v40, s45, v41
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v23, v43, v23
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v43, s44, v41
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v21, v39, v21
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v39, s46, v41
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v20, v38, v20
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v38, s47, v41
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v19, v37, v19
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v37, s48, v41
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v16, v40, v16
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v40, s50, v41
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v18, v43, v18
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v43, s49, v41
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v41, s51, v41
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v13, v39, v13
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v11, v38, v11
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v17, v37, v17
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v14, v40, v14
; TONGA-GCNTRACKERS-NEXT: s_waitcnt vmcnt(2) lgkmcnt(0)
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v37, s0, v42
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v38, s1, v42
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v39, s2, v42
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v40, s3, v42
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v15, v43, v15
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v12, v41, v12
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v41, s4, v42
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v43, s5, v42
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v7, v37, v7
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v37, s6, v42
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v8, v38, v8
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v38, s7, v42
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v9, v39, v9
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v39, s8, v42
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v10, v40, v10
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v40, s9, v42
; TONGA-GCNTRACKERS-NEXT: s_load_dwordx16 s[36:51], s[68:69], 0x3000
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v36, v41, v36
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v41, s10, v42
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v4, v43, v4
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v43, s11, v42
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v3, v37, v3
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v37, s12, v42
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v35, v38, v35
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v38, s13, v42
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v34, v39, v34
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v39, s14, v42
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v33, v40, v33
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v40, s15, v42
; TONGA-GCNTRACKERS-NEXT: s_load_dwordx16 s[0:15], s[68:69], 0x1040
; TONGA-GCNTRACKERS-NEXT: s_load_dwordx16 s[16:31], s[68:69], 0x3040
; TONGA-GCNTRACKERS-NEXT: s_add_u32 s68, s68, 0x4000
; TONGA-GCNTRACKERS-NEXT: s_addc_u32 s69, s69, 0
; TONGA-GCNTRACKERS-NEXT: v_writelane_b32 v47, s68, 0
; TONGA-GCNTRACKERS-NEXT: v_writelane_b32 v47, s69, 1
; TONGA-GCNTRACKERS-NEXT: v_writelane_b32 v47, s70, 2
; TONGA-GCNTRACKERS-NEXT: v_writelane_b32 v47, s71, 3
; TONGA-GCNTRACKERS-NEXT: v_writelane_b32 v47, s72, 4
; TONGA-GCNTRACKERS-NEXT: v_writelane_b32 v47, s73, 5
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v32, v41, v32
; TONGA-GCNTRACKERS-NEXT: s_waitcnt lgkmcnt(0)
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v41, s0, v42
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v31, v43, v31
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v43, s1, v42
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v30, v37, v30
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v37, s2, v42
; TONGA-GCNTRACKERS-NEXT: v_writelane_b32 v47, s74, 6
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v29, v38, v29
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v38, s3, v42
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v28, v39, v28
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v39, s4, v42
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v27, v40, v27
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v40, s5, v42
; TONGA-GCNTRACKERS-NEXT: v_writelane_b32 v47, s75, 7
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v26, v41, v26
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v41, s6, v42
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v25, v43, v25
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v43, s7, v42
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v24, v37, v24
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v37, s8, v42
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v23, v38, v23
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v38, s9, v42
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v22, v39, v22
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v39, s10, v42
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v21, v40, v21
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v40, s11, v42
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v20, v41, v20
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v41, s12, v42
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v19, v43, v19
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v43, s13, v42
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v18, v37, v18
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v37, s14, v42
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v42, s15, v42
; TONGA-GCNTRACKERS-NEXT: v_readlane_b32 s0, v47, 8
; TONGA-GCNTRACKERS-NEXT: v_readlane_b32 s1, v47, 9
; TONGA-GCNTRACKERS-NEXT: v_readlane_b32 s2, v47, 10
; TONGA-GCNTRACKERS-NEXT: v_readlane_b32 s3, v47, 11
; TONGA-GCNTRACKERS-NEXT: v_readlane_b32 s4, v47, 12
; TONGA-GCNTRACKERS-NEXT: v_readlane_b32 s5, v47, 13
; TONGA-GCNTRACKERS-NEXT: v_readlane_b32 s6, v47, 14
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v16, v38, v16
; TONGA-GCNTRACKERS-NEXT: v_readlane_b32 s7, v47, 15
; TONGA-GCNTRACKERS-NEXT: v_readlane_b32 s8, v47, 16
; TONGA-GCNTRACKERS-NEXT: v_readlane_b32 s9, v47, 17
; TONGA-GCNTRACKERS-NEXT: v_readlane_b32 s10, v47, 18
; TONGA-GCNTRACKERS-NEXT: v_readlane_b32 s11, v47, 19
; TONGA-GCNTRACKERS-NEXT: v_readlane_b32 s12, v47, 20
; TONGA-GCNTRACKERS-NEXT: v_readlane_b32 s13, v47, 21
; TONGA-GCNTRACKERS-NEXT: s_waitcnt vmcnt(1)
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v38, s0, v6
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v13, v39, v13
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v39, s1, v6
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v11, v40, v11
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v40, s2, v6
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v17, v41, v17
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v41, s3, v6
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v15, v43, v15
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v43, s4, v6
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v14, v37, v14
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v37, s5, v6
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v12, v42, v12
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v42, s6, v6
; TONGA-GCNTRACKERS-NEXT: v_readlane_b32 s14, v47, 22
; TONGA-GCNTRACKERS-NEXT: v_readlane_b32 s15, v47, 23
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v7, v38, v7
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v38, s7, v6
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v8, v39, v8
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v39, s8, v6
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v9, v40, v9
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v40, s9, v6
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v10, v41, v10
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v41, s10, v6
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v36, v43, v36
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v43, s11, v6
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v4, v37, v4
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v37, s12, v6
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v3, v42, v3
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v42, s13, v6
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v35, v38, v35
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v38, s14, v6
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v34, v39, v34
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v39, s15, v6
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v33, v40, v33
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v40, s52, v6
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v32, v41, v32
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v41, s53, v6
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v31, v43, v31
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v43, s54, v6
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v30, v37, v30
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v37, s55, v6
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v29, v42, v29
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v42, s56, v6
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v28, v38, v28
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v38, s57, v6
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v27, v39, v27
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v39, s58, v6
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v26, v40, v26
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v40, s59, v6
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v25, v41, v25
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v41, s60, v6
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v24, v43, v24
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v43, s61, v6
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v23, v37, v23
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v37, s62, v6
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v22, v42, v22
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v42, s63, v6
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v21, v38, v21
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v38, s64, v6
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v20, v39, v20
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v39, s65, v6
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v19, v40, v19
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v40, s66, v6
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v6, s67, v6
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v18, v41, v18
; TONGA-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0)
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v41, s36, v5
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v16, v43, v16
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v43, s37, v5
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v13, v37, v13
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v37, s38, v5
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v11, v42, v11
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v42, s39, v5
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v17, v38, v17
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v15, v39, v15
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v14, v40, v14
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v6, v6, v12
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v12, s43, v5
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v40, v41, v7
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v7, s44, v5
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v39, v43, v8
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v8, s45, v5
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v38, v37, v9
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v9, s46, v5
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v37, v42, v10
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v10, s47, v5
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v41, s48, v5
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v42, s49, v5
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v43, s50, v5
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v35, v12, v35
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v12, s51, v5
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v34, v7, v34
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v7, s16, v5
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v33, v8, v33
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v8, s17, v5
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v32, v9, v32
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v9, s18, v5
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v31, v10, v31
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v10, s19, v5
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v30, v41, v30
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v41, s20, v5
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v29, v42, v29
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v42, s21, v5
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v28, v43, v28
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v43, s22, v5
; TONGA-GCNTRACKERS-NEXT: s_add_i32 s34, s34, -1
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v44, s40, v5
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v45, s41, v5
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v46, s42, v5
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v27, v12, v27
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v12, s23, v5
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v26, v7, v26
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v7, s24, v5
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v25, v8, v25
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v8, s25, v5
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v24, v9, v24
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v9, s26, v5
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v23, v10, v23
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v10, s27, v5
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v22, v41, v22
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v41, s28, v5
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v21, v42, v21
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v42, s29, v5
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v20, v43, v20
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v43, s30, v5
; TONGA-GCNTRACKERS-NEXT: v_mul_f32_e32 v5, s31, v5
; TONGA-GCNTRACKERS-NEXT: s_cmp_lg_u32 s34, 0
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v36, v44, v36
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v4, v45, v4
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v3, v46, v3
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v19, v12, v19
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v18, v7, v18
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v16, v8, v16
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v13, v9, v13
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v11, v10, v11
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v17, v41, v17
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v15, v42, v15
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v14, v43, v14
; TONGA-GCNTRACKERS-NEXT: v_add_f32_e32 v12, v5, v6
; TONGA-GCNTRACKERS-NEXT: s_cbranch_scc1 .LBB4_1
; TONGA-GCNTRACKERS-NEXT: ; %bb.2: ; %for.cond.cleanup26
; TONGA-GCNTRACKERS-NEXT: v_readlane_b32 s4, v47, 0
; TONGA-GCNTRACKERS-NEXT: s_lshl_b32 s0, s33, 8
; TONGA-GCNTRACKERS-NEXT: v_readlane_b32 s10, v47, 6
; TONGA-GCNTRACKERS-NEXT: s_add_i32 s0, s10, s0
; TONGA-GCNTRACKERS-NEXT: v_add_u32_e32 v0, vcc, s0, v0
; TONGA-GCNTRACKERS-NEXT: s_mov_b32 s0, 0x5397829d
; TONGA-GCNTRACKERS-NEXT: v_mul_hi_u32 v0, v0, s0
; TONGA-GCNTRACKERS-NEXT: s_mov_b32 s0, 0x31000
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v1, 0
; TONGA-GCNTRACKERS-NEXT: v_readlane_b32 s6, v47, 2
; TONGA-GCNTRACKERS-NEXT: v_lshrrev_b32_e32 v0, 4, v0
; TONGA-GCNTRACKERS-NEXT: v_mul_lo_u32 v0, v0, s0
; TONGA-GCNTRACKERS-NEXT: v_readlane_b32 s7, v47, 3
; TONGA-GCNTRACKERS-NEXT: s_mov_b64 s[2:3], s[6:7]
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v2, s3
; TONGA-GCNTRACKERS-NEXT: v_lshlrev_b64 v[0:1], 2, v[0:1]
; TONGA-GCNTRACKERS-NEXT: s_movk_i32 s0, 0x310
; TONGA-GCNTRACKERS-NEXT: v_add_u32_e32 v0, vcc, s2, v0
; TONGA-GCNTRACKERS-NEXT: v_addc_u32_e32 v1, vcc, v2, v1, vcc
; TONGA-GCNTRACKERS-NEXT: v_add_u32_e32 v5, vcc, s0, v0
; TONGA-GCNTRACKERS-NEXT: v_addc_u32_e32 v6, vcc, 0, v1, vcc
; TONGA-GCNTRACKERS-NEXT: s_movk_i32 s0, 0x348
; TONGA-GCNTRACKERS-NEXT: flat_store_dword v[5:6], v39
; TONGA-GCNTRACKERS-NEXT: v_add_u32_e32 v5, vcc, s0, v0
; TONGA-GCNTRACKERS-NEXT: s_mov_b32 s0, 0
; TONGA-GCNTRACKERS-NEXT: s_mov_b32 s1, s0
; TONGA-GCNTRACKERS-NEXT: v_addc_u32_e32 v6, vcc, 0, v1, vcc
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v8, s1
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v7, s0
; TONGA-GCNTRACKERS-NEXT: s_movk_i32 s1, 0x620
; TONGA-GCNTRACKERS-NEXT: flat_store_dwordx2 v[5:6], v[7:8]
; TONGA-GCNTRACKERS-NEXT: v_add_u32_e32 v5, vcc, s1, v0
; TONGA-GCNTRACKERS-NEXT: v_addc_u32_e32 v6, vcc, 0, v1, vcc
; TONGA-GCNTRACKERS-NEXT: s_movk_i32 s1, 0x930
; TONGA-GCNTRACKERS-NEXT: flat_store_dword v[5:6], v38
; TONGA-GCNTRACKERS-NEXT: v_add_u32_e32 v5, vcc, s1, v0
; TONGA-GCNTRACKERS-NEXT: v_addc_u32_e32 v6, vcc, 0, v1, vcc
; TONGA-GCNTRACKERS-NEXT: s_movk_i32 s1, 0xc40
; TONGA-GCNTRACKERS-NEXT: flat_store_dword v[5:6], v37
; TONGA-GCNTRACKERS-NEXT: v_add_u32_e32 v5, vcc, s1, v0
; TONGA-GCNTRACKERS-NEXT: v_addc_u32_e32 v6, vcc, 0, v1, vcc
; TONGA-GCNTRACKERS-NEXT: s_movk_i32 s1, 0xf50
; TONGA-GCNTRACKERS-NEXT: v_add_u32_e32 v9, vcc, s1, v0
; TONGA-GCNTRACKERS-NEXT: flat_store_dword v[5:6], v36
; TONGA-GCNTRACKERS-NEXT: v_addc_u32_e32 v10, vcc, 0, v1, vcc
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v5, s0
; TONGA-GCNTRACKERS-NEXT: s_movk_i32 s1, 0x1260
; TONGA-GCNTRACKERS-NEXT: flat_store_dwordx2 v[9:10], v[4:5]
; TONGA-GCNTRACKERS-NEXT: v_add_u32_e32 v5, vcc, s1, v0
; TONGA-GCNTRACKERS-NEXT: v_addc_u32_e32 v6, vcc, 0, v1, vcc
; TONGA-GCNTRACKERS-NEXT: v_mov_b32_e32 v4, s0
; TONGA-GCNTRACKERS-NEXT: s_movk_i32 s0, 0x1570
; TONGA-GCNTRACKERS-NEXT: v_add_u32_e32 v2, vcc, s0, v0
; TONGA-GCNTRACKERS-NEXT: flat_store_dwordx2 v[5:6], v[3:4]
; TONGA-GCNTRACKERS-NEXT: v_addc_u32_e32 v3, vcc, 0, v1, vcc
; TONGA-GCNTRACKERS-NEXT: s_movk_i32 s0, 0x1880
; TONGA-GCNTRACKERS-NEXT: flat_store_dword v[2:3], v35
; TONGA-GCNTRACKERS-NEXT: v_add_u32_e32 v2, vcc, s0, v0
; TONGA-GCNTRACKERS-NEXT: v_addc_u32_e32 v3, vcc, 0, v1, vcc
; TONGA-GCNTRACKERS-NEXT: s_movk_i32 s0, 0x18b8
; TONGA-GCNTRACKERS-NEXT: flat_store_dword v[2:3], v34
; TONGA-GCNTRACKERS-NEXT: v_add_u32_e32 v2, vcc, s0, v0
; TONGA-GCNTRACKERS-NEXT: v_addc_u32_e32 v3, vcc, 0, v1, vcc
; TONGA-GCNTRACKERS-NEXT: s_movk_i32 s0, 0x1b90
; TONGA-GCNTRACKERS-NEXT: flat_store_dwordx2 v[2:3], v[7:8]
; TONGA-GCNTRACKERS-NEXT: v_add_u32_e32 v2, vcc, s0, v0
; TONGA-GCNTRACKERS-NEXT: v_addc_u32_e32 v3, vcc, 0, v1, vcc
; TONGA-GCNTRACKERS-NEXT: s_movk_i32 s0, 0x1ea0
; TONGA-GCNTRACKERS-NEXT: flat_store_dword v[2:3], v33
; TONGA-GCNTRACKERS-NEXT: v_add_u32_e32 v2, vcc, s0, v0
; TONGA-GCNTRACKERS-NEXT: v_addc_u32_e32 v3, vcc, 0, v1, vcc
; TONGA-GCNTRACKERS-NEXT: s_movk_i32 s0, 0x21b0
; TONGA-GCNTRACKERS-NEXT: flat_store_dword v[0:1], v40
; TONGA-GCNTRACKERS-NEXT: flat_store_dword v[2:3], v32
; TONGA-GCNTRACKERS-NEXT: v_add_u32_e32 v2, vcc, s0, v0
; TONGA-GCNTRACKERS-NEXT: v_addc_u32_e32 v3, vcc, 0, v1, vcc
; TONGA-GCNTRACKERS-NEXT: s_movk_i32 s0, 0x24c0
; TONGA-GCNTRACKERS-NEXT: flat_store_dword v[2:3], v31
; TONGA-GCNTRACKERS-NEXT: v_add_u32_e32 v2, vcc, s0, v0
; TONGA-GCNTRACKERS-NEXT: v_addc_u32_e32 v3, vcc, 0, v1, vcc
; TONGA-GCNTRACKERS-NEXT: s_movk_i32 s0, 0x27d0
; TONGA-GCNTRACKERS-NEXT: flat_store_dword v[2:3], v30
; TONGA-GCNTRACKERS-NEXT: v_add_u32_e32 v2, vcc, s0, v0
; TONGA-GCNTRACKERS-NEXT: v_addc_u32_e32 v3, vcc, 0, v1, vcc
; TONGA-GCNTRACKERS-NEXT: s_movk_i32 s0, 0x2ae0
; TONGA-GCNTRACKERS-NEXT: flat_store_dword v[2:3], v29
; TONGA-GCNTRACKERS-NEXT: v_add_u32_e32 v2, vcc, s0, v0
; TONGA-GCNTRACKERS-NEXT: v_addc_u32_e32 v3, vcc, 0, v1, vcc
; TONGA-GCNTRACKERS-NEXT: s_movk_i32 s0, 0x2df0
; TONGA-GCNTRACKERS-NEXT: flat_store_dword v[2:3], v28
; TONGA-GCNTRACKERS-NEXT: v_add_u32_e32 v2, vcc, s0, v0
; TONGA-GCNTRACKERS-NEXT: v_addc_u32_e32 v3, vcc, 0, v1, vcc
; TONGA-GCNTRACKERS-NEXT: s_movk_i32 s0, 0x3100
; TONGA-GCNTRACKERS-NEXT: flat_store_dword v[2:3], v27
; TONGA-GCNTRACKERS-NEXT: v_add_u32_e32 v2, vcc, s0, v0
; TONGA-GCNTRACKERS-NEXT: v_addc_u32_e32 v3, vcc, 0, v1, vcc
; TONGA-GCNTRACKERS-NEXT: s_movk_i32 s0, 0x3410
; TONGA-GCNTRACKERS-NEXT: flat_store_dword v[2:3], v26
; TONGA-GCNTRACKERS-NEXT: v_add_u32_e32 v2, vcc, s0, v0
; TONGA-GCNTRACKERS-NEXT: v_addc_u32_e32 v3, vcc, 0, v1, vcc
; TONGA-GCNTRACKERS-NEXT: s_movk_i32 s0, 0x3720
; TONGA-GCNTRACKERS-NEXT: flat_store_dword v[2:3], v25
; TONGA-GCNTRACKERS-NEXT: v_add_u32_e32 v2, vcc, s0, v0
; TONGA-GCNTRACKERS-NEXT: v_addc_u32_e32 v3, vcc, 0, v1, vcc
; TONGA-GCNTRACKERS-NEXT: s_movk_i32 s0, 0x3a30
; TONGA-GCNTRACKERS-NEXT: flat_store_dword v[2:3], v24
; TONGA-GCNTRACKERS-NEXT: v_add_u32_e32 v2, vcc, s0, v0
; TONGA-GCNTRACKERS-NEXT: v_addc_u32_e32 v3, vcc, 0, v1, vcc
; TONGA-GCNTRACKERS-NEXT: s_movk_i32 s0, 0x3d40
; TONGA-GCNTRACKERS-NEXT: flat_store_dword v[2:3], v23
; TONGA-GCNTRACKERS-NEXT: v_add_u32_e32 v2, vcc, s0, v0
; TONGA-GCNTRACKERS-NEXT: v_addc_u32_e32 v3, vcc, 0, v1, vcc
; TONGA-GCNTRACKERS-NEXT: s_movk_i32 s0, 0x4050
; TONGA-GCNTRACKERS-NEXT: flat_store_dword v[2:3], v22
; TONGA-GCNTRACKERS-NEXT: v_add_u32_e32 v2, vcc, s0, v0
; TONGA-GCNTRACKERS-NEXT: v_addc_u32_e32 v3, vcc, 0, v1, vcc
; TONGA-GCNTRACKERS-NEXT: s_movk_i32 s0, 0x4360
; TONGA-GCNTRACKERS-NEXT: flat_store_dword v[2:3], v21
; TONGA-GCNTRACKERS-NEXT: v_add_u32_e32 v2, vcc, s0, v0
; TONGA-GCNTRACKERS-NEXT: v_addc_u32_e32 v3, vcc, 0, v1, vcc
; TONGA-GCNTRACKERS-NEXT: s_movk_i32 s0, 0x4670
; TONGA-GCNTRACKERS-NEXT: flat_store_dword v[2:3], v20
; TONGA-GCNTRACKERS-NEXT: v_add_u32_e32 v2, vcc, s0, v0
; TONGA-GCNTRACKERS-NEXT: v_addc_u32_e32 v3, vcc, 0, v1, vcc
; TONGA-GCNTRACKERS-NEXT: s_movk_i32 s0, 0x4980
; TONGA-GCNTRACKERS-NEXT: flat_store_dword v[2:3], v19
; TONGA-GCNTRACKERS-NEXT: v_add_u32_e32 v2, vcc, s0, v0
; TONGA-GCNTRACKERS-NEXT: v_addc_u32_e32 v3, vcc, 0, v1, vcc
; TONGA-GCNTRACKERS-NEXT: s_movk_i32 s0, 0x4c90
; TONGA-GCNTRACKERS-NEXT: flat_store_dword v[2:3], v18
; TONGA-GCNTRACKERS-NEXT: v_add_u32_e32 v2, vcc, s0, v0
; TONGA-GCNTRACKERS-NEXT: v_addc_u32_e32 v3, vcc, 0, v1, vcc
; TONGA-GCNTRACKERS-NEXT: s_movk_i32 s0, 0x4fa0
; TONGA-GCNTRACKERS-NEXT: flat_store_dword v[2:3], v16
; TONGA-GCNTRACKERS-NEXT: v_add_u32_e32 v2, vcc, s0, v0
; TONGA-GCNTRACKERS-NEXT: v_addc_u32_e32 v3, vcc, 0, v1, vcc
; TONGA-GCNTRACKERS-NEXT: s_movk_i32 s0, 0x52b0
; TONGA-GCNTRACKERS-NEXT: flat_store_dword v[2:3], v13
; TONGA-GCNTRACKERS-NEXT: v_add_u32_e32 v2, vcc, s0, v0
; TONGA-GCNTRACKERS-NEXT: v_addc_u32_e32 v3, vcc, 0, v1, vcc
; TONGA-GCNTRACKERS-NEXT: s_movk_i32 s0, 0x55c0
; TONGA-GCNTRACKERS-NEXT: flat_store_dword v[2:3], v11
; TONGA-GCNTRACKERS-NEXT: v_add_u32_e32 v2, vcc, s0, v0
; TONGA-GCNTRACKERS-NEXT: v_addc_u32_e32 v3, vcc, 0, v1, vcc
; TONGA-GCNTRACKERS-NEXT: s_movk_i32 s0, 0x58d0
; TONGA-GCNTRACKERS-NEXT: flat_store_dword v[2:3], v17
; TONGA-GCNTRACKERS-NEXT: v_add_u32_e32 v2, vcc, s0, v0
; TONGA-GCNTRACKERS-NEXT: v_addc_u32_e32 v3, vcc, 0, v1, vcc
; TONGA-GCNTRACKERS-NEXT: flat_store_dword v[2:3], v15
; TONGA-GCNTRACKERS-NEXT: v_add_u32_e32 v2, vcc, 0x5be0, v0
; TONGA-GCNTRACKERS-NEXT: v_addc_u32_e32 v3, vcc, 0, v1, vcc
; TONGA-GCNTRACKERS-NEXT: v_add_u32_e32 v0, vcc, 0x5ef0, v0
; TONGA-GCNTRACKERS-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
; TONGA-GCNTRACKERS-NEXT: v_readlane_b32 s5, v47, 1
; TONGA-GCNTRACKERS-NEXT: v_readlane_b32 s8, v47, 4
; TONGA-GCNTRACKERS-NEXT: v_readlane_b32 s9, v47, 5
; TONGA-GCNTRACKERS-NEXT: v_readlane_b32 s11, v47, 7
; TONGA-GCNTRACKERS-NEXT: flat_store_dword v[2:3], v14
; TONGA-GCNTRACKERS-NEXT: flat_store_dword v[0:1], v12
; TONGA-GCNTRACKERS-NEXT: s_endpgm
;
; GFX908-LABEL: excess_soft_clause_reg_pressure:
; GFX908: ; %bb.0: ; %entry
; GFX908-NEXT: s_load_dwordx8 s[52:59], s[4:5], 0x24
; GFX908-NEXT: v_lshlrev_b32_e32 v1, 2, v0
; GFX908-NEXT: s_movk_i32 s0, 0x200
; GFX908-NEXT: v_mov_b32_e32 v5, 0
; GFX908-NEXT: v_mov_b32_e32 v6, 0
; GFX908-NEXT: s_waitcnt lgkmcnt(0)
; GFX908-NEXT: v_mov_b32_e32 v2, s57
; GFX908-NEXT: v_add_co_u32_e32 v1, vcc, s56, v1
; GFX908-NEXT: v_addc_co_u32_e32 v2, vcc, 0, v2, vcc
; GFX908-NEXT: v_add_co_u32_e32 v3, vcc, 0x188, v1
; GFX908-NEXT: v_addc_co_u32_e32 v4, vcc, 0, v2, vcc
; GFX908-NEXT: v_mov_b32_e32 v7, 0
; GFX908-NEXT: v_mov_b32_e32 v8, 0
; GFX908-NEXT: v_mov_b32_e32 v9, 0
; GFX908-NEXT: v_mov_b32_e32 v14, 0
; GFX908-NEXT: v_mov_b32_e32 v15, 0
; GFX908-NEXT: v_mov_b32_e32 v16, 0
; GFX908-NEXT: v_mov_b32_e32 v17, 0
; GFX908-NEXT: v_mov_b32_e32 v18, 0
; GFX908-NEXT: v_mov_b32_e32 v19, 0
; GFX908-NEXT: v_mov_b32_e32 v20, 0
; GFX908-NEXT: v_mov_b32_e32 v21, 0
; GFX908-NEXT: v_mov_b32_e32 v22, 0
; GFX908-NEXT: v_mov_b32_e32 v23, 0
; GFX908-NEXT: v_mov_b32_e32 v24, 0
; GFX908-NEXT: v_mov_b32_e32 v25, 0
; GFX908-NEXT: v_mov_b32_e32 v26, 0
; GFX908-NEXT: v_mov_b32_e32 v27, 0
; GFX908-NEXT: v_mov_b32_e32 v28, 0
; GFX908-NEXT: v_mov_b32_e32 v29, 0
; GFX908-NEXT: v_mov_b32_e32 v30, 0
; GFX908-NEXT: v_mov_b32_e32 v31, 0
; GFX908-NEXT: v_mov_b32_e32 v32, 0
; GFX908-NEXT: v_mov_b32_e32 v33, 0
; GFX908-NEXT: v_mov_b32_e32 v34, 0
; GFX908-NEXT: v_mov_b32_e32 v10, 0
; GFX908-NEXT: v_mov_b32_e32 v11, 0
; GFX908-NEXT: v_mov_b32_e32 v12, 0
; GFX908-NEXT: v_mov_b32_e32 v13, 0
; GFX908-NEXT: v_mov_b32_e32 v1, 0
; GFX908-NEXT: v_mov_b32_e32 v2, 0
; GFX908-NEXT: .LBB4_1: ; %for.cond28.preheader
; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX908-NEXT: global_load_dword v38, v[3:4], off offset:-392
; GFX908-NEXT: global_load_dword v37, v[3:4], off offset:-196
; GFX908-NEXT: global_load_dword v36, v[3:4], off
; GFX908-NEXT: global_load_dword v35, v[3:4], off offset:196
; GFX908-NEXT: s_load_dwordx16 s[36:51], s[52:53], 0x0
; GFX908-NEXT: s_load_dwordx16 s[12:27], s[52:53], 0x40
; GFX908-NEXT: v_add_co_u32_e32 v3, vcc, 0x310, v3
; GFX908-NEXT: v_addc_co_u32_e32 v4, vcc, 0, v4, vcc
; GFX908-NEXT: s_waitcnt vmcnt(3) lgkmcnt(0)
; GFX908-NEXT: v_fmac_f32_e32 v34, s36, v38
; GFX908-NEXT: v_fmac_f32_e32 v33, s37, v38
; GFX908-NEXT: v_fmac_f32_e32 v32, s38, v38
; GFX908-NEXT: v_fmac_f32_e32 v31, s39, v38
; GFX908-NEXT: v_fmac_f32_e32 v30, s40, v38
; GFX908-NEXT: v_fmac_f32_e32 v2, s41, v38
; GFX908-NEXT: v_fmac_f32_e32 v1, s42, v38
; GFX908-NEXT: v_fmac_f32_e32 v29, s43, v38
; GFX908-NEXT: v_fmac_f32_e32 v28, s44, v38
; GFX908-NEXT: v_fmac_f32_e32 v27, s45, v38
; GFX908-NEXT: v_fmac_f32_e32 v26, s46, v38
; GFX908-NEXT: v_fmac_f32_e32 v25, s47, v38
; GFX908-NEXT: v_fmac_f32_e32 v24, s48, v38
; GFX908-NEXT: v_fmac_f32_e32 v23, s49, v38
; GFX908-NEXT: v_fmac_f32_e32 v22, s50, v38
; GFX908-NEXT: v_fmac_f32_e32 v21, s51, v38
; GFX908-NEXT: v_fmac_f32_e32 v20, s12, v38
; GFX908-NEXT: v_fmac_f32_e32 v19, s13, v38
; GFX908-NEXT: v_fmac_f32_e32 v18, s14, v38
; GFX908-NEXT: v_fmac_f32_e32 v17, s15, v38
; GFX908-NEXT: v_fmac_f32_e32 v16, s16, v38
; GFX908-NEXT: v_fmac_f32_e32 v15, s17, v38
; GFX908-NEXT: v_fmac_f32_e32 v14, s18, v38
; GFX908-NEXT: v_fmac_f32_e32 v9, s19, v38
; GFX908-NEXT: v_fmac_f32_e32 v8, s20, v38
; GFX908-NEXT: v_fmac_f32_e32 v7, s21, v38
; GFX908-NEXT: v_fmac_f32_e32 v6, s22, v38
; GFX908-NEXT: v_fmac_f32_e32 v5, s23, v38
; GFX908-NEXT: v_fmac_f32_e32 v10, s24, v38
; GFX908-NEXT: v_fmac_f32_e32 v11, s25, v38
; GFX908-NEXT: v_fmac_f32_e32 v12, s26, v38
; GFX908-NEXT: v_fmac_f32_e32 v13, s27, v38
; GFX908-NEXT: s_load_dwordx16 s[36:51], s[52:53], 0x1000
; GFX908-NEXT: s_load_dwordx16 s[12:27], s[52:53], 0x1040
; GFX908-NEXT: s_waitcnt vmcnt(2) lgkmcnt(0)
; GFX908-NEXT: v_fmac_f32_e32 v34, s36, v37
; GFX908-NEXT: v_fmac_f32_e32 v33, s37, v37
; GFX908-NEXT: v_fmac_f32_e32 v32, s38, v37
; GFX908-NEXT: v_fmac_f32_e32 v31, s39, v37
; GFX908-NEXT: v_fmac_f32_e32 v30, s40, v37
; GFX908-NEXT: v_fmac_f32_e32 v2, s41, v37
; GFX908-NEXT: v_fmac_f32_e32 v1, s42, v37
; GFX908-NEXT: v_fmac_f32_e32 v29, s43, v37
; GFX908-NEXT: v_fmac_f32_e32 v28, s44, v37
; GFX908-NEXT: v_fmac_f32_e32 v27, s45, v37
; GFX908-NEXT: v_fmac_f32_e32 v26, s46, v37
; GFX908-NEXT: v_fmac_f32_e32 v25, s47, v37
; GFX908-NEXT: v_fmac_f32_e32 v24, s48, v37
; GFX908-NEXT: v_fmac_f32_e32 v23, s49, v37
; GFX908-NEXT: v_fmac_f32_e32 v22, s50, v37
; GFX908-NEXT: v_fmac_f32_e32 v21, s51, v37
; GFX908-NEXT: v_fmac_f32_e32 v20, s12, v37
; GFX908-NEXT: v_fmac_f32_e32 v19, s13, v37
; GFX908-NEXT: v_fmac_f32_e32 v18, s14, v37
; GFX908-NEXT: v_fmac_f32_e32 v17, s15, v37
; GFX908-NEXT: v_fmac_f32_e32 v16, s16, v37
; GFX908-NEXT: v_fmac_f32_e32 v15, s17, v37
; GFX908-NEXT: v_fmac_f32_e32 v14, s18, v37
; GFX908-NEXT: v_fmac_f32_e32 v9, s19, v37
; GFX908-NEXT: v_fmac_f32_e32 v8, s20, v37
; GFX908-NEXT: v_fmac_f32_e32 v7, s21, v37
; GFX908-NEXT: v_fmac_f32_e32 v6, s22, v37
; GFX908-NEXT: v_fmac_f32_e32 v5, s23, v37
; GFX908-NEXT: v_fmac_f32_e32 v10, s24, v37
; GFX908-NEXT: v_fmac_f32_e32 v11, s25, v37
; GFX908-NEXT: v_fmac_f32_e32 v12, s26, v37
; GFX908-NEXT: v_fmac_f32_e32 v13, s27, v37
; GFX908-NEXT: s_load_dwordx16 s[36:51], s[52:53], 0x2000
; GFX908-NEXT: s_load_dwordx16 s[12:27], s[52:53], 0x2040
; GFX908-NEXT: s_waitcnt vmcnt(1) lgkmcnt(0)
; GFX908-NEXT: v_fmac_f32_e32 v34, s36, v36
; GFX908-NEXT: v_fmac_f32_e32 v33, s37, v36
; GFX908-NEXT: v_fmac_f32_e32 v32, s38, v36
; GFX908-NEXT: v_fmac_f32_e32 v31, s39, v36
; GFX908-NEXT: v_fmac_f32_e32 v30, s40, v36
; GFX908-NEXT: v_fmac_f32_e32 v2, s41, v36
; GFX908-NEXT: v_fmac_f32_e32 v1, s42, v36
; GFX908-NEXT: v_fmac_f32_e32 v29, s43, v36
; GFX908-NEXT: v_fmac_f32_e32 v28, s44, v36
; GFX908-NEXT: v_fmac_f32_e32 v27, s45, v36
; GFX908-NEXT: v_fmac_f32_e32 v26, s46, v36
; GFX908-NEXT: v_fmac_f32_e32 v25, s47, v36
; GFX908-NEXT: v_fmac_f32_e32 v24, s48, v36
; GFX908-NEXT: v_fmac_f32_e32 v23, s49, v36
; GFX908-NEXT: v_fmac_f32_e32 v22, s50, v36
; GFX908-NEXT: v_fmac_f32_e32 v21, s51, v36
; GFX908-NEXT: v_fmac_f32_e32 v20, s12, v36
; GFX908-NEXT: v_fmac_f32_e32 v19, s13, v36
; GFX908-NEXT: v_fmac_f32_e32 v18, s14, v36
; GFX908-NEXT: v_fmac_f32_e32 v17, s15, v36
; GFX908-NEXT: v_fmac_f32_e32 v16, s16, v36
; GFX908-NEXT: v_fmac_f32_e32 v15, s17, v36
; GFX908-NEXT: v_fmac_f32_e32 v14, s18, v36
; GFX908-NEXT: v_fmac_f32_e32 v9, s19, v36
; GFX908-NEXT: v_fmac_f32_e32 v8, s20, v36
; GFX908-NEXT: v_fmac_f32_e32 v7, s21, v36
; GFX908-NEXT: v_fmac_f32_e32 v6, s22, v36
; GFX908-NEXT: v_fmac_f32_e32 v5, s23, v36
; GFX908-NEXT: v_fmac_f32_e32 v10, s24, v36
; GFX908-NEXT: v_fmac_f32_e32 v11, s25, v36
; GFX908-NEXT: v_fmac_f32_e32 v12, s26, v36
; GFX908-NEXT: v_fmac_f32_e32 v13, s27, v36
; GFX908-NEXT: s_load_dwordx16 s[36:51], s[52:53], 0x3000
; GFX908-NEXT: s_load_dwordx16 s[12:27], s[52:53], 0x3040
; GFX908-NEXT: s_add_u32 s52, s52, 0x4000
; GFX908-NEXT: s_addc_u32 s53, s53, 0
; GFX908-NEXT: s_add_i32 s0, s0, -1
; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX908-NEXT: v_fmac_f32_e32 v34, s36, v35
; GFX908-NEXT: v_fmac_f32_e32 v33, s37, v35
; GFX908-NEXT: v_fmac_f32_e32 v32, s38, v35
; GFX908-NEXT: v_fmac_f32_e32 v31, s39, v35
; GFX908-NEXT: v_fmac_f32_e32 v30, s40, v35
; GFX908-NEXT: v_fmac_f32_e32 v2, s41, v35
; GFX908-NEXT: v_fmac_f32_e32 v1, s42, v35
; GFX908-NEXT: v_fmac_f32_e32 v29, s43, v35
; GFX908-NEXT: v_fmac_f32_e32 v28, s44, v35
; GFX908-NEXT: v_fmac_f32_e32 v27, s45, v35
; GFX908-NEXT: v_fmac_f32_e32 v26, s46, v35
; GFX908-NEXT: v_fmac_f32_e32 v25, s47, v35
; GFX908-NEXT: v_fmac_f32_e32 v24, s48, v35
; GFX908-NEXT: v_fmac_f32_e32 v23, s49, v35
; GFX908-NEXT: v_fmac_f32_e32 v22, s50, v35
; GFX908-NEXT: v_fmac_f32_e32 v21, s51, v35
; GFX908-NEXT: v_fmac_f32_e32 v20, s12, v35
; GFX908-NEXT: v_fmac_f32_e32 v19, s13, v35
; GFX908-NEXT: v_fmac_f32_e32 v18, s14, v35
; GFX908-NEXT: v_fmac_f32_e32 v17, s15, v35
; GFX908-NEXT: v_fmac_f32_e32 v16, s16, v35
; GFX908-NEXT: v_fmac_f32_e32 v15, s17, v35
; GFX908-NEXT: v_fmac_f32_e32 v14, s18, v35
; GFX908-NEXT: v_fmac_f32_e32 v9, s19, v35
; GFX908-NEXT: v_fmac_f32_e32 v8, s20, v35
; GFX908-NEXT: v_fmac_f32_e32 v7, s21, v35
; GFX908-NEXT: v_fmac_f32_e32 v6, s22, v35
; GFX908-NEXT: v_fmac_f32_e32 v5, s23, v35
; GFX908-NEXT: v_fmac_f32_e32 v10, s24, v35
; GFX908-NEXT: v_fmac_f32_e32 v11, s25, v35
; GFX908-NEXT: v_fmac_f32_e32 v12, s26, v35
; GFX908-NEXT: v_fmac_f32_e32 v13, s27, v35
; GFX908-NEXT: s_cmp_lg_u32 s0, 0
; GFX908-NEXT: s_cbranch_scc1 .LBB4_1
; GFX908-NEXT: ; %bb.2: ; %for.cond.cleanup26
; GFX908-NEXT: s_lshl_b32 s0, s8, 8
; GFX908-NEXT: s_add_i32 s0, s58, s0
; GFX908-NEXT: v_add_u32_e32 v0, s0, v0
; GFX908-NEXT: s_mov_b32 s0, 0x5397829d
; GFX908-NEXT: v_mul_hi_u32 v0, v0, s0
; GFX908-NEXT: s_mov_b32 s0, 0x31000
; GFX908-NEXT: v_mov_b32_e32 v4, 0
; GFX908-NEXT: s_movk_i32 s1, 0x310
; GFX908-NEXT: v_lshrrev_b32_e32 v0, 4, v0
; GFX908-NEXT: v_mul_lo_u32 v3, v0, s0
; GFX908-NEXT: v_mov_b32_e32 v0, s55
; GFX908-NEXT: s_mov_b32 s0, 0
; GFX908-NEXT: v_lshlrev_b64 v[3:4], 2, v[3:4]
; GFX908-NEXT: v_add_co_u32_e32 v35, vcc, s54, v3
; GFX908-NEXT: v_addc_co_u32_e32 v36, vcc, v0, v4, vcc
; GFX908-NEXT: v_mov_b32_e32 v3, s0
; GFX908-NEXT: global_store_dwordx2 v[35:36], v[2:3], off offset:3920
; GFX908-NEXT: global_store_dword v[35:36], v34, off
; GFX908-NEXT: global_store_dword v[35:36], v33, off offset:784
; GFX908-NEXT: global_store_dword v[35:36], v32, off offset:1568
; GFX908-NEXT: global_store_dword v[35:36], v31, off offset:2352
; GFX908-NEXT: global_store_dword v[35:36], v30, off offset:3136
; GFX908-NEXT: v_add_co_u32_e32 v3, vcc, s1, v35
; GFX908-NEXT: s_mov_b32 s1, s0
; GFX908-NEXT: v_addc_co_u32_e32 v4, vcc, 0, v36, vcc
; GFX908-NEXT: v_mov_b32_e32 v31, s1
; GFX908-NEXT: v_mov_b32_e32 v30, s0
; GFX908-NEXT: s_movk_i32 s1, 0x620
; GFX908-NEXT: v_add_co_u32_e32 v32, vcc, s1, v35
; GFX908-NEXT: v_addc_co_u32_e32 v33, vcc, 0, v36, vcc
; GFX908-NEXT: s_movk_i32 s1, 0x930
; GFX908-NEXT: v_add_co_u32_e32 v37, vcc, s1, v35
; GFX908-NEXT: v_addc_co_u32_e32 v38, vcc, 0, v36, vcc
; GFX908-NEXT: s_movk_i32 s1, 0xc40
; GFX908-NEXT: v_add_co_u32_e32 v39, vcc, s1, v35
; GFX908-NEXT: v_addc_co_u32_e32 v40, vcc, 0, v36, vcc
; GFX908-NEXT: s_movk_i32 s1, 0xf50
; GFX908-NEXT: v_add_co_u32_e32 v34, vcc, s1, v35
; GFX908-NEXT: global_store_dwordx2 v[35:36], v[30:31], off offset:840
; GFX908-NEXT: v_addc_co_u32_e32 v35, vcc, 0, v36, vcc
; GFX908-NEXT: v_mov_b32_e32 v2, s0
; GFX908-NEXT: v_add_co_u32_e32 v0, vcc, s1, v3
; GFX908-NEXT: global_store_dwordx2 v[3:4], v[1:2], off offset:3920
; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v4, vcc
; GFX908-NEXT: v_add_co_u32_e32 v2, vcc, s1, v32
; GFX908-NEXT: v_addc_co_u32_e32 v3, vcc, 0, v33, vcc
; GFX908-NEXT: global_store_dwordx2 v[37:38], v[30:31], off offset:3976
; GFX908-NEXT: global_store_dword v[37:38], v28, off offset:3920
; GFX908-NEXT: v_add_co_u32_e32 v28, vcc, s1, v37
; GFX908-NEXT: global_store_dword v[32:33], v29, off offset:3920
; GFX908-NEXT: v_addc_co_u32_e32 v29, vcc, 0, v38, vcc
; GFX908-NEXT: v_add_co_u32_e32 v30, vcc, s1, v39
; GFX908-NEXT: v_addc_co_u32_e32 v31, vcc, 0, v40, vcc
; GFX908-NEXT: global_store_dword v[39:40], v27, off offset:3920
; GFX908-NEXT: global_store_dword v[34:35], v26, off offset:3920
; GFX908-NEXT: v_add_co_u32_e32 v26, vcc, s1, v34
; GFX908-NEXT: v_addc_co_u32_e32 v27, vcc, 0, v35, vcc
; GFX908-NEXT: global_store_dword v[0:1], v25, off offset:3920
; GFX908-NEXT: v_add_co_u32_e32 v0, vcc, s1, v0
; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
; GFX908-NEXT: global_store_dword v[2:3], v24, off offset:3920
; GFX908-NEXT: v_add_co_u32_e32 v2, vcc, s1, v2
; GFX908-NEXT: v_addc_co_u32_e32 v3, vcc, 0, v3, vcc
; GFX908-NEXT: global_store_dword v[28:29], v23, off offset:3920
; GFX908-NEXT: v_add_co_u32_e32 v23, vcc, s1, v28
; GFX908-NEXT: v_addc_co_u32_e32 v24, vcc, 0, v29, vcc
; GFX908-NEXT: v_add_co_u32_e32 v28, vcc, s1, v30
; GFX908-NEXT: v_addc_co_u32_e32 v29, vcc, 0, v31, vcc
; GFX908-NEXT: global_store_dword v[26:27], v21, off offset:3920
; GFX908-NEXT: v_add_co_u32_e32 v21, vcc, s1, v26
; GFX908-NEXT: global_store_dword v[30:31], v22, off offset:3920
; GFX908-NEXT: v_addc_co_u32_e32 v22, vcc, 0, v27, vcc
; GFX908-NEXT: global_store_dword v[0:1], v20, off offset:3920
; GFX908-NEXT: v_add_co_u32_e32 v0, vcc, s1, v0
; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
; GFX908-NEXT: global_store_dword v[2:3], v19, off offset:3920
; GFX908-NEXT: v_add_co_u32_e32 v2, vcc, s1, v2
; GFX908-NEXT: v_addc_co_u32_e32 v3, vcc, 0, v3, vcc
; GFX908-NEXT: global_store_dword v[23:24], v18, off offset:3920
; GFX908-NEXT: v_add_co_u32_e32 v18, vcc, s1, v23
; GFX908-NEXT: v_addc_co_u32_e32 v19, vcc, 0, v24, vcc
; GFX908-NEXT: v_add_co_u32_e32 v23, vcc, s1, v28
; GFX908-NEXT: v_addc_co_u32_e32 v24, vcc, 0, v29, vcc
; GFX908-NEXT: v_add_co_u32_e32 v25, vcc, s1, v21
; GFX908-NEXT: v_addc_co_u32_e32 v26, vcc, 0, v22, vcc
; GFX908-NEXT: global_store_dword v[28:29], v17, off offset:3920
; GFX908-NEXT: global_store_dword v[21:22], v16, off offset:3920
; GFX908-NEXT: v_add_co_u32_e32 v16, vcc, s1, v0
; GFX908-NEXT: v_addc_co_u32_e32 v17, vcc, 0, v1, vcc
; GFX908-NEXT: global_store_dword v[0:1], v15, off offset:3920
; GFX908-NEXT: v_add_co_u32_e32 v0, vcc, s1, v2
; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v3, vcc
; GFX908-NEXT: global_store_dword v[2:3], v14, off offset:3920
; GFX908-NEXT: v_add_co_u32_e32 v2, vcc, s1, v18
; GFX908-NEXT: v_addc_co_u32_e32 v3, vcc, 0, v19, vcc
; GFX908-NEXT: v_add_co_u32_e32 v14, vcc, s1, v23
; GFX908-NEXT: v_addc_co_u32_e32 v15, vcc, 0, v24, vcc
; GFX908-NEXT: global_store_dword v[23:24], v8, off offset:3920
; GFX908-NEXT: v_add_co_u32_e32 v8, vcc, 0xf50, v25
; GFX908-NEXT: global_store_dword v[18:19], v9, off offset:3920
; GFX908-NEXT: v_addc_co_u32_e32 v9, vcc, 0, v26, vcc
; GFX908-NEXT: v_add_co_u32_e32 v18, vcc, 0xf50, v16
; GFX908-NEXT: global_store_dword v[25:26], v7, off offset:3920
; GFX908-NEXT: v_addc_co_u32_e32 v19, vcc, 0, v17, vcc
; GFX908-NEXT: global_store_dword v[16:17], v6, off offset:3920
; GFX908-NEXT: global_store_dword v[0:1], v5, off offset:3920
; GFX908-NEXT: global_store_dword v[2:3], v10, off offset:3920
; GFX908-NEXT: global_store_dword v[14:15], v11, off offset:3920
; GFX908-NEXT: global_store_dword v[8:9], v12, off offset:3920
; GFX908-NEXT: global_store_dword v[18:19], v13, off offset:3920
; GFX908-NEXT: s_endpgm
;
; GFX908-GCNTRACKERS-LABEL: excess_soft_clause_reg_pressure:
; GFX908-GCNTRACKERS: ; %bb.0: ; %entry
; GFX908-GCNTRACKERS-NEXT: s_load_dwordx8 s[52:59], s[4:5], 0x24
; GFX908-GCNTRACKERS-NEXT: v_lshlrev_b32_e32 v2, 2, v0
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v1, 0
; GFX908-GCNTRACKERS-NEXT: s_movk_i32 s0, 0x200
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v5, 0
; GFX908-GCNTRACKERS-NEXT: s_waitcnt lgkmcnt(0)
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v3, s57
; GFX908-GCNTRACKERS-NEXT: v_add_co_u32_e32 v2, vcc, s56, v2
; GFX908-GCNTRACKERS-NEXT: v_addc_co_u32_e32 v4, vcc, 0, v3, vcc
; GFX908-GCNTRACKERS-NEXT: v_add_co_u32_e32 v3, vcc, 0x188, v2
; GFX908-GCNTRACKERS-NEXT: v_addc_co_u32_e32 v4, vcc, 0, v4, vcc
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v6, 0
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v7, 0
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v8, 0
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v9, 0
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v14, 0
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v15, 0
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v16, 0
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v17, 0
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v18, 0
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v19, 0
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v20, 0
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v21, 0
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v22, 0
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v23, 0
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v24, 0
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v25, 0
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v26, 0
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v27, 0
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v28, 0
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v29, 0
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v2, 0
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v30, 0
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v31, 0
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v32, 0
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v33, 0
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v34, 0
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v10, 0
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v11, 0
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v12, 0
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v13, 0
; GFX908-GCNTRACKERS-NEXT: .LBB4_1: ; %for.cond28.preheader
; GFX908-GCNTRACKERS-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX908-GCNTRACKERS-NEXT: global_load_dword v38, v[3:4], off offset:-392
; GFX908-GCNTRACKERS-NEXT: global_load_dword v37, v[3:4], off offset:-196
; GFX908-GCNTRACKERS-NEXT: global_load_dword v36, v[3:4], off
; GFX908-GCNTRACKERS-NEXT: global_load_dword v35, v[3:4], off offset:196
; GFX908-GCNTRACKERS-NEXT: s_load_dwordx16 s[36:51], s[52:53], 0x0
; GFX908-GCNTRACKERS-NEXT: s_load_dwordx16 s[12:27], s[52:53], 0x40
; GFX908-GCNTRACKERS-NEXT: v_add_co_u32_e32 v3, vcc, 0x310, v3
; GFX908-GCNTRACKERS-NEXT: v_addc_co_u32_e32 v4, vcc, 0, v4, vcc
; GFX908-GCNTRACKERS-NEXT: s_waitcnt vmcnt(3) lgkmcnt(0)
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v34, s36, v38
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v33, s37, v38
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v32, s38, v38
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v31, s39, v38
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v30, s40, v38
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v2, s41, v38
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v1, s42, v38
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v29, s43, v38
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v28, s44, v38
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v27, s45, v38
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v26, s46, v38
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v25, s47, v38
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v24, s48, v38
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v23, s49, v38
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v22, s50, v38
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v21, s51, v38
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v20, s12, v38
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v19, s13, v38
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v18, s14, v38
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v17, s15, v38
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v16, s16, v38
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v15, s17, v38
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v14, s18, v38
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v9, s19, v38
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v8, s20, v38
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v7, s21, v38
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v6, s22, v38
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v5, s23, v38
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v10, s24, v38
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v11, s25, v38
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v12, s26, v38
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v13, s27, v38
; GFX908-GCNTRACKERS-NEXT: s_load_dwordx16 s[36:51], s[52:53], 0x1000
; GFX908-GCNTRACKERS-NEXT: s_load_dwordx16 s[12:27], s[52:53], 0x1040
; GFX908-GCNTRACKERS-NEXT: s_waitcnt vmcnt(2) lgkmcnt(0)
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v34, s36, v37
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v33, s37, v37
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v32, s38, v37
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v31, s39, v37
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v30, s40, v37
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v2, s41, v37
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v1, s42, v37
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v29, s43, v37
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v28, s44, v37
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v27, s45, v37
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v26, s46, v37
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v25, s47, v37
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v24, s48, v37
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v23, s49, v37
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v22, s50, v37
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v21, s51, v37
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v20, s12, v37
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v19, s13, v37
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v18, s14, v37
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v17, s15, v37
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v16, s16, v37
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v15, s17, v37
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v14, s18, v37
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v9, s19, v37
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v8, s20, v37
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v7, s21, v37
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v6, s22, v37
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v5, s23, v37
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v10, s24, v37
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v11, s25, v37
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v12, s26, v37
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v13, s27, v37
; GFX908-GCNTRACKERS-NEXT: s_load_dwordx16 s[36:51], s[52:53], 0x2000
; GFX908-GCNTRACKERS-NEXT: s_load_dwordx16 s[12:27], s[52:53], 0x2040
; GFX908-GCNTRACKERS-NEXT: s_waitcnt vmcnt(1) lgkmcnt(0)
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v34, s36, v36
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v33, s37, v36
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v32, s38, v36
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v31, s39, v36
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v30, s40, v36
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v2, s41, v36
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v1, s42, v36
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v29, s43, v36
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v28, s44, v36
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v27, s45, v36
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v26, s46, v36
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v25, s47, v36
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v24, s48, v36
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v23, s49, v36
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v22, s50, v36
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v21, s51, v36
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v20, s12, v36
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v19, s13, v36
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v18, s14, v36
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v17, s15, v36
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v16, s16, v36
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v15, s17, v36
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v14, s18, v36
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v9, s19, v36
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v8, s20, v36
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v7, s21, v36
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v6, s22, v36
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v5, s23, v36
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v10, s24, v36
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v11, s25, v36
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v12, s26, v36
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v13, s27, v36
; GFX908-GCNTRACKERS-NEXT: s_load_dwordx16 s[36:51], s[52:53], 0x3000
; GFX908-GCNTRACKERS-NEXT: s_load_dwordx16 s[12:27], s[52:53], 0x3040
; GFX908-GCNTRACKERS-NEXT: s_add_u32 s52, s52, 0x4000
; GFX908-GCNTRACKERS-NEXT: s_addc_u32 s53, s53, 0
; GFX908-GCNTRACKERS-NEXT: s_add_i32 s0, s0, -1
; GFX908-GCNTRACKERS-NEXT: s_cmp_lg_u32 s0, 0
; GFX908-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v34, s36, v35
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v33, s37, v35
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v32, s38, v35
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v31, s39, v35
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v30, s40, v35
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v2, s41, v35
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v1, s42, v35
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v29, s43, v35
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v28, s44, v35
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v27, s45, v35
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v26, s46, v35
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v25, s47, v35
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v24, s48, v35
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v23, s49, v35
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v22, s50, v35
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v21, s51, v35
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v20, s12, v35
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v19, s13, v35
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v18, s14, v35
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v17, s15, v35
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v16, s16, v35
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v15, s17, v35
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v14, s18, v35
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v9, s19, v35
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v8, s20, v35
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v7, s21, v35
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v6, s22, v35
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v5, s23, v35
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v10, s24, v35
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v11, s25, v35
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v12, s26, v35
; GFX908-GCNTRACKERS-NEXT: v_fmac_f32_e32 v13, s27, v35
; GFX908-GCNTRACKERS-NEXT: s_cbranch_scc1 .LBB4_1
; GFX908-GCNTRACKERS-NEXT: ; %bb.2: ; %for.cond.cleanup26
; GFX908-GCNTRACKERS-NEXT: s_lshl_b32 s0, s8, 8
; GFX908-GCNTRACKERS-NEXT: s_add_i32 s0, s58, s0
; GFX908-GCNTRACKERS-NEXT: v_add_u32_e32 v0, s0, v0
; GFX908-GCNTRACKERS-NEXT: s_mov_b32 s0, 0x5397829d
; GFX908-GCNTRACKERS-NEXT: v_mul_hi_u32 v0, v0, s0
; GFX908-GCNTRACKERS-NEXT: s_mov_b32 s0, 0x31000
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v4, 0
; GFX908-GCNTRACKERS-NEXT: s_movk_i32 s1, 0x310
; GFX908-GCNTRACKERS-NEXT: v_lshrrev_b32_e32 v0, 4, v0
; GFX908-GCNTRACKERS-NEXT: v_mul_lo_u32 v3, v0, s0
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v0, s55
; GFX908-GCNTRACKERS-NEXT: s_mov_b32 s0, 0
; GFX908-GCNTRACKERS-NEXT: v_lshlrev_b64 v[3:4], 2, v[3:4]
; GFX908-GCNTRACKERS-NEXT: v_add_co_u32_e32 v35, vcc, s54, v3
; GFX908-GCNTRACKERS-NEXT: v_addc_co_u32_e32 v36, vcc, v0, v4, vcc
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v3, s0
; GFX908-GCNTRACKERS-NEXT: global_store_dwordx2 v[35:36], v[2:3], off offset:3920
; GFX908-GCNTRACKERS-NEXT: v_add_co_u32_e32 v3, vcc, s1, v35
; GFX908-GCNTRACKERS-NEXT: v_addc_co_u32_e32 v4, vcc, 0, v36, vcc
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v2, s0
; GFX908-GCNTRACKERS-NEXT: s_mov_b32 s1, s0
; GFX908-GCNTRACKERS-NEXT: global_store_dwordx2 v[3:4], v[1:2], off offset:3920
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v0, s0
; GFX908-GCNTRACKERS-NEXT: v_mov_b32_e32 v1, s1
; GFX908-GCNTRACKERS-NEXT: s_movk_i32 s0, 0x930
; GFX908-GCNTRACKERS-NEXT: v_add_co_u32_e32 v37, vcc, s0, v35
; GFX908-GCNTRACKERS-NEXT: v_addc_co_u32_e32 v38, vcc, 0, v36, vcc
; GFX908-GCNTRACKERS-NEXT: s_movk_i32 s0, 0x620
; GFX908-GCNTRACKERS-NEXT: global_store_dwordx2 v[35:36], v[0:1], off offset:840
; GFX908-GCNTRACKERS-NEXT: global_store_dwordx2 v[37:38], v[0:1], off offset:3976
; GFX908-GCNTRACKERS-NEXT: global_store_dword v[35:36], v34, off
; GFX908-GCNTRACKERS-NEXT: global_store_dword v[35:36], v33, off offset:784
; GFX908-GCNTRACKERS-NEXT: global_store_dword v[35:36], v32, off offset:1568
; GFX908-GCNTRACKERS-NEXT: global_store_dword v[35:36], v31, off offset:2352
; GFX908-GCNTRACKERS-NEXT: global_store_dword v[35:36], v30, off offset:3136
; GFX908-GCNTRACKERS-NEXT: v_add_co_u32_e32 v0, vcc, s0, v35
; GFX908-GCNTRACKERS-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v36, vcc
; GFX908-GCNTRACKERS-NEXT: s_movk_i32 s0, 0xc40
; GFX908-GCNTRACKERS-NEXT: global_store_dword v[0:1], v29, off offset:3920
; GFX908-GCNTRACKERS-NEXT: global_store_dword v[37:38], v28, off offset:3920
; GFX908-GCNTRACKERS-NEXT: v_add_co_u32_e32 v28, vcc, s0, v35
; GFX908-GCNTRACKERS-NEXT: v_addc_co_u32_e32 v29, vcc, 0, v36, vcc
; GFX908-GCNTRACKERS-NEXT: s_movk_i32 s0, 0xf50
; GFX908-GCNTRACKERS-NEXT: v_add_co_u32_e32 v30, vcc, s0, v35
; GFX908-GCNTRACKERS-NEXT: v_addc_co_u32_e32 v31, vcc, 0, v36, vcc
; GFX908-GCNTRACKERS-NEXT: v_add_co_u32_e32 v2, vcc, s0, v3
; GFX908-GCNTRACKERS-NEXT: v_addc_co_u32_e32 v3, vcc, 0, v4, vcc
; GFX908-GCNTRACKERS-NEXT: v_add_co_u32_e32 v0, vcc, s0, v0
; GFX908-GCNTRACKERS-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
; GFX908-GCNTRACKERS-NEXT: global_store_dword v[28:29], v27, off offset:3920
; GFX908-GCNTRACKERS-NEXT: global_store_dword v[0:1], v24, off offset:3920
; GFX908-GCNTRACKERS-NEXT: v_add_co_u32_e32 v24, vcc, s0, v37
; GFX908-GCNTRACKERS-NEXT: global_store_dword v[2:3], v25, off offset:3920
; GFX908-GCNTRACKERS-NEXT: v_addc_co_u32_e32 v25, vcc, 0, v38, vcc
; GFX908-GCNTRACKERS-NEXT: global_store_dword v[30:31], v26, off offset:3920
; GFX908-GCNTRACKERS-NEXT: v_add_co_u32_e32 v26, vcc, s0, v28
; GFX908-GCNTRACKERS-NEXT: v_addc_co_u32_e32 v27, vcc, 0, v29, vcc
; GFX908-GCNTRACKERS-NEXT: global_store_dword v[26:27], v22, off offset:3920
; GFX908-GCNTRACKERS-NEXT: v_add_co_u32_e32 v22, vcc, s0, v30
; GFX908-GCNTRACKERS-NEXT: global_store_dword v[24:25], v23, off offset:3920
; GFX908-GCNTRACKERS-NEXT: v_addc_co_u32_e32 v23, vcc, 0, v31, vcc
; GFX908-GCNTRACKERS-NEXT: v_add_co_u32_e32 v2, vcc, s0, v2
; GFX908-GCNTRACKERS-NEXT: v_addc_co_u32_e32 v3, vcc, 0, v3, vcc
; GFX908-GCNTRACKERS-NEXT: v_add_co_u32_e32 v0, vcc, s0, v0
; GFX908-GCNTRACKERS-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
; GFX908-GCNTRACKERS-NEXT: global_store_dword v[2:3], v20, off offset:3920
; GFX908-GCNTRACKERS-NEXT: v_add_co_u32_e32 v20, vcc, s0, v24
; GFX908-GCNTRACKERS-NEXT: global_store_dword v[22:23], v21, off offset:3920
; GFX908-GCNTRACKERS-NEXT: v_addc_co_u32_e32 v21, vcc, 0, v25, vcc
; GFX908-GCNTRACKERS-NEXT: v_add_co_u32_e32 v24, vcc, s0, v26
; GFX908-GCNTRACKERS-NEXT: v_addc_co_u32_e32 v25, vcc, 0, v27, vcc
; GFX908-GCNTRACKERS-NEXT: v_add_co_u32_e32 v22, vcc, s0, v22
; GFX908-GCNTRACKERS-NEXT: v_addc_co_u32_e32 v23, vcc, 0, v23, vcc
; GFX908-GCNTRACKERS-NEXT: v_add_co_u32_e32 v2, vcc, s0, v2
; GFX908-GCNTRACKERS-NEXT: v_addc_co_u32_e32 v3, vcc, 0, v3, vcc
; GFX908-GCNTRACKERS-NEXT: global_store_dword v[0:1], v19, off offset:3920
; GFX908-GCNTRACKERS-NEXT: v_add_co_u32_e32 v0, vcc, s0, v0
; GFX908-GCNTRACKERS-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
; GFX908-GCNTRACKERS-NEXT: global_store_dword v[20:21], v18, off offset:3920
; GFX908-GCNTRACKERS-NEXT: v_add_co_u32_e32 v18, vcc, s0, v20
; GFX908-GCNTRACKERS-NEXT: v_addc_co_u32_e32 v19, vcc, 0, v21, vcc
; GFX908-GCNTRACKERS-NEXT: v_add_co_u32_e32 v20, vcc, s0, v24
; GFX908-GCNTRACKERS-NEXT: v_addc_co_u32_e32 v21, vcc, 0, v25, vcc
; GFX908-GCNTRACKERS-NEXT: global_store_dword v[24:25], v17, off offset:3920
; GFX908-GCNTRACKERS-NEXT: v_add_co_u32_e32 v24, vcc, s0, v22
; GFX908-GCNTRACKERS-NEXT: v_addc_co_u32_e32 v25, vcc, 0, v23, vcc
; GFX908-GCNTRACKERS-NEXT: global_store_dword v[22:23], v16, off offset:3920
; GFX908-GCNTRACKERS-NEXT: v_add_co_u32_e32 v16, vcc, s0, v2
; GFX908-GCNTRACKERS-NEXT: v_addc_co_u32_e32 v17, vcc, 0, v3, vcc
; GFX908-GCNTRACKERS-NEXT: global_store_dword v[2:3], v15, off offset:3920
; GFX908-GCNTRACKERS-NEXT: v_add_co_u32_e32 v2, vcc, s0, v0
; GFX908-GCNTRACKERS-NEXT: v_addc_co_u32_e32 v3, vcc, 0, v1, vcc
; GFX908-GCNTRACKERS-NEXT: global_store_dword v[0:1], v14, off offset:3920
; GFX908-GCNTRACKERS-NEXT: v_add_co_u32_e32 v0, vcc, s0, v18
; GFX908-GCNTRACKERS-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v19, vcc
; GFX908-GCNTRACKERS-NEXT: v_add_co_u32_e32 v14, vcc, s0, v20
; GFX908-GCNTRACKERS-NEXT: v_addc_co_u32_e32 v15, vcc, 0, v21, vcc
; GFX908-GCNTRACKERS-NEXT: global_store_dword v[20:21], v8, off offset:3920
; GFX908-GCNTRACKERS-NEXT: v_add_co_u32_e32 v8, vcc, 0xf50, v24
; GFX908-GCNTRACKERS-NEXT: global_store_dword v[18:19], v9, off offset:3920
; GFX908-GCNTRACKERS-NEXT: v_addc_co_u32_e32 v9, vcc, 0, v25, vcc
; GFX908-GCNTRACKERS-NEXT: v_add_co_u32_e32 v18, vcc, 0xf50, v16
; GFX908-GCNTRACKERS-NEXT: global_store_dword v[24:25], v7, off offset:3920
; GFX908-GCNTRACKERS-NEXT: v_addc_co_u32_e32 v19, vcc, 0, v17, vcc
; GFX908-GCNTRACKERS-NEXT: global_store_dword v[16:17], v6, off offset:3920
; GFX908-GCNTRACKERS-NEXT: global_store_dword v[2:3], v5, off offset:3920
; GFX908-GCNTRACKERS-NEXT: global_store_dword v[0:1], v10, off offset:3920
; GFX908-GCNTRACKERS-NEXT: global_store_dword v[14:15], v11, off offset:3920
; GFX908-GCNTRACKERS-NEXT: global_store_dword v[8:9], v12, off offset:3920
; GFX908-GCNTRACKERS-NEXT: global_store_dword v[18:19], v13, off offset:3920
; GFX908-GCNTRACKERS-NEXT: s_endpgm
;
; GENERIC-LABEL: excess_soft_clause_reg_pressure:
; GENERIC: ; %bb.0: ; %entry
; GENERIC-NEXT: s_load_dwordx8 s[0:7], s[4:5], 0x9
; GENERIC-NEXT: v_lshlrev_b32_e32 v1, 2, v0
; GENERIC-NEXT: s_mov_b32 s30, 0
; GENERIC-NEXT: s_mov_b32 s31, 0xf000
; GENERIC-NEXT: s_waitcnt lgkmcnt(0)
; GENERIC-NEXT: s_movk_i32 s7, 0x3000
; GENERIC-NEXT: v_mov_b32_e32 v2, s5
; GENERIC-NEXT: v_add_i32_e32 v1, vcc, s4, v1
; GENERIC-NEXT: v_addc_u32_e32 v2, vcc, 0, v2, vcc
; GENERIC-NEXT: s_movk_i32 s4, 0x1000
; GENERIC-NEXT: s_movk_i32 s5, 0x2000
; GENERIC-NEXT: s_movk_i32 s9, 0x1040
; GENERIC-NEXT: s_movk_i32 s10, 0x2040
; GENERIC-NEXT: s_movk_i32 s11, 0x3040
; GENERIC-NEXT: s_movk_i32 s33, 0x310
; GENERIC-NEXT: v_mov_b32_e32 v6, 0
; GENERIC-NEXT: v_mov_b32_e32 v7, 0
; GENERIC-NEXT: v_mov_b32_e32 v8, 0
; GENERIC-NEXT: v_mov_b32_e32 v9, 0
; GENERIC-NEXT: v_mov_b32_e32 v10, 0
; GENERIC-NEXT: v_mov_b32_e32 v11, 0
; GENERIC-NEXT: v_mov_b32_e32 v12, 0
; GENERIC-NEXT: v_mov_b32_e32 v16, 0
; GENERIC-NEXT: v_mov_b32_e32 v18, 0
; GENERIC-NEXT: v_mov_b32_e32 v19, 0
; GENERIC-NEXT: v_mov_b32_e32 v20, 0
; GENERIC-NEXT: v_mov_b32_e32 v21, 0
; GENERIC-NEXT: v_mov_b32_e32 v22, 0
; GENERIC-NEXT: v_mov_b32_e32 v23, 0
; GENERIC-NEXT: v_mov_b32_e32 v24, 0
; GENERIC-NEXT: v_mov_b32_e32 v25, 0
; GENERIC-NEXT: v_mov_b32_e32 v26, 0
; GENERIC-NEXT: s_mov_b32 s34, s30
; GENERIC-NEXT: v_mov_b32_e32 v27, 0
; GENERIC-NEXT: v_mov_b32_e32 v28, 0
; GENERIC-NEXT: v_mov_b32_e32 v29, 0
; GENERIC-NEXT: v_mov_b32_e32 v30, 0
; GENERIC-NEXT: v_mov_b32_e32 v32, 0
; GENERIC-NEXT: v_mov_b32_e32 v33, 0
; GENERIC-NEXT: v_mov_b32_e32 v34, 0
; GENERIC-NEXT: v_mov_b32_e32 v35, 0
; GENERIC-NEXT: v_mov_b32_e32 v36, 0
; GENERIC-NEXT: v_mov_b32_e32 v17, 0
; GENERIC-NEXT: v_mov_b32_e32 v15, 0
; GENERIC-NEXT: v_mov_b32_e32 v14, 0
; GENERIC-NEXT: v_mov_b32_e32 v13, 0
; GENERIC-NEXT: v_mov_b32_e32 v3, 0
; GENERIC-NEXT: v_mov_b32_e32 v4, 0
; GENERIC-NEXT: .LBB4_1: ; %for.cond28.preheader
; GENERIC-NEXT: ; =>This Inner Loop Header: Depth=1
; GENERIC-NEXT: s_mov_b32 s28, s30
; GENERIC-NEXT: s_mov_b32 s29, s30
; GENERIC-NEXT: buffer_load_dword v38, v[1:2], s[28:31], 0 addr64
; GENERIC-NEXT: buffer_load_dword v37, v[1:2], s[28:31], 0 addr64 offset:196
; GENERIC-NEXT: buffer_load_dword v31, v[1:2], s[28:31], 0 addr64 offset:392
; GENERIC-NEXT: buffer_load_dword v5, v[1:2], s[28:31], 0 addr64 offset:588
; GENERIC-NEXT: s_load_dwordx16 s[36:51], s[0:1], 0x0
; GENERIC-NEXT: s_load_dwordx16 s[12:27], s[0:1], 0x10
; GENERIC-NEXT: v_add_i32_e32 v1, vcc, s33, v1
; GENERIC-NEXT: v_addc_u32_e32 v2, vcc, 0, v2, vcc
; GENERIC-NEXT: s_waitcnt vmcnt(3) lgkmcnt(0)
; GENERIC-NEXT: v_fma_f32 v4, v38, s41, v4
; GENERIC-NEXT: v_fma_f32 v3, v38, s42, v3
; GENERIC-NEXT: v_fma_f32 v36, v38, s36, v36
; GENERIC-NEXT: v_fma_f32 v35, v38, s37, v35
; GENERIC-NEXT: v_fma_f32 v34, v38, s38, v34
; GENERIC-NEXT: v_fma_f32 v33, v38, s39, v33
; GENERIC-NEXT: v_fma_f32 v32, v38, s40, v32
; GENERIC-NEXT: v_fma_f32 v30, v38, s43, v30
; GENERIC-NEXT: v_fma_f32 v29, v38, s44, v29
; GENERIC-NEXT: v_fma_f32 v28, v38, s45, v28
; GENERIC-NEXT: v_fma_f32 v27, v38, s46, v27
; GENERIC-NEXT: v_fma_f32 v26, v38, s47, v26
; GENERIC-NEXT: v_fma_f32 v25, v38, s48, v25
; GENERIC-NEXT: v_fma_f32 v24, v38, s49, v24
; GENERIC-NEXT: v_fma_f32 v23, v38, s50, v23
; GENERIC-NEXT: v_fma_f32 v22, v38, s51, v22
; GENERIC-NEXT: v_fma_f32 v21, v38, s12, v21
; GENERIC-NEXT: v_fma_f32 v20, v38, s13, v20
; GENERIC-NEXT: v_fma_f32 v19, v38, s14, v19
; GENERIC-NEXT: v_fma_f32 v18, v38, s15, v18
; GENERIC-NEXT: v_fma_f32 v16, v38, s16, v16
; GENERIC-NEXT: v_fma_f32 v12, v38, s17, v12
; GENERIC-NEXT: v_fma_f32 v11, v38, s18, v11
; GENERIC-NEXT: v_fma_f32 v10, v38, s19, v10
; GENERIC-NEXT: v_fma_f32 v9, v38, s20, v9
; GENERIC-NEXT: v_fma_f32 v8, v38, s21, v8
; GENERIC-NEXT: v_fma_f32 v7, v38, s22, v7
; GENERIC-NEXT: v_fma_f32 v6, v38, s23, v6
; GENERIC-NEXT: v_fma_f32 v17, v38, s24, v17
; GENERIC-NEXT: v_fma_f32 v15, v38, s25, v15
; GENERIC-NEXT: v_fma_f32 v14, v38, s26, v14
; GENERIC-NEXT: v_fma_f32 v13, v38, s27, v13
; GENERIC-NEXT: s_load_dwordx16 s[36:51], s[0:1], s4
; GENERIC-NEXT: s_load_dwordx16 s[12:27], s[0:1], s9
; GENERIC-NEXT: s_waitcnt vmcnt(2) lgkmcnt(0)
; GENERIC-NEXT: v_fma_f32 v36, v37, s36, v36
; GENERIC-NEXT: v_fma_f32 v35, v37, s37, v35
; GENERIC-NEXT: v_fma_f32 v34, v37, s38, v34
; GENERIC-NEXT: v_fma_f32 v33, v37, s39, v33
; GENERIC-NEXT: v_fma_f32 v32, v37, s40, v32
; GENERIC-NEXT: v_fma_f32 v4, v37, s41, v4
; GENERIC-NEXT: v_fma_f32 v3, v37, s42, v3
; GENERIC-NEXT: v_fma_f32 v30, v37, s43, v30
; GENERIC-NEXT: v_fma_f32 v29, v37, s44, v29
; GENERIC-NEXT: v_fma_f32 v28, v37, s45, v28
; GENERIC-NEXT: v_fma_f32 v27, v37, s46, v27
; GENERIC-NEXT: v_fma_f32 v26, v37, s47, v26
; GENERIC-NEXT: v_fma_f32 v25, v37, s48, v25
; GENERIC-NEXT: v_fma_f32 v24, v37, s49, v24
; GENERIC-NEXT: v_fma_f32 v23, v37, s50, v23
; GENERIC-NEXT: v_fma_f32 v22, v37, s51, v22
; GENERIC-NEXT: v_fma_f32 v21, v37, s12, v21
; GENERIC-NEXT: v_fma_f32 v20, v37, s13, v20
; GENERIC-NEXT: v_fma_f32 v19, v37, s14, v19
; GENERIC-NEXT: v_fma_f32 v18, v37, s15, v18
; GENERIC-NEXT: v_fma_f32 v16, v37, s16, v16
; GENERIC-NEXT: v_fma_f32 v12, v37, s17, v12
; GENERIC-NEXT: v_fma_f32 v11, v37, s18, v11
; GENERIC-NEXT: v_fma_f32 v10, v37, s19, v10
; GENERIC-NEXT: v_fma_f32 v9, v37, s20, v9
; GENERIC-NEXT: v_fma_f32 v8, v37, s21, v8
; GENERIC-NEXT: v_fma_f32 v7, v37, s22, v7
; GENERIC-NEXT: v_fma_f32 v6, v37, s23, v6
; GENERIC-NEXT: v_fma_f32 v17, v37, s24, v17
; GENERIC-NEXT: v_fma_f32 v15, v37, s25, v15
; GENERIC-NEXT: v_fma_f32 v14, v37, s26, v14
; GENERIC-NEXT: v_fma_f32 v13, v37, s27, v13
; GENERIC-NEXT: s_load_dwordx16 s[52:67], s[0:1], s5
; GENERIC-NEXT: s_load_dwordx16 s[12:27], s[0:1], s10
; GENERIC-NEXT: s_load_dwordx16 s[36:51], s[0:1], s7
; GENERIC-NEXT: s_waitcnt vmcnt(1) lgkmcnt(0)
; GENERIC-NEXT: v_fma_f32 v36, v31, s52, v36
; GENERIC-NEXT: v_fma_f32 v35, v31, s53, v35
; GENERIC-NEXT: v_fma_f32 v34, v31, s54, v34
; GENERIC-NEXT: v_fma_f32 v33, v31, s55, v33
; GENERIC-NEXT: v_fma_f32 v32, v31, s56, v32
; GENERIC-NEXT: v_fma_f32 v4, v31, s57, v4
; GENERIC-NEXT: v_fma_f32 v3, v31, s58, v3
; GENERIC-NEXT: v_fma_f32 v30, v31, s59, v30
; GENERIC-NEXT: v_fma_f32 v29, v31, s60, v29
; GENERIC-NEXT: v_fma_f32 v28, v31, s61, v28
; GENERIC-NEXT: v_fma_f32 v27, v31, s62, v27
; GENERIC-NEXT: v_fma_f32 v26, v31, s63, v26
; GENERIC-NEXT: v_fma_f32 v25, v31, s64, v25
; GENERIC-NEXT: v_fma_f32 v24, v31, s65, v24
; GENERIC-NEXT: v_fma_f32 v23, v31, s66, v23
; GENERIC-NEXT: v_fma_f32 v22, v31, s67, v22
; GENERIC-NEXT: s_load_dwordx16 s[52:67], s[0:1], s11
; GENERIC-NEXT: s_add_u32 s0, s0, 0x4000
; GENERIC-NEXT: s_addc_u32 s1, s1, 0
; GENERIC-NEXT: s_add_i32 s34, s34, 1
; GENERIC-NEXT: v_fma_f32 v21, v31, s12, v21
; GENERIC-NEXT: v_fma_f32 v20, v31, s13, v20
; GENERIC-NEXT: v_fma_f32 v19, v31, s14, v19
; GENERIC-NEXT: v_fma_f32 v18, v31, s15, v18
; GENERIC-NEXT: v_fma_f32 v16, v31, s16, v16
; GENERIC-NEXT: v_fma_f32 v12, v31, s17, v12
; GENERIC-NEXT: v_fma_f32 v11, v31, s18, v11
; GENERIC-NEXT: v_fma_f32 v10, v31, s19, v10
; GENERIC-NEXT: v_fma_f32 v9, v31, s20, v9
; GENERIC-NEXT: v_fma_f32 v8, v31, s21, v8
; GENERIC-NEXT: v_fma_f32 v7, v31, s22, v7
; GENERIC-NEXT: v_fma_f32 v6, v31, s23, v6
; GENERIC-NEXT: v_fma_f32 v17, v31, s24, v17
; GENERIC-NEXT: v_fma_f32 v15, v31, s25, v15
; GENERIC-NEXT: v_fma_f32 v14, v31, s26, v14
; GENERIC-NEXT: v_fma_f32 v13, v31, s27, v13
; GENERIC-NEXT: s_cmpk_lg_i32 s34, 0x200
; GENERIC-NEXT: s_waitcnt vmcnt(0)
; GENERIC-NEXT: v_fma_f32 v36, v5, s36, v36
; GENERIC-NEXT: v_fma_f32 v35, v5, s37, v35
; GENERIC-NEXT: v_fma_f32 v34, v5, s38, v34
; GENERIC-NEXT: v_fma_f32 v33, v5, s39, v33
; GENERIC-NEXT: v_fma_f32 v32, v5, s40, v32
; GENERIC-NEXT: v_fma_f32 v30, v5, s43, v30
; GENERIC-NEXT: v_fma_f32 v29, v5, s44, v29
; GENERIC-NEXT: v_fma_f32 v28, v5, s45, v28
; GENERIC-NEXT: v_fma_f32 v27, v5, s46, v27
; GENERIC-NEXT: v_fma_f32 v26, v5, s47, v26
; GENERIC-NEXT: v_fma_f32 v25, v5, s48, v25
; GENERIC-NEXT: v_fma_f32 v24, v5, s49, v24
; GENERIC-NEXT: v_fma_f32 v23, v5, s50, v23
; GENERIC-NEXT: v_fma_f32 v22, v5, s51, v22
; GENERIC-NEXT: s_waitcnt lgkmcnt(0)
; GENERIC-NEXT: v_fma_f32 v21, v5, s52, v21
; GENERIC-NEXT: v_fma_f32 v20, v5, s53, v20
; GENERIC-NEXT: v_fma_f32 v19, v5, s54, v19
; GENERIC-NEXT: v_fma_f32 v18, v5, s55, v18
; GENERIC-NEXT: v_fma_f32 v16, v5, s56, v16
; GENERIC-NEXT: v_fma_f32 v12, v5, s57, v12
; GENERIC-NEXT: v_fma_f32 v11, v5, s58, v11
; GENERIC-NEXT: v_fma_f32 v10, v5, s59, v10
; GENERIC-NEXT: v_fma_f32 v9, v5, s60, v9
; GENERIC-NEXT: v_fma_f32 v8, v5, s61, v8
; GENERIC-NEXT: v_fma_f32 v7, v5, s62, v7
; GENERIC-NEXT: v_fma_f32 v6, v5, s63, v6
; GENERIC-NEXT: v_fma_f32 v17, v5, s64, v17
; GENERIC-NEXT: v_fma_f32 v15, v5, s65, v15
; GENERIC-NEXT: v_fma_f32 v14, v5, s66, v14
; GENERIC-NEXT: v_fma_f32 v13, v5, s67, v13
; GENERIC-NEXT: v_fma_f32 v4, v5, s41, v4
; GENERIC-NEXT: v_fma_f32 v3, v5, s42, v3
; GENERIC-NEXT: s_cbranch_scc1 .LBB4_1
; GENERIC-NEXT: ; %bb.2: ; %for.cond.cleanup26
; GENERIC-NEXT: s_lshl_b32 s0, s8, 8
; GENERIC-NEXT: s_add_i32 s0, s6, s0
; GENERIC-NEXT: v_add_i32_e32 v0, vcc, s0, v0
; GENERIC-NEXT: s_mov_b32 s0, 0x5397829d
; GENERIC-NEXT: v_mul_hi_u32 v0, v0, s0
; GENERIC-NEXT: s_mov_b32 s0, 0x31000
; GENERIC-NEXT: v_mov_b32_e32 v1, 0
; GENERIC-NEXT: v_mov_b32_e32 v2, s3
; GENERIC-NEXT: v_lshrrev_b32_e32 v0, 4, v0
; GENERIC-NEXT: v_mul_lo_u32 v0, v0, s0
; GENERIC-NEXT: s_mov_b32 s7, 0xf000
; GENERIC-NEXT: s_mov_b32 s6, 0
; GENERIC-NEXT: s_mov_b64 s[4:5], s[2:3]
; GENERIC-NEXT: v_lshl_b64 v[0:1], v[0:1], 2
; GENERIC-NEXT: buffer_store_dword v34, v[0:1], s[4:7], 0 addr64 offset:1568
; GENERIC-NEXT: buffer_store_dword v33, v[0:1], s[4:7], 0 addr64 offset:2352
; GENERIC-NEXT: v_add_i32_e32 v37, vcc, s2, v0
; GENERIC-NEXT: v_addc_u32_e32 v38, vcc, v2, v1, vcc
; GENERIC-NEXT: s_movk_i32 s2, 0x310
; GENERIC-NEXT: s_waitcnt expcnt(0)
; GENERIC-NEXT: v_add_i32_e32 v33, vcc, s2, v37
; GENERIC-NEXT: s_mov_b32 s2, s6
; GENERIC-NEXT: s_mov_b32 s3, s6
; GENERIC-NEXT: buffer_store_dword v36, v[0:1], s[4:7], 0 addr64
; GENERIC-NEXT: buffer_store_dword v35, v[0:1], s[4:7], 0 addr64 offset:784
; GENERIC-NEXT: v_addc_u32_e32 v34, vcc, 0, v38, vcc
; GENERIC-NEXT: v_mov_b32_e32 v5, s6
; GENERIC-NEXT: s_waitcnt expcnt(1)
; GENERIC-NEXT: v_mov_b32_e32 v36, s3
; GENERIC-NEXT: s_waitcnt expcnt(0)
; GENERIC-NEXT: v_mov_b32_e32 v35, s2
; GENERIC-NEXT: s_movk_i32 s2, 0x620
; GENERIC-NEXT: buffer_store_dwordx2 v[4:5], v[0:1], s[4:7], 0 addr64 offset:3920
; GENERIC-NEXT: buffer_store_dwordx2 v[35:36], v[0:1], s[4:7], 0 addr64 offset:840
; GENERIC-NEXT: buffer_store_dword v32, v[0:1], s[4:7], 0 addr64 offset:3136
; GENERIC-NEXT: v_add_i32_e32 v0, vcc, s2, v37
; GENERIC-NEXT: v_addc_u32_e32 v1, vcc, 0, v38, vcc
; GENERIC-NEXT: s_movk_i32 s4, 0x930
; GENERIC-NEXT: s_mov_b64 s[0:1], 0x310
; GENERIC-NEXT: s_mov_b64 s[2:3], s[6:7]
; GENERIC-NEXT: s_waitcnt expcnt(2)
; GENERIC-NEXT: v_mov_b32_e32 v4, s6
; GENERIC-NEXT: v_add_i32_e32 v2, vcc, s4, v37
; GENERIC-NEXT: buffer_store_dwordx2 v[3:4], v[37:38], s[0:3], 0 addr64 offset:3920
; GENERIC-NEXT: s_waitcnt expcnt(0)
; GENERIC-NEXT: v_addc_u32_e32 v3, vcc, 0, v38, vcc
; GENERIC-NEXT: s_movk_i32 s0, 0xc40
; GENERIC-NEXT: v_add_i32_e32 v4, vcc, s0, v37
; GENERIC-NEXT: s_mov_b64 s[8:9], 0x620
; GENERIC-NEXT: s_mov_b64 s[10:11], s[6:7]
; GENERIC-NEXT: v_addc_u32_e32 v5, vcc, 0, v38, vcc
; GENERIC-NEXT: s_movk_i32 s4, 0xf50
; GENERIC-NEXT: buffer_store_dword v30, v[37:38], s[8:11], 0 addr64 offset:3920
; GENERIC-NEXT: s_waitcnt expcnt(0)
; GENERIC-NEXT: v_add_i32_e32 v30, vcc, s4, v37
; GENERIC-NEXT: s_mov_b64 s[16:17], 0xc40
; GENERIC-NEXT: v_addc_u32_e32 v31, vcc, 0, v38, vcc
; GENERIC-NEXT: s_mov_b64 s[18:19], s[6:7]
; GENERIC-NEXT: s_mov_b64 s[12:13], 0x930
; GENERIC-NEXT: s_mov_b64 s[14:15], s[6:7]
; GENERIC-NEXT: buffer_store_dword v28, v[37:38], s[16:19], 0 addr64 offset:3920
; GENERIC-NEXT: s_waitcnt expcnt(0)
; GENERIC-NEXT: v_add_i32_e32 v28, vcc, s4, v33
; GENERIC-NEXT: buffer_store_dword v29, v[37:38], s[12:15], 0 addr64 offset:3920
; GENERIC-NEXT: s_waitcnt expcnt(0)
; GENERIC-NEXT: v_addc_u32_e32 v29, vcc, 0, v34, vcc
; GENERIC-NEXT: buffer_store_dwordx2 v[35:36], v[37:38], s[12:15], 0 addr64 offset:3976
; GENERIC-NEXT: s_waitcnt expcnt(0)
; GENERIC-NEXT: v_add_i32_e32 v35, vcc, s4, v0
; GENERIC-NEXT: s_mov_b64 s[0:1], 0xf50
; GENERIC-NEXT: v_addc_u32_e32 v36, vcc, 0, v1, vcc
; GENERIC-NEXT: buffer_store_dword v26, v[33:34], s[0:3], 0 addr64 offset:3920
; GENERIC-NEXT: s_waitcnt expcnt(0)
; GENERIC-NEXT: v_add_i32_e32 v26, vcc, s4, v2
; GENERIC-NEXT: buffer_store_dword v27, v[37:38], s[0:3], 0 addr64 offset:3920
; GENERIC-NEXT: s_waitcnt expcnt(0)
; GENERIC-NEXT: v_addc_u32_e32 v27, vcc, 0, v3, vcc
; GENERIC-NEXT: v_add_i32_e32 v32, vcc, s4, v4
; GENERIC-NEXT: v_addc_u32_e32 v33, vcc, 0, v5, vcc
; GENERIC-NEXT: buffer_store_dword v25, v[0:1], s[0:3], 0 addr64 offset:3920
; GENERIC-NEXT: v_add_i32_e32 v0, vcc, s4, v30
; GENERIC-NEXT: v_addc_u32_e32 v1, vcc, 0, v31, vcc
; GENERIC-NEXT: buffer_store_dword v24, v[2:3], s[0:3], 0 addr64 offset:3920
; GENERIC-NEXT: v_add_i32_e32 v2, vcc, s4, v28
; GENERIC-NEXT: v_addc_u32_e32 v3, vcc, 0, v29, vcc
; GENERIC-NEXT: s_waitcnt expcnt(0)
; GENERIC-NEXT: v_add_i32_e32 v24, vcc, s4, v35
; GENERIC-NEXT: v_addc_u32_e32 v25, vcc, 0, v36, vcc
; GENERIC-NEXT: buffer_store_dword v23, v[4:5], s[0:3], 0 addr64 offset:3920
; GENERIC-NEXT: v_add_i32_e32 v4, vcc, s4, v26
; GENERIC-NEXT: v_addc_u32_e32 v5, vcc, 0, v27, vcc
; GENERIC-NEXT: buffer_store_dword v22, v[30:31], s[0:3], 0 addr64 offset:3920
; GENERIC-NEXT: s_waitcnt expcnt(0)
; GENERIC-NEXT: v_add_i32_e32 v22, vcc, s4, v32
; GENERIC-NEXT: v_addc_u32_e32 v23, vcc, 0, v33, vcc
; GENERIC-NEXT: v_add_i32_e32 v30, vcc, s4, v0
; GENERIC-NEXT: v_addc_u32_e32 v31, vcc, 0, v1, vcc
; GENERIC-NEXT: buffer_store_dword v21, v[28:29], s[0:3], 0 addr64 offset:3920
; GENERIC-NEXT: v_add_i32_e32 v28, vcc, s4, v2
; GENERIC-NEXT: v_addc_u32_e32 v29, vcc, 0, v3, vcc
; GENERIC-NEXT: buffer_store_dword v20, v[35:36], s[0:3], 0 addr64 offset:3920
; GENERIC-NEXT: s_waitcnt expcnt(0)
; GENERIC-NEXT: v_add_i32_e32 v20, vcc, s4, v24
; GENERIC-NEXT: v_addc_u32_e32 v21, vcc, 0, v25, vcc
; GENERIC-NEXT: v_add_i32_e32 v34, vcc, s4, v4
; GENERIC-NEXT: v_addc_u32_e32 v35, vcc, 0, v5, vcc
; GENERIC-NEXT: buffer_store_dword v19, v[26:27], s[0:3], 0 addr64 offset:3920
; GENERIC-NEXT: v_add_i32_e32 v26, vcc, s4, v22
; GENERIC-NEXT: v_addc_u32_e32 v27, vcc, 0, v23, vcc
; GENERIC-NEXT: buffer_store_dword v18, v[32:33], s[0:3], 0 addr64 offset:3920
; GENERIC-NEXT: s_waitcnt expcnt(0)
; GENERIC-NEXT: v_add_i32_e32 v18, vcc, s4, v30
; GENERIC-NEXT: v_addc_u32_e32 v19, vcc, 0, v31, vcc
; GENERIC-NEXT: buffer_store_dword v16, v[0:1], s[0:3], 0 addr64 offset:3920
; GENERIC-NEXT: v_add_i32_e32 v0, vcc, 0xf50, v28
; GENERIC-NEXT: v_addc_u32_e32 v1, vcc, 0, v29, vcc
; GENERIC-NEXT: buffer_store_dword v12, v[2:3], s[0:3], 0 addr64 offset:3920
; GENERIC-NEXT: buffer_store_dword v11, v[24:25], s[0:3], 0 addr64 offset:3920
; GENERIC-NEXT: buffer_store_dword v10, v[4:5], s[0:3], 0 addr64 offset:3920
; GENERIC-NEXT: buffer_store_dword v9, v[22:23], s[0:3], 0 addr64 offset:3920
; GENERIC-NEXT: buffer_store_dword v8, v[30:31], s[0:3], 0 addr64 offset:3920
; GENERIC-NEXT: buffer_store_dword v7, v[28:29], s[0:3], 0 addr64 offset:3920
; GENERIC-NEXT: buffer_store_dword v6, v[20:21], s[0:3], 0 addr64 offset:3920
; GENERIC-NEXT: buffer_store_dword v17, v[34:35], s[0:3], 0 addr64 offset:3920
; GENERIC-NEXT: buffer_store_dword v15, v[26:27], s[0:3], 0 addr64 offset:3920
; GENERIC-NEXT: buffer_store_dword v14, v[18:19], s[0:3], 0 addr64 offset:3920
; GENERIC-NEXT: buffer_store_dword v13, v[0:1], s[0:3], 0 addr64 offset:3920
; GENERIC-NEXT: s_endpgm
;
; GENERIC-GCNTRACKERS-LABEL: excess_soft_clause_reg_pressure:
; GENERIC-GCNTRACKERS: ; %bb.0: ; %entry
; GENERIC-GCNTRACKERS-NEXT: s_load_dwordx8 s[0:7], s[4:5], 0x9
; GENERIC-GCNTRACKERS-NEXT: v_lshlrev_b32_e32 v1, 2, v0
; GENERIC-GCNTRACKERS-NEXT: s_mov_b32 s30, 0
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v3, 0
; GENERIC-GCNTRACKERS-NEXT: s_mov_b32 s31, 0xf000
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt lgkmcnt(0)
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v2, s5
; GENERIC-GCNTRACKERS-NEXT: v_add_i32_e32 v1, vcc, s4, v1
; GENERIC-GCNTRACKERS-NEXT: v_addc_u32_e32 v2, vcc, 0, v2, vcc
; GENERIC-GCNTRACKERS-NEXT: s_movk_i32 s4, 0x1000
; GENERIC-GCNTRACKERS-NEXT: s_movk_i32 s5, 0x2000
; GENERIC-GCNTRACKERS-NEXT: s_movk_i32 s7, 0x3000
; GENERIC-GCNTRACKERS-NEXT: s_movk_i32 s9, 0x1040
; GENERIC-GCNTRACKERS-NEXT: s_movk_i32 s10, 0x2040
; GENERIC-GCNTRACKERS-NEXT: s_movk_i32 s11, 0x3040
; GENERIC-GCNTRACKERS-NEXT: s_movk_i32 s33, 0x310
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v6, 0
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v7, 0
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v8, 0
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v9, 0
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v10, 0
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v11, 0
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v12, 0
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v16, 0
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v18, 0
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v19, 0
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v20, 0
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v21, 0
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v22, 0
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v23, 0
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v24, 0
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v25, 0
; GENERIC-GCNTRACKERS-NEXT: s_mov_b32 s34, s30
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v26, 0
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v27, 0
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v28, 0
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v29, 0
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v30, 0
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v4, 0
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v31, 0
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v33, 0
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v34, 0
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v35, 0
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v37, 0
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v17, 0
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v15, 0
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v14, 0
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v13, 0
; GENERIC-GCNTRACKERS-NEXT: .LBB4_1: ; %for.cond28.preheader
; GENERIC-GCNTRACKERS-NEXT: ; =>This Inner Loop Header: Depth=1
; GENERIC-GCNTRACKERS-NEXT: s_mov_b32 s28, s30
; GENERIC-GCNTRACKERS-NEXT: s_mov_b32 s29, s30
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v38, v[1:2], s[28:31], 0 addr64
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v36, v[1:2], s[28:31], 0 addr64 offset:196
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v32, v[1:2], s[28:31], 0 addr64 offset:392
; GENERIC-GCNTRACKERS-NEXT: buffer_load_dword v5, v[1:2], s[28:31], 0 addr64 offset:588
; GENERIC-GCNTRACKERS-NEXT: s_load_dwordx16 s[36:51], s[0:1], 0x0
; GENERIC-GCNTRACKERS-NEXT: s_load_dwordx16 s[12:27], s[0:1], 0x10
; GENERIC-GCNTRACKERS-NEXT: v_add_i32_e32 v1, vcc, s33, v1
; GENERIC-GCNTRACKERS-NEXT: v_addc_u32_e32 v2, vcc, 0, v2, vcc
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt vmcnt(3) lgkmcnt(0)
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v37, v38, s36, v37
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v35, v38, s37, v35
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v34, v38, s38, v34
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v33, v38, s39, v33
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v31, v38, s40, v31
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v4, v38, s41, v4
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v3, v38, s42, v3
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v30, v38, s43, v30
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v29, v38, s44, v29
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v28, v38, s45, v28
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v27, v38, s46, v27
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v26, v38, s47, v26
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v25, v38, s48, v25
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v24, v38, s49, v24
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v23, v38, s50, v23
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v22, v38, s51, v22
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v21, v38, s12, v21
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v20, v38, s13, v20
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v19, v38, s14, v19
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v18, v38, s15, v18
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v16, v38, s16, v16
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v12, v38, s17, v12
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v11, v38, s18, v11
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v10, v38, s19, v10
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v9, v38, s20, v9
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v8, v38, s21, v8
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v7, v38, s22, v7
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v6, v38, s23, v6
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v17, v38, s24, v17
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v15, v38, s25, v15
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v14, v38, s26, v14
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v13, v38, s27, v13
; GENERIC-GCNTRACKERS-NEXT: s_load_dwordx16 s[36:51], s[0:1], s4
; GENERIC-GCNTRACKERS-NEXT: s_load_dwordx16 s[12:27], s[0:1], s9
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt vmcnt(2) lgkmcnt(0)
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v37, v36, s36, v37
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v35, v36, s37, v35
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v34, v36, s38, v34
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v33, v36, s39, v33
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v31, v36, s40, v31
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v4, v36, s41, v4
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v3, v36, s42, v3
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v30, v36, s43, v30
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v29, v36, s44, v29
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v28, v36, s45, v28
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v27, v36, s46, v27
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v26, v36, s47, v26
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v25, v36, s48, v25
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v24, v36, s49, v24
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v23, v36, s50, v23
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v22, v36, s51, v22
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v21, v36, s12, v21
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v20, v36, s13, v20
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v19, v36, s14, v19
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v18, v36, s15, v18
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v16, v36, s16, v16
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v12, v36, s17, v12
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v11, v36, s18, v11
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v10, v36, s19, v10
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v9, v36, s20, v9
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v8, v36, s21, v8
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v7, v36, s22, v7
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v6, v36, s23, v6
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v17, v36, s24, v17
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v15, v36, s25, v15
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v14, v36, s26, v14
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v13, v36, s27, v13
; GENERIC-GCNTRACKERS-NEXT: s_load_dwordx16 s[36:51], s[0:1], s5
; GENERIC-GCNTRACKERS-NEXT: s_load_dwordx16 s[12:27], s[0:1], s10
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt vmcnt(1) lgkmcnt(0)
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v36, v32, s36, v37
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v35, v32, s37, v35
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v34, v32, s38, v34
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v33, v32, s39, v33
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v31, v32, s40, v31
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v4, v32, s41, v4
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v3, v32, s42, v3
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v30, v32, s43, v30
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v29, v32, s44, v29
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v28, v32, s45, v28
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v27, v32, s46, v27
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v26, v32, s47, v26
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v25, v32, s48, v25
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v24, v32, s49, v24
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v23, v32, s50, v23
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v22, v32, s51, v22
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v21, v32, s12, v21
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v20, v32, s13, v20
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v19, v32, s14, v19
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v18, v32, s15, v18
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v16, v32, s16, v16
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v12, v32, s17, v12
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v11, v32, s18, v11
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v10, v32, s19, v10
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v9, v32, s20, v9
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v8, v32, s21, v8
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v7, v32, s22, v7
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v6, v32, s23, v6
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v17, v32, s24, v17
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v15, v32, s25, v15
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v14, v32, s26, v14
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v13, v32, s27, v13
; GENERIC-GCNTRACKERS-NEXT: s_load_dwordx16 s[36:51], s[0:1], s7
; GENERIC-GCNTRACKERS-NEXT: s_load_dwordx16 s[12:27], s[0:1], s11
; GENERIC-GCNTRACKERS-NEXT: s_add_u32 s0, s0, 0x4000
; GENERIC-GCNTRACKERS-NEXT: s_addc_u32 s1, s1, 0
; GENERIC-GCNTRACKERS-NEXT: s_add_i32 s34, s34, 1
; GENERIC-GCNTRACKERS-NEXT: s_cmpk_lg_i32 s34, 0x200
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v37, v5, s36, v36
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v35, v5, s37, v35
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v34, v5, s38, v34
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v33, v5, s39, v33
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v31, v5, s40, v31
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v4, v5, s41, v4
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v3, v5, s42, v3
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v30, v5, s43, v30
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v29, v5, s44, v29
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v28, v5, s45, v28
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v27, v5, s46, v27
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v26, v5, s47, v26
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v25, v5, s48, v25
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v24, v5, s49, v24
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v23, v5, s50, v23
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v22, v5, s51, v22
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v21, v5, s12, v21
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v20, v5, s13, v20
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v19, v5, s14, v19
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v18, v5, s15, v18
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v16, v5, s16, v16
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v12, v5, s17, v12
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v11, v5, s18, v11
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v10, v5, s19, v10
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v9, v5, s20, v9
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v8, v5, s21, v8
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v7, v5, s22, v7
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v6, v5, s23, v6
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v17, v5, s24, v17
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v15, v5, s25, v15
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v14, v5, s26, v14
; GENERIC-GCNTRACKERS-NEXT: v_fma_f32 v13, v5, s27, v13
; GENERIC-GCNTRACKERS-NEXT: s_cbranch_scc1 .LBB4_1
; GENERIC-GCNTRACKERS-NEXT: ; %bb.2: ; %for.cond.cleanup26
; GENERIC-GCNTRACKERS-NEXT: s_lshl_b32 s0, s8, 8
; GENERIC-GCNTRACKERS-NEXT: s_add_i32 s0, s6, s0
; GENERIC-GCNTRACKERS-NEXT: v_add_i32_e32 v0, vcc, s0, v0
; GENERIC-GCNTRACKERS-NEXT: s_mov_b32 s0, 0x5397829d
; GENERIC-GCNTRACKERS-NEXT: v_mul_hi_u32 v0, v0, s0
; GENERIC-GCNTRACKERS-NEXT: s_mov_b32 s0, 0x31000
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v1, 0
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v2, s3
; GENERIC-GCNTRACKERS-NEXT: v_lshrrev_b32_e32 v0, 4, v0
; GENERIC-GCNTRACKERS-NEXT: v_mul_lo_u32 v0, v0, s0
; GENERIC-GCNTRACKERS-NEXT: s_mov_b32 s7, 0xf000
; GENERIC-GCNTRACKERS-NEXT: s_mov_b32 s6, 0
; GENERIC-GCNTRACKERS-NEXT: s_mov_b64 s[4:5], s[2:3]
; GENERIC-GCNTRACKERS-NEXT: v_lshl_b64 v[0:1], v[0:1], 2
; GENERIC-GCNTRACKERS-NEXT: s_mov_b32 s3, s6
; GENERIC-GCNTRACKERS-NEXT: v_add_i32_e32 v38, vcc, s2, v0
; GENERIC-GCNTRACKERS-NEXT: v_addc_u32_e32 v39, vcc, v2, v1, vcc
; GENERIC-GCNTRACKERS-NEXT: s_movk_i32 s2, 0x310
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v35, v[0:1], s[4:7], 0 addr64 offset:784
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt expcnt(0)
; GENERIC-GCNTRACKERS-NEXT: v_add_i32_e32 v35, vcc, s2, v38
; GENERIC-GCNTRACKERS-NEXT: s_mov_b32 s2, s6
; GENERIC-GCNTRACKERS-NEXT: v_addc_u32_e32 v36, vcc, 0, v39, vcc
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v34, v[0:1], s[4:7], 0 addr64 offset:1568
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v33, v[0:1], s[4:7], 0 addr64 offset:2352
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt expcnt(0)
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v33, s3
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v32, s2
; GENERIC-GCNTRACKERS-NEXT: s_movk_i32 s2, 0x620
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v5, s6
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v37, v[0:1], s[4:7], 0 addr64
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dwordx2 v[32:33], v[0:1], s[4:7], 0 addr64 offset:840
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v31, v[0:1], s[4:7], 0 addr64 offset:3136
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dwordx2 v[4:5], v[0:1], s[4:7], 0 addr64 offset:3920
; GENERIC-GCNTRACKERS-NEXT: v_add_i32_e32 v0, vcc, s2, v38
; GENERIC-GCNTRACKERS-NEXT: v_addc_u32_e32 v1, vcc, 0, v39, vcc
; GENERIC-GCNTRACKERS-NEXT: s_movk_i32 s4, 0x930
; GENERIC-GCNTRACKERS-NEXT: s_mov_b64 s[0:1], 0x310
; GENERIC-GCNTRACKERS-NEXT: s_mov_b64 s[2:3], s[6:7]
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt expcnt(0)
; GENERIC-GCNTRACKERS-NEXT: v_mov_b32_e32 v4, s6
; GENERIC-GCNTRACKERS-NEXT: v_add_i32_e32 v2, vcc, s4, v38
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dwordx2 v[3:4], v[38:39], s[0:3], 0 addr64 offset:3920
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt expcnt(0)
; GENERIC-GCNTRACKERS-NEXT: v_addc_u32_e32 v3, vcc, 0, v39, vcc
; GENERIC-GCNTRACKERS-NEXT: s_movk_i32 s0, 0xc40
; GENERIC-GCNTRACKERS-NEXT: v_add_i32_e32 v4, vcc, s0, v38
; GENERIC-GCNTRACKERS-NEXT: s_mov_b64 s[8:9], 0x620
; GENERIC-GCNTRACKERS-NEXT: s_mov_b64 s[10:11], s[6:7]
; GENERIC-GCNTRACKERS-NEXT: v_addc_u32_e32 v5, vcc, 0, v39, vcc
; GENERIC-GCNTRACKERS-NEXT: s_movk_i32 s4, 0xf50
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v30, v[38:39], s[8:11], 0 addr64 offset:3920
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt expcnt(0)
; GENERIC-GCNTRACKERS-NEXT: v_add_i32_e32 v30, vcc, s4, v38
; GENERIC-GCNTRACKERS-NEXT: s_mov_b64 s[16:17], 0xc40
; GENERIC-GCNTRACKERS-NEXT: v_addc_u32_e32 v31, vcc, 0, v39, vcc
; GENERIC-GCNTRACKERS-NEXT: s_mov_b64 s[18:19], s[6:7]
; GENERIC-GCNTRACKERS-NEXT: s_mov_b64 s[12:13], 0x930
; GENERIC-GCNTRACKERS-NEXT: s_mov_b64 s[14:15], s[6:7]
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v28, v[38:39], s[16:19], 0 addr64 offset:3920
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt expcnt(0)
; GENERIC-GCNTRACKERS-NEXT: v_add_i32_e32 v28, vcc, s4, v35
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v29, v[38:39], s[12:15], 0 addr64 offset:3920
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt expcnt(0)
; GENERIC-GCNTRACKERS-NEXT: v_addc_u32_e32 v29, vcc, 0, v36, vcc
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dwordx2 v[32:33], v[38:39], s[12:15], 0 addr64 offset:3976
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt expcnt(0)
; GENERIC-GCNTRACKERS-NEXT: v_add_i32_e32 v32, vcc, s4, v0
; GENERIC-GCNTRACKERS-NEXT: s_mov_b64 s[0:1], 0xf50
; GENERIC-GCNTRACKERS-NEXT: v_addc_u32_e32 v33, vcc, 0, v1, vcc
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v26, v[35:36], s[0:3], 0 addr64 offset:3920
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt expcnt(0)
; GENERIC-GCNTRACKERS-NEXT: v_add_i32_e32 v26, vcc, s4, v2
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v27, v[38:39], s[0:3], 0 addr64 offset:3920
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt expcnt(0)
; GENERIC-GCNTRACKERS-NEXT: v_addc_u32_e32 v27, vcc, 0, v3, vcc
; GENERIC-GCNTRACKERS-NEXT: v_add_i32_e32 v34, vcc, s4, v4
; GENERIC-GCNTRACKERS-NEXT: v_addc_u32_e32 v35, vcc, 0, v5, vcc
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v25, v[0:1], s[0:3], 0 addr64 offset:3920
; GENERIC-GCNTRACKERS-NEXT: v_add_i32_e32 v0, vcc, s4, v30
; GENERIC-GCNTRACKERS-NEXT: v_addc_u32_e32 v1, vcc, 0, v31, vcc
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v24, v[2:3], s[0:3], 0 addr64 offset:3920
; GENERIC-GCNTRACKERS-NEXT: v_add_i32_e32 v2, vcc, s4, v28
; GENERIC-GCNTRACKERS-NEXT: v_addc_u32_e32 v3, vcc, 0, v29, vcc
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt expcnt(0)
; GENERIC-GCNTRACKERS-NEXT: v_add_i32_e32 v24, vcc, s4, v32
; GENERIC-GCNTRACKERS-NEXT: v_addc_u32_e32 v25, vcc, 0, v33, vcc
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v23, v[4:5], s[0:3], 0 addr64 offset:3920
; GENERIC-GCNTRACKERS-NEXT: v_add_i32_e32 v4, vcc, s4, v26
; GENERIC-GCNTRACKERS-NEXT: v_addc_u32_e32 v5, vcc, 0, v27, vcc
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v22, v[30:31], s[0:3], 0 addr64 offset:3920
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt expcnt(0)
; GENERIC-GCNTRACKERS-NEXT: v_add_i32_e32 v22, vcc, s4, v34
; GENERIC-GCNTRACKERS-NEXT: v_addc_u32_e32 v23, vcc, 0, v35, vcc
; GENERIC-GCNTRACKERS-NEXT: v_add_i32_e32 v30, vcc, s4, v0
; GENERIC-GCNTRACKERS-NEXT: v_addc_u32_e32 v31, vcc, 0, v1, vcc
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v21, v[28:29], s[0:3], 0 addr64 offset:3920
; GENERIC-GCNTRACKERS-NEXT: v_add_i32_e32 v28, vcc, s4, v2
; GENERIC-GCNTRACKERS-NEXT: v_addc_u32_e32 v29, vcc, 0, v3, vcc
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v20, v[32:33], s[0:3], 0 addr64 offset:3920
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt expcnt(0)
; GENERIC-GCNTRACKERS-NEXT: v_add_i32_e32 v20, vcc, s4, v24
; GENERIC-GCNTRACKERS-NEXT: v_addc_u32_e32 v21, vcc, 0, v25, vcc
; GENERIC-GCNTRACKERS-NEXT: v_add_i32_e32 v32, vcc, s4, v4
; GENERIC-GCNTRACKERS-NEXT: v_addc_u32_e32 v33, vcc, 0, v5, vcc
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v19, v[26:27], s[0:3], 0 addr64 offset:3920
; GENERIC-GCNTRACKERS-NEXT: v_add_i32_e32 v26, vcc, s4, v22
; GENERIC-GCNTRACKERS-NEXT: v_addc_u32_e32 v27, vcc, 0, v23, vcc
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v18, v[34:35], s[0:3], 0 addr64 offset:3920
; GENERIC-GCNTRACKERS-NEXT: s_waitcnt expcnt(0)
; GENERIC-GCNTRACKERS-NEXT: v_add_i32_e32 v18, vcc, s4, v30
; GENERIC-GCNTRACKERS-NEXT: v_addc_u32_e32 v19, vcc, 0, v31, vcc
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v16, v[0:1], s[0:3], 0 addr64 offset:3920
; GENERIC-GCNTRACKERS-NEXT: v_add_i32_e32 v0, vcc, 0xf50, v28
; GENERIC-GCNTRACKERS-NEXT: v_addc_u32_e32 v1, vcc, 0, v29, vcc
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v12, v[2:3], s[0:3], 0 addr64 offset:3920
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v11, v[24:25], s[0:3], 0 addr64 offset:3920
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v10, v[4:5], s[0:3], 0 addr64 offset:3920
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v9, v[22:23], s[0:3], 0 addr64 offset:3920
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v8, v[30:31], s[0:3], 0 addr64 offset:3920
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v7, v[28:29], s[0:3], 0 addr64 offset:3920
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v6, v[20:21], s[0:3], 0 addr64 offset:3920
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v17, v[32:33], s[0:3], 0 addr64 offset:3920
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v15, v[26:27], s[0:3], 0 addr64 offset:3920
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v14, v[18:19], s[0:3], 0 addr64 offset:3920
; GENERIC-GCNTRACKERS-NEXT: buffer_store_dword v13, v[0:1], s[0:3], 0 addr64 offset:3920
; GENERIC-GCNTRACKERS-NEXT: s_endpgm
entry:
%i = tail call ptr addrspace(4) @llvm.amdgcn.implicitarg.ptr()
%i2 = load i64, ptr addrspace(4) %i, align 8
%i3 = tail call i32 @llvm.amdgcn.workgroup.id.x()
%i4 = shl i32 %i3, 8
%i5 = tail call i32 @llvm.amdgcn.workitem.id.x(), !range !5
%i6 = add i32 %i4, %i5
%i7 = trunc i64 %i2 to i32
%conv = add i32 %i6, %i7
%conv.frozen = freeze i32 %conv
%div = udiv i32 %conv.frozen, 49
%add.ptr22 = getelementptr inbounds float, ptr addrspace(4) %wei_ptr, i64 0
%in.ptr1 = getelementptr inbounds float, ptr addrspace(1) %in, i32 %i5
br label %for.cond28.preheader
for.cond28.preheader: ; preds = %for.cond28.preheader, %entry
%accum.sroa.110.0 = phi float [ 0.000000e+00, %entry ], [ %i251, %for.cond28.preheader ]
%accum.sroa.106.0 = phi float [ 0.000000e+00, %entry ], [ %i247, %for.cond28.preheader ]
%accum.sroa.102.0 = phi float [ 0.000000e+00, %entry ], [ %i243, %for.cond28.preheader ]
%accum.sroa.98.0 = phi float [ 0.000000e+00, %entry ], [ %i239, %for.cond28.preheader ]
%accum.sroa.94.0 = phi float [ 0.000000e+00, %entry ], [ %i235, %for.cond28.preheader ]
%accum.sroa.90.0 = phi float [ 0.000000e+00, %entry ], [ %i231, %for.cond28.preheader ]
%accum.sroa.86.0 = phi float [ 0.000000e+00, %entry ], [ %i227, %for.cond28.preheader ]
%accum.sroa.82.0 = phi float [ 0.000000e+00, %entry ], [ %i223, %for.cond28.preheader ]
%accum.sroa.78.0 = phi float [ 0.000000e+00, %entry ], [ %i219, %for.cond28.preheader ]
%accum.sroa.74.0 = phi float [ 0.000000e+00, %entry ], [ %i215, %for.cond28.preheader ]
%accum.sroa.70.0 = phi float [ 0.000000e+00, %entry ], [ %i211, %for.cond28.preheader ]
%accum.sroa.66.0 = phi float [ 0.000000e+00, %entry ], [ %i207, %for.cond28.preheader ]
%accum.sroa.62.0 = phi float [ 0.000000e+00, %entry ], [ %i203, %for.cond28.preheader ]
%accum.sroa.58.0 = phi float [ 0.000000e+00, %entry ], [ %i199, %for.cond28.preheader ]
%accum.sroa.54.0 = phi float [ 0.000000e+00, %entry ], [ %i195, %for.cond28.preheader ]
%accum.sroa.50.0 = phi float [ 0.000000e+00, %entry ], [ %i191, %for.cond28.preheader ]
%accum.sroa.46.0 = phi float [ 0.000000e+00, %entry ], [ %i187, %for.cond28.preheader ]
%accum.sroa.42.0 = phi float [ 0.000000e+00, %entry ], [ %i183, %for.cond28.preheader ]
%accum.sroa.38.0 = phi float [ 0.000000e+00, %entry ], [ %i179, %for.cond28.preheader ]
%accum.sroa.34.0 = phi float [ 0.000000e+00, %entry ], [ %i175, %for.cond28.preheader ]
%accum.sroa.30.0 = phi float [ 0.000000e+00, %entry ], [ %i171, %for.cond28.preheader ]
%accum.sroa.26.0 = phi float [ 0.000000e+00, %entry ], [ %i167, %for.cond28.preheader ]
%accum.sroa.22.0 = phi float [ 0.000000e+00, %entry ], [ %i163, %for.cond28.preheader ]
%accum.sroa.18.0 = phi float [ 0.000000e+00, %entry ], [ %i159, %for.cond28.preheader ]
%accum.sroa.14.0 = phi float [ 0.000000e+00, %entry ], [ %i155, %for.cond28.preheader ]
%accum.sroa.10.0 = phi float [ 0.000000e+00, %entry ], [ %i151, %for.cond28.preheader ]
%accum.sroa.6.0 = phi float [ 0.000000e+00, %entry ], [ %i147, %for.cond28.preheader ]
%accum.sroa.0.0 = phi float [ 0.000000e+00, %entry ], [ %i143, %for.cond28.preheader ]
%accum.sroa.114.0 = phi float [ 0.000000e+00, %entry ], [ %i255, %for.cond28.preheader ]
%accum.sroa.118.0 = phi float [ 0.000000e+00, %entry ], [ %i259, %for.cond28.preheader ]
%accum.sroa.122.0 = phi float [ 0.000000e+00, %entry ], [ %i263, %for.cond28.preheader ]
%accum.sroa.126.0 = phi float [ 0.000000e+00, %entry ], [ %i267, %for.cond28.preheader ]
%i_ptr.0288 = phi ptr addrspace(1) [ %in.ptr1, %entry ], [ %add.ptr47.3, %for.cond28.preheader ]
%w_ptr.0287 = phi ptr addrspace(4) [ %add.ptr22, %entry ], [ %add.ptr74, %for.cond28.preheader ]
%ci.0286 = phi i32 [ 0, %entry ], [ %inc116, %for.cond28.preheader ]
%i8 = load float, ptr addrspace(1) %i_ptr.0288, align 4
%add.ptr47 = getelementptr inbounds float, ptr addrspace(1) %i_ptr.0288, i64 49
%i9 = load float, ptr addrspace(1) %add.ptr47, align 4
%add.ptr47.1 = getelementptr inbounds float, ptr addrspace(1) %i_ptr.0288, i64 98
%i10 = load float, ptr addrspace(1) %add.ptr47.1, align 4
%add.ptr47.2 = getelementptr inbounds float, ptr addrspace(1) %i_ptr.0288, i64 147
%i11 = load float, ptr addrspace(1) %add.ptr47.2, align 4
%i12 = load float, ptr addrspace(4) %w_ptr.0287, align 4
%add.ptr66 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 1024
%i13 = load float, ptr addrspace(4) %add.ptr66, align 4
%add.ptr66.1 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 2048
%i14 = load float, ptr addrspace(4) %add.ptr66.1, align 4
%add.ptr66.2 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 3072
%i15 = load float, ptr addrspace(4) %add.ptr66.2, align 4
%add.ptr70 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 1
%i16 = load float, ptr addrspace(4) %add.ptr70, align 4
%add.ptr66.1291 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 1025
%i17 = load float, ptr addrspace(4) %add.ptr66.1291, align 4
%add.ptr66.1.1 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 2049
%i18 = load float, ptr addrspace(4) %add.ptr66.1.1, align 4
%add.ptr66.2.1 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 3073
%i19 = load float, ptr addrspace(4) %add.ptr66.2.1, align 4
%add.ptr70.1 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 2
%i20 = load float, ptr addrspace(4) %add.ptr70.1, align 4
%add.ptr66.2293 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 1026
%i21 = load float, ptr addrspace(4) %add.ptr66.2293, align 4
%add.ptr66.1.2 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 2050
%i22 = load float, ptr addrspace(4) %add.ptr66.1.2, align 4
%add.ptr66.2.2 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 3074
%i23 = load float, ptr addrspace(4) %add.ptr66.2.2, align 4
%add.ptr70.2 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 3
%i24 = load float, ptr addrspace(4) %add.ptr70.2, align 4
%add.ptr66.3 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 1027
%i25 = load float, ptr addrspace(4) %add.ptr66.3, align 4
%add.ptr66.1.3 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 2051
%i26 = load float, ptr addrspace(4) %add.ptr66.1.3, align 4
%add.ptr66.2.3 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 3075
%i27 = load float, ptr addrspace(4) %add.ptr66.2.3, align 4
%add.ptr70.3 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 4
%i28 = load float, ptr addrspace(4) %add.ptr70.3, align 4
%add.ptr66.4 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 1028
%i29 = load float, ptr addrspace(4) %add.ptr66.4, align 4
%add.ptr66.1.4 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 2052
%i30 = load float, ptr addrspace(4) %add.ptr66.1.4, align 4
%add.ptr66.2.4 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 3076
%i31 = load float, ptr addrspace(4) %add.ptr66.2.4, align 4
%add.ptr70.4 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 5
%i32 = load float, ptr addrspace(4) %add.ptr70.4, align 4
%add.ptr66.5 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 1029
%i33 = load float, ptr addrspace(4) %add.ptr66.5, align 4
%add.ptr66.1.5 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 2053
%i34 = load float, ptr addrspace(4) %add.ptr66.1.5, align 4
%add.ptr66.2.5 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 3077
%i35 = load float, ptr addrspace(4) %add.ptr66.2.5, align 4
%add.ptr70.5 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 6
%i36 = load float, ptr addrspace(4) %add.ptr70.5, align 4
%add.ptr66.6 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 1030
%i37 = load float, ptr addrspace(4) %add.ptr66.6, align 4
%add.ptr66.1.6 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 2054
%i38 = load float, ptr addrspace(4) %add.ptr66.1.6, align 4
%add.ptr66.2.6 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 3078
%i39 = load float, ptr addrspace(4) %add.ptr66.2.6, align 4
%add.ptr70.6 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 7
%i40 = load float, ptr addrspace(4) %add.ptr70.6, align 4
%add.ptr66.7 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 1031
%i41 = load float, ptr addrspace(4) %add.ptr66.7, align 4
%add.ptr66.1.7 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 2055
%i42 = load float, ptr addrspace(4) %add.ptr66.1.7, align 4
%add.ptr66.2.7 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 3079
%i43 = load float, ptr addrspace(4) %add.ptr66.2.7, align 4
%add.ptr70.7 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 8
%i44 = load float, ptr addrspace(4) %add.ptr70.7, align 4
%add.ptr66.8 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 1032
%i45 = load float, ptr addrspace(4) %add.ptr66.8, align 4
%add.ptr66.1.8 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 2056
%i46 = load float, ptr addrspace(4) %add.ptr66.1.8, align 4
%add.ptr66.2.8 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 3080
%i47 = load float, ptr addrspace(4) %add.ptr66.2.8, align 4
%add.ptr70.8 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 9
%i48 = load float, ptr addrspace(4) %add.ptr70.8, align 4
%add.ptr66.9 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 1033
%i49 = load float, ptr addrspace(4) %add.ptr66.9, align 4
%add.ptr66.1.9 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 2057
%i50 = load float, ptr addrspace(4) %add.ptr66.1.9, align 4
%add.ptr66.2.9 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 3081
%i51 = load float, ptr addrspace(4) %add.ptr66.2.9, align 4
%add.ptr70.9 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 10
%i52 = load float, ptr addrspace(4) %add.ptr70.9, align 4
%add.ptr66.10 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 1034
%i53 = load float, ptr addrspace(4) %add.ptr66.10, align 4
%add.ptr66.1.10 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 2058
%i54 = load float, ptr addrspace(4) %add.ptr66.1.10, align 4
%add.ptr66.2.10 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 3082
%i55 = load float, ptr addrspace(4) %add.ptr66.2.10, align 4
%add.ptr70.10 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 11
%i56 = load float, ptr addrspace(4) %add.ptr70.10, align 4
%add.ptr66.11 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 1035
%i57 = load float, ptr addrspace(4) %add.ptr66.11, align 4
%add.ptr66.1.11 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 2059
%i58 = load float, ptr addrspace(4) %add.ptr66.1.11, align 4
%add.ptr66.2.11 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 3083
%i59 = load float, ptr addrspace(4) %add.ptr66.2.11, align 4
%add.ptr70.11 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 12
%i60 = load float, ptr addrspace(4) %add.ptr70.11, align 4
%add.ptr66.12 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 1036
%i61 = load float, ptr addrspace(4) %add.ptr66.12, align 4
%add.ptr66.1.12 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 2060
%i62 = load float, ptr addrspace(4) %add.ptr66.1.12, align 4
%add.ptr66.2.12 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 3084
%i63 = load float, ptr addrspace(4) %add.ptr66.2.12, align 4
%add.ptr70.12 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 13
%i64 = load float, ptr addrspace(4) %add.ptr70.12, align 4
%add.ptr66.13 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 1037
%i65 = load float, ptr addrspace(4) %add.ptr66.13, align 4
%add.ptr66.1.13 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 2061
%i66 = load float, ptr addrspace(4) %add.ptr66.1.13, align 4
%add.ptr66.2.13 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 3085
%i67 = load float, ptr addrspace(4) %add.ptr66.2.13, align 4
%add.ptr70.13 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 14
%i68 = load float, ptr addrspace(4) %add.ptr70.13, align 4
%add.ptr66.14 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 1038
%i69 = load float, ptr addrspace(4) %add.ptr66.14, align 4
%add.ptr66.1.14 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 2062
%i70 = load float, ptr addrspace(4) %add.ptr66.1.14, align 4
%add.ptr66.2.14 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 3086
%i71 = load float, ptr addrspace(4) %add.ptr66.2.14, align 4
%add.ptr70.14 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 15
%i72 = load float, ptr addrspace(4) %add.ptr70.14, align 4
%add.ptr66.15 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 1039
%i73 = load float, ptr addrspace(4) %add.ptr66.15, align 4
%add.ptr66.1.15 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 2063
%i74 = load float, ptr addrspace(4) %add.ptr66.1.15, align 4
%add.ptr66.2.15 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 3087
%i75 = load float, ptr addrspace(4) %add.ptr66.2.15, align 4
%add.ptr70.15 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 16
%i76 = load float, ptr addrspace(4) %add.ptr70.15, align 4
%add.ptr66.16 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 1040
%i77 = load float, ptr addrspace(4) %add.ptr66.16, align 4
%add.ptr66.1.16 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 2064
%i78 = load float, ptr addrspace(4) %add.ptr66.1.16, align 4
%add.ptr66.2.16 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 3088
%i79 = load float, ptr addrspace(4) %add.ptr66.2.16, align 4
%add.ptr70.16 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 17
%i80 = load float, ptr addrspace(4) %add.ptr70.16, align 4
%add.ptr66.17 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 1041
%i81 = load float, ptr addrspace(4) %add.ptr66.17, align 4
%add.ptr66.1.17 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 2065
%i82 = load float, ptr addrspace(4) %add.ptr66.1.17, align 4
%add.ptr66.2.17 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 3089
%i83 = load float, ptr addrspace(4) %add.ptr66.2.17, align 4
%add.ptr70.17 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 18
%i84 = load float, ptr addrspace(4) %add.ptr70.17, align 4
%add.ptr66.18 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 1042
%i85 = load float, ptr addrspace(4) %add.ptr66.18, align 4
%add.ptr66.1.18 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 2066
%i86 = load float, ptr addrspace(4) %add.ptr66.1.18, align 4
%add.ptr66.2.18 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 3090
%i87 = load float, ptr addrspace(4) %add.ptr66.2.18, align 4
%add.ptr70.18 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 19
%i88 = load float, ptr addrspace(4) %add.ptr70.18, align 4
%add.ptr66.19 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 1043
%i89 = load float, ptr addrspace(4) %add.ptr66.19, align 4
%add.ptr66.1.19 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 2067
%i90 = load float, ptr addrspace(4) %add.ptr66.1.19, align 4
%add.ptr66.2.19 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 3091
%i91 = load float, ptr addrspace(4) %add.ptr66.2.19, align 4
%add.ptr70.19 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 20
%i92 = load float, ptr addrspace(4) %add.ptr70.19, align 4
%add.ptr66.20 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 1044
%i93 = load float, ptr addrspace(4) %add.ptr66.20, align 4
%add.ptr66.1.20 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 2068
%i94 = load float, ptr addrspace(4) %add.ptr66.1.20, align 4
%add.ptr66.2.20 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 3092
%i95 = load float, ptr addrspace(4) %add.ptr66.2.20, align 4
%add.ptr70.20 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 21
%i96 = load float, ptr addrspace(4) %add.ptr70.20, align 4
%add.ptr66.21 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 1045
%i97 = load float, ptr addrspace(4) %add.ptr66.21, align 4
%add.ptr66.1.21 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 2069
%i98 = load float, ptr addrspace(4) %add.ptr66.1.21, align 4
%add.ptr66.2.21 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 3093
%i99 = load float, ptr addrspace(4) %add.ptr66.2.21, align 4
%add.ptr70.21 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 22
%i100 = load float, ptr addrspace(4) %add.ptr70.21, align 4
%add.ptr66.22 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 1046
%i101 = load float, ptr addrspace(4) %add.ptr66.22, align 4
%add.ptr66.1.22 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 2070
%i102 = load float, ptr addrspace(4) %add.ptr66.1.22, align 4
%add.ptr66.2.22 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 3094
%i103 = load float, ptr addrspace(4) %add.ptr66.2.22, align 4
%add.ptr70.22 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 23
%i104 = load float, ptr addrspace(4) %add.ptr70.22, align 4
%add.ptr66.23 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 1047
%i105 = load float, ptr addrspace(4) %add.ptr66.23, align 4
%add.ptr66.1.23 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 2071
%i106 = load float, ptr addrspace(4) %add.ptr66.1.23, align 4
%add.ptr66.2.23 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 3095
%i107 = load float, ptr addrspace(4) %add.ptr66.2.23, align 4
%add.ptr70.23 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 24
%i108 = load float, ptr addrspace(4) %add.ptr70.23, align 4
%add.ptr66.24 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 1048
%i109 = load float, ptr addrspace(4) %add.ptr66.24, align 4
%add.ptr66.1.24 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 2072
%i110 = load float, ptr addrspace(4) %add.ptr66.1.24, align 4
%add.ptr66.2.24 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 3096
%i111 = load float, ptr addrspace(4) %add.ptr66.2.24, align 4
%add.ptr70.24 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 25
%i112 = load float, ptr addrspace(4) %add.ptr70.24, align 4
%add.ptr66.25 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 1049
%i113 = load float, ptr addrspace(4) %add.ptr66.25, align 4
%add.ptr66.1.25 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 2073
%i114 = load float, ptr addrspace(4) %add.ptr66.1.25, align 4
%add.ptr66.2.25 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 3097
%i115 = load float, ptr addrspace(4) %add.ptr66.2.25, align 4
%add.ptr70.25 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 26
%i116 = load float, ptr addrspace(4) %add.ptr70.25, align 4
%add.ptr66.26 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 1050
%i117 = load float, ptr addrspace(4) %add.ptr66.26, align 4
%add.ptr66.1.26 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 2074
%i118 = load float, ptr addrspace(4) %add.ptr66.1.26, align 4
%add.ptr66.2.26 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 3098
%i119 = load float, ptr addrspace(4) %add.ptr66.2.26, align 4
%add.ptr70.26 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 27
%i120 = load float, ptr addrspace(4) %add.ptr70.26, align 4
%add.ptr66.27 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 1051
%i121 = load float, ptr addrspace(4) %add.ptr66.27, align 4
%add.ptr66.1.27 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 2075
%i122 = load float, ptr addrspace(4) %add.ptr66.1.27, align 4
%add.ptr66.2.27 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 3099
%i123 = load float, ptr addrspace(4) %add.ptr66.2.27, align 4
%add.ptr70.27 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 28
%i124 = load float, ptr addrspace(4) %add.ptr70.27, align 4
%add.ptr66.28 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 1052
%i125 = load float, ptr addrspace(4) %add.ptr66.28, align 4
%add.ptr66.1.28 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 2076
%i126 = load float, ptr addrspace(4) %add.ptr66.1.28, align 4
%add.ptr66.2.28 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 3100
%i127 = load float, ptr addrspace(4) %add.ptr66.2.28, align 4
%add.ptr70.28 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 29
%i128 = load float, ptr addrspace(4) %add.ptr70.28, align 4
%add.ptr66.29 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 1053
%i129 = load float, ptr addrspace(4) %add.ptr66.29, align 4
%add.ptr66.1.29 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 2077
%i130 = load float, ptr addrspace(4) %add.ptr66.1.29, align 4
%add.ptr66.2.29 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 3101
%i131 = load float, ptr addrspace(4) %add.ptr66.2.29, align 4
%add.ptr70.29 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 30
%i132 = load float, ptr addrspace(4) %add.ptr70.29, align 4
%add.ptr66.30 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 1054
%i133 = load float, ptr addrspace(4) %add.ptr66.30, align 4
%add.ptr66.1.30 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 2078
%i134 = load float, ptr addrspace(4) %add.ptr66.1.30, align 4
%add.ptr66.2.30 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 3102
%i135 = load float, ptr addrspace(4) %add.ptr66.2.30, align 4
%add.ptr70.30 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 31
%i136 = load float, ptr addrspace(4) %add.ptr70.30, align 4
%add.ptr66.31 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 1055
%i137 = load float, ptr addrspace(4) %add.ptr66.31, align 4
%add.ptr66.1.31 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 2079
%i138 = load float, ptr addrspace(4) %add.ptr66.1.31, align 4
%add.ptr66.2.31 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 3103
%i139 = load float, ptr addrspace(4) %add.ptr66.2.31, align 4
%add.ptr47.3 = getelementptr inbounds float, ptr addrspace(1) %i_ptr.0288, i64 196
%i140 = tail call float @llvm.fmuladd.f32(float %i8, float %i12, float %accum.sroa.0.0)
%i141 = tail call float @llvm.fmuladd.f32(float %i9, float %i13, float %i140)
%i142 = tail call float @llvm.fmuladd.f32(float %i10, float %i14, float %i141)
%i143 = tail call float @llvm.fmuladd.f32(float %i11, float %i15, float %i142)
%i144 = tail call float @llvm.fmuladd.f32(float %i8, float %i16, float %accum.sroa.6.0)
%i145 = tail call float @llvm.fmuladd.f32(float %i9, float %i17, float %i144)
%i146 = tail call float @llvm.fmuladd.f32(float %i10, float %i18, float %i145)
%i147 = tail call float @llvm.fmuladd.f32(float %i11, float %i19, float %i146)
%i148 = tail call float @llvm.fmuladd.f32(float %i8, float %i20, float %accum.sroa.10.0)
%i149 = tail call float @llvm.fmuladd.f32(float %i9, float %i21, float %i148)
%i150 = tail call float @llvm.fmuladd.f32(float %i10, float %i22, float %i149)
%i151 = tail call float @llvm.fmuladd.f32(float %i11, float %i23, float %i150)
%i152 = tail call float @llvm.fmuladd.f32(float %i8, float %i24, float %accum.sroa.14.0)
%i153 = tail call float @llvm.fmuladd.f32(float %i9, float %i25, float %i152)
%i154 = tail call float @llvm.fmuladd.f32(float %i10, float %i26, float %i153)
%i155 = tail call float @llvm.fmuladd.f32(float %i11, float %i27, float %i154)
%i156 = tail call float @llvm.fmuladd.f32(float %i8, float %i28, float %accum.sroa.18.0)
%i157 = tail call float @llvm.fmuladd.f32(float %i9, float %i29, float %i156)
%i158 = tail call float @llvm.fmuladd.f32(float %i10, float %i30, float %i157)
%i159 = tail call float @llvm.fmuladd.f32(float %i11, float %i31, float %i158)
%i160 = tail call float @llvm.fmuladd.f32(float %i8, float %i32, float %accum.sroa.22.0)
%i161 = tail call float @llvm.fmuladd.f32(float %i9, float %i33, float %i160)
%i162 = tail call float @llvm.fmuladd.f32(float %i10, float %i34, float %i161)
%i163 = tail call float @llvm.fmuladd.f32(float %i11, float %i35, float %i162)
%i164 = tail call float @llvm.fmuladd.f32(float %i8, float %i36, float %accum.sroa.26.0)
%i165 = tail call float @llvm.fmuladd.f32(float %i9, float %i37, float %i164)
%i166 = tail call float @llvm.fmuladd.f32(float %i10, float %i38, float %i165)
%i167 = tail call float @llvm.fmuladd.f32(float %i11, float %i39, float %i166)
%i168 = tail call float @llvm.fmuladd.f32(float %i8, float %i40, float %accum.sroa.30.0)
%i169 = tail call float @llvm.fmuladd.f32(float %i9, float %i41, float %i168)
%i170 = tail call float @llvm.fmuladd.f32(float %i10, float %i42, float %i169)
%i171 = tail call float @llvm.fmuladd.f32(float %i11, float %i43, float %i170)
%i172 = tail call float @llvm.fmuladd.f32(float %i8, float %i44, float %accum.sroa.34.0)
%i173 = tail call float @llvm.fmuladd.f32(float %i9, float %i45, float %i172)
%i174 = tail call float @llvm.fmuladd.f32(float %i10, float %i46, float %i173)
%i175 = tail call float @llvm.fmuladd.f32(float %i11, float %i47, float %i174)
%i176 = tail call float @llvm.fmuladd.f32(float %i8, float %i48, float %accum.sroa.38.0)
%i177 = tail call float @llvm.fmuladd.f32(float %i9, float %i49, float %i176)
%i178 = tail call float @llvm.fmuladd.f32(float %i10, float %i50, float %i177)
%i179 = tail call float @llvm.fmuladd.f32(float %i11, float %i51, float %i178)
%i180 = tail call float @llvm.fmuladd.f32(float %i8, float %i52, float %accum.sroa.42.0)
%i181 = tail call float @llvm.fmuladd.f32(float %i9, float %i53, float %i180)
%i182 = tail call float @llvm.fmuladd.f32(float %i10, float %i54, float %i181)
%i183 = tail call float @llvm.fmuladd.f32(float %i11, float %i55, float %i182)
%i184 = tail call float @llvm.fmuladd.f32(float %i8, float %i56, float %accum.sroa.46.0)
%i185 = tail call float @llvm.fmuladd.f32(float %i9, float %i57, float %i184)
%i186 = tail call float @llvm.fmuladd.f32(float %i10, float %i58, float %i185)
%i187 = tail call float @llvm.fmuladd.f32(float %i11, float %i59, float %i186)
%i188 = tail call float @llvm.fmuladd.f32(float %i8, float %i60, float %accum.sroa.50.0)
%i189 = tail call float @llvm.fmuladd.f32(float %i9, float %i61, float %i188)
%i190 = tail call float @llvm.fmuladd.f32(float %i10, float %i62, float %i189)
%i191 = tail call float @llvm.fmuladd.f32(float %i11, float %i63, float %i190)
%i192 = tail call float @llvm.fmuladd.f32(float %i8, float %i64, float %accum.sroa.54.0)
%i193 = tail call float @llvm.fmuladd.f32(float %i9, float %i65, float %i192)
%i194 = tail call float @llvm.fmuladd.f32(float %i10, float %i66, float %i193)
%i195 = tail call float @llvm.fmuladd.f32(float %i11, float %i67, float %i194)
%i196 = tail call float @llvm.fmuladd.f32(float %i8, float %i68, float %accum.sroa.58.0)
%i197 = tail call float @llvm.fmuladd.f32(float %i9, float %i69, float %i196)
%i198 = tail call float @llvm.fmuladd.f32(float %i10, float %i70, float %i197)
%i199 = tail call float @llvm.fmuladd.f32(float %i11, float %i71, float %i198)
%i200 = tail call float @llvm.fmuladd.f32(float %i8, float %i72, float %accum.sroa.62.0)
%i201 = tail call float @llvm.fmuladd.f32(float %i9, float %i73, float %i200)
%i202 = tail call float @llvm.fmuladd.f32(float %i10, float %i74, float %i201)
%i203 = tail call float @llvm.fmuladd.f32(float %i11, float %i75, float %i202)
%i204 = tail call float @llvm.fmuladd.f32(float %i8, float %i76, float %accum.sroa.66.0)
%i205 = tail call float @llvm.fmuladd.f32(float %i9, float %i77, float %i204)
%i206 = tail call float @llvm.fmuladd.f32(float %i10, float %i78, float %i205)
%i207 = tail call float @llvm.fmuladd.f32(float %i11, float %i79, float %i206)
%i208 = tail call float @llvm.fmuladd.f32(float %i8, float %i80, float %accum.sroa.70.0)
%i209 = tail call float @llvm.fmuladd.f32(float %i9, float %i81, float %i208)
%i210 = tail call float @llvm.fmuladd.f32(float %i10, float %i82, float %i209)
%i211 = tail call float @llvm.fmuladd.f32(float %i11, float %i83, float %i210)
%i212 = tail call float @llvm.fmuladd.f32(float %i8, float %i84, float %accum.sroa.74.0)
%i213 = tail call float @llvm.fmuladd.f32(float %i9, float %i85, float %i212)
%i214 = tail call float @llvm.fmuladd.f32(float %i10, float %i86, float %i213)
%i215 = tail call float @llvm.fmuladd.f32(float %i11, float %i87, float %i214)
%i216 = tail call float @llvm.fmuladd.f32(float %i8, float %i88, float %accum.sroa.78.0)
%i217 = tail call float @llvm.fmuladd.f32(float %i9, float %i89, float %i216)
%i218 = tail call float @llvm.fmuladd.f32(float %i10, float %i90, float %i217)
%i219 = tail call float @llvm.fmuladd.f32(float %i11, float %i91, float %i218)
%i220 = tail call float @llvm.fmuladd.f32(float %i8, float %i92, float %accum.sroa.82.0)
%i221 = tail call float @llvm.fmuladd.f32(float %i9, float %i93, float %i220)
%i222 = tail call float @llvm.fmuladd.f32(float %i10, float %i94, float %i221)
%i223 = tail call float @llvm.fmuladd.f32(float %i11, float %i95, float %i222)
%i224 = tail call float @llvm.fmuladd.f32(float %i8, float %i96, float %accum.sroa.86.0)
%i225 = tail call float @llvm.fmuladd.f32(float %i9, float %i97, float %i224)
%i226 = tail call float @llvm.fmuladd.f32(float %i10, float %i98, float %i225)
%i227 = tail call float @llvm.fmuladd.f32(float %i11, float %i99, float %i226)
%i228 = tail call float @llvm.fmuladd.f32(float %i8, float %i100, float %accum.sroa.90.0)
%i229 = tail call float @llvm.fmuladd.f32(float %i9, float %i101, float %i228)
%i230 = tail call float @llvm.fmuladd.f32(float %i10, float %i102, float %i229)
%i231 = tail call float @llvm.fmuladd.f32(float %i11, float %i103, float %i230)
%i232 = tail call float @llvm.fmuladd.f32(float %i8, float %i104, float %accum.sroa.94.0)
%i233 = tail call float @llvm.fmuladd.f32(float %i9, float %i105, float %i232)
%i234 = tail call float @llvm.fmuladd.f32(float %i10, float %i106, float %i233)
%i235 = tail call float @llvm.fmuladd.f32(float %i11, float %i107, float %i234)
%i236 = tail call float @llvm.fmuladd.f32(float %i8, float %i108, float %accum.sroa.98.0)
%i237 = tail call float @llvm.fmuladd.f32(float %i9, float %i109, float %i236)
%i238 = tail call float @llvm.fmuladd.f32(float %i10, float %i110, float %i237)
%i239 = tail call float @llvm.fmuladd.f32(float %i11, float %i111, float %i238)
%i240 = tail call float @llvm.fmuladd.f32(float %i8, float %i112, float %accum.sroa.102.0)
%i241 = tail call float @llvm.fmuladd.f32(float %i9, float %i113, float %i240)
%i242 = tail call float @llvm.fmuladd.f32(float %i10, float %i114, float %i241)
%i243 = tail call float @llvm.fmuladd.f32(float %i11, float %i115, float %i242)
%i244 = tail call float @llvm.fmuladd.f32(float %i8, float %i116, float %accum.sroa.106.0)
%i245 = tail call float @llvm.fmuladd.f32(float %i9, float %i117, float %i244)
%i246 = tail call float @llvm.fmuladd.f32(float %i10, float %i118, float %i245)
%i247 = tail call float @llvm.fmuladd.f32(float %i11, float %i119, float %i246)
%i248 = tail call float @llvm.fmuladd.f32(float %i8, float %i120, float %accum.sroa.110.0)
%i249 = tail call float @llvm.fmuladd.f32(float %i9, float %i121, float %i248)
%i250 = tail call float @llvm.fmuladd.f32(float %i10, float %i122, float %i249)
%i251 = tail call float @llvm.fmuladd.f32(float %i11, float %i123, float %i250)
%i252 = tail call float @llvm.fmuladd.f32(float %i8, float %i124, float %accum.sroa.114.0)
%i253 = tail call float @llvm.fmuladd.f32(float %i9, float %i125, float %i252)
%i254 = tail call float @llvm.fmuladd.f32(float %i10, float %i126, float %i253)
%i255 = tail call float @llvm.fmuladd.f32(float %i11, float %i127, float %i254)
%i256 = tail call float @llvm.fmuladd.f32(float %i8, float %i128, float %accum.sroa.118.0)
%i257 = tail call float @llvm.fmuladd.f32(float %i9, float %i129, float %i256)
%i258 = tail call float @llvm.fmuladd.f32(float %i10, float %i130, float %i257)
%i259 = tail call float @llvm.fmuladd.f32(float %i11, float %i131, float %i258)
%i260 = tail call float @llvm.fmuladd.f32(float %i8, float %i132, float %accum.sroa.122.0)
%i261 = tail call float @llvm.fmuladd.f32(float %i9, float %i133, float %i260)
%i262 = tail call float @llvm.fmuladd.f32(float %i10, float %i134, float %i261)
%i263 = tail call float @llvm.fmuladd.f32(float %i11, float %i135, float %i262)
%i264 = tail call float @llvm.fmuladd.f32(float %i8, float %i136, float %accum.sroa.126.0)
%i265 = tail call float @llvm.fmuladd.f32(float %i9, float %i137, float %i264)
%i266 = tail call float @llvm.fmuladd.f32(float %i10, float %i138, float %i265)
%i267 = tail call float @llvm.fmuladd.f32(float %i11, float %i139, float %i266)
%add.ptr74 = getelementptr inbounds float, ptr addrspace(4) %w_ptr.0287, i64 4096
%inc116 = add nuw nsw i32 %ci.0286, 1
%exitcond.not = icmp eq i32 %inc116, 512
br i1 %exitcond.not, label %for.cond.cleanup26, label %for.cond28.preheader
for.cond.cleanup26: ; preds = %for.cond28.preheader
%mul119 = shl nuw nsw i32 0, 1
%mul120 = mul i32 %div, 200704
%mul121 = mul i32 0, 6272
%add122 = add i32 %mul120, %mul121
%mul123 = mul nuw nsw i32 0, 28
%add124 = add i32 %add122, %mul123
%add126 = add i32 %add124, %mul119
%idx.ext127 = zext i32 %add126 to i64
%add.ptr128 = getelementptr inbounds float, ptr addrspace(1) %out_ptr, i64 %idx.ext127
store float %i143, ptr addrspace(1) %add.ptr128, align 4
%add.ptr184 = getelementptr inbounds float, ptr addrspace(1) %add.ptr128, i64 196
store float %i147, ptr addrspace(1) %add.ptr184, align 4
%add.ptr167.1 = getelementptr inbounds float, ptr addrspace(1) %add.ptr184, i64 14
store float 0.000000e+00, ptr addrspace(1) %add.ptr167.1, align 4
%add.ptr175.1.1 = getelementptr inbounds float, ptr addrspace(1) %add.ptr167.1, i64 1
store float 0.000000e+00, ptr addrspace(1) %add.ptr175.1.1, align 4
%add.ptr184.1 = getelementptr inbounds float, ptr addrspace(1) %add.ptr184, i64 196
store float %i151, ptr addrspace(1) %add.ptr184.1, align 4
%add.ptr184.2 = getelementptr inbounds float, ptr addrspace(1) %add.ptr184.1, i64 196
store float %i155, ptr addrspace(1) %add.ptr184.2, align 4
%add.ptr184.3 = getelementptr inbounds float, ptr addrspace(1) %add.ptr184.2, i64 196
store float %i159, ptr addrspace(1) %add.ptr184.3, align 4
%add.ptr184.4 = getelementptr inbounds float, ptr addrspace(1) %add.ptr184.3, i64 196
store float %i163, ptr addrspace(1) %add.ptr184.4, align 4
%add.ptr154.5 = getelementptr inbounds float, ptr addrspace(1) %add.ptr184.4, i64 1
store float 0.000000e+00, ptr addrspace(1) %add.ptr154.5, align 4
%add.ptr184.5 = getelementptr inbounds float, ptr addrspace(1) %add.ptr184.4, i64 196
store float %i167, ptr addrspace(1) %add.ptr184.5, align 4
%add.ptr154.6 = getelementptr inbounds float, ptr addrspace(1) %add.ptr184.5, i64 1
store float 0.000000e+00, ptr addrspace(1) %add.ptr154.6, align 4
%add.ptr184.6 = getelementptr inbounds float, ptr addrspace(1) %add.ptr184.5, i64 196
store float %i171, ptr addrspace(1) %add.ptr184.6, align 4
%add.ptr184.7 = getelementptr inbounds float, ptr addrspace(1) %add.ptr184.6, i64 196
store float %i175, ptr addrspace(1) %add.ptr184.7, align 4
%add.ptr167.8 = getelementptr inbounds float, ptr addrspace(1) %add.ptr184.7, i64 14
store float 0.000000e+00, ptr addrspace(1) %add.ptr167.8, align 4
%add.ptr175.1.8 = getelementptr inbounds float, ptr addrspace(1) %add.ptr167.8, i64 1
store float 0.000000e+00, ptr addrspace(1) %add.ptr175.1.8, align 4
%add.ptr184.8 = getelementptr inbounds float, ptr addrspace(1) %add.ptr184.7, i64 196
store float %i179, ptr addrspace(1) %add.ptr184.8, align 4
%add.ptr184.9 = getelementptr inbounds float, ptr addrspace(1) %add.ptr184.8, i64 196
store float %i183, ptr addrspace(1) %add.ptr184.9, align 4
%add.ptr184.10 = getelementptr inbounds float, ptr addrspace(1) %add.ptr184.9, i64 196
store float %i187, ptr addrspace(1) %add.ptr184.10, align 4
%add.ptr184.11 = getelementptr inbounds float, ptr addrspace(1) %add.ptr184.10, i64 196
store float %i191, ptr addrspace(1) %add.ptr184.11, align 4
%add.ptr184.12 = getelementptr inbounds float, ptr addrspace(1) %add.ptr184.11, i64 196
store float %i195, ptr addrspace(1) %add.ptr184.12, align 4
%add.ptr184.13 = getelementptr inbounds float, ptr addrspace(1) %add.ptr184.12, i64 196
store float %i199, ptr addrspace(1) %add.ptr184.13, align 4
%add.ptr184.14 = getelementptr inbounds float, ptr addrspace(1) %add.ptr184.13, i64 196
store float %i203, ptr addrspace(1) %add.ptr184.14, align 4
%add.ptr184.15 = getelementptr inbounds float, ptr addrspace(1) %add.ptr184.14, i64 196
store float %i207, ptr addrspace(1) %add.ptr184.15, align 4
%add.ptr184.16 = getelementptr inbounds float, ptr addrspace(1) %add.ptr184.15, i64 196
store float %i211, ptr addrspace(1) %add.ptr184.16, align 4
%add.ptr184.17 = getelementptr inbounds float, ptr addrspace(1) %add.ptr184.16, i64 196
store float %i215, ptr addrspace(1) %add.ptr184.17, align 4
%add.ptr184.18 = getelementptr inbounds float, ptr addrspace(1) %add.ptr184.17, i64 196
store float %i219, ptr addrspace(1) %add.ptr184.18, align 4
%add.ptr184.19 = getelementptr inbounds float, ptr addrspace(1) %add.ptr184.18, i64 196
store float %i223, ptr addrspace(1) %add.ptr184.19, align 4
%add.ptr184.20 = getelementptr inbounds float, ptr addrspace(1) %add.ptr184.19, i64 196
store float %i227, ptr addrspace(1) %add.ptr184.20, align 4
%add.ptr184.21 = getelementptr inbounds float, ptr addrspace(1) %add.ptr184.20, i64 196
store float %i231, ptr addrspace(1) %add.ptr184.21, align 4
%add.ptr184.22 = getelementptr inbounds float, ptr addrspace(1) %add.ptr184.21, i64 196
store float %i235, ptr addrspace(1) %add.ptr184.22, align 4
%add.ptr184.23 = getelementptr inbounds float, ptr addrspace(1) %add.ptr184.22, i64 196
store float %i239, ptr addrspace(1) %add.ptr184.23, align 4
%add.ptr184.24 = getelementptr inbounds float, ptr addrspace(1) %add.ptr184.23, i64 196
store float %i243, ptr addrspace(1) %add.ptr184.24, align 4
%add.ptr184.25 = getelementptr inbounds float, ptr addrspace(1) %add.ptr184.24, i64 196
store float %i247, ptr addrspace(1) %add.ptr184.25, align 4
%add.ptr184.26 = getelementptr inbounds float, ptr addrspace(1) %add.ptr184.25, i64 196
store float %i251, ptr addrspace(1) %add.ptr184.26, align 4
%add.ptr184.27 = getelementptr inbounds float, ptr addrspace(1) %add.ptr184.26, i64 196
store float %i255, ptr addrspace(1) %add.ptr184.27, align 4
%add.ptr184.28 = getelementptr inbounds float, ptr addrspace(1) %add.ptr184.27, i64 196
store float %i259, ptr addrspace(1) %add.ptr184.28, align 4
%add.ptr184.29 = getelementptr inbounds float, ptr addrspace(1) %add.ptr184.28, i64 196
store float %i263, ptr addrspace(1) %add.ptr184.29, align 4
%add.ptr184.30 = getelementptr inbounds float, ptr addrspace(1) %add.ptr184.29, i64 196
store float %i267, ptr addrspace(1) %add.ptr184.30, align 4
ret void
}
declare float @llvm.fmuladd.f32(float, float, float) #2
declare i32 @llvm.amdgcn.workitem.id.x() #3
declare i32 @llvm.amdgcn.workgroup.id.x() #3
declare align 4 ptr addrspace(4) @llvm.amdgcn.implicitarg.ptr() #3
!0 = !{i32 1, i32 2, i32 1, i32 0}
!1 = !{!"none", !"none", !"none", !"none"}
!2 = !{!"ptr", !"ptr", !"ptr", !"float"}
!3 = !{!"restrict const", !"restrict const", !"restrict", !""}
!4 = !{i32 256, i32 1, i32 1}
!5 = !{i32 0, i32 1024}
attributes #0 = { nounwind "amdgpu-waves-per-eu"="1,1" "amdgpu-flat-work-group-size"="1,1" }
attributes #1 = { nounwind "amdgpu-num-vgpr"="64" }
attributes #2 = { nofree nosync nounwind readnone speculatable willreturn }
attributes #3 = { nounwind readnone speculatable willreturn }