; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx700 < %s | FileCheck -check-prefix=GFX7 %s
; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx803 < %s | FileCheck -check-prefix=GFX8 %s
; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx900 < %s | FileCheck -check-prefixes=GFX9,GFX900 %s
; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx950 < %s | FileCheck -check-prefixes=GFX9,GFX950 %s
; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1030 < %s | FileCheck -check-prefix=GFX10 %s
; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1100 -mattr=+real-true16 < %s | FileCheck -check-prefixes=GFX11,GFX11-TRUE16 %s
; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1100 -mattr=-real-true16 < %s | FileCheck -check-prefixes=GFX11,GFX11-FAKE16 %s
; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1200 -mattr=+real-true16 < %s | FileCheck -check-prefixes=GFX12,GFX12-TRUE16 %s
; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1200 -mattr=-real-true16 < %s | FileCheck -check-prefixes=GFX12,GFX12-FAKE16 %s

define bfloat @v_maximumnum_bf16(bfloat %x, bfloat %y) {
; GFX7-LABEL: v_maximumnum_bf16:
; GFX7:       ; %bb.0:
; GFX7-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT:    v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT:    v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT:    v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT:    v_max_f32_e32 v0, v0, v1
; GFX7-NEXT:    v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT:    s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_maximumnum_bf16:
; GFX8:       ; %bb.0:
; GFX8-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT:    v_lshlrev_b32_e32 v2, 16, v0
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v2, v2
; GFX8-NEXT:    v_lshlrev_b32_e32 v2, 16, v1
; GFX8-NEXT:    v_cndmask_b32_e32 v0, v0, v1, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v2, v2
; GFX8-NEXT:    v_cndmask_b32_e32 v1, v1, v0, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v2, 16, v1
; GFX8-NEXT:    v_lshlrev_b32_e32 v3, 16, v0
; GFX8-NEXT:    v_cmp_gt_f32_e32 vcc, v3, v2
; GFX8-NEXT:    v_cndmask_b32_e32 v2, v1, v0, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; GFX8-NEXT:    v_mul_f32_e32 v2, 1.0, v2
; GFX8-NEXT:    v_bfe_u32 v3, v2, 16, 1
; GFX8-NEXT:    v_add_u32_e32 v3, vcc, v3, v2
; GFX8-NEXT:    s_movk_i32 s4, 0x7fff
; GFX8-NEXT:    v_add_u32_e32 v3, vcc, s4, v3
; GFX8-NEXT:    v_or_b32_e32 v4, 0x400000, v2
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v2, v2
; GFX8-NEXT:    v_cndmask_b32_e32 v2, v3, v4, vcc
; GFX8-NEXT:    v_lshrrev_b32_e32 v3, 16, v2
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v0
; GFX8-NEXT:    v_cndmask_b32_e32 v0, v3, v0, vcc
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v1
; GFX8-NEXT:    v_cndmask_b32_e32 v0, v0, v1, vcc
; GFX8-NEXT:    v_and_b32_e32 v1, 0xffff0000, v2
; GFX8-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v1
; GFX8-NEXT:    v_cndmask_b32_e32 v0, v3, v0, vcc
; GFX8-NEXT:    s_setpc_b64 s[30:31]
;
; GFX900-LABEL: v_maximumnum_bf16:
; GFX900:       ; %bb.0:
; GFX900-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX900-NEXT:    v_lshlrev_b32_e32 v2, 16, v0
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v2, v2
; GFX900-NEXT:    v_lshlrev_b32_e32 v3, 16, v1
; GFX900-NEXT:    v_cndmask_b32_e32 v0, v0, v1, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; GFX900-NEXT:    v_cndmask_b32_e32 v1, v1, v0, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v2, 16, v0
; GFX900-NEXT:    v_lshlrev_b32_e32 v3, 16, v1
; GFX900-NEXT:    v_cmp_gt_f32_e32 vcc, v2, v3
; GFX900-NEXT:    v_cndmask_b32_e32 v2, v1, v0, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; GFX900-NEXT:    v_max_f32_e32 v2, v2, v2
; GFX900-NEXT:    v_bfe_u32 v3, v2, 16, 1
; GFX900-NEXT:    s_movk_i32 s4, 0x7fff
; GFX900-NEXT:    v_add3_u32 v3, v3, v2, s4
; GFX900-NEXT:    v_or_b32_e32 v4, 0x400000, v2
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v2, v2
; GFX900-NEXT:    v_cndmask_b32_e32 v2, v3, v4, vcc
; GFX900-NEXT:    v_lshrrev_b32_e32 v3, 16, v2
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v0
; GFX900-NEXT:    v_cndmask_b32_e32 v0, v3, v0, vcc
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v1
; GFX900-NEXT:    v_cndmask_b32_e32 v0, v0, v1, vcc
; GFX900-NEXT:    v_and_b32_e32 v1, 0xffff0000, v2
; GFX900-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v1
; GFX900-NEXT:    v_cndmask_b32_e32 v0, v3, v0, vcc
; GFX900-NEXT:    s_setpc_b64 s[30:31]
;
; GFX950-LABEL: v_maximumnum_bf16:
; GFX950:       ; %bb.0:
; GFX950-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX950-NEXT:    v_lshlrev_b32_e32 v2, 16, v0
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v2, v2
; GFX950-NEXT:    v_lshlrev_b32_e32 v3, 16, v1
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v0, v0, v1, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; GFX950-NEXT:    v_lshlrev_b32_e32 v2, 16, v0
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v1, v1, v0, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v3, 16, v1
; GFX950-NEXT:    v_cmp_gt_f32_e32 vcc, v2, v3
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v2, v1, v0, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; GFX950-NEXT:    v_max_f32_e32 v2, v2, v2
; GFX950-NEXT:    v_cvt_pk_bf16_f32 v2, v2, s0
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v0
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v0, v2, v0, vcc
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v1
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v0, v0, v1, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v1, 16, v2
; GFX950-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v1
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v0, v2, v0, vcc
; GFX950-NEXT:    s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_maximumnum_bf16:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT:    v_lshlrev_b32_e32 v2, 16, v0
; GFX10-NEXT:    v_lshlrev_b32_e32 v3, 16, v1
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX10-NEXT:    v_cndmask_b32_e32 v0, v0, v1, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX10-NEXT:    v_lshlrev_b32_e32 v2, 16, v0
; GFX10-NEXT:    v_cndmask_b32_e32 v1, v1, v0, vcc_lo
; GFX10-NEXT:    v_lshlrev_b32_e32 v3, 16, v1
; GFX10-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v2, v3
; GFX10-NEXT:    v_cndmask_b32_e32 v2, v1, v0, vcc_lo
; GFX10-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; GFX10-NEXT:    v_max_f32_e32 v2, v2, v2
; GFX10-NEXT:    v_bfe_u32 v3, v2, 16, 1
; GFX10-NEXT:    v_or_b32_e32 v4, 0x400000, v2
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX10-NEXT:    v_add3_u32 v3, v3, v2, 0x7fff
; GFX10-NEXT:    v_cndmask_b32_e32 v2, v3, v4, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v0
; GFX10-NEXT:    v_lshrrev_b32_e32 v3, 16, v2
; GFX10-NEXT:    v_and_b32_e32 v2, 0xffff0000, v2
; GFX10-NEXT:    v_cndmask_b32_e32 v0, v3, v0, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v1
; GFX10-NEXT:    v_cndmask_b32_e32 v0, v0, v1, vcc_lo
; GFX10-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v2
; GFX10-NEXT:    v_cndmask_b32_e32 v0, v3, v0, vcc_lo
; GFX10-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-TRUE16-LABEL: v_maximumnum_bf16:
; GFX11-TRUE16:       ; %bb.0:
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v0
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.h, v1.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s0, v1, v1
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v0.l, v0.l, v0.h, vcc_lo
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v1.l, v0.h, v0.l, s0
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, v0.l
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v3.l, v1.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; GFX11-TRUE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v2, v3
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v2.l, v1.l, v0.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v0.l
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; GFX11-TRUE16-NEXT:    v_max_f32_e32 v2, v2, v2
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v3, v2, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, 0x400000, v2
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s0, v2, v2
; GFX11-TRUE16-NEXT:    v_add3_u32 v3, v3, v2, 0x7fff
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e64 v2, v3, v4, s0
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s0, 0, v1.l
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v0.l, v2.h, v0.l, vcc_lo
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff0000, v2
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v0.l, v0.l, v1.l, s0
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v3
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v0.l, v2.h, v0.l, vcc_lo
; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-FAKE16-LABEL: v_maximumnum_bf16:
; GFX11-FAKE16:       ; %bb.0:
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v0
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v0, v1, vcc_lo
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v1
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v0
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v1, v0, vcc_lo
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v1
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v2, v3
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v1, v0, vcc_lo
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_max_f32_e32 v2, v2, v2
; GFX11-FAKE16-NEXT:    v_bfe_u32 v3, v2, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, 0x400000, v2
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_add3_u32 v3, v3, v2, 0x7fff
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v3, v4, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v0
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v3, 16, v2
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xffff0000, v2
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v3, v0, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v1
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v0, v1, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v2
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v3, v0, vcc_lo
; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
;
; GFX12-TRUE16-LABEL: v_maximumnum_bf16:
; GFX12-TRUE16:       ; %bb.0:
; GFX12-TRUE16-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX12-TRUE16-NEXT:    s_wait_expcnt 0x0
; GFX12-TRUE16-NEXT:    s_wait_samplecnt 0x0
; GFX12-TRUE16-NEXT:    s_wait_bvhcnt 0x0
; GFX12-TRUE16-NEXT:    s_wait_kmcnt 0x0
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v0
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v0.h, v1.l
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s0, v1, v1
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v0.l, v0.l, v0.h, vcc_lo
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v1.l, v0.h, v0.l, s0
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v2.l, v0.l
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v3.l, v1.l
; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_2)
; GFX12-TRUE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v2, v3
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v2.l, v1.l, v0.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v0.l
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT:    v_max_num_f32_e32 v2, v2, v2
; GFX12-TRUE16-NEXT:    v_bfe_u32 v3, v2, 16, 1
; GFX12-TRUE16-NEXT:    v_or_b32_e32 v4, 0x400000, v2
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s0, v2, v2
; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT:    v_add3_u32 v3, v3, v2, 0x7fff
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b32_e64 v2, v3, v4, s0
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s0, 0, v1.l
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v0.l, v2.h, v0.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff0000, v2
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v0.l, v0.l, v1.l, s0
; GFX12-TRUE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v3
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v0.l, v2.h, v0.l, vcc_lo
; GFX12-TRUE16-NEXT:    s_setpc_b64 s[30:31]
;
; GFX12-FAKE16-LABEL: v_maximumnum_bf16:
; GFX12-FAKE16:       ; %bb.0:
; GFX12-FAKE16-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX12-FAKE16-NEXT:    s_wait_expcnt 0x0
; GFX12-FAKE16-NEXT:    s_wait_samplecnt 0x0
; GFX12-FAKE16-NEXT:    s_wait_bvhcnt 0x0
; GFX12-FAKE16-NEXT:    s_wait_kmcnt 0x0
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v0
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_2)
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v0, v1, vcc_lo
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v1
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v0
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1)
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v1, v0, vcc_lo
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v1
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1)
; GFX12-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v2, v3
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v1, v0, vcc_lo
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-FAKE16-NEXT:    v_max_num_f32_e32 v2, v2, v2
; GFX12-FAKE16-NEXT:    v_bfe_u32 v3, v2, 16, 1
; GFX12-FAKE16-NEXT:    v_or_b32_e32 v4, 0x400000, v2
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX12-FAKE16-NEXT:    v_add3_u32 v3, v3, v2, 0x7fff
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v3, v4, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v0
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_2)
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v3, 16, v2
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v2, 0xffff0000, v2
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v3, v0, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v1
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_2)
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v0, v1, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v2
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v3, v0, vcc_lo
; GFX12-FAKE16-NEXT:    s_setpc_b64 s[30:31]
  %result = call bfloat @llvm.maximumnum.bf16(bfloat %x, bfloat %y)
  ret bfloat %result
}

define bfloat @v_maximumnum_bf16_nnan(bfloat %x, bfloat %y) {
; GFX7-LABEL: v_maximumnum_bf16_nnan:
; GFX7:       ; %bb.0:
; GFX7-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT:    v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT:    v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT:    v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT:    v_max_f32_e32 v0, v0, v1
; GFX7-NEXT:    v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT:    s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_maximumnum_bf16_nnan:
; GFX8:       ; %bb.0:
; GFX8-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT:    v_lshlrev_b32_e32 v2, 16, v1
; GFX8-NEXT:    v_lshlrev_b32_e32 v3, 16, v0
; GFX8-NEXT:    v_cmp_gt_f32_e32 vcc, v3, v2
; GFX8-NEXT:    v_cndmask_b32_e32 v2, v1, v0, vcc
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v0
; GFX8-NEXT:    v_cndmask_b32_e32 v0, v2, v0, vcc
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v1
; GFX8-NEXT:    v_cndmask_b32_e32 v0, v0, v1, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v1, 16, v2
; GFX8-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v1
; GFX8-NEXT:    v_cndmask_b32_e32 v0, v2, v0, vcc
; GFX8-NEXT:    s_setpc_b64 s[30:31]
;
; GFX900-LABEL: v_maximumnum_bf16_nnan:
; GFX900:       ; %bb.0:
; GFX900-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX900-NEXT:    v_lshlrev_b32_e32 v2, 16, v1
; GFX900-NEXT:    v_lshlrev_b32_e32 v3, 16, v0
; GFX900-NEXT:    v_cmp_gt_f32_e32 vcc, v3, v2
; GFX900-NEXT:    v_cndmask_b32_e32 v2, v1, v0, vcc
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v0
; GFX900-NEXT:    v_cndmask_b32_e32 v0, v2, v0, vcc
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v1
; GFX900-NEXT:    v_cndmask_b32_e32 v0, v0, v1, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v1, 16, v2
; GFX900-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v1
; GFX900-NEXT:    v_cndmask_b32_e32 v0, v2, v0, vcc
; GFX900-NEXT:    s_setpc_b64 s[30:31]
;
; GFX950-LABEL: v_maximumnum_bf16_nnan:
; GFX950:       ; %bb.0:
; GFX950-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX950-NEXT:    v_lshlrev_b32_e32 v2, 16, v1
; GFX950-NEXT:    v_lshlrev_b32_e32 v3, 16, v0
; GFX950-NEXT:    v_cmp_gt_f32_e32 vcc, v3, v2
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v2, v1, v0, vcc
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v0
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v0, v2, v0, vcc
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v1
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v0, v0, v1, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v1, 16, v2
; GFX950-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v1
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v0, v2, v0, vcc
; GFX950-NEXT:    s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_maximumnum_bf16_nnan:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT:    v_lshlrev_b32_e32 v2, 16, v1
; GFX10-NEXT:    v_lshlrev_b32_e32 v3, 16, v0
; GFX10-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v3, v2
; GFX10-NEXT:    v_cndmask_b32_e32 v2, v1, v0, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v0
; GFX10-NEXT:    v_lshlrev_b32_e32 v3, 16, v2
; GFX10-NEXT:    v_cndmask_b32_e32 v0, v2, v0, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v1
; GFX10-NEXT:    v_cndmask_b32_e32 v0, v0, v1, vcc_lo
; GFX10-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v3
; GFX10-NEXT:    v_cndmask_b32_e32 v0, v2, v0, vcc_lo
; GFX10-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-TRUE16-LABEL: v_maximumnum_bf16_nnan:
; GFX11-TRUE16:       ; %bb.0:
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v1
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v0
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v3, v2
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v2.l, v1.l, v0.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v0.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v3.l, v2.l
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v0.l, v2.l, v0.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v1.l
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v0.l, v0.l, v1.l, vcc_lo
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v3
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v0.l, v2.l, v0.l, vcc_lo
; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-FAKE16-LABEL: v_maximumnum_bf16_nnan:
; GFX11-FAKE16:       ; %bb.0:
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v1
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v0
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v3, v2
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v1, v0, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v0
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v0, v2, v0 :: v_dual_lshlrev_b32 v3, 16, v2
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v1
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v0, v1, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v3
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v2, v0, vcc_lo
; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
;
; GFX12-TRUE16-LABEL: v_maximumnum_bf16_nnan:
; GFX12-TRUE16:       ; %bb.0:
; GFX12-TRUE16-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX12-TRUE16-NEXT:    s_wait_expcnt 0x0
; GFX12-TRUE16-NEXT:    s_wait_samplecnt 0x0
; GFX12-TRUE16-NEXT:    s_wait_bvhcnt 0x0
; GFX12-TRUE16-NEXT:    s_wait_kmcnt 0x0
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v1
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v0
; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_2)
; GFX12-TRUE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v3, v2
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v2.l, v1.l, v0.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v0.l
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v3.l, v2.l
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v0.l, v2.l, v0.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v1.l
; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v0.l, v0.l, v1.l, vcc_lo
; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX12-TRUE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v3
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v0.l, v2.l, v0.l, vcc_lo
; GFX12-TRUE16-NEXT:    s_setpc_b64 s[30:31]
;
; GFX12-FAKE16-LABEL: v_maximumnum_bf16_nnan:
; GFX12-FAKE16:       ; %bb.0:
; GFX12-FAKE16-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX12-FAKE16-NEXT:    s_wait_expcnt 0x0
; GFX12-FAKE16-NEXT:    s_wait_samplecnt 0x0
; GFX12-FAKE16-NEXT:    s_wait_bvhcnt 0x0
; GFX12-FAKE16-NEXT:    s_wait_kmcnt 0x0
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v1
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v0
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_4) | instid1(VALU_DEP_2)
; GFX12-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v3, v2
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v1, v0, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v0
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v0, v2, v0 :: v_dual_lshlrev_b32 v3, 16, v2
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v1
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v0, v1, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v3
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v2, v0, vcc_lo
; GFX12-FAKE16-NEXT:    s_setpc_b64 s[30:31]
  %result = call nnan bfloat @llvm.maximumnum.bf16(bfloat %x, bfloat %y)
  ret bfloat %result
}

define <2 x bfloat> @v_maximumnum_v2bf16(<2 x bfloat> %x, <2 x bfloat> %y) {
; GFX7-LABEL: v_maximumnum_v2bf16:
; GFX7:       ; %bb.0:
; GFX7-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT:    v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT:    v_mul_f32_e32 v2, 1.0, v2
; GFX7-NEXT:    v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT:    v_mul_f32_e32 v3, 1.0, v3
; GFX7-NEXT:    v_and_b32_e32 v3, 0xffff0000, v3
; GFX7-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT:    v_and_b32_e32 v2, 0xffff0000, v2
; GFX7-NEXT:    v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT:    v_max_f32_e32 v1, v1, v3
; GFX7-NEXT:    v_max_f32_e32 v0, v0, v2
; GFX7-NEXT:    v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT:    s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_maximumnum_v2bf16:
; GFX8:       ; %bb.0:
; GFX8-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT:    v_and_b32_e32 v2, 0xffff0000, v0
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v2, v2
; GFX8-NEXT:    v_lshrrev_b32_e32 v2, 16, v1
; GFX8-NEXT:    v_and_b32_e32 v5, 0xffff0000, v1
; GFX8-NEXT:    v_cndmask_b32_sdwa v3, v0, v2, vcc dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; GFX8-NEXT:    v_cndmask_b32_e32 v2, v2, v3, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v4, 16, v3
; GFX8-NEXT:    v_lshlrev_b32_e32 v5, 16, v2
; GFX8-NEXT:    v_cmp_gt_f32_e32 vcc, v4, v5
; GFX8-NEXT:    v_cndmask_b32_e32 v4, v2, v3, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; GFX8-NEXT:    v_mul_f32_e32 v4, 1.0, v4
; GFX8-NEXT:    v_bfe_u32 v5, v4, 16, 1
; GFX8-NEXT:    v_add_u32_e32 v5, vcc, v5, v4
; GFX8-NEXT:    s_movk_i32 s4, 0x7fff
; GFX8-NEXT:    v_add_u32_e32 v5, vcc, s4, v5
; GFX8-NEXT:    v_or_b32_e32 v6, 0x400000, v4
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v4, v4
; GFX8-NEXT:    v_cndmask_b32_e32 v4, v5, v6, vcc
; GFX8-NEXT:    v_lshrrev_b32_e32 v5, 16, v4
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v3
; GFX8-NEXT:    v_cndmask_b32_e32 v3, v5, v3, vcc
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v2
; GFX8-NEXT:    v_cndmask_b32_e32 v2, v3, v2, vcc
; GFX8-NEXT:    v_and_b32_e32 v3, 0xffff0000, v4
; GFX8-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v3
; GFX8-NEXT:    v_lshlrev_b32_e32 v3, 16, v0
; GFX8-NEXT:    v_cndmask_b32_e32 v2, v5, v2, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; GFX8-NEXT:    v_lshlrev_b32_e32 v3, 16, v1
; GFX8-NEXT:    v_cndmask_b32_e32 v0, v0, v1, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; GFX8-NEXT:    v_cndmask_b32_e32 v1, v1, v0, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v3, 16, v1
; GFX8-NEXT:    v_lshlrev_b32_e32 v4, 16, v0
; GFX8-NEXT:    v_cmp_gt_f32_e32 vcc, v4, v3
; GFX8-NEXT:    v_cndmask_b32_e32 v3, v1, v0, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; GFX8-NEXT:    v_mul_f32_e32 v3, 1.0, v3
; GFX8-NEXT:    v_bfe_u32 v4, v3, 16, 1
; GFX8-NEXT:    v_add_u32_e32 v4, vcc, v4, v3
; GFX8-NEXT:    v_add_u32_e32 v4, vcc, s4, v4
; GFX8-NEXT:    v_or_b32_e32 v5, 0x400000, v3
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; GFX8-NEXT:    v_cndmask_b32_e32 v3, v4, v5, vcc
; GFX8-NEXT:    v_lshrrev_b32_e32 v4, 16, v3
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v0
; GFX8-NEXT:    v_cndmask_b32_e32 v0, v4, v0, vcc
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v1
; GFX8-NEXT:    v_cndmask_b32_e32 v0, v0, v1, vcc
; GFX8-NEXT:    v_and_b32_e32 v1, 0xffff0000, v3
; GFX8-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v1
; GFX8-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; GFX8-NEXT:    v_cndmask_b32_e32 v0, v4, v0, vcc
; GFX8-NEXT:    v_or_b32_sdwa v0, v0, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX8-NEXT:    s_setpc_b64 s[30:31]
;
; GFX900-LABEL: v_maximumnum_v2bf16:
; GFX900:       ; %bb.0:
; GFX900-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX900-NEXT:    v_and_b32_e32 v2, 0xffff0000, v0
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v2, v2
; GFX900-NEXT:    v_lshrrev_b32_e32 v2, 16, v1
; GFX900-NEXT:    v_and_b32_e32 v5, 0xffff0000, v1
; GFX900-NEXT:    v_cndmask_b32_sdwa v3, v0, v2, vcc dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; GFX900-NEXT:    v_cndmask_b32_e32 v2, v2, v3, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v4, 16, v3
; GFX900-NEXT:    v_lshlrev_b32_e32 v5, 16, v2
; GFX900-NEXT:    v_cmp_gt_f32_e32 vcc, v4, v5
; GFX900-NEXT:    v_cndmask_b32_e32 v4, v2, v3, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; GFX900-NEXT:    v_max_f32_e32 v4, v4, v4
; GFX900-NEXT:    v_bfe_u32 v5, v4, 16, 1
; GFX900-NEXT:    s_movk_i32 s4, 0x7fff
; GFX900-NEXT:    v_add3_u32 v5, v5, v4, s4
; GFX900-NEXT:    v_or_b32_e32 v6, 0x400000, v4
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v4, v4
; GFX900-NEXT:    v_cndmask_b32_e32 v4, v5, v6, vcc
; GFX900-NEXT:    v_lshrrev_b32_e32 v5, 16, v4
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v3
; GFX900-NEXT:    v_cndmask_b32_e32 v3, v5, v3, vcc
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v2
; GFX900-NEXT:    v_cndmask_b32_e32 v2, v3, v2, vcc
; GFX900-NEXT:    v_and_b32_e32 v3, 0xffff0000, v4
; GFX900-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v3
; GFX900-NEXT:    v_lshlrev_b32_e32 v3, 16, v0
; GFX900-NEXT:    v_cndmask_b32_e32 v2, v5, v2, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; GFX900-NEXT:    v_lshlrev_b32_e32 v3, 16, v1
; GFX900-NEXT:    v_cndmask_b32_e32 v0, v0, v1, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; GFX900-NEXT:    v_cndmask_b32_e32 v1, v1, v0, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v3, 16, v1
; GFX900-NEXT:    v_lshlrev_b32_e32 v4, 16, v0
; GFX900-NEXT:    v_cmp_gt_f32_e32 vcc, v4, v3
; GFX900-NEXT:    v_cndmask_b32_e32 v3, v1, v0, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; GFX900-NEXT:    v_max_f32_e32 v3, v3, v3
; GFX900-NEXT:    v_bfe_u32 v4, v3, 16, 1
; GFX900-NEXT:    v_add3_u32 v4, v4, v3, s4
; GFX900-NEXT:    v_or_b32_e32 v5, 0x400000, v3
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; GFX900-NEXT:    v_cndmask_b32_e32 v3, v4, v5, vcc
; GFX900-NEXT:    v_lshrrev_b32_e32 v4, 16, v3
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v0
; GFX900-NEXT:    v_cndmask_b32_e32 v0, v4, v0, vcc
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v1
; GFX900-NEXT:    v_cndmask_b32_e32 v0, v0, v1, vcc
; GFX900-NEXT:    v_and_b32_e32 v1, 0xffff0000, v3
; GFX900-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v1
; GFX900-NEXT:    v_cndmask_b32_e32 v0, v4, v0, vcc
; GFX900-NEXT:    s_mov_b32 s4, 0x5040100
; GFX900-NEXT:    v_perm_b32 v0, v2, v0, s4
; GFX900-NEXT:    s_setpc_b64 s[30:31]
;
; GFX950-LABEL: v_maximumnum_v2bf16:
; GFX950:       ; %bb.0:
; GFX950-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX950-NEXT:    v_and_b32_e32 v2, 0xffff0000, v0
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v2, v2
; GFX950-NEXT:    v_lshrrev_b32_e32 v2, 16, v1
; GFX950-NEXT:    v_and_b32_e32 v5, 0xffff0000, v1
; GFX950-NEXT:    v_cndmask_b32_sdwa v3, v0, v2, vcc dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; GFX950-NEXT:    v_lshlrev_b32_e32 v4, 16, v3
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v2, v2, v3, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v5, 16, v2
; GFX950-NEXT:    v_cmp_gt_f32_e32 vcc, v4, v5
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v4, v2, v3, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; GFX950-NEXT:    v_max_f32_e32 v4, v4, v4
; GFX950-NEXT:    v_cvt_pk_bf16_f32 v4, v4, s0
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v3
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v3, v4, v3, vcc
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v2
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v2, v3, v2, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v3, 16, v4
; GFX950-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v3
; GFX950-NEXT:    v_lshlrev_b32_e32 v3, 16, v0
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v2, v4, v2, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; GFX950-NEXT:    v_lshlrev_b32_e32 v3, 16, v1
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v0, v0, v1, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; GFX950-NEXT:    v_lshlrev_b32_e32 v4, 16, v0
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v1, v1, v0, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v3, 16, v1
; GFX950-NEXT:    v_cmp_gt_f32_e32 vcc, v4, v3
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v3, v1, v0, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; GFX950-NEXT:    v_max_f32_e32 v3, v3, v3
; GFX950-NEXT:    v_cvt_pk_bf16_f32 v3, v3, s0
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v0
; GFX950-NEXT:    s_mov_b32 s0, 0x5040100
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v0, v3, v0, vcc
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v1
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v0, v0, v1, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v1, 16, v3
; GFX950-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v1
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v0, v3, v0, vcc
; GFX950-NEXT:    v_perm_b32 v0, v2, v0, s0
; GFX950-NEXT:    s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_maximumnum_v2bf16:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT:    v_and_b32_e32 v2, 0xffff0000, v0
; GFX10-NEXT:    v_lshlrev_b32_e32 v4, 16, v0
; GFX10-NEXT:    v_and_b32_e32 v3, 0xffff0000, v1
; GFX10-NEXT:    v_lshlrev_b32_e32 v5, 16, v1
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX10-NEXT:    v_lshrrev_b32_e32 v2, 16, v1
; GFX10-NEXT:    v_cndmask_b32_sdwa v6, v0, v2, vcc_lo dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX10-NEXT:    v_cndmask_b32_e32 v0, v0, v1, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX10-NEXT:    v_lshlrev_b32_e32 v3, 16, v6
; GFX10-NEXT:    v_cndmask_b32_e32 v2, v2, v6, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX10-NEXT:    v_lshlrev_b32_e32 v5, 16, v0
; GFX10-NEXT:    v_lshlrev_b32_e32 v4, 16, v2
; GFX10-NEXT:    v_cndmask_b32_e32 v1, v1, v0, vcc_lo
; GFX10-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v3, v4
; GFX10-NEXT:    v_lshlrev_b32_e32 v7, 16, v1
; GFX10-NEXT:    v_cndmask_b32_e32 v3, v2, v6, vcc_lo
; GFX10-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v5, v7
; GFX10-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; GFX10-NEXT:    v_cndmask_b32_e32 v4, v1, v0, vcc_lo
; GFX10-NEXT:    v_max_f32_e32 v3, v3, v3
; GFX10-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; GFX10-NEXT:    v_bfe_u32 v5, v3, 16, 1
; GFX10-NEXT:    v_max_f32_e32 v4, v4, v4
; GFX10-NEXT:    v_or_b32_e32 v8, 0x400000, v3
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX10-NEXT:    v_add3_u32 v5, v5, v3, 0x7fff
; GFX10-NEXT:    v_bfe_u32 v7, v4, 16, 1
; GFX10-NEXT:    v_or_b32_e32 v9, 0x400000, v4
; GFX10-NEXT:    v_cndmask_b32_e32 v3, v5, v8, vcc_lo
; GFX10-NEXT:    v_add3_u32 v7, v7, v4, 0x7fff
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX10-NEXT:    v_lshrrev_b32_e32 v5, 16, v3
; GFX10-NEXT:    v_and_b32_e32 v3, 0xffff0000, v3
; GFX10-NEXT:    v_cndmask_b32_e32 v4, v7, v9, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v6
; GFX10-NEXT:    v_lshrrev_b32_e32 v7, 16, v4
; GFX10-NEXT:    v_cndmask_b32_e32 v6, v5, v6, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v0
; GFX10-NEXT:    v_and_b32_e32 v4, 0xffff0000, v4
; GFX10-NEXT:    v_cndmask_b32_e32 v0, v7, v0, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v2
; GFX10-NEXT:    v_cndmask_b32_e32 v2, v6, v2, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v1
; GFX10-NEXT:    v_cndmask_b32_e32 v0, v0, v1, vcc_lo
; GFX10-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v3
; GFX10-NEXT:    v_cndmask_b32_e32 v1, v5, v2, vcc_lo
; GFX10-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v4
; GFX10-NEXT:    v_cndmask_b32_e32 v0, v7, v0, vcc_lo
; GFX10-NEXT:    v_perm_b32 v0, v1, v0, 0x5040100
; GFX10-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-TRUE16-LABEL: v_maximumnum_v2bf16:
; GFX11-TRUE16:       ; %bb.0:
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff0000, v0
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v0
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xffff0000, v1
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 16, v1
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s0, v3, v3
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s1, v4, v4
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v2.l, v0.h, v1.h, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v0.l, v0.l, v1.l, s0
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v3.l, v1.h, v2.l, s1
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, v2.l
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v1.l, v1.l, v0.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, v0.l
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v5.l, v3.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.l, v1.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s2, 0, v1.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; GFX11-TRUE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v4, v5
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_cmp_gt_f32_e64 s0, v6, v7
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v4.l, v3.l, v2.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v2.l
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v5.l, v1.l, v0.l, s0
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s0, 0, v0.l
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; GFX11-TRUE16-NEXT:    v_dual_max_f32 v4, v4, v4 :: v_dual_lshlrev_b32 v5, 16, v5
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT:    v_max_f32_e32 v5, v5, v5
; GFX11-TRUE16-NEXT:    v_bfe_u32 v6, v4, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, 0x400000, v4
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v7, v5, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s1, v4, v4
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, 0x400000, v5
; GFX11-TRUE16-NEXT:    v_add3_u32 v6, v6, v4, 0x7fff
; GFX11-TRUE16-NEXT:    v_add3_u32 v7, v7, v5, 0x7fff
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e64 v4, v6, v8, s1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s1, v5, v5
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v0.h, v4.h, v2.l, vcc_lo
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e64 v5, v7, v9, s1
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s1, 0, v3.l
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff0000, v4
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v0.l, v5.h, v0.l, s0
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xffff0000, v5
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v0.h, v0.h, v3.l, s1
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v2
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v0.l, v0.l, v1.l, s2
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_cmp_eq_f32_e64 s0, 0, v6
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v0.h, v4.h, v0.h, vcc_lo
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v0.l, v5.h, v0.l, s0
; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-FAKE16-LABEL: v_maximumnum_v2bf16:
; GFX11-FAKE16:       ; %bb.0:
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xffff0000, v0
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v3, 16, v1
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v4, 16, v0
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v2, v4, v3 :: v_dual_lshlrev_b32 v5, 16, v0
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v6, 0xffff0000, v1
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v4, 16, v2
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v0, v1, vcc_lo
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v3, v3, v2, vcc_lo
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v7, 16, v1
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v5, 16, v3
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v6, 16, v0
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v1, v0, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v4, v5
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v4, v3, v2, vcc_lo
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; GFX11-FAKE16-NEXT:    v_dual_max_f32 v4, v4, v4 :: v_dual_lshlrev_b32 v7, 16, v1
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v6, v7
; GFX11-FAKE16-NEXT:    v_bfe_u32 v6, v4, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v5, v1, v0, vcc_lo
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, 0x400000, v4
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_add3_u32 v6, v6, v4, 0x7fff
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v4, v6, v8 :: v_dual_lshlrev_b32 v5, 16, v5
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_max_f32_e32 v5, v5, v5
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v6, 16, v4
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v7, v5, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, 0x400000, v5
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xffff0000, v4
; GFX11-FAKE16-NEXT:    v_add3_u32 v7, v7, v5, 0x7fff
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v5, v7, v9, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v2
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v7, 16, v5
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v6, v2, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v0
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v0, v7, v0 :: v_dual_and_b32 v5, 0xffff0000, v5
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v3
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v2, v3, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v1
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v0, v1, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v4
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v6, v2, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v5
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v7, v0, vcc_lo
; GFX11-FAKE16-NEXT:    v_perm_b32 v0, v1, v0, 0x5040100
; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
;
; GFX12-TRUE16-LABEL: v_maximumnum_v2bf16:
; GFX12-TRUE16:       ; %bb.0:
; GFX12-TRUE16-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX12-TRUE16-NEXT:    s_wait_expcnt 0x0
; GFX12-TRUE16-NEXT:    s_wait_samplecnt 0x0
; GFX12-TRUE16-NEXT:    s_wait_bvhcnt 0x0
; GFX12-TRUE16-NEXT:    s_wait_kmcnt 0x0
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff0000, v0
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v0
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v4, 0xffff0000, v1
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 16, v1
; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s0, v3, v3
; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s1, v4, v4
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v2.l, v0.h, v1.h, vcc_lo
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v0.l, v0.l, v1.l, s0
; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v3.l, v1.h, v2.l, s1
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v4.l, v2.l
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v1.l, v1.l, v0.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v6.l, v0.l
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v5.l, v3.l
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v7.l, v1.l
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s2, 0, v1.l
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; GFX12-TRUE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v4, v5
; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
; GFX12-TRUE16-NEXT:    v_cmp_gt_f32_e64 s0, v6, v7
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v4.l, v3.l, v2.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v2.l
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v5.l, v1.l, v0.l, s0
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s0, 0, v0.l
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT:    v_dual_max_num_f32 v4, v4, v4 :: v_dual_lshlrev_b32 v5, 16, v5
; GFX12-TRUE16-NEXT:    v_max_num_f32_e32 v5, v5, v5
; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX12-TRUE16-NEXT:    v_bfe_u32 v6, v4, 16, 1
; GFX12-TRUE16-NEXT:    v_or_b32_e32 v8, 0x400000, v4
; GFX12-TRUE16-NEXT:    v_bfe_u32 v7, v5, 16, 1
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s1, v4, v4
; GFX12-TRUE16-NEXT:    v_or_b32_e32 v9, 0x400000, v5
; GFX12-TRUE16-NEXT:    v_add3_u32 v6, v6, v4, 0x7fff
; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX12-TRUE16-NEXT:    v_add3_u32 v7, v7, v5, 0x7fff
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b32_e64 v4, v6, v8, s1
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s1, v5, v5
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v0.h, v4.h, v2.l, vcc_lo
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b32_e64 v5, v7, v9, s1
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s1, 0, v3.l
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff0000, v4
; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v0.l, v5.h, v0.l, s0
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v6, 0xffff0000, v5
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v0.h, v0.h, v3.l, s1
; GFX12-TRUE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v2
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v0.l, v0.l, v1.l, s2
; GFX12-TRUE16-NEXT:    v_cmp_eq_f32_e64 s0, 0, v6
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v0.h, v4.h, v0.h, vcc_lo
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v0.l, v5.h, v0.l, s0
; GFX12-TRUE16-NEXT:    s_setpc_b64 s[30:31]
;
; GFX12-FAKE16-LABEL: v_maximumnum_v2bf16:
; GFX12-FAKE16:       ; %bb.0:
; GFX12-FAKE16-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX12-FAKE16-NEXT:    s_wait_expcnt 0x0
; GFX12-FAKE16-NEXT:    s_wait_samplecnt 0x0
; GFX12-FAKE16-NEXT:    s_wait_bvhcnt 0x0
; GFX12-FAKE16-NEXT:    s_wait_kmcnt 0x0
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v2, 0xffff0000, v0
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v3, 16, v1
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v4, 16, v0
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v2, v4, v3 :: v_dual_lshlrev_b32 v5, 16, v0
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v6, 0xffff0000, v1
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v4, 16, v2
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v0, v1, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v3, v3, v2, vcc_lo
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v7, 16, v1
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v5, 16, v3
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v6, 16, v0
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v1, v0, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v4, v5
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v4, v3, v2, vcc_lo
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; GFX12-FAKE16-NEXT:    v_dual_max_num_f32 v4, v4, v4 :: v_dual_lshlrev_b32 v7, 16, v1
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX12-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v6, v7
; GFX12-FAKE16-NEXT:    v_bfe_u32 v6, v4, 16, 1
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v5, v1, v0, vcc_lo
; GFX12-FAKE16-NEXT:    v_or_b32_e32 v8, 0x400000, v4
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX12-FAKE16-NEXT:    v_add3_u32 v6, v6, v4, 0x7fff
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v4, v6, v8 :: v_dual_lshlrev_b32 v5, 16, v5
; GFX12-FAKE16-NEXT:    v_max_num_f32_e32 v5, v5, v5
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v6, 16, v4
; GFX12-FAKE16-NEXT:    v_bfe_u32 v7, v5, 16, 1
; GFX12-FAKE16-NEXT:    v_or_b32_e32 v9, 0x400000, v5
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v4, 0xffff0000, v4
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX12-FAKE16-NEXT:    v_add3_u32 v7, v7, v5, 0x7fff
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v5, v7, v9, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v2
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v7, 16, v5
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v6, v2, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v0
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v0, v7, v0 :: v_dual_and_b32 v5, 0xffff0000, v5
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v3
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v2, v3, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v1
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v0, v1, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v4
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v6, v2, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v5
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v7, v0, vcc_lo
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX12-FAKE16-NEXT:    v_perm_b32 v0, v1, v0, 0x5040100
; GFX12-FAKE16-NEXT:    s_setpc_b64 s[30:31]
  %result = call <2 x bfloat> @llvm.maximumnum.v2bf16(<2 x bfloat> %x, <2 x bfloat> %y)
  ret <2 x bfloat> %result
}

define <2 x bfloat> @v_maximumnum_v2bf16_nnan(<2 x bfloat> %x, <2 x bfloat> %y) {
; GFX7-LABEL: v_maximumnum_v2bf16_nnan:
; GFX7:       ; %bb.0:
; GFX7-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT:    v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT:    v_mul_f32_e32 v2, 1.0, v2
; GFX7-NEXT:    v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT:    v_mul_f32_e32 v3, 1.0, v3
; GFX7-NEXT:    v_and_b32_e32 v3, 0xffff0000, v3
; GFX7-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT:    v_and_b32_e32 v2, 0xffff0000, v2
; GFX7-NEXT:    v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT:    v_max_f32_e32 v1, v1, v3
; GFX7-NEXT:    v_max_f32_e32 v0, v0, v2
; GFX7-NEXT:    v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT:    s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_maximumnum_v2bf16_nnan:
; GFX8:       ; %bb.0:
; GFX8-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT:    v_lshlrev_b32_e32 v2, 16, v1
; GFX8-NEXT:    v_lshlrev_b32_e32 v3, 16, v0
; GFX8-NEXT:    v_cmp_gt_f32_e32 vcc, v3, v2
; GFX8-NEXT:    v_cndmask_b32_e32 v2, v1, v0, vcc
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v0
; GFX8-NEXT:    v_cndmask_b32_e32 v3, v2, v0, vcc
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v1
; GFX8-NEXT:    v_lshlrev_b32_e32 v4, 16, v2
; GFX8-NEXT:    v_cndmask_b32_e32 v3, v3, v1, vcc
; GFX8-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v4
; GFX8-NEXT:    v_cndmask_b32_e32 v2, v2, v3, vcc
; GFX8-NEXT:    v_and_b32_e32 v3, 0xffff0000, v1
; GFX8-NEXT:    v_and_b32_e32 v4, 0xffff0000, v0
; GFX8-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; GFX8-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
; GFX8-NEXT:    v_cmp_gt_f32_e32 vcc, v4, v3
; GFX8-NEXT:    v_cndmask_b32_e32 v3, v1, v0, vcc
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v0
; GFX8-NEXT:    v_cndmask_b32_e32 v0, v3, v0, vcc
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v1
; GFX8-NEXT:    v_cndmask_b32_e32 v0, v0, v1, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v1, 16, v3
; GFX8-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v1
; GFX8-NEXT:    v_cndmask_b32_sdwa v0, v3, v0, vcc dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX8-NEXT:    v_or_b32_sdwa v0, v2, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX8-NEXT:    s_setpc_b64 s[30:31]
;
; GFX900-LABEL: v_maximumnum_v2bf16_nnan:
; GFX900:       ; %bb.0:
; GFX900-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX900-NEXT:    v_lshlrev_b32_e32 v2, 16, v1
; GFX900-NEXT:    v_lshlrev_b32_e32 v3, 16, v0
; GFX900-NEXT:    v_cmp_gt_f32_e32 vcc, v3, v2
; GFX900-NEXT:    v_cndmask_b32_e32 v2, v1, v0, vcc
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v0
; GFX900-NEXT:    v_cndmask_b32_e32 v3, v2, v0, vcc
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v1
; GFX900-NEXT:    v_lshlrev_b32_e32 v4, 16, v2
; GFX900-NEXT:    v_cndmask_b32_e32 v3, v3, v1, vcc
; GFX900-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v4
; GFX900-NEXT:    v_cndmask_b32_e32 v2, v2, v3, vcc
; GFX900-NEXT:    v_and_b32_e32 v3, 0xffff0000, v1
; GFX900-NEXT:    v_and_b32_e32 v4, 0xffff0000, v0
; GFX900-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; GFX900-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
; GFX900-NEXT:    v_cmp_gt_f32_e32 vcc, v4, v3
; GFX900-NEXT:    v_cndmask_b32_e32 v3, v1, v0, vcc
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v0
; GFX900-NEXT:    v_cndmask_b32_e32 v0, v3, v0, vcc
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v1
; GFX900-NEXT:    v_cndmask_b32_e32 v0, v0, v1, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v1, 16, v3
; GFX900-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v1
; GFX900-NEXT:    v_cndmask_b32_e32 v0, v3, v0, vcc
; GFX900-NEXT:    s_mov_b32 s4, 0x5040100
; GFX900-NEXT:    v_perm_b32 v0, v0, v2, s4
; GFX900-NEXT:    s_setpc_b64 s[30:31]
;
; GFX950-LABEL: v_maximumnum_v2bf16_nnan:
; GFX950:       ; %bb.0:
; GFX950-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX950-NEXT:    v_lshlrev_b32_e32 v2, 16, v1
; GFX950-NEXT:    v_lshlrev_b32_e32 v3, 16, v0
; GFX950-NEXT:    v_cmp_gt_f32_e32 vcc, v3, v2
; GFX950-NEXT:    s_mov_b32 s0, 0x5040100
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v2, v1, v0, vcc
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v0
; GFX950-NEXT:    v_lshlrev_b32_e32 v4, 16, v2
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v3, v2, v0, vcc
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v1
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v3, v3, v1, vcc
; GFX950-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v4
; GFX950-NEXT:    v_and_b32_e32 v4, 0xffff0000, v0
; GFX950-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; GFX950-NEXT:    v_cndmask_b32_e32 v2, v2, v3, vcc
; GFX950-NEXT:    v_and_b32_e32 v3, 0xffff0000, v1
; GFX950-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
; GFX950-NEXT:    v_cmp_gt_f32_e32 vcc, v4, v3
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v3, v1, v0, vcc
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v0
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v0, v3, v0, vcc
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v1
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v0, v0, v1, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v1, 16, v3
; GFX950-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v1
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v0, v3, v0, vcc
; GFX950-NEXT:    v_perm_b32 v0, v0, v2, s0
; GFX950-NEXT:    s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_maximumnum_v2bf16_nnan:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT:    v_lshlrev_b32_e32 v2, 16, v1
; GFX10-NEXT:    v_lshlrev_b32_e32 v3, 16, v0
; GFX10-NEXT:    v_and_b32_e32 v4, 0xffff0000, v1
; GFX10-NEXT:    v_and_b32_e32 v5, 0xffff0000, v0
; GFX10-NEXT:    v_lshrrev_b32_e32 v6, 16, v0
; GFX10-NEXT:    v_lshrrev_b32_e32 v7, 16, v1
; GFX10-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v3, v2
; GFX10-NEXT:    v_cndmask_b32_e32 v2, v1, v0, vcc_lo
; GFX10-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v5, v4
; GFX10-NEXT:    v_lshlrev_b32_e32 v4, 16, v2
; GFX10-NEXT:    v_cndmask_b32_e32 v3, v7, v6, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v0
; GFX10-NEXT:    v_cndmask_b32_e32 v0, v2, v0, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v6
; GFX10-NEXT:    v_cndmask_b32_e32 v5, v3, v6, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v1
; GFX10-NEXT:    v_lshlrev_b32_e32 v6, 16, v3
; GFX10-NEXT:    v_cndmask_b32_e32 v0, v0, v1, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v7
; GFX10-NEXT:    v_cndmask_b32_e32 v1, v5, v7, vcc_lo
; GFX10-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v4
; GFX10-NEXT:    v_cndmask_b32_e32 v0, v2, v0, vcc_lo
; GFX10-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v6
; GFX10-NEXT:    v_cndmask_b32_e32 v1, v3, v1, vcc_lo
; GFX10-NEXT:    v_perm_b32 v0, v1, v0, 0x5040100
; GFX10-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-TRUE16-LABEL: v_maximumnum_v2bf16_nnan:
; GFX11-TRUE16:       ; %bb.0:
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v1
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v0
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xffff0000, v1
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xffff0000, v0
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s1, 0, v0.h
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s2, 0, v1.h
; GFX11-TRUE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v3, v2
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_cmp_gt_f32_e64 s0, v5, v4
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v2.l, v1.l, v0.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v0.l
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v3.l, v1.h, v0.h, s0
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s0, 0, v1.l
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, v2.l
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v0.l, v2.l, v0.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v5.l, v3.l
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v0.h, v3.l, v0.h, s1
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v0.l, v0.l, v1.l, s0
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v0.h, v0.h, v1.h, s2
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v4
; GFX11-TRUE16-NEXT:    v_cmp_eq_f32_e64 s0, 0, v5
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v0.l, v2.l, v0.l, vcc_lo
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v0.h, v3.l, v0.h, s0
; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-FAKE16-LABEL: v_maximumnum_v2bf16_nnan:
; GFX11-FAKE16:       ; %bb.0:
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v0
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v6, 16, v0
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v7, 16, v1
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v1
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xffff0000, v1
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v5, 0xffff0000, v0
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v3, v2
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v1, v0, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v5, v4
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_4) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v4, 16, v2
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v3, v7, v6, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v0
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v2, v0, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v6
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v5, v3, v6, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v1
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v6, 16, v3
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v0, v1, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v7
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v5, v7, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v4
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v2, v0, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v6
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v3, v1, vcc_lo
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_perm_b32 v0, v1, v0, 0x5040100
; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
;
; GFX12-TRUE16-LABEL: v_maximumnum_v2bf16_nnan:
; GFX12-TRUE16:       ; %bb.0:
; GFX12-TRUE16-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX12-TRUE16-NEXT:    s_wait_expcnt 0x0
; GFX12-TRUE16-NEXT:    s_wait_samplecnt 0x0
; GFX12-TRUE16-NEXT:    s_wait_bvhcnt 0x0
; GFX12-TRUE16-NEXT:    s_wait_kmcnt 0x0
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v1
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v0
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v4, 0xffff0000, v1
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v5, 0xffff0000, v0
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s1, 0, v0.h
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s2, 0, v1.h
; GFX12-TRUE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v3, v2
; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX12-TRUE16-NEXT:    v_cmp_gt_f32_e64 s0, v5, v4
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v2.l, v1.l, v0.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v0.l
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v3.l, v1.h, v0.h, s0
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s0, 0, v1.l
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v4.l, v2.l
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v0.l, v2.l, v0.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v5.l, v3.l
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v0.h, v3.l, v0.h, s1
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v0.l, v0.l, v1.l, s0
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v0.h, v0.h, v1.h, s2
; GFX12-TRUE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v4
; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX12-TRUE16-NEXT:    v_cmp_eq_f32_e64 s0, 0, v5
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v0.l, v2.l, v0.l, vcc_lo
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v0.h, v3.l, v0.h, s0
; GFX12-TRUE16-NEXT:    s_setpc_b64 s[30:31]
;
; GFX12-FAKE16-LABEL: v_maximumnum_v2bf16_nnan:
; GFX12-FAKE16:       ; %bb.0:
; GFX12-FAKE16-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX12-FAKE16-NEXT:    s_wait_expcnt 0x0
; GFX12-FAKE16-NEXT:    s_wait_samplecnt 0x0
; GFX12-FAKE16-NEXT:    s_wait_bvhcnt 0x0
; GFX12-FAKE16-NEXT:    s_wait_kmcnt 0x0
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v0
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v6, 16, v0
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v7, 16, v1
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v1
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v4, 0xffff0000, v1
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v5, 0xffff0000, v0
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX12-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v3, v2
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v1, v0, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v5, v4
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v4, 16, v2
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v3, v7, v6, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v0
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v2, v0, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v6
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v5, v3, v6, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v1
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v6, 16, v3
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v0, v1, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v7
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v5, v7, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v4
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v2, v0, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v6
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v3, v1, vcc_lo
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX12-FAKE16-NEXT:    v_perm_b32 v0, v1, v0, 0x5040100
; GFX12-FAKE16-NEXT:    s_setpc_b64 s[30:31]
  %result = call nnan <2 x bfloat> @llvm.maximumnum.v2bf16(<2 x bfloat> %x, <2 x bfloat> %y)
  ret <2 x bfloat> %result
}

define <3 x bfloat> @v_maximumnum_v3bf16(<3 x bfloat> %x, <3 x bfloat> %y) {
; GFX7-LABEL: v_maximumnum_v3bf16:
; GFX7:       ; %bb.0:
; GFX7-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT:    v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT:    v_mul_f32_e32 v3, 1.0, v3
; GFX7-NEXT:    v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT:    v_mul_f32_e32 v4, 1.0, v4
; GFX7-NEXT:    v_mul_f32_e32 v2, 1.0, v2
; GFX7-NEXT:    v_mul_f32_e32 v5, 1.0, v5
; GFX7-NEXT:    v_and_b32_e32 v5, 0xffff0000, v5
; GFX7-NEXT:    v_and_b32_e32 v2, 0xffff0000, v2
; GFX7-NEXT:    v_and_b32_e32 v4, 0xffff0000, v4
; GFX7-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT:    v_and_b32_e32 v3, 0xffff0000, v3
; GFX7-NEXT:    v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT:    v_max_f32_e32 v2, v2, v5
; GFX7-NEXT:    v_max_f32_e32 v1, v1, v4
; GFX7-NEXT:    v_max_f32_e32 v0, v0, v3
; GFX7-NEXT:    v_and_b32_e32 v2, 0xffff0000, v2
; GFX7-NEXT:    v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT:    s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_maximumnum_v3bf16:
; GFX8:       ; %bb.0:
; GFX8-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT:    v_and_b32_e32 v4, 0xffff0000, v0
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v4, v4
; GFX8-NEXT:    v_lshrrev_b32_e32 v4, 16, v2
; GFX8-NEXT:    v_and_b32_e32 v7, 0xffff0000, v2
; GFX8-NEXT:    v_cndmask_b32_sdwa v5, v0, v4, vcc dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v7, v7
; GFX8-NEXT:    v_cndmask_b32_e32 v4, v4, v5, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v6, 16, v5
; GFX8-NEXT:    v_lshlrev_b32_e32 v7, 16, v4
; GFX8-NEXT:    v_cmp_gt_f32_e32 vcc, v6, v7
; GFX8-NEXT:    v_cndmask_b32_e32 v6, v4, v5, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; GFX8-NEXT:    v_mul_f32_e32 v6, 1.0, v6
; GFX8-NEXT:    v_bfe_u32 v7, v6, 16, 1
; GFX8-NEXT:    v_add_u32_e32 v7, vcc, v7, v6
; GFX8-NEXT:    s_movk_i32 s4, 0x7fff
; GFX8-NEXT:    v_add_u32_e32 v7, vcc, s4, v7
; GFX8-NEXT:    v_or_b32_e32 v8, 0x400000, v6
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v6, v6
; GFX8-NEXT:    v_cndmask_b32_e32 v6, v7, v8, vcc
; GFX8-NEXT:    v_lshrrev_b32_e32 v7, 16, v6
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v5
; GFX8-NEXT:    v_cndmask_b32_e32 v5, v7, v5, vcc
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v4
; GFX8-NEXT:    v_cndmask_b32_e32 v4, v5, v4, vcc
; GFX8-NEXT:    v_and_b32_e32 v5, 0xffff0000, v6
; GFX8-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v5
; GFX8-NEXT:    v_lshlrev_b32_e32 v5, 16, v1
; GFX8-NEXT:    v_cndmask_b32_e32 v4, v7, v4, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; GFX8-NEXT:    v_lshlrev_b32_e32 v5, 16, v3
; GFX8-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; GFX8-NEXT:    v_cndmask_b32_e32 v3, v3, v1, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v5, 16, v3
; GFX8-NEXT:    v_lshlrev_b32_e32 v6, 16, v1
; GFX8-NEXT:    v_cmp_gt_f32_e32 vcc, v6, v5
; GFX8-NEXT:    v_cndmask_b32_e32 v5, v3, v1, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; GFX8-NEXT:    v_mul_f32_e32 v5, 1.0, v5
; GFX8-NEXT:    v_bfe_u32 v6, v5, 16, 1
; GFX8-NEXT:    v_add_u32_e32 v6, vcc, v6, v5
; GFX8-NEXT:    v_add_u32_e32 v6, vcc, s4, v6
; GFX8-NEXT:    v_or_b32_e32 v7, 0x400000, v5
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; GFX8-NEXT:    v_cndmask_b32_e32 v5, v6, v7, vcc
; GFX8-NEXT:    v_lshrrev_b32_e32 v6, 16, v5
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v1
; GFX8-NEXT:    v_cndmask_b32_e32 v1, v6, v1, vcc
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v3
; GFX8-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc
; GFX8-NEXT:    v_and_b32_e32 v3, 0xffff0000, v5
; GFX8-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v3
; GFX8-NEXT:    v_lshlrev_b32_e32 v3, 16, v0
; GFX8-NEXT:    v_cndmask_b32_e32 v1, v6, v1, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; GFX8-NEXT:    v_lshlrev_b32_e32 v3, 16, v2
; GFX8-NEXT:    v_cndmask_b32_e32 v0, v0, v2, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; GFX8-NEXT:    v_cndmask_b32_e32 v2, v2, v0, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v3, 16, v2
; GFX8-NEXT:    v_lshlrev_b32_e32 v5, 16, v0
; GFX8-NEXT:    v_cmp_gt_f32_e32 vcc, v5, v3
; GFX8-NEXT:    v_cndmask_b32_e32 v3, v2, v0, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; GFX8-NEXT:    v_mul_f32_e32 v3, 1.0, v3
; GFX8-NEXT:    v_bfe_u32 v5, v3, 16, 1
; GFX8-NEXT:    v_add_u32_e32 v5, vcc, v5, v3
; GFX8-NEXT:    v_add_u32_e32 v5, vcc, s4, v5
; GFX8-NEXT:    v_or_b32_e32 v6, 0x400000, v3
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; GFX8-NEXT:    v_cndmask_b32_e32 v3, v5, v6, vcc
; GFX8-NEXT:    v_lshrrev_b32_e32 v5, 16, v3
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v0
; GFX8-NEXT:    v_cndmask_b32_e32 v0, v5, v0, vcc
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v2
; GFX8-NEXT:    v_cndmask_b32_e32 v0, v0, v2, vcc
; GFX8-NEXT:    v_and_b32_e32 v2, 0xffff0000, v3
; GFX8-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v2
; GFX8-NEXT:    v_cndmask_b32_e32 v0, v5, v0, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v2, 16, v4
; GFX8-NEXT:    v_or_b32_sdwa v0, v0, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX8-NEXT:    s_setpc_b64 s[30:31]
;
; GFX900-LABEL: v_maximumnum_v3bf16:
; GFX900:       ; %bb.0:
; GFX900-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX900-NEXT:    v_and_b32_e32 v4, 0xffff0000, v0
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v4, v4
; GFX900-NEXT:    v_lshrrev_b32_e32 v4, 16, v2
; GFX900-NEXT:    v_and_b32_e32 v7, 0xffff0000, v2
; GFX900-NEXT:    v_cndmask_b32_sdwa v5, v0, v4, vcc dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v7, v7
; GFX900-NEXT:    v_cndmask_b32_e32 v4, v4, v5, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v6, 16, v5
; GFX900-NEXT:    v_lshlrev_b32_e32 v7, 16, v4
; GFX900-NEXT:    v_cmp_gt_f32_e32 vcc, v6, v7
; GFX900-NEXT:    v_cndmask_b32_e32 v6, v4, v5, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; GFX900-NEXT:    v_max_f32_e32 v6, v6, v6
; GFX900-NEXT:    v_bfe_u32 v7, v6, 16, 1
; GFX900-NEXT:    s_movk_i32 s4, 0x7fff
; GFX900-NEXT:    v_add3_u32 v7, v7, v6, s4
; GFX900-NEXT:    v_or_b32_e32 v8, 0x400000, v6
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v6, v6
; GFX900-NEXT:    v_cndmask_b32_e32 v6, v7, v8, vcc
; GFX900-NEXT:    v_lshrrev_b32_e32 v7, 16, v6
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v5
; GFX900-NEXT:    v_cndmask_b32_e32 v5, v7, v5, vcc
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v4
; GFX900-NEXT:    v_cndmask_b32_e32 v4, v5, v4, vcc
; GFX900-NEXT:    v_and_b32_e32 v5, 0xffff0000, v6
; GFX900-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v5
; GFX900-NEXT:    v_lshlrev_b32_e32 v5, 16, v1
; GFX900-NEXT:    v_cndmask_b32_e32 v4, v7, v4, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; GFX900-NEXT:    v_lshlrev_b32_e32 v5, 16, v3
; GFX900-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; GFX900-NEXT:    v_cndmask_b32_e32 v3, v3, v1, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v5, 16, v3
; GFX900-NEXT:    v_lshlrev_b32_e32 v6, 16, v1
; GFX900-NEXT:    v_cmp_gt_f32_e32 vcc, v6, v5
; GFX900-NEXT:    v_cndmask_b32_e32 v5, v3, v1, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; GFX900-NEXT:    v_max_f32_e32 v5, v5, v5
; GFX900-NEXT:    v_bfe_u32 v6, v5, 16, 1
; GFX900-NEXT:    v_add3_u32 v6, v6, v5, s4
; GFX900-NEXT:    v_or_b32_e32 v7, 0x400000, v5
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; GFX900-NEXT:    v_cndmask_b32_e32 v5, v6, v7, vcc
; GFX900-NEXT:    v_lshrrev_b32_e32 v6, 16, v5
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v1
; GFX900-NEXT:    v_cndmask_b32_e32 v1, v6, v1, vcc
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v3
; GFX900-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc
; GFX900-NEXT:    v_and_b32_e32 v3, 0xffff0000, v5
; GFX900-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v3
; GFX900-NEXT:    v_lshlrev_b32_e32 v3, 16, v0
; GFX900-NEXT:    v_cndmask_b32_e32 v1, v6, v1, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; GFX900-NEXT:    v_lshlrev_b32_e32 v3, 16, v2
; GFX900-NEXT:    v_cndmask_b32_e32 v0, v0, v2, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; GFX900-NEXT:    v_cndmask_b32_e32 v2, v2, v0, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v3, 16, v2
; GFX900-NEXT:    v_lshlrev_b32_e32 v5, 16, v0
; GFX900-NEXT:    v_cmp_gt_f32_e32 vcc, v5, v3
; GFX900-NEXT:    v_cndmask_b32_e32 v3, v2, v0, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; GFX900-NEXT:    v_max_f32_e32 v3, v3, v3
; GFX900-NEXT:    v_bfe_u32 v5, v3, 16, 1
; GFX900-NEXT:    v_add3_u32 v5, v5, v3, s4
; GFX900-NEXT:    v_or_b32_e32 v6, 0x400000, v3
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; GFX900-NEXT:    v_cndmask_b32_e32 v3, v5, v6, vcc
; GFX900-NEXT:    v_lshrrev_b32_e32 v5, 16, v3
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v0
; GFX900-NEXT:    v_cndmask_b32_e32 v0, v5, v0, vcc
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v2
; GFX900-NEXT:    v_cndmask_b32_e32 v0, v0, v2, vcc
; GFX900-NEXT:    v_and_b32_e32 v2, 0xffff0000, v3
; GFX900-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v2
; GFX900-NEXT:    v_cndmask_b32_e32 v0, v5, v0, vcc
; GFX900-NEXT:    s_mov_b32 s4, 0x5040100
; GFX900-NEXT:    v_perm_b32 v0, v4, v0, s4
; GFX900-NEXT:    s_setpc_b64 s[30:31]
;
; GFX950-LABEL: v_maximumnum_v3bf16:
; GFX950:       ; %bb.0:
; GFX950-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX950-NEXT:    v_and_b32_e32 v4, 0xffff0000, v0
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v4, v4
; GFX950-NEXT:    v_lshrrev_b32_e32 v4, 16, v2
; GFX950-NEXT:    v_and_b32_e32 v7, 0xffff0000, v2
; GFX950-NEXT:    v_cndmask_b32_sdwa v5, v0, v4, vcc dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v7, v7
; GFX950-NEXT:    v_lshlrev_b32_e32 v6, 16, v5
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v4, v4, v5, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v7, 16, v4
; GFX950-NEXT:    v_cmp_gt_f32_e32 vcc, v6, v7
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v6, v4, v5, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; GFX950-NEXT:    v_max_f32_e32 v6, v6, v6
; GFX950-NEXT:    v_cvt_pk_bf16_f32 v6, v6, s0
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v5
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v5, v6, v5, vcc
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v4
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v4, v5, v4, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v5, 16, v6
; GFX950-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v5
; GFX950-NEXT:    v_lshlrev_b32_e32 v5, 16, v1
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v4, v6, v4, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; GFX950-NEXT:    v_lshlrev_b32_e32 v5, 16, v3
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; GFX950-NEXT:    v_lshlrev_b32_e32 v6, 16, v1
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v3, v3, v1, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v5, 16, v3
; GFX950-NEXT:    v_cmp_gt_f32_e32 vcc, v6, v5
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v5, v3, v1, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; GFX950-NEXT:    v_max_f32_e32 v5, v5, v5
; GFX950-NEXT:    v_cvt_pk_bf16_f32 v5, v5, s0
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v1
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v1, v5, v1, vcc
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v3
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v3, 16, v5
; GFX950-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v3
; GFX950-NEXT:    v_lshlrev_b32_e32 v3, 16, v0
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v1, v5, v1, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; GFX950-NEXT:    v_lshlrev_b32_e32 v3, 16, v2
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v0, v0, v2, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; GFX950-NEXT:    v_lshlrev_b32_e32 v5, 16, v0
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v2, v2, v0, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v3, 16, v2
; GFX950-NEXT:    v_cmp_gt_f32_e32 vcc, v5, v3
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v3, v2, v0, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; GFX950-NEXT:    v_max_f32_e32 v3, v3, v3
; GFX950-NEXT:    v_cvt_pk_bf16_f32 v3, v3, s0
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v0
; GFX950-NEXT:    s_mov_b32 s0, 0x5040100
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v0, v3, v0, vcc
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v2
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v0, v0, v2, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v2, 16, v3
; GFX950-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v2
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v0, v3, v0, vcc
; GFX950-NEXT:    v_perm_b32 v0, v4, v0, s0
; GFX950-NEXT:    s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_maximumnum_v3bf16:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT:    v_lshlrev_b32_e32 v5, 16, v1
; GFX10-NEXT:    v_and_b32_e32 v4, 0xffff0000, v0
; GFX10-NEXT:    v_lshlrev_b32_e32 v8, 16, v0
; GFX10-NEXT:    v_lshrrev_b32_e32 v6, 16, v2
; GFX10-NEXT:    v_and_b32_e32 v7, 0xffff0000, v2
; GFX10-NEXT:    v_cmp_u_f32_e64 s4, v5, v5
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX10-NEXT:    v_lshlrev_b32_e32 v4, 16, v3
; GFX10-NEXT:    v_cndmask_b32_e64 v1, v1, v3, s4
; GFX10-NEXT:    v_cmp_u_f32_e64 s4, v8, v8
; GFX10-NEXT:    v_lshlrev_b32_e32 v8, 16, v2
; GFX10-NEXT:    v_cndmask_b32_e64 v5, v0, v2, s4
; GFX10-NEXT:    v_cndmask_b32_sdwa v0, v0, v6, vcc_lo dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX10-NEXT:    v_lshlrev_b32_e32 v9, 16, v5
; GFX10-NEXT:    v_lshlrev_b32_e32 v10, 16, v0
; GFX10-NEXT:    v_cndmask_b32_e32 v3, v3, v1, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
; GFX10-NEXT:    v_cndmask_b32_e32 v2, v2, v5, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
; GFX10-NEXT:    v_lshlrev_b32_e32 v7, 16, v1
; GFX10-NEXT:    v_lshlrev_b32_e32 v8, 16, v2
; GFX10-NEXT:    v_cndmask_b32_e32 v4, v6, v0, vcc_lo
; GFX10-NEXT:    v_lshlrev_b32_e32 v6, 16, v3
; GFX10-NEXT:    v_lshlrev_b32_e32 v11, 16, v4
; GFX10-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v7, v6
; GFX10-NEXT:    v_cndmask_b32_e32 v6, v3, v1, vcc_lo
; GFX10-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v9, v8
; GFX10-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; GFX10-NEXT:    v_cndmask_b32_e32 v7, v2, v5, vcc_lo
; GFX10-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v10, v11
; GFX10-NEXT:    v_max_f32_e32 v6, v6, v6
; GFX10-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; GFX10-NEXT:    v_cndmask_b32_e32 v8, v4, v0, vcc_lo
; GFX10-NEXT:    v_bfe_u32 v9, v6, 16, 1
; GFX10-NEXT:    v_max_f32_e32 v7, v7, v7
; GFX10-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; GFX10-NEXT:    v_or_b32_e32 v11, 0x400000, v6
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX10-NEXT:    v_add3_u32 v9, v9, v6, 0x7fff
; GFX10-NEXT:    v_bfe_u32 v10, v7, 16, 1
; GFX10-NEXT:    v_max_f32_e32 v8, v8, v8
; GFX10-NEXT:    v_or_b32_e32 v13, 0x400000, v7
; GFX10-NEXT:    v_cndmask_b32_e32 v6, v9, v11, vcc_lo
; GFX10-NEXT:    v_add3_u32 v10, v10, v7, 0x7fff
; GFX10-NEXT:    v_bfe_u32 v12, v8, 16, 1
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
; GFX10-NEXT:    v_or_b32_e32 v14, 0x400000, v8
; GFX10-NEXT:    v_lshrrev_b32_e32 v9, 16, v6
; GFX10-NEXT:    v_and_b32_e32 v6, 0xffff0000, v6
; GFX10-NEXT:    v_add3_u32 v12, v12, v8, 0x7fff
; GFX10-NEXT:    v_cndmask_b32_e32 v7, v10, v13, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
; GFX10-NEXT:    v_lshrrev_b32_e32 v10, 16, v7
; GFX10-NEXT:    v_cndmask_b32_e32 v8, v12, v14, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v5
; GFX10-NEXT:    v_and_b32_e32 v7, 0xffff0000, v7
; GFX10-NEXT:    v_lshrrev_b32_e32 v11, 16, v8
; GFX10-NEXT:    v_cndmask_b32_e32 v5, v10, v5, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v0
; GFX10-NEXT:    v_and_b32_e32 v8, 0xffff0000, v8
; GFX10-NEXT:    v_cndmask_b32_e32 v0, v11, v0, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v2
; GFX10-NEXT:    v_cndmask_b32_e32 v2, v5, v2, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v4
; GFX10-NEXT:    v_cndmask_b32_e32 v0, v0, v4, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v1
; GFX10-NEXT:    v_cndmask_b32_e32 v1, v9, v1, vcc_lo
; GFX10-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v7
; GFX10-NEXT:    v_cndmask_b32_e32 v2, v10, v2, vcc_lo
; GFX10-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v8
; GFX10-NEXT:    v_cndmask_b32_e32 v0, v11, v0, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v3
; GFX10-NEXT:    v_perm_b32 v0, v0, v2, 0x5040100
; GFX10-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc_lo
; GFX10-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v6
; GFX10-NEXT:    v_cndmask_b32_e32 v1, v9, v1, vcc_lo
; GFX10-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-TRUE16-LABEL: v_maximumnum_v3bf16:
; GFX11-TRUE16:       ; %bb.0:
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xffff0000, v0
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xffff0000, v2
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 16, v1
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 16, v2
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 16, v3
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s1, v5, v5
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s0, v6, v6
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 16, v0
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v4.l, v0.h, v2.h, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v1.l, v1.l, v3.l, s0
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s0, v9, v9
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s2, v8, v8
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v5.l, v2.h, v4.l, s1
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, v4.l
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v3.l, v3.l, v1.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.l, v1.l
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v0.l, v0.l, v2.l, s2
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v9.l, v5.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v10.l, v3.l
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s2, 0, v5.l
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v2.l, v2.l, v0.l, s0
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, v0.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s3, 0, v3.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v11.l, v2.l
; GFX11-TRUE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v6, v9
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s4, 0, v2.l
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v6.l, v5.l, v4.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v4.l
; GFX11-TRUE16-NEXT:    v_cmp_gt_f32_e64 s1, v8, v11
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v8.l, v2.l, v0.l, s1
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_dual_max_f32 v6, v6, v6 :: v_dual_lshlrev_b32 v7, 16, v7
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s1, 0, v0.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v9, v6, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_gt_f32_e64 s0, v7, v10
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, 0x400000, v6
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s5, v6, v6
; GFX11-TRUE16-NEXT:    v_max_f32_e32 v8, v8, v8
; GFX11-TRUE16-NEXT:    v_add3_u32 v9, v9, v6, 0x7fff
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v7.l, v3.l, v1.l, s0
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s0, 0, v1.l
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v12, v8, 16, 1
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e64 v6, v9, v10, s5
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, 0x400000, v8
; GFX11-TRUE16-NEXT:    v_add3_u32 v12, v12, v8, 0x7fff
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v0.h, v6.h, v4.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_dual_max_f32 v7, v7, v7 :: v_dual_and_b32 v4, 0xffff0000, v6
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v0.h, v0.h, v5.l, s2
; GFX11-TRUE16-NEXT:    v_bfe_u32 v11, v7, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v13, 0x400000, v7
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s5, v7, v7
; GFX11-TRUE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v4
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT:    v_add3_u32 v11, v11, v7, 0x7fff
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v0.h, v6.h, v0.h, vcc_lo
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e64 v7, v11, v13, s5
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s5, v8, v8
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v1.l, v7.h, v1.l, s0
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e64 v8, v12, v14, s5
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xffff0000, v7
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v1.l, v1.l, v3.l, s3
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v0.l, v8.h, v0.l, s1
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xffff0000, v8
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_cmp_eq_f32_e64 s1, 0, v9
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v0.l, v0.l, v2.l, s4
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_cmp_eq_f32_e64 s0, 0, v10
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v1.l, v7.h, v1.l, s1
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v0.l, v8.h, v0.l, s0
; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-FAKE16-LABEL: v_maximumnum_v3bf16:
; GFX11-FAKE16:       ; %bb.0:
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xffff0000, v0
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v5, 16, v2
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v6, 16, v0
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v9, 16, v0
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v4, v6, v5, vcc_lo
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v6, 0xffff0000, v2
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v10, 16, v4
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v5, v5, v4, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v0, v0, v2 :: v_dual_lshlrev_b32 v7, 16, v1
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v6, 16, v2
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc_lo
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v8, 16, v3
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v2, v2, v0 :: v_dual_lshlrev_b32 v7, 16, v5
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v3, v3, v1 :: v_dual_lshlrev_b32 v8, 16, v0
; GFX11-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v10, v7
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v7, 16, v2
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v10, 16, v1
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v6, v5, v4 :: v_dual_lshlrev_b32 v9, 16, v3
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v8, v7
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v7, v2, v0 :: v_dual_lshlrev_b32 v6, 16, v6
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v10, v9
; GFX11-FAKE16-NEXT:    v_dual_max_f32 v6, v6, v6 :: v_dual_lshlrev_b32 v7, 16, v7
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v8, v3, v1, vcc_lo
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v9, v6, 16, 1
; GFX11-FAKE16-NEXT:    v_max_f32_e32 v7, v7, v7
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v10, 0x400000, v6
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; GFX11-FAKE16-NEXT:    v_add3_u32 v9, v9, v6, 0x7fff
; GFX11-FAKE16-NEXT:    v_bfe_u32 v11, v7, 16, 1
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v6, v9, v10, vcc_lo
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, 0x400000, v7
; GFX11-FAKE16-NEXT:    v_add3_u32 v10, v11, v7, 0x7fff
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v13, 16, v6
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v7, v10, v9 :: v_dual_and_b32 v6, 0xffff0000, v6
; GFX11-FAKE16-NEXT:    v_max_f32_e32 v8, v8, v8
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v9, 16, v7
; GFX11-FAKE16-NEXT:    v_bfe_u32 v12, v8, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v11, 0x400000, v8
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_add3_u32 v12, v12, v8, 0x7fff
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v8, v12, v11, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v4
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v10, 16, v8
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v4, v13, v4, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v5
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v4, v4, v5 :: v_dual_and_b32 v5, 0xffff0000, v7
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v0
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v9, v0, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v6
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v4, v13, v4, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v2
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v0, v2, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v1
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v1, v10, v1 :: v_dual_and_b32 v2, 0xffff0000, v8
; GFX11-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v5
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v9, v0, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v3
; GFX11-FAKE16-NEXT:    v_perm_b32 v0, v4, v0, 0x5040100
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v2
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v10, v1, vcc_lo
; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
;
; GFX12-TRUE16-LABEL: v_maximumnum_v3bf16:
; GFX12-TRUE16:       ; %bb.0:
; GFX12-TRUE16-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX12-TRUE16-NEXT:    s_wait_expcnt 0x0
; GFX12-TRUE16-NEXT:    s_wait_samplecnt 0x0
; GFX12-TRUE16-NEXT:    s_wait_bvhcnt 0x0
; GFX12-TRUE16-NEXT:    s_wait_kmcnt 0x0
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v4, 0xffff0000, v0
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v5, 0xffff0000, v2
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 16, v1
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 16, v2
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 16, v3
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s1, v5, v5
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s0, v6, v6
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 16, v0
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v4.l, v0.h, v2.h, vcc_lo
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v1.l, v1.l, v3.l, s0
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s0, v9, v9
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s2, v8, v8
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v5.l, v2.h, v4.l, s1
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v6.l, v4.l
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v3.l, v3.l, v1.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v7.l, v1.l
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v0.l, v0.l, v2.l, s2
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v9.l, v5.l
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v10.l, v3.l
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s2, 0, v5.l
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v2.l, v2.l, v0.l, s0
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v8.l, v0.l
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s3, 0, v3.l
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v11.l, v2.l
; GFX12-TRUE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v6, v9
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s4, 0, v2.l
; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v6.l, v5.l, v4.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v4.l
; GFX12-TRUE16-NEXT:    v_cmp_gt_f32_e64 s1, v8, v11
; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_2)
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v8.l, v2.l, v0.l, s1
; GFX12-TRUE16-NEXT:    v_dual_max_num_f32 v6, v6, v6 :: v_dual_lshlrev_b32 v7, 16, v7
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s1, 0, v0.l
; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; GFX12-TRUE16-NEXT:    v_bfe_u32 v9, v6, 16, 1
; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX12-TRUE16-NEXT:    v_cmp_gt_f32_e64 s0, v7, v10
; GFX12-TRUE16-NEXT:    v_or_b32_e32 v10, 0x400000, v6
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s5, v6, v6
; GFX12-TRUE16-NEXT:    v_max_num_f32_e32 v8, v8, v8
; GFX12-TRUE16-NEXT:    v_add3_u32 v9, v9, v6, 0x7fff
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v7.l, v3.l, v1.l, s0
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s0, 0, v1.l
; GFX12-TRUE16-NEXT:    v_bfe_u32 v12, v8, 16, 1
; GFX12-TRUE16-NEXT:    v_cndmask_b32_e64 v6, v9, v10, s5
; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; GFX12-TRUE16-NEXT:    v_or_b32_e32 v14, 0x400000, v8
; GFX12-TRUE16-NEXT:    v_add3_u32 v12, v12, v8, 0x7fff
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v0.h, v6.h, v4.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_dual_max_num_f32 v7, v7, v7 :: v_dual_and_b32 v4, 0xffff0000, v6
; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v0.h, v0.h, v5.l, s2
; GFX12-TRUE16-NEXT:    v_bfe_u32 v11, v7, 16, 1
; GFX12-TRUE16-NEXT:    v_or_b32_e32 v13, 0x400000, v7
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s5, v7, v7
; GFX12-TRUE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v4
; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_2)
; GFX12-TRUE16-NEXT:    v_add3_u32 v11, v11, v7, 0x7fff
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v0.h, v6.h, v0.h, vcc_lo
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b32_e64 v7, v11, v13, s5
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s5, v8, v8
; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v1.l, v7.h, v1.l, s0
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b32_e64 v8, v12, v14, s5
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v9, 0xffff0000, v7
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v1.l, v1.l, v3.l, s3
; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v0.l, v8.h, v0.l, s1
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v10, 0xffff0000, v8
; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX12-TRUE16-NEXT:    v_cmp_eq_f32_e64 s1, 0, v9
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v0.l, v0.l, v2.l, s4
; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX12-TRUE16-NEXT:    v_cmp_eq_f32_e64 s0, 0, v10
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v1.l, v7.h, v1.l, s1
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v0.l, v8.h, v0.l, s0
; GFX12-TRUE16-NEXT:    s_setpc_b64 s[30:31]
;
; GFX12-FAKE16-LABEL: v_maximumnum_v3bf16:
; GFX12-FAKE16:       ; %bb.0:
; GFX12-FAKE16-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX12-FAKE16-NEXT:    s_wait_expcnt 0x0
; GFX12-FAKE16-NEXT:    s_wait_samplecnt 0x0
; GFX12-FAKE16-NEXT:    s_wait_bvhcnt 0x0
; GFX12-FAKE16-NEXT:    s_wait_kmcnt 0x0
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v4, 0xffff0000, v0
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v5, 16, v2
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v6, 16, v0
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v9, 16, v0
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v4, v6, v5, vcc_lo
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v6, 0xffff0000, v2
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v10, 16, v4
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v5, v5, v4, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v0, v0, v2 :: v_dual_lshlrev_b32 v7, 16, v1
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v6, 16, v2
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc_lo
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v8, 16, v3
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v2, v2, v0 :: v_dual_lshlrev_b32 v7, 16, v5
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v3, v3, v1 :: v_dual_lshlrev_b32 v8, 16, v0
; GFX12-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v10, v7
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v7, 16, v2
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v10, 16, v1
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v6, v5, v4 :: v_dual_lshlrev_b32 v9, 16, v3
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX12-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v8, v7
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v7, v2, v0 :: v_dual_lshlrev_b32 v6, 16, v6
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX12-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v10, v9
; GFX12-FAKE16-NEXT:    v_dual_max_num_f32 v6, v6, v6 :: v_dual_lshlrev_b32 v7, 16, v7
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v8, v3, v1, vcc_lo
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX12-FAKE16-NEXT:    v_bfe_u32 v9, v6, 16, 1
; GFX12-FAKE16-NEXT:    v_max_num_f32_e32 v7, v7, v7
; GFX12-FAKE16-NEXT:    v_or_b32_e32 v10, 0x400000, v6
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; GFX12-FAKE16-NEXT:    v_add3_u32 v9, v9, v6, 0x7fff
; GFX12-FAKE16-NEXT:    v_bfe_u32 v11, v7, 16, 1
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v6, v9, v10, vcc_lo
; GFX12-FAKE16-NEXT:    v_or_b32_e32 v9, 0x400000, v7
; GFX12-FAKE16-NEXT:    v_add3_u32 v10, v11, v7, 0x7fff
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v13, 16, v6
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v7, v10, v9 :: v_dual_and_b32 v6, 0xffff0000, v6
; GFX12-FAKE16-NEXT:    v_max_num_f32_e32 v8, v8, v8
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v9, 16, v7
; GFX12-FAKE16-NEXT:    v_bfe_u32 v12, v8, 16, 1
; GFX12-FAKE16-NEXT:    v_or_b32_e32 v11, 0x400000, v8
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX12-FAKE16-NEXT:    v_add3_u32 v12, v12, v8, 0x7fff
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v8, v12, v11, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v4
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_4) | instid1(VALU_DEP_2)
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v10, 16, v8
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v4, v13, v4, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v5
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v4, v4, v5 :: v_dual_and_b32 v5, 0xffff0000, v7
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v0
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v9, v0, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v6
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v4, v13, v4, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v2
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v0, v2, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v1
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v1, v10, v1 :: v_dual_and_b32 v2, 0xffff0000, v8
; GFX12-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v5
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v9, v0, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v3
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_4) | instid1(VALU_DEP_2)
; GFX12-FAKE16-NEXT:    v_perm_b32 v0, v4, v0, 0x5040100
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v2
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v10, v1, vcc_lo
; GFX12-FAKE16-NEXT:    s_setpc_b64 s[30:31]
  %result = call <3 x bfloat> @llvm.maximumnum.v3bf16(<3 x bfloat> %x, <3 x bfloat> %y)
  ret <3 x bfloat> %result
}

define <3 x bfloat> @v_maximumnum_v3bf16_nnan(<3 x bfloat> %x, <3 x bfloat> %y) {
; GFX7-LABEL: v_maximumnum_v3bf16_nnan:
; GFX7:       ; %bb.0:
; GFX7-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT:    v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT:    v_mul_f32_e32 v3, 1.0, v3
; GFX7-NEXT:    v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT:    v_mul_f32_e32 v4, 1.0, v4
; GFX7-NEXT:    v_mul_f32_e32 v2, 1.0, v2
; GFX7-NEXT:    v_mul_f32_e32 v5, 1.0, v5
; GFX7-NEXT:    v_and_b32_e32 v5, 0xffff0000, v5
; GFX7-NEXT:    v_and_b32_e32 v2, 0xffff0000, v2
; GFX7-NEXT:    v_and_b32_e32 v4, 0xffff0000, v4
; GFX7-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT:    v_and_b32_e32 v3, 0xffff0000, v3
; GFX7-NEXT:    v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT:    v_max_f32_e32 v2, v2, v5
; GFX7-NEXT:    v_max_f32_e32 v1, v1, v4
; GFX7-NEXT:    v_max_f32_e32 v0, v0, v3
; GFX7-NEXT:    v_and_b32_e32 v2, 0xffff0000, v2
; GFX7-NEXT:    v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT:    s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_maximumnum_v3bf16_nnan:
; GFX8:       ; %bb.0:
; GFX8-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT:    v_lshlrev_b32_e32 v4, 16, v3
; GFX8-NEXT:    v_lshlrev_b32_e32 v5, 16, v1
; GFX8-NEXT:    v_cmp_gt_f32_e32 vcc, v5, v4
; GFX8-NEXT:    v_cndmask_b32_e32 v4, v3, v1, vcc
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v1
; GFX8-NEXT:    v_cndmask_b32_e32 v1, v4, v1, vcc
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v3
; GFX8-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v3, 16, v4
; GFX8-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v3
; GFX8-NEXT:    v_cndmask_b32_e32 v1, v4, v1, vcc
; GFX8-NEXT:    v_and_b32_e32 v3, 0xffff0000, v2
; GFX8-NEXT:    v_and_b32_e32 v4, 0xffff0000, v0
; GFX8-NEXT:    v_lshrrev_b32_e32 v5, 16, v0
; GFX8-NEXT:    v_lshrrev_b32_e32 v6, 16, v2
; GFX8-NEXT:    v_cmp_gt_f32_e32 vcc, v4, v3
; GFX8-NEXT:    v_cndmask_b32_e32 v3, v6, v5, vcc
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v5
; GFX8-NEXT:    v_cndmask_b32_e32 v4, v3, v5, vcc
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v6
; GFX8-NEXT:    v_lshlrev_b32_e32 v5, 16, v3
; GFX8-NEXT:    v_cndmask_b32_e32 v4, v4, v6, vcc
; GFX8-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v5
; GFX8-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v4, 16, v2
; GFX8-NEXT:    v_lshlrev_b32_e32 v5, 16, v0
; GFX8-NEXT:    v_cmp_gt_f32_e32 vcc, v5, v4
; GFX8-NEXT:    v_cndmask_b32_e32 v4, v2, v0, vcc
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v0
; GFX8-NEXT:    v_cndmask_b32_e32 v0, v4, v0, vcc
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v2
; GFX8-NEXT:    v_cndmask_b32_e32 v0, v0, v2, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v2, 16, v4
; GFX8-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v2
; GFX8-NEXT:    v_cndmask_b32_e32 v0, v4, v0, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v2, 16, v3
; GFX8-NEXT:    v_or_b32_sdwa v0, v0, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX8-NEXT:    s_setpc_b64 s[30:31]
;
; GFX900-LABEL: v_maximumnum_v3bf16_nnan:
; GFX900:       ; %bb.0:
; GFX900-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX900-NEXT:    v_lshlrev_b32_e32 v4, 16, v3
; GFX900-NEXT:    v_lshlrev_b32_e32 v5, 16, v1
; GFX900-NEXT:    v_cmp_gt_f32_e32 vcc, v5, v4
; GFX900-NEXT:    v_cndmask_b32_e32 v4, v3, v1, vcc
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v1
; GFX900-NEXT:    v_cndmask_b32_e32 v1, v4, v1, vcc
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v3
; GFX900-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v3, 16, v4
; GFX900-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v3
; GFX900-NEXT:    v_cndmask_b32_e32 v1, v4, v1, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v3, 16, v2
; GFX900-NEXT:    v_lshlrev_b32_e32 v4, 16, v0
; GFX900-NEXT:    v_cmp_gt_f32_e32 vcc, v4, v3
; GFX900-NEXT:    v_cndmask_b32_e32 v3, v2, v0, vcc
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v0
; GFX900-NEXT:    v_cndmask_b32_e32 v4, v3, v0, vcc
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v2
; GFX900-NEXT:    v_lshlrev_b32_e32 v5, 16, v3
; GFX900-NEXT:    v_cndmask_b32_e32 v4, v4, v2, vcc
; GFX900-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v5
; GFX900-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc
; GFX900-NEXT:    v_and_b32_e32 v4, 0xffff0000, v2
; GFX900-NEXT:    v_and_b32_e32 v5, 0xffff0000, v0
; GFX900-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; GFX900-NEXT:    v_lshrrev_b32_e32 v2, 16, v2
; GFX900-NEXT:    v_cmp_gt_f32_e32 vcc, v5, v4
; GFX900-NEXT:    v_cndmask_b32_e32 v4, v2, v0, vcc
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v0
; GFX900-NEXT:    v_cndmask_b32_e32 v0, v4, v0, vcc
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v2
; GFX900-NEXT:    v_cndmask_b32_e32 v0, v0, v2, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v2, 16, v4
; GFX900-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v2
; GFX900-NEXT:    v_cndmask_b32_e32 v0, v4, v0, vcc
; GFX900-NEXT:    s_mov_b32 s4, 0x5040100
; GFX900-NEXT:    v_perm_b32 v0, v0, v3, s4
; GFX900-NEXT:    s_setpc_b64 s[30:31]
;
; GFX950-LABEL: v_maximumnum_v3bf16_nnan:
; GFX950:       ; %bb.0:
; GFX950-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX950-NEXT:    v_lshlrev_b32_e32 v4, 16, v3
; GFX950-NEXT:    v_lshlrev_b32_e32 v5, 16, v1
; GFX950-NEXT:    v_cmp_gt_f32_e32 vcc, v5, v4
; GFX950-NEXT:    s_mov_b32 s0, 0x5040100
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v4, v3, v1, vcc
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v1
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v1, v4, v1, vcc
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v3
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v3, 16, v4
; GFX950-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v3
; GFX950-NEXT:    v_lshlrev_b32_e32 v3, 16, v2
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v1, v4, v1, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v4, 16, v0
; GFX950-NEXT:    v_cmp_gt_f32_e32 vcc, v4, v3
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v3, v2, v0, vcc
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v0
; GFX950-NEXT:    v_lshlrev_b32_e32 v5, 16, v3
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v4, v3, v0, vcc
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v2
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v4, v4, v2, vcc
; GFX950-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v5
; GFX950-NEXT:    v_and_b32_e32 v5, 0xffff0000, v0
; GFX950-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; GFX950-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc
; GFX950-NEXT:    v_and_b32_e32 v4, 0xffff0000, v2
; GFX950-NEXT:    v_lshrrev_b32_e32 v2, 16, v2
; GFX950-NEXT:    v_cmp_gt_f32_e32 vcc, v5, v4
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v4, v2, v0, vcc
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v0
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v0, v4, v0, vcc
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v2
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v0, v0, v2, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v2, 16, v4
; GFX950-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v2
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v0, v4, v0, vcc
; GFX950-NEXT:    v_perm_b32 v0, v0, v3, s0
; GFX950-NEXT:    s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_maximumnum_v3bf16_nnan:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT:    v_lshlrev_b32_e32 v6, 16, v2
; GFX10-NEXT:    v_lshlrev_b32_e32 v7, 16, v0
; GFX10-NEXT:    v_lshlrev_b32_e32 v4, 16, v3
; GFX10-NEXT:    v_lshlrev_b32_e32 v5, 16, v1
; GFX10-NEXT:    v_and_b32_e32 v8, 0xffff0000, v2
; GFX10-NEXT:    v_and_b32_e32 v9, 0xffff0000, v0
; GFX10-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v7, v6
; GFX10-NEXT:    v_lshrrev_b32_e32 v10, 16, v0
; GFX10-NEXT:    v_lshrrev_b32_e32 v11, 16, v2
; GFX10-NEXT:    v_cndmask_b32_e32 v6, v2, v0, vcc_lo
; GFX10-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v5, v4
; GFX10-NEXT:    v_cndmask_b32_e32 v4, v3, v1, vcc_lo
; GFX10-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v9, v8
; GFX10-NEXT:    v_lshlrev_b32_e32 v9, 16, v4
; GFX10-NEXT:    v_cndmask_b32_e32 v5, v11, v10, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v0
; GFX10-NEXT:    v_lshlrev_b32_e32 v8, 16, v5
; GFX10-NEXT:    v_cndmask_b32_e32 v0, v6, v0, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v1
; GFX10-NEXT:    v_cndmask_b32_e32 v1, v4, v1, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v10
; GFX10-NEXT:    v_cndmask_b32_e32 v7, v5, v10, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v2
; GFX10-NEXT:    v_cndmask_b32_e32 v0, v0, v2, vcc_lo
; GFX10-NEXT:    v_lshlrev_b32_e32 v2, 16, v6
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v11
; GFX10-NEXT:    v_cndmask_b32_e32 v7, v7, v11, vcc_lo
; GFX10-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v2
; GFX10-NEXT:    v_cndmask_b32_e32 v0, v6, v0, vcc_lo
; GFX10-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v8
; GFX10-NEXT:    v_cndmask_b32_e32 v2, v5, v7, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v3
; GFX10-NEXT:    v_perm_b32 v0, v2, v0, 0x5040100
; GFX10-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc_lo
; GFX10-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v9
; GFX10-NEXT:    v_cndmask_b32_e32 v1, v4, v1, vcc_lo
; GFX10-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-TRUE16-LABEL: v_maximumnum_v3bf16_nnan:
; GFX11-TRUE16:       ; %bb.0:
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v4, 16, v3
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 16, v1
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 16, v2
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 16, v0
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xffff0000, v2
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xffff0000, v0
; GFX11-TRUE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v5, v4
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s0, 0, v1.l
; GFX11-TRUE16-NEXT:    v_cmp_gt_f32_e64 s1, v7, v6
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s2, 0, v3.l
; GFX11-TRUE16-NEXT:    v_cmp_gt_f32_e64 s3, v9, v8
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v4.l, v3.l, v1.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v0.l
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v5.l, v2.l, v0.l, s1
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s1, 0, v0.h
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v6.l, v2.h, v0.h, s3
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v1.l, v4.l, v1.l, s0
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.l, v4.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, v5.l
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v0.l, v5.l, v0.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v9.l, v6.l
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v1.l, v1.l, v3.l, s2
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v2.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v8
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v0.h, v6.l, v0.h, s1
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s0, 0, v2.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 16, v9
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v0.l, v0.l, v2.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v3
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v0.h, v0.h, v2.h, s0
; GFX11-TRUE16-NEXT:    v_cmp_eq_f32_e64 s0, 0, v8
; GFX11-TRUE16-NEXT:    v_cmp_eq_f32_e64 s1, 0, v7
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v0.l, v5.l, v0.l, vcc_lo
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v0.h, v6.l, v0.h, s0
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v1.l, v4.l, v1.l, s1
; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-FAKE16-LABEL: v_maximumnum_v3bf16_nnan:
; GFX11-FAKE16:       ; %bb.0:
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v6, 16, v2
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v7, 16, v0
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v10, 16, v0
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v11, 16, v2
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v4, 16, v3
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v8, 0xffff0000, v2
; GFX11-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v7, v6
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v5, 16, v1
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v9, 0xffff0000, v0
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v6, v2, v0, vcc_lo
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v5, v4
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v4, v3, v1, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v9, v8
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v5, v11, v10, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v0
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v8, 16, v5
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v6, v0, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v4, v1, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v10
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v7, v5, v10, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v2
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v0, v2, vcc_lo
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v6
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v11
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v7, v7, v11, vcc_lo
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v2
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v6, v0, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v8
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v2, v5, v7 :: v_dual_lshlrev_b32 v9, 16, v4
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v3
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc_lo
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v9
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v4, v1, vcc_lo
; GFX11-FAKE16-NEXT:    v_perm_b32 v0, v2, v0, 0x5040100
; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
;
; GFX12-TRUE16-LABEL: v_maximumnum_v3bf16_nnan:
; GFX12-TRUE16:       ; %bb.0:
; GFX12-TRUE16-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX12-TRUE16-NEXT:    s_wait_expcnt 0x0
; GFX12-TRUE16-NEXT:    s_wait_samplecnt 0x0
; GFX12-TRUE16-NEXT:    s_wait_bvhcnt 0x0
; GFX12-TRUE16-NEXT:    s_wait_kmcnt 0x0
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v4, 16, v3
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 16, v1
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 16, v2
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 16, v0
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v8, 0xffff0000, v2
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v9, 0xffff0000, v0
; GFX12-TRUE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v5, v4
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s0, 0, v1.l
; GFX12-TRUE16-NEXT:    v_cmp_gt_f32_e64 s1, v7, v6
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s2, 0, v3.l
; GFX12-TRUE16-NEXT:    v_cmp_gt_f32_e64 s3, v9, v8
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v4.l, v3.l, v1.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v0.l
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v5.l, v2.l, v0.l, s1
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s1, 0, v0.h
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v6.l, v2.h, v0.h, s3
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v1.l, v4.l, v1.l, s0
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v7.l, v4.l
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v8.l, v5.l
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v0.l, v5.l, v0.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v9.l, v6.l
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v1.l, v1.l, v3.l, s2
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v2.l
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v8
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v0.h, v6.l, v0.h, s1
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s0, 0, v2.h
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 16, v9
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v0.l, v0.l, v2.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v3
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v0.h, v0.h, v2.h, s0
; GFX12-TRUE16-NEXT:    v_cmp_eq_f32_e64 s0, 0, v8
; GFX12-TRUE16-NEXT:    v_cmp_eq_f32_e64 s1, 0, v7
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v0.l, v5.l, v0.l, vcc_lo
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v0.h, v6.l, v0.h, s0
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v1.l, v4.l, v1.l, s1
; GFX12-TRUE16-NEXT:    s_setpc_b64 s[30:31]
;
; GFX12-FAKE16-LABEL: v_maximumnum_v3bf16_nnan:
; GFX12-FAKE16:       ; %bb.0:
; GFX12-FAKE16-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX12-FAKE16-NEXT:    s_wait_expcnt 0x0
; GFX12-FAKE16-NEXT:    s_wait_samplecnt 0x0
; GFX12-FAKE16-NEXT:    s_wait_bvhcnt 0x0
; GFX12-FAKE16-NEXT:    s_wait_kmcnt 0x0
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v6, 16, v2
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v7, 16, v0
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v10, 16, v0
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v11, 16, v2
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v4, 16, v3
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v8, 0xffff0000, v2
; GFX12-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v7, v6
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v5, 16, v1
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v9, 0xffff0000, v0
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v6, v2, v0, vcc_lo
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX12-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v5, v4
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v4, v3, v1, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v9, v8
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v5, v11, v10, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v0
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v8, 16, v5
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v6, v0, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v1
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v4, v1, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v10
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v7, v5, v10, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v2
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v0, v2, vcc_lo
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v6
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v11
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v7, v7, v11, vcc_lo
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX12-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v2
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v6, v0, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v8
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v2, v5, v7 :: v_dual_lshlrev_b32 v9, 16, v4
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v3
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc_lo
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX12-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v9
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v4, v1, vcc_lo
; GFX12-FAKE16-NEXT:    v_perm_b32 v0, v2, v0, 0x5040100
; GFX12-FAKE16-NEXT:    s_setpc_b64 s[30:31]
  %result = call nnan <3 x bfloat> @llvm.maximumnum.v3bf16(<3 x bfloat> %x, <3 x bfloat> %y)
  ret <3 x bfloat> %result
}

define <4 x bfloat> @v_maximumnum_v4bf16(<4 x bfloat> %x, <4 x bfloat> %y) {
; GFX7-LABEL: v_maximumnum_v4bf16:
; GFX7:       ; %bb.0:
; GFX7-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT:    v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT:    v_mul_f32_e32 v4, 1.0, v4
; GFX7-NEXT:    v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT:    v_mul_f32_e32 v5, 1.0, v5
; GFX7-NEXT:    v_mul_f32_e32 v2, 1.0, v2
; GFX7-NEXT:    v_mul_f32_e32 v6, 1.0, v6
; GFX7-NEXT:    v_mul_f32_e32 v3, 1.0, v3
; GFX7-NEXT:    v_mul_f32_e32 v7, 1.0, v7
; GFX7-NEXT:    v_and_b32_e32 v7, 0xffff0000, v7
; GFX7-NEXT:    v_and_b32_e32 v3, 0xffff0000, v3
; GFX7-NEXT:    v_and_b32_e32 v6, 0xffff0000, v6
; GFX7-NEXT:    v_and_b32_e32 v2, 0xffff0000, v2
; GFX7-NEXT:    v_and_b32_e32 v5, 0xffff0000, v5
; GFX7-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT:    v_and_b32_e32 v4, 0xffff0000, v4
; GFX7-NEXT:    v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT:    v_max_f32_e32 v3, v3, v7
; GFX7-NEXT:    v_max_f32_e32 v2, v2, v6
; GFX7-NEXT:    v_max_f32_e32 v1, v1, v5
; GFX7-NEXT:    v_max_f32_e32 v0, v0, v4
; GFX7-NEXT:    v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT:    v_and_b32_e32 v2, 0xffff0000, v2
; GFX7-NEXT:    v_and_b32_e32 v3, 0xffff0000, v3
; GFX7-NEXT:    s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_maximumnum_v4bf16:
; GFX8:       ; %bb.0:
; GFX8-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT:    v_and_b32_e32 v4, 0xffff0000, v1
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v4, v4
; GFX8-NEXT:    v_lshrrev_b32_e32 v4, 16, v3
; GFX8-NEXT:    v_and_b32_e32 v7, 0xffff0000, v3
; GFX8-NEXT:    v_cndmask_b32_sdwa v5, v1, v4, vcc dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v7, v7
; GFX8-NEXT:    v_cndmask_b32_e32 v4, v4, v5, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v6, 16, v5
; GFX8-NEXT:    v_lshlrev_b32_e32 v7, 16, v4
; GFX8-NEXT:    v_cmp_gt_f32_e32 vcc, v6, v7
; GFX8-NEXT:    v_cndmask_b32_e32 v6, v4, v5, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; GFX8-NEXT:    v_mul_f32_e32 v6, 1.0, v6
; GFX8-NEXT:    v_bfe_u32 v7, v6, 16, 1
; GFX8-NEXT:    v_add_u32_e32 v7, vcc, v7, v6
; GFX8-NEXT:    s_movk_i32 s4, 0x7fff
; GFX8-NEXT:    v_add_u32_e32 v7, vcc, s4, v7
; GFX8-NEXT:    v_or_b32_e32 v8, 0x400000, v6
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v6, v6
; GFX8-NEXT:    v_cndmask_b32_e32 v6, v7, v8, vcc
; GFX8-NEXT:    v_lshrrev_b32_e32 v7, 16, v6
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v5
; GFX8-NEXT:    v_cndmask_b32_e32 v5, v7, v5, vcc
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v4
; GFX8-NEXT:    v_cndmask_b32_e32 v4, v5, v4, vcc
; GFX8-NEXT:    v_and_b32_e32 v5, 0xffff0000, v6
; GFX8-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v5
; GFX8-NEXT:    v_and_b32_e32 v5, 0xffff0000, v0
; GFX8-NEXT:    v_cndmask_b32_e32 v4, v7, v4, vcc
; GFX8-NEXT:    v_lshrrev_b32_e32 v6, 16, v2
; GFX8-NEXT:    v_lshrrev_b32_e32 v7, 16, v0
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; GFX8-NEXT:    v_and_b32_e32 v8, 0xffff0000, v2
; GFX8-NEXT:    v_cndmask_b32_e32 v5, v7, v6, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v8, v8
; GFX8-NEXT:    v_cndmask_b32_e32 v6, v6, v5, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v7, 16, v5
; GFX8-NEXT:    v_lshlrev_b32_e32 v8, 16, v6
; GFX8-NEXT:    v_cmp_gt_f32_e32 vcc, v7, v8
; GFX8-NEXT:    v_cndmask_b32_e32 v7, v6, v5, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; GFX8-NEXT:    v_mul_f32_e32 v7, 1.0, v7
; GFX8-NEXT:    v_bfe_u32 v8, v7, 16, 1
; GFX8-NEXT:    v_add_u32_e32 v8, vcc, v8, v7
; GFX8-NEXT:    v_add_u32_e32 v8, vcc, s4, v8
; GFX8-NEXT:    v_or_b32_e32 v9, 0x400000, v7
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v7, v7
; GFX8-NEXT:    v_cndmask_b32_e32 v7, v8, v9, vcc
; GFX8-NEXT:    v_lshrrev_b32_e32 v8, 16, v7
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v5
; GFX8-NEXT:    v_cndmask_b32_e32 v5, v8, v5, vcc
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v6
; GFX8-NEXT:    v_cndmask_b32_e32 v5, v5, v6, vcc
; GFX8-NEXT:    v_and_b32_e32 v6, 0xffff0000, v7
; GFX8-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v6
; GFX8-NEXT:    v_lshlrev_b32_e32 v6, 16, v1
; GFX8-NEXT:    v_cndmask_b32_e32 v5, v8, v5, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v6, v6
; GFX8-NEXT:    v_lshlrev_b32_e32 v6, 16, v3
; GFX8-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v6, v6
; GFX8-NEXT:    v_cndmask_b32_e32 v3, v3, v1, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v6, 16, v3
; GFX8-NEXT:    v_lshlrev_b32_e32 v7, 16, v1
; GFX8-NEXT:    v_cmp_gt_f32_e32 vcc, v7, v6
; GFX8-NEXT:    v_cndmask_b32_e32 v6, v3, v1, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; GFX8-NEXT:    v_mul_f32_e32 v6, 1.0, v6
; GFX8-NEXT:    v_bfe_u32 v7, v6, 16, 1
; GFX8-NEXT:    v_add_u32_e32 v7, vcc, v7, v6
; GFX8-NEXT:    v_add_u32_e32 v7, vcc, s4, v7
; GFX8-NEXT:    v_or_b32_e32 v8, 0x400000, v6
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v6, v6
; GFX8-NEXT:    v_cndmask_b32_e32 v6, v7, v8, vcc
; GFX8-NEXT:    v_lshrrev_b32_e32 v7, 16, v6
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v1
; GFX8-NEXT:    v_cndmask_b32_e32 v1, v7, v1, vcc
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v3
; GFX8-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc
; GFX8-NEXT:    v_and_b32_e32 v3, 0xffff0000, v6
; GFX8-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v3
; GFX8-NEXT:    v_lshlrev_b32_e32 v3, 16, v0
; GFX8-NEXT:    v_cndmask_b32_e32 v1, v7, v1, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; GFX8-NEXT:    v_lshlrev_b32_e32 v3, 16, v2
; GFX8-NEXT:    v_cndmask_b32_e32 v0, v0, v2, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; GFX8-NEXT:    v_cndmask_b32_e32 v2, v2, v0, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v3, 16, v2
; GFX8-NEXT:    v_lshlrev_b32_e32 v6, 16, v0
; GFX8-NEXT:    v_cmp_gt_f32_e32 vcc, v6, v3
; GFX8-NEXT:    v_cndmask_b32_e32 v3, v2, v0, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; GFX8-NEXT:    v_mul_f32_e32 v3, 1.0, v3
; GFX8-NEXT:    v_bfe_u32 v6, v3, 16, 1
; GFX8-NEXT:    v_add_u32_e32 v6, vcc, v6, v3
; GFX8-NEXT:    v_add_u32_e32 v6, vcc, s4, v6
; GFX8-NEXT:    v_or_b32_e32 v7, 0x400000, v3
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; GFX8-NEXT:    v_cndmask_b32_e32 v3, v6, v7, vcc
; GFX8-NEXT:    v_lshrrev_b32_e32 v6, 16, v3
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v0
; GFX8-NEXT:    v_cndmask_b32_e32 v0, v6, v0, vcc
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v2
; GFX8-NEXT:    v_cndmask_b32_e32 v0, v0, v2, vcc
; GFX8-NEXT:    v_and_b32_e32 v2, 0xffff0000, v3
; GFX8-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v2
; GFX8-NEXT:    v_cndmask_b32_e32 v0, v6, v0, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v2, 16, v5
; GFX8-NEXT:    v_or_b32_sdwa v0, v0, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX8-NEXT:    v_lshlrev_b32_e32 v2, 16, v4
; GFX8-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX8-NEXT:    s_setpc_b64 s[30:31]
;
; GFX900-LABEL: v_maximumnum_v4bf16:
; GFX900:       ; %bb.0:
; GFX900-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX900-NEXT:    v_and_b32_e32 v4, 0xffff0000, v1
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v4, v4
; GFX900-NEXT:    v_lshrrev_b32_e32 v4, 16, v3
; GFX900-NEXT:    v_and_b32_e32 v7, 0xffff0000, v3
; GFX900-NEXT:    v_cndmask_b32_sdwa v5, v1, v4, vcc dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v7, v7
; GFX900-NEXT:    v_cndmask_b32_e32 v4, v4, v5, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v6, 16, v5
; GFX900-NEXT:    v_lshlrev_b32_e32 v7, 16, v4
; GFX900-NEXT:    v_cmp_gt_f32_e32 vcc, v6, v7
; GFX900-NEXT:    v_cndmask_b32_e32 v6, v4, v5, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; GFX900-NEXT:    v_max_f32_e32 v6, v6, v6
; GFX900-NEXT:    v_bfe_u32 v7, v6, 16, 1
; GFX900-NEXT:    s_movk_i32 s4, 0x7fff
; GFX900-NEXT:    v_add3_u32 v7, v7, v6, s4
; GFX900-NEXT:    v_or_b32_e32 v8, 0x400000, v6
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v6, v6
; GFX900-NEXT:    v_cndmask_b32_e32 v6, v7, v8, vcc
; GFX900-NEXT:    v_lshrrev_b32_e32 v7, 16, v6
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v5
; GFX900-NEXT:    v_cndmask_b32_e32 v5, v7, v5, vcc
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v4
; GFX900-NEXT:    v_cndmask_b32_e32 v4, v5, v4, vcc
; GFX900-NEXT:    v_and_b32_e32 v5, 0xffff0000, v6
; GFX900-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v5
; GFX900-NEXT:    v_and_b32_e32 v5, 0xffff0000, v0
; GFX900-NEXT:    v_cndmask_b32_e32 v4, v7, v4, vcc
; GFX900-NEXT:    v_lshrrev_b32_e32 v6, 16, v2
; GFX900-NEXT:    v_lshrrev_b32_e32 v7, 16, v0
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; GFX900-NEXT:    v_and_b32_e32 v8, 0xffff0000, v2
; GFX900-NEXT:    v_cndmask_b32_e32 v5, v7, v6, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v8, v8
; GFX900-NEXT:    v_cndmask_b32_e32 v6, v6, v5, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v7, 16, v5
; GFX900-NEXT:    v_lshlrev_b32_e32 v8, 16, v6
; GFX900-NEXT:    v_cmp_gt_f32_e32 vcc, v7, v8
; GFX900-NEXT:    v_cndmask_b32_e32 v7, v6, v5, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; GFX900-NEXT:    v_max_f32_e32 v7, v7, v7
; GFX900-NEXT:    v_bfe_u32 v8, v7, 16, 1
; GFX900-NEXT:    v_add3_u32 v8, v8, v7, s4
; GFX900-NEXT:    v_or_b32_e32 v9, 0x400000, v7
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v7, v7
; GFX900-NEXT:    v_cndmask_b32_e32 v7, v8, v9, vcc
; GFX900-NEXT:    v_lshrrev_b32_e32 v8, 16, v7
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v5
; GFX900-NEXT:    v_cndmask_b32_e32 v5, v8, v5, vcc
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v6
; GFX900-NEXT:    v_cndmask_b32_e32 v5, v5, v6, vcc
; GFX900-NEXT:    v_and_b32_e32 v6, 0xffff0000, v7
; GFX900-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v6
; GFX900-NEXT:    v_lshlrev_b32_e32 v6, 16, v1
; GFX900-NEXT:    v_cndmask_b32_e32 v5, v8, v5, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v6, v6
; GFX900-NEXT:    v_lshlrev_b32_e32 v6, 16, v3
; GFX900-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v6, v6
; GFX900-NEXT:    v_cndmask_b32_e32 v3, v3, v1, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v6, 16, v3
; GFX900-NEXT:    v_lshlrev_b32_e32 v7, 16, v1
; GFX900-NEXT:    v_cmp_gt_f32_e32 vcc, v7, v6
; GFX900-NEXT:    v_cndmask_b32_e32 v6, v3, v1, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; GFX900-NEXT:    v_max_f32_e32 v6, v6, v6
; GFX900-NEXT:    v_bfe_u32 v7, v6, 16, 1
; GFX900-NEXT:    v_add3_u32 v7, v7, v6, s4
; GFX900-NEXT:    v_or_b32_e32 v8, 0x400000, v6
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v6, v6
; GFX900-NEXT:    v_cndmask_b32_e32 v6, v7, v8, vcc
; GFX900-NEXT:    v_lshrrev_b32_e32 v7, 16, v6
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v1
; GFX900-NEXT:    v_cndmask_b32_e32 v1, v7, v1, vcc
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v3
; GFX900-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc
; GFX900-NEXT:    v_and_b32_e32 v3, 0xffff0000, v6
; GFX900-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v3
; GFX900-NEXT:    v_lshlrev_b32_e32 v3, 16, v0
; GFX900-NEXT:    v_cndmask_b32_e32 v1, v7, v1, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; GFX900-NEXT:    v_lshlrev_b32_e32 v3, 16, v2
; GFX900-NEXT:    v_cndmask_b32_e32 v0, v0, v2, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; GFX900-NEXT:    v_cndmask_b32_e32 v2, v2, v0, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v3, 16, v2
; GFX900-NEXT:    v_lshlrev_b32_e32 v6, 16, v0
; GFX900-NEXT:    v_cmp_gt_f32_e32 vcc, v6, v3
; GFX900-NEXT:    v_cndmask_b32_e32 v3, v2, v0, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; GFX900-NEXT:    v_max_f32_e32 v3, v3, v3
; GFX900-NEXT:    v_bfe_u32 v6, v3, 16, 1
; GFX900-NEXT:    v_add3_u32 v6, v6, v3, s4
; GFX900-NEXT:    v_or_b32_e32 v7, 0x400000, v3
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; GFX900-NEXT:    v_cndmask_b32_e32 v3, v6, v7, vcc
; GFX900-NEXT:    v_lshrrev_b32_e32 v6, 16, v3
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v0
; GFX900-NEXT:    v_cndmask_b32_e32 v0, v6, v0, vcc
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v2
; GFX900-NEXT:    v_cndmask_b32_e32 v0, v0, v2, vcc
; GFX900-NEXT:    v_and_b32_e32 v2, 0xffff0000, v3
; GFX900-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v2
; GFX900-NEXT:    v_cndmask_b32_e32 v0, v6, v0, vcc
; GFX900-NEXT:    s_mov_b32 s4, 0x5040100
; GFX900-NEXT:    v_perm_b32 v0, v5, v0, s4
; GFX900-NEXT:    v_perm_b32 v1, v4, v1, s4
; GFX900-NEXT:    s_setpc_b64 s[30:31]
;
; GFX950-LABEL: v_maximumnum_v4bf16:
; GFX950:       ; %bb.0:
; GFX950-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX950-NEXT:    v_and_b32_e32 v4, 0xffff0000, v1
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v4, v4
; GFX950-NEXT:    v_lshrrev_b32_e32 v4, 16, v3
; GFX950-NEXT:    v_and_b32_e32 v7, 0xffff0000, v3
; GFX950-NEXT:    v_cndmask_b32_sdwa v5, v1, v4, vcc dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v7, v7
; GFX950-NEXT:    v_lshlrev_b32_e32 v6, 16, v5
; GFX950-NEXT:    v_and_b32_e32 v8, 0xffff0000, v2
; GFX950-NEXT:    v_cndmask_b32_e32 v4, v4, v5, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v7, 16, v4
; GFX950-NEXT:    v_cmp_gt_f32_e32 vcc, v6, v7
; GFX950-NEXT:    v_lshrrev_b32_e32 v7, 16, v0
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v6, v4, v5, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; GFX950-NEXT:    v_max_f32_e32 v6, v6, v6
; GFX950-NEXT:    v_cvt_pk_bf16_f32 v6, v6, s0
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v5
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v5, v6, v5, vcc
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v4
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v4, v5, v4, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v5, 16, v6
; GFX950-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v5
; GFX950-NEXT:    v_and_b32_e32 v5, 0xffff0000, v0
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v4, v6, v4, vcc
; GFX950-NEXT:    v_lshrrev_b32_e32 v6, 16, v2
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v5, v7, v6, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v8, v8
; GFX950-NEXT:    v_lshlrev_b32_e32 v7, 16, v5
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v6, v6, v5, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v8, 16, v6
; GFX950-NEXT:    v_cmp_gt_f32_e32 vcc, v7, v8
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v7, v6, v5, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; GFX950-NEXT:    v_max_f32_e32 v7, v7, v7
; GFX950-NEXT:    v_cvt_pk_bf16_f32 v7, v7, s0
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v5
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v5, v7, v5, vcc
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v6
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v5, v5, v6, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v6, 16, v7
; GFX950-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v6
; GFX950-NEXT:    v_lshlrev_b32_e32 v6, 16, v1
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v5, v7, v5, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v6, v6
; GFX950-NEXT:    v_lshlrev_b32_e32 v6, 16, v3
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v6, v6
; GFX950-NEXT:    v_lshlrev_b32_e32 v7, 16, v1
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v3, v3, v1, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v6, 16, v3
; GFX950-NEXT:    v_cmp_gt_f32_e32 vcc, v7, v6
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v6, v3, v1, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; GFX950-NEXT:    v_max_f32_e32 v6, v6, v6
; GFX950-NEXT:    v_cvt_pk_bf16_f32 v6, v6, s0
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v1
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v1, v6, v1, vcc
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v3
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v3, 16, v6
; GFX950-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v3
; GFX950-NEXT:    v_lshlrev_b32_e32 v3, 16, v0
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v1, v6, v1, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; GFX950-NEXT:    v_lshlrev_b32_e32 v3, 16, v2
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v0, v0, v2, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; GFX950-NEXT:    v_lshlrev_b32_e32 v6, 16, v0
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v2, v2, v0, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v3, 16, v2
; GFX950-NEXT:    v_cmp_gt_f32_e32 vcc, v6, v3
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v3, v2, v0, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; GFX950-NEXT:    v_max_f32_e32 v3, v3, v3
; GFX950-NEXT:    v_cvt_pk_bf16_f32 v3, v3, s0
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v0
; GFX950-NEXT:    s_mov_b32 s0, 0x5040100
; GFX950-NEXT:    v_perm_b32 v1, v4, v1, s0
; GFX950-NEXT:    v_cndmask_b32_e32 v0, v3, v0, vcc
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v2
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v0, v0, v2, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v2, 16, v3
; GFX950-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v2
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v0, v3, v0, vcc
; GFX950-NEXT:    v_perm_b32 v0, v5, v0, s0
; GFX950-NEXT:    s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_maximumnum_v4bf16:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT:    v_and_b32_e32 v4, 0xffff0000, v1
; GFX10-NEXT:    v_lshrrev_b32_e32 v5, 16, v3
; GFX10-NEXT:    v_and_b32_e32 v6, 0xffff0000, v3
; GFX10-NEXT:    v_lshrrev_b32_e32 v7, 16, v2
; GFX10-NEXT:    v_lshrrev_b32_e32 v8, 16, v0
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX10-NEXT:    v_and_b32_e32 v4, 0xffff0000, v0
; GFX10-NEXT:    v_lshlrev_b32_e32 v11, 16, v1
; GFX10-NEXT:    v_lshlrev_b32_e32 v13, 16, v0
; GFX10-NEXT:    v_and_b32_e32 v9, 0xffff0000, v2
; GFX10-NEXT:    v_cndmask_b32_sdwa v10, v1, v5, vcc_lo dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX10-NEXT:    v_lshlrev_b32_e32 v12, 16, v3
; GFX10-NEXT:    v_lshlrev_b32_e32 v6, 16, v2
; GFX10-NEXT:    v_cndmask_b32_e32 v5, v5, v10, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX10-NEXT:    v_lshlrev_b32_e32 v14, 16, v5
; GFX10-NEXT:    v_cndmask_b32_e32 v4, v8, v7, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v11, v11
; GFX10-NEXT:    v_lshlrev_b32_e32 v8, 16, v10
; GFX10-NEXT:    v_lshlrev_b32_e32 v11, 16, v4
; GFX10-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v13, v13
; GFX10-NEXT:    v_cndmask_b32_e32 v0, v0, v2, vcc_lo
; GFX10-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v8, v14
; GFX10-NEXT:    v_cndmask_b32_e32 v8, v5, v10, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
; GFX10-NEXT:    v_lshlrev_b32_e32 v9, 16, v1
; GFX10-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; GFX10-NEXT:    v_cndmask_b32_e32 v7, v7, v4, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v12, v12
; GFX10-NEXT:    v_max_f32_e32 v8, v8, v8
; GFX10-NEXT:    v_lshlrev_b32_e32 v12, 16, v7
; GFX10-NEXT:    v_cndmask_b32_e32 v3, v3, v1, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX10-NEXT:    v_lshlrev_b32_e32 v6, 16, v0
; GFX10-NEXT:    v_bfe_u32 v14, v8, 16, 1
; GFX10-NEXT:    v_lshlrev_b32_e32 v13, 16, v3
; GFX10-NEXT:    v_cndmask_b32_e32 v2, v2, v0, vcc_lo
; GFX10-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v11, v12
; GFX10-NEXT:    v_lshlrev_b32_e32 v12, 16, v2
; GFX10-NEXT:    v_cndmask_b32_e32 v11, v7, v4, vcc_lo
; GFX10-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v9, v13
; GFX10-NEXT:    v_or_b32_e32 v13, 0x400000, v8
; GFX10-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
; GFX10-NEXT:    v_cndmask_b32_e32 v9, v3, v1, vcc_lo
; GFX10-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v6, v12
; GFX10-NEXT:    v_add3_u32 v12, v14, v8, 0x7fff
; GFX10-NEXT:    v_max_f32_e32 v11, v11, v11
; GFX10-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; GFX10-NEXT:    v_cndmask_b32_e32 v6, v2, v0, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
; GFX10-NEXT:    v_or_b32_e32 v14, 0x400000, v11
; GFX10-NEXT:    v_max_f32_e32 v9, v9, v9
; GFX10-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; GFX10-NEXT:    v_cndmask_b32_e32 v8, v12, v13, vcc_lo
; GFX10-NEXT:    v_bfe_u32 v12, v11, 16, 1
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v10
; GFX10-NEXT:    v_bfe_u32 v15, v9, 16, 1
; GFX10-NEXT:    v_max_f32_e32 v6, v6, v6
; GFX10-NEXT:    v_lshrrev_b32_e32 v13, 16, v8
; GFX10-NEXT:    v_add3_u32 v12, v12, v11, 0x7fff
; GFX10-NEXT:    v_and_b32_e32 v8, 0xffff0000, v8
; GFX10-NEXT:    v_or_b32_e32 v16, 0x400000, v6
; GFX10-NEXT:    v_cndmask_b32_e32 v10, v13, v10, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v11, v11
; GFX10-NEXT:    v_cndmask_b32_e32 v11, v12, v14, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v5
; GFX10-NEXT:    v_bfe_u32 v12, v6, 16, 1
; GFX10-NEXT:    v_add3_u32 v14, v15, v9, 0x7fff
; GFX10-NEXT:    v_or_b32_e32 v15, 0x400000, v9
; GFX10-NEXT:    v_cndmask_b32_e32 v5, v10, v5, vcc_lo
; GFX10-NEXT:    v_lshrrev_b32_e32 v10, 16, v11
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v4
; GFX10-NEXT:    v_add3_u32 v12, v12, v6, 0x7fff
; GFX10-NEXT:    v_and_b32_e32 v11, 0xffff0000, v11
; GFX10-NEXT:    v_cndmask_b32_e32 v4, v10, v4, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
; GFX10-NEXT:    v_cndmask_b32_e32 v9, v14, v15, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX10-NEXT:    v_cndmask_b32_e32 v6, v12, v16, vcc_lo
; GFX10-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v8
; GFX10-NEXT:    v_lshrrev_b32_e32 v8, 16, v9
; GFX10-NEXT:    v_cndmask_b32_e32 v5, v13, v5, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v7
; GFX10-NEXT:    v_cndmask_b32_e32 v4, v4, v7, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v1
; GFX10-NEXT:    v_lshrrev_b32_e32 v7, 16, v6
; GFX10-NEXT:    v_cndmask_b32_e32 v1, v8, v1, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v0
; GFX10-NEXT:    v_cndmask_b32_e32 v0, v7, v0, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v3
; GFX10-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v2
; GFX10-NEXT:    v_and_b32_e32 v3, 0xffff0000, v6
; GFX10-NEXT:    v_cndmask_b32_e32 v0, v0, v2, vcc_lo
; GFX10-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v11
; GFX10-NEXT:    v_and_b32_e32 v2, 0xffff0000, v9
; GFX10-NEXT:    v_cndmask_b32_e32 v4, v10, v4, vcc_lo
; GFX10-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v3
; GFX10-NEXT:    v_cndmask_b32_e32 v0, v7, v0, vcc_lo
; GFX10-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v2
; GFX10-NEXT:    v_perm_b32 v0, v4, v0, 0x5040100
; GFX10-NEXT:    v_cndmask_b32_e32 v1, v8, v1, vcc_lo
; GFX10-NEXT:    v_perm_b32 v1, v5, v1, 0x5040100
; GFX10-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-TRUE16-LABEL: v_maximumnum_v4bf16:
; GFX11-TRUE16:       ; %bb.0:
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xffff0000, v1
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xffff0000, v3
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xffff0000, v0
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff0000, v2
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 16, v1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s0, v5, v5
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 16, v3
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s1, v6, v6
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s2, v8, v8
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v4.l, v1.h, v3.h, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s3, v9, v9
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v6.l, v0.h, v2.h, s1
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v10, 16, v0
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v5.l, v3.h, v4.l, s0
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, v4.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v11, 16, v2
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v7.l, v2.h, v6.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s4, v10, v10
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v9.l, v5.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v1.l, v1.l, v3.l, s2
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v10.l, v6.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v12.l, v7.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s0, v11, v11
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v0.l, v0.l, v2.l, s4
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v3.l, v3.l, v1.l, s3
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
; GFX11-TRUE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v8, v9
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v11.l, v1.l
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v2.l, v2.l, v0.l, s0
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v13.l, v3.l
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v8.l, v5.l, v4.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v10, v12
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v9.l, v0.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v14.l, v2.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v10.l, v8.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 16, v13
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v8.l, v7.l, v6.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v13, 16, v14
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
; GFX11-TRUE16-NEXT:    v_cmp_gt_f32_e64 s0, v11, v12
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v11.l, v8.l
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v4.l
; GFX11-TRUE16-NEXT:    v_cmp_gt_f32_e64 s1, v9, v13
; GFX11-TRUE16-NEXT:    v_max_f32_e32 v10, v10, v10
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v8.l, v3.l, v1.l, s0
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s2, 0, v6.l
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v9.l, v2.l, v0.l, s1
; GFX11-TRUE16-NEXT:    v_bfe_u32 v12, v10, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v13, 0x400000, v10
; GFX11-TRUE16-NEXT:    v_max_f32_e32 v11, v11, v11
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s1, v10, v10
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; GFX11-TRUE16-NEXT:    v_add3_u32 v12, v12, v10, 0x7fff
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; GFX11-TRUE16-NEXT:    v_bfe_u32 v14, v11, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s0, 0, v5.l
; GFX11-TRUE16-NEXT:    v_max_f32_e32 v8, v8, v8
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e64 v10, v12, v13, s1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, 0x400000, v11
; GFX11-TRUE16-NEXT:    v_add3_u32 v13, v14, v11, 0x7fff
; GFX11-TRUE16-NEXT:    v_max_f32_e32 v9, v9, v9
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s1, v8, v8
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v0.h, v10.h, v4.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xffff0000, v10
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v11, v11
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v0.h, v0.h, v5.l, s0
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s0, 0, v7.l
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v11, v13, v12, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v12, v8, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v4
; GFX11-TRUE16-NEXT:    v_bfe_u32 v4, v9, 16, 1
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v1.h, v11.h, v6.l, s2
; GFX11-TRUE16-NEXT:    v_add3_u32 v5, v12, v8, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, 0x400000, v8
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_add3_u32 v4, v4, v9, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, 0x400000, v9
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s2, v9, v9
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v2.h, v1.h, v7.l, s0
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e64 v5, v5, v6, s1
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s1, 0, v1.l
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xffff0000, v11
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e64 v4, v4, v12, s2
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s2, 0, v0.l
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s0, 0, v3.l
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v1.l, v5.h, v1.l, s1
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s1, 0, v2.l
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff0000, v4
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v0.l, v4.h, v0.l, s2
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xffff0000, v5
; GFX11-TRUE16-NEXT:    v_cmp_eq_f32_e64 s2, 0, v6
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v1.l, v1.l, v3.l, s0
; GFX11-TRUE16-NEXT:    v_cmp_eq_f32_e64 s0, 0, v7
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v0.l, v0.l, v2.l, s1
; GFX11-TRUE16-NEXT:    v_cmp_eq_f32_e64 s1, 0, v8
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v1.h, v10.h, v0.h, vcc_lo
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v0.h, v11.h, v2.h, s2
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v0.l, v4.h, v0.l, s0
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v1.l, v5.h, v1.l, s1
; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-FAKE16-LABEL: v_maximumnum_v4bf16:
; GFX11-FAKE16:       ; %bb.0:
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v8, 16, v2
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xffff0000, v1
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v5, 16, v3
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v6, 16, v1
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v9, 16, v0
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v12, 16, v3
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v4, v6, v5 :: v_dual_and_b32 v7, 0xffff0000, v3
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v6, 0xffff0000, v0
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v5, v5, v4, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v7, 0xffff0000, v2
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v10, 16, v5
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v6, v9, v8, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v7, v8, v6, vcc_lo
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v9, 16, v4
; GFX11-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v9, v10
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v11, 16, v1
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v9, 16, v0
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v8, v5, v4, vcc_lo
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v11, v11
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v13, 16, v7
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v1, v1, v3 :: v_dual_lshlrev_b32 v10, 16, v6
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_max_f32_e32 v8, v8, v8
; GFX11-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v10, v13
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v13, 16, v1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v11, 0x400000, v8
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v10, v7, v6, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
; GFX11-FAKE16-NEXT:    v_bfe_u32 v9, v8, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v0, v2, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v12, v12
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_add3_u32 v9, v9, v8, 0x7fff
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v3, v3, v1 :: v_dual_lshlrev_b32 v14, 16, v2
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v12, 16, v3
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v14, v14
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v2, v0, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v15, 16, v0
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v14, 16, v2
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v8, v9, v11, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v13, v12
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v13, 16, v8
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v9, v3, v1, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v15, v14
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v8, 0xffff0000, v8
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v12, v2, v0 :: v_dual_lshlrev_b32 v9, 16, v9
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v4
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_dual_max_f32 v9, v9, v9 :: v_dual_lshlrev_b32 v12, 16, v12
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v4, v13, v4, vcc_lo
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v15, 0x400000, v9
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_max_f32_e32 v12, v12, v12
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v16, 0x400000, v12
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_max_f32_e32 v10, v10, v10
; GFX11-FAKE16-NEXT:    v_bfe_u32 v11, v10, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v14, 0x400000, v10
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v10, v10
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_add3_u32 v11, v11, v10, 0x7fff
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v10, v11, v14, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v5
; GFX11-FAKE16-NEXT:    v_bfe_u32 v11, v9, 16, 1
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v14, 16, v10
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v4, v4, v5, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v6
; GFX11-FAKE16-NEXT:    v_bfe_u32 v5, v12, 16, 1
; GFX11-FAKE16-NEXT:    v_add3_u32 v11, v11, v9, 0x7fff
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xffff0000, v10
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v6, v14, v6, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
; GFX11-FAKE16-NEXT:    v_add3_u32 v5, v5, v12, 0x7fff
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v9, v11, v15, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v12, v12
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v5, v5, v16, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v8
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v8, 16, v9
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v4, v13, v4, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v7
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v6, v6, v7, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v1
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v7, 16, v5
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v8, v1, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v0
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v7, v0, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v3
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v2
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v0, v0, v2 :: v_dual_and_b32 v3, 0xffff0000, v5
; GFX11-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v10
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v5, v14, v6 :: v_dual_and_b32 v2, 0xffff0000, v9
; GFX11-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v3
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v7, v0, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v2
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_perm_b32 v0, v5, v0, 0x5040100
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v8, v1, vcc_lo
; GFX11-FAKE16-NEXT:    v_perm_b32 v1, v4, v1, 0x5040100
; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
;
; GFX12-TRUE16-LABEL: v_maximumnum_v4bf16:
; GFX12-TRUE16:       ; %bb.0:
; GFX12-TRUE16-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX12-TRUE16-NEXT:    s_wait_expcnt 0x0
; GFX12-TRUE16-NEXT:    s_wait_samplecnt 0x0
; GFX12-TRUE16-NEXT:    s_wait_bvhcnt 0x0
; GFX12-TRUE16-NEXT:    s_wait_kmcnt 0x0
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v4, 0xffff0000, v1
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v5, 0xffff0000, v3
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v6, 0xffff0000, v0
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff0000, v2
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 16, v1
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s0, v5, v5
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 16, v3
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s1, v6, v6
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s2, v8, v8
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v4.l, v1.h, v3.h, vcc_lo
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s3, v9, v9
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v6.l, v0.h, v2.h, s1
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v10, 16, v0
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v5.l, v3.h, v4.l, s0
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v8.l, v4.l
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v11, 16, v2
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v7.l, v2.h, v6.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s4, v10, v10
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v9.l, v5.l
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v1.l, v1.l, v3.l, s2
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v10.l, v6.l
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v12.l, v7.l
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s0, v11, v11
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v0.l, v0.l, v2.l, s4
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v3.l, v3.l, v1.l, s3
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
; GFX12-TRUE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v8, v9
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v11.l, v1.l
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v2.l, v2.l, v0.l, s0
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v13.l, v3.l
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v8.l, v5.l, v4.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v10, v12
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v9.l, v0.l
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v14.l, v2.l
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v10.l, v8.l
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 16, v13
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v8.l, v7.l, v6.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v13, 16, v14
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
; GFX12-TRUE16-NEXT:    v_cmp_gt_f32_e64 s0, v11, v12
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v11.l, v8.l
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v4.l
; GFX12-TRUE16-NEXT:    v_cmp_gt_f32_e64 s1, v9, v13
; GFX12-TRUE16-NEXT:    v_max_num_f32_e32 v10, v10, v10
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v8.l, v3.l, v1.l, s0
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s2, 0, v6.l
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v9.l, v2.l, v0.l, s1
; GFX12-TRUE16-NEXT:    v_bfe_u32 v12, v10, 16, 1
; GFX12-TRUE16-NEXT:    v_or_b32_e32 v13, 0x400000, v10
; GFX12-TRUE16-NEXT:    v_max_num_f32_e32 v11, v11, v11
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s1, v10, v10
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; GFX12-TRUE16-NEXT:    v_add3_u32 v12, v12, v10, 0x7fff
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; GFX12-TRUE16-NEXT:    v_bfe_u32 v14, v11, 16, 1
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s0, 0, v5.l
; GFX12-TRUE16-NEXT:    v_max_num_f32_e32 v8, v8, v8
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b32_e64 v10, v12, v13, s1
; GFX12-TRUE16-NEXT:    v_or_b32_e32 v12, 0x400000, v11
; GFX12-TRUE16-NEXT:    v_add3_u32 v13, v14, v11, 0x7fff
; GFX12-TRUE16-NEXT:    v_max_num_f32_e32 v9, v9, v9
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s1, v8, v8
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v0.h, v10.h, v4.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v4, 0xffff0000, v10
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v11, v11
; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v0.h, v0.h, v5.l, s0
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s0, 0, v7.l
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b32_e32 v11, v13, v12, vcc_lo
; GFX12-TRUE16-NEXT:    v_bfe_u32 v12, v8, 16, 1
; GFX12-TRUE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v4
; GFX12-TRUE16-NEXT:    v_bfe_u32 v4, v9, 16, 1
; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v1.h, v11.h, v6.l, s2
; GFX12-TRUE16-NEXT:    v_add3_u32 v5, v12, v8, 0x7fff
; GFX12-TRUE16-NEXT:    v_or_b32_e32 v6, 0x400000, v8
; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX12-TRUE16-NEXT:    v_add3_u32 v4, v4, v9, 0x7fff
; GFX12-TRUE16-NEXT:    v_or_b32_e32 v12, 0x400000, v9
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s2, v9, v9
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v2.h, v1.h, v7.l, s0
; GFX12-TRUE16-NEXT:    v_cndmask_b32_e64 v5, v5, v6, s1
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s1, 0, v1.l
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v6, 0xffff0000, v11
; GFX12-TRUE16-NEXT:    v_cndmask_b32_e64 v4, v4, v12, s2
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s2, 0, v0.l
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s0, 0, v3.l
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v1.l, v5.h, v1.l, s1
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s1, 0, v2.l
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff0000, v4
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v0.l, v4.h, v0.l, s2
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v8, 0xffff0000, v5
; GFX12-TRUE16-NEXT:    v_cmp_eq_f32_e64 s2, 0, v6
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v1.l, v1.l, v3.l, s0
; GFX12-TRUE16-NEXT:    v_cmp_eq_f32_e64 s0, 0, v7
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v0.l, v0.l, v2.l, s1
; GFX12-TRUE16-NEXT:    v_cmp_eq_f32_e64 s1, 0, v8
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v1.h, v10.h, v0.h, vcc_lo
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v0.h, v11.h, v2.h, s2
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v0.l, v4.h, v0.l, s0
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v1.l, v5.h, v1.l, s1
; GFX12-TRUE16-NEXT:    s_setpc_b64 s[30:31]
;
; GFX12-FAKE16-LABEL: v_maximumnum_v4bf16:
; GFX12-FAKE16:       ; %bb.0:
; GFX12-FAKE16-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX12-FAKE16-NEXT:    s_wait_expcnt 0x0
; GFX12-FAKE16-NEXT:    s_wait_samplecnt 0x0
; GFX12-FAKE16-NEXT:    s_wait_bvhcnt 0x0
; GFX12-FAKE16-NEXT:    s_wait_kmcnt 0x0
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v8, 16, v2
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v4, 0xffff0000, v1
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v5, 16, v3
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v6, 16, v1
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v9, 16, v0
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v12, 16, v3
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v4, v6, v5 :: v_dual_and_b32 v7, 0xffff0000, v3
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v6, 0xffff0000, v0
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v5, v5, v4, vcc_lo
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v7, 0xffff0000, v2
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v10, 16, v5
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v6, v9, v8, vcc_lo
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v7, v8, v6, vcc_lo
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v9, 16, v4
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_4) | instid1(VALU_DEP_3)
; GFX12-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v9, v10
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v11, 16, v1
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v9, 16, v0
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v8, v5, v4, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v11, v11
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v13, 16, v7
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v1, v1, v3 :: v_dual_lshlrev_b32 v10, 16, v6
; GFX12-FAKE16-NEXT:    v_max_num_f32_e32 v8, v8, v8
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX12-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v10, v13
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v13, 16, v1
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX12-FAKE16-NEXT:    v_or_b32_e32 v11, 0x400000, v8
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v10, v7, v6, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
; GFX12-FAKE16-NEXT:    v_bfe_u32 v9, v8, 16, 1
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v0, v2, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v12, v12
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_1)
; GFX12-FAKE16-NEXT:    v_add3_u32 v9, v9, v8, 0x7fff
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v3, v3, v1 :: v_dual_lshlrev_b32 v14, 16, v2
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v12, 16, v3
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_4) | instid1(VALU_DEP_3)
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v14, v14
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v2, v0, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v15, 16, v0
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v14, 16, v2
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v8, v9, v11, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v13, v12
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v13, 16, v8
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v9, v3, v1, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v15, v14
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v8, 0xffff0000, v8
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v12, v2, v0 :: v_dual_lshlrev_b32 v9, 16, v9
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v4
; GFX12-FAKE16-NEXT:    v_dual_max_num_f32 v9, v9, v9 :: v_dual_lshlrev_b32 v12, 16, v12
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v4, v13, v4, vcc_lo
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX12-FAKE16-NEXT:    v_or_b32_e32 v15, 0x400000, v9
; GFX12-FAKE16-NEXT:    v_max_num_f32_e32 v12, v12, v12
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX12-FAKE16-NEXT:    v_or_b32_e32 v16, 0x400000, v12
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
; GFX12-FAKE16-NEXT:    v_max_num_f32_e32 v10, v10, v10
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX12-FAKE16-NEXT:    v_bfe_u32 v11, v10, 16, 1
; GFX12-FAKE16-NEXT:    v_or_b32_e32 v14, 0x400000, v10
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v10, v10
; GFX12-FAKE16-NEXT:    v_add3_u32 v11, v11, v10, 0x7fff
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v10, v11, v14, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v5
; GFX12-FAKE16-NEXT:    v_bfe_u32 v11, v9, 16, 1
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v14, 16, v10
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v4, v4, v5, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v6
; GFX12-FAKE16-NEXT:    v_bfe_u32 v5, v12, 16, 1
; GFX12-FAKE16-NEXT:    v_add3_u32 v11, v11, v9, 0x7fff
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v10, 0xffff0000, v10
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v6, v14, v6, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
; GFX12-FAKE16-NEXT:    v_add3_u32 v5, v5, v12, 0x7fff
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v9, v11, v15, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v12, v12
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v5, v5, v16, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v8
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v8, 16, v9
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v4, v13, v4, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v7
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v6, v6, v7, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v1
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v7, 16, v5
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v8, v1, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v0
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v7, v0, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v3
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v2
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v0, v0, v2 :: v_dual_and_b32 v3, 0xffff0000, v5
; GFX12-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v10
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v5, v14, v6 :: v_dual_and_b32 v2, 0xffff0000, v9
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX12-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v3
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v7, v0, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v2
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1)
; GFX12-FAKE16-NEXT:    v_perm_b32 v0, v5, v0, 0x5040100
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v8, v1, vcc_lo
; GFX12-FAKE16-NEXT:    v_perm_b32 v1, v4, v1, 0x5040100
; GFX12-FAKE16-NEXT:    s_setpc_b64 s[30:31]
  %result = call <4 x bfloat> @llvm.maximumnum.v4bf16(<4 x bfloat> %x, <4 x bfloat> %y)
  ret <4 x bfloat> %result
}

define <4 x bfloat> @v_maximumnum_v4bf16_nnan(<4 x bfloat> %x, <4 x bfloat> %y) {
; GFX7-LABEL: v_maximumnum_v4bf16_nnan:
; GFX7:       ; %bb.0:
; GFX7-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT:    v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT:    v_mul_f32_e32 v4, 1.0, v4
; GFX7-NEXT:    v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT:    v_mul_f32_e32 v5, 1.0, v5
; GFX7-NEXT:    v_mul_f32_e32 v2, 1.0, v2
; GFX7-NEXT:    v_mul_f32_e32 v6, 1.0, v6
; GFX7-NEXT:    v_mul_f32_e32 v3, 1.0, v3
; GFX7-NEXT:    v_mul_f32_e32 v7, 1.0, v7
; GFX7-NEXT:    v_and_b32_e32 v7, 0xffff0000, v7
; GFX7-NEXT:    v_and_b32_e32 v3, 0xffff0000, v3
; GFX7-NEXT:    v_and_b32_e32 v6, 0xffff0000, v6
; GFX7-NEXT:    v_and_b32_e32 v2, 0xffff0000, v2
; GFX7-NEXT:    v_and_b32_e32 v5, 0xffff0000, v5
; GFX7-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT:    v_and_b32_e32 v4, 0xffff0000, v4
; GFX7-NEXT:    v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT:    v_max_f32_e32 v3, v3, v7
; GFX7-NEXT:    v_max_f32_e32 v2, v2, v6
; GFX7-NEXT:    v_max_f32_e32 v1, v1, v5
; GFX7-NEXT:    v_max_f32_e32 v0, v0, v4
; GFX7-NEXT:    v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT:    v_and_b32_e32 v2, 0xffff0000, v2
; GFX7-NEXT:    v_and_b32_e32 v3, 0xffff0000, v3
; GFX7-NEXT:    s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_maximumnum_v4bf16_nnan:
; GFX8:       ; %bb.0:
; GFX8-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT:    v_and_b32_e32 v4, 0xffff0000, v3
; GFX8-NEXT:    v_and_b32_e32 v5, 0xffff0000, v1
; GFX8-NEXT:    v_lshrrev_b32_e32 v6, 16, v1
; GFX8-NEXT:    v_lshrrev_b32_e32 v7, 16, v3
; GFX8-NEXT:    v_cmp_gt_f32_e32 vcc, v5, v4
; GFX8-NEXT:    v_cndmask_b32_e32 v4, v7, v6, vcc
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v6
; GFX8-NEXT:    v_cndmask_b32_e32 v5, v4, v6, vcc
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v7
; GFX8-NEXT:    v_lshlrev_b32_e32 v6, 16, v4
; GFX8-NEXT:    v_cndmask_b32_e32 v5, v5, v7, vcc
; GFX8-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v6
; GFX8-NEXT:    v_cndmask_b32_e32 v4, v4, v5, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v5, 16, v3
; GFX8-NEXT:    v_lshlrev_b32_e32 v6, 16, v1
; GFX8-NEXT:    v_cmp_gt_f32_e32 vcc, v6, v5
; GFX8-NEXT:    v_cndmask_b32_e32 v5, v3, v1, vcc
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v1
; GFX8-NEXT:    v_cndmask_b32_e32 v1, v5, v1, vcc
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v3
; GFX8-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v3, 16, v5
; GFX8-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v3
; GFX8-NEXT:    v_cndmask_b32_e32 v1, v5, v1, vcc
; GFX8-NEXT:    v_and_b32_e32 v3, 0xffff0000, v2
; GFX8-NEXT:    v_and_b32_e32 v5, 0xffff0000, v0
; GFX8-NEXT:    v_lshrrev_b32_e32 v6, 16, v0
; GFX8-NEXT:    v_lshrrev_b32_e32 v7, 16, v2
; GFX8-NEXT:    v_cmp_gt_f32_e32 vcc, v5, v3
; GFX8-NEXT:    v_cndmask_b32_e32 v3, v7, v6, vcc
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v6
; GFX8-NEXT:    v_cndmask_b32_e32 v5, v3, v6, vcc
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v7
; GFX8-NEXT:    v_lshlrev_b32_e32 v6, 16, v3
; GFX8-NEXT:    v_cndmask_b32_e32 v5, v5, v7, vcc
; GFX8-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v6
; GFX8-NEXT:    v_cndmask_b32_e32 v3, v3, v5, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v5, 16, v2
; GFX8-NEXT:    v_lshlrev_b32_e32 v6, 16, v0
; GFX8-NEXT:    v_cmp_gt_f32_e32 vcc, v6, v5
; GFX8-NEXT:    v_cndmask_b32_e32 v5, v2, v0, vcc
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v0
; GFX8-NEXT:    v_cndmask_b32_e32 v0, v5, v0, vcc
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v2
; GFX8-NEXT:    v_cndmask_b32_e32 v0, v0, v2, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v2, 16, v5
; GFX8-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v2
; GFX8-NEXT:    v_cndmask_b32_e32 v0, v5, v0, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v2, 16, v3
; GFX8-NEXT:    v_or_b32_sdwa v0, v0, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX8-NEXT:    v_lshlrev_b32_e32 v2, 16, v4
; GFX8-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX8-NEXT:    s_setpc_b64 s[30:31]
;
; GFX900-LABEL: v_maximumnum_v4bf16_nnan:
; GFX900:       ; %bb.0:
; GFX900-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX900-NEXT:    v_lshlrev_b32_e32 v4, 16, v3
; GFX900-NEXT:    v_lshlrev_b32_e32 v5, 16, v1
; GFX900-NEXT:    v_cmp_gt_f32_e32 vcc, v5, v4
; GFX900-NEXT:    v_cndmask_b32_e32 v4, v3, v1, vcc
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v1
; GFX900-NEXT:    v_cndmask_b32_e32 v5, v4, v1, vcc
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v3
; GFX900-NEXT:    v_lshlrev_b32_e32 v6, 16, v4
; GFX900-NEXT:    v_cndmask_b32_e32 v5, v5, v3, vcc
; GFX900-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v6
; GFX900-NEXT:    v_cndmask_b32_e32 v4, v4, v5, vcc
; GFX900-NEXT:    v_and_b32_e32 v5, 0xffff0000, v3
; GFX900-NEXT:    v_and_b32_e32 v6, 0xffff0000, v1
; GFX900-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
; GFX900-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
; GFX900-NEXT:    v_cmp_gt_f32_e32 vcc, v6, v5
; GFX900-NEXT:    v_cndmask_b32_e32 v5, v3, v1, vcc
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v1
; GFX900-NEXT:    v_cndmask_b32_e32 v1, v5, v1, vcc
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v3
; GFX900-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v3, 16, v5
; GFX900-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v3
; GFX900-NEXT:    v_cndmask_b32_e32 v1, v5, v1, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v3, 16, v2
; GFX900-NEXT:    v_lshlrev_b32_e32 v5, 16, v0
; GFX900-NEXT:    v_cmp_gt_f32_e32 vcc, v5, v3
; GFX900-NEXT:    v_cndmask_b32_e32 v3, v2, v0, vcc
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v0
; GFX900-NEXT:    v_cndmask_b32_e32 v5, v3, v0, vcc
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v2
; GFX900-NEXT:    v_lshlrev_b32_e32 v6, 16, v3
; GFX900-NEXT:    v_cndmask_b32_e32 v5, v5, v2, vcc
; GFX900-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v6
; GFX900-NEXT:    v_cndmask_b32_e32 v3, v3, v5, vcc
; GFX900-NEXT:    v_and_b32_e32 v5, 0xffff0000, v2
; GFX900-NEXT:    v_and_b32_e32 v6, 0xffff0000, v0
; GFX900-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; GFX900-NEXT:    v_lshrrev_b32_e32 v2, 16, v2
; GFX900-NEXT:    v_cmp_gt_f32_e32 vcc, v6, v5
; GFX900-NEXT:    v_cndmask_b32_e32 v5, v2, v0, vcc
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v0
; GFX900-NEXT:    v_cndmask_b32_e32 v0, v5, v0, vcc
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v2
; GFX900-NEXT:    v_cndmask_b32_e32 v0, v0, v2, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v2, 16, v5
; GFX900-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v2
; GFX900-NEXT:    v_cndmask_b32_e32 v0, v5, v0, vcc
; GFX900-NEXT:    s_mov_b32 s4, 0x5040100
; GFX900-NEXT:    v_perm_b32 v0, v0, v3, s4
; GFX900-NEXT:    v_perm_b32 v1, v1, v4, s4
; GFX900-NEXT:    s_setpc_b64 s[30:31]
;
; GFX950-LABEL: v_maximumnum_v4bf16_nnan:
; GFX950:       ; %bb.0:
; GFX950-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX950-NEXT:    v_lshlrev_b32_e32 v4, 16, v3
; GFX950-NEXT:    v_lshlrev_b32_e32 v5, 16, v1
; GFX950-NEXT:    v_cmp_gt_f32_e32 vcc, v5, v4
; GFX950-NEXT:    s_mov_b32 s0, 0x5040100
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v4, v3, v1, vcc
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v1
; GFX950-NEXT:    v_lshlrev_b32_e32 v6, 16, v4
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v5, v4, v1, vcc
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v3
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v5, v5, v3, vcc
; GFX950-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v6
; GFX950-NEXT:    v_and_b32_e32 v6, 0xffff0000, v1
; GFX950-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
; GFX950-NEXT:    v_cndmask_b32_e32 v4, v4, v5, vcc
; GFX950-NEXT:    v_and_b32_e32 v5, 0xffff0000, v3
; GFX950-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
; GFX950-NEXT:    v_cmp_gt_f32_e32 vcc, v6, v5
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v5, v3, v1, vcc
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v1
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v1, v5, v1, vcc
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v3
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v3, 16, v5
; GFX950-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v3
; GFX950-NEXT:    v_lshlrev_b32_e32 v3, 16, v2
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v1, v5, v1, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v5, 16, v0
; GFX950-NEXT:    v_cmp_gt_f32_e32 vcc, v5, v3
; GFX950-NEXT:    v_perm_b32 v1, v1, v4, s0
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v3, v2, v0, vcc
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v0
; GFX950-NEXT:    v_lshlrev_b32_e32 v6, 16, v3
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v5, v3, v0, vcc
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v2
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v5, v5, v2, vcc
; GFX950-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v6
; GFX950-NEXT:    v_and_b32_e32 v6, 0xffff0000, v0
; GFX950-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; GFX950-NEXT:    v_cndmask_b32_e32 v3, v3, v5, vcc
; GFX950-NEXT:    v_and_b32_e32 v5, 0xffff0000, v2
; GFX950-NEXT:    v_lshrrev_b32_e32 v2, 16, v2
; GFX950-NEXT:    v_cmp_gt_f32_e32 vcc, v6, v5
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v5, v2, v0, vcc
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v0
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v0, v5, v0, vcc
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v2
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v0, v0, v2, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v2, 16, v5
; GFX950-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v2
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v0, v5, v0, vcc
; GFX950-NEXT:    v_perm_b32 v0, v0, v3, s0
; GFX950-NEXT:    s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_maximumnum_v4bf16_nnan:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT:    v_lshlrev_b32_e32 v4, 16, v3
; GFX10-NEXT:    v_lshlrev_b32_e32 v5, 16, v1
; GFX10-NEXT:    v_and_b32_e32 v6, 0xffff0000, v3
; GFX10-NEXT:    v_and_b32_e32 v7, 0xffff0000, v1
; GFX10-NEXT:    v_lshrrev_b32_e32 v8, 16, v1
; GFX10-NEXT:    v_lshlrev_b32_e32 v9, 16, v2
; GFX10-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v5, v4
; GFX10-NEXT:    v_lshrrev_b32_e32 v5, 16, v3
; GFX10-NEXT:    v_lshlrev_b32_e32 v10, 16, v0
; GFX10-NEXT:    v_and_b32_e32 v11, 0xffff0000, v2
; GFX10-NEXT:    v_and_b32_e32 v12, 0xffff0000, v0
; GFX10-NEXT:    v_cndmask_b32_e32 v4, v3, v1, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v1
; GFX10-NEXT:    v_lshlrev_b32_e32 v13, 16, v4
; GFX10-NEXT:    v_cndmask_b32_e32 v1, v4, v1, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v3
; GFX10-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc_lo
; GFX10-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v7, v6
; GFX10-NEXT:    v_lshrrev_b32_e32 v3, 16, v0
; GFX10-NEXT:    v_cndmask_b32_e32 v6, v5, v8, vcc_lo
; GFX10-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v13
; GFX10-NEXT:    v_cndmask_b32_e32 v1, v4, v1, vcc_lo
; GFX10-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v10, v9
; GFX10-NEXT:    v_lshrrev_b32_e32 v4, 16, v2
; GFX10-NEXT:    v_cndmask_b32_e32 v7, v2, v0, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v8
; GFX10-NEXT:    v_cndmask_b32_e32 v8, v6, v8, vcc_lo
; GFX10-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v12, v11
; GFX10-NEXT:    v_cndmask_b32_e32 v9, v4, v3, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v0
; GFX10-NEXT:    v_cndmask_b32_e32 v0, v7, v0, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v5
; GFX10-NEXT:    v_cndmask_b32_e32 v5, v8, v5, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v3
; GFX10-NEXT:    v_lshlrev_b32_e32 v8, 16, v9
; GFX10-NEXT:    v_cndmask_b32_e32 v3, v9, v3, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v2
; GFX10-NEXT:    v_cndmask_b32_e32 v0, v0, v2, vcc_lo
; GFX10-NEXT:    v_lshlrev_b32_e32 v2, 16, v7
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v4
; GFX10-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc_lo
; GFX10-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v2
; GFX10-NEXT:    v_lshlrev_b32_e32 v4, 16, v6
; GFX10-NEXT:    v_cndmask_b32_e32 v0, v7, v0, vcc_lo
; GFX10-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v8
; GFX10-NEXT:    v_cndmask_b32_e32 v2, v9, v3, vcc_lo
; GFX10-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v4
; GFX10-NEXT:    v_perm_b32 v0, v2, v0, 0x5040100
; GFX10-NEXT:    v_cndmask_b32_e32 v3, v6, v5, vcc_lo
; GFX10-NEXT:    v_perm_b32 v1, v3, v1, 0x5040100
; GFX10-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-TRUE16-LABEL: v_maximumnum_v4bf16_nnan:
; GFX11-TRUE16:       ; %bb.0:
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v4, 16, v3
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 16, v1
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xffff0000, v3
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff0000, v1
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v1.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 16, v2
; GFX11-TRUE16-NEXT:    v_cmp_gt_f32_e64 s1, v5, v4
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xffff0000, v0
; GFX11-TRUE16-NEXT:    v_cmp_gt_f32_e64 s4, v7, v6
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 16, v0
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff0000, v2
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v4.l, v3.l, v1.l, s1
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s0, 0, v3.l
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v5.l, v3.h, v1.h, s4
; GFX11-TRUE16-NEXT:    v_cmp_gt_f32_e64 s4, v6, v8
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s2, 0, v1.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v9.l, v4.l
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v1.l, v4.l, v1.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v10, v7
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v6.l, v2.l, v0.l, s4
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s1, 0, v0.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 16, v9
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v1.l, v1.l, v3.l, s0
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v3.l, v2.h, v0.h, vcc_lo
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.l, v5.l
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v0.h
; GFX11-TRUE16-NEXT:    v_cmp_eq_f32_e64 s0, 0, v8
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, v6.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v9.l, v3.l
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s3, 0, v3.h
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v1.h, v5.l, v1.h, s2
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v0.l, v6.l, v0.l, s1
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s1, 0, v2.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v0.h, v3.l, v0.h, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v2.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v1.h, v1.h, v3.h, s3
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v0.l, v0.l, v2.l, s1
; GFX11-TRUE16-NEXT:    v_cmp_eq_f32_e64 s1, 0, v8
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v0.h, v0.h, v2.h, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v9
; GFX11-TRUE16-NEXT:    v_cmp_eq_f32_e64 s2, 0, v7
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v1.l, v4.l, v1.l, s0
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v0.l, v6.l, v0.l, s1
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v0.h, v3.l, v0.h, vcc_lo
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v1.h, v5.l, v1.h, s2
; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-FAKE16-LABEL: v_maximumnum_v4bf16_nnan:
; GFX11-FAKE16:       ; %bb.0:
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v4, 16, v3
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v8, 16, v1
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v5, 16, v1
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v9, 16, v2
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v5, v4
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v5, 16, v3
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v7, 0xffff0000, v1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v4, v3, v1, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v1
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v6, 0xffff0000, v3
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v13, 16, v4
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v4, v1, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v3
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v1, v1, v3 :: v_dual_lshlrev_b32 v10, 16, v0
; GFX11-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v7, v6
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v3, 16, v0
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v6, v5, v8, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v13
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v1, v4, v1 :: v_dual_and_b32 v12, 0xffff0000, v0
; GFX11-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v10, v9
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v4, 16, v2
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v7, v2, v0, vcc_lo
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v11, 0xffff0000, v2
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v8
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v8, v6, v8, vcc_lo
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v12, v11
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v9, v4, v3, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v0
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v7, v0, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v5
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v5, v8, v5, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v3
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v3, v9, v3, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v2
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v0, v2, vcc_lo
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v7
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v4
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v3, v3, v4 :: v_dual_lshlrev_b32 v4, 16, v6
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v8, 16, v9
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v2
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v7, v0, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v8
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v9, v3, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v4
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v3, v6, v5, vcc_lo
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_perm_b32 v0, v2, v0, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v1, v3, v1, 0x5040100
; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
;
; GFX12-TRUE16-LABEL: v_maximumnum_v4bf16_nnan:
; GFX12-TRUE16:       ; %bb.0:
; GFX12-TRUE16-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX12-TRUE16-NEXT:    s_wait_expcnt 0x0
; GFX12-TRUE16-NEXT:    s_wait_samplecnt 0x0
; GFX12-TRUE16-NEXT:    s_wait_bvhcnt 0x0
; GFX12-TRUE16-NEXT:    s_wait_kmcnt 0x0
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v4, 16, v3
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 16, v1
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v6, 0xffff0000, v3
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff0000, v1
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v1.l
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 16, v2
; GFX12-TRUE16-NEXT:    v_cmp_gt_f32_e64 s1, v5, v4
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v10, 0xffff0000, v0
; GFX12-TRUE16-NEXT:    v_cmp_gt_f32_e64 s4, v7, v6
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 16, v0
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff0000, v2
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v4.l, v3.l, v1.l, s1
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s0, 0, v3.l
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v5.l, v3.h, v1.h, s4
; GFX12-TRUE16-NEXT:    v_cmp_gt_f32_e64 s4, v6, v8
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s2, 0, v1.h
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v9.l, v4.l
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v1.l, v4.l, v1.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v10, v7
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v6.l, v2.l, v0.l, s4
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s1, 0, v0.l
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 16, v9
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v1.l, v1.l, v3.l, s0
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v3.l, v2.h, v0.h, vcc_lo
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v7.l, v5.l
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v0.h
; GFX12-TRUE16-NEXT:    v_cmp_eq_f32_e64 s0, 0, v8
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v8.l, v6.l
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v9.l, v3.l
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s3, 0, v3.h
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v1.h, v5.l, v1.h, s2
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v0.l, v6.l, v0.l, s1
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s1, 0, v2.l
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v0.h, v3.l, v0.h, vcc_lo
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v2.h
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v1.h, v1.h, v3.h, s3
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v0.l, v0.l, v2.l, s1
; GFX12-TRUE16-NEXT:    v_cmp_eq_f32_e64 s1, 0, v8
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v0.h, v0.h, v2.h, vcc_lo
; GFX12-TRUE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v9
; GFX12-TRUE16-NEXT:    v_cmp_eq_f32_e64 s2, 0, v7
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v1.l, v4.l, v1.l, s0
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v0.l, v6.l, v0.l, s1
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v0.h, v3.l, v0.h, vcc_lo
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v1.h, v5.l, v1.h, s2
; GFX12-TRUE16-NEXT:    s_setpc_b64 s[30:31]
;
; GFX12-FAKE16-LABEL: v_maximumnum_v4bf16_nnan:
; GFX12-FAKE16:       ; %bb.0:
; GFX12-FAKE16-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX12-FAKE16-NEXT:    s_wait_expcnt 0x0
; GFX12-FAKE16-NEXT:    s_wait_samplecnt 0x0
; GFX12-FAKE16-NEXT:    s_wait_bvhcnt 0x0
; GFX12-FAKE16-NEXT:    s_wait_kmcnt 0x0
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v4, 16, v3
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v8, 16, v1
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v5, 16, v1
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v9, 16, v2
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
; GFX12-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v5, v4
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v5, 16, v3
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v7, 0xffff0000, v1
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v4, v3, v1, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v1
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v6, 0xffff0000, v3
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_2)
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v13, 16, v4
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v4, v1, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v3
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v1, v1, v3 :: v_dual_lshlrev_b32 v10, 16, v0
; GFX12-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v7, v6
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v3, 16, v0
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v6, v5, v8, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v13
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v1, v4, v1 :: v_dual_and_b32 v12, 0xffff0000, v0
; GFX12-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v10, v9
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v4, 16, v2
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v7, v2, v0, vcc_lo
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v11, 0xffff0000, v2
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v8
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v8, v6, v8, vcc_lo
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX12-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v12, v11
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v9, v4, v3, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v0
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v7, v0, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v5
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v5, v8, v5, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v3
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v3, v9, v3, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v2
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v0, v2, vcc_lo
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v7
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v4
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v3, v3, v4 :: v_dual_lshlrev_b32 v4, 16, v6
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v8, 16, v9
; GFX12-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v2
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v7, v0, vcc_lo
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX12-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v8
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v9, v3, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v4
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v3, v6, v5, vcc_lo
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX12-FAKE16-NEXT:    v_perm_b32 v0, v2, v0, 0x5040100
; GFX12-FAKE16-NEXT:    v_perm_b32 v1, v3, v1, 0x5040100
; GFX12-FAKE16-NEXT:    s_setpc_b64 s[30:31]
  %result = call nnan <4 x bfloat> @llvm.maximumnum.v4bf16(<4 x bfloat> %x, <4 x bfloat> %y)
  ret <4 x bfloat> %result
}

define <6 x bfloat> @v_maximumnum_v6bf16(<6 x bfloat> %x, <6 x bfloat> %y) {
; GFX7-LABEL: v_maximumnum_v6bf16:
; GFX7:       ; %bb.0:
; GFX7-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT:    v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT:    v_mul_f32_e32 v6, 1.0, v6
; GFX7-NEXT:    v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT:    v_mul_f32_e32 v7, 1.0, v7
; GFX7-NEXT:    v_mul_f32_e32 v2, 1.0, v2
; GFX7-NEXT:    v_mul_f32_e32 v8, 1.0, v8
; GFX7-NEXT:    v_mul_f32_e32 v3, 1.0, v3
; GFX7-NEXT:    v_mul_f32_e32 v9, 1.0, v9
; GFX7-NEXT:    v_mul_f32_e32 v4, 1.0, v4
; GFX7-NEXT:    v_mul_f32_e32 v10, 1.0, v10
; GFX7-NEXT:    v_mul_f32_e32 v5, 1.0, v5
; GFX7-NEXT:    v_mul_f32_e32 v11, 1.0, v11
; GFX7-NEXT:    v_and_b32_e32 v11, 0xffff0000, v11
; GFX7-NEXT:    v_and_b32_e32 v5, 0xffff0000, v5
; GFX7-NEXT:    v_and_b32_e32 v10, 0xffff0000, v10
; GFX7-NEXT:    v_and_b32_e32 v4, 0xffff0000, v4
; GFX7-NEXT:    v_and_b32_e32 v9, 0xffff0000, v9
; GFX7-NEXT:    v_and_b32_e32 v3, 0xffff0000, v3
; GFX7-NEXT:    v_and_b32_e32 v8, 0xffff0000, v8
; GFX7-NEXT:    v_and_b32_e32 v2, 0xffff0000, v2
; GFX7-NEXT:    v_and_b32_e32 v7, 0xffff0000, v7
; GFX7-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT:    v_and_b32_e32 v6, 0xffff0000, v6
; GFX7-NEXT:    v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT:    v_max_f32_e32 v5, v5, v11
; GFX7-NEXT:    v_max_f32_e32 v4, v4, v10
; GFX7-NEXT:    v_max_f32_e32 v3, v3, v9
; GFX7-NEXT:    v_max_f32_e32 v2, v2, v8
; GFX7-NEXT:    v_max_f32_e32 v1, v1, v7
; GFX7-NEXT:    v_max_f32_e32 v0, v0, v6
; GFX7-NEXT:    v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT:    v_and_b32_e32 v2, 0xffff0000, v2
; GFX7-NEXT:    v_and_b32_e32 v3, 0xffff0000, v3
; GFX7-NEXT:    v_and_b32_e32 v4, 0xffff0000, v4
; GFX7-NEXT:    v_and_b32_e32 v5, 0xffff0000, v5
; GFX7-NEXT:    s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_maximumnum_v6bf16:
; GFX8:       ; %bb.0:
; GFX8-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT:    v_and_b32_e32 v6, 0xffff0000, v2
; GFX8-NEXT:    v_lshrrev_b32_e32 v7, 16, v5
; GFX8-NEXT:    v_lshrrev_b32_e32 v8, 16, v2
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v6, v6
; GFX8-NEXT:    v_and_b32_e32 v9, 0xffff0000, v5
; GFX8-NEXT:    v_cndmask_b32_e32 v6, v8, v7, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v9, v9
; GFX8-NEXT:    v_cndmask_b32_e32 v7, v7, v6, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v8, 16, v6
; GFX8-NEXT:    v_lshlrev_b32_e32 v9, 16, v7
; GFX8-NEXT:    v_cmp_gt_f32_e32 vcc, v8, v9
; GFX8-NEXT:    v_cndmask_b32_e32 v8, v7, v6, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; GFX8-NEXT:    v_mul_f32_e32 v8, 1.0, v8
; GFX8-NEXT:    v_bfe_u32 v9, v8, 16, 1
; GFX8-NEXT:    v_add_u32_e32 v9, vcc, v9, v8
; GFX8-NEXT:    s_movk_i32 s4, 0x7fff
; GFX8-NEXT:    v_add_u32_e32 v9, vcc, s4, v9
; GFX8-NEXT:    v_or_b32_e32 v10, 0x400000, v8
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v8, v8
; GFX8-NEXT:    v_cndmask_b32_e32 v8, v9, v10, vcc
; GFX8-NEXT:    v_lshrrev_b32_e32 v9, 16, v8
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v6
; GFX8-NEXT:    v_cndmask_b32_e32 v6, v9, v6, vcc
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v7
; GFX8-NEXT:    v_cndmask_b32_e32 v6, v6, v7, vcc
; GFX8-NEXT:    v_and_b32_e32 v7, 0xffff0000, v8
; GFX8-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v7
; GFX8-NEXT:    v_and_b32_e32 v7, 0xffff0000, v1
; GFX8-NEXT:    v_cndmask_b32_e32 v6, v9, v6, vcc
; GFX8-NEXT:    v_lshrrev_b32_e32 v8, 16, v4
; GFX8-NEXT:    v_lshrrev_b32_e32 v9, 16, v1
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v7, v7
; GFX8-NEXT:    v_and_b32_e32 v10, 0xffff0000, v4
; GFX8-NEXT:    v_cndmask_b32_e32 v7, v9, v8, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v10, v10
; GFX8-NEXT:    v_cndmask_b32_e32 v8, v8, v7, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v9, 16, v7
; GFX8-NEXT:    v_lshlrev_b32_e32 v10, 16, v8
; GFX8-NEXT:    v_cmp_gt_f32_e32 vcc, v9, v10
; GFX8-NEXT:    v_cndmask_b32_e32 v9, v8, v7, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; GFX8-NEXT:    v_mul_f32_e32 v9, 1.0, v9
; GFX8-NEXT:    v_bfe_u32 v10, v9, 16, 1
; GFX8-NEXT:    v_add_u32_e32 v10, vcc, v10, v9
; GFX8-NEXT:    v_add_u32_e32 v10, vcc, s4, v10
; GFX8-NEXT:    v_or_b32_e32 v11, 0x400000, v9
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v9, v9
; GFX8-NEXT:    v_cndmask_b32_e32 v9, v10, v11, vcc
; GFX8-NEXT:    v_lshrrev_b32_e32 v10, 16, v9
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v7
; GFX8-NEXT:    v_cndmask_b32_e32 v7, v10, v7, vcc
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v8
; GFX8-NEXT:    v_cndmask_b32_e32 v7, v7, v8, vcc
; GFX8-NEXT:    v_and_b32_e32 v8, 0xffff0000, v9
; GFX8-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v8
; GFX8-NEXT:    v_and_b32_e32 v8, 0xffff0000, v0
; GFX8-NEXT:    v_cndmask_b32_e32 v7, v10, v7, vcc
; GFX8-NEXT:    v_lshrrev_b32_e32 v9, 16, v3
; GFX8-NEXT:    v_lshrrev_b32_e32 v10, 16, v0
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v8, v8
; GFX8-NEXT:    v_and_b32_e32 v11, 0xffff0000, v3
; GFX8-NEXT:    v_cndmask_b32_e32 v8, v10, v9, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v11, v11
; GFX8-NEXT:    v_cndmask_b32_e32 v9, v9, v8, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v10, 16, v8
; GFX8-NEXT:    v_lshlrev_b32_e32 v11, 16, v9
; GFX8-NEXT:    v_cmp_gt_f32_e32 vcc, v10, v11
; GFX8-NEXT:    v_cndmask_b32_e32 v10, v9, v8, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
; GFX8-NEXT:    v_mul_f32_e32 v10, 1.0, v10
; GFX8-NEXT:    v_bfe_u32 v11, v10, 16, 1
; GFX8-NEXT:    v_add_u32_e32 v11, vcc, v11, v10
; GFX8-NEXT:    v_add_u32_e32 v11, vcc, s4, v11
; GFX8-NEXT:    v_or_b32_e32 v12, 0x400000, v10
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v10, v10
; GFX8-NEXT:    v_cndmask_b32_e32 v10, v11, v12, vcc
; GFX8-NEXT:    v_lshrrev_b32_e32 v11, 16, v10
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v8
; GFX8-NEXT:    v_cndmask_b32_e32 v8, v11, v8, vcc
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v9
; GFX8-NEXT:    v_cndmask_b32_e32 v8, v8, v9, vcc
; GFX8-NEXT:    v_and_b32_e32 v9, 0xffff0000, v10
; GFX8-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v9
; GFX8-NEXT:    v_lshlrev_b32_e32 v9, 16, v2
; GFX8-NEXT:    v_cndmask_b32_e32 v8, v11, v8, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v9, v9
; GFX8-NEXT:    v_lshlrev_b32_e32 v9, 16, v5
; GFX8-NEXT:    v_cndmask_b32_e32 v2, v2, v5, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v9, v9
; GFX8-NEXT:    v_cndmask_b32_e32 v5, v5, v2, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v9, 16, v5
; GFX8-NEXT:    v_lshlrev_b32_e32 v10, 16, v2
; GFX8-NEXT:    v_cmp_gt_f32_e32 vcc, v10, v9
; GFX8-NEXT:    v_cndmask_b32_e32 v9, v5, v2, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; GFX8-NEXT:    v_mul_f32_e32 v9, 1.0, v9
; GFX8-NEXT:    v_bfe_u32 v10, v9, 16, 1
; GFX8-NEXT:    v_add_u32_e32 v10, vcc, v10, v9
; GFX8-NEXT:    v_add_u32_e32 v10, vcc, s4, v10
; GFX8-NEXT:    v_or_b32_e32 v11, 0x400000, v9
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v9, v9
; GFX8-NEXT:    v_cndmask_b32_e32 v9, v10, v11, vcc
; GFX8-NEXT:    v_lshrrev_b32_e32 v10, 16, v9
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v2
; GFX8-NEXT:    v_cndmask_b32_e32 v2, v10, v2, vcc
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v5
; GFX8-NEXT:    v_cndmask_b32_e32 v2, v2, v5, vcc
; GFX8-NEXT:    v_and_b32_e32 v5, 0xffff0000, v9
; GFX8-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v5
; GFX8-NEXT:    v_lshlrev_b32_e32 v5, 16, v1
; GFX8-NEXT:    v_cndmask_b32_e32 v2, v10, v2, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; GFX8-NEXT:    v_lshlrev_b32_e32 v5, 16, v4
; GFX8-NEXT:    v_cndmask_b32_e32 v1, v1, v4, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; GFX8-NEXT:    v_cndmask_b32_e32 v4, v4, v1, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v5, 16, v4
; GFX8-NEXT:    v_lshlrev_b32_e32 v9, 16, v1
; GFX8-NEXT:    v_cmp_gt_f32_e32 vcc, v9, v5
; GFX8-NEXT:    v_cndmask_b32_e32 v5, v4, v1, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; GFX8-NEXT:    v_mul_f32_e32 v5, 1.0, v5
; GFX8-NEXT:    v_bfe_u32 v9, v5, 16, 1
; GFX8-NEXT:    v_add_u32_e32 v9, vcc, v9, v5
; GFX8-NEXT:    v_add_u32_e32 v9, vcc, s4, v9
; GFX8-NEXT:    v_or_b32_e32 v10, 0x400000, v5
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; GFX8-NEXT:    v_cndmask_b32_e32 v5, v9, v10, vcc
; GFX8-NEXT:    v_lshrrev_b32_e32 v9, 16, v5
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v1
; GFX8-NEXT:    v_cndmask_b32_e32 v1, v9, v1, vcc
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v4
; GFX8-NEXT:    v_cndmask_b32_e32 v1, v1, v4, vcc
; GFX8-NEXT:    v_and_b32_e32 v4, 0xffff0000, v5
; GFX8-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v4
; GFX8-NEXT:    v_lshlrev_b32_e32 v4, 16, v0
; GFX8-NEXT:    v_cndmask_b32_e32 v1, v9, v1, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v4, v4
; GFX8-NEXT:    v_lshlrev_b32_e32 v4, 16, v3
; GFX8-NEXT:    v_cndmask_b32_e32 v0, v0, v3, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v4, v4
; GFX8-NEXT:    v_cndmask_b32_e32 v3, v3, v0, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v4, 16, v3
; GFX8-NEXT:    v_lshlrev_b32_e32 v5, 16, v0
; GFX8-NEXT:    v_cmp_gt_f32_e32 vcc, v5, v4
; GFX8-NEXT:    v_cndmask_b32_e32 v4, v3, v0, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; GFX8-NEXT:    v_mul_f32_e32 v4, 1.0, v4
; GFX8-NEXT:    v_bfe_u32 v5, v4, 16, 1
; GFX8-NEXT:    v_add_u32_e32 v5, vcc, v5, v4
; GFX8-NEXT:    v_add_u32_e32 v5, vcc, s4, v5
; GFX8-NEXT:    v_or_b32_e32 v9, 0x400000, v4
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v4, v4
; GFX8-NEXT:    v_cndmask_b32_e32 v4, v5, v9, vcc
; GFX8-NEXT:    v_lshrrev_b32_e32 v5, 16, v4
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v0
; GFX8-NEXT:    v_cndmask_b32_e32 v0, v5, v0, vcc
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v3
; GFX8-NEXT:    v_cndmask_b32_e32 v0, v0, v3, vcc
; GFX8-NEXT:    v_and_b32_e32 v3, 0xffff0000, v4
; GFX8-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v3
; GFX8-NEXT:    v_cndmask_b32_e32 v0, v5, v0, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v3, 16, v8
; GFX8-NEXT:    v_or_b32_sdwa v0, v0, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX8-NEXT:    v_lshlrev_b32_e32 v3, 16, v7
; GFX8-NEXT:    v_or_b32_sdwa v1, v1, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX8-NEXT:    v_lshlrev_b32_e32 v3, 16, v6
; GFX8-NEXT:    v_or_b32_sdwa v2, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX8-NEXT:    s_setpc_b64 s[30:31]
;
; GFX900-LABEL: v_maximumnum_v6bf16:
; GFX900:       ; %bb.0:
; GFX900-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX900-NEXT:    v_and_b32_e32 v6, 0xffff0000, v2
; GFX900-NEXT:    v_lshrrev_b32_e32 v7, 16, v5
; GFX900-NEXT:    v_lshrrev_b32_e32 v8, 16, v2
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v6, v6
; GFX900-NEXT:    v_and_b32_e32 v9, 0xffff0000, v5
; GFX900-NEXT:    v_cndmask_b32_e32 v6, v8, v7, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v9, v9
; GFX900-NEXT:    v_cndmask_b32_e32 v7, v7, v6, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v8, 16, v6
; GFX900-NEXT:    v_lshlrev_b32_e32 v9, 16, v7
; GFX900-NEXT:    v_cmp_gt_f32_e32 vcc, v8, v9
; GFX900-NEXT:    v_cndmask_b32_e32 v8, v7, v6, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; GFX900-NEXT:    v_max_f32_e32 v8, v8, v8
; GFX900-NEXT:    v_bfe_u32 v9, v8, 16, 1
; GFX900-NEXT:    s_movk_i32 s4, 0x7fff
; GFX900-NEXT:    v_add3_u32 v9, v9, v8, s4
; GFX900-NEXT:    v_or_b32_e32 v10, 0x400000, v8
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v8, v8
; GFX900-NEXT:    v_cndmask_b32_e32 v8, v9, v10, vcc
; GFX900-NEXT:    v_lshrrev_b32_e32 v9, 16, v8
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v6
; GFX900-NEXT:    v_cndmask_b32_e32 v6, v9, v6, vcc
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v7
; GFX900-NEXT:    v_cndmask_b32_e32 v6, v6, v7, vcc
; GFX900-NEXT:    v_and_b32_e32 v7, 0xffff0000, v8
; GFX900-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v7
; GFX900-NEXT:    v_and_b32_e32 v7, 0xffff0000, v1
; GFX900-NEXT:    v_cndmask_b32_e32 v6, v9, v6, vcc
; GFX900-NEXT:    v_lshrrev_b32_e32 v8, 16, v4
; GFX900-NEXT:    v_lshrrev_b32_e32 v9, 16, v1
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v7, v7
; GFX900-NEXT:    v_and_b32_e32 v10, 0xffff0000, v4
; GFX900-NEXT:    v_cndmask_b32_e32 v7, v9, v8, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v10, v10
; GFX900-NEXT:    v_cndmask_b32_e32 v8, v8, v7, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v9, 16, v7
; GFX900-NEXT:    v_lshlrev_b32_e32 v10, 16, v8
; GFX900-NEXT:    v_cmp_gt_f32_e32 vcc, v9, v10
; GFX900-NEXT:    v_cndmask_b32_e32 v9, v8, v7, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; GFX900-NEXT:    v_max_f32_e32 v9, v9, v9
; GFX900-NEXT:    v_bfe_u32 v10, v9, 16, 1
; GFX900-NEXT:    v_add3_u32 v10, v10, v9, s4
; GFX900-NEXT:    v_or_b32_e32 v11, 0x400000, v9
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v9, v9
; GFX900-NEXT:    v_cndmask_b32_e32 v9, v10, v11, vcc
; GFX900-NEXT:    v_lshrrev_b32_e32 v10, 16, v9
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v7
; GFX900-NEXT:    v_cndmask_b32_e32 v7, v10, v7, vcc
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v8
; GFX900-NEXT:    v_cndmask_b32_e32 v7, v7, v8, vcc
; GFX900-NEXT:    v_and_b32_e32 v8, 0xffff0000, v9
; GFX900-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v8
; GFX900-NEXT:    v_and_b32_e32 v8, 0xffff0000, v0
; GFX900-NEXT:    v_cndmask_b32_e32 v7, v10, v7, vcc
; GFX900-NEXT:    v_lshrrev_b32_e32 v9, 16, v3
; GFX900-NEXT:    v_lshrrev_b32_e32 v10, 16, v0
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v8, v8
; GFX900-NEXT:    v_and_b32_e32 v11, 0xffff0000, v3
; GFX900-NEXT:    v_cndmask_b32_e32 v8, v10, v9, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v11, v11
; GFX900-NEXT:    v_cndmask_b32_e32 v9, v9, v8, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v10, 16, v8
; GFX900-NEXT:    v_lshlrev_b32_e32 v11, 16, v9
; GFX900-NEXT:    v_cmp_gt_f32_e32 vcc, v10, v11
; GFX900-NEXT:    v_cndmask_b32_e32 v10, v9, v8, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
; GFX900-NEXT:    v_max_f32_e32 v10, v10, v10
; GFX900-NEXT:    v_bfe_u32 v11, v10, 16, 1
; GFX900-NEXT:    v_add3_u32 v11, v11, v10, s4
; GFX900-NEXT:    v_or_b32_e32 v12, 0x400000, v10
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v10, v10
; GFX900-NEXT:    v_cndmask_b32_e32 v10, v11, v12, vcc
; GFX900-NEXT:    v_lshrrev_b32_e32 v11, 16, v10
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v8
; GFX900-NEXT:    v_cndmask_b32_e32 v8, v11, v8, vcc
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v9
; GFX900-NEXT:    v_cndmask_b32_e32 v8, v8, v9, vcc
; GFX900-NEXT:    v_and_b32_e32 v9, 0xffff0000, v10
; GFX900-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v9
; GFX900-NEXT:    v_lshlrev_b32_e32 v9, 16, v2
; GFX900-NEXT:    v_cndmask_b32_e32 v8, v11, v8, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v9, v9
; GFX900-NEXT:    v_lshlrev_b32_e32 v9, 16, v5
; GFX900-NEXT:    v_cndmask_b32_e32 v2, v2, v5, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v9, v9
; GFX900-NEXT:    v_cndmask_b32_e32 v5, v5, v2, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v9, 16, v5
; GFX900-NEXT:    v_lshlrev_b32_e32 v10, 16, v2
; GFX900-NEXT:    v_cmp_gt_f32_e32 vcc, v10, v9
; GFX900-NEXT:    v_cndmask_b32_e32 v9, v5, v2, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; GFX900-NEXT:    v_max_f32_e32 v9, v9, v9
; GFX900-NEXT:    v_bfe_u32 v10, v9, 16, 1
; GFX900-NEXT:    v_add3_u32 v10, v10, v9, s4
; GFX900-NEXT:    v_or_b32_e32 v11, 0x400000, v9
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v9, v9
; GFX900-NEXT:    v_cndmask_b32_e32 v9, v10, v11, vcc
; GFX900-NEXT:    v_lshrrev_b32_e32 v10, 16, v9
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v2
; GFX900-NEXT:    v_cndmask_b32_e32 v2, v10, v2, vcc
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v5
; GFX900-NEXT:    v_cndmask_b32_e32 v2, v2, v5, vcc
; GFX900-NEXT:    v_and_b32_e32 v5, 0xffff0000, v9
; GFX900-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v5
; GFX900-NEXT:    v_lshlrev_b32_e32 v5, 16, v1
; GFX900-NEXT:    v_cndmask_b32_e32 v2, v10, v2, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; GFX900-NEXT:    v_lshlrev_b32_e32 v5, 16, v4
; GFX900-NEXT:    v_cndmask_b32_e32 v1, v1, v4, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; GFX900-NEXT:    v_cndmask_b32_e32 v4, v4, v1, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v5, 16, v4
; GFX900-NEXT:    v_lshlrev_b32_e32 v9, 16, v1
; GFX900-NEXT:    v_cmp_gt_f32_e32 vcc, v9, v5
; GFX900-NEXT:    v_cndmask_b32_e32 v5, v4, v1, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; GFX900-NEXT:    v_max_f32_e32 v5, v5, v5
; GFX900-NEXT:    v_bfe_u32 v9, v5, 16, 1
; GFX900-NEXT:    v_add3_u32 v9, v9, v5, s4
; GFX900-NEXT:    v_or_b32_e32 v10, 0x400000, v5
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; GFX900-NEXT:    v_cndmask_b32_e32 v5, v9, v10, vcc
; GFX900-NEXT:    v_lshrrev_b32_e32 v9, 16, v5
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v1
; GFX900-NEXT:    v_cndmask_b32_e32 v1, v9, v1, vcc
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v4
; GFX900-NEXT:    v_cndmask_b32_e32 v1, v1, v4, vcc
; GFX900-NEXT:    v_and_b32_e32 v4, 0xffff0000, v5
; GFX900-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v4
; GFX900-NEXT:    v_lshlrev_b32_e32 v4, 16, v0
; GFX900-NEXT:    v_cndmask_b32_e32 v1, v9, v1, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v4, v4
; GFX900-NEXT:    v_lshlrev_b32_e32 v4, 16, v3
; GFX900-NEXT:    v_cndmask_b32_e32 v0, v0, v3, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v4, v4
; GFX900-NEXT:    v_cndmask_b32_e32 v3, v3, v0, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v4, 16, v3
; GFX900-NEXT:    v_lshlrev_b32_e32 v5, 16, v0
; GFX900-NEXT:    v_cmp_gt_f32_e32 vcc, v5, v4
; GFX900-NEXT:    v_cndmask_b32_e32 v4, v3, v0, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; GFX900-NEXT:    v_max_f32_e32 v4, v4, v4
; GFX900-NEXT:    v_bfe_u32 v5, v4, 16, 1
; GFX900-NEXT:    v_add3_u32 v5, v5, v4, s4
; GFX900-NEXT:    v_or_b32_e32 v9, 0x400000, v4
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v4, v4
; GFX900-NEXT:    v_cndmask_b32_e32 v4, v5, v9, vcc
; GFX900-NEXT:    v_lshrrev_b32_e32 v5, 16, v4
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v0
; GFX900-NEXT:    v_cndmask_b32_e32 v0, v5, v0, vcc
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v3
; GFX900-NEXT:    v_cndmask_b32_e32 v0, v0, v3, vcc
; GFX900-NEXT:    v_and_b32_e32 v3, 0xffff0000, v4
; GFX900-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v3
; GFX900-NEXT:    v_cndmask_b32_e32 v0, v5, v0, vcc
; GFX900-NEXT:    s_mov_b32 s4, 0x5040100
; GFX900-NEXT:    v_perm_b32 v0, v8, v0, s4
; GFX900-NEXT:    v_perm_b32 v1, v7, v1, s4
; GFX900-NEXT:    v_perm_b32 v2, v6, v2, s4
; GFX900-NEXT:    s_setpc_b64 s[30:31]
;
; GFX950-LABEL: v_maximumnum_v6bf16:
; GFX950:       ; %bb.0:
; GFX950-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX950-NEXT:    v_and_b32_e32 v6, 0xffff0000, v2
; GFX950-NEXT:    v_lshrrev_b32_e32 v7, 16, v5
; GFX950-NEXT:    v_lshrrev_b32_e32 v8, 16, v2
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v6, v6
; GFX950-NEXT:    v_and_b32_e32 v9, 0xffff0000, v5
; GFX950-NEXT:    v_and_b32_e32 v10, 0xffff0000, v4
; GFX950-NEXT:    v_cndmask_b32_e32 v6, v8, v7, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v9, v9
; GFX950-NEXT:    v_lshlrev_b32_e32 v8, 16, v6
; GFX950-NEXT:    v_and_b32_e32 v11, 0xffff0000, v3
; GFX950-NEXT:    v_cndmask_b32_e32 v7, v7, v6, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v9, 16, v7
; GFX950-NEXT:    v_cmp_gt_f32_e32 vcc, v8, v9
; GFX950-NEXT:    v_lshrrev_b32_e32 v9, 16, v1
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v8, v7, v6, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; GFX950-NEXT:    v_max_f32_e32 v8, v8, v8
; GFX950-NEXT:    v_cvt_pk_bf16_f32 v8, v8, s0
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v6
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v6, v8, v6, vcc
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v7
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v6, v6, v7, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v7, 16, v8
; GFX950-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v7
; GFX950-NEXT:    v_and_b32_e32 v7, 0xffff0000, v1
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v6, v8, v6, vcc
; GFX950-NEXT:    v_lshrrev_b32_e32 v8, 16, v4
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v7, v7
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v7, v9, v8, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v10, v10
; GFX950-NEXT:    v_lshlrev_b32_e32 v9, 16, v7
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v8, v8, v7, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v10, 16, v8
; GFX950-NEXT:    v_cmp_gt_f32_e32 vcc, v9, v10
; GFX950-NEXT:    v_lshrrev_b32_e32 v10, 16, v0
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v9, v8, v7, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; GFX950-NEXT:    v_max_f32_e32 v9, v9, v9
; GFX950-NEXT:    v_cvt_pk_bf16_f32 v9, v9, s0
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v7
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v7, v9, v7, vcc
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v8
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v7, v7, v8, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v8, 16, v9
; GFX950-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v8
; GFX950-NEXT:    v_and_b32_e32 v8, 0xffff0000, v0
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v7, v9, v7, vcc
; GFX950-NEXT:    v_lshrrev_b32_e32 v9, 16, v3
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v8, v8
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v8, v10, v9, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v11, v11
; GFX950-NEXT:    v_lshlrev_b32_e32 v10, 16, v8
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v9, v9, v8, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v11, 16, v9
; GFX950-NEXT:    v_cmp_gt_f32_e32 vcc, v10, v11
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v10, v9, v8, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
; GFX950-NEXT:    v_max_f32_e32 v10, v10, v10
; GFX950-NEXT:    v_cvt_pk_bf16_f32 v10, v10, s0
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v8
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v8, v10, v8, vcc
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v9
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v8, v8, v9, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v9, 16, v10
; GFX950-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v9
; GFX950-NEXT:    v_lshlrev_b32_e32 v9, 16, v2
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v8, v10, v8, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v9, v9
; GFX950-NEXT:    v_lshlrev_b32_e32 v9, 16, v5
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v2, v2, v5, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v9, v9
; GFX950-NEXT:    v_lshlrev_b32_e32 v10, 16, v2
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v5, v5, v2, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v9, 16, v5
; GFX950-NEXT:    v_cmp_gt_f32_e32 vcc, v10, v9
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v9, v5, v2, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; GFX950-NEXT:    v_max_f32_e32 v9, v9, v9
; GFX950-NEXT:    v_cvt_pk_bf16_f32 v9, v9, s0
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v2
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v2, v9, v2, vcc
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v5
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v2, v2, v5, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v5, 16, v9
; GFX950-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v5
; GFX950-NEXT:    v_lshlrev_b32_e32 v5, 16, v1
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v2, v9, v2, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; GFX950-NEXT:    v_lshlrev_b32_e32 v5, 16, v4
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v1, v1, v4, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; GFX950-NEXT:    v_lshlrev_b32_e32 v9, 16, v1
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v4, v4, v1, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v5, 16, v4
; GFX950-NEXT:    v_cmp_gt_f32_e32 vcc, v9, v5
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v5, v4, v1, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; GFX950-NEXT:    v_max_f32_e32 v5, v5, v5
; GFX950-NEXT:    v_cvt_pk_bf16_f32 v5, v5, s0
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v1
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v1, v5, v1, vcc
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v4
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v1, v1, v4, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v4, 16, v5
; GFX950-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v4
; GFX950-NEXT:    v_lshlrev_b32_e32 v4, 16, v0
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v1, v5, v1, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v4, v4
; GFX950-NEXT:    v_lshlrev_b32_e32 v4, 16, v3
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v0, v0, v3, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v4, v4
; GFX950-NEXT:    v_lshlrev_b32_e32 v5, 16, v0
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v3, v3, v0, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v4, 16, v3
; GFX950-NEXT:    v_cmp_gt_f32_e32 vcc, v5, v4
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v4, v3, v0, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; GFX950-NEXT:    v_max_f32_e32 v4, v4, v4
; GFX950-NEXT:    v_cvt_pk_bf16_f32 v4, v4, s0
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v0
; GFX950-NEXT:    s_mov_b32 s0, 0x5040100
; GFX950-NEXT:    v_perm_b32 v1, v7, v1, s0
; GFX950-NEXT:    v_cndmask_b32_e32 v0, v4, v0, vcc
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v3
; GFX950-NEXT:    v_perm_b32 v2, v6, v2, s0
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v0, v0, v3, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v3, 16, v4
; GFX950-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v3
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v0, v4, v0, vcc
; GFX950-NEXT:    v_perm_b32 v0, v8, v0, s0
; GFX950-NEXT:    s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_maximumnum_v6bf16:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT:    v_and_b32_e32 v6, 0xffff0000, v2
; GFX10-NEXT:    v_lshrrev_b32_e32 v7, 16, v5
; GFX10-NEXT:    v_and_b32_e32 v8, 0xffff0000, v5
; GFX10-NEXT:    v_lshrrev_b32_e32 v9, 16, v4
; GFX10-NEXT:    v_lshrrev_b32_e32 v10, 16, v1
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX10-NEXT:    v_and_b32_e32 v6, 0xffff0000, v1
; GFX10-NEXT:    v_and_b32_e32 v11, 0xffff0000, v4
; GFX10-NEXT:    v_and_b32_e32 v13, 0xffff0000, v0
; GFX10-NEXT:    v_lshrrev_b32_e32 v14, 16, v0
; GFX10-NEXT:    v_cndmask_b32_sdwa v12, v2, v7, vcc_lo dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX10-NEXT:    v_and_b32_e32 v15, 0xffff0000, v3
; GFX10-NEXT:    v_cndmask_b32_e32 v6, v10, v9, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
; GFX10-NEXT:    v_lshlrev_b32_e32 v10, 16, v12
; GFX10-NEXT:    v_lshrrev_b32_e32 v8, 16, v3
; GFX10-NEXT:    v_lshlrev_b32_e32 v16, 16, v6
; GFX10-NEXT:    v_cndmask_b32_e32 v7, v7, v12, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v11, v11
; GFX10-NEXT:    v_lshlrev_b32_e32 v11, 16, v7
; GFX10-NEXT:    v_cndmask_b32_e32 v9, v9, v6, vcc_lo
; GFX10-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v10, v11
; GFX10-NEXT:    v_lshlrev_b32_e32 v17, 16, v9
; GFX10-NEXT:    v_cndmask_b32_e32 v10, v7, v12, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v13, v13
; GFX10-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
; GFX10-NEXT:    v_cndmask_b32_e32 v11, v14, v8, vcc_lo
; GFX10-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v16, v17
; GFX10-NEXT:    v_max_f32_e32 v10, v10, v10
; GFX10-NEXT:    v_lshlrev_b32_e32 v14, 16, v11
; GFX10-NEXT:    v_cndmask_b32_e32 v13, v9, v6, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v15, v15
; GFX10-NEXT:    v_bfe_u32 v16, v10, 16, 1
; GFX10-NEXT:    v_or_b32_e32 v18, 0x400000, v10
; GFX10-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
; GFX10-NEXT:    v_cndmask_b32_e32 v8, v8, v11, vcc_lo
; GFX10-NEXT:    v_add3_u32 v16, v16, v10, 0x7fff
; GFX10-NEXT:    v_max_f32_e32 v13, v13, v13
; GFX10-NEXT:    v_lshlrev_b32_e32 v15, 16, v8
; GFX10-NEXT:    v_bfe_u32 v17, v13, 16, 1
; GFX10-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v14, v15
; GFX10-NEXT:    v_lshlrev_b32_e32 v15, 16, v2
; GFX10-NEXT:    v_or_b32_e32 v19, 0x400000, v13
; GFX10-NEXT:    v_add3_u32 v17, v17, v13, 0x7fff
; GFX10-NEXT:    v_cndmask_b32_e32 v14, v8, v11, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v10, v10
; GFX10-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
; GFX10-NEXT:    v_cndmask_b32_e32 v10, v16, v18, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v13, v13
; GFX10-NEXT:    v_lshlrev_b32_e32 v16, 16, v5
; GFX10-NEXT:    v_max_f32_e32 v14, v14, v14
; GFX10-NEXT:    v_cndmask_b32_e32 v13, v17, v19, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v15, v15
; GFX10-NEXT:    v_lshrrev_b32_e32 v17, 16, v10
; GFX10-NEXT:    v_and_b32_e32 v10, 0xffff0000, v10
; GFX10-NEXT:    v_bfe_u32 v18, v14, 16, 1
; GFX10-NEXT:    v_lshrrev_b32_e32 v15, 16, v13
; GFX10-NEXT:    v_cndmask_b32_e32 v2, v2, v5, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v12
; GFX10-NEXT:    v_cndmask_b32_e32 v12, v17, v12, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v6
; GFX10-NEXT:    v_cndmask_b32_e32 v6, v15, v6, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v7
; GFX10-NEXT:    v_cndmask_b32_e32 v7, v12, v7, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v9
; GFX10-NEXT:    v_and_b32_e32 v12, 0xffff0000, v13
; GFX10-NEXT:    v_cndmask_b32_e32 v6, v6, v9, vcc_lo
; GFX10-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v10
; GFX10-NEXT:    v_add3_u32 v9, v18, v14, 0x7fff
; GFX10-NEXT:    v_or_b32_e32 v10, 0x400000, v14
; GFX10-NEXT:    v_cndmask_b32_e32 v7, v17, v7, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v16, v16
; GFX10-NEXT:    v_lshlrev_b32_e32 v17, 16, v4
; GFX10-NEXT:    v_lshlrev_b32_e32 v16, 16, v2
; GFX10-NEXT:    v_cndmask_b32_e32 v5, v5, v2, vcc_lo
; GFX10-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v12
; GFX10-NEXT:    v_lshlrev_b32_e32 v12, 16, v1
; GFX10-NEXT:    v_lshlrev_b32_e32 v13, 16, v5
; GFX10-NEXT:    v_cndmask_b32_e32 v6, v15, v6, vcc_lo
; GFX10-NEXT:    v_lshlrev_b32_e32 v15, 16, v0
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v12, v12
; GFX10-NEXT:    v_lshlrev_b32_e32 v12, 16, v3
; GFX10-NEXT:    v_cndmask_b32_e32 v1, v1, v4, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v15, v15
; GFX10-NEXT:    v_lshlrev_b32_e32 v15, 16, v1
; GFX10-NEXT:    v_cndmask_b32_e32 v0, v0, v3, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v17, v17
; GFX10-NEXT:    v_cndmask_b32_e32 v4, v4, v1, vcc_lo
; GFX10-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v16, v13
; GFX10-NEXT:    v_cndmask_b32_e32 v13, v5, v2, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v12, v12
; GFX10-NEXT:    v_lshlrev_b32_e32 v12, 16, v4
; GFX10-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
; GFX10-NEXT:    v_cndmask_b32_e32 v3, v3, v0, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v14, v14
; GFX10-NEXT:    v_lshlrev_b32_e32 v14, 16, v0
; GFX10-NEXT:    v_max_f32_e32 v13, v13, v13
; GFX10-NEXT:    v_cndmask_b32_e32 v9, v9, v10, vcc_lo
; GFX10-NEXT:    v_lshlrev_b32_e32 v10, 16, v3
; GFX10-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v15, v12
; GFX10-NEXT:    v_lshrrev_b32_e32 v15, 16, v9
; GFX10-NEXT:    v_and_b32_e32 v9, 0xffff0000, v9
; GFX10-NEXT:    v_cndmask_b32_e32 v12, v4, v1, vcc_lo
; GFX10-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v14, v10
; GFX10-NEXT:    v_bfe_u32 v14, v13, 16, 1
; GFX10-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; GFX10-NEXT:    v_cndmask_b32_e32 v10, v3, v0, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v11
; GFX10-NEXT:    v_max_f32_e32 v12, v12, v12
; GFX10-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
; GFX10-NEXT:    v_cndmask_b32_e32 v11, v15, v11, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v8
; GFX10-NEXT:    v_bfe_u32 v16, v12, 16, 1
; GFX10-NEXT:    v_max_f32_e32 v10, v10, v10
; GFX10-NEXT:    v_cndmask_b32_e32 v8, v11, v8, vcc_lo
; GFX10-NEXT:    v_add3_u32 v11, v14, v13, 0x7fff
; GFX10-NEXT:    v_or_b32_e32 v14, 0x400000, v13
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v13, v13
; GFX10-NEXT:    v_bfe_u32 v17, v10, 16, 1
; GFX10-NEXT:    v_add3_u32 v13, v16, v12, 0x7fff
; GFX10-NEXT:    v_cndmask_b32_e32 v11, v11, v14, vcc_lo
; GFX10-NEXT:    v_or_b32_e32 v14, 0x400000, v12
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v12, v12
; GFX10-NEXT:    v_add3_u32 v16, v17, v10, 0x7fff
; GFX10-NEXT:    v_or_b32_e32 v17, 0x400000, v10
; GFX10-NEXT:    v_lshrrev_b32_e32 v18, 16, v11
; GFX10-NEXT:    v_cndmask_b32_e32 v12, v13, v14, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v10, v10
; GFX10-NEXT:    v_cndmask_b32_e32 v10, v16, v17, vcc_lo
; GFX10-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v9
; GFX10-NEXT:    v_lshrrev_b32_e32 v9, 16, v12
; GFX10-NEXT:    v_lshrrev_b32_e32 v13, 16, v10
; GFX10-NEXT:    v_cndmask_b32_e32 v8, v15, v8, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v2
; GFX10-NEXT:    v_cndmask_b32_e32 v2, v18, v2, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v1
; GFX10-NEXT:    v_cndmask_b32_e32 v1, v9, v1, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v5
; GFX10-NEXT:    v_cndmask_b32_e32 v2, v2, v5, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v0
; GFX10-NEXT:    v_and_b32_e32 v5, 0xffff0000, v10
; GFX10-NEXT:    v_cndmask_b32_e32 v0, v13, v0, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v4
; GFX10-NEXT:    v_cndmask_b32_e32 v1, v1, v4, vcc_lo
; GFX10-NEXT:    v_and_b32_e32 v4, 0xffff0000, v12
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v3
; GFX10-NEXT:    v_cndmask_b32_e32 v0, v0, v3, vcc_lo
; GFX10-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v4
; GFX10-NEXT:    v_and_b32_e32 v3, 0xffff0000, v11
; GFX10-NEXT:    v_cndmask_b32_e32 v1, v9, v1, vcc_lo
; GFX10-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v5
; GFX10-NEXT:    v_perm_b32 v1, v6, v1, 0x5040100
; GFX10-NEXT:    v_cndmask_b32_e32 v0, v13, v0, vcc_lo
; GFX10-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v3
; GFX10-NEXT:    v_perm_b32 v0, v8, v0, 0x5040100
; GFX10-NEXT:    v_cndmask_b32_e32 v2, v18, v2, vcc_lo
; GFX10-NEXT:    v_perm_b32 v2, v7, v2, 0x5040100
; GFX10-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-TRUE16-LABEL: v_maximumnum_v6bf16:
; GFX11-TRUE16:       ; %bb.0:
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xffff0000, v2
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xffff0000, v5
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff0000, v1
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xffff0000, v4
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xffff0000, v0
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s1, v8, v8
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s0, v7, v7
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 16, v2
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xffff0000, v3
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v6.l, v2.h, v5.h, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v7.l, v1.h, v4.h, s0
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s0, v10, v10
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s3, v12, v12
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v8.l, v5.h, v6.l, s1
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v15.l, v6.l
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v9.l, v4.h, v7.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.l, v7.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v14, 16, v1
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v16.l, v8.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.l, v9.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s1, v11, v11
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s4, v14, v14
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v13, 16, v5
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v20, 16, v0
; GFX11-TRUE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v15, v16
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v2.l, v2.l, v5.l, s3
; GFX11-TRUE16-NEXT:    v_cmp_gt_f32_e64 s2, v18, v19
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s3, 0, v7.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v17, 16, v4
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v10.l, v8.l, v6.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v13, v13
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v11.l, v9.l, v7.l, s2
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s2, v20, v20
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v21, 16, v3
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v12.l, v10.l
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v10.l, v0.h, v3.h, s0
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v14, 16, v11
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s0, 0, v6.l
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v5.l, v5.l, v2.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v11.l, v3.h, v10.l, s1
; GFX11-TRUE16-NEXT:    v_max_f32_e32 v14, v14, v14
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v13.l, v10.l
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v8.l
; GFX11-TRUE16-NEXT:    v_max_f32_e32 v12, v12, v12
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.l, v11.l
; GFX11-TRUE16-NEXT:    v_bfe_u32 v19, v14, 16, 1
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v22, 0x400000, v14
; GFX11-TRUE16-NEXT:    v_bfe_u32 v16, v12, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v20, 0x400000, v12
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s1, v12, v12
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
; GFX11-TRUE16-NEXT:    v_add3_u32 v19, v19, v14, 0x7fff
; GFX11-TRUE16-NEXT:    v_add3_u32 v16, v16, v12, 0x7fff
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s6, v14, v14
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v15.l, v2.l
; GFX11-TRUE16-NEXT:    v_cmp_gt_f32_e64 s5, v13, v18
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v1.l, v1.l, v4.l, s4
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e64 v12, v16, v20, s1
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e64 v13, v19, v22, s6
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s1, 0, v9.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v16.l, v5.l
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v0.l, v0.l, v3.l, s2
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v0.h, v12.h, v6.l, s0
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xffff0000, v12
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v1.h, v13.h, v7.l, s3
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v6.l, v11.l, v10.l, s5
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 16, v15
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v0.h, v0.h, v8.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v14
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v1.h, v1.h, v9.l, s1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s1, v21, v21
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 16, v16
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v12.l, v1.l
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v2.h, v12.h, v0.h, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v17, v17
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v3.l, v3.l, v0.l, s1
; GFX11-TRUE16-NEXT:    v_cmp_gt_f32_e64 s0, v7, v8
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 16, v6
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v15.l, v0.l
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v4.l, v4.l, v1.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v14.l, v3.l
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v6.l, v5.l, v2.l, s0
; GFX11-TRUE16-NEXT:    v_dual_max_f32 v7, v7, v7 :: v_dual_lshlrev_b32 v12, 16, v12
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v9.l, v4.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v16, v7, 16, 1
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xffff0000, v13
; GFX11-TRUE16-NEXT:    v_cmp_gt_f32_e64 s0, v15, v14
; GFX11-TRUE16-NEXT:    v_max_f32_e32 v14, v6, v6
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v12, v9
; GFX11-TRUE16-NEXT:    v_add3_u32 v9, v16, v7, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, 0x400000, v7
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s1, v14, v14
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v6.l, v4.l, v1.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v7.l, v3.l, v0.l, s0
; GFX11-TRUE16-NEXT:    v_cmp_eq_f32_e64 s0, 0, v8
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v9, v9, v12 :: v_dual_lshlrev_b32 v6, 16, v6
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v10.l
; GFX11-TRUE16-NEXT:    v_bfe_u32 v12, v14, 16, 1
; GFX11-TRUE16-NEXT:    v_dual_max_f32 v6, v6, v6 :: v_dual_lshlrev_b32 v7, 16, v7
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v1.h, v13.h, v1.h, s0
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v0.h, v9.h, v10.l, vcc_lo
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_add3_u32 v8, v12, v14, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, 0x400000, v14
; GFX11-TRUE16-NEXT:    v_max_f32_e32 v7, v7, v7
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v11.l
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, 0x400000, v6
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e64 v8, v8, v10, s1
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s1, 0, v2.l
; GFX11-TRUE16-NEXT:    v_bfe_u32 v10, v6, 16, 1
; GFX11-TRUE16-NEXT:    v_bfe_u32 v12, v7, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v15, 0x400000, v7
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s2, v7, v7
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v2.l, v8.h, v2.l, s1
; GFX11-TRUE16-NEXT:    v_add3_u32 v10, v10, v6, 0x7fff
; GFX11-TRUE16-NEXT:    v_add3_u32 v12, v12, v7, 0x7fff
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s1, v6, v6
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v0.h, v0.h, v11.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xffff0000, v9
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v5.l
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e64 v7, v12, v15, s2
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e64 v6, v10, v14, s1
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s1, 0, v1.l
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s2, 0, v0.l
; GFX11-TRUE16-NEXT:    v_cmp_eq_f32_e64 s3, 0, v11
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v2.l, v2.l, v5.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v4.l
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v1.l, v6.h, v1.l, s1
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v0.l, v7.h, v0.l, s2
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s1, 0, v3.l
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xffff0000, v7
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xffff0000, v6
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xffff0000, v8
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v1.l, v1.l, v4.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v0.l, v0.l, v3.l, s1
; GFX11-TRUE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v5
; GFX11-TRUE16-NEXT:    v_cmp_eq_f32_e64 s1, 0, v10
; GFX11-TRUE16-NEXT:    v_cmp_eq_f32_e64 s2, 0, v11
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v0.h, v9.h, v0.h, s3
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v0.l, v7.h, v0.l, vcc_lo
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v1.l, v6.h, v1.l, s1
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v2.l, v8.h, v2.l, s2
; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-FAKE16-LABEL: v_maximumnum_v6bf16:
; GFX11-FAKE16:       ; %bb.0:
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v6, 0xffff0000, v2
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v7, 16, v5
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v8, 16, v2
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v10, 16, v4
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v11, 16, v1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v9, 0xffff0000, v5
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v14, 16, v3
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v15, 16, v0
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v16, 0xffff0000, v3
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v6, v8, v7, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v7, v7, v6 :: v_dual_and_b32 v8, 0xffff0000, v1
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v12, 0xffff0000, v4
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v8, v11, v10 :: v_dual_lshlrev_b32 v11, 16, v7
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v12, v12
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v9, 16, v6
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v12, 16, v8
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v10, v10, v8, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v9, v11
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v11, 0xffff0000, v0
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v13, 16, v10
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v9, v7, v6, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v12, v13
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v12, v10, v8, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v11, v11
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v11, v15, v14, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v16, v16
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v15, 16, v11
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v14, v14, v11 :: v_dual_max_f32 v9, v9, v9
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v17, 16, v14
; GFX11-FAKE16-NEXT:    v_bfe_u32 v13, v9, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v16, 0x400000, v9
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_add3_u32 v13, v13, v9, 0x7fff
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v9, v13, v16, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v15, v17
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v13, v14, v11 :: v_dual_lshlrev_b32 v12, 16, v12
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_max_f32_e32 v12, v12, v12
; GFX11-FAKE16-NEXT:    v_bfe_u32 v18, v12, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v15, 0x400000, v12
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v12, v12
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_add3_u32 v16, v18, v12, 0x7fff
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v18, 16, v9
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v12, v16, v15, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v6
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v15, 16, v12
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v6, v18, v6, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v7
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v6, v6, v7, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v8
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v8, v15, v8 :: v_dual_lshlrev_b32 v17, 16, v2
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v9, 0xffff0000, v9
; GFX11-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v9
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v6, v18, v6 :: v_dual_lshlrev_b32 v9, 16, v5
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v17, v17
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v7, 0xffff0000, v12
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_max_f32_e32 v13, v13, v13
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v2, v5, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v10
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v8, v8, v10, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
; GFX11-FAKE16-NEXT:    v_bfe_u32 v10, v13, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, 0x400000, v13
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v5, v5, v2, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v7
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v7, v15, v8, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v8, v10, v13, 0x7fff
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v10, 16, v1
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v12, 16, v5
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v17, 16, v4
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v10, v10
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v1, v1, v4 :: v_dual_lshlrev_b32 v16, 16, v2
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v15, 16, v0
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v15, v15
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v0, v3, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v17, v17
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v10, 16, v3
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v4, v4, v1, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v16, v12
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v12, v5, v2, vcc_lo
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v10, v10
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v10, 16, v4
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v3, v3, v0, vcc_lo
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v15, 16, v1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v13, v13
; GFX11-FAKE16-NEXT:    v_max_f32_e32 v12, v12, v12
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v8, v8, v9 :: v_dual_lshlrev_b32 v9, 16, v3
; GFX11-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v15, v10
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v13, 16, v0
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v15, 16, v8
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v8, 0xffff0000, v8
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v10, v4, v1, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v13, v9
; GFX11-FAKE16-NEXT:    v_bfe_u32 v13, v12, 16, 1
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v9, v3, v0 :: v_dual_lshlrev_b32 v10, 16, v10
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v11
; GFX11-FAKE16-NEXT:    v_add3_u32 v13, v13, v12, 0x7fff
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_dual_max_f32 v10, v10, v10 :: v_dual_lshlrev_b32 v9, 16, v9
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v11, v15, v11, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v14
; GFX11-FAKE16-NEXT:    v_bfe_u32 v16, v10, 16, 1
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_max_f32_e32 v9, v9, v9
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v11, v11, v14, vcc_lo
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v14, 0x400000, v12
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v12, v12
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v17, v9, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v12, v13, v14, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v13, v16, v10, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v14, 0x400000, v10
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v10, v10
; GFX11-FAKE16-NEXT:    v_add3_u32 v16, v17, v9, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v17, 0x400000, v9
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v10, v13, v14, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v9, v16, v17, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v8
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v18, 16, v12
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v13, 16, v9
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v8, v15, v11, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v2
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v11, 16, v10
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v18, v2, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v1
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v11, v1, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v5
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v2, v5, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v0
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v0, v13, v0 :: v_dual_and_b32 v5, 0xffff0000, v9
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v4
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v1, v1, v4 :: v_dual_and_b32 v4, 0xffff0000, v10
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v3
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v0, v3, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v4
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xffff0000, v12
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v11, v1, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v5
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_perm_b32 v1, v7, v1, 0x5040100
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v13, v0, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v3
; GFX11-FAKE16-NEXT:    v_perm_b32 v0, v8, v0, 0x5040100
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v18, v2, vcc_lo
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_perm_b32 v2, v6, v2, 0x5040100
; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
;
; GFX12-TRUE16-LABEL: v_maximumnum_v6bf16:
; GFX12-TRUE16:       ; %bb.0:
; GFX12-TRUE16-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX12-TRUE16-NEXT:    s_wait_expcnt 0x0
; GFX12-TRUE16-NEXT:    s_wait_samplecnt 0x0
; GFX12-TRUE16-NEXT:    s_wait_bvhcnt 0x0
; GFX12-TRUE16-NEXT:    s_wait_kmcnt 0x0
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v6, 0xffff0000, v2
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v8, 0xffff0000, v5
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff0000, v1
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v9, 0xffff0000, v4
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v10, 0xffff0000, v0
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s1, v8, v8
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s0, v7, v7
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 16, v2
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v11, 0xffff0000, v3
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v6.l, v2.h, v5.h, vcc_lo
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v7.l, v1.h, v4.h, s0
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s0, v10, v10
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s3, v12, v12
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v8.l, v5.h, v6.l, s1
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v15.l, v6.l
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v9.l, v4.h, v7.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v18.l, v7.l
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v14, 16, v1
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v16.l, v8.l
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v19.l, v9.l
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s1, v11, v11
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s4, v14, v14
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v13, 16, v5
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v20, 16, v0
; GFX12-TRUE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v15, v16
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v2.l, v2.l, v5.l, s3
; GFX12-TRUE16-NEXT:    v_cmp_gt_f32_e64 s2, v18, v19
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s3, 0, v7.l
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v17, 16, v4
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v10.l, v8.l, v6.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v13, v13
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v11.l, v9.l, v7.l, s2
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s2, v20, v20
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v21, 16, v3
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v12.l, v10.l
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v10.l, v0.h, v3.h, s0
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v14, 16, v11
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s0, 0, v6.l
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v5.l, v5.l, v2.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v11.l, v3.h, v10.l, s1
; GFX12-TRUE16-NEXT:    v_max_num_f32_e32 v14, v14, v14
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v13.l, v10.l
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v8.l
; GFX12-TRUE16-NEXT:    v_max_num_f32_e32 v12, v12, v12
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v18.l, v11.l
; GFX12-TRUE16-NEXT:    v_bfe_u32 v19, v14, 16, 1
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
; GFX12-TRUE16-NEXT:    v_or_b32_e32 v22, 0x400000, v14
; GFX12-TRUE16-NEXT:    v_bfe_u32 v16, v12, 16, 1
; GFX12-TRUE16-NEXT:    v_or_b32_e32 v20, 0x400000, v12
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s1, v12, v12
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
; GFX12-TRUE16-NEXT:    v_add3_u32 v19, v19, v14, 0x7fff
; GFX12-TRUE16-NEXT:    v_add3_u32 v16, v16, v12, 0x7fff
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s6, v14, v14
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v15.l, v2.l
; GFX12-TRUE16-NEXT:    v_cmp_gt_f32_e64 s5, v13, v18
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v1.l, v1.l, v4.l, s4
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b32_e64 v12, v16, v20, s1
; GFX12-TRUE16-NEXT:    v_cndmask_b32_e64 v13, v19, v22, s6
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s1, 0, v9.l
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v16.l, v5.l
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v0.l, v0.l, v3.l, s2
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v0.h, v12.h, v6.l, s0
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v14, 0xffff0000, v12
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v1.h, v13.h, v7.l, s3
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v6.l, v11.l, v10.l, s5
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 16, v15
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v0.h, v0.h, v8.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v14
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v1.h, v1.h, v9.l, s1
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s1, v21, v21
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 16, v16
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v12.l, v1.l
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v2.h, v12.h, v0.h, vcc_lo
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v17, v17
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v3.l, v3.l, v0.l, s1
; GFX12-TRUE16-NEXT:    v_cmp_gt_f32_e64 s0, v7, v8
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 16, v6
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v15.l, v0.l
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v4.l, v4.l, v1.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v14.l, v3.l
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v6.l, v5.l, v2.l, s0
; GFX12-TRUE16-NEXT:    v_dual_max_num_f32 v7, v7, v7 :: v_dual_lshlrev_b32 v12, 16, v12
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v9.l, v4.l
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_4)
; GFX12-TRUE16-NEXT:    v_bfe_u32 v16, v7, 16, 1
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v8, 0xffff0000, v13
; GFX12-TRUE16-NEXT:    v_cmp_gt_f32_e64 s0, v15, v14
; GFX12-TRUE16-NEXT:    v_max_num_f32_e32 v14, v6, v6
; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX12-TRUE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v12, v9
; GFX12-TRUE16-NEXT:    v_add3_u32 v9, v16, v7, 0x7fff
; GFX12-TRUE16-NEXT:    v_or_b32_e32 v12, 0x400000, v7
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s1, v14, v14
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v6.l, v4.l, v1.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v7.l, v3.l, v0.l, s0
; GFX12-TRUE16-NEXT:    v_cmp_eq_f32_e64 s0, 0, v8
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_dual_cndmask_b32 v9, v9, v12 :: v_dual_lshlrev_b32 v6, 16, v6
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v10.l
; GFX12-TRUE16-NEXT:    v_bfe_u32 v12, v14, 16, 1
; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX12-TRUE16-NEXT:    v_dual_max_num_f32 v6, v6, v6 :: v_dual_lshlrev_b32 v7, 16, v7
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v1.h, v13.h, v1.h, s0
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v0.h, v9.h, v10.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_add3_u32 v8, v12, v14, 0x7fff
; GFX12-TRUE16-NEXT:    v_or_b32_e32 v10, 0x400000, v14
; GFX12-TRUE16-NEXT:    v_max_num_f32_e32 v7, v7, v7
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v11.l
; GFX12-TRUE16-NEXT:    v_or_b32_e32 v14, 0x400000, v6
; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX12-TRUE16-NEXT:    v_cndmask_b32_e64 v8, v8, v10, s1
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s1, 0, v2.l
; GFX12-TRUE16-NEXT:    v_bfe_u32 v10, v6, 16, 1
; GFX12-TRUE16-NEXT:    v_bfe_u32 v12, v7, 16, 1
; GFX12-TRUE16-NEXT:    v_or_b32_e32 v15, 0x400000, v7
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s2, v7, v7
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v2.l, v8.h, v2.l, s1
; GFX12-TRUE16-NEXT:    v_add3_u32 v10, v10, v6, 0x7fff
; GFX12-TRUE16-NEXT:    v_add3_u32 v12, v12, v7, 0x7fff
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s1, v6, v6
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v0.h, v0.h, v11.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v11, 0xffff0000, v9
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v5.l
; GFX12-TRUE16-NEXT:    v_cndmask_b32_e64 v7, v12, v15, s2
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b32_e64 v6, v10, v14, s1
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s1, 0, v1.l
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s2, 0, v0.l
; GFX12-TRUE16-NEXT:    v_cmp_eq_f32_e64 s3, 0, v11
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v2.l, v2.l, v5.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v4.l
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v1.l, v6.h, v1.l, s1
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v0.l, v7.h, v0.l, s2
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s1, 0, v3.l
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v5, 0xffff0000, v7
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v10, 0xffff0000, v6
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v11, 0xffff0000, v8
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v1.l, v1.l, v4.l, vcc_lo
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v0.l, v0.l, v3.l, s1
; GFX12-TRUE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v5
; GFX12-TRUE16-NEXT:    v_cmp_eq_f32_e64 s1, 0, v10
; GFX12-TRUE16-NEXT:    v_cmp_eq_f32_e64 s2, 0, v11
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v0.h, v9.h, v0.h, s3
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v0.l, v7.h, v0.l, vcc_lo
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v1.l, v6.h, v1.l, s1
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v2.l, v8.h, v2.l, s2
; GFX12-TRUE16-NEXT:    s_setpc_b64 s[30:31]
;
; GFX12-FAKE16-LABEL: v_maximumnum_v6bf16:
; GFX12-FAKE16:       ; %bb.0:
; GFX12-FAKE16-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX12-FAKE16-NEXT:    s_wait_expcnt 0x0
; GFX12-FAKE16-NEXT:    s_wait_samplecnt 0x0
; GFX12-FAKE16-NEXT:    s_wait_bvhcnt 0x0
; GFX12-FAKE16-NEXT:    s_wait_kmcnt 0x0
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v6, 0xffff0000, v2
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v7, 16, v5
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v8, 16, v2
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v10, 16, v4
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v11, 16, v1
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v9, 0xffff0000, v5
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v14, 16, v3
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v15, 16, v0
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v16, 0xffff0000, v3
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v6, v8, v7, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v7, v7, v6 :: v_dual_and_b32 v8, 0xffff0000, v1
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v12, 0xffff0000, v4
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v8, v11, v10 :: v_dual_lshlrev_b32 v11, 16, v7
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v12, v12
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v9, 16, v6
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v12, 16, v8
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v10, v10, v8, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v9, v11
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v11, 0xffff0000, v0
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_2)
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v13, 16, v10
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v9, v7, v6, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v12, v13
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v12, v10, v8, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v11, v11
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v11, v15, v14, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v16, v16
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v15, 16, v11
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v14, v14, v11 :: v_dual_max_num_f32 v9, v9, v9
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v17, 16, v14
; GFX12-FAKE16-NEXT:    v_bfe_u32 v13, v9, 16, 1
; GFX12-FAKE16-NEXT:    v_or_b32_e32 v16, 0x400000, v9
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX12-FAKE16-NEXT:    v_add3_u32 v13, v13, v9, 0x7fff
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v9, v13, v16, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v15, v17
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v13, v14, v11 :: v_dual_lshlrev_b32 v12, 16, v12
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-FAKE16-NEXT:    v_max_num_f32_e32 v12, v12, v12
; GFX12-FAKE16-NEXT:    v_bfe_u32 v18, v12, 16, 1
; GFX12-FAKE16-NEXT:    v_or_b32_e32 v15, 0x400000, v12
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v12, v12
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_2)
; GFX12-FAKE16-NEXT:    v_add3_u32 v16, v18, v12, 0x7fff
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v18, 16, v9
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v12, v16, v15, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v6
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_4) | instid1(VALU_DEP_2)
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v15, 16, v12
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v6, v18, v6, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v7
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v6, v6, v7, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v8
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v8, v15, v8 :: v_dual_lshlrev_b32 v17, 16, v2
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v9, 0xffff0000, v9
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX12-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v9
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v6, v18, v6 :: v_dual_lshlrev_b32 v9, 16, v5
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v17, v17
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v7, 0xffff0000, v12
; GFX12-FAKE16-NEXT:    v_max_num_f32_e32 v13, v13, v13
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v2, v5, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v10
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v8, v8, v10, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
; GFX12-FAKE16-NEXT:    v_bfe_u32 v10, v13, 16, 1
; GFX12-FAKE16-NEXT:    v_or_b32_e32 v9, 0x400000, v13
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v5, v5, v2, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v7
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v7, v15, v8, vcc_lo
; GFX12-FAKE16-NEXT:    v_add3_u32 v8, v10, v13, 0x7fff
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v10, 16, v1
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v12, 16, v5
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v17, 16, v4
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_1)
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v10, v10
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v1, v1, v4 :: v_dual_lshlrev_b32 v16, 16, v2
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v15, 16, v0
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v15, v15
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v0, v3, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v17, v17
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v10, 16, v3
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v4, v4, v1, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v16, v12
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v12, v5, v2, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v10, v10
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v10, 16, v4
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v3, v3, v0, vcc_lo
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v15, 16, v1
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v13, v13
; GFX12-FAKE16-NEXT:    v_max_num_f32_e32 v12, v12, v12
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v8, v8, v9 :: v_dual_lshlrev_b32 v9, 16, v3
; GFX12-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v15, v10
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v13, 16, v0
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v15, 16, v8
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v8, 0xffff0000, v8
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v10, v4, v1, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v13, v9
; GFX12-FAKE16-NEXT:    v_bfe_u32 v13, v12, 16, 1
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v9, v3, v0 :: v_dual_lshlrev_b32 v10, 16, v10
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v11
; GFX12-FAKE16-NEXT:    v_add3_u32 v13, v13, v12, 0x7fff
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX12-FAKE16-NEXT:    v_dual_max_num_f32 v10, v10, v10 :: v_dual_lshlrev_b32 v9, 16, v9
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v11, v15, v11, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v14
; GFX12-FAKE16-NEXT:    v_bfe_u32 v16, v10, 16, 1
; GFX12-FAKE16-NEXT:    v_max_num_f32_e32 v9, v9, v9
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v11, v11, v14, vcc_lo
; GFX12-FAKE16-NEXT:    v_or_b32_e32 v14, 0x400000, v12
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v12, v12
; GFX12-FAKE16-NEXT:    v_bfe_u32 v17, v9, 16, 1
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v12, v13, v14, vcc_lo
; GFX12-FAKE16-NEXT:    v_add3_u32 v13, v16, v10, 0x7fff
; GFX12-FAKE16-NEXT:    v_or_b32_e32 v14, 0x400000, v10
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v10, v10
; GFX12-FAKE16-NEXT:    v_add3_u32 v16, v17, v9, 0x7fff
; GFX12-FAKE16-NEXT:    v_or_b32_e32 v17, 0x400000, v9
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v10, v13, v14, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v9, v16, v17, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v8
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v18, 16, v12
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v13, 16, v9
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v8, v15, v11, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v2
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v11, 16, v10
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v18, v2, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v1
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v11, v1, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v5
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v2, v5, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v0
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v0, v13, v0 :: v_dual_and_b32 v5, 0xffff0000, v9
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v4
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v1, v1, v4 :: v_dual_and_b32 v4, 0xffff0000, v10
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v3
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v0, v3, vcc_lo
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_2)
; GFX12-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v4
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v3, 0xffff0000, v12
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v11, v1, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v5
; GFX12-FAKE16-NEXT:    v_perm_b32 v1, v7, v1, 0x5040100
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v13, v0, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v3
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1)
; GFX12-FAKE16-NEXT:    v_perm_b32 v0, v8, v0, 0x5040100
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v18, v2, vcc_lo
; GFX12-FAKE16-NEXT:    v_perm_b32 v2, v6, v2, 0x5040100
; GFX12-FAKE16-NEXT:    s_setpc_b64 s[30:31]
  %result = call <6 x bfloat> @llvm.maximumnum.v6bf16(<6 x bfloat> %x, <6 x bfloat> %y)
  ret <6 x bfloat> %result
}

define <8 x bfloat> @v_maximumnum_v8bf16(<8 x bfloat> %x, <8 x bfloat> %y) {
; GFX7-LABEL: v_maximumnum_v8bf16:
; GFX7:       ; %bb.0:
; GFX7-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT:    v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT:    v_mul_f32_e32 v8, 1.0, v8
; GFX7-NEXT:    v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT:    v_mul_f32_e32 v9, 1.0, v9
; GFX7-NEXT:    v_mul_f32_e32 v2, 1.0, v2
; GFX7-NEXT:    v_mul_f32_e32 v10, 1.0, v10
; GFX7-NEXT:    v_mul_f32_e32 v3, 1.0, v3
; GFX7-NEXT:    v_mul_f32_e32 v11, 1.0, v11
; GFX7-NEXT:    v_mul_f32_e32 v4, 1.0, v4
; GFX7-NEXT:    v_mul_f32_e32 v12, 1.0, v12
; GFX7-NEXT:    v_mul_f32_e32 v5, 1.0, v5
; GFX7-NEXT:    v_mul_f32_e32 v13, 1.0, v13
; GFX7-NEXT:    v_mul_f32_e32 v6, 1.0, v6
; GFX7-NEXT:    v_mul_f32_e32 v14, 1.0, v14
; GFX7-NEXT:    v_mul_f32_e32 v7, 1.0, v7
; GFX7-NEXT:    v_mul_f32_e32 v15, 1.0, v15
; GFX7-NEXT:    v_and_b32_e32 v15, 0xffff0000, v15
; GFX7-NEXT:    v_and_b32_e32 v7, 0xffff0000, v7
; GFX7-NEXT:    v_and_b32_e32 v14, 0xffff0000, v14
; GFX7-NEXT:    v_and_b32_e32 v6, 0xffff0000, v6
; GFX7-NEXT:    v_and_b32_e32 v13, 0xffff0000, v13
; GFX7-NEXT:    v_and_b32_e32 v5, 0xffff0000, v5
; GFX7-NEXT:    v_and_b32_e32 v12, 0xffff0000, v12
; GFX7-NEXT:    v_and_b32_e32 v4, 0xffff0000, v4
; GFX7-NEXT:    v_and_b32_e32 v11, 0xffff0000, v11
; GFX7-NEXT:    v_and_b32_e32 v3, 0xffff0000, v3
; GFX7-NEXT:    v_and_b32_e32 v10, 0xffff0000, v10
; GFX7-NEXT:    v_and_b32_e32 v2, 0xffff0000, v2
; GFX7-NEXT:    v_and_b32_e32 v9, 0xffff0000, v9
; GFX7-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT:    v_and_b32_e32 v8, 0xffff0000, v8
; GFX7-NEXT:    v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT:    v_max_f32_e32 v7, v7, v15
; GFX7-NEXT:    v_max_f32_e32 v6, v6, v14
; GFX7-NEXT:    v_max_f32_e32 v5, v5, v13
; GFX7-NEXT:    v_max_f32_e32 v4, v4, v12
; GFX7-NEXT:    v_max_f32_e32 v3, v3, v11
; GFX7-NEXT:    v_max_f32_e32 v2, v2, v10
; GFX7-NEXT:    v_max_f32_e32 v1, v1, v9
; GFX7-NEXT:    v_max_f32_e32 v0, v0, v8
; GFX7-NEXT:    v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT:    v_and_b32_e32 v2, 0xffff0000, v2
; GFX7-NEXT:    v_and_b32_e32 v3, 0xffff0000, v3
; GFX7-NEXT:    v_and_b32_e32 v4, 0xffff0000, v4
; GFX7-NEXT:    v_and_b32_e32 v5, 0xffff0000, v5
; GFX7-NEXT:    v_and_b32_e32 v6, 0xffff0000, v6
; GFX7-NEXT:    v_and_b32_e32 v7, 0xffff0000, v7
; GFX7-NEXT:    s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_maximumnum_v8bf16:
; GFX8:       ; %bb.0:
; GFX8-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT:    v_and_b32_e32 v8, 0xffff0000, v3
; GFX8-NEXT:    v_lshrrev_b32_e32 v9, 16, v7
; GFX8-NEXT:    v_lshrrev_b32_e32 v10, 16, v3
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v8, v8
; GFX8-NEXT:    v_and_b32_e32 v11, 0xffff0000, v7
; GFX8-NEXT:    v_cndmask_b32_e32 v8, v10, v9, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v11, v11
; GFX8-NEXT:    v_cndmask_b32_e32 v9, v9, v8, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v10, 16, v8
; GFX8-NEXT:    v_lshlrev_b32_e32 v11, 16, v9
; GFX8-NEXT:    v_cmp_gt_f32_e32 vcc, v10, v11
; GFX8-NEXT:    v_cndmask_b32_e32 v10, v9, v8, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
; GFX8-NEXT:    v_mul_f32_e32 v10, 1.0, v10
; GFX8-NEXT:    v_bfe_u32 v11, v10, 16, 1
; GFX8-NEXT:    v_add_u32_e32 v11, vcc, v11, v10
; GFX8-NEXT:    s_movk_i32 s4, 0x7fff
; GFX8-NEXT:    v_add_u32_e32 v11, vcc, s4, v11
; GFX8-NEXT:    v_or_b32_e32 v12, 0x400000, v10
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v10, v10
; GFX8-NEXT:    v_cndmask_b32_e32 v10, v11, v12, vcc
; GFX8-NEXT:    v_lshrrev_b32_e32 v11, 16, v10
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v8
; GFX8-NEXT:    v_cndmask_b32_e32 v8, v11, v8, vcc
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v9
; GFX8-NEXT:    v_cndmask_b32_e32 v8, v8, v9, vcc
; GFX8-NEXT:    v_and_b32_e32 v9, 0xffff0000, v10
; GFX8-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v9
; GFX8-NEXT:    v_and_b32_e32 v9, 0xffff0000, v2
; GFX8-NEXT:    v_cndmask_b32_e32 v8, v11, v8, vcc
; GFX8-NEXT:    v_lshrrev_b32_e32 v10, 16, v6
; GFX8-NEXT:    v_lshrrev_b32_e32 v11, 16, v2
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v9, v9
; GFX8-NEXT:    v_and_b32_e32 v12, 0xffff0000, v6
; GFX8-NEXT:    v_cndmask_b32_e32 v9, v11, v10, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v12, v12
; GFX8-NEXT:    v_cndmask_b32_e32 v10, v10, v9, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v11, 16, v9
; GFX8-NEXT:    v_lshlrev_b32_e32 v12, 16, v10
; GFX8-NEXT:    v_cmp_gt_f32_e32 vcc, v11, v12
; GFX8-NEXT:    v_cndmask_b32_e32 v11, v10, v9, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
; GFX8-NEXT:    v_mul_f32_e32 v11, 1.0, v11
; GFX8-NEXT:    v_bfe_u32 v12, v11, 16, 1
; GFX8-NEXT:    v_add_u32_e32 v12, vcc, v12, v11
; GFX8-NEXT:    v_add_u32_e32 v12, vcc, s4, v12
; GFX8-NEXT:    v_or_b32_e32 v13, 0x400000, v11
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v11, v11
; GFX8-NEXT:    v_cndmask_b32_e32 v11, v12, v13, vcc
; GFX8-NEXT:    v_lshrrev_b32_e32 v12, 16, v11
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v9
; GFX8-NEXT:    v_cndmask_b32_e32 v9, v12, v9, vcc
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v10
; GFX8-NEXT:    v_cndmask_b32_e32 v9, v9, v10, vcc
; GFX8-NEXT:    v_and_b32_e32 v10, 0xffff0000, v11
; GFX8-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v10
; GFX8-NEXT:    v_and_b32_e32 v10, 0xffff0000, v1
; GFX8-NEXT:    v_cndmask_b32_e32 v9, v12, v9, vcc
; GFX8-NEXT:    v_lshrrev_b32_e32 v11, 16, v5
; GFX8-NEXT:    v_lshrrev_b32_e32 v12, 16, v1
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v10, v10
; GFX8-NEXT:    v_and_b32_e32 v13, 0xffff0000, v5
; GFX8-NEXT:    v_cndmask_b32_e32 v10, v12, v11, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v13, v13
; GFX8-NEXT:    v_cndmask_b32_e32 v11, v11, v10, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v12, 16, v10
; GFX8-NEXT:    v_lshlrev_b32_e32 v13, 16, v11
; GFX8-NEXT:    v_cmp_gt_f32_e32 vcc, v12, v13
; GFX8-NEXT:    v_cndmask_b32_e32 v12, v11, v10, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; GFX8-NEXT:    v_mul_f32_e32 v12, 1.0, v12
; GFX8-NEXT:    v_bfe_u32 v13, v12, 16, 1
; GFX8-NEXT:    v_add_u32_e32 v13, vcc, v13, v12
; GFX8-NEXT:    v_add_u32_e32 v13, vcc, s4, v13
; GFX8-NEXT:    v_or_b32_e32 v14, 0x400000, v12
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v12, v12
; GFX8-NEXT:    v_cndmask_b32_e32 v12, v13, v14, vcc
; GFX8-NEXT:    v_lshrrev_b32_e32 v13, 16, v12
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v10
; GFX8-NEXT:    v_cndmask_b32_e32 v10, v13, v10, vcc
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v11
; GFX8-NEXT:    v_cndmask_b32_e32 v10, v10, v11, vcc
; GFX8-NEXT:    v_and_b32_e32 v11, 0xffff0000, v12
; GFX8-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v11
; GFX8-NEXT:    v_and_b32_e32 v11, 0xffff0000, v0
; GFX8-NEXT:    v_cndmask_b32_e32 v10, v13, v10, vcc
; GFX8-NEXT:    v_lshrrev_b32_e32 v12, 16, v4
; GFX8-NEXT:    v_lshrrev_b32_e32 v13, 16, v0
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v11, v11
; GFX8-NEXT:    v_and_b32_e32 v14, 0xffff0000, v4
; GFX8-NEXT:    v_cndmask_b32_e32 v11, v13, v12, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v14, v14
; GFX8-NEXT:    v_cndmask_b32_e32 v12, v12, v11, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v13, 16, v11
; GFX8-NEXT:    v_lshlrev_b32_e32 v14, 16, v12
; GFX8-NEXT:    v_cmp_gt_f32_e32 vcc, v13, v14
; GFX8-NEXT:    v_cndmask_b32_e32 v13, v12, v11, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
; GFX8-NEXT:    v_mul_f32_e32 v13, 1.0, v13
; GFX8-NEXT:    v_bfe_u32 v14, v13, 16, 1
; GFX8-NEXT:    v_add_u32_e32 v14, vcc, v14, v13
; GFX8-NEXT:    v_add_u32_e32 v14, vcc, s4, v14
; GFX8-NEXT:    v_or_b32_e32 v15, 0x400000, v13
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v13, v13
; GFX8-NEXT:    v_cndmask_b32_e32 v13, v14, v15, vcc
; GFX8-NEXT:    v_lshrrev_b32_e32 v14, 16, v13
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v11
; GFX8-NEXT:    v_cndmask_b32_e32 v11, v14, v11, vcc
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v12
; GFX8-NEXT:    v_cndmask_b32_e32 v11, v11, v12, vcc
; GFX8-NEXT:    v_and_b32_e32 v12, 0xffff0000, v13
; GFX8-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v12
; GFX8-NEXT:    v_lshlrev_b32_e32 v12, 16, v3
; GFX8-NEXT:    v_cndmask_b32_e32 v11, v14, v11, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v12, v12
; GFX8-NEXT:    v_lshlrev_b32_e32 v12, 16, v7
; GFX8-NEXT:    v_cndmask_b32_e32 v3, v3, v7, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v12, v12
; GFX8-NEXT:    v_cndmask_b32_e32 v7, v7, v3, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v12, 16, v7
; GFX8-NEXT:    v_lshlrev_b32_e32 v13, 16, v3
; GFX8-NEXT:    v_cmp_gt_f32_e32 vcc, v13, v12
; GFX8-NEXT:    v_cndmask_b32_e32 v12, v7, v3, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; GFX8-NEXT:    v_mul_f32_e32 v12, 1.0, v12
; GFX8-NEXT:    v_bfe_u32 v13, v12, 16, 1
; GFX8-NEXT:    v_add_u32_e32 v13, vcc, v13, v12
; GFX8-NEXT:    v_add_u32_e32 v13, vcc, s4, v13
; GFX8-NEXT:    v_or_b32_e32 v14, 0x400000, v12
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v12, v12
; GFX8-NEXT:    v_cndmask_b32_e32 v12, v13, v14, vcc
; GFX8-NEXT:    v_lshrrev_b32_e32 v13, 16, v12
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v3
; GFX8-NEXT:    v_cndmask_b32_e32 v3, v13, v3, vcc
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v7
; GFX8-NEXT:    v_cndmask_b32_e32 v3, v3, v7, vcc
; GFX8-NEXT:    v_and_b32_e32 v7, 0xffff0000, v12
; GFX8-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v7
; GFX8-NEXT:    v_lshlrev_b32_e32 v7, 16, v2
; GFX8-NEXT:    v_cndmask_b32_e32 v3, v13, v3, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v7, v7
; GFX8-NEXT:    v_lshlrev_b32_e32 v7, 16, v6
; GFX8-NEXT:    v_cndmask_b32_e32 v2, v2, v6, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v7, v7
; GFX8-NEXT:    v_cndmask_b32_e32 v6, v6, v2, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v7, 16, v6
; GFX8-NEXT:    v_lshlrev_b32_e32 v12, 16, v2
; GFX8-NEXT:    v_cmp_gt_f32_e32 vcc, v12, v7
; GFX8-NEXT:    v_cndmask_b32_e32 v7, v6, v2, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; GFX8-NEXT:    v_mul_f32_e32 v7, 1.0, v7
; GFX8-NEXT:    v_bfe_u32 v12, v7, 16, 1
; GFX8-NEXT:    v_add_u32_e32 v12, vcc, v12, v7
; GFX8-NEXT:    v_add_u32_e32 v12, vcc, s4, v12
; GFX8-NEXT:    v_or_b32_e32 v13, 0x400000, v7
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v7, v7
; GFX8-NEXT:    v_cndmask_b32_e32 v7, v12, v13, vcc
; GFX8-NEXT:    v_lshrrev_b32_e32 v12, 16, v7
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v2
; GFX8-NEXT:    v_cndmask_b32_e32 v2, v12, v2, vcc
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v6
; GFX8-NEXT:    v_cndmask_b32_e32 v2, v2, v6, vcc
; GFX8-NEXT:    v_and_b32_e32 v6, 0xffff0000, v7
; GFX8-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v6
; GFX8-NEXT:    v_lshlrev_b32_e32 v6, 16, v1
; GFX8-NEXT:    v_cndmask_b32_e32 v2, v12, v2, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v6, v6
; GFX8-NEXT:    v_lshlrev_b32_e32 v6, 16, v5
; GFX8-NEXT:    v_cndmask_b32_e32 v1, v1, v5, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v6, v6
; GFX8-NEXT:    v_cndmask_b32_e32 v5, v5, v1, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v6, 16, v5
; GFX8-NEXT:    v_lshlrev_b32_e32 v7, 16, v1
; GFX8-NEXT:    v_cmp_gt_f32_e32 vcc, v7, v6
; GFX8-NEXT:    v_cndmask_b32_e32 v6, v5, v1, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; GFX8-NEXT:    v_mul_f32_e32 v6, 1.0, v6
; GFX8-NEXT:    v_bfe_u32 v7, v6, 16, 1
; GFX8-NEXT:    v_add_u32_e32 v7, vcc, v7, v6
; GFX8-NEXT:    v_add_u32_e32 v7, vcc, s4, v7
; GFX8-NEXT:    v_or_b32_e32 v12, 0x400000, v6
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v6, v6
; GFX8-NEXT:    v_cndmask_b32_e32 v6, v7, v12, vcc
; GFX8-NEXT:    v_lshrrev_b32_e32 v7, 16, v6
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v1
; GFX8-NEXT:    v_cndmask_b32_e32 v1, v7, v1, vcc
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v5
; GFX8-NEXT:    v_cndmask_b32_e32 v1, v1, v5, vcc
; GFX8-NEXT:    v_and_b32_e32 v5, 0xffff0000, v6
; GFX8-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v5
; GFX8-NEXT:    v_lshlrev_b32_e32 v5, 16, v0
; GFX8-NEXT:    v_cndmask_b32_e32 v1, v7, v1, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; GFX8-NEXT:    v_lshlrev_b32_e32 v5, 16, v4
; GFX8-NEXT:    v_cndmask_b32_e32 v0, v0, v4, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; GFX8-NEXT:    v_cndmask_b32_e32 v4, v4, v0, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v5, 16, v4
; GFX8-NEXT:    v_lshlrev_b32_e32 v6, 16, v0
; GFX8-NEXT:    v_cmp_gt_f32_e32 vcc, v6, v5
; GFX8-NEXT:    v_cndmask_b32_e32 v5, v4, v0, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; GFX8-NEXT:    v_mul_f32_e32 v5, 1.0, v5
; GFX8-NEXT:    v_bfe_u32 v6, v5, 16, 1
; GFX8-NEXT:    v_add_u32_e32 v6, vcc, v6, v5
; GFX8-NEXT:    v_add_u32_e32 v6, vcc, s4, v6
; GFX8-NEXT:    v_or_b32_e32 v7, 0x400000, v5
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; GFX8-NEXT:    v_cndmask_b32_e32 v5, v6, v7, vcc
; GFX8-NEXT:    v_lshrrev_b32_e32 v6, 16, v5
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v0
; GFX8-NEXT:    v_cndmask_b32_e32 v0, v6, v0, vcc
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v4
; GFX8-NEXT:    v_cndmask_b32_e32 v0, v0, v4, vcc
; GFX8-NEXT:    v_and_b32_e32 v4, 0xffff0000, v5
; GFX8-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v4
; GFX8-NEXT:    v_cndmask_b32_e32 v0, v6, v0, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v4, 16, v11
; GFX8-NEXT:    v_or_b32_sdwa v0, v0, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX8-NEXT:    v_lshlrev_b32_e32 v4, 16, v10
; GFX8-NEXT:    v_or_b32_sdwa v1, v1, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX8-NEXT:    v_lshlrev_b32_e32 v4, 16, v9
; GFX8-NEXT:    v_or_b32_sdwa v2, v2, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX8-NEXT:    v_lshlrev_b32_e32 v4, 16, v8
; GFX8-NEXT:    v_or_b32_sdwa v3, v3, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX8-NEXT:    s_setpc_b64 s[30:31]
;
; GFX900-LABEL: v_maximumnum_v8bf16:
; GFX900:       ; %bb.0:
; GFX900-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX900-NEXT:    v_and_b32_e32 v8, 0xffff0000, v3
; GFX900-NEXT:    v_lshrrev_b32_e32 v9, 16, v7
; GFX900-NEXT:    v_lshrrev_b32_e32 v10, 16, v3
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v8, v8
; GFX900-NEXT:    v_and_b32_e32 v11, 0xffff0000, v7
; GFX900-NEXT:    v_cndmask_b32_e32 v8, v10, v9, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v11, v11
; GFX900-NEXT:    v_cndmask_b32_e32 v9, v9, v8, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v10, 16, v8
; GFX900-NEXT:    v_lshlrev_b32_e32 v11, 16, v9
; GFX900-NEXT:    v_cmp_gt_f32_e32 vcc, v10, v11
; GFX900-NEXT:    v_cndmask_b32_e32 v10, v9, v8, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
; GFX900-NEXT:    v_max_f32_e32 v10, v10, v10
; GFX900-NEXT:    v_bfe_u32 v11, v10, 16, 1
; GFX900-NEXT:    s_movk_i32 s4, 0x7fff
; GFX900-NEXT:    v_add3_u32 v11, v11, v10, s4
; GFX900-NEXT:    v_or_b32_e32 v12, 0x400000, v10
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v10, v10
; GFX900-NEXT:    v_cndmask_b32_e32 v10, v11, v12, vcc
; GFX900-NEXT:    v_lshrrev_b32_e32 v11, 16, v10
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v8
; GFX900-NEXT:    v_cndmask_b32_e32 v8, v11, v8, vcc
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v9
; GFX900-NEXT:    v_cndmask_b32_e32 v8, v8, v9, vcc
; GFX900-NEXT:    v_and_b32_e32 v9, 0xffff0000, v10
; GFX900-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v9
; GFX900-NEXT:    v_and_b32_e32 v9, 0xffff0000, v2
; GFX900-NEXT:    v_cndmask_b32_e32 v8, v11, v8, vcc
; GFX900-NEXT:    v_lshrrev_b32_e32 v10, 16, v6
; GFX900-NEXT:    v_lshrrev_b32_e32 v11, 16, v2
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v9, v9
; GFX900-NEXT:    v_and_b32_e32 v12, 0xffff0000, v6
; GFX900-NEXT:    v_cndmask_b32_e32 v9, v11, v10, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v12, v12
; GFX900-NEXT:    v_cndmask_b32_e32 v10, v10, v9, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v11, 16, v9
; GFX900-NEXT:    v_lshlrev_b32_e32 v12, 16, v10
; GFX900-NEXT:    v_cmp_gt_f32_e32 vcc, v11, v12
; GFX900-NEXT:    v_cndmask_b32_e32 v11, v10, v9, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
; GFX900-NEXT:    v_max_f32_e32 v11, v11, v11
; GFX900-NEXT:    v_bfe_u32 v12, v11, 16, 1
; GFX900-NEXT:    v_add3_u32 v12, v12, v11, s4
; GFX900-NEXT:    v_or_b32_e32 v13, 0x400000, v11
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v11, v11
; GFX900-NEXT:    v_cndmask_b32_e32 v11, v12, v13, vcc
; GFX900-NEXT:    v_lshrrev_b32_e32 v12, 16, v11
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v9
; GFX900-NEXT:    v_cndmask_b32_e32 v9, v12, v9, vcc
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v10
; GFX900-NEXT:    v_cndmask_b32_e32 v9, v9, v10, vcc
; GFX900-NEXT:    v_and_b32_e32 v10, 0xffff0000, v11
; GFX900-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v10
; GFX900-NEXT:    v_and_b32_e32 v10, 0xffff0000, v1
; GFX900-NEXT:    v_cndmask_b32_e32 v9, v12, v9, vcc
; GFX900-NEXT:    v_lshrrev_b32_e32 v11, 16, v5
; GFX900-NEXT:    v_lshrrev_b32_e32 v12, 16, v1
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v10, v10
; GFX900-NEXT:    v_and_b32_e32 v13, 0xffff0000, v5
; GFX900-NEXT:    v_cndmask_b32_e32 v10, v12, v11, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v13, v13
; GFX900-NEXT:    v_cndmask_b32_e32 v11, v11, v10, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v12, 16, v10
; GFX900-NEXT:    v_lshlrev_b32_e32 v13, 16, v11
; GFX900-NEXT:    v_cmp_gt_f32_e32 vcc, v12, v13
; GFX900-NEXT:    v_cndmask_b32_e32 v12, v11, v10, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; GFX900-NEXT:    v_max_f32_e32 v12, v12, v12
; GFX900-NEXT:    v_bfe_u32 v13, v12, 16, 1
; GFX900-NEXT:    v_add3_u32 v13, v13, v12, s4
; GFX900-NEXT:    v_or_b32_e32 v14, 0x400000, v12
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v12, v12
; GFX900-NEXT:    v_cndmask_b32_e32 v12, v13, v14, vcc
; GFX900-NEXT:    v_lshrrev_b32_e32 v13, 16, v12
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v10
; GFX900-NEXT:    v_cndmask_b32_e32 v10, v13, v10, vcc
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v11
; GFX900-NEXT:    v_cndmask_b32_e32 v10, v10, v11, vcc
; GFX900-NEXT:    v_and_b32_e32 v11, 0xffff0000, v12
; GFX900-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v11
; GFX900-NEXT:    v_and_b32_e32 v11, 0xffff0000, v0
; GFX900-NEXT:    v_cndmask_b32_e32 v10, v13, v10, vcc
; GFX900-NEXT:    v_lshrrev_b32_e32 v12, 16, v4
; GFX900-NEXT:    v_lshrrev_b32_e32 v13, 16, v0
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v11, v11
; GFX900-NEXT:    v_and_b32_e32 v14, 0xffff0000, v4
; GFX900-NEXT:    v_cndmask_b32_e32 v11, v13, v12, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v14, v14
; GFX900-NEXT:    v_cndmask_b32_e32 v12, v12, v11, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v13, 16, v11
; GFX900-NEXT:    v_lshlrev_b32_e32 v14, 16, v12
; GFX900-NEXT:    v_cmp_gt_f32_e32 vcc, v13, v14
; GFX900-NEXT:    v_cndmask_b32_e32 v13, v12, v11, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
; GFX900-NEXT:    v_max_f32_e32 v13, v13, v13
; GFX900-NEXT:    v_bfe_u32 v14, v13, 16, 1
; GFX900-NEXT:    v_add3_u32 v14, v14, v13, s4
; GFX900-NEXT:    v_or_b32_e32 v15, 0x400000, v13
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v13, v13
; GFX900-NEXT:    v_cndmask_b32_e32 v13, v14, v15, vcc
; GFX900-NEXT:    v_lshrrev_b32_e32 v14, 16, v13
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v11
; GFX900-NEXT:    v_cndmask_b32_e32 v11, v14, v11, vcc
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v12
; GFX900-NEXT:    v_cndmask_b32_e32 v11, v11, v12, vcc
; GFX900-NEXT:    v_and_b32_e32 v12, 0xffff0000, v13
; GFX900-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v12
; GFX900-NEXT:    v_lshlrev_b32_e32 v12, 16, v3
; GFX900-NEXT:    v_cndmask_b32_e32 v11, v14, v11, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v12, v12
; GFX900-NEXT:    v_lshlrev_b32_e32 v12, 16, v7
; GFX900-NEXT:    v_cndmask_b32_e32 v3, v3, v7, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v12, v12
; GFX900-NEXT:    v_cndmask_b32_e32 v7, v7, v3, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v12, 16, v7
; GFX900-NEXT:    v_lshlrev_b32_e32 v13, 16, v3
; GFX900-NEXT:    v_cmp_gt_f32_e32 vcc, v13, v12
; GFX900-NEXT:    v_cndmask_b32_e32 v12, v7, v3, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; GFX900-NEXT:    v_max_f32_e32 v12, v12, v12
; GFX900-NEXT:    v_bfe_u32 v13, v12, 16, 1
; GFX900-NEXT:    v_add3_u32 v13, v13, v12, s4
; GFX900-NEXT:    v_or_b32_e32 v14, 0x400000, v12
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v12, v12
; GFX900-NEXT:    v_cndmask_b32_e32 v12, v13, v14, vcc
; GFX900-NEXT:    v_lshrrev_b32_e32 v13, 16, v12
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v3
; GFX900-NEXT:    v_cndmask_b32_e32 v3, v13, v3, vcc
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v7
; GFX900-NEXT:    v_cndmask_b32_e32 v3, v3, v7, vcc
; GFX900-NEXT:    v_and_b32_e32 v7, 0xffff0000, v12
; GFX900-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v7
; GFX900-NEXT:    v_lshlrev_b32_e32 v7, 16, v2
; GFX900-NEXT:    v_cndmask_b32_e32 v3, v13, v3, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v7, v7
; GFX900-NEXT:    v_lshlrev_b32_e32 v7, 16, v6
; GFX900-NEXT:    v_cndmask_b32_e32 v2, v2, v6, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v7, v7
; GFX900-NEXT:    v_cndmask_b32_e32 v6, v6, v2, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v7, 16, v6
; GFX900-NEXT:    v_lshlrev_b32_e32 v12, 16, v2
; GFX900-NEXT:    v_cmp_gt_f32_e32 vcc, v12, v7
; GFX900-NEXT:    v_cndmask_b32_e32 v7, v6, v2, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; GFX900-NEXT:    v_max_f32_e32 v7, v7, v7
; GFX900-NEXT:    v_bfe_u32 v12, v7, 16, 1
; GFX900-NEXT:    v_add3_u32 v12, v12, v7, s4
; GFX900-NEXT:    v_or_b32_e32 v13, 0x400000, v7
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v7, v7
; GFX900-NEXT:    v_cndmask_b32_e32 v7, v12, v13, vcc
; GFX900-NEXT:    v_lshrrev_b32_e32 v12, 16, v7
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v2
; GFX900-NEXT:    v_cndmask_b32_e32 v2, v12, v2, vcc
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v6
; GFX900-NEXT:    v_cndmask_b32_e32 v2, v2, v6, vcc
; GFX900-NEXT:    v_and_b32_e32 v6, 0xffff0000, v7
; GFX900-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v6
; GFX900-NEXT:    v_lshlrev_b32_e32 v6, 16, v1
; GFX900-NEXT:    v_cndmask_b32_e32 v2, v12, v2, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v6, v6
; GFX900-NEXT:    v_lshlrev_b32_e32 v6, 16, v5
; GFX900-NEXT:    v_cndmask_b32_e32 v1, v1, v5, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v6, v6
; GFX900-NEXT:    v_cndmask_b32_e32 v5, v5, v1, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v6, 16, v5
; GFX900-NEXT:    v_lshlrev_b32_e32 v7, 16, v1
; GFX900-NEXT:    v_cmp_gt_f32_e32 vcc, v7, v6
; GFX900-NEXT:    v_cndmask_b32_e32 v6, v5, v1, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; GFX900-NEXT:    v_max_f32_e32 v6, v6, v6
; GFX900-NEXT:    v_bfe_u32 v7, v6, 16, 1
; GFX900-NEXT:    v_add3_u32 v7, v7, v6, s4
; GFX900-NEXT:    v_or_b32_e32 v12, 0x400000, v6
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v6, v6
; GFX900-NEXT:    v_cndmask_b32_e32 v6, v7, v12, vcc
; GFX900-NEXT:    v_lshrrev_b32_e32 v7, 16, v6
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v1
; GFX900-NEXT:    v_cndmask_b32_e32 v1, v7, v1, vcc
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v5
; GFX900-NEXT:    v_cndmask_b32_e32 v1, v1, v5, vcc
; GFX900-NEXT:    v_and_b32_e32 v5, 0xffff0000, v6
; GFX900-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v5
; GFX900-NEXT:    v_lshlrev_b32_e32 v5, 16, v0
; GFX900-NEXT:    v_cndmask_b32_e32 v1, v7, v1, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; GFX900-NEXT:    v_lshlrev_b32_e32 v5, 16, v4
; GFX900-NEXT:    v_cndmask_b32_e32 v0, v0, v4, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; GFX900-NEXT:    v_cndmask_b32_e32 v4, v4, v0, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v5, 16, v4
; GFX900-NEXT:    v_lshlrev_b32_e32 v6, 16, v0
; GFX900-NEXT:    v_cmp_gt_f32_e32 vcc, v6, v5
; GFX900-NEXT:    v_cndmask_b32_e32 v5, v4, v0, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; GFX900-NEXT:    v_max_f32_e32 v5, v5, v5
; GFX900-NEXT:    v_bfe_u32 v6, v5, 16, 1
; GFX900-NEXT:    v_add3_u32 v6, v6, v5, s4
; GFX900-NEXT:    v_or_b32_e32 v7, 0x400000, v5
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; GFX900-NEXT:    v_cndmask_b32_e32 v5, v6, v7, vcc
; GFX900-NEXT:    v_lshrrev_b32_e32 v6, 16, v5
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v0
; GFX900-NEXT:    v_cndmask_b32_e32 v0, v6, v0, vcc
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v4
; GFX900-NEXT:    v_cndmask_b32_e32 v0, v0, v4, vcc
; GFX900-NEXT:    v_and_b32_e32 v4, 0xffff0000, v5
; GFX900-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v4
; GFX900-NEXT:    v_cndmask_b32_e32 v0, v6, v0, vcc
; GFX900-NEXT:    s_mov_b32 s4, 0x5040100
; GFX900-NEXT:    v_perm_b32 v0, v11, v0, s4
; GFX900-NEXT:    v_perm_b32 v1, v10, v1, s4
; GFX900-NEXT:    v_perm_b32 v2, v9, v2, s4
; GFX900-NEXT:    v_perm_b32 v3, v8, v3, s4
; GFX900-NEXT:    s_setpc_b64 s[30:31]
;
; GFX950-LABEL: v_maximumnum_v8bf16:
; GFX950:       ; %bb.0:
; GFX950-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX950-NEXT:    v_and_b32_e32 v8, 0xffff0000, v3
; GFX950-NEXT:    v_lshrrev_b32_e32 v9, 16, v7
; GFX950-NEXT:    v_lshrrev_b32_e32 v10, 16, v3
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v8, v8
; GFX950-NEXT:    v_and_b32_e32 v11, 0xffff0000, v7
; GFX950-NEXT:    v_and_b32_e32 v12, 0xffff0000, v6
; GFX950-NEXT:    v_cndmask_b32_e32 v8, v10, v9, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v11, v11
; GFX950-NEXT:    v_lshlrev_b32_e32 v10, 16, v8
; GFX950-NEXT:    v_and_b32_e32 v13, 0xffff0000, v5
; GFX950-NEXT:    v_cndmask_b32_e32 v9, v9, v8, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v11, 16, v9
; GFX950-NEXT:    v_cmp_gt_f32_e32 vcc, v10, v11
; GFX950-NEXT:    v_lshrrev_b32_e32 v11, 16, v2
; GFX950-NEXT:    v_and_b32_e32 v14, 0xffff0000, v4
; GFX950-NEXT:    v_cndmask_b32_e32 v10, v9, v8, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
; GFX950-NEXT:    v_max_f32_e32 v10, v10, v10
; GFX950-NEXT:    v_cvt_pk_bf16_f32 v10, v10, s0
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v8
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v8, v10, v8, vcc
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v9
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v8, v8, v9, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v9, 16, v10
; GFX950-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v9
; GFX950-NEXT:    v_and_b32_e32 v9, 0xffff0000, v2
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v8, v10, v8, vcc
; GFX950-NEXT:    v_lshrrev_b32_e32 v10, 16, v6
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v9, v9
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v9, v11, v10, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v12, v12
; GFX950-NEXT:    v_lshlrev_b32_e32 v11, 16, v9
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v10, v10, v9, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v12, 16, v10
; GFX950-NEXT:    v_cmp_gt_f32_e32 vcc, v11, v12
; GFX950-NEXT:    v_lshrrev_b32_e32 v12, 16, v1
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v11, v10, v9, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
; GFX950-NEXT:    v_max_f32_e32 v11, v11, v11
; GFX950-NEXT:    v_cvt_pk_bf16_f32 v11, v11, s0
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v9
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v9, v11, v9, vcc
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v10
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v9, v9, v10, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v10, 16, v11
; GFX950-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v10
; GFX950-NEXT:    v_and_b32_e32 v10, 0xffff0000, v1
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v9, v11, v9, vcc
; GFX950-NEXT:    v_lshrrev_b32_e32 v11, 16, v5
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v10, v10
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v10, v12, v11, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v13, v13
; GFX950-NEXT:    v_lshlrev_b32_e32 v12, 16, v10
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v11, v11, v10, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v13, 16, v11
; GFX950-NEXT:    v_cmp_gt_f32_e32 vcc, v12, v13
; GFX950-NEXT:    v_lshrrev_b32_e32 v13, 16, v0
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v12, v11, v10, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; GFX950-NEXT:    v_max_f32_e32 v12, v12, v12
; GFX950-NEXT:    v_cvt_pk_bf16_f32 v12, v12, s0
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v10
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v10, v12, v10, vcc
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v11
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v10, v10, v11, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v11, 16, v12
; GFX950-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v11
; GFX950-NEXT:    v_and_b32_e32 v11, 0xffff0000, v0
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v10, v12, v10, vcc
; GFX950-NEXT:    v_lshrrev_b32_e32 v12, 16, v4
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v11, v11
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v11, v13, v12, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v14, v14
; GFX950-NEXT:    v_lshlrev_b32_e32 v13, 16, v11
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v12, v12, v11, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v14, 16, v12
; GFX950-NEXT:    v_cmp_gt_f32_e32 vcc, v13, v14
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v13, v12, v11, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
; GFX950-NEXT:    v_max_f32_e32 v13, v13, v13
; GFX950-NEXT:    v_cvt_pk_bf16_f32 v13, v13, s0
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v11
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v11, v13, v11, vcc
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v12
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v11, v11, v12, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v12, 16, v13
; GFX950-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v12
; GFX950-NEXT:    v_lshlrev_b32_e32 v12, 16, v3
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v11, v13, v11, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v12, v12
; GFX950-NEXT:    v_lshlrev_b32_e32 v12, 16, v7
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v3, v3, v7, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v12, v12
; GFX950-NEXT:    v_lshlrev_b32_e32 v13, 16, v3
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v7, v7, v3, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v12, 16, v7
; GFX950-NEXT:    v_cmp_gt_f32_e32 vcc, v13, v12
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v12, v7, v3, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; GFX950-NEXT:    v_max_f32_e32 v12, v12, v12
; GFX950-NEXT:    v_cvt_pk_bf16_f32 v12, v12, s0
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v3
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v3, v12, v3, vcc
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v7
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v3, v3, v7, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v7, 16, v12
; GFX950-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v7
; GFX950-NEXT:    v_lshlrev_b32_e32 v7, 16, v2
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v3, v12, v3, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v7, v7
; GFX950-NEXT:    v_lshlrev_b32_e32 v7, 16, v6
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v2, v2, v6, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v7, v7
; GFX950-NEXT:    v_lshlrev_b32_e32 v12, 16, v2
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v6, v6, v2, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v7, 16, v6
; GFX950-NEXT:    v_cmp_gt_f32_e32 vcc, v12, v7
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v7, v6, v2, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; GFX950-NEXT:    v_max_f32_e32 v7, v7, v7
; GFX950-NEXT:    v_cvt_pk_bf16_f32 v7, v7, s0
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v2
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v2, v7, v2, vcc
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v6
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v2, v2, v6, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v6, 16, v7
; GFX950-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v6
; GFX950-NEXT:    v_lshlrev_b32_e32 v6, 16, v1
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v2, v7, v2, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v6, v6
; GFX950-NEXT:    v_lshlrev_b32_e32 v6, 16, v5
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v1, v1, v5, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v6, v6
; GFX950-NEXT:    v_lshlrev_b32_e32 v7, 16, v1
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v5, v5, v1, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v6, 16, v5
; GFX950-NEXT:    v_cmp_gt_f32_e32 vcc, v7, v6
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v6, v5, v1, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; GFX950-NEXT:    v_max_f32_e32 v6, v6, v6
; GFX950-NEXT:    v_cvt_pk_bf16_f32 v6, v6, s0
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v1
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v1, v6, v1, vcc
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v5
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v1, v1, v5, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v5, 16, v6
; GFX950-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v5
; GFX950-NEXT:    v_lshlrev_b32_e32 v5, 16, v0
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v1, v6, v1, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; GFX950-NEXT:    v_lshlrev_b32_e32 v5, 16, v4
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v0, v0, v4, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; GFX950-NEXT:    v_lshlrev_b32_e32 v6, 16, v0
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v4, v4, v0, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v5, 16, v4
; GFX950-NEXT:    v_cmp_gt_f32_e32 vcc, v6, v5
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v5, v4, v0, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; GFX950-NEXT:    v_max_f32_e32 v5, v5, v5
; GFX950-NEXT:    v_cvt_pk_bf16_f32 v5, v5, s0
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v0
; GFX950-NEXT:    s_mov_b32 s0, 0x5040100
; GFX950-NEXT:    v_perm_b32 v1, v10, v1, s0
; GFX950-NEXT:    v_cndmask_b32_e32 v0, v5, v0, vcc
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v4
; GFX950-NEXT:    v_perm_b32 v2, v9, v2, s0
; GFX950-NEXT:    v_perm_b32 v3, v8, v3, s0
; GFX950-NEXT:    v_cndmask_b32_e32 v0, v0, v4, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v4, 16, v5
; GFX950-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v4
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v0, v5, v0, vcc
; GFX950-NEXT:    v_perm_b32 v0, v11, v0, s0
; GFX950-NEXT:    s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_maximumnum_v8bf16:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT:    v_and_b32_e32 v8, 0xffff0000, v3
; GFX10-NEXT:    v_lshrrev_b32_e32 v9, 16, v7
; GFX10-NEXT:    v_lshrrev_b32_e32 v10, 16, v3
; GFX10-NEXT:    v_and_b32_e32 v11, 0xffff0000, v7
; GFX10-NEXT:    v_lshrrev_b32_e32 v13, 16, v6
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
; GFX10-NEXT:    v_lshrrev_b32_e32 v14, 16, v2
; GFX10-NEXT:    v_and_b32_e32 v15, 0xffff0000, v6
; GFX10-NEXT:    v_and_b32_e32 v16, 0xffff0000, v1
; GFX10-NEXT:    v_and_b32_e32 v17, 0xffff0000, v0
; GFX10-NEXT:    v_cndmask_b32_e32 v8, v10, v9, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v11, v11
; GFX10-NEXT:    v_and_b32_e32 v10, 0xffff0000, v2
; GFX10-NEXT:    v_lshrrev_b32_e32 v20, 16, v4
; GFX10-NEXT:    v_lshrrev_b32_e32 v21, 16, v0
; GFX10-NEXT:    v_lshlrev_b32_e32 v11, 16, v8
; GFX10-NEXT:    v_cndmask_b32_e32 v9, v9, v8, vcc_lo
; GFX10-NEXT:    v_and_b32_e32 v23, 0xffff0000, v4
; GFX10-NEXT:    v_lshlrev_b32_e32 v12, 16, v9
; GFX10-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v11, v12
; GFX10-NEXT:    v_lshrrev_b32_e32 v12, 16, v5
; GFX10-NEXT:    v_cndmask_b32_e32 v11, v9, v8, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v10, v10
; GFX10-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
; GFX10-NEXT:    v_cndmask_b32_e32 v10, v14, v13, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v15, v15
; GFX10-NEXT:    v_lshrrev_b32_e32 v14, 16, v1
; GFX10-NEXT:    v_and_b32_e32 v15, 0xffff0000, v5
; GFX10-NEXT:    v_max_f32_e32 v11, v11, v11
; GFX10-NEXT:    v_lshlrev_b32_e32 v18, 16, v10
; GFX10-NEXT:    v_cndmask_b32_e32 v13, v13, v10, vcc_lo
; GFX10-NEXT:    v_bfe_u32 v22, v11, 16, 1
; GFX10-NEXT:    v_lshlrev_b32_e32 v19, 16, v13
; GFX10-NEXT:    v_add3_u32 v22, v22, v11, 0x7fff
; GFX10-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v18, v19
; GFX10-NEXT:    v_or_b32_e32 v19, 0x400000, v11
; GFX10-NEXT:    v_cndmask_b32_e32 v18, v13, v10, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v16, v16
; GFX10-NEXT:    v_lshlrev_b32_e32 v16, 16, v18
; GFX10-NEXT:    v_cndmask_b32_e32 v14, v14, v12, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v11, v11
; GFX10-NEXT:    v_lshlrev_b32_e32 v18, 16, v14
; GFX10-NEXT:    v_cndmask_b32_e32 v11, v22, v19, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v15, v15
; GFX10-NEXT:    v_max_f32_e32 v15, v16, v16
; GFX10-NEXT:    v_cndmask_b32_e32 v12, v12, v14, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v17, v17
; GFX10-NEXT:    v_lshrrev_b32_e32 v17, 16, v11
; GFX10-NEXT:    v_bfe_u32 v19, v15, 16, 1
; GFX10-NEXT:    v_or_b32_e32 v22, 0x400000, v15
; GFX10-NEXT:    v_and_b32_e32 v11, 0xffff0000, v11
; GFX10-NEXT:    v_cndmask_b32_e32 v16, v21, v20, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v8
; GFX10-NEXT:    v_lshlrev_b32_e32 v21, 16, v12
; GFX10-NEXT:    v_add3_u32 v19, v19, v15, 0x7fff
; GFX10-NEXT:    v_cndmask_b32_e32 v8, v17, v8, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v9
; GFX10-NEXT:    v_cndmask_b32_e32 v8, v8, v9, vcc_lo
; GFX10-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v18, v21
; GFX10-NEXT:    v_cndmask_b32_e32 v9, v12, v14, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v23, v23
; GFX10-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; GFX10-NEXT:    v_cndmask_b32_e32 v18, v20, v16, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v15, v15
; GFX10-NEXT:    v_max_f32_e32 v9, v9, v9
; GFX10-NEXT:    v_lshlrev_b32_e32 v20, 16, v18
; GFX10-NEXT:    v_cndmask_b32_e32 v15, v19, v22, vcc_lo
; GFX10-NEXT:    v_lshlrev_b32_e32 v19, 16, v16
; GFX10-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v11
; GFX10-NEXT:    v_lshrrev_b32_e32 v11, 16, v15
; GFX10-NEXT:    v_and_b32_e32 v15, 0xffff0000, v15
; GFX10-NEXT:    v_cndmask_b32_e32 v8, v17, v8, vcc_lo
; GFX10-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v19, v20
; GFX10-NEXT:    v_bfe_u32 v19, v9, 16, 1
; GFX10-NEXT:    v_cndmask_b32_e32 v17, v18, v16, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v10
; GFX10-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
; GFX10-NEXT:    v_cndmask_b32_e32 v10, v11, v10, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v13
; GFX10-NEXT:    v_max_f32_e32 v17, v17, v17
; GFX10-NEXT:    v_cndmask_b32_e32 v10, v10, v13, vcc_lo
; GFX10-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v15
; GFX10-NEXT:    v_add3_u32 v13, v19, v9, 0x7fff
; GFX10-NEXT:    v_or_b32_e32 v19, 0x400000, v9
; GFX10-NEXT:    v_lshlrev_b32_e32 v15, 16, v7
; GFX10-NEXT:    v_cndmask_b32_e32 v10, v11, v10, vcc_lo
; GFX10-NEXT:    v_lshlrev_b32_e32 v11, 16, v3
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
; GFX10-NEXT:    v_cndmask_b32_e32 v9, v13, v19, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v11, v11
; GFX10-NEXT:    v_bfe_u32 v13, v17, 16, 1
; GFX10-NEXT:    v_or_b32_e32 v19, 0x400000, v17
; GFX10-NEXT:    v_lshrrev_b32_e32 v11, 16, v9
; GFX10-NEXT:    v_cndmask_b32_e32 v3, v3, v7, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v15, v15
; GFX10-NEXT:    v_add3_u32 v13, v13, v17, 0x7fff
; GFX10-NEXT:    v_and_b32_e32 v9, 0xffff0000, v9
; GFX10-NEXT:    v_cndmask_b32_e32 v7, v7, v3, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v14
; GFX10-NEXT:    v_lshlrev_b32_e32 v15, 16, v7
; GFX10-NEXT:    v_cndmask_b32_e32 v14, v11, v14, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v17, v17
; GFX10-NEXT:    v_lshlrev_b32_e32 v17, 16, v3
; GFX10-NEXT:    v_cndmask_b32_e32 v13, v13, v19, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v12
; GFX10-NEXT:    v_lshrrev_b32_e32 v19, 16, v13
; GFX10-NEXT:    v_cndmask_b32_e32 v12, v14, v12, vcc_lo
; GFX10-NEXT:    v_lshlrev_b32_e32 v14, 16, v2
; GFX10-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v17, v15
; GFX10-NEXT:    v_lshlrev_b32_e32 v17, 16, v6
; GFX10-NEXT:    v_and_b32_e32 v13, 0xffff0000, v13
; GFX10-NEXT:    v_cndmask_b32_e32 v15, v7, v3, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v14, v14
; GFX10-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; GFX10-NEXT:    v_cndmask_b32_e32 v2, v2, v6, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v16
; GFX10-NEXT:    v_cndmask_b32_e32 v14, v19, v16, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v17, v17
; GFX10-NEXT:    v_lshlrev_b32_e32 v17, 16, v1
; GFX10-NEXT:    v_cndmask_b32_e32 v6, v6, v2, vcc_lo
; GFX10-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v9
; GFX10-NEXT:    v_cndmask_b32_e32 v9, v11, v12, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v18
; GFX10-NEXT:    v_max_f32_e32 v11, v15, v15
; GFX10-NEXT:    v_lshlrev_b32_e32 v15, 16, v2
; GFX10-NEXT:    v_cndmask_b32_e32 v12, v14, v18, vcc_lo
; GFX10-NEXT:    v_lshlrev_b32_e32 v14, 16, v6
; GFX10-NEXT:    v_lshlrev_b32_e32 v18, 16, v0
; GFX10-NEXT:    v_bfe_u32 v16, v11, 16, 1
; GFX10-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v15, v14
; GFX10-NEXT:    v_lshlrev_b32_e32 v15, 16, v5
; GFX10-NEXT:    v_cndmask_b32_e32 v14, v6, v2, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v17, v17
; GFX10-NEXT:    v_lshlrev_b32_e32 v17, 16, v4
; GFX10-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
; GFX10-NEXT:    v_cndmask_b32_e32 v1, v1, v5, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v18, v18
; GFX10-NEXT:    v_or_b32_e32 v18, 0x400000, v11
; GFX10-NEXT:    v_max_f32_e32 v14, v14, v14
; GFX10-NEXT:    v_cndmask_b32_e32 v0, v0, v4, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v15, v15
; GFX10-NEXT:    v_add3_u32 v15, v16, v11, 0x7fff
; GFX10-NEXT:    v_lshlrev_b32_e32 v21, 16, v0
; GFX10-NEXT:    v_cndmask_b32_e32 v5, v5, v1, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v17, v17
; GFX10-NEXT:    v_lshlrev_b32_e32 v17, 16, v1
; GFX10-NEXT:    v_lshlrev_b32_e32 v16, 16, v5
; GFX10-NEXT:    v_cndmask_b32_e32 v4, v4, v0, vcc_lo
; GFX10-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v17, v16
; GFX10-NEXT:    v_lshlrev_b32_e32 v20, 16, v4
; GFX10-NEXT:    v_bfe_u32 v17, v14, 16, 1
; GFX10-NEXT:    v_cndmask_b32_e32 v16, v5, v1, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v11, v11
; GFX10-NEXT:    v_add3_u32 v17, v17, v14, 0x7fff
; GFX10-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
; GFX10-NEXT:    v_cndmask_b32_e32 v11, v15, v18, vcc_lo
; GFX10-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v21, v20
; GFX10-NEXT:    v_or_b32_e32 v18, 0x400000, v14
; GFX10-NEXT:    v_max_f32_e32 v16, v16, v16
; GFX10-NEXT:    v_cndmask_b32_e32 v15, v4, v0, vcc_lo
; GFX10-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v13
; GFX10-NEXT:    v_lshlrev_b32_e32 v13, 16, v15
; GFX10-NEXT:    v_cndmask_b32_e32 v12, v19, v12, vcc_lo
; GFX10-NEXT:    v_lshrrev_b32_e32 v15, 16, v11
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v3
; GFX10-NEXT:    v_bfe_u32 v19, v16, 16, 1
; GFX10-NEXT:    v_max_f32_e32 v13, v13, v13
; GFX10-NEXT:    v_and_b32_e32 v11, 0xffff0000, v11
; GFX10-NEXT:    v_cndmask_b32_e32 v3, v15, v3, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v14, v14
; GFX10-NEXT:    v_or_b32_e32 v20, 0x400000, v13
; GFX10-NEXT:    v_cndmask_b32_e32 v14, v17, v18, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v7
; GFX10-NEXT:    v_or_b32_e32 v18, 0x400000, v16
; GFX10-NEXT:    v_bfe_u32 v17, v13, 16, 1
; GFX10-NEXT:    v_cndmask_b32_e32 v3, v3, v7, vcc_lo
; GFX10-NEXT:    v_add3_u32 v7, v19, v16, 0x7fff
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v16, v16
; GFX10-NEXT:    v_lshrrev_b32_e32 v19, 16, v14
; GFX10-NEXT:    v_add3_u32 v17, v17, v13, 0x7fff
; GFX10-NEXT:    v_and_b32_e32 v14, 0xffff0000, v14
; GFX10-NEXT:    v_cndmask_b32_e32 v7, v7, v18, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v2
; GFX10-NEXT:    v_lshrrev_b32_e32 v16, 16, v7
; GFX10-NEXT:    v_cndmask_b32_e32 v2, v19, v2, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v13, v13
; GFX10-NEXT:    v_cndmask_b32_e32 v13, v17, v20, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v6
; GFX10-NEXT:    v_cndmask_b32_e32 v2, v2, v6, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v1
; GFX10-NEXT:    v_lshrrev_b32_e32 v6, 16, v13
; GFX10-NEXT:    v_cndmask_b32_e32 v1, v16, v1, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v0
; GFX10-NEXT:    v_cndmask_b32_e32 v0, v6, v0, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v5
; GFX10-NEXT:    v_cndmask_b32_e32 v1, v1, v5, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v4
; GFX10-NEXT:    v_and_b32_e32 v5, 0xffff0000, v13
; GFX10-NEXT:    v_cndmask_b32_e32 v0, v0, v4, vcc_lo
; GFX10-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v14
; GFX10-NEXT:    v_and_b32_e32 v4, 0xffff0000, v7
; GFX10-NEXT:    v_cndmask_b32_e32 v2, v19, v2, vcc_lo
; GFX10-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v5
; GFX10-NEXT:    v_perm_b32 v2, v10, v2, 0x5040100
; GFX10-NEXT:    v_cndmask_b32_e32 v0, v6, v0, vcc_lo
; GFX10-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v4
; GFX10-NEXT:    v_perm_b32 v0, v12, v0, 0x5040100
; GFX10-NEXT:    v_cndmask_b32_e32 v1, v16, v1, vcc_lo
; GFX10-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v11
; GFX10-NEXT:    v_perm_b32 v1, v9, v1, 0x5040100
; GFX10-NEXT:    v_cndmask_b32_e32 v3, v15, v3, vcc_lo
; GFX10-NEXT:    v_perm_b32 v3, v8, v3, 0x5040100
; GFX10-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-TRUE16-LABEL: v_maximumnum_v8bf16:
; GFX11-TRUE16:       ; %bb.0:
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xffff0000, v3
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xffff0000, v7
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xffff0000, v2
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xffff0000, v6
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xffff0000, v1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s0, v9, v9
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v13, 0xffff0000, v5
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xffff0000, v0
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v15, 0xffff0000, v4
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v8.l, v3.h, v7.h, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v10, v10
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s1, v13, v13
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s2, v15, v15
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v9.l, v7.h, v8.l, s0
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s0, v11, v11
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v10.l, v2.h, v6.h, vcc_lo
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v16.l, v8.l
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v12, v12
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v17.l, v9.l
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v11.l, v6.h, v10.l, s0
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 16, v16
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s0, v14, v14
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v13, 16, v17
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v14.l, v10.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v16.l, v11.l
; GFX11-TRUE16-NEXT:    v_cmp_gt_f32_e64 s3, v12, v13
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v12.l, v1.h, v5.h, vcc_lo
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v17, 16, v14
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v13.l, v0.h, v4.h, s0
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v14.l, v9.l, v8.l, s3
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v15.l, v5.h, v12.l, s1
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v8.l
; GFX11-TRUE16-NEXT:    v_cmp_gt_f32_e64 s0, v17, v16
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v17.l, v12.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v16.l, v14.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.l, v15.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.l, v13.l
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v14.l, v11.l, v10.l, s0
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s1, 0, v10.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.l, v14.l
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v14.l, v4.h, v13.l, s2
; GFX11-TRUE16-NEXT:    v_max_f32_e32 v21, v16, v16
; GFX11-TRUE16-NEXT:    v_cmp_gt_f32_e64 s2, v17, v18
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v18, 16, v19
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v17, 16, v20
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.l, v14.l
; GFX11-TRUE16-NEXT:    v_bfe_u32 v19, v21, 16, 1
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v16.l, v15.l, v12.l, s2
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v22, 0x400000, v21
; GFX11-TRUE16-NEXT:    v_max_f32_e32 v17, v17, v17
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s2, v21, v21
; GFX11-TRUE16-NEXT:    v_add3_u32 v19, v19, v21, 0x7fff
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
; GFX11-TRUE16-NEXT:    v_bfe_u32 v23, v17, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s0, 0, v9.l
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e64 v19, v19, v22, s2
; GFX11-TRUE16-NEXT:    v_cmp_gt_f32_e64 s2, v18, v20
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v18, 0x400000, v17
; GFX11-TRUE16-NEXT:    v_add3_u32 v20, v23, v17, 0x7fff
; GFX11-TRUE16-NEXT:    v_max_f32_e32 v16, v16, v16
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v8.l, v19.h, v8.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v17, v17
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xffff0000, v19
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s3, 0, v13.l
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v8.l, v8.l, v9.l, s0
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v17, v20, v18, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v21
; GFX11-TRUE16-NEXT:    v_bfe_u32 v18, v16, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s0, 0, v11.l
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v9.l, v17.h, v10.l, s1
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v10.l, v14.l, v13.l, s2
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v8.h, v19.h, v8.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_add3_u32 v18, v18, v16, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v19, 0x400000, v16
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v8.l, v9.l, v11.l, s0
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 16, v3
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v16, v16
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v16, 0xffff0000, v17
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s0, v9, v9
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v11, v18, v19, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v12.l
; GFX11-TRUE16-NEXT:    v_dual_max_f32 v10, v10, v10 :: v_dual_lshlrev_b32 v9, 16, v7
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v3.l, v3.l, v7.l, s0
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s0, 0, v15.l
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v3.h, v11.h, v12.l, vcc_lo
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xffff0000, v11
; GFX11-TRUE16-NEXT:    v_bfe_u32 v12, v10, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_eq_f32_e64 s1, 0, v16
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v3.h, v3.h, v15.l, s0
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v7.l, v7.l, v3.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v15, 16, v2
; GFX11-TRUE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v9
; GFX11-TRUE16-NEXT:    v_add3_u32 v9, v12, v10, 0x7fff
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v16.l, v3.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v12.l, v7.l
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s0, v15, v15
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v15, 16, v6
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v18, 0x400000, v10
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v2.l, v2.l, v6.l, s0
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s0, v15, v15
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s2, v10, v10
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v3.h, v11.h, v3.h, vcc_lo
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v11, 16, v1
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v15, 16, v0
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v6.l, v6.l, v2.l, s0
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e64 v10, v9, v18, s2
; GFX11-TRUE16-NEXT:    v_cmp_gt_f32_e64 s2, v16, v12
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s0, v11, v11
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v11, 16, v5
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v2.h, v17.h, v8.l, s1
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v6.h, v10.h, v13.l, s3
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v8.l, v7.l, v3.l, s2
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v9.l, v6.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v13.l, v2.l
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s1, v15, v15
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v15, 16, v4
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v1.l, v1.l, v5.l, s0
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s0, v11, v11
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v12.l, v8.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v0.l, v0.l, v4.l, s1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s1, v15, v15
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v5.l, v5.l, v1.l, s0
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v14.l
; GFX11-TRUE16-NEXT:    v_cmp_gt_f32_e64 s2, v13, v9
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 16, v12
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v4.l, v4.l, v0.l, s1
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v11.l, v5.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v12.l, v1.l
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v0.h, v6.h, v14.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v8.l, v6.l, v2.l, s2
; GFX11-TRUE16-NEXT:    v_max_f32_e32 v9, v9, v9
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v14.l, v4.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v15.l, v0.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v13.l, v8.l
; GFX11-TRUE16-NEXT:    v_bfe_u32 v16, v9, 16, 1
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; GFX11-TRUE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v12, v11
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
; GFX11-TRUE16-NEXT:    v_add3_u32 v11, v16, v9, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, 0x400000, v9
; GFX11-TRUE16-NEXT:    v_cmp_gt_f32_e64 s0, v15, v14
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v8.l, v5.l, v1.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
; GFX11-TRUE16-NEXT:    v_max_f32_e32 v13, v13, v13
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v15, 0xffff0000, v10
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v9.l, v4.l, v0.l, s0
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v11, v11, v12, vcc_lo
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v12.l, v8.l
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v3.l
; GFX11-TRUE16-NEXT:    v_bfe_u32 v14, v13, 16, 1
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v16, 0x400000, v13
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v1.h, v11.h, v3.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_add3_u32 v14, v14, v13, 0x7fff
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v7.l
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s0, v13, v13
; GFX11-TRUE16-NEXT:    v_dual_max_f32 v12, v12, v12 :: v_dual_max_f32 v9, v9, v9
; GFX11-TRUE16-NEXT:    v_cmp_eq_f32_e64 s1, 0, v15
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v1.h, v1.h, v7.l, vcc_lo
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e64 v13, v14, v16, s0
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s0, 0, v2.l
; GFX11-TRUE16-NEXT:    v_bfe_u32 v14, v12, 16, 1
; GFX11-TRUE16-NEXT:    v_bfe_u32 v7, v9, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v15, 0x400000, v12
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v16, 0x400000, v9
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v2.l, v13.h, v2.l, s0
; GFX11-TRUE16-NEXT:    v_add3_u32 v14, v14, v12, 0x7fff
; GFX11-TRUE16-NEXT:    v_add3_u32 v7, v7, v9, 0x7fff
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s0, v12, v12
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s2, v9, v9
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v6.l
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xffff0000, v11
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v0.h, v10.h, v0.h, s1
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e64 v12, v14, v15, s0
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s0, 0, v1.l
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e64 v7, v7, v16, s2
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s2, 0, v0.l
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v2.l, v2.l, v6.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v5.l
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v1.l, v12.h, v1.l, s0
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s0, 0, v4.l
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v0.l, v7.h, v0.l, s2
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xffff0000, v12
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xffff0000, v7
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v15, 0xffff0000, v13
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v1.l, v1.l, v5.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v0.l, v0.l, v4.l, s0
; GFX11-TRUE16-NEXT:    v_cmp_eq_f32_e64 s0, 0, v14
; GFX11-TRUE16-NEXT:    v_cmp_eq_f32_e64 s3, 0, v9
; GFX11-TRUE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v6
; GFX11-TRUE16-NEXT:    v_cmp_eq_f32_e64 s2, 0, v15
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v3.l, v12.h, v1.l, s0
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v8.l, v11.h, v1.h, s3
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v0.l, v7.h, v0.l, vcc_lo
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v2.l, v13.h, v2.l, s2
; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v1, v3
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v3, v8
; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-FAKE16-LABEL: v_maximumnum_v8bf16:
; GFX11-FAKE16:       ; %bb.0:
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v13, 16, v6
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v8, 0xffff0000, v3
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v9, 16, v7
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v10, 16, v3
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v14, 16, v2
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v20, 16, v4
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v11, 0xffff0000, v7
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v21, 16, v0
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v8, v10, v9, vcc_lo
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v11, v11
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xffff0000, v2
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v11, 16, v8
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v9, v9, v8, vcc_lo
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v17, 0xffff0000, v0
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v12, 16, v9
; GFX11-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v11, v12
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v12, 16, v5
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v11, v9, v8, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v10, v10
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v10, v14, v13, vcc_lo
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v14, 16, v1
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v15, 0xffff0000, v6
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v23, 0xffff0000, v4
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v18, 16, v10
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v15, v15
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v13, v13, v10, vcc_lo
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v19, 16, v13
; GFX11-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v18, v19
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v18, v13, v10 :: v_dual_lshlrev_b32 v11, 16, v11
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_dual_max_f32 v11, v11, v11 :: v_dual_and_b32 v16, 0xffff0000, v1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v16, v16
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v22, v11, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v19, 0x400000, v11
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v16, 16, v18
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v14, v14, v12 :: v_dual_and_b32 v15, 0xffff0000, v5
; GFX11-FAKE16-NEXT:    v_add3_u32 v22, v22, v11, 0x7fff
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v11, v11
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v11, v22, v19 :: v_dual_lshlrev_b32 v18, 16, v14
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v15, v15
; GFX11-FAKE16-NEXT:    v_max_f32_e32 v15, v16, v16
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v12, v12, v14, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v17, v17
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v17, 16, v11
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v11, 0xffff0000, v11
; GFX11-FAKE16-NEXT:    v_bfe_u32 v19, v15, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v22, 0x400000, v15
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v16, v21, v20, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v8
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v21, 16, v12
; GFX11-FAKE16-NEXT:    v_add3_u32 v19, v19, v15, 0x7fff
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v8, v17, v8, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v9
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v8, v8, v9, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v18, v21
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v9, v12, v14, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v23, v23
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v18, v20, v16, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v15, v15
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v15, v19, v22, vcc_lo
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v19, 16, v16
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v20, 16, v18
; GFX11-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v11
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v11, 16, v15
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v8, v17, v8 :: v_dual_and_b32 v15, 0xffff0000, v15
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v19, v20
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v17, v18, v16, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v10
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v10, v11, v10 :: v_dual_lshlrev_b32 v17, 16, v17
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v13
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_max_f32_e32 v17, v17, v17
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v10, v10, v13, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v15
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v10, v11, v10 :: v_dual_lshlrev_b32 v15, 16, v7
; GFX11-FAKE16-NEXT:    v_max_f32_e32 v9, v9, v9
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v11, 16, v3
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v19, v9, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
; GFX11-FAKE16-NEXT:    v_add3_u32 v13, v19, v9, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v19, 0x400000, v9
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v9, v13, v19, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v11, v11
; GFX11-FAKE16-NEXT:    v_bfe_u32 v13, v17, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v19, 0x400000, v17
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v11, 16, v9
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v3, v3, v7, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v15, v15
; GFX11-FAKE16-NEXT:    v_add3_u32 v13, v13, v17, 0x7fff
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v7, v7, v3, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v14
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v14, v11, v14 :: v_dual_lshlrev_b32 v15, 16, v7
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v17, v17
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v17, 16, v3
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v13, v13, v19, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v12
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v19, 16, v13
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v12, v14, v12, vcc_lo
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v14, 16, v2
; GFX11-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v17, v15
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v17, 16, v6
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v15, v7, v3, vcc_lo
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v14, v14
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v2, v6, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v16
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v14, v19, v16, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v17, v17
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v6, v6, v2 :: v_dual_lshlrev_b32 v17, 16, v1
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v9, 0xffff0000, v9
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_4) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v9
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v9, v11, v12, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v18
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v12, v14, v18 :: v_dual_lshlrev_b32 v15, 16, v15
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v14, 16, v6
; GFX11-FAKE16-NEXT:    v_dual_max_f32 v11, v15, v15 :: v_dual_lshlrev_b32 v18, 16, v0
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v15, 16, v2
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v15, v14
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v14, v6, v2 :: v_dual_lshlrev_b32 v15, 16, v5
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v17, v17
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v17, 16, v4
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v1, v1, v5 :: v_dual_lshlrev_b32 v14, 16, v14
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v18, v18
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_max_f32_e32 v14, v14, v14
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v0, v4, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v15, v15
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v21, 16, v0
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v5, v5, v1, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v17, v17
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v4, v4, v0 :: v_dual_lshlrev_b32 v17, 16, v1
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v20, 16, v4
; GFX11-FAKE16-NEXT:    v_bfe_u32 v16, v11, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v18, 0x400000, v11
; GFX11-FAKE16-NEXT:    v_add3_u32 v15, v16, v11, 0x7fff
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v16, 16, v5
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v17, v16
; GFX11-FAKE16-NEXT:    v_bfe_u32 v17, v14, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v16, v5, v1, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v11, v11
; GFX11-FAKE16-NEXT:    v_add3_u32 v17, v17, v14, 0x7fff
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v11, v15, v18 :: v_dual_lshlrev_b32 v16, 16, v16
; GFX11-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v21, v20
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v18, 0x400000, v14
; GFX11-FAKE16-NEXT:    v_dual_max_f32 v16, v16, v16 :: v_dual_and_b32 v13, 0xffff0000, v13
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v15, v4, v0, vcc_lo
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v13
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v13, 16, v15
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v15, 16, v11
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v12, v19, v12, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v3
; GFX11-FAKE16-NEXT:    v_bfe_u32 v19, v16, 16, 1
; GFX11-FAKE16-NEXT:    v_max_f32_e32 v13, v13, v13
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v3, v15, v3, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v14, v14
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v20, 0x400000, v13
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v14, v17, v18, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v7
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v18, 0x400000, v16
; GFX11-FAKE16-NEXT:    v_bfe_u32 v17, v13, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v3, v3, v7, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v7, v19, v16, 0x7fff
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v16, v16
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v19, 16, v14
; GFX11-FAKE16-NEXT:    v_add3_u32 v17, v17, v13, 0x7fff
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v14, 0xffff0000, v14
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v7, v7, v18, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v2
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_4) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v16, 16, v7
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v19, v2, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v13, v13
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v13, v17, v20, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v6
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v2, v6, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v1
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v6, 16, v13
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v16, v1, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v0
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v6, v0, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v5
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v1, v5, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v4
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v0, v0, v4 :: v_dual_and_b32 v5, 0xffff0000, v13
; GFX11-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v14
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xffff0000, v7
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v2, v19, v2 :: v_dual_and_b32 v11, 0xffff0000, v11
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v5
; GFX11-FAKE16-NEXT:    v_perm_b32 v2, v10, v2, 0x5040100
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v6, v0, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v4
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_perm_b32 v0, v12, v0, 0x5040100
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v16, v1, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v11
; GFX11-FAKE16-NEXT:    v_perm_b32 v1, v9, v1, 0x5040100
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v3, v15, v3, vcc_lo
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_perm_b32 v3, v8, v3, 0x5040100
; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
;
; GFX12-TRUE16-LABEL: v_maximumnum_v8bf16:
; GFX12-TRUE16:       ; %bb.0:
; GFX12-TRUE16-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX12-TRUE16-NEXT:    s_wait_expcnt 0x0
; GFX12-TRUE16-NEXT:    s_wait_samplecnt 0x0
; GFX12-TRUE16-NEXT:    s_wait_bvhcnt 0x0
; GFX12-TRUE16-NEXT:    s_wait_kmcnt 0x0
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v8, 0xffff0000, v3
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v9, 0xffff0000, v7
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v10, 0xffff0000, v2
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v11, 0xffff0000, v6
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v12, 0xffff0000, v1
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s0, v9, v9
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v13, 0xffff0000, v5
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v14, 0xffff0000, v0
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v15, 0xffff0000, v4
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v8.l, v3.h, v7.h, vcc_lo
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v10, v10
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s1, v13, v13
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s2, v15, v15
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v9.l, v7.h, v8.l, s0
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s0, v11, v11
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v10.l, v2.h, v6.h, vcc_lo
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v16.l, v8.l
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v12, v12
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v17.l, v9.l
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v11.l, v6.h, v10.l, s0
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 16, v16
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s0, v14, v14
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v13, 16, v17
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v14.l, v10.l
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v16.l, v11.l
; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX12-TRUE16-NEXT:    v_cmp_gt_f32_e64 s3, v12, v13
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v12.l, v1.h, v5.h, vcc_lo
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v17, 16, v14
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v13.l, v0.h, v4.h, s0
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v14.l, v9.l, v8.l, s3
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v15.l, v5.h, v12.l, s1
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v8.l
; GFX12-TRUE16-NEXT:    v_cmp_gt_f32_e64 s0, v17, v16
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v17.l, v12.l
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v16.l, v14.l
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v18.l, v15.l
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v19.l, v13.l
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v14.l, v11.l, v10.l, s0
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s1, 0, v10.l
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v20.l, v14.l
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v14.l, v4.h, v13.l, s2
; GFX12-TRUE16-NEXT:    v_max_num_f32_e32 v21, v16, v16
; GFX12-TRUE16-NEXT:    v_cmp_gt_f32_e64 s2, v17, v18
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v18, 16, v19
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v17, 16, v20
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v20.l, v14.l
; GFX12-TRUE16-NEXT:    v_bfe_u32 v19, v21, 16, 1
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v16.l, v15.l, v12.l, s2
; GFX12-TRUE16-NEXT:    v_or_b32_e32 v22, 0x400000, v21
; GFX12-TRUE16-NEXT:    v_max_num_f32_e32 v17, v17, v17
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s2, v21, v21
; GFX12-TRUE16-NEXT:    v_add3_u32 v19, v19, v21, 0x7fff
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
; GFX12-TRUE16-NEXT:    v_bfe_u32 v23, v17, 16, 1
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s0, 0, v9.l
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b32_e64 v19, v19, v22, s2
; GFX12-TRUE16-NEXT:    v_cmp_gt_f32_e64 s2, v18, v20
; GFX12-TRUE16-NEXT:    v_or_b32_e32 v18, 0x400000, v17
; GFX12-TRUE16-NEXT:    v_add3_u32 v20, v23, v17, 0x7fff
; GFX12-TRUE16-NEXT:    v_max_num_f32_e32 v16, v16, v16
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v8.l, v19.h, v8.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v17, v17
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v21, 0xffff0000, v19
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s3, 0, v13.l
; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v8.l, v8.l, v9.l, s0
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b32_e32 v17, v20, v18, vcc_lo
; GFX12-TRUE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v21
; GFX12-TRUE16-NEXT:    v_bfe_u32 v18, v16, 16, 1
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s0, 0, v11.l
; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v9.l, v17.h, v10.l, s1
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v10.l, v14.l, v13.l, s2
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v8.h, v19.h, v8.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_add3_u32 v18, v18, v16, 0x7fff
; GFX12-TRUE16-NEXT:    v_or_b32_e32 v19, 0x400000, v16
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v8.l, v9.l, v11.l, s0
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 16, v3
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v16, v16
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v16, 0xffff0000, v17
; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s0, v9, v9
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b32_e32 v11, v18, v19, vcc_lo
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v12.l
; GFX12-TRUE16-NEXT:    v_dual_max_num_f32 v10, v10, v10 :: v_dual_lshlrev_b32 v9, 16, v7
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v3.l, v3.l, v7.l, s0
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s0, 0, v15.l
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v3.h, v11.h, v12.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v9, 0xffff0000, v11
; GFX12-TRUE16-NEXT:    v_bfe_u32 v12, v10, 16, 1
; GFX12-TRUE16-NEXT:    v_cmp_eq_f32_e64 s1, 0, v16
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v3.h, v3.h, v15.l, s0
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v7.l, v7.l, v3.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v15, 16, v2
; GFX12-TRUE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v9
; GFX12-TRUE16-NEXT:    v_add3_u32 v9, v12, v10, 0x7fff
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v16.l, v3.l
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v12.l, v7.l
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s0, v15, v15
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v15, 16, v6
; GFX12-TRUE16-NEXT:    v_or_b32_e32 v18, 0x400000, v10
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v2.l, v2.l, v6.l, s0
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s0, v15, v15
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s2, v10, v10
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v3.h, v11.h, v3.h, vcc_lo
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v11, 16, v1
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v15, 16, v0
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v6.l, v6.l, v2.l, s0
; GFX12-TRUE16-NEXT:    v_cndmask_b32_e64 v10, v9, v18, s2
; GFX12-TRUE16-NEXT:    v_cmp_gt_f32_e64 s2, v16, v12
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s0, v11, v11
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v11, 16, v5
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v2.h, v17.h, v8.l, s1
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v6.h, v10.h, v13.l, s3
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v8.l, v7.l, v3.l, s2
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v9.l, v6.l
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v13.l, v2.l
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s1, v15, v15
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v15, 16, v4
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v1.l, v1.l, v5.l, s0
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s0, v11, v11
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v12.l, v8.l
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v0.l, v0.l, v4.l, s1
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s1, v15, v15
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v5.l, v5.l, v1.l, s0
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v14.l
; GFX12-TRUE16-NEXT:    v_cmp_gt_f32_e64 s2, v13, v9
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 16, v12
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v4.l, v4.l, v0.l, s1
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v11.l, v5.l
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v12.l, v1.l
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v0.h, v6.h, v14.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v8.l, v6.l, v2.l, s2
; GFX12-TRUE16-NEXT:    v_max_num_f32_e32 v9, v9, v9
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v14.l, v4.l
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v15.l, v0.l
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v13.l, v8.l
; GFX12-TRUE16-NEXT:    v_bfe_u32 v16, v9, 16, 1
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; GFX12-TRUE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v12, v11
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
; GFX12-TRUE16-NEXT:    v_add3_u32 v11, v16, v9, 0x7fff
; GFX12-TRUE16-NEXT:    v_or_b32_e32 v12, 0x400000, v9
; GFX12-TRUE16-NEXT:    v_cmp_gt_f32_e64 s0, v15, v14
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v8.l, v5.l, v1.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
; GFX12-TRUE16-NEXT:    v_max_num_f32_e32 v13, v13, v13
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v15, 0xffff0000, v10
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v9.l, v4.l, v0.l, s0
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b32_e32 v11, v11, v12, vcc_lo
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v12.l, v8.l
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v3.l
; GFX12-TRUE16-NEXT:    v_bfe_u32 v14, v13, 16, 1
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; GFX12-TRUE16-NEXT:    v_or_b32_e32 v16, 0x400000, v13
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v1.h, v11.h, v3.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_add3_u32 v14, v14, v13, 0x7fff
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v7.l
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s0, v13, v13
; GFX12-TRUE16-NEXT:    v_dual_max_num_f32 v12, v12, v12 :: v_dual_max_num_f32 v9, v9, v9
; GFX12-TRUE16-NEXT:    v_cmp_eq_f32_e64 s1, 0, v15
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v1.h, v1.h, v7.l, vcc_lo
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b32_e64 v13, v14, v16, s0
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s0, 0, v2.l
; GFX12-TRUE16-NEXT:    v_bfe_u32 v14, v12, 16, 1
; GFX12-TRUE16-NEXT:    v_bfe_u32 v7, v9, 16, 1
; GFX12-TRUE16-NEXT:    v_or_b32_e32 v15, 0x400000, v12
; GFX12-TRUE16-NEXT:    v_or_b32_e32 v16, 0x400000, v9
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v2.l, v13.h, v2.l, s0
; GFX12-TRUE16-NEXT:    v_add3_u32 v14, v14, v12, 0x7fff
; GFX12-TRUE16-NEXT:    v_add3_u32 v7, v7, v9, 0x7fff
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s0, v12, v12
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s2, v9, v9
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v6.l
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v9, 0xffff0000, v11
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v0.h, v10.h, v0.h, s1
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b32_e64 v12, v14, v15, s0
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s0, 0, v1.l
; GFX12-TRUE16-NEXT:    v_cndmask_b32_e64 v7, v7, v16, s2
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s2, 0, v0.l
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v2.l, v2.l, v6.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v5.l
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v1.l, v12.h, v1.l, s0
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s0, 0, v4.l
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v0.l, v7.h, v0.l, s2
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v14, 0xffff0000, v12
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v6, 0xffff0000, v7
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v15, 0xffff0000, v13
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v1.l, v1.l, v5.l, vcc_lo
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v0.l, v0.l, v4.l, s0
; GFX12-TRUE16-NEXT:    v_cmp_eq_f32_e64 s0, 0, v14
; GFX12-TRUE16-NEXT:    v_cmp_eq_f32_e64 s3, 0, v9
; GFX12-TRUE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v6
; GFX12-TRUE16-NEXT:    v_cmp_eq_f32_e64 s2, 0, v15
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v3.l, v12.h, v1.l, s0
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v8.l, v11.h, v1.h, s3
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v0.l, v7.h, v0.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v2.l, v13.h, v2.l, s2
; GFX12-TRUE16-NEXT:    v_mov_b32_e32 v1, v3
; GFX12-TRUE16-NEXT:    v_mov_b32_e32 v3, v8
; GFX12-TRUE16-NEXT:    s_setpc_b64 s[30:31]
;
; GFX12-FAKE16-LABEL: v_maximumnum_v8bf16:
; GFX12-FAKE16:       ; %bb.0:
; GFX12-FAKE16-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX12-FAKE16-NEXT:    s_wait_expcnt 0x0
; GFX12-FAKE16-NEXT:    s_wait_samplecnt 0x0
; GFX12-FAKE16-NEXT:    s_wait_bvhcnt 0x0
; GFX12-FAKE16-NEXT:    s_wait_kmcnt 0x0
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v13, 16, v6
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v8, 0xffff0000, v3
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v9, 16, v7
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v10, 16, v3
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v14, 16, v2
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v20, 16, v4
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v11, 0xffff0000, v7
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v21, 16, v0
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v8, v10, v9, vcc_lo
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v11, v11
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v10, 0xffff0000, v2
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v11, 16, v8
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v9, v9, v8, vcc_lo
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v17, 0xffff0000, v0
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v12, 16, v9
; GFX12-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v11, v12
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v12, 16, v5
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v11, v9, v8, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v10, v10
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v10, v14, v13, vcc_lo
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v14, 16, v1
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v15, 0xffff0000, v6
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v23, 0xffff0000, v4
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v18, 16, v10
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v15, v15
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v13, v13, v10, vcc_lo
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v19, 16, v13
; GFX12-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v18, v19
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v18, v13, v10 :: v_dual_lshlrev_b32 v11, 16, v11
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-FAKE16-NEXT:    v_dual_max_num_f32 v11, v11, v11 :: v_dual_and_b32 v16, 0xffff0000, v1
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v16, v16
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
; GFX12-FAKE16-NEXT:    v_bfe_u32 v22, v11, 16, 1
; GFX12-FAKE16-NEXT:    v_or_b32_e32 v19, 0x400000, v11
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v16, 16, v18
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v14, v14, v12 :: v_dual_and_b32 v15, 0xffff0000, v5
; GFX12-FAKE16-NEXT:    v_add3_u32 v22, v22, v11, 0x7fff
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v11, v11
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v11, v22, v19 :: v_dual_lshlrev_b32 v18, 16, v14
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v15, v15
; GFX12-FAKE16-NEXT:    v_max_num_f32_e32 v15, v16, v16
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v12, v12, v14, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v17, v17
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v17, 16, v11
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v11, 0xffff0000, v11
; GFX12-FAKE16-NEXT:    v_bfe_u32 v19, v15, 16, 1
; GFX12-FAKE16-NEXT:    v_or_b32_e32 v22, 0x400000, v15
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v16, v21, v20, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v8
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v21, 16, v12
; GFX12-FAKE16-NEXT:    v_add3_u32 v19, v19, v15, 0x7fff
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v8, v17, v8, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v9
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v8, v8, v9, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v18, v21
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v9, v12, v14, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v23, v23
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v18, v20, v16, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v15, v15
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v15, v19, v22, vcc_lo
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v19, 16, v16
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v20, 16, v18
; GFX12-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v11
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v11, 16, v15
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v8, v17, v8 :: v_dual_and_b32 v15, 0xffff0000, v15
; GFX12-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v19, v20
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v17, v18, v16, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v10
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v10, v11, v10 :: v_dual_lshlrev_b32 v17, 16, v17
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v13
; GFX12-FAKE16-NEXT:    v_max_num_f32_e32 v17, v17, v17
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v10, v10, v13, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v15
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v10, v11, v10 :: v_dual_lshlrev_b32 v15, 16, v7
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX12-FAKE16-NEXT:    v_max_num_f32_e32 v9, v9, v9
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v11, 16, v3
; GFX12-FAKE16-NEXT:    v_bfe_u32 v19, v9, 16, 1
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1)
; GFX12-FAKE16-NEXT:    v_add3_u32 v13, v19, v9, 0x7fff
; GFX12-FAKE16-NEXT:    v_or_b32_e32 v19, 0x400000, v9
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v9, v13, v19, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v11, v11
; GFX12-FAKE16-NEXT:    v_bfe_u32 v13, v17, 16, 1
; GFX12-FAKE16-NEXT:    v_or_b32_e32 v19, 0x400000, v17
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v11, 16, v9
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v3, v3, v7, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v15, v15
; GFX12-FAKE16-NEXT:    v_add3_u32 v13, v13, v17, 0x7fff
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_2)
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v7, v7, v3, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v14
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v14, v11, v14 :: v_dual_lshlrev_b32 v15, 16, v7
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v17, v17
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v17, 16, v3
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v13, v13, v19, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v12
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v19, 16, v13
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v12, v14, v12, vcc_lo
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v14, 16, v2
; GFX12-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v17, v15
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v17, 16, v6
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v15, v7, v3, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v14, v14
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v2, v6, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v16
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v14, v19, v16, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v17, v17
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v6, v6, v2 :: v_dual_lshlrev_b32 v17, 16, v1
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v9, 0xffff0000, v9
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX12-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v9
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v9, v11, v12, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v18
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v12, v14, v18 :: v_dual_lshlrev_b32 v15, 16, v15
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v14, 16, v6
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX12-FAKE16-NEXT:    v_dual_max_num_f32 v11, v15, v15 :: v_dual_lshlrev_b32 v18, 16, v0
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v15, 16, v2
; GFX12-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v15, v14
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v14, v6, v2 :: v_dual_lshlrev_b32 v15, 16, v5
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v17, v17
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v17, 16, v4
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v1, v1, v5 :: v_dual_lshlrev_b32 v14, 16, v14
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v18, v18
; GFX12-FAKE16-NEXT:    v_max_num_f32_e32 v14, v14, v14
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v0, v4, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v15, v15
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v21, 16, v0
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v5, v5, v1, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v17, v17
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v4, v4, v0 :: v_dual_lshlrev_b32 v17, 16, v1
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_2)
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v20, 16, v4
; GFX12-FAKE16-NEXT:    v_bfe_u32 v16, v11, 16, 1
; GFX12-FAKE16-NEXT:    v_or_b32_e32 v18, 0x400000, v11
; GFX12-FAKE16-NEXT:    v_add3_u32 v15, v16, v11, 0x7fff
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v16, 16, v5
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_4) | instid1(VALU_DEP_3)
; GFX12-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v17, v16
; GFX12-FAKE16-NEXT:    v_bfe_u32 v17, v14, 16, 1
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v16, v5, v1, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v11, v11
; GFX12-FAKE16-NEXT:    v_add3_u32 v17, v17, v14, 0x7fff
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v11, v15, v18 :: v_dual_lshlrev_b32 v16, 16, v16
; GFX12-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v21, v20
; GFX12-FAKE16-NEXT:    v_or_b32_e32 v18, 0x400000, v14
; GFX12-FAKE16-NEXT:    v_dual_max_num_f32 v16, v16, v16 :: v_dual_and_b32 v13, 0xffff0000, v13
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v15, v4, v0, vcc_lo
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX12-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v13
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v13, 16, v15
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v15, 16, v11
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v12, v19, v12, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v3
; GFX12-FAKE16-NEXT:    v_bfe_u32 v19, v16, 16, 1
; GFX12-FAKE16-NEXT:    v_max_num_f32_e32 v13, v13, v13
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v3, v15, v3, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v14, v14
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX12-FAKE16-NEXT:    v_or_b32_e32 v20, 0x400000, v13
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v14, v17, v18, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v7
; GFX12-FAKE16-NEXT:    v_or_b32_e32 v18, 0x400000, v16
; GFX12-FAKE16-NEXT:    v_bfe_u32 v17, v13, 16, 1
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v3, v3, v7, vcc_lo
; GFX12-FAKE16-NEXT:    v_add3_u32 v7, v19, v16, 0x7fff
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v16, v16
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v19, 16, v14
; GFX12-FAKE16-NEXT:    v_add3_u32 v17, v17, v13, 0x7fff
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v14, 0xffff0000, v14
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v7, v7, v18, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v2
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v16, 16, v7
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v19, v2, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v13, v13
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v13, v17, v20, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v6
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v2, v6, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v1
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v6, 16, v13
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v16, v1, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v0
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v6, v0, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v5
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v1, v5, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v4
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v0, v0, v4 :: v_dual_and_b32 v5, 0xffff0000, v13
; GFX12-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v14
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v4, 0xffff0000, v7
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v2, v19, v2 :: v_dual_and_b32 v11, 0xffff0000, v11
; GFX12-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v5
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_2)
; GFX12-FAKE16-NEXT:    v_perm_b32 v2, v10, v2, 0x5040100
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v6, v0, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v4
; GFX12-FAKE16-NEXT:    v_perm_b32 v0, v12, v0, 0x5040100
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v16, v1, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v11
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1)
; GFX12-FAKE16-NEXT:    v_perm_b32 v1, v9, v1, 0x5040100
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v3, v15, v3, vcc_lo
; GFX12-FAKE16-NEXT:    v_perm_b32 v3, v8, v3, 0x5040100
; GFX12-FAKE16-NEXT:    s_setpc_b64 s[30:31]
  %result = call <8 x bfloat> @llvm.maximumnum.v8bf16(<8 x bfloat> %x, <8 x bfloat> %y)
  ret <8 x bfloat> %result
}

define <16 x bfloat> @v_maximumnum_v16bf16(<16 x bfloat> %x, <16 x bfloat> %y) {
; GFX7-LABEL: v_maximumnum_v16bf16:
; GFX7:       ; %bb.0:
; GFX7-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT:    v_mul_f32_e32 v11, 1.0, v11
; GFX7-NEXT:    v_mul_f32_e32 v27, 1.0, v27
; GFX7-NEXT:    v_and_b32_e32 v27, 0xffff0000, v27
; GFX7-NEXT:    v_and_b32_e32 v11, 0xffff0000, v11
; GFX7-NEXT:    v_max_f32_e32 v11, v11, v27
; GFX7-NEXT:    buffer_load_dword v27, off, s[0:3], s32
; GFX7-NEXT:    v_mul_f32_e32 v6, 1.0, v6
; GFX7-NEXT:    v_mul_f32_e32 v22, 1.0, v22
; GFX7-NEXT:    v_and_b32_e32 v22, 0xffff0000, v22
; GFX7-NEXT:    v_and_b32_e32 v6, 0xffff0000, v6
; GFX7-NEXT:    v_mul_f32_e32 v14, 1.0, v14
; GFX7-NEXT:    v_mul_f32_e32 v30, 1.0, v30
; GFX7-NEXT:    v_mul_f32_e32 v13, 1.0, v13
; GFX7-NEXT:    v_mul_f32_e32 v29, 1.0, v29
; GFX7-NEXT:    v_mul_f32_e32 v12, 1.0, v12
; GFX7-NEXT:    v_mul_f32_e32 v28, 1.0, v28
; GFX7-NEXT:    v_mul_f32_e32 v10, 1.0, v10
; GFX7-NEXT:    v_mul_f32_e32 v26, 1.0, v26
; GFX7-NEXT:    v_mul_f32_e32 v9, 1.0, v9
; GFX7-NEXT:    v_mul_f32_e32 v25, 1.0, v25
; GFX7-NEXT:    v_mul_f32_e32 v8, 1.0, v8
; GFX7-NEXT:    v_mul_f32_e32 v24, 1.0, v24
; GFX7-NEXT:    v_mul_f32_e32 v7, 1.0, v7
; GFX7-NEXT:    v_mul_f32_e32 v23, 1.0, v23
; GFX7-NEXT:    v_mul_f32_e32 v15, 1.0, v15
; GFX7-NEXT:    v_mul_f32_e32 v5, 1.0, v5
; GFX7-NEXT:    v_max_f32_e32 v6, v6, v22
; GFX7-NEXT:    v_mul_f32_e32 v21, 1.0, v21
; GFX7-NEXT:    v_mul_f32_e32 v4, 1.0, v4
; GFX7-NEXT:    v_mul_f32_e32 v20, 1.0, v20
; GFX7-NEXT:    v_mul_f32_e32 v3, 1.0, v3
; GFX7-NEXT:    v_mul_f32_e32 v19, 1.0, v19
; GFX7-NEXT:    v_mul_f32_e32 v2, 1.0, v2
; GFX7-NEXT:    v_mul_f32_e32 v18, 1.0, v18
; GFX7-NEXT:    v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT:    v_mul_f32_e32 v17, 1.0, v17
; GFX7-NEXT:    v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT:    v_mul_f32_e32 v16, 1.0, v16
; GFX7-NEXT:    v_and_b32_e32 v30, 0xffff0000, v30
; GFX7-NEXT:    v_and_b32_e32 v14, 0xffff0000, v14
; GFX7-NEXT:    v_and_b32_e32 v29, 0xffff0000, v29
; GFX7-NEXT:    v_and_b32_e32 v13, 0xffff0000, v13
; GFX7-NEXT:    v_and_b32_e32 v28, 0xffff0000, v28
; GFX7-NEXT:    v_and_b32_e32 v12, 0xffff0000, v12
; GFX7-NEXT:    v_and_b32_e32 v26, 0xffff0000, v26
; GFX7-NEXT:    v_and_b32_e32 v10, 0xffff0000, v10
; GFX7-NEXT:    v_and_b32_e32 v25, 0xffff0000, v25
; GFX7-NEXT:    v_and_b32_e32 v9, 0xffff0000, v9
; GFX7-NEXT:    v_and_b32_e32 v24, 0xffff0000, v24
; GFX7-NEXT:    v_and_b32_e32 v8, 0xffff0000, v8
; GFX7-NEXT:    v_and_b32_e32 v23, 0xffff0000, v23
; GFX7-NEXT:    v_and_b32_e32 v7, 0xffff0000, v7
; GFX7-NEXT:    v_and_b32_e32 v15, 0xffff0000, v15
; GFX7-NEXT:    v_and_b32_e32 v21, 0xffff0000, v21
; GFX7-NEXT:    v_and_b32_e32 v5, 0xffff0000, v5
; GFX7-NEXT:    v_and_b32_e32 v20, 0xffff0000, v20
; GFX7-NEXT:    v_and_b32_e32 v4, 0xffff0000, v4
; GFX7-NEXT:    v_and_b32_e32 v19, 0xffff0000, v19
; GFX7-NEXT:    v_and_b32_e32 v3, 0xffff0000, v3
; GFX7-NEXT:    v_and_b32_e32 v18, 0xffff0000, v18
; GFX7-NEXT:    v_and_b32_e32 v2, 0xffff0000, v2
; GFX7-NEXT:    v_and_b32_e32 v17, 0xffff0000, v17
; GFX7-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT:    v_and_b32_e32 v16, 0xffff0000, v16
; GFX7-NEXT:    v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT:    v_max_f32_e32 v14, v14, v30
; GFX7-NEXT:    v_max_f32_e32 v13, v13, v29
; GFX7-NEXT:    v_max_f32_e32 v12, v12, v28
; GFX7-NEXT:    v_max_f32_e32 v10, v10, v26
; GFX7-NEXT:    v_max_f32_e32 v9, v9, v25
; GFX7-NEXT:    v_max_f32_e32 v8, v8, v24
; GFX7-NEXT:    v_max_f32_e32 v7, v7, v23
; GFX7-NEXT:    v_max_f32_e32 v5, v5, v21
; GFX7-NEXT:    v_max_f32_e32 v4, v4, v20
; GFX7-NEXT:    v_max_f32_e32 v3, v3, v19
; GFX7-NEXT:    v_max_f32_e32 v2, v2, v18
; GFX7-NEXT:    v_max_f32_e32 v1, v1, v17
; GFX7-NEXT:    v_max_f32_e32 v0, v0, v16
; GFX7-NEXT:    v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT:    v_and_b32_e32 v2, 0xffff0000, v2
; GFX7-NEXT:    v_and_b32_e32 v3, 0xffff0000, v3
; GFX7-NEXT:    v_and_b32_e32 v4, 0xffff0000, v4
; GFX7-NEXT:    v_and_b32_e32 v5, 0xffff0000, v5
; GFX7-NEXT:    v_and_b32_e32 v6, 0xffff0000, v6
; GFX7-NEXT:    s_waitcnt vmcnt(0)
; GFX7-NEXT:    v_mul_f32_e32 v22, 1.0, v27
; GFX7-NEXT:    v_and_b32_e32 v22, 0xffff0000, v22
; GFX7-NEXT:    v_max_f32_e32 v15, v15, v22
; GFX7-NEXT:    v_and_b32_e32 v7, 0xffff0000, v7
; GFX7-NEXT:    v_and_b32_e32 v8, 0xffff0000, v8
; GFX7-NEXT:    v_and_b32_e32 v9, 0xffff0000, v9
; GFX7-NEXT:    v_and_b32_e32 v10, 0xffff0000, v10
; GFX7-NEXT:    v_and_b32_e32 v11, 0xffff0000, v11
; GFX7-NEXT:    v_and_b32_e32 v12, 0xffff0000, v12
; GFX7-NEXT:    v_and_b32_e32 v13, 0xffff0000, v13
; GFX7-NEXT:    v_and_b32_e32 v14, 0xffff0000, v14
; GFX7-NEXT:    v_and_b32_e32 v15, 0xffff0000, v15
; GFX7-NEXT:    s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_maximumnum_v16bf16:
; GFX8:       ; %bb.0:
; GFX8-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT:    v_and_b32_e32 v16, 0xffff0000, v7
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v16, v16
; GFX8-NEXT:    v_lshrrev_b32_e32 v16, 16, v15
; GFX8-NEXT:    v_lshrrev_b32_e32 v17, 16, v7
; GFX8-NEXT:    v_and_b32_e32 v19, 0xffff0000, v15
; GFX8-NEXT:    v_cndmask_b32_e32 v17, v17, v16, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v19, v19
; GFX8-NEXT:    v_cndmask_b32_e32 v16, v16, v17, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v18, 16, v17
; GFX8-NEXT:    v_lshlrev_b32_e32 v19, 16, v16
; GFX8-NEXT:    v_cmp_gt_f32_e32 vcc, v18, v19
; GFX8-NEXT:    v_cndmask_b32_e32 v18, v16, v17, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
; GFX8-NEXT:    v_mul_f32_e32 v18, 1.0, v18
; GFX8-NEXT:    v_bfe_u32 v19, v18, 16, 1
; GFX8-NEXT:    v_add_u32_e32 v19, vcc, v19, v18
; GFX8-NEXT:    s_movk_i32 s4, 0x7fff
; GFX8-NEXT:    v_add_u32_e32 v19, vcc, s4, v19
; GFX8-NEXT:    v_or_b32_e32 v20, 0x400000, v18
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; GFX8-NEXT:    v_cndmask_b32_e32 v18, v19, v20, vcc
; GFX8-NEXT:    v_lshrrev_b32_e32 v19, 16, v18
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v17
; GFX8-NEXT:    v_cndmask_b32_e32 v17, v19, v17, vcc
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v16
; GFX8-NEXT:    v_cndmask_b32_e32 v16, v17, v16, vcc
; GFX8-NEXT:    v_and_b32_e32 v17, 0xffff0000, v18
; GFX8-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v17
; GFX8-NEXT:    v_and_b32_e32 v17, 0xffff0000, v6
; GFX8-NEXT:    v_cndmask_b32_e32 v16, v19, v16, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX8-NEXT:    v_lshrrev_b32_e32 v17, 16, v14
; GFX8-NEXT:    v_lshrrev_b32_e32 v18, 16, v6
; GFX8-NEXT:    v_and_b32_e32 v20, 0xffff0000, v14
; GFX8-NEXT:    v_cndmask_b32_e32 v18, v18, v17, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v20, v20
; GFX8-NEXT:    v_cndmask_b32_e32 v17, v17, v18, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v19, 16, v18
; GFX8-NEXT:    v_lshlrev_b32_e32 v20, 16, v17
; GFX8-NEXT:    v_cmp_gt_f32_e32 vcc, v19, v20
; GFX8-NEXT:    v_cndmask_b32_e32 v19, v17, v18, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
; GFX8-NEXT:    v_mul_f32_e32 v19, 1.0, v19
; GFX8-NEXT:    v_bfe_u32 v20, v19, 16, 1
; GFX8-NEXT:    v_add_u32_e32 v20, vcc, v20, v19
; GFX8-NEXT:    v_add_u32_e32 v20, vcc, s4, v20
; GFX8-NEXT:    v_or_b32_e32 v21, 0x400000, v19
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v19, v19
; GFX8-NEXT:    v_cndmask_b32_e32 v19, v20, v21, vcc
; GFX8-NEXT:    v_lshrrev_b32_e32 v20, 16, v19
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v18
; GFX8-NEXT:    v_cndmask_b32_e32 v18, v20, v18, vcc
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v17
; GFX8-NEXT:    v_cndmask_b32_e32 v17, v18, v17, vcc
; GFX8-NEXT:    v_and_b32_e32 v18, 0xffff0000, v19
; GFX8-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v18
; GFX8-NEXT:    v_and_b32_e32 v18, 0xffff0000, v5
; GFX8-NEXT:    v_cndmask_b32_e32 v17, v20, v17, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; GFX8-NEXT:    v_lshrrev_b32_e32 v18, 16, v13
; GFX8-NEXT:    v_lshrrev_b32_e32 v19, 16, v5
; GFX8-NEXT:    v_and_b32_e32 v21, 0xffff0000, v13
; GFX8-NEXT:    v_cndmask_b32_e32 v19, v19, v18, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v21, v21
; GFX8-NEXT:    v_cndmask_b32_e32 v18, v18, v19, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v20, 16, v19
; GFX8-NEXT:    v_lshlrev_b32_e32 v21, 16, v18
; GFX8-NEXT:    v_cmp_gt_f32_e32 vcc, v20, v21
; GFX8-NEXT:    v_cndmask_b32_e32 v20, v18, v19, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
; GFX8-NEXT:    v_mul_f32_e32 v20, 1.0, v20
; GFX8-NEXT:    v_bfe_u32 v21, v20, 16, 1
; GFX8-NEXT:    v_add_u32_e32 v21, vcc, v21, v20
; GFX8-NEXT:    v_add_u32_e32 v21, vcc, s4, v21
; GFX8-NEXT:    v_or_b32_e32 v22, 0x400000, v20
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v20, v20
; GFX8-NEXT:    v_cndmask_b32_e32 v20, v21, v22, vcc
; GFX8-NEXT:    v_lshrrev_b32_e32 v21, 16, v20
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v19
; GFX8-NEXT:    v_cndmask_b32_e32 v19, v21, v19, vcc
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v18
; GFX8-NEXT:    v_cndmask_b32_e32 v18, v19, v18, vcc
; GFX8-NEXT:    v_and_b32_e32 v19, 0xffff0000, v20
; GFX8-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v19
; GFX8-NEXT:    v_and_b32_e32 v19, 0xffff0000, v4
; GFX8-NEXT:    v_cndmask_b32_e32 v18, v21, v18, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v19, v19
; GFX8-NEXT:    v_lshrrev_b32_e32 v19, 16, v12
; GFX8-NEXT:    v_lshrrev_b32_e32 v20, 16, v4
; GFX8-NEXT:    v_and_b32_e32 v22, 0xffff0000, v12
; GFX8-NEXT:    v_cndmask_b32_e32 v20, v20, v19, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v22, v22
; GFX8-NEXT:    v_cndmask_b32_e32 v19, v19, v20, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v21, 16, v20
; GFX8-NEXT:    v_lshlrev_b32_e32 v22, 16, v19
; GFX8-NEXT:    v_cmp_gt_f32_e32 vcc, v21, v22
; GFX8-NEXT:    v_cndmask_b32_e32 v21, v19, v20, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
; GFX8-NEXT:    v_mul_f32_e32 v21, 1.0, v21
; GFX8-NEXT:    v_bfe_u32 v22, v21, 16, 1
; GFX8-NEXT:    v_add_u32_e32 v22, vcc, v22, v21
; GFX8-NEXT:    v_add_u32_e32 v22, vcc, s4, v22
; GFX8-NEXT:    v_or_b32_e32 v23, 0x400000, v21
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v21, v21
; GFX8-NEXT:    v_cndmask_b32_e32 v21, v22, v23, vcc
; GFX8-NEXT:    v_lshrrev_b32_e32 v22, 16, v21
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v20
; GFX8-NEXT:    v_cndmask_b32_e32 v20, v22, v20, vcc
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v19
; GFX8-NEXT:    v_cndmask_b32_e32 v19, v20, v19, vcc
; GFX8-NEXT:    v_and_b32_e32 v20, 0xffff0000, v21
; GFX8-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v20
; GFX8-NEXT:    v_and_b32_e32 v20, 0xffff0000, v3
; GFX8-NEXT:    v_cndmask_b32_e32 v19, v22, v19, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v20, v20
; GFX8-NEXT:    v_lshrrev_b32_e32 v20, 16, v11
; GFX8-NEXT:    v_lshrrev_b32_e32 v21, 16, v3
; GFX8-NEXT:    v_and_b32_e32 v23, 0xffff0000, v11
; GFX8-NEXT:    v_cndmask_b32_e32 v21, v21, v20, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v23, v23
; GFX8-NEXT:    v_cndmask_b32_e32 v20, v20, v21, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v22, 16, v21
; GFX8-NEXT:    v_lshlrev_b32_e32 v23, 16, v20
; GFX8-NEXT:    v_cmp_gt_f32_e32 vcc, v22, v23
; GFX8-NEXT:    v_cndmask_b32_e32 v22, v20, v21, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v22, 16, v22
; GFX8-NEXT:    v_mul_f32_e32 v22, 1.0, v22
; GFX8-NEXT:    v_bfe_u32 v23, v22, 16, 1
; GFX8-NEXT:    v_add_u32_e32 v23, vcc, v23, v22
; GFX8-NEXT:    v_add_u32_e32 v23, vcc, s4, v23
; GFX8-NEXT:    v_or_b32_e32 v24, 0x400000, v22
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v22, v22
; GFX8-NEXT:    v_cndmask_b32_e32 v22, v23, v24, vcc
; GFX8-NEXT:    v_lshrrev_b32_e32 v23, 16, v22
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v21
; GFX8-NEXT:    v_cndmask_b32_e32 v21, v23, v21, vcc
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v20
; GFX8-NEXT:    v_cndmask_b32_e32 v20, v21, v20, vcc
; GFX8-NEXT:    v_and_b32_e32 v21, 0xffff0000, v22
; GFX8-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v21
; GFX8-NEXT:    v_and_b32_e32 v21, 0xffff0000, v2
; GFX8-NEXT:    v_cndmask_b32_e32 v20, v23, v20, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v21, v21
; GFX8-NEXT:    v_lshrrev_b32_e32 v21, 16, v10
; GFX8-NEXT:    v_lshrrev_b32_e32 v22, 16, v2
; GFX8-NEXT:    v_and_b32_e32 v24, 0xffff0000, v10
; GFX8-NEXT:    v_cndmask_b32_e32 v22, v22, v21, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v24, v24
; GFX8-NEXT:    v_cndmask_b32_e32 v21, v21, v22, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v23, 16, v22
; GFX8-NEXT:    v_lshlrev_b32_e32 v24, 16, v21
; GFX8-NEXT:    v_cmp_gt_f32_e32 vcc, v23, v24
; GFX8-NEXT:    v_cndmask_b32_e32 v23, v21, v22, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; GFX8-NEXT:    v_mul_f32_e32 v23, 1.0, v23
; GFX8-NEXT:    v_bfe_u32 v24, v23, 16, 1
; GFX8-NEXT:    v_add_u32_e32 v24, vcc, v24, v23
; GFX8-NEXT:    v_add_u32_e32 v24, vcc, s4, v24
; GFX8-NEXT:    v_or_b32_e32 v25, 0x400000, v23
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v23, v23
; GFX8-NEXT:    v_cndmask_b32_e32 v23, v24, v25, vcc
; GFX8-NEXT:    v_lshrrev_b32_e32 v24, 16, v23
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v22
; GFX8-NEXT:    v_cndmask_b32_e32 v22, v24, v22, vcc
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v21
; GFX8-NEXT:    v_cndmask_b32_e32 v21, v22, v21, vcc
; GFX8-NEXT:    v_and_b32_e32 v22, 0xffff0000, v23
; GFX8-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v22
; GFX8-NEXT:    v_and_b32_e32 v22, 0xffff0000, v1
; GFX8-NEXT:    v_cndmask_b32_e32 v21, v24, v21, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v22, v22
; GFX8-NEXT:    v_lshrrev_b32_e32 v22, 16, v9
; GFX8-NEXT:    v_lshrrev_b32_e32 v23, 16, v1
; GFX8-NEXT:    v_and_b32_e32 v25, 0xffff0000, v9
; GFX8-NEXT:    v_cndmask_b32_e32 v23, v23, v22, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v25, v25
; GFX8-NEXT:    v_cndmask_b32_e32 v22, v22, v23, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v24, 16, v23
; GFX8-NEXT:    v_lshlrev_b32_e32 v25, 16, v22
; GFX8-NEXT:    v_cmp_gt_f32_e32 vcc, v24, v25
; GFX8-NEXT:    v_cndmask_b32_e32 v24, v22, v23, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
; GFX8-NEXT:    v_mul_f32_e32 v24, 1.0, v24
; GFX8-NEXT:    v_bfe_u32 v25, v24, 16, 1
; GFX8-NEXT:    v_add_u32_e32 v25, vcc, v25, v24
; GFX8-NEXT:    v_add_u32_e32 v25, vcc, s4, v25
; GFX8-NEXT:    v_or_b32_e32 v26, 0x400000, v24
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v24, v24
; GFX8-NEXT:    v_cndmask_b32_e32 v24, v25, v26, vcc
; GFX8-NEXT:    v_lshrrev_b32_e32 v25, 16, v24
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v23
; GFX8-NEXT:    v_cndmask_b32_e32 v23, v25, v23, vcc
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v22
; GFX8-NEXT:    v_cndmask_b32_e32 v22, v23, v22, vcc
; GFX8-NEXT:    v_and_b32_e32 v23, 0xffff0000, v24
; GFX8-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v23
; GFX8-NEXT:    v_and_b32_e32 v23, 0xffff0000, v0
; GFX8-NEXT:    v_cndmask_b32_e32 v22, v25, v22, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v23, v23
; GFX8-NEXT:    v_lshrrev_b32_e32 v23, 16, v8
; GFX8-NEXT:    v_lshrrev_b32_e32 v24, 16, v0
; GFX8-NEXT:    v_and_b32_e32 v26, 0xffff0000, v8
; GFX8-NEXT:    v_cndmask_b32_e32 v24, v24, v23, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v26, v26
; GFX8-NEXT:    v_cndmask_b32_e32 v23, v23, v24, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v25, 16, v24
; GFX8-NEXT:    v_lshlrev_b32_e32 v26, 16, v23
; GFX8-NEXT:    v_cmp_gt_f32_e32 vcc, v25, v26
; GFX8-NEXT:    v_cndmask_b32_e32 v25, v23, v24, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v25, 16, v25
; GFX8-NEXT:    v_mul_f32_e32 v25, 1.0, v25
; GFX8-NEXT:    v_bfe_u32 v26, v25, 16, 1
; GFX8-NEXT:    v_add_u32_e32 v26, vcc, v26, v25
; GFX8-NEXT:    v_add_u32_e32 v26, vcc, s4, v26
; GFX8-NEXT:    v_or_b32_e32 v27, 0x400000, v25
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v25, v25
; GFX8-NEXT:    v_cndmask_b32_e32 v25, v26, v27, vcc
; GFX8-NEXT:    v_lshrrev_b32_e32 v26, 16, v25
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v24
; GFX8-NEXT:    v_cndmask_b32_e32 v24, v26, v24, vcc
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v23
; GFX8-NEXT:    v_cndmask_b32_e32 v23, v24, v23, vcc
; GFX8-NEXT:    v_and_b32_e32 v24, 0xffff0000, v25
; GFX8-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v24
; GFX8-NEXT:    v_lshlrev_b32_e32 v24, 16, v7
; GFX8-NEXT:    v_cndmask_b32_e32 v23, v26, v23, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v24, v24
; GFX8-NEXT:    v_lshlrev_b32_e32 v24, 16, v15
; GFX8-NEXT:    v_cndmask_b32_e32 v7, v7, v15, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v24, v24
; GFX8-NEXT:    v_cndmask_b32_e32 v15, v15, v7, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v24, 16, v15
; GFX8-NEXT:    v_lshlrev_b32_e32 v25, 16, v7
; GFX8-NEXT:    v_cmp_gt_f32_e32 vcc, v25, v24
; GFX8-NEXT:    v_cndmask_b32_e32 v24, v15, v7, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
; GFX8-NEXT:    v_mul_f32_e32 v24, 1.0, v24
; GFX8-NEXT:    v_bfe_u32 v25, v24, 16, 1
; GFX8-NEXT:    v_add_u32_e32 v25, vcc, v25, v24
; GFX8-NEXT:    v_add_u32_e32 v25, vcc, s4, v25
; GFX8-NEXT:    v_or_b32_e32 v26, 0x400000, v24
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v24, v24
; GFX8-NEXT:    v_cndmask_b32_e32 v24, v25, v26, vcc
; GFX8-NEXT:    v_lshrrev_b32_e32 v25, 16, v24
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v7
; GFX8-NEXT:    v_cndmask_b32_e32 v7, v25, v7, vcc
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v15
; GFX8-NEXT:    v_cndmask_b32_e32 v7, v7, v15, vcc
; GFX8-NEXT:    v_and_b32_e32 v15, 0xffff0000, v24
; GFX8-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v15
; GFX8-NEXT:    v_lshlrev_b32_e32 v15, 16, v6
; GFX8-NEXT:    v_cndmask_b32_e32 v7, v25, v7, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v15, v15
; GFX8-NEXT:    v_lshlrev_b32_e32 v15, 16, v14
; GFX8-NEXT:    v_cndmask_b32_e32 v6, v6, v14, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v15, v15
; GFX8-NEXT:    v_cndmask_b32_e32 v14, v14, v6, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v15, 16, v14
; GFX8-NEXT:    v_lshlrev_b32_e32 v24, 16, v6
; GFX8-NEXT:    v_cmp_gt_f32_e32 vcc, v24, v15
; GFX8-NEXT:    v_cndmask_b32_e32 v15, v14, v6, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; GFX8-NEXT:    v_mul_f32_e32 v15, 1.0, v15
; GFX8-NEXT:    v_bfe_u32 v24, v15, 16, 1
; GFX8-NEXT:    v_add_u32_e32 v24, vcc, v24, v15
; GFX8-NEXT:    v_add_u32_e32 v24, vcc, s4, v24
; GFX8-NEXT:    v_or_b32_e32 v25, 0x400000, v15
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v15, v15
; GFX8-NEXT:    v_cndmask_b32_e32 v15, v24, v25, vcc
; GFX8-NEXT:    v_lshrrev_b32_e32 v24, 16, v15
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v6
; GFX8-NEXT:    v_cndmask_b32_e32 v6, v24, v6, vcc
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v14
; GFX8-NEXT:    v_cndmask_b32_e32 v6, v6, v14, vcc
; GFX8-NEXT:    v_and_b32_e32 v14, 0xffff0000, v15
; GFX8-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v14
; GFX8-NEXT:    v_lshlrev_b32_e32 v14, 16, v5
; GFX8-NEXT:    v_cndmask_b32_e32 v6, v24, v6, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v14, v14
; GFX8-NEXT:    v_lshlrev_b32_e32 v14, 16, v13
; GFX8-NEXT:    v_cndmask_b32_e32 v5, v5, v13, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v14, v14
; GFX8-NEXT:    v_cndmask_b32_e32 v13, v13, v5, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v14, 16, v13
; GFX8-NEXT:    v_lshlrev_b32_e32 v15, 16, v5
; GFX8-NEXT:    v_cmp_gt_f32_e32 vcc, v15, v14
; GFX8-NEXT:    v_cndmask_b32_e32 v14, v13, v5, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
; GFX8-NEXT:    v_mul_f32_e32 v14, 1.0, v14
; GFX8-NEXT:    v_bfe_u32 v15, v14, 16, 1
; GFX8-NEXT:    v_add_u32_e32 v15, vcc, v15, v14
; GFX8-NEXT:    v_add_u32_e32 v15, vcc, s4, v15
; GFX8-NEXT:    v_or_b32_e32 v24, 0x400000, v14
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v14, v14
; GFX8-NEXT:    v_cndmask_b32_e32 v14, v15, v24, vcc
; GFX8-NEXT:    v_lshrrev_b32_e32 v15, 16, v14
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v5
; GFX8-NEXT:    v_cndmask_b32_e32 v5, v15, v5, vcc
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v13
; GFX8-NEXT:    v_cndmask_b32_e32 v5, v5, v13, vcc
; GFX8-NEXT:    v_and_b32_e32 v13, 0xffff0000, v14
; GFX8-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v13
; GFX8-NEXT:    v_lshlrev_b32_e32 v13, 16, v4
; GFX8-NEXT:    v_cndmask_b32_e32 v5, v15, v5, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v13, v13
; GFX8-NEXT:    v_lshlrev_b32_e32 v13, 16, v12
; GFX8-NEXT:    v_cndmask_b32_e32 v4, v4, v12, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v13, v13
; GFX8-NEXT:    v_cndmask_b32_e32 v12, v12, v4, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v13, 16, v12
; GFX8-NEXT:    v_lshlrev_b32_e32 v14, 16, v4
; GFX8-NEXT:    v_cmp_gt_f32_e32 vcc, v14, v13
; GFX8-NEXT:    v_cndmask_b32_e32 v13, v12, v4, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
; GFX8-NEXT:    v_mul_f32_e32 v13, 1.0, v13
; GFX8-NEXT:    v_bfe_u32 v14, v13, 16, 1
; GFX8-NEXT:    v_add_u32_e32 v14, vcc, v14, v13
; GFX8-NEXT:    v_add_u32_e32 v14, vcc, s4, v14
; GFX8-NEXT:    v_or_b32_e32 v15, 0x400000, v13
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v13, v13
; GFX8-NEXT:    v_cndmask_b32_e32 v13, v14, v15, vcc
; GFX8-NEXT:    v_lshrrev_b32_e32 v14, 16, v13
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v4
; GFX8-NEXT:    v_cndmask_b32_e32 v4, v14, v4, vcc
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v12
; GFX8-NEXT:    v_cndmask_b32_e32 v4, v4, v12, vcc
; GFX8-NEXT:    v_and_b32_e32 v12, 0xffff0000, v13
; GFX8-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v12
; GFX8-NEXT:    v_lshlrev_b32_e32 v12, 16, v3
; GFX8-NEXT:    v_cndmask_b32_e32 v4, v14, v4, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v12, v12
; GFX8-NEXT:    v_lshlrev_b32_e32 v12, 16, v11
; GFX8-NEXT:    v_cndmask_b32_e32 v3, v3, v11, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v12, v12
; GFX8-NEXT:    v_cndmask_b32_e32 v11, v11, v3, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v12, 16, v11
; GFX8-NEXT:    v_lshlrev_b32_e32 v13, 16, v3
; GFX8-NEXT:    v_cmp_gt_f32_e32 vcc, v13, v12
; GFX8-NEXT:    v_cndmask_b32_e32 v12, v11, v3, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; GFX8-NEXT:    v_mul_f32_e32 v12, 1.0, v12
; GFX8-NEXT:    v_bfe_u32 v13, v12, 16, 1
; GFX8-NEXT:    v_add_u32_e32 v13, vcc, v13, v12
; GFX8-NEXT:    v_add_u32_e32 v13, vcc, s4, v13
; GFX8-NEXT:    v_or_b32_e32 v14, 0x400000, v12
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v12, v12
; GFX8-NEXT:    v_cndmask_b32_e32 v12, v13, v14, vcc
; GFX8-NEXT:    v_lshrrev_b32_e32 v13, 16, v12
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v3
; GFX8-NEXT:    v_cndmask_b32_e32 v3, v13, v3, vcc
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v11
; GFX8-NEXT:    v_cndmask_b32_e32 v3, v3, v11, vcc
; GFX8-NEXT:    v_and_b32_e32 v11, 0xffff0000, v12
; GFX8-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v11
; GFX8-NEXT:    v_lshlrev_b32_e32 v11, 16, v2
; GFX8-NEXT:    v_cndmask_b32_e32 v3, v13, v3, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v11, v11
; GFX8-NEXT:    v_lshlrev_b32_e32 v11, 16, v10
; GFX8-NEXT:    v_cndmask_b32_e32 v2, v2, v10, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v11, v11
; GFX8-NEXT:    v_cndmask_b32_e32 v10, v10, v2, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v11, 16, v10
; GFX8-NEXT:    v_lshlrev_b32_e32 v12, 16, v2
; GFX8-NEXT:    v_cmp_gt_f32_e32 vcc, v12, v11
; GFX8-NEXT:    v_cndmask_b32_e32 v11, v10, v2, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
; GFX8-NEXT:    v_mul_f32_e32 v11, 1.0, v11
; GFX8-NEXT:    v_bfe_u32 v12, v11, 16, 1
; GFX8-NEXT:    v_add_u32_e32 v12, vcc, v12, v11
; GFX8-NEXT:    v_add_u32_e32 v12, vcc, s4, v12
; GFX8-NEXT:    v_or_b32_e32 v13, 0x400000, v11
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v11, v11
; GFX8-NEXT:    v_cndmask_b32_e32 v11, v12, v13, vcc
; GFX8-NEXT:    v_lshrrev_b32_e32 v12, 16, v11
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v2
; GFX8-NEXT:    v_cndmask_b32_e32 v2, v12, v2, vcc
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v10
; GFX8-NEXT:    v_cndmask_b32_e32 v2, v2, v10, vcc
; GFX8-NEXT:    v_and_b32_e32 v10, 0xffff0000, v11
; GFX8-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v10
; GFX8-NEXT:    v_lshlrev_b32_e32 v10, 16, v1
; GFX8-NEXT:    v_cndmask_b32_e32 v2, v12, v2, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v10, v10
; GFX8-NEXT:    v_lshlrev_b32_e32 v10, 16, v9
; GFX8-NEXT:    v_cndmask_b32_e32 v1, v1, v9, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v10, v10
; GFX8-NEXT:    v_cndmask_b32_e32 v9, v9, v1, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v10, 16, v9
; GFX8-NEXT:    v_lshlrev_b32_e32 v11, 16, v1
; GFX8-NEXT:    v_cmp_gt_f32_e32 vcc, v11, v10
; GFX8-NEXT:    v_cndmask_b32_e32 v10, v9, v1, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
; GFX8-NEXT:    v_mul_f32_e32 v10, 1.0, v10
; GFX8-NEXT:    v_bfe_u32 v11, v10, 16, 1
; GFX8-NEXT:    v_add_u32_e32 v11, vcc, v11, v10
; GFX8-NEXT:    v_add_u32_e32 v11, vcc, s4, v11
; GFX8-NEXT:    v_or_b32_e32 v12, 0x400000, v10
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v10, v10
; GFX8-NEXT:    v_cndmask_b32_e32 v10, v11, v12, vcc
; GFX8-NEXT:    v_lshrrev_b32_e32 v11, 16, v10
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v1
; GFX8-NEXT:    v_cndmask_b32_e32 v1, v11, v1, vcc
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v9
; GFX8-NEXT:    v_cndmask_b32_e32 v1, v1, v9, vcc
; GFX8-NEXT:    v_and_b32_e32 v9, 0xffff0000, v10
; GFX8-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v9
; GFX8-NEXT:    v_lshlrev_b32_e32 v9, 16, v0
; GFX8-NEXT:    v_cndmask_b32_e32 v1, v11, v1, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v9, v9
; GFX8-NEXT:    v_lshlrev_b32_e32 v9, 16, v8
; GFX8-NEXT:    v_cndmask_b32_e32 v0, v0, v8, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v9, v9
; GFX8-NEXT:    v_cndmask_b32_e32 v8, v8, v0, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v9, 16, v8
; GFX8-NEXT:    v_lshlrev_b32_e32 v10, 16, v0
; GFX8-NEXT:    v_cmp_gt_f32_e32 vcc, v10, v9
; GFX8-NEXT:    v_cndmask_b32_e32 v9, v8, v0, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; GFX8-NEXT:    v_mul_f32_e32 v9, 1.0, v9
; GFX8-NEXT:    v_bfe_u32 v10, v9, 16, 1
; GFX8-NEXT:    v_add_u32_e32 v10, vcc, v10, v9
; GFX8-NEXT:    v_add_u32_e32 v10, vcc, s4, v10
; GFX8-NEXT:    v_or_b32_e32 v11, 0x400000, v9
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v9, v9
; GFX8-NEXT:    v_cndmask_b32_e32 v9, v10, v11, vcc
; GFX8-NEXT:    v_lshrrev_b32_e32 v10, 16, v9
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v0
; GFX8-NEXT:    v_cndmask_b32_e32 v0, v10, v0, vcc
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v8
; GFX8-NEXT:    v_cndmask_b32_e32 v0, v0, v8, vcc
; GFX8-NEXT:    v_and_b32_e32 v8, 0xffff0000, v9
; GFX8-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v8
; GFX8-NEXT:    v_cndmask_b32_e32 v0, v10, v0, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v8, 16, v23
; GFX8-NEXT:    v_or_b32_sdwa v0, v0, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX8-NEXT:    v_lshlrev_b32_e32 v8, 16, v22
; GFX8-NEXT:    v_or_b32_sdwa v1, v1, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX8-NEXT:    v_lshlrev_b32_e32 v8, 16, v21
; GFX8-NEXT:    v_or_b32_sdwa v2, v2, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX8-NEXT:    v_lshlrev_b32_e32 v8, 16, v20
; GFX8-NEXT:    v_or_b32_sdwa v3, v3, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX8-NEXT:    v_lshlrev_b32_e32 v8, 16, v19
; GFX8-NEXT:    v_or_b32_sdwa v4, v4, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX8-NEXT:    v_lshlrev_b32_e32 v8, 16, v18
; GFX8-NEXT:    v_or_b32_sdwa v5, v5, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX8-NEXT:    v_lshlrev_b32_e32 v8, 16, v17
; GFX8-NEXT:    v_or_b32_sdwa v6, v6, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX8-NEXT:    v_lshlrev_b32_e32 v8, 16, v16
; GFX8-NEXT:    v_or_b32_sdwa v7, v7, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX8-NEXT:    s_setpc_b64 s[30:31]
;
; GFX900-LABEL: v_maximumnum_v16bf16:
; GFX900:       ; %bb.0:
; GFX900-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX900-NEXT:    v_and_b32_e32 v16, 0xffff0000, v7
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v16, v16
; GFX900-NEXT:    v_lshrrev_b32_e32 v16, 16, v15
; GFX900-NEXT:    v_lshrrev_b32_e32 v17, 16, v7
; GFX900-NEXT:    v_and_b32_e32 v19, 0xffff0000, v15
; GFX900-NEXT:    v_cndmask_b32_e32 v17, v17, v16, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v19, v19
; GFX900-NEXT:    v_cndmask_b32_e32 v16, v16, v17, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v18, 16, v17
; GFX900-NEXT:    v_lshlrev_b32_e32 v19, 16, v16
; GFX900-NEXT:    v_cmp_gt_f32_e32 vcc, v18, v19
; GFX900-NEXT:    v_cndmask_b32_e32 v18, v16, v17, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
; GFX900-NEXT:    v_max_f32_e32 v18, v18, v18
; GFX900-NEXT:    v_bfe_u32 v19, v18, 16, 1
; GFX900-NEXT:    s_movk_i32 s4, 0x7fff
; GFX900-NEXT:    v_add3_u32 v19, v19, v18, s4
; GFX900-NEXT:    v_or_b32_e32 v20, 0x400000, v18
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; GFX900-NEXT:    v_cndmask_b32_e32 v18, v19, v20, vcc
; GFX900-NEXT:    v_lshrrev_b32_e32 v19, 16, v18
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v17
; GFX900-NEXT:    v_cndmask_b32_e32 v17, v19, v17, vcc
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v16
; GFX900-NEXT:    v_cndmask_b32_e32 v16, v17, v16, vcc
; GFX900-NEXT:    v_and_b32_e32 v17, 0xffff0000, v18
; GFX900-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v17
; GFX900-NEXT:    v_and_b32_e32 v17, 0xffff0000, v6
; GFX900-NEXT:    v_cndmask_b32_e32 v16, v19, v16, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX900-NEXT:    v_lshrrev_b32_e32 v17, 16, v14
; GFX900-NEXT:    v_lshrrev_b32_e32 v18, 16, v6
; GFX900-NEXT:    v_and_b32_e32 v20, 0xffff0000, v14
; GFX900-NEXT:    v_cndmask_b32_e32 v18, v18, v17, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v20, v20
; GFX900-NEXT:    v_cndmask_b32_e32 v17, v17, v18, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v19, 16, v18
; GFX900-NEXT:    v_lshlrev_b32_e32 v20, 16, v17
; GFX900-NEXT:    v_cmp_gt_f32_e32 vcc, v19, v20
; GFX900-NEXT:    v_cndmask_b32_e32 v19, v17, v18, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
; GFX900-NEXT:    v_max_f32_e32 v19, v19, v19
; GFX900-NEXT:    v_bfe_u32 v20, v19, 16, 1
; GFX900-NEXT:    v_add3_u32 v20, v20, v19, s4
; GFX900-NEXT:    v_or_b32_e32 v21, 0x400000, v19
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v19, v19
; GFX900-NEXT:    v_cndmask_b32_e32 v19, v20, v21, vcc
; GFX900-NEXT:    v_lshrrev_b32_e32 v20, 16, v19
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v18
; GFX900-NEXT:    v_cndmask_b32_e32 v18, v20, v18, vcc
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v17
; GFX900-NEXT:    v_cndmask_b32_e32 v17, v18, v17, vcc
; GFX900-NEXT:    v_and_b32_e32 v18, 0xffff0000, v19
; GFX900-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v18
; GFX900-NEXT:    v_and_b32_e32 v18, 0xffff0000, v5
; GFX900-NEXT:    v_cndmask_b32_e32 v17, v20, v17, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; GFX900-NEXT:    v_lshrrev_b32_e32 v18, 16, v13
; GFX900-NEXT:    v_lshrrev_b32_e32 v19, 16, v5
; GFX900-NEXT:    v_and_b32_e32 v21, 0xffff0000, v13
; GFX900-NEXT:    v_cndmask_b32_e32 v19, v19, v18, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v21, v21
; GFX900-NEXT:    v_cndmask_b32_e32 v18, v18, v19, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v20, 16, v19
; GFX900-NEXT:    v_lshlrev_b32_e32 v21, 16, v18
; GFX900-NEXT:    v_cmp_gt_f32_e32 vcc, v20, v21
; GFX900-NEXT:    v_cndmask_b32_e32 v20, v18, v19, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
; GFX900-NEXT:    v_max_f32_e32 v20, v20, v20
; GFX900-NEXT:    v_bfe_u32 v21, v20, 16, 1
; GFX900-NEXT:    v_add3_u32 v21, v21, v20, s4
; GFX900-NEXT:    v_or_b32_e32 v22, 0x400000, v20
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v20, v20
; GFX900-NEXT:    v_cndmask_b32_e32 v20, v21, v22, vcc
; GFX900-NEXT:    v_lshrrev_b32_e32 v21, 16, v20
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v19
; GFX900-NEXT:    v_cndmask_b32_e32 v19, v21, v19, vcc
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v18
; GFX900-NEXT:    v_cndmask_b32_e32 v18, v19, v18, vcc
; GFX900-NEXT:    v_and_b32_e32 v19, 0xffff0000, v20
; GFX900-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v19
; GFX900-NEXT:    v_and_b32_e32 v19, 0xffff0000, v4
; GFX900-NEXT:    v_cndmask_b32_e32 v18, v21, v18, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v19, v19
; GFX900-NEXT:    v_lshrrev_b32_e32 v19, 16, v12
; GFX900-NEXT:    v_lshrrev_b32_e32 v20, 16, v4
; GFX900-NEXT:    v_and_b32_e32 v22, 0xffff0000, v12
; GFX900-NEXT:    v_cndmask_b32_e32 v20, v20, v19, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v22, v22
; GFX900-NEXT:    v_cndmask_b32_e32 v19, v19, v20, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v21, 16, v20
; GFX900-NEXT:    v_lshlrev_b32_e32 v22, 16, v19
; GFX900-NEXT:    v_cmp_gt_f32_e32 vcc, v21, v22
; GFX900-NEXT:    v_cndmask_b32_e32 v21, v19, v20, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
; GFX900-NEXT:    v_max_f32_e32 v21, v21, v21
; GFX900-NEXT:    v_bfe_u32 v22, v21, 16, 1
; GFX900-NEXT:    v_add3_u32 v22, v22, v21, s4
; GFX900-NEXT:    v_or_b32_e32 v23, 0x400000, v21
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v21, v21
; GFX900-NEXT:    v_cndmask_b32_e32 v21, v22, v23, vcc
; GFX900-NEXT:    v_lshrrev_b32_e32 v22, 16, v21
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v20
; GFX900-NEXT:    v_cndmask_b32_e32 v20, v22, v20, vcc
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v19
; GFX900-NEXT:    v_cndmask_b32_e32 v19, v20, v19, vcc
; GFX900-NEXT:    v_and_b32_e32 v20, 0xffff0000, v21
; GFX900-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v20
; GFX900-NEXT:    v_and_b32_e32 v20, 0xffff0000, v3
; GFX900-NEXT:    v_cndmask_b32_e32 v19, v22, v19, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v20, v20
; GFX900-NEXT:    v_lshrrev_b32_e32 v20, 16, v11
; GFX900-NEXT:    v_lshrrev_b32_e32 v21, 16, v3
; GFX900-NEXT:    v_and_b32_e32 v23, 0xffff0000, v11
; GFX900-NEXT:    v_cndmask_b32_e32 v21, v21, v20, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v23, v23
; GFX900-NEXT:    v_cndmask_b32_e32 v20, v20, v21, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v22, 16, v21
; GFX900-NEXT:    v_lshlrev_b32_e32 v23, 16, v20
; GFX900-NEXT:    v_cmp_gt_f32_e32 vcc, v22, v23
; GFX900-NEXT:    v_cndmask_b32_e32 v22, v20, v21, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v22, 16, v22
; GFX900-NEXT:    v_max_f32_e32 v22, v22, v22
; GFX900-NEXT:    v_bfe_u32 v23, v22, 16, 1
; GFX900-NEXT:    v_add3_u32 v23, v23, v22, s4
; GFX900-NEXT:    v_or_b32_e32 v24, 0x400000, v22
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v22, v22
; GFX900-NEXT:    v_cndmask_b32_e32 v22, v23, v24, vcc
; GFX900-NEXT:    v_lshrrev_b32_e32 v23, 16, v22
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v21
; GFX900-NEXT:    v_cndmask_b32_e32 v21, v23, v21, vcc
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v20
; GFX900-NEXT:    v_cndmask_b32_e32 v20, v21, v20, vcc
; GFX900-NEXT:    v_and_b32_e32 v21, 0xffff0000, v22
; GFX900-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v21
; GFX900-NEXT:    v_and_b32_e32 v21, 0xffff0000, v2
; GFX900-NEXT:    v_cndmask_b32_e32 v20, v23, v20, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v21, v21
; GFX900-NEXT:    v_lshrrev_b32_e32 v21, 16, v10
; GFX900-NEXT:    v_lshrrev_b32_e32 v22, 16, v2
; GFX900-NEXT:    v_and_b32_e32 v24, 0xffff0000, v10
; GFX900-NEXT:    v_cndmask_b32_e32 v22, v22, v21, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v24, v24
; GFX900-NEXT:    v_cndmask_b32_e32 v21, v21, v22, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v23, 16, v22
; GFX900-NEXT:    v_lshlrev_b32_e32 v24, 16, v21
; GFX900-NEXT:    v_cmp_gt_f32_e32 vcc, v23, v24
; GFX900-NEXT:    v_cndmask_b32_e32 v23, v21, v22, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; GFX900-NEXT:    v_max_f32_e32 v23, v23, v23
; GFX900-NEXT:    v_bfe_u32 v24, v23, 16, 1
; GFX900-NEXT:    v_add3_u32 v24, v24, v23, s4
; GFX900-NEXT:    v_or_b32_e32 v25, 0x400000, v23
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v23, v23
; GFX900-NEXT:    v_cndmask_b32_e32 v23, v24, v25, vcc
; GFX900-NEXT:    v_lshrrev_b32_e32 v24, 16, v23
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v22
; GFX900-NEXT:    v_cndmask_b32_e32 v22, v24, v22, vcc
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v21
; GFX900-NEXT:    v_cndmask_b32_e32 v21, v22, v21, vcc
; GFX900-NEXT:    v_and_b32_e32 v22, 0xffff0000, v23
; GFX900-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v22
; GFX900-NEXT:    v_and_b32_e32 v22, 0xffff0000, v1
; GFX900-NEXT:    v_cndmask_b32_e32 v21, v24, v21, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v22, v22
; GFX900-NEXT:    v_lshrrev_b32_e32 v22, 16, v9
; GFX900-NEXT:    v_lshrrev_b32_e32 v23, 16, v1
; GFX900-NEXT:    v_and_b32_e32 v25, 0xffff0000, v9
; GFX900-NEXT:    v_cndmask_b32_e32 v23, v23, v22, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v25, v25
; GFX900-NEXT:    v_cndmask_b32_e32 v22, v22, v23, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v24, 16, v23
; GFX900-NEXT:    v_lshlrev_b32_e32 v25, 16, v22
; GFX900-NEXT:    v_cmp_gt_f32_e32 vcc, v24, v25
; GFX900-NEXT:    v_cndmask_b32_e32 v24, v22, v23, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
; GFX900-NEXT:    v_max_f32_e32 v24, v24, v24
; GFX900-NEXT:    v_bfe_u32 v25, v24, 16, 1
; GFX900-NEXT:    v_add3_u32 v25, v25, v24, s4
; GFX900-NEXT:    v_or_b32_e32 v26, 0x400000, v24
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v24, v24
; GFX900-NEXT:    v_cndmask_b32_e32 v24, v25, v26, vcc
; GFX900-NEXT:    v_lshrrev_b32_e32 v25, 16, v24
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v23
; GFX900-NEXT:    v_cndmask_b32_e32 v23, v25, v23, vcc
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v22
; GFX900-NEXT:    v_cndmask_b32_e32 v22, v23, v22, vcc
; GFX900-NEXT:    v_and_b32_e32 v23, 0xffff0000, v24
; GFX900-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v23
; GFX900-NEXT:    v_and_b32_e32 v23, 0xffff0000, v0
; GFX900-NEXT:    v_cndmask_b32_e32 v22, v25, v22, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v23, v23
; GFX900-NEXT:    v_lshrrev_b32_e32 v23, 16, v8
; GFX900-NEXT:    v_lshrrev_b32_e32 v24, 16, v0
; GFX900-NEXT:    v_and_b32_e32 v26, 0xffff0000, v8
; GFX900-NEXT:    v_cndmask_b32_e32 v24, v24, v23, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v26, v26
; GFX900-NEXT:    v_cndmask_b32_e32 v23, v23, v24, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v25, 16, v24
; GFX900-NEXT:    v_lshlrev_b32_e32 v26, 16, v23
; GFX900-NEXT:    v_cmp_gt_f32_e32 vcc, v25, v26
; GFX900-NEXT:    v_cndmask_b32_e32 v25, v23, v24, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v25, 16, v25
; GFX900-NEXT:    v_max_f32_e32 v25, v25, v25
; GFX900-NEXT:    v_bfe_u32 v26, v25, 16, 1
; GFX900-NEXT:    v_add3_u32 v26, v26, v25, s4
; GFX900-NEXT:    v_or_b32_e32 v27, 0x400000, v25
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v25, v25
; GFX900-NEXT:    v_cndmask_b32_e32 v25, v26, v27, vcc
; GFX900-NEXT:    v_lshrrev_b32_e32 v26, 16, v25
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v24
; GFX900-NEXT:    v_cndmask_b32_e32 v24, v26, v24, vcc
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v23
; GFX900-NEXT:    v_cndmask_b32_e32 v23, v24, v23, vcc
; GFX900-NEXT:    v_and_b32_e32 v24, 0xffff0000, v25
; GFX900-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v24
; GFX900-NEXT:    v_lshlrev_b32_e32 v24, 16, v7
; GFX900-NEXT:    v_cndmask_b32_e32 v23, v26, v23, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v24, v24
; GFX900-NEXT:    v_lshlrev_b32_e32 v24, 16, v15
; GFX900-NEXT:    v_cndmask_b32_e32 v7, v7, v15, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v24, v24
; GFX900-NEXT:    v_cndmask_b32_e32 v15, v15, v7, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v24, 16, v15
; GFX900-NEXT:    v_lshlrev_b32_e32 v25, 16, v7
; GFX900-NEXT:    v_cmp_gt_f32_e32 vcc, v25, v24
; GFX900-NEXT:    v_cndmask_b32_e32 v24, v15, v7, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
; GFX900-NEXT:    v_max_f32_e32 v24, v24, v24
; GFX900-NEXT:    v_bfe_u32 v25, v24, 16, 1
; GFX900-NEXT:    v_add3_u32 v25, v25, v24, s4
; GFX900-NEXT:    v_or_b32_e32 v26, 0x400000, v24
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v24, v24
; GFX900-NEXT:    v_cndmask_b32_e32 v24, v25, v26, vcc
; GFX900-NEXT:    v_lshrrev_b32_e32 v25, 16, v24
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v7
; GFX900-NEXT:    v_cndmask_b32_e32 v7, v25, v7, vcc
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v15
; GFX900-NEXT:    v_cndmask_b32_e32 v7, v7, v15, vcc
; GFX900-NEXT:    v_and_b32_e32 v15, 0xffff0000, v24
; GFX900-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v15
; GFX900-NEXT:    v_lshlrev_b32_e32 v15, 16, v6
; GFX900-NEXT:    v_cndmask_b32_e32 v7, v25, v7, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v15, v15
; GFX900-NEXT:    v_lshlrev_b32_e32 v15, 16, v14
; GFX900-NEXT:    v_cndmask_b32_e32 v6, v6, v14, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v15, v15
; GFX900-NEXT:    v_cndmask_b32_e32 v14, v14, v6, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v15, 16, v14
; GFX900-NEXT:    v_lshlrev_b32_e32 v24, 16, v6
; GFX900-NEXT:    v_cmp_gt_f32_e32 vcc, v24, v15
; GFX900-NEXT:    v_cndmask_b32_e32 v15, v14, v6, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; GFX900-NEXT:    v_max_f32_e32 v15, v15, v15
; GFX900-NEXT:    v_bfe_u32 v24, v15, 16, 1
; GFX900-NEXT:    v_add3_u32 v24, v24, v15, s4
; GFX900-NEXT:    v_or_b32_e32 v25, 0x400000, v15
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v15, v15
; GFX900-NEXT:    v_cndmask_b32_e32 v15, v24, v25, vcc
; GFX900-NEXT:    v_lshrrev_b32_e32 v24, 16, v15
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v6
; GFX900-NEXT:    v_cndmask_b32_e32 v6, v24, v6, vcc
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v14
; GFX900-NEXT:    v_cndmask_b32_e32 v6, v6, v14, vcc
; GFX900-NEXT:    v_and_b32_e32 v14, 0xffff0000, v15
; GFX900-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v14
; GFX900-NEXT:    v_lshlrev_b32_e32 v14, 16, v5
; GFX900-NEXT:    v_cndmask_b32_e32 v6, v24, v6, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v14, v14
; GFX900-NEXT:    v_lshlrev_b32_e32 v14, 16, v13
; GFX900-NEXT:    v_cndmask_b32_e32 v5, v5, v13, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v14, v14
; GFX900-NEXT:    v_cndmask_b32_e32 v13, v13, v5, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v14, 16, v13
; GFX900-NEXT:    v_lshlrev_b32_e32 v15, 16, v5
; GFX900-NEXT:    v_cmp_gt_f32_e32 vcc, v15, v14
; GFX900-NEXT:    v_cndmask_b32_e32 v14, v13, v5, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
; GFX900-NEXT:    v_max_f32_e32 v14, v14, v14
; GFX900-NEXT:    v_bfe_u32 v15, v14, 16, 1
; GFX900-NEXT:    v_add3_u32 v15, v15, v14, s4
; GFX900-NEXT:    v_or_b32_e32 v24, 0x400000, v14
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v14, v14
; GFX900-NEXT:    v_cndmask_b32_e32 v14, v15, v24, vcc
; GFX900-NEXT:    v_lshrrev_b32_e32 v15, 16, v14
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v5
; GFX900-NEXT:    v_cndmask_b32_e32 v5, v15, v5, vcc
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v13
; GFX900-NEXT:    v_cndmask_b32_e32 v5, v5, v13, vcc
; GFX900-NEXT:    v_and_b32_e32 v13, 0xffff0000, v14
; GFX900-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v13
; GFX900-NEXT:    v_lshlrev_b32_e32 v13, 16, v4
; GFX900-NEXT:    v_cndmask_b32_e32 v5, v15, v5, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v13, v13
; GFX900-NEXT:    v_lshlrev_b32_e32 v13, 16, v12
; GFX900-NEXT:    v_cndmask_b32_e32 v4, v4, v12, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v13, v13
; GFX900-NEXT:    v_cndmask_b32_e32 v12, v12, v4, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v13, 16, v12
; GFX900-NEXT:    v_lshlrev_b32_e32 v14, 16, v4
; GFX900-NEXT:    v_cmp_gt_f32_e32 vcc, v14, v13
; GFX900-NEXT:    v_cndmask_b32_e32 v13, v12, v4, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
; GFX900-NEXT:    v_max_f32_e32 v13, v13, v13
; GFX900-NEXT:    v_bfe_u32 v14, v13, 16, 1
; GFX900-NEXT:    v_add3_u32 v14, v14, v13, s4
; GFX900-NEXT:    v_or_b32_e32 v15, 0x400000, v13
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v13, v13
; GFX900-NEXT:    v_cndmask_b32_e32 v13, v14, v15, vcc
; GFX900-NEXT:    v_lshrrev_b32_e32 v14, 16, v13
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v4
; GFX900-NEXT:    v_cndmask_b32_e32 v4, v14, v4, vcc
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v12
; GFX900-NEXT:    v_cndmask_b32_e32 v4, v4, v12, vcc
; GFX900-NEXT:    v_and_b32_e32 v12, 0xffff0000, v13
; GFX900-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v12
; GFX900-NEXT:    v_lshlrev_b32_e32 v12, 16, v3
; GFX900-NEXT:    v_cndmask_b32_e32 v4, v14, v4, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v12, v12
; GFX900-NEXT:    v_lshlrev_b32_e32 v12, 16, v11
; GFX900-NEXT:    v_cndmask_b32_e32 v3, v3, v11, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v12, v12
; GFX900-NEXT:    v_cndmask_b32_e32 v11, v11, v3, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v12, 16, v11
; GFX900-NEXT:    v_lshlrev_b32_e32 v13, 16, v3
; GFX900-NEXT:    v_cmp_gt_f32_e32 vcc, v13, v12
; GFX900-NEXT:    v_cndmask_b32_e32 v12, v11, v3, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; GFX900-NEXT:    v_max_f32_e32 v12, v12, v12
; GFX900-NEXT:    v_bfe_u32 v13, v12, 16, 1
; GFX900-NEXT:    v_add3_u32 v13, v13, v12, s4
; GFX900-NEXT:    v_or_b32_e32 v14, 0x400000, v12
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v12, v12
; GFX900-NEXT:    v_cndmask_b32_e32 v12, v13, v14, vcc
; GFX900-NEXT:    v_lshrrev_b32_e32 v13, 16, v12
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v3
; GFX900-NEXT:    v_cndmask_b32_e32 v3, v13, v3, vcc
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v11
; GFX900-NEXT:    v_cndmask_b32_e32 v3, v3, v11, vcc
; GFX900-NEXT:    v_and_b32_e32 v11, 0xffff0000, v12
; GFX900-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v11
; GFX900-NEXT:    v_lshlrev_b32_e32 v11, 16, v2
; GFX900-NEXT:    v_cndmask_b32_e32 v3, v13, v3, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v11, v11
; GFX900-NEXT:    v_lshlrev_b32_e32 v11, 16, v10
; GFX900-NEXT:    v_cndmask_b32_e32 v2, v2, v10, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v11, v11
; GFX900-NEXT:    v_cndmask_b32_e32 v10, v10, v2, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v11, 16, v10
; GFX900-NEXT:    v_lshlrev_b32_e32 v12, 16, v2
; GFX900-NEXT:    v_cmp_gt_f32_e32 vcc, v12, v11
; GFX900-NEXT:    v_cndmask_b32_e32 v11, v10, v2, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
; GFX900-NEXT:    v_max_f32_e32 v11, v11, v11
; GFX900-NEXT:    v_bfe_u32 v12, v11, 16, 1
; GFX900-NEXT:    v_add3_u32 v12, v12, v11, s4
; GFX900-NEXT:    v_or_b32_e32 v13, 0x400000, v11
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v11, v11
; GFX900-NEXT:    v_cndmask_b32_e32 v11, v12, v13, vcc
; GFX900-NEXT:    v_lshrrev_b32_e32 v12, 16, v11
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v2
; GFX900-NEXT:    v_cndmask_b32_e32 v2, v12, v2, vcc
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v10
; GFX900-NEXT:    v_cndmask_b32_e32 v2, v2, v10, vcc
; GFX900-NEXT:    v_and_b32_e32 v10, 0xffff0000, v11
; GFX900-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v10
; GFX900-NEXT:    v_lshlrev_b32_e32 v10, 16, v1
; GFX900-NEXT:    v_cndmask_b32_e32 v2, v12, v2, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v10, v10
; GFX900-NEXT:    v_lshlrev_b32_e32 v10, 16, v9
; GFX900-NEXT:    v_cndmask_b32_e32 v1, v1, v9, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v10, v10
; GFX900-NEXT:    v_cndmask_b32_e32 v9, v9, v1, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v10, 16, v9
; GFX900-NEXT:    v_lshlrev_b32_e32 v11, 16, v1
; GFX900-NEXT:    v_cmp_gt_f32_e32 vcc, v11, v10
; GFX900-NEXT:    v_cndmask_b32_e32 v10, v9, v1, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
; GFX900-NEXT:    v_max_f32_e32 v10, v10, v10
; GFX900-NEXT:    v_bfe_u32 v11, v10, 16, 1
; GFX900-NEXT:    v_add3_u32 v11, v11, v10, s4
; GFX900-NEXT:    v_or_b32_e32 v12, 0x400000, v10
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v10, v10
; GFX900-NEXT:    v_cndmask_b32_e32 v10, v11, v12, vcc
; GFX900-NEXT:    v_lshrrev_b32_e32 v11, 16, v10
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v1
; GFX900-NEXT:    v_cndmask_b32_e32 v1, v11, v1, vcc
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v9
; GFX900-NEXT:    v_cndmask_b32_e32 v1, v1, v9, vcc
; GFX900-NEXT:    v_and_b32_e32 v9, 0xffff0000, v10
; GFX900-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v9
; GFX900-NEXT:    v_lshlrev_b32_e32 v9, 16, v0
; GFX900-NEXT:    v_cndmask_b32_e32 v1, v11, v1, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v9, v9
; GFX900-NEXT:    v_lshlrev_b32_e32 v9, 16, v8
; GFX900-NEXT:    v_cndmask_b32_e32 v0, v0, v8, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v9, v9
; GFX900-NEXT:    v_cndmask_b32_e32 v8, v8, v0, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v9, 16, v8
; GFX900-NEXT:    v_lshlrev_b32_e32 v10, 16, v0
; GFX900-NEXT:    v_cmp_gt_f32_e32 vcc, v10, v9
; GFX900-NEXT:    v_cndmask_b32_e32 v9, v8, v0, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; GFX900-NEXT:    v_max_f32_e32 v9, v9, v9
; GFX900-NEXT:    v_bfe_u32 v10, v9, 16, 1
; GFX900-NEXT:    v_add3_u32 v10, v10, v9, s4
; GFX900-NEXT:    v_or_b32_e32 v11, 0x400000, v9
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v9, v9
; GFX900-NEXT:    v_cndmask_b32_e32 v9, v10, v11, vcc
; GFX900-NEXT:    v_lshrrev_b32_e32 v10, 16, v9
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v0
; GFX900-NEXT:    v_cndmask_b32_e32 v0, v10, v0, vcc
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v8
; GFX900-NEXT:    v_cndmask_b32_e32 v0, v0, v8, vcc
; GFX900-NEXT:    v_and_b32_e32 v8, 0xffff0000, v9
; GFX900-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v8
; GFX900-NEXT:    v_cndmask_b32_e32 v0, v10, v0, vcc
; GFX900-NEXT:    s_mov_b32 s4, 0x5040100
; GFX900-NEXT:    v_perm_b32 v0, v23, v0, s4
; GFX900-NEXT:    v_perm_b32 v1, v22, v1, s4
; GFX900-NEXT:    v_perm_b32 v2, v21, v2, s4
; GFX900-NEXT:    v_perm_b32 v3, v20, v3, s4
; GFX900-NEXT:    v_perm_b32 v4, v19, v4, s4
; GFX900-NEXT:    v_perm_b32 v5, v18, v5, s4
; GFX900-NEXT:    v_perm_b32 v6, v17, v6, s4
; GFX900-NEXT:    v_perm_b32 v7, v16, v7, s4
; GFX900-NEXT:    s_setpc_b64 s[30:31]
;
; GFX950-LABEL: v_maximumnum_v16bf16:
; GFX950:       ; %bb.0:
; GFX950-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX950-NEXT:    v_and_b32_e32 v16, 0xffff0000, v7
; GFX950-NEXT:    v_lshrrev_b32_e32 v17, 16, v15
; GFX950-NEXT:    v_lshrrev_b32_e32 v18, 16, v7
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v16, v16
; GFX950-NEXT:    v_and_b32_e32 v19, 0xffff0000, v15
; GFX950-NEXT:    v_and_b32_e32 v20, 0xffff0000, v14
; GFX950-NEXT:    v_cndmask_b32_e32 v16, v18, v17, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v19, v19
; GFX950-NEXT:    v_lshlrev_b32_e32 v18, 16, v16
; GFX950-NEXT:    v_and_b32_e32 v21, 0xffff0000, v13
; GFX950-NEXT:    v_cndmask_b32_e32 v17, v17, v16, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v19, 16, v17
; GFX950-NEXT:    v_cmp_gt_f32_e32 vcc, v18, v19
; GFX950-NEXT:    v_lshrrev_b32_e32 v19, 16, v6
; GFX950-NEXT:    v_and_b32_e32 v22, 0xffff0000, v12
; GFX950-NEXT:    v_cndmask_b32_e32 v18, v17, v16, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
; GFX950-NEXT:    v_max_f32_e32 v18, v18, v18
; GFX950-NEXT:    v_cvt_pk_bf16_f32 v18, v18, s0
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v16
; GFX950-NEXT:    v_and_b32_e32 v23, 0xffff0000, v11
; GFX950-NEXT:    v_and_b32_e32 v24, 0xffff0000, v10
; GFX950-NEXT:    v_cndmask_b32_e32 v16, v18, v16, vcc
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v17
; GFX950-NEXT:    v_and_b32_e32 v25, 0xffff0000, v9
; GFX950-NEXT:    v_and_b32_e32 v26, 0xffff0000, v8
; GFX950-NEXT:    v_cndmask_b32_e32 v16, v16, v17, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v17, 16, v18
; GFX950-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v17
; GFX950-NEXT:    v_and_b32_e32 v17, 0xffff0000, v6
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v16, v18, v16, vcc
; GFX950-NEXT:    v_lshrrev_b32_e32 v18, 16, v14
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v17, v19, v18, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v20, v20
; GFX950-NEXT:    v_lshlrev_b32_e32 v19, 16, v17
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v18, v18, v17, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v20, 16, v18
; GFX950-NEXT:    v_cmp_gt_f32_e32 vcc, v19, v20
; GFX950-NEXT:    v_lshrrev_b32_e32 v20, 16, v5
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v19, v18, v17, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
; GFX950-NEXT:    v_max_f32_e32 v19, v19, v19
; GFX950-NEXT:    v_cvt_pk_bf16_f32 v19, v19, s0
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v17
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v17, v19, v17, vcc
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v18
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v17, v17, v18, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v18, 16, v19
; GFX950-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v18
; GFX950-NEXT:    v_and_b32_e32 v18, 0xffff0000, v5
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v17, v19, v17, vcc
; GFX950-NEXT:    v_lshrrev_b32_e32 v19, 16, v13
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v18, v20, v19, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v21, v21
; GFX950-NEXT:    v_lshlrev_b32_e32 v20, 16, v18
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v19, v19, v18, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v21, 16, v19
; GFX950-NEXT:    v_cmp_gt_f32_e32 vcc, v20, v21
; GFX950-NEXT:    v_lshrrev_b32_e32 v21, 16, v4
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v20, v19, v18, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
; GFX950-NEXT:    v_max_f32_e32 v20, v20, v20
; GFX950-NEXT:    v_cvt_pk_bf16_f32 v20, v20, s0
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v18
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v18, v20, v18, vcc
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v19
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v18, v18, v19, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v19, 16, v20
; GFX950-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v19
; GFX950-NEXT:    v_and_b32_e32 v19, 0xffff0000, v4
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v18, v20, v18, vcc
; GFX950-NEXT:    v_lshrrev_b32_e32 v20, 16, v12
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v19, v19
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v19, v21, v20, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v22, v22
; GFX950-NEXT:    v_lshlrev_b32_e32 v21, 16, v19
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v20, v20, v19, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v22, 16, v20
; GFX950-NEXT:    v_cmp_gt_f32_e32 vcc, v21, v22
; GFX950-NEXT:    v_lshrrev_b32_e32 v22, 16, v3
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v21, v20, v19, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
; GFX950-NEXT:    v_max_f32_e32 v21, v21, v21
; GFX950-NEXT:    v_cvt_pk_bf16_f32 v21, v21, s0
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v19
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v19, v21, v19, vcc
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v20
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v19, v19, v20, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v20, 16, v21
; GFX950-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v20
; GFX950-NEXT:    v_and_b32_e32 v20, 0xffff0000, v3
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v19, v21, v19, vcc
; GFX950-NEXT:    v_lshrrev_b32_e32 v21, 16, v11
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v20, v20
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v20, v22, v21, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v23, v23
; GFX950-NEXT:    v_lshlrev_b32_e32 v22, 16, v20
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v21, v21, v20, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v23, 16, v21
; GFX950-NEXT:    v_cmp_gt_f32_e32 vcc, v22, v23
; GFX950-NEXT:    v_lshrrev_b32_e32 v23, 16, v2
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v22, v21, v20, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v22, 16, v22
; GFX950-NEXT:    v_max_f32_e32 v22, v22, v22
; GFX950-NEXT:    v_cvt_pk_bf16_f32 v22, v22, s0
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v20
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v20, v22, v20, vcc
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v21
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v20, v20, v21, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v21, 16, v22
; GFX950-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v21
; GFX950-NEXT:    v_and_b32_e32 v21, 0xffff0000, v2
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v20, v22, v20, vcc
; GFX950-NEXT:    v_lshrrev_b32_e32 v22, 16, v10
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v21, v21
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v21, v23, v22, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v24, v24
; GFX950-NEXT:    v_lshlrev_b32_e32 v23, 16, v21
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v22, v22, v21, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v24, 16, v22
; GFX950-NEXT:    v_cmp_gt_f32_e32 vcc, v23, v24
; GFX950-NEXT:    v_lshrrev_b32_e32 v24, 16, v1
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v23, v22, v21, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; GFX950-NEXT:    v_max_f32_e32 v23, v23, v23
; GFX950-NEXT:    v_cvt_pk_bf16_f32 v23, v23, s0
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v21
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v21, v23, v21, vcc
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v22
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v21, v21, v22, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v22, 16, v23
; GFX950-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v22
; GFX950-NEXT:    v_and_b32_e32 v22, 0xffff0000, v1
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v21, v23, v21, vcc
; GFX950-NEXT:    v_lshrrev_b32_e32 v23, 16, v9
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v22, v22
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v22, v24, v23, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v25, v25
; GFX950-NEXT:    v_lshlrev_b32_e32 v24, 16, v22
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v23, v23, v22, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v25, 16, v23
; GFX950-NEXT:    v_cmp_gt_f32_e32 vcc, v24, v25
; GFX950-NEXT:    v_lshrrev_b32_e32 v25, 16, v0
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v24, v23, v22, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
; GFX950-NEXT:    v_max_f32_e32 v24, v24, v24
; GFX950-NEXT:    v_cvt_pk_bf16_f32 v24, v24, s0
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v22
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v22, v24, v22, vcc
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v23
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v22, v22, v23, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v23, 16, v24
; GFX950-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v23
; GFX950-NEXT:    v_and_b32_e32 v23, 0xffff0000, v0
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v22, v24, v22, vcc
; GFX950-NEXT:    v_lshrrev_b32_e32 v24, 16, v8
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v23, v23
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v23, v25, v24, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v26, v26
; GFX950-NEXT:    v_lshlrev_b32_e32 v25, 16, v23
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v24, v24, v23, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v26, 16, v24
; GFX950-NEXT:    v_cmp_gt_f32_e32 vcc, v25, v26
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v25, v24, v23, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v25, 16, v25
; GFX950-NEXT:    v_max_f32_e32 v25, v25, v25
; GFX950-NEXT:    v_cvt_pk_bf16_f32 v25, v25, s0
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v23
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v23, v25, v23, vcc
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v24
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v23, v23, v24, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v24, 16, v25
; GFX950-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v24
; GFX950-NEXT:    v_lshlrev_b32_e32 v24, 16, v7
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v23, v25, v23, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v24, v24
; GFX950-NEXT:    v_lshlrev_b32_e32 v24, 16, v15
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v7, v7, v15, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v24, v24
; GFX950-NEXT:    v_lshlrev_b32_e32 v25, 16, v7
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v15, v15, v7, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v24, 16, v15
; GFX950-NEXT:    v_cmp_gt_f32_e32 vcc, v25, v24
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v24, v15, v7, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
; GFX950-NEXT:    v_max_f32_e32 v24, v24, v24
; GFX950-NEXT:    v_cvt_pk_bf16_f32 v24, v24, s0
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v7
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v7, v24, v7, vcc
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v15
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v7, v7, v15, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v15, 16, v24
; GFX950-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v15
; GFX950-NEXT:    v_lshlrev_b32_e32 v15, 16, v6
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v7, v24, v7, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v15, v15
; GFX950-NEXT:    v_lshlrev_b32_e32 v15, 16, v14
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v6, v6, v14, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v15, v15
; GFX950-NEXT:    v_lshlrev_b32_e32 v24, 16, v6
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v14, v14, v6, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v15, 16, v14
; GFX950-NEXT:    v_cmp_gt_f32_e32 vcc, v24, v15
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v15, v14, v6, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; GFX950-NEXT:    v_max_f32_e32 v15, v15, v15
; GFX950-NEXT:    v_cvt_pk_bf16_f32 v15, v15, s0
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v6
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v6, v15, v6, vcc
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v14
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v6, v6, v14, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v14, 16, v15
; GFX950-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v14
; GFX950-NEXT:    v_lshlrev_b32_e32 v14, 16, v5
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v6, v15, v6, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v14, v14
; GFX950-NEXT:    v_lshlrev_b32_e32 v14, 16, v13
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v5, v5, v13, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v14, v14
; GFX950-NEXT:    v_lshlrev_b32_e32 v15, 16, v5
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v13, v13, v5, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v14, 16, v13
; GFX950-NEXT:    v_cmp_gt_f32_e32 vcc, v15, v14
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v14, v13, v5, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
; GFX950-NEXT:    v_max_f32_e32 v14, v14, v14
; GFX950-NEXT:    v_cvt_pk_bf16_f32 v14, v14, s0
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v5
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v5, v14, v5, vcc
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v13
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v5, v5, v13, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v13, 16, v14
; GFX950-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v13
; GFX950-NEXT:    v_lshlrev_b32_e32 v13, 16, v4
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v5, v14, v5, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v13, v13
; GFX950-NEXT:    v_lshlrev_b32_e32 v13, 16, v12
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v4, v4, v12, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v13, v13
; GFX950-NEXT:    v_lshlrev_b32_e32 v14, 16, v4
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v12, v12, v4, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v13, 16, v12
; GFX950-NEXT:    v_cmp_gt_f32_e32 vcc, v14, v13
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v13, v12, v4, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
; GFX950-NEXT:    v_max_f32_e32 v13, v13, v13
; GFX950-NEXT:    v_cvt_pk_bf16_f32 v13, v13, s0
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v4
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v4, v13, v4, vcc
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v12
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v4, v4, v12, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v12, 16, v13
; GFX950-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v12
; GFX950-NEXT:    v_lshlrev_b32_e32 v12, 16, v3
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v4, v13, v4, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v12, v12
; GFX950-NEXT:    v_lshlrev_b32_e32 v12, 16, v11
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v3, v3, v11, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v12, v12
; GFX950-NEXT:    v_lshlrev_b32_e32 v13, 16, v3
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v11, v11, v3, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v12, 16, v11
; GFX950-NEXT:    v_cmp_gt_f32_e32 vcc, v13, v12
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v12, v11, v3, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; GFX950-NEXT:    v_max_f32_e32 v12, v12, v12
; GFX950-NEXT:    v_cvt_pk_bf16_f32 v12, v12, s0
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v3
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v3, v12, v3, vcc
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v11
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v3, v3, v11, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v11, 16, v12
; GFX950-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v11
; GFX950-NEXT:    v_lshlrev_b32_e32 v11, 16, v2
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v3, v12, v3, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v11, v11
; GFX950-NEXT:    v_lshlrev_b32_e32 v11, 16, v10
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v2, v2, v10, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v11, v11
; GFX950-NEXT:    v_lshlrev_b32_e32 v12, 16, v2
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v10, v10, v2, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v11, 16, v10
; GFX950-NEXT:    v_cmp_gt_f32_e32 vcc, v12, v11
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v11, v10, v2, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
; GFX950-NEXT:    v_max_f32_e32 v11, v11, v11
; GFX950-NEXT:    v_cvt_pk_bf16_f32 v11, v11, s0
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v2
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v2, v11, v2, vcc
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v10
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v2, v2, v10, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v10, 16, v11
; GFX950-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v10
; GFX950-NEXT:    v_lshlrev_b32_e32 v10, 16, v1
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v2, v11, v2, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v10, v10
; GFX950-NEXT:    v_lshlrev_b32_e32 v10, 16, v9
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v1, v1, v9, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v10, v10
; GFX950-NEXT:    v_lshlrev_b32_e32 v11, 16, v1
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v9, v9, v1, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v10, 16, v9
; GFX950-NEXT:    v_cmp_gt_f32_e32 vcc, v11, v10
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v10, v9, v1, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
; GFX950-NEXT:    v_max_f32_e32 v10, v10, v10
; GFX950-NEXT:    v_cvt_pk_bf16_f32 v10, v10, s0
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v1
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v1, v10, v1, vcc
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v9
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v1, v1, v9, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v9, 16, v10
; GFX950-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v9
; GFX950-NEXT:    v_lshlrev_b32_e32 v9, 16, v0
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v1, v10, v1, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v9, v9
; GFX950-NEXT:    v_lshlrev_b32_e32 v9, 16, v8
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v0, v0, v8, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v9, v9
; GFX950-NEXT:    v_lshlrev_b32_e32 v10, 16, v0
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v8, v8, v0, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v9, 16, v8
; GFX950-NEXT:    v_cmp_gt_f32_e32 vcc, v10, v9
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v9, v8, v0, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; GFX950-NEXT:    v_max_f32_e32 v9, v9, v9
; GFX950-NEXT:    v_cvt_pk_bf16_f32 v9, v9, s0
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v0
; GFX950-NEXT:    s_mov_b32 s0, 0x5040100
; GFX950-NEXT:    v_perm_b32 v1, v22, v1, s0
; GFX950-NEXT:    v_cndmask_b32_e32 v0, v9, v0, vcc
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v8
; GFX950-NEXT:    v_perm_b32 v2, v21, v2, s0
; GFX950-NEXT:    v_perm_b32 v3, v20, v3, s0
; GFX950-NEXT:    v_cndmask_b32_e32 v0, v0, v8, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v8, 16, v9
; GFX950-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v8
; GFX950-NEXT:    v_perm_b32 v4, v19, v4, s0
; GFX950-NEXT:    v_perm_b32 v5, v18, v5, s0
; GFX950-NEXT:    v_cndmask_b32_e32 v0, v9, v0, vcc
; GFX950-NEXT:    v_perm_b32 v0, v23, v0, s0
; GFX950-NEXT:    v_perm_b32 v6, v17, v6, s0
; GFX950-NEXT:    v_perm_b32 v7, v16, v7, s0
; GFX950-NEXT:    s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_maximumnum_v16bf16:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT:    v_and_b32_e32 v16, 0xffff0000, v7
; GFX10-NEXT:    v_lshrrev_b32_e32 v17, 16, v15
; GFX10-NEXT:    v_lshrrev_b32_e32 v18, 16, v7
; GFX10-NEXT:    v_and_b32_e32 v19, 0xffff0000, v15
; GFX10-NEXT:    v_and_b32_e32 v20, 0xffff0000, v6
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v16, v16
; GFX10-NEXT:    v_lshrrev_b32_e32 v21, 16, v14
; GFX10-NEXT:    v_lshrrev_b32_e32 v22, 16, v6
; GFX10-NEXT:    v_and_b32_e32 v24, 0xffff0000, v5
; GFX10-NEXT:    v_lshrrev_b32_e32 v25, 16, v13
; GFX10-NEXT:    v_cndmask_b32_e32 v16, v18, v17, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v19, v19
; GFX10-NEXT:    v_lshrrev_b32_e32 v26, 16, v5
; GFX10-NEXT:    v_lshrrev_b32_e32 v28, 16, v12
; GFX10-NEXT:    v_lshrrev_b32_e32 v29, 16, v4
; GFX10-NEXT:    v_lshlrev_b32_e32 v18, 16, v16
; GFX10-NEXT:    v_cndmask_b32_e32 v17, v17, v16, vcc_lo
; GFX10-NEXT:    v_lshrrev_b32_e32 v30, 16, v3
; GFX10-NEXT:    v_lshrrev_b32_e32 v31, 16, v2
; GFX10-NEXT:    v_lshrrev_b32_e32 v32, 16, v1
; GFX10-NEXT:    v_and_b32_e32 v34, 0xffff0000, v8
; GFX10-NEXT:    v_lshlrev_b32_e32 v19, 16, v17
; GFX10-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v18, v19
; GFX10-NEXT:    v_and_b32_e32 v19, 0xffff0000, v14
; GFX10-NEXT:    v_cndmask_b32_e32 v18, v17, v16, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v20, v20
; GFX10-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
; GFX10-NEXT:    v_cndmask_b32_e32 v20, v22, v21, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v19, v19
; GFX10-NEXT:    v_max_f32_e32 v18, v18, v18
; GFX10-NEXT:    v_cndmask_b32_e32 v19, v21, v20, vcc_lo
; GFX10-NEXT:    v_lshlrev_b32_e32 v21, 16, v20
; GFX10-NEXT:    v_bfe_u32 v23, v18, 16, 1
; GFX10-NEXT:    v_lshlrev_b32_e32 v22, 16, v19
; GFX10-NEXT:    v_add3_u32 v23, v23, v18, 0x7fff
; GFX10-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v21, v22
; GFX10-NEXT:    v_or_b32_e32 v22, 0x400000, v18
; GFX10-NEXT:    v_cndmask_b32_e32 v21, v19, v20, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v18, v18
; GFX10-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
; GFX10-NEXT:    v_cndmask_b32_e32 v18, v23, v22, vcc_lo
; GFX10-NEXT:    v_and_b32_e32 v22, 0xffff0000, v13
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v24, v24
; GFX10-NEXT:    v_max_f32_e32 v21, v21, v21
; GFX10-NEXT:    v_lshrrev_b32_e32 v24, 16, v18
; GFX10-NEXT:    v_and_b32_e32 v18, 0xffff0000, v18
; GFX10-NEXT:    v_cndmask_b32_e32 v23, v26, v25, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v22, v22
; GFX10-NEXT:    v_bfe_u32 v26, v21, 16, 1
; GFX10-NEXT:    v_cndmask_b32_e32 v22, v25, v23, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v16
; GFX10-NEXT:    v_lshlrev_b32_e32 v25, 16, v23
; GFX10-NEXT:    v_lshlrev_b32_e32 v27, 16, v22
; GFX10-NEXT:    v_cndmask_b32_e32 v16, v24, v16, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v17
; GFX10-NEXT:    v_cndmask_b32_e32 v16, v16, v17, vcc_lo
; GFX10-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v25, v27
; GFX10-NEXT:    v_add3_u32 v17, v26, v21, 0x7fff
; GFX10-NEXT:    v_or_b32_e32 v26, 0x400000, v21
; GFX10-NEXT:    v_and_b32_e32 v27, 0xffff0000, v4
; GFX10-NEXT:    v_cndmask_b32_e32 v25, v22, v23, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v21, v21
; GFX10-NEXT:    v_lshlrev_b32_e32 v21, 16, v25
; GFX10-NEXT:    v_cndmask_b32_e32 v17, v17, v26, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v27, v27
; GFX10-NEXT:    v_and_b32_e32 v26, 0xffff0000, v12
; GFX10-NEXT:    v_max_f32_e32 v21, v21, v21
; GFX10-NEXT:    v_cndmask_b32_e32 v25, v29, v28, vcc_lo
; GFX10-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v18
; GFX10-NEXT:    v_lshrrev_b32_e32 v18, 16, v17
; GFX10-NEXT:    v_and_b32_e32 v17, 0xffff0000, v17
; GFX10-NEXT:    v_lshrrev_b32_e32 v29, 16, v11
; GFX10-NEXT:    v_lshlrev_b32_e32 v27, 16, v25
; GFX10-NEXT:    v_cndmask_b32_e32 v16, v24, v16, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v26, v26
; GFX10-NEXT:    v_bfe_u32 v26, v21, 16, 1
; GFX10-NEXT:    v_cndmask_b32_e32 v24, v28, v25, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v20
; GFX10-NEXT:    v_lshlrev_b32_e32 v28, 16, v24
; GFX10-NEXT:    v_cndmask_b32_e32 v20, v18, v20, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v19
; GFX10-NEXT:    v_cndmask_b32_e32 v19, v20, v19, vcc_lo
; GFX10-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v27, v28
; GFX10-NEXT:    v_add3_u32 v20, v26, v21, 0x7fff
; GFX10-NEXT:    v_or_b32_e32 v26, 0x400000, v21
; GFX10-NEXT:    v_and_b32_e32 v28, 0xffff0000, v3
; GFX10-NEXT:    v_cndmask_b32_e32 v27, v24, v25, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v21, v21
; GFX10-NEXT:    v_lshlrev_b32_e32 v21, 16, v27
; GFX10-NEXT:    v_cndmask_b32_e32 v20, v20, v26, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v28, v28
; GFX10-NEXT:    v_and_b32_e32 v27, 0xffff0000, v11
; GFX10-NEXT:    v_cndmask_b32_e32 v26, v30, v29, vcc_lo
; GFX10-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v17
; GFX10-NEXT:    v_lshrrev_b32_e32 v30, 16, v10
; GFX10-NEXT:    v_lshlrev_b32_e32 v28, 16, v26
; GFX10-NEXT:    v_cndmask_b32_e32 v17, v18, v19, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v27, v27
; GFX10-NEXT:    v_lshrrev_b32_e32 v18, 16, v20
; GFX10-NEXT:    v_max_f32_e32 v19, v21, v21
; GFX10-NEXT:    v_and_b32_e32 v20, 0xffff0000, v20
; GFX10-NEXT:    v_cndmask_b32_e32 v21, v29, v26, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v23
; GFX10-NEXT:    v_bfe_u32 v27, v19, 16, 1
; GFX10-NEXT:    v_lshlrev_b32_e32 v29, 16, v21
; GFX10-NEXT:    v_cndmask_b32_e32 v23, v18, v23, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v22
; GFX10-NEXT:    v_cndmask_b32_e32 v22, v23, v22, vcc_lo
; GFX10-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v28, v29
; GFX10-NEXT:    v_add3_u32 v23, v27, v19, 0x7fff
; GFX10-NEXT:    v_or_b32_e32 v27, 0x400000, v19
; GFX10-NEXT:    v_and_b32_e32 v29, 0xffff0000, v2
; GFX10-NEXT:    v_cndmask_b32_e32 v28, v21, v26, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v19, v19
; GFX10-NEXT:    v_cndmask_b32_e32 v19, v23, v27, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v29, v29
; GFX10-NEXT:    v_lshlrev_b32_e32 v23, 16, v28
; GFX10-NEXT:    v_and_b32_e32 v28, 0xffff0000, v10
; GFX10-NEXT:    v_cndmask_b32_e32 v27, v31, v30, vcc_lo
; GFX10-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v20
; GFX10-NEXT:    v_lshrrev_b32_e32 v20, 16, v19
; GFX10-NEXT:    v_and_b32_e32 v19, 0xffff0000, v19
; GFX10-NEXT:    v_lshrrev_b32_e32 v31, 16, v9
; GFX10-NEXT:    v_lshlrev_b32_e32 v29, 16, v27
; GFX10-NEXT:    v_cndmask_b32_e32 v18, v18, v22, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v28, v28
; GFX10-NEXT:    v_max_f32_e32 v22, v23, v23
; GFX10-NEXT:    v_cndmask_b32_e32 v23, v30, v27, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v25
; GFX10-NEXT:    v_bfe_u32 v28, v22, 16, 1
; GFX10-NEXT:    v_lshlrev_b32_e32 v30, 16, v23
; GFX10-NEXT:    v_cndmask_b32_e32 v25, v20, v25, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v24
; GFX10-NEXT:    v_cndmask_b32_e32 v24, v25, v24, vcc_lo
; GFX10-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v29, v30
; GFX10-NEXT:    v_add3_u32 v25, v28, v22, 0x7fff
; GFX10-NEXT:    v_or_b32_e32 v28, 0x400000, v22
; GFX10-NEXT:    v_and_b32_e32 v30, 0xffff0000, v1
; GFX10-NEXT:    v_cndmask_b32_e32 v29, v23, v27, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v22, v22
; GFX10-NEXT:    v_cndmask_b32_e32 v22, v25, v28, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v30, v30
; GFX10-NEXT:    v_lshlrev_b32_e32 v25, 16, v29
; GFX10-NEXT:    v_and_b32_e32 v29, 0xffff0000, v9
; GFX10-NEXT:    v_cndmask_b32_e32 v28, v32, v31, vcc_lo
; GFX10-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v19
; GFX10-NEXT:    v_lshrrev_b32_e32 v32, 16, v0
; GFX10-NEXT:    v_lshlrev_b32_e32 v30, 16, v28
; GFX10-NEXT:    v_cndmask_b32_e32 v19, v20, v24, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v29, v29
; GFX10-NEXT:    v_lshrrev_b32_e32 v20, 16, v22
; GFX10-NEXT:    v_max_f32_e32 v24, v25, v25
; GFX10-NEXT:    v_and_b32_e32 v22, 0xffff0000, v22
; GFX10-NEXT:    v_cndmask_b32_e32 v25, v31, v28, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v26
; GFX10-NEXT:    v_bfe_u32 v29, v24, 16, 1
; GFX10-NEXT:    v_or_b32_e32 v33, 0x400000, v24
; GFX10-NEXT:    v_lshlrev_b32_e32 v31, 16, v25
; GFX10-NEXT:    v_cndmask_b32_e32 v26, v20, v26, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v21
; GFX10-NEXT:    v_cndmask_b32_e32 v21, v26, v21, vcc_lo
; GFX10-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v30, v31
; GFX10-NEXT:    v_and_b32_e32 v30, 0xffff0000, v0
; GFX10-NEXT:    v_add3_u32 v26, v29, v24, 0x7fff
; GFX10-NEXT:    v_lshrrev_b32_e32 v31, 16, v8
; GFX10-NEXT:    v_cndmask_b32_e32 v29, v25, v28, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v30, v30
; GFX10-NEXT:    v_lshlrev_b32_e32 v29, 16, v29
; GFX10-NEXT:    v_cndmask_b32_e32 v30, v32, v31, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v24, v24
; GFX10-NEXT:    v_cndmask_b32_e32 v24, v26, v33, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v34, v34
; GFX10-NEXT:    v_max_f32_e32 v26, v29, v29
; GFX10-NEXT:    v_lshlrev_b32_e32 v33, 16, v7
; GFX10-NEXT:    v_cndmask_b32_e32 v29, v31, v30, vcc_lo
; GFX10-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v22
; GFX10-NEXT:    v_lshlrev_b32_e32 v31, 16, v30
; GFX10-NEXT:    v_bfe_u32 v22, v26, 16, 1
; GFX10-NEXT:    v_or_b32_e32 v34, 0x400000, v26
; GFX10-NEXT:    v_lshlrev_b32_e32 v32, 16, v29
; GFX10-NEXT:    v_cndmask_b32_e32 v20, v20, v21, vcc_lo
; GFX10-NEXT:    v_lshrrev_b32_e32 v21, 16, v24
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v27
; GFX10-NEXT:    v_add3_u32 v22, v22, v26, 0x7fff
; GFX10-NEXT:    v_and_b32_e32 v24, 0xffff0000, v24
; GFX10-NEXT:    v_cndmask_b32_e32 v27, v21, v27, vcc_lo
; GFX10-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v31, v32
; GFX10-NEXT:    v_lshlrev_b32_e32 v32, 16, v15
; GFX10-NEXT:    v_cndmask_b32_e32 v31, v29, v30, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v33, v33
; GFX10-NEXT:    v_cndmask_b32_e32 v33, v7, v15, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v26, v26
; GFX10-NEXT:    v_lshlrev_b32_e32 v7, 16, v31
; GFX10-NEXT:    v_cndmask_b32_e32 v22, v22, v34, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v32, v32
; GFX10-NEXT:    v_lshlrev_b32_e32 v32, 16, v33
; GFX10-NEXT:    v_lshrrev_b32_e32 v26, 16, v22
; GFX10-NEXT:    v_cndmask_b32_e32 v15, v15, v33, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v23
; GFX10-NEXT:    v_and_b32_e32 v22, 0xffff0000, v22
; GFX10-NEXT:    v_lshlrev_b32_e32 v31, 16, v15
; GFX10-NEXT:    v_cndmask_b32_e32 v23, v27, v23, vcc_lo
; GFX10-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v24
; GFX10-NEXT:    v_max_f32_e32 v27, v7, v7
; GFX10-NEXT:    v_cndmask_b32_e32 v7, v21, v23, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v28
; GFX10-NEXT:    v_bfe_u32 v23, v27, 16, 1
; GFX10-NEXT:    v_cndmask_b32_e32 v21, v26, v28, vcc_lo
; GFX10-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v32, v31
; GFX10-NEXT:    v_add3_u32 v23, v23, v27, 0x7fff
; GFX10-NEXT:    v_lshlrev_b32_e32 v28, 16, v6
; GFX10-NEXT:    v_cndmask_b32_e32 v24, v15, v33, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v25
; GFX10-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
; GFX10-NEXT:    v_cndmask_b32_e32 v21, v21, v25, vcc_lo
; GFX10-NEXT:    v_or_b32_e32 v25, 0x400000, v27
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v27, v27
; GFX10-NEXT:    v_lshlrev_b32_e32 v27, 16, v14
; GFX10-NEXT:    v_max_f32_e32 v24, v24, v24
; GFX10-NEXT:    v_cndmask_b32_e32 v23, v23, v25, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v28, v28
; GFX10-NEXT:    v_or_b32_e32 v28, 0x400000, v24
; GFX10-NEXT:    v_cndmask_b32_e32 v25, v6, v14, vcc_lo
; GFX10-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v22
; GFX10-NEXT:    v_bfe_u32 v22, v24, 16, 1
; GFX10-NEXT:    v_lshlrev_b32_e32 v31, 16, v25
; GFX10-NEXT:    v_cndmask_b32_e32 v6, v26, v21, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v27, v27
; GFX10-NEXT:    v_lshrrev_b32_e32 v21, 16, v23
; GFX10-NEXT:    v_lshlrev_b32_e32 v27, 16, v5
; GFX10-NEXT:    v_add3_u32 v22, v22, v24, 0x7fff
; GFX10-NEXT:    v_and_b32_e32 v23, 0xffff0000, v23
; GFX10-NEXT:    v_cndmask_b32_e32 v14, v14, v25, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v30
; GFX10-NEXT:    v_cndmask_b32_e32 v26, v21, v30, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v27, v27
; GFX10-NEXT:    v_lshlrev_b32_e32 v30, 16, v14
; GFX10-NEXT:    v_cndmask_b32_e32 v27, v5, v13, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v24, v24
; GFX10-NEXT:    v_lshlrev_b32_e32 v5, 16, v13
; GFX10-NEXT:    v_cndmask_b32_e32 v22, v22, v28, vcc_lo
; GFX10-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v31, v30
; GFX10-NEXT:    v_cndmask_b32_e32 v24, v14, v25, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX10-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
; GFX10-NEXT:    v_cndmask_b32_e32 v13, v13, v27, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v29
; GFX10-NEXT:    v_lshlrev_b32_e32 v28, 16, v13
; GFX10-NEXT:    v_cndmask_b32_e32 v5, v26, v29, vcc_lo
; GFX10-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v23
; GFX10-NEXT:    v_lshrrev_b32_e32 v26, 16, v22
; GFX10-NEXT:    v_lshlrev_b32_e32 v29, 16, v27
; GFX10-NEXT:    v_max_f32_e32 v23, v24, v24
; GFX10-NEXT:    v_and_b32_e32 v22, 0xffff0000, v22
; GFX10-NEXT:    v_cndmask_b32_e32 v5, v21, v5, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v33
; GFX10-NEXT:    v_cndmask_b32_e32 v21, v26, v33, vcc_lo
; GFX10-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v29, v28
; GFX10-NEXT:    v_lshlrev_b32_e32 v28, 16, v4
; GFX10-NEXT:    v_lshlrev_b32_e32 v29, 16, v12
; GFX10-NEXT:    v_cndmask_b32_e32 v24, v13, v27, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v15
; GFX10-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
; GFX10-NEXT:    v_cndmask_b32_e32 v15, v21, v15, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v28, v28
; GFX10-NEXT:    v_bfe_u32 v21, v23, 16, 1
; GFX10-NEXT:    v_or_b32_e32 v28, 0x400000, v23
; GFX10-NEXT:    v_max_f32_e32 v24, v24, v24
; GFX10-NEXT:    v_cndmask_b32_e32 v4, v4, v12, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v29, v29
; GFX10-NEXT:    v_add3_u32 v21, v21, v23, 0x7fff
; GFX10-NEXT:    v_lshlrev_b32_e32 v29, 16, v4
; GFX10-NEXT:    v_cndmask_b32_e32 v12, v12, v4, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v23, v23
; GFX10-NEXT:    v_bfe_u32 v23, v24, 16, 1
; GFX10-NEXT:    v_cndmask_b32_e32 v21, v21, v28, vcc_lo
; GFX10-NEXT:    v_lshlrev_b32_e32 v28, 16, v12
; GFX10-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v22
; GFX10-NEXT:    v_add3_u32 v23, v23, v24, 0x7fff
; GFX10-NEXT:    v_lshrrev_b32_e32 v22, 16, v21
; GFX10-NEXT:    v_and_b32_e32 v21, 0xffff0000, v21
; GFX10-NEXT:    v_cndmask_b32_e32 v15, v26, v15, vcc_lo
; GFX10-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v29, v28
; GFX10-NEXT:    v_or_b32_e32 v26, 0x400000, v24
; GFX10-NEXT:    v_lshlrev_b32_e32 v29, 16, v3
; GFX10-NEXT:    v_cndmask_b32_e32 v28, v12, v4, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v25
; GFX10-NEXT:    v_cndmask_b32_e32 v25, v22, v25, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v24, v24
; GFX10-NEXT:    v_lshlrev_b32_e32 v24, 16, v28
; GFX10-NEXT:    v_cndmask_b32_e32 v23, v23, v26, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v29, v29
; GFX10-NEXT:    v_lshlrev_b32_e32 v26, 16, v11
; GFX10-NEXT:    v_max_f32_e32 v24, v24, v24
; GFX10-NEXT:    v_cndmask_b32_e32 v3, v3, v11, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v14
; GFX10-NEXT:    v_or_b32_e32 v30, 0x400000, v24
; GFX10-NEXT:    v_lshlrev_b32_e32 v29, 16, v3
; GFX10-NEXT:    v_cndmask_b32_e32 v14, v25, v14, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v26, v26
; GFX10-NEXT:    v_lshrrev_b32_e32 v25, 16, v23
; GFX10-NEXT:    v_bfe_u32 v26, v24, 16, 1
; GFX10-NEXT:    v_and_b32_e32 v23, 0xffff0000, v23
; GFX10-NEXT:    v_cndmask_b32_e32 v11, v11, v3, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v27
; GFX10-NEXT:    v_add3_u32 v26, v26, v24, 0x7fff
; GFX10-NEXT:    v_lshlrev_b32_e32 v28, 16, v11
; GFX10-NEXT:    v_cndmask_b32_e32 v27, v25, v27, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v13
; GFX10-NEXT:    v_cndmask_b32_e32 v13, v27, v13, vcc_lo
; GFX10-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v29, v28
; GFX10-NEXT:    v_lshlrev_b32_e32 v29, 16, v9
; GFX10-NEXT:    v_cndmask_b32_e32 v27, v11, v3, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v24, v24
; GFX10-NEXT:    v_cndmask_b32_e32 v24, v26, v30, vcc_lo
; GFX10-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v21
; GFX10-NEXT:    v_lshlrev_b32_e32 v21, 16, v27
; GFX10-NEXT:    v_lshlrev_b32_e32 v27, 16, v0
; GFX10-NEXT:    v_cndmask_b32_e32 v14, v22, v14, vcc_lo
; GFX10-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v23
; GFX10-NEXT:    v_lshrrev_b32_e32 v22, 16, v24
; GFX10-NEXT:    v_lshlrev_b32_e32 v23, 16, v2
; GFX10-NEXT:    v_max_f32_e32 v21, v21, v21
; GFX10-NEXT:    v_cndmask_b32_e32 v13, v25, v13, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v4
; GFX10-NEXT:    v_lshlrev_b32_e32 v25, 16, v10
; GFX10-NEXT:    v_cndmask_b32_e32 v4, v22, v4, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v23, v23
; GFX10-NEXT:    v_bfe_u32 v23, v21, 16, 1
; GFX10-NEXT:    v_cndmask_b32_e32 v2, v2, v10, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v12
; GFX10-NEXT:    v_add3_u32 v23, v23, v21, 0x7fff
; GFX10-NEXT:    v_lshlrev_b32_e32 v28, 16, v2
; GFX10-NEXT:    v_cndmask_b32_e32 v4, v4, v12, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v25, v25
; GFX10-NEXT:    v_lshlrev_b32_e32 v25, 16, v1
; GFX10-NEXT:    v_and_b32_e32 v12, 0xffff0000, v24
; GFX10-NEXT:    v_or_b32_e32 v24, 0x400000, v21
; GFX10-NEXT:    v_cndmask_b32_e32 v10, v10, v2, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v25, v25
; GFX10-NEXT:    v_lshlrev_b32_e32 v25, 16, v8
; GFX10-NEXT:    v_lshlrev_b32_e32 v26, 16, v10
; GFX10-NEXT:    v_cndmask_b32_e32 v1, v1, v9, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v27, v27
; GFX10-NEXT:    v_lshlrev_b32_e32 v27, 16, v1
; GFX10-NEXT:    v_cndmask_b32_e32 v0, v0, v8, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v29, v29
; GFX10-NEXT:    v_cndmask_b32_e32 v9, v9, v1, vcc_lo
; GFX10-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v28, v26
; GFX10-NEXT:    v_cndmask_b32_e32 v26, v10, v2, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v25, v25
; GFX10-NEXT:    v_lshlrev_b32_e32 v25, 16, v9
; GFX10-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
; GFX10-NEXT:    v_cndmask_b32_e32 v8, v8, v0, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v21, v21
; GFX10-NEXT:    v_cndmask_b32_e32 v21, v23, v24, vcc_lo
; GFX10-NEXT:    v_lshlrev_b32_e32 v23, 16, v8
; GFX10-NEXT:    v_lshlrev_b32_e32 v24, 16, v0
; GFX10-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v27, v25
; GFX10-NEXT:    v_lshrrev_b32_e32 v27, 16, v21
; GFX10-NEXT:    v_and_b32_e32 v21, 0xffff0000, v21
; GFX10-NEXT:    v_cndmask_b32_e32 v25, v9, v1, vcc_lo
; GFX10-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v24, v23
; GFX10-NEXT:    v_max_f32_e32 v24, v26, v26
; GFX10-NEXT:    v_lshlrev_b32_e32 v25, 16, v25
; GFX10-NEXT:    v_cndmask_b32_e32 v23, v8, v0, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v3
; GFX10-NEXT:    v_bfe_u32 v26, v24, 16, 1
; GFX10-NEXT:    v_max_f32_e32 v25, v25, v25
; GFX10-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; GFX10-NEXT:    v_cndmask_b32_e32 v3, v27, v3, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v11
; GFX10-NEXT:    v_bfe_u32 v28, v25, 16, 1
; GFX10-NEXT:    v_cndmask_b32_e32 v3, v3, v11, vcc_lo
; GFX10-NEXT:    v_max_f32_e32 v11, v23, v23
; GFX10-NEXT:    v_add3_u32 v23, v26, v24, 0x7fff
; GFX10-NEXT:    v_or_b32_e32 v26, 0x400000, v24
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v24, v24
; GFX10-NEXT:    v_add3_u32 v24, v28, v25, 0x7fff
; GFX10-NEXT:    v_bfe_u32 v29, v11, 16, 1
; GFX10-NEXT:    v_cndmask_b32_e32 v23, v23, v26, vcc_lo
; GFX10-NEXT:    v_or_b32_e32 v26, 0x400000, v25
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v25, v25
; GFX10-NEXT:    v_add3_u32 v28, v29, v11, 0x7fff
; GFX10-NEXT:    v_or_b32_e32 v29, 0x400000, v11
; GFX10-NEXT:    v_lshrrev_b32_e32 v30, 16, v23
; GFX10-NEXT:    v_cndmask_b32_e32 v24, v24, v26, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v11, v11
; GFX10-NEXT:    v_cndmask_b32_e32 v11, v28, v29, vcc_lo
; GFX10-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v21
; GFX10-NEXT:    v_lshrrev_b32_e32 v21, 16, v24
; GFX10-NEXT:    v_lshrrev_b32_e32 v25, 16, v11
; GFX10-NEXT:    v_cndmask_b32_e32 v3, v27, v3, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v2
; GFX10-NEXT:    v_perm_b32 v3, v20, v3, 0x5040100
; GFX10-NEXT:    v_cndmask_b32_e32 v2, v30, v2, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v1
; GFX10-NEXT:    v_cndmask_b32_e32 v1, v21, v1, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v10
; GFX10-NEXT:    v_cndmask_b32_e32 v2, v2, v10, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v0
; GFX10-NEXT:    v_and_b32_e32 v10, 0xffff0000, v11
; GFX10-NEXT:    v_cndmask_b32_e32 v0, v25, v0, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v9
; GFX10-NEXT:    v_cndmask_b32_e32 v1, v1, v9, vcc_lo
; GFX10-NEXT:    v_and_b32_e32 v9, 0xffff0000, v24
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v8
; GFX10-NEXT:    v_cndmask_b32_e32 v0, v0, v8, vcc_lo
; GFX10-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v9
; GFX10-NEXT:    v_and_b32_e32 v8, 0xffff0000, v23
; GFX10-NEXT:    v_cndmask_b32_e32 v1, v21, v1, vcc_lo
; GFX10-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v10
; GFX10-NEXT:    v_perm_b32 v1, v6, v1, 0x5040100
; GFX10-NEXT:    v_cndmask_b32_e32 v0, v25, v0, vcc_lo
; GFX10-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v8
; GFX10-NEXT:    v_perm_b32 v6, v17, v14, 0x5040100
; GFX10-NEXT:    v_perm_b32 v0, v5, v0, 0x5040100
; GFX10-NEXT:    v_cndmask_b32_e32 v2, v30, v2, vcc_lo
; GFX10-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v12
; GFX10-NEXT:    v_perm_b32 v5, v18, v13, 0x5040100
; GFX10-NEXT:    v_perm_b32 v2, v7, v2, 0x5040100
; GFX10-NEXT:    v_cndmask_b32_e32 v4, v22, v4, vcc_lo
; GFX10-NEXT:    v_perm_b32 v7, v16, v15, 0x5040100
; GFX10-NEXT:    v_perm_b32 v4, v19, v4, 0x5040100
; GFX10-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-TRUE16-LABEL: v_maximumnum_v16bf16:
; GFX11-TRUE16:       ; %bb.0:
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v18, v7 :: v_dual_mov_b32 v17, v6
; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v20, v5
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xffff0000, v15
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v16, 0xffff0000, v14
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xffff0000, v18
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff0000, v17
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v24, 0xffff0000, v13
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v25, 0xffff0000, v4
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s0, v7, v7
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s1, v24, v24
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v5.l, v18.h, v15.h, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v7.l, v17.h, v14.h, s0
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.l, v5.l
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v6.l, v15.h, v5.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v16, v16
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v22.l, v7.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.l, v6.l
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v16.l, v14.h, v7.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v22, 16, v22
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.l, v16.l
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v19, v21
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xffff0000, v20
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v19.l, v6.l, v5.l, vcc_lo
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v21, v21
; GFX11-TRUE16-NEXT:    v_cmp_gt_f32_e64 s0, v22, v23
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v22, 0xffff0000, v12
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.l, v19.l
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v19.l, v20.h, v13.h, vcc_lo
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v21.l, v16.l, v7.l, s0
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s0, v22, v22
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v25, v25
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v22.l, v13.h, v19.l, s1
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v24.l, v21.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v25.l, v19.l
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v21.l, v4.h, v12.h, vcc_lo
; GFX11-TRUE16-NEXT:    v_max_f32_e32 v23, v23, v23
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v26.l, v22.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v25, 16, v25
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v5.l
; GFX11-TRUE16-NEXT:    v_bfe_u32 v27, v23, 16, 1
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
; GFX11-TRUE16-NEXT:    v_max_f32_e32 v24, v24, v24
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v28, 0x400000, v23
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s3, v23, v23
; GFX11-TRUE16-NEXT:    v_add3_u32 v27, v27, v23, 0x7fff
; GFX11-TRUE16-NEXT:    v_cmp_gt_f32_e64 s2, v25, v26
; GFX11-TRUE16-NEXT:    v_bfe_u32 v25, v24, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s1, 0, v7.l
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e64 v26, v27, v28, s3
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s3, 0, v6.l
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v23.l, v22.l, v19.l, s2
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v27, 0x400000, v24
; GFX11-TRUE16-NEXT:    v_add3_u32 v25, v25, v24, 0x7fff
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v5.l, v26.h, v5.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v24, v24
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v28, 0xffff0000, v26
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v5.l, v5.l, v6.l, s3
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v24, v25, v27, vcc_lo
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v6.l, v12.h, v21.l, s0
; GFX11-TRUE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v28
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s0, 0, v16.l
; GFX11-TRUE16-NEXT:    v_max_f32_e32 v23, v23, v23
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v5.h, v24.h, v7.l, s1
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v25.l, v21.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v26.l, v6.l
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v7.h, v26.h, v5.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v27, v23, 16, 1
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v5.l, v5.h, v16.l, s0
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v16, 0xffff0000, v24
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v25, 16, v25
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v28, 0xffff0000, v3
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s2, v23, v23
; GFX11-TRUE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v16
; GFX11-TRUE16-NEXT:    v_add3_u32 v16, v27, v23, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v27, 0x400000, v23
; GFX11-TRUE16-NEXT:    v_cmp_gt_f32_e64 s0, v25, v26
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s1, v28, v28
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v25, 0xffff0000, v11
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v6.h, v24.h, v5.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e64 v23, v16, v27, s2
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s2, 0, v19.l
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v7.l, v6.l, v21.l, s0
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v16.l, v3.h, v11.h, s1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s0, v25, v25
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v24, 0xffff0000, v2
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v5.l, v23.h, v19.l, s2
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.l, v7.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v26.l, v16.l
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v7.l, v11.h, v16.l, s0
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s0, v24, v24
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v24, 0xffff0000, v10
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v22.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v27.l, v7.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v25, 0xffff0000, v23
; GFX11-TRUE16-NEXT:    v_max_f32_e32 v28, v19, v19
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v19.l, v2.h, v10.h, s0
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v27, 16, v27
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s0, v24, v24
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v5.l, v5.l, v22.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v24, v28, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_eq_f32_e64 s1, 0, v25
; GFX11-TRUE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v26, v27
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v22.l, v10.h, v19.l, s0
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v25, 0x400000, v28
; GFX11-TRUE16-NEXT:    v_add3_u32 v24, v24, v28, 0x7fff
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v26.l, v19.l
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v23.l, v7.l, v16.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v27.l, v22.l
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v28, v28
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v28, 0xffff0000, v1
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v24, v24, v25 :: v_dual_lshlrev_b32 v27, 16, v27
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v21.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v25.l, v23.l
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s2, v28, v28
; GFX11-TRUE16-NEXT:    v_cmp_gt_f32_e64 s0, v26, v27
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v27, 0xffff0000, v24
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v5.h, v24.h, v21.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v21, 16, v25
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v25, 0xffff0000, v9
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v23.l, v1.h, v9.h, s2
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v6.l
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_max_f32_e32 v26, v21, v21
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v21.l, v22.l, v19.l, s0
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s0, v25, v25
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v6.l, v5.h, v6.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v27
; GFX11-TRUE16-NEXT:    v_bfe_u32 v25, v26, 16, 1
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v28.l, v21.l
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v21.l, v9.h, v23.l, s0
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v27, 0x400000, v26
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v29.l, v23.l
; GFX11-TRUE16-NEXT:    v_add3_u32 v25, v25, v26, 0x7fff
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v30.l, v21.l
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s0, v26, v26
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v16.h, v24.h, v6.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v5.h, v23.h, v5.l, s1
; GFX11-TRUE16-NEXT:    v_max_f32_e32 v26, v28, v28
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v28, 16, v30
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e64 v25, v25, v27, s0
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v27, 16, v29
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s0, 0, v16.l
; GFX11-TRUE16-NEXT:    v_bfe_u32 v24, v26, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s1, v26, v26
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v27, v28
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v27, 0xffff0000, v0
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v5.l, v25.h, v16.l, s0
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s0, 0, v7.l
; GFX11-TRUE16-NEXT:    v_add3_u32 v24, v24, v26, 0x7fff
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v6.l, v21.l, v23.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v27, v27
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v27, 0xffff0000, v8
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v28, 0x400000, v26
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v5.l, v5.l, v7.l, s0
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v26.l, v6.l
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v6.l, v0.h, v8.h, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v27, v27
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e64 v24, v24, v28, s1
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s1, 0, v19.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v27, 0xffff0000, v25
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v7.l, v8.h, v6.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v22.l
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v16.l, v24.h, v19.l, s1
; GFX11-TRUE16-NEXT:    v_max_f32_e32 v19, v26, v26
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v26.l, v6.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v28.l, v7.l
; GFX11-TRUE16-NEXT:    v_cmp_eq_f32_e64 s0, 0, v27
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v16.l, v16.l, v22.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v22, v19, 16, 1
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v27, 16, v28
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v28, 16, v18
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v30, 0x400000, v19
; GFX11-TRUE16-NEXT:    v_add3_u32 v22, v22, v19, 0x7fff
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s2, v19, v19
; GFX11-TRUE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v26, v27
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s1, v28, v28
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v26, 16, v15
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v29, 0xffff0000, v24
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e64 v22, v22, v30, s2
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s2, 0, v23.l
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v19.l, v7.l, v6.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v18.l, v18.l, v15.l, s1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v26, v26
; GFX11-TRUE16-NEXT:    v_cmp_eq_f32_e64 s1, 0, v29
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v15.h, v22.h, v23.l, s2
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s2, 0, v21.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.l, v19.l
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v15.l, v15.l, v18.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v18.h, v25.h, v5.l, s0
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v24.l, v18.l
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v5.l, v15.h, v21.l, s2
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v21, 16, v23
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.l, v15.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v25, 16, v17
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v19.h, v24.h, v16.l, s1
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_dual_max_f32 v21, v21, v21 :: v_dual_and_b32 v26, 0xffff0000, v22
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v25, v25
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v25, 16, v14
; GFX11-TRUE16-NEXT:    v_bfe_u32 v27, v21, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_eq_f32_e64 s0, 0, v26
; GFX11-TRUE16-NEXT:    v_cmp_gt_f32_e64 s1, v24, v23
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v16.l, v17.l, v14.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v25, v25
; GFX11-TRUE16-NEXT:    v_add3_u32 v17, v27, v21, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v23, 0x400000, v21
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v14.h, v22.h, v5.l, s0
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v5.l, v15.l, v18.l, s1
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v14.l, v14.l, v16.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v21, v21
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v24, 16, v20
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s0, 0, v7.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.l, v5.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v22.l, v14.l
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v17, v17, v23, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v6.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.l, v16.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v22, 16, v22
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v5.l, v17.h, v6.l, vcc_lo
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v24, v24
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v24, 16, v13
; GFX11-TRUE16-NEXT:    v_max_f32_e32 v21, v21, v21
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v5.l, v5.l, v7.l, s0
; GFX11-TRUE16-NEXT:    v_cmp_gt_f32_e64 s1, v23, v22
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v6.l, v20.l, v13.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v24, v24
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v20, 0xffff0000, v17
; GFX11-TRUE16-NEXT:    v_bfe_u32 v22, v21, 16, 1
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v7.l, v14.l, v16.l, s1
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v25.l, v6.l
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v13.l, v13.l, v6.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v20
; GFX11-TRUE16-NEXT:    v_add3_u32 v20, v22, v21, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v22, 0x400000, v21
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.l, v7.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v24.l, v13.l
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s0, v21, v21
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s1, 0, v15.l
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v21, 16, v23
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v23, 16, v25
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e64 v20, v20, v22, s0
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s0, 0, v18.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v22, 16, v24
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v24, 16, v4
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v4.h, v17.h, v5.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_max_f32_e32 v17, v21, v21
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v5.l, v20.h, v18.l, s0
; GFX11-TRUE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v23, v22
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s0, v24, v24
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v21, 16, v12
; GFX11-TRUE16-NEXT:    v_bfe_u32 v22, v17, 16, 1
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v5.l, v5.l, v15.l, s1
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v7.l, v13.l, v6.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v4.l, v4.l, v12.l, s0
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v21, v21
; GFX11-TRUE16-NEXT:    v_add3_u32 v15, v22, v17, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v21, 0x400000, v17
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v22.l, v7.l
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v23, 0xffff0000, v20
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v12.l, v12.l, v4.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v17, v17
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v17, 16, v22
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v22.l, v4.l
; GFX11-TRUE16-NEXT:    v_cmp_eq_f32_e64 s0, 0, v23
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v15, v15, v21, vcc_lo
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.l, v12.l
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v16.l
; GFX11-TRUE16-NEXT:    v_dual_max_f32 v17, v17, v17 :: v_dual_lshlrev_b32 v22, 16, v22
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v7.l, v20.h, v5.l, s0
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v12.h, v15.h, v16.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v14.l
; GFX11-TRUE16-NEXT:    v_bfe_u32 v23, v17, 16, 1
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v20, 16, v3
; GFX11-TRUE16-NEXT:    v_cmp_gt_f32_e64 s0, v22, v21
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v22, 0x400000, v17
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v3.h, v12.h, v14.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_add3_u32 v21, v23, v17, 0x7fff
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v20, v20
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v5.l, v12.l, v4.l, s0
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v20, 16, v11
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s0, v17, v17
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v3.l, v3.l, v11.l, vcc_lo
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v20, v20
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e64 v17, v21, v22, s0
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s0, 0, v6.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.l, v5.l
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v20, 0xffff0000, v15
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v11.l, v11.l, v3.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v22, 16, v2
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v5.l, v17.h, v6.l, s0
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s0, 0, v13.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
; GFX11-TRUE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v20
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v20, 0xffff0000, v17
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s1, v22, v22
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v5.l, v5.l, v13.l, s0
; GFX11-TRUE16-NEXT:    v_dual_max_f32 v13, v21, v21 :: v_dual_lshlrev_b32 v22, 16, v10
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_cmp_eq_f32_e64 s0, 0, v20
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.l, v11.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.l, v3.l
; GFX11-TRUE16-NEXT:    v_bfe_u32 v23, v13, 16, 1
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v2.l, v2.l, v10.l, s1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s1, v22, v22
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v5.l, v17.h, v5.l, s0
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v17, 16, v1
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
; GFX11-TRUE16-NEXT:    v_add3_u32 v23, v23, v13, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v24, 0x400000, v13
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s3, v13, v13
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v10.l, v10.l, v2.l, s1
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v6.l, v15.h, v3.h, vcc_lo
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v22, 16, v0
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v17, v17
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v17, 16, v9
; GFX11-TRUE16-NEXT:    v_cmp_gt_f32_e64 s2, v20, v21
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e64 v20, v23, v24, s3
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s3, 0, v4.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v15.l, v10.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.l, v2.l
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s0, v22, v22
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v22, 16, v8
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v1.l, v1.l, v9.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v17, v17
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v13.l, v11.l, v3.l, s2
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v2.h, v20.h, v4.l, s3
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v0.l, v0.l, v8.l, s0
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s0, v22, v22
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v4.l, v9.l, v1.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
; GFX11-TRUE16-NEXT:    v_cmp_gt_f32_e64 s1, v21, v15
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v17.l, v1.l
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v8.l, v8.l, v0.l, s0
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v15.l, v4.l
; GFX11-TRUE16-NEXT:    v_max_f32_e32 v13, v13, v13
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v9.l, v10.l, v2.l, s1
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v22.l, v0.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.l, v8.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
; GFX11-TRUE16-NEXT:    v_bfe_u32 v23, v13, 16, 1
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v22, 16, v22
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; GFX11-TRUE16-NEXT:    v_cmp_gt_f32_e64 s0, v17, v15
; GFX11-TRUE16-NEXT:    v_add3_u32 v15, v23, v13, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v17, 0x400000, v13
; GFX11-TRUE16-NEXT:    v_cmp_gt_f32_e64 s1, v22, v21
; GFX11-TRUE16-NEXT:    v_max_f32_e32 v21, v9, v9
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v9.l, v4.l, v1.l, s0
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s0, v13, v13
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v12.l
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v13.l, v8.l, v0.l, s1
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e64 v15, v15, v17, s0
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s0, 0, v3.l
; GFX11-TRUE16-NEXT:    v_bfe_u32 v17, v21, 16, 1
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v0.h, v2.h, v12.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, 0x400000, v21
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v1.h, v15.h, v3.l, s0
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v9
; GFX11-TRUE16-NEXT:    v_add3_u32 v9, v17, v21, 0x7fff
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s0, v21, v21
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v11.l
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xffff0000, v20
; GFX11-TRUE16-NEXT:    v_max_f32_e32 v3, v3, v3
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e64 v9, v9, v12, s0
; GFX11-TRUE16-NEXT:    v_max_f32_e32 v12, v13, v13
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s0, 0, v2.l
; GFX11-TRUE16-NEXT:    v_bfe_u32 v13, v3, 16, 1
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v1.h, v1.h, v11.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v21, 0x400000, v3
; GFX11-TRUE16-NEXT:    v_bfe_u32 v11, v12, 16, 1
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v2.l, v9.h, v2.l, s0
; GFX11-TRUE16-NEXT:    v_add3_u32 v13, v13, v3, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v22, 0x400000, v12
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s0, v3, v3
; GFX11-TRUE16-NEXT:    v_add3_u32 v11, v11, v12, 0x7fff
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s1, v12, v12
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v10.l
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xffff0000, v15
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e64 v3, v13, v21, s0
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s0, 0, v1.l
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e64 v11, v11, v22, s1
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s1, 0, v0.l
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v2.l, v2.l, v10.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v4.l
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v1.l, v3.h, v1.l, s0
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s0, 0, v8.l
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v0.l, v11.h, v0.l, s1
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xffff0000, v11
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v13, 0xffff0000, v3
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xffff0000, v9
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v1.l, v1.l, v4.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v0.l, v0.l, v8.l, s0
; GFX11-TRUE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v10
; GFX11-TRUE16-NEXT:    v_cmp_eq_f32_e64 s0, 0, v13
; GFX11-TRUE16-NEXT:    v_cmp_eq_f32_e64 s1, 0, v21
; GFX11-TRUE16-NEXT:    v_cmp_eq_f32_e64 s2, 0, v12
; GFX11-TRUE16-NEXT:    v_cmp_eq_f32_e64 s3, 0, v17
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v4.l, v11.h, v0.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v14.l, v3.h, v1.l, s0
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v19.l, v9.h, v2.l, s1
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v18.l, v15.h, v1.h, s2
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v16.l, v20.h, v0.h, s3
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, v4 :: v_dual_mov_b32 v1, v14
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, v19 :: v_dual_mov_b32 v3, v18
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v4, v16
; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-FAKE16-LABEL: v_maximumnum_v16bf16:
; GFX11-FAKE16:       ; %bb.0:
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v21, 16, v14
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v16, 0xffff0000, v7
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v17, 16, v15
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v18, 16, v7
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v25, 16, v13
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v26, 16, v5
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v16, v16
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xffff0000, v15
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v22, 16, v6
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v24, 0xffff0000, v5
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v28, 16, v12
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v16, v18, v17, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v19, v19
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v4
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v3
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v20, 0xffff0000, v6
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v18, 16, v16
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v17, v17, v16, vcc_lo
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v1
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v2
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v19, 16, v17
; GFX11-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v18, v19
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v18, v17, v16, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v20, v20
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v20, v22, v21 :: v_dual_and_b32 v19, 0xffff0000, v14
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v19, v19
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v19, v21, v20, vcc_lo
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v21, 16, v20
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v22, 16, v19
; GFX11-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v21, v22
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v21, v19, v20 :: v_dual_lshlrev_b32 v18, 16, v18
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_dual_max_f32 v18, v18, v18 :: v_dual_lshlrev_b32 v21, 16, v21
; GFX11-FAKE16-NEXT:    v_bfe_u32 v23, v18, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v22, 0x400000, v18
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v18, v18
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_max_f32_e32 v21, v21, v21
; GFX11-FAKE16-NEXT:    v_add3_u32 v23, v23, v18, 0x7fff
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v18, v23, v22, vcc_lo
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v22, 0xffff0000, v13
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v24, v24
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v24, 16, v18
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v23, v26, v25, vcc_lo
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v22, v22
; GFX11-FAKE16-NEXT:    v_bfe_u32 v26, v21, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v22, v25, v23, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v16
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v16, v24, v16 :: v_dual_lshlrev_b32 v25, 16, v23
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v17
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v16, v16, v17 :: v_dual_lshlrev_b32 v27, 16, v22
; GFX11-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v25, v27
; GFX11-FAKE16-NEXT:    v_add3_u32 v17, v26, v21, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v26, 0x400000, v21
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v27, 0xffff0000, v4
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v25, v22, v23, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v21, v21
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v21, 16, v25
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v17, v17, v26, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v27, v27
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v25, v29, v28 :: v_dual_and_b32 v18, 0xffff0000, v18
; GFX11-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v18
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v18, 16, v17
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v11
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v16, v24, v16 :: v_dual_lshlrev_b32 v27, 16, v25
; GFX11-FAKE16-NEXT:    v_dual_max_f32 v21, v21, v21 :: v_dual_and_b32 v26, 0xffff0000, v12
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v26, v26
; GFX11-FAKE16-NEXT:    v_bfe_u32 v26, v21, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v24, v28, v25, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v20
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v20, v18, v20, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v19
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v19, v20, v19 :: v_dual_lshlrev_b32 v28, 16, v24
; GFX11-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v27, v28
; GFX11-FAKE16-NEXT:    v_add3_u32 v20, v26, v21, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v26, 0x400000, v21
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v27, v24, v25 :: v_dual_and_b32 v28, 0xffff0000, v3
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v21, v21
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v17, 0xffff0000, v17
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v20, v20, v26 :: v_dual_lshlrev_b32 v21, 16, v27
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v28, v28
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v26, v30, v29 :: v_dual_and_b32 v27, 0xffff0000, v11
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v17
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v10
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v17, v18, v19 :: v_dual_lshlrev_b32 v28, 16, v26
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v27, v27
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v18, 16, v20
; GFX11-FAKE16-NEXT:    v_dual_max_f32 v19, v21, v21 :: v_dual_and_b32 v20, 0xffff0000, v20
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v21, v29, v26, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v23
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v23, v18, v23, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v22
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v22, v23, v22 :: v_dual_lshlrev_b32 v29, 16, v21
; GFX11-FAKE16-NEXT:    v_bfe_u32 v27, v19, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v28, v29
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_add3_u32 v23, v27, v19, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v27, 0x400000, v19
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v28, v21, v26, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v19, v19
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v19, v23, v27, vcc_lo
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v23, 16, v28
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v29, 0xffff0000, v2
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v28, 0xffff0000, v10
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v29, v29
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v27, v31, v30, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v20
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v20, 16, v19
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v9
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v18, v18, v22, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v28, v28
; GFX11-FAKE16-NEXT:    v_max_f32_e32 v22, v23, v23
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v23, v30, v27, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v25
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v28, v22, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v25, v20, v25, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v24
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v24, v25, v24, vcc_lo
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v30, 16, v23
; GFX11-FAKE16-NEXT:    v_add3_u32 v25, v28, v22, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v28, 0x400000, v22
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v29, 16, v27
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v29, v30
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v29, v23, v27 :: v_dual_and_b32 v30, 0xffff0000, v1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v22, v22
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v22, v25, v28 :: v_dual_lshlrev_b32 v25, 16, v29
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xffff0000, v19
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v30, v30
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v28, v32, v31, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v19
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v0
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v29, 0xffff0000, v9
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v30, 16, v28
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v19, v20, v24 :: v_dual_max_f32 v24, v25, v25
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v29, v29
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v20, 16, v22
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v22, 0xffff0000, v22
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v33, 0x400000, v24
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v25, v31, v28, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v26
; GFX11-FAKE16-NEXT:    v_bfe_u32 v29, v24, 16, 1
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v26, v20, v26 :: v_dual_lshlrev_b32 v31, 16, v25
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v21
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v21, v26, v21, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v30, v31
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v30, 0xffff0000, v0
; GFX11-FAKE16-NEXT:    v_add3_u32 v26, v29, v24, 0x7fff
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v8
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v29, v25, v28, vcc_lo
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v30, v30
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v30, v32, v31 :: v_dual_lshlrev_b32 v29, 16, v29
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v34, 0xffff0000, v8
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v24, v24
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v24, v26, v33, vcc_lo
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v34, v34
; GFX11-FAKE16-NEXT:    v_dual_max_f32 v26, v29, v29 :: v_dual_lshlrev_b32 v33, 16, v7
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v29, v31, v30, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v22
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v31, 16, v30
; GFX11-FAKE16-NEXT:    v_bfe_u32 v22, v26, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v34, 0x400000, v26
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v20, v20, v21, vcc_lo
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v21, 16, v24
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v27
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v32, 16, v29
; GFX11-FAKE16-NEXT:    v_add3_u32 v22, v22, v26, 0x7fff
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v27, v21, v27 :: v_dual_and_b32 v24, 0xffff0000, v24
; GFX11-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v31, v32
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v31, v29, v30 :: v_dual_lshlrev_b32 v32, 16, v15
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v33, v33
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v7, v7, v15, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v26, v26
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v22, v22, v34, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v32, v32
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v15, v15, v7, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v23
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v23, v27, v23, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v24
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v27, 16, v22
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v21, v21, v23 :: v_dual_and_b32 v22, 0xffff0000, v22
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v28
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v26, 16, v31
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v31, 16, v15
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v23, v27, v28 :: v_dual_lshlrev_b32 v32, 16, v7
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_max_f32_e32 v26, v26, v26
; GFX11-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v32, v31
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v24, v26, 16, 1
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v28, v15, v7 :: v_dual_lshlrev_b32 v31, 16, v6
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v25
; GFX11-FAKE16-NEXT:    v_add3_u32 v24, v24, v26, 0x7fff
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v23, v23, v25 :: v_dual_lshlrev_b32 v28, 16, v28
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v25, 0x400000, v26
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v26, v26
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v26, 16, v14
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v24, v24, v25, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v31, v31
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v6, v6, v14, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v22
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_dual_max_f32 v25, v28, v28 :: v_dual_lshlrev_b32 v32, 16, v6
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v22, v27, v23, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v26, v26
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v23, 16, v24
; GFX11-FAKE16-NEXT:    v_bfe_u32 v27, v25, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v14, v14, v6, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v30
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_add3_u32 v27, v27, v25, 0x7fff
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v26, v23, v30, vcc_lo
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v30, 0x400000, v25
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v28, 16, v5
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v28, v28
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v5, v5, v13, vcc_lo
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v31, 16, v14
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v25, v25
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v25, v27, v30, vcc_lo
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v32, v31
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v27, v14, v6 :: v_dual_lshlrev_b32 v28, 16, v13
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v28, v28
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v28, 16, v25
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v25, 0xffff0000, v25
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v13, v13, v5, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v29
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v26, v26, v29, vcc_lo
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v24, 0xffff0000, v24
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v29, 16, v13
; GFX11-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v24
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v23, v23, v26, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v7
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v7, v28, v7 :: v_dual_lshlrev_b32 v30, 16, v5
; GFX11-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v30, v29
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v27, 16, v27
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v26, v13, v5 :: v_dual_lshlrev_b32 v29, 16, v12
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_max_f32_e32 v24, v27, v27
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v15
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v7, v7, v15, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v15, v24, 16, 1
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_add3_u32 v15, v15, v24, 0x7fff
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
; GFX11-FAKE16-NEXT:    v_dual_max_f32 v26, v26, v26 :: v_dual_lshlrev_b32 v27, 16, v4
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v27, v27
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v27, 0x400000, v24
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v4, v4, v12, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v29, v29
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v29, 16, v4
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v12, v12, v4, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v24, v24
; GFX11-FAKE16-NEXT:    v_bfe_u32 v24, v26, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v15, v15, v27, vcc_lo
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v27, 16, v12
; GFX11-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v25
; GFX11-FAKE16-NEXT:    v_add3_u32 v24, v24, v26, 0x7fff
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v25, 16, v15
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v15, 0xffff0000, v15
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v7, v28, v7, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v29, v27
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v29, 16, v3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v28, 0x400000, v26
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v27, v12, v4, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v6
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v6, v25, v6, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v26, v26
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v26, 16, v27
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v24, v24, v28, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v29, v29
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v3, v3, v11, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v14
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v6, v6, v14 :: v_dual_lshlrev_b32 v29, 16, v3
; GFX11-FAKE16-NEXT:    v_dual_max_f32 v14, v26, v26 :: v_dual_lshlrev_b32 v27, 16, v11
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v26, 16, v24
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v24, 0xffff0000, v24
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v27, v27
; GFX11-FAKE16-NEXT:    v_bfe_u32 v27, v14, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v30, 0x400000, v14
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v11, v11, v3, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v5
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_add3_u32 v27, v27, v14, 0x7fff
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v5, v26, v5 :: v_dual_lshlrev_b32 v28, 16, v11
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v13
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v5, v5, v13, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v29, v28
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v13, v11, v3, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v14, v14
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v14, v27, v30, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v15
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v27, 16, v0
; GFX11-FAKE16-NEXT:    v_perm_b32 v7, v16, v7, 0x5040100
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v15, 16, v14
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v6, v25, v6, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v24
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v24, 16, v2
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v25, 16, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v6, v17, v6, 0x5040100
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v5, v26, v5, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v4
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v4, v15, v4, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v24, v24
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v2, v10, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v12
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v4, v4, v12, vcc_lo
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v12, 0xffff0000, v14
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v25, v25
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
; GFX11-FAKE16-NEXT:    v_perm_b32 v5, v18, v5, 0x5040100
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v10, v10, v2 :: v_dual_lshlrev_b32 v25, 16, v1
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_dual_max_f32 v13, v13, v13 :: v_dual_lshlrev_b32 v26, 16, v10
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v25, v25
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v28, 16, v2
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v25, 16, v8
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v24, v13, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v1, v9, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v27, v27
; GFX11-FAKE16-NEXT:    v_add3_u32 v14, v24, v13, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v24, 0x400000, v13
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v0, v0, v8 :: v_dual_lshlrev_b32 v27, 16, v1
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v29, 16, v9
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v29, v29
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v9, v9, v1, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v28, v26
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v26, v10, v2, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v25, v25
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v8, v8, v0, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v13, v13
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v13, v14, v24, vcc_lo
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v14, 16, v8
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v25, 16, v9
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v27, v25
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v24, 16, v0
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v27, 16, v13
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v25, v9, v1, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v24, v14
; GFX11-FAKE16-NEXT:    v_max_f32_e32 v24, v26, v26
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v14, v8, v0, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v3
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v26, v24, 16, 1
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v3, v27, v3 :: v_dual_lshlrev_b32 v14, 16, v14
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v11
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v3, v3, v11, vcc_lo
; GFX11-FAKE16-NEXT:    v_max_f32_e32 v11, v14, v14
; GFX11-FAKE16-NEXT:    v_add3_u32 v14, v26, v24, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v26, 0x400000, v24
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v24, v24
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v29, v11, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v14, v14, v26, vcc_lo
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v14
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v25, 16, v25
; GFX11-FAKE16-NEXT:    v_max_f32_e32 v25, v25, v25
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v28, v25, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v26, 0x400000, v25
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v25, v25
; GFX11-FAKE16-NEXT:    v_add3_u32 v24, v28, v25, 0x7fff
; GFX11-FAKE16-NEXT:    v_add3_u32 v28, v29, v11, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v29, 0x400000, v11
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v24, v24, v26, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v11, v11
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v11, v28, v29, vcc_lo
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v25, 16, v11
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v13, 0xffff0000, v13
; GFX11-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v13
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v13, 16, v24
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v3, v27, v3, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v2
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v30, v2, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v13, v1, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v10
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v2, v10, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v0
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v25, v0, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v9
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v1, v9, vcc_lo
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v9, 0xffff0000, v24
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v8
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v0, v8, vcc_lo
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v8, 0xffff0000, v14
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xffff0000, v11
; GFX11-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v9
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v13, v1, vcc_lo
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v1, v22, v1, 0x5040100
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v25, v0, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v8
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_perm_b32 v0, v23, v0, 0x5040100
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v30, v2, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v12
; GFX11-FAKE16-NEXT:    v_perm_b32 v3, v20, v3, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v2, v21, v2, 0x5040100
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v4, v15, v4, vcc_lo
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_perm_b32 v4, v19, v4, 0x5040100
; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
;
; GFX12-TRUE16-LABEL: v_maximumnum_v16bf16:
; GFX12-TRUE16:       ; %bb.0:
; GFX12-TRUE16-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX12-TRUE16-NEXT:    s_wait_expcnt 0x0
; GFX12-TRUE16-NEXT:    s_wait_samplecnt 0x0
; GFX12-TRUE16-NEXT:    s_wait_bvhcnt 0x0
; GFX12-TRUE16-NEXT:    s_wait_kmcnt 0x0
; GFX12-TRUE16-NEXT:    v_dual_mov_b32 v18, v7 :: v_dual_mov_b32 v17, v6
; GFX12-TRUE16-NEXT:    v_mov_b32_e32 v20, v5
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v6, 0xffff0000, v15
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v16, 0xffff0000, v14
; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_4)
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v5, 0xffff0000, v18
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff0000, v17
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v24, 0xffff0000, v13
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v25, 0xffff0000, v4
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s0, v7, v7
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s1, v24, v24
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v5.l, v18.h, v15.h, vcc_lo
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v7.l, v17.h, v14.h, s0
; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v19.l, v5.l
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v6.l, v15.h, v5.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v16, v16
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v22.l, v7.l
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v21.l, v6.l
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v16.l, v14.h, v7.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v22, 16, v22
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v23.l, v16.l
; GFX12-TRUE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v19, v21
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v21, 0xffff0000, v20
; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v19.l, v6.l, v5.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v21, v21
; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX12-TRUE16-NEXT:    v_cmp_gt_f32_e64 s0, v22, v23
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v22, 0xffff0000, v12
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v23.l, v19.l
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v19.l, v20.h, v13.h, vcc_lo
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v21.l, v16.l, v7.l, s0
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s0, v22, v22
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v25, v25
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v22.l, v13.h, v19.l, s1
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v24.l, v21.l
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v25.l, v19.l
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v21.l, v4.h, v12.h, vcc_lo
; GFX12-TRUE16-NEXT:    v_max_num_f32_e32 v23, v23, v23
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v26.l, v22.l
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v25, 16, v25
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v5.l
; GFX12-TRUE16-NEXT:    v_bfe_u32 v27, v23, 16, 1
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
; GFX12-TRUE16-NEXT:    v_max_num_f32_e32 v24, v24, v24
; GFX12-TRUE16-NEXT:    v_or_b32_e32 v28, 0x400000, v23
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s3, v23, v23
; GFX12-TRUE16-NEXT:    v_add3_u32 v27, v27, v23, 0x7fff
; GFX12-TRUE16-NEXT:    v_cmp_gt_f32_e64 s2, v25, v26
; GFX12-TRUE16-NEXT:    v_bfe_u32 v25, v24, 16, 1
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s1, 0, v7.l
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b32_e64 v26, v27, v28, s3
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s3, 0, v6.l
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v23.l, v22.l, v19.l, s2
; GFX12-TRUE16-NEXT:    v_or_b32_e32 v27, 0x400000, v24
; GFX12-TRUE16-NEXT:    v_add3_u32 v25, v25, v24, 0x7fff
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v5.l, v26.h, v5.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v24, v24
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v28, 0xffff0000, v26
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v5.l, v5.l, v6.l, s3
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b32_e32 v24, v25, v27, vcc_lo
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v6.l, v12.h, v21.l, s0
; GFX12-TRUE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v28
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s0, 0, v16.l
; GFX12-TRUE16-NEXT:    v_max_num_f32_e32 v23, v23, v23
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v5.h, v24.h, v7.l, s1
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v25.l, v21.l
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v26.l, v6.l
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v7.h, v26.h, v5.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_bfe_u32 v27, v23, 16, 1
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v5.l, v5.h, v16.l, s0
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v16, 0xffff0000, v24
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v25, 16, v25
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v28, 0xffff0000, v3
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s2, v23, v23
; GFX12-TRUE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v16
; GFX12-TRUE16-NEXT:    v_add3_u32 v16, v27, v23, 0x7fff
; GFX12-TRUE16-NEXT:    v_or_b32_e32 v27, 0x400000, v23
; GFX12-TRUE16-NEXT:    v_cmp_gt_f32_e64 s0, v25, v26
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s1, v28, v28
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v25, 0xffff0000, v11
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v6.h, v24.h, v5.l, vcc_lo
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b32_e64 v23, v16, v27, s2
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s2, 0, v19.l
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v7.l, v6.l, v21.l, s0
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v16.l, v3.h, v11.h, s1
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s0, v25, v25
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v24, 0xffff0000, v2
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v5.l, v23.h, v19.l, s2
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v19.l, v7.l
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v26.l, v16.l
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v7.l, v11.h, v16.l, s0
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s0, v24, v24
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v24, 0xffff0000, v10
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v22.l
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v27.l, v7.l
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v25, 0xffff0000, v23
; GFX12-TRUE16-NEXT:    v_max_num_f32_e32 v28, v19, v19
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v19.l, v2.h, v10.h, s0
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v27, 16, v27
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s0, v24, v24
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v5.l, v5.l, v22.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_bfe_u32 v24, v28, 16, 1
; GFX12-TRUE16-NEXT:    v_cmp_eq_f32_e64 s1, 0, v25
; GFX12-TRUE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v26, v27
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v22.l, v10.h, v19.l, s0
; GFX12-TRUE16-NEXT:    v_or_b32_e32 v25, 0x400000, v28
; GFX12-TRUE16-NEXT:    v_add3_u32 v24, v24, v28, 0x7fff
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v26.l, v19.l
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v23.l, v7.l, v16.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v27.l, v22.l
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v28, v28
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v28, 0xffff0000, v1
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_dual_cndmask_b32 v24, v24, v25 :: v_dual_lshlrev_b32 v27, 16, v27
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v21.l
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v25.l, v23.l
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s2, v28, v28
; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX12-TRUE16-NEXT:    v_cmp_gt_f32_e64 s0, v26, v27
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v27, 0xffff0000, v24
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v5.h, v24.h, v21.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v21, 16, v25
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v25, 0xffff0000, v9
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v23.l, v1.h, v9.h, s2
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v6.l
; GFX12-TRUE16-NEXT:    v_max_num_f32_e32 v26, v21, v21
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v21.l, v22.l, v19.l, s0
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s0, v25, v25
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v6.l, v5.h, v6.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v27
; GFX12-TRUE16-NEXT:    v_bfe_u32 v25, v26, 16, 1
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v28.l, v21.l
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v21.l, v9.h, v23.l, s0
; GFX12-TRUE16-NEXT:    v_or_b32_e32 v27, 0x400000, v26
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v29.l, v23.l
; GFX12-TRUE16-NEXT:    v_add3_u32 v25, v25, v26, 0x7fff
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v30.l, v21.l
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s0, v26, v26
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v16.h, v24.h, v6.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v5.h, v23.h, v5.l, s1
; GFX12-TRUE16-NEXT:    v_max_num_f32_e32 v26, v28, v28
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v28, 16, v30
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b32_e64 v25, v25, v27, s0
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v27, 16, v29
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s0, 0, v16.l
; GFX12-TRUE16-NEXT:    v_bfe_u32 v24, v26, 16, 1
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s1, v26, v26
; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX12-TRUE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v27, v28
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v27, 0xffff0000, v0
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v5.l, v25.h, v16.l, s0
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s0, 0, v7.l
; GFX12-TRUE16-NEXT:    v_add3_u32 v24, v24, v26, 0x7fff
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v6.l, v21.l, v23.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v27, v27
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v27, 0xffff0000, v8
; GFX12-TRUE16-NEXT:    v_or_b32_e32 v28, 0x400000, v26
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v5.l, v5.l, v7.l, s0
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v26.l, v6.l
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v6.l, v0.h, v8.h, vcc_lo
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v27, v27
; GFX12-TRUE16-NEXT:    v_cndmask_b32_e64 v24, v24, v28, s1
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s1, 0, v19.l
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v27, 0xffff0000, v25
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v7.l, v8.h, v6.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v22.l
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v16.l, v24.h, v19.l, s1
; GFX12-TRUE16-NEXT:    v_max_num_f32_e32 v19, v26, v26
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v26.l, v6.l
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v28.l, v7.l
; GFX12-TRUE16-NEXT:    v_cmp_eq_f32_e64 s0, 0, v27
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v16.l, v16.l, v22.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_bfe_u32 v22, v19, 16, 1
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v27, 16, v28
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v28, 16, v18
; GFX12-TRUE16-NEXT:    v_or_b32_e32 v30, 0x400000, v19
; GFX12-TRUE16-NEXT:    v_add3_u32 v22, v22, v19, 0x7fff
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s2, v19, v19
; GFX12-TRUE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v26, v27
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s1, v28, v28
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v26, 16, v15
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v29, 0xffff0000, v24
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b32_e64 v22, v22, v30, s2
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s2, 0, v23.l
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v19.l, v7.l, v6.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v18.l, v18.l, v15.l, s1
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v26, v26
; GFX12-TRUE16-NEXT:    v_cmp_eq_f32_e64 s1, 0, v29
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v15.h, v22.h, v23.l, s2
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s2, 0, v21.l
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v23.l, v19.l
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v15.l, v15.l, v18.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v18.h, v25.h, v5.l, s0
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v24.l, v18.l
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v5.l, v15.h, v21.l, s2
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v21, 16, v23
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v23.l, v15.l
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v25, 16, v17
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v19.h, v24.h, v16.l, s1
; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX12-TRUE16-NEXT:    v_dual_max_num_f32 v21, v21, v21 :: v_dual_and_b32 v26, 0xffff0000, v22
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v25, v25
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v25, 16, v14
; GFX12-TRUE16-NEXT:    v_bfe_u32 v27, v21, 16, 1
; GFX12-TRUE16-NEXT:    v_cmp_eq_f32_e64 s0, 0, v26
; GFX12-TRUE16-NEXT:    v_cmp_gt_f32_e64 s1, v24, v23
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v16.l, v17.l, v14.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v25, v25
; GFX12-TRUE16-NEXT:    v_add3_u32 v17, v27, v21, 0x7fff
; GFX12-TRUE16-NEXT:    v_or_b32_e32 v23, 0x400000, v21
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v14.h, v22.h, v5.l, s0
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v5.l, v15.l, v18.l, s1
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v14.l, v14.l, v16.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v21, v21
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v24, 16, v20
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s0, 0, v7.l
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v21.l, v5.l
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v22.l, v14.l
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b32_e32 v17, v17, v23, vcc_lo
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v6.l
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v23.l, v16.l
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v22, 16, v22
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v5.l, v17.h, v6.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v24, v24
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v24, 16, v13
; GFX12-TRUE16-NEXT:    v_max_num_f32_e32 v21, v21, v21
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v5.l, v5.l, v7.l, s0
; GFX12-TRUE16-NEXT:    v_cmp_gt_f32_e64 s1, v23, v22
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v6.l, v20.l, v13.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v24, v24
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v20, 0xffff0000, v17
; GFX12-TRUE16-NEXT:    v_bfe_u32 v22, v21, 16, 1
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v7.l, v14.l, v16.l, s1
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v25.l, v6.l
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v13.l, v13.l, v6.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v20
; GFX12-TRUE16-NEXT:    v_add3_u32 v20, v22, v21, 0x7fff
; GFX12-TRUE16-NEXT:    v_or_b32_e32 v22, 0x400000, v21
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v23.l, v7.l
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v24.l, v13.l
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s0, v21, v21
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s1, 0, v15.l
; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v21, 16, v23
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v23, 16, v25
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b32_e64 v20, v20, v22, s0
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s0, 0, v18.l
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v22, 16, v24
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v24, 16, v4
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v4.h, v17.h, v5.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_max_num_f32_e32 v17, v21, v21
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v5.l, v20.h, v18.l, s0
; GFX12-TRUE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v23, v22
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s0, v24, v24
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v21, 16, v12
; GFX12-TRUE16-NEXT:    v_bfe_u32 v22, v17, 16, 1
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v5.l, v5.l, v15.l, s1
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v7.l, v13.l, v6.l, vcc_lo
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v4.l, v4.l, v12.l, s0
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v21, v21
; GFX12-TRUE16-NEXT:    v_add3_u32 v15, v22, v17, 0x7fff
; GFX12-TRUE16-NEXT:    v_or_b32_e32 v21, 0x400000, v17
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v22.l, v7.l
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v23, 0xffff0000, v20
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v12.l, v12.l, v4.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v17, v17
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v17, 16, v22
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v22.l, v4.l
; GFX12-TRUE16-NEXT:    v_cmp_eq_f32_e64 s0, 0, v23
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b32_e32 v15, v15, v21, vcc_lo
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v21.l, v12.l
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v16.l
; GFX12-TRUE16-NEXT:    v_dual_max_num_f32 v17, v17, v17 :: v_dual_lshlrev_b32 v22, 16, v22
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v7.l, v20.h, v5.l, s0
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v12.h, v15.h, v16.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v14.l
; GFX12-TRUE16-NEXT:    v_bfe_u32 v23, v17, 16, 1
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v20, 16, v3
; GFX12-TRUE16-NEXT:    v_cmp_gt_f32_e64 s0, v22, v21
; GFX12-TRUE16-NEXT:    v_or_b32_e32 v22, 0x400000, v17
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v3.h, v12.h, v14.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_add3_u32 v21, v23, v17, 0x7fff
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v20, v20
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v5.l, v12.l, v4.l, s0
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v20, 16, v11
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s0, v17, v17
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v3.l, v3.l, v11.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v20, v20
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b32_e64 v17, v21, v22, s0
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s0, 0, v6.l
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v21.l, v5.l
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v20, 0xffff0000, v15
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v11.l, v11.l, v3.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v22, 16, v2
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v5.l, v17.h, v6.l, s0
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s0, 0, v13.l
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
; GFX12-TRUE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v20
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v20, 0xffff0000, v17
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s1, v22, v22
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v5.l, v5.l, v13.l, s0
; GFX12-TRUE16-NEXT:    v_dual_max_num_f32 v13, v21, v21 :: v_dual_lshlrev_b32 v22, 16, v10
; GFX12-TRUE16-NEXT:    v_cmp_eq_f32_e64 s0, 0, v20
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v21.l, v11.l
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v20.l, v3.l
; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX12-TRUE16-NEXT:    v_bfe_u32 v23, v13, 16, 1
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v2.l, v2.l, v10.l, s1
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s1, v22, v22
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v5.l, v17.h, v5.l, s0
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v17, 16, v1
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
; GFX12-TRUE16-NEXT:    v_add3_u32 v23, v23, v13, 0x7fff
; GFX12-TRUE16-NEXT:    v_or_b32_e32 v24, 0x400000, v13
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s3, v13, v13
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v10.l, v10.l, v2.l, s1
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v6.l, v15.h, v3.h, vcc_lo
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v22, 16, v0
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v17, v17
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v17, 16, v9
; GFX12-TRUE16-NEXT:    v_cmp_gt_f32_e64 s2, v20, v21
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b32_e64 v20, v23, v24, s3
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s3, 0, v4.l
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v15.l, v10.l
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v21.l, v2.l
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s0, v22, v22
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v22, 16, v8
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v1.l, v1.l, v9.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v17, v17
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v13.l, v11.l, v3.l, s2
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v2.h, v20.h, v4.l, s3
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v0.l, v0.l, v8.l, s0
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s0, v22, v22
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v4.l, v9.l, v1.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
; GFX12-TRUE16-NEXT:    v_cmp_gt_f32_e64 s1, v21, v15
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v17.l, v1.l
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v8.l, v8.l, v0.l, s0
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v15.l, v4.l
; GFX12-TRUE16-NEXT:    v_max_num_f32_e32 v13, v13, v13
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v9.l, v10.l, v2.l, s1
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v22.l, v0.l
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v21.l, v8.l
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
; GFX12-TRUE16-NEXT:    v_bfe_u32 v23, v13, 16, 1
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v22, 16, v22
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; GFX12-TRUE16-NEXT:    v_cmp_gt_f32_e64 s0, v17, v15
; GFX12-TRUE16-NEXT:    v_add3_u32 v15, v23, v13, 0x7fff
; GFX12-TRUE16-NEXT:    v_or_b32_e32 v17, 0x400000, v13
; GFX12-TRUE16-NEXT:    v_cmp_gt_f32_e64 s1, v22, v21
; GFX12-TRUE16-NEXT:    v_max_num_f32_e32 v21, v9, v9
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v9.l, v4.l, v1.l, s0
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s0, v13, v13
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v12.l
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v13.l, v8.l, v0.l, s1
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b32_e64 v15, v15, v17, s0
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s0, 0, v3.l
; GFX12-TRUE16-NEXT:    v_bfe_u32 v17, v21, 16, 1
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v0.h, v2.h, v12.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_or_b32_e32 v12, 0x400000, v21
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v1.h, v15.h, v3.l, s0
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v9
; GFX12-TRUE16-NEXT:    v_add3_u32 v9, v17, v21, 0x7fff
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s0, v21, v21
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v11.l
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v17, 0xffff0000, v20
; GFX12-TRUE16-NEXT:    v_max_num_f32_e32 v3, v3, v3
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b32_e64 v9, v9, v12, s0
; GFX12-TRUE16-NEXT:    v_max_num_f32_e32 v12, v13, v13
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s0, 0, v2.l
; GFX12-TRUE16-NEXT:    v_bfe_u32 v13, v3, 16, 1
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v1.h, v1.h, v11.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_or_b32_e32 v21, 0x400000, v3
; GFX12-TRUE16-NEXT:    v_bfe_u32 v11, v12, 16, 1
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v2.l, v9.h, v2.l, s0
; GFX12-TRUE16-NEXT:    v_add3_u32 v13, v13, v3, 0x7fff
; GFX12-TRUE16-NEXT:    v_or_b32_e32 v22, 0x400000, v12
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s0, v3, v3
; GFX12-TRUE16-NEXT:    v_add3_u32 v11, v11, v12, 0x7fff
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s1, v12, v12
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v10.l
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v12, 0xffff0000, v15
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b32_e64 v3, v13, v21, s0
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s0, 0, v1.l
; GFX12-TRUE16-NEXT:    v_cndmask_b32_e64 v11, v11, v22, s1
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s1, 0, v0.l
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v2.l, v2.l, v10.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v4.l
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v1.l, v3.h, v1.l, s0
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s0, 0, v8.l
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v0.l, v11.h, v0.l, s1
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v10, 0xffff0000, v11
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v13, 0xffff0000, v3
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v21, 0xffff0000, v9
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v1.l, v1.l, v4.l, vcc_lo
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v0.l, v0.l, v8.l, s0
; GFX12-TRUE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v10
; GFX12-TRUE16-NEXT:    v_cmp_eq_f32_e64 s0, 0, v13
; GFX12-TRUE16-NEXT:    v_cmp_eq_f32_e64 s1, 0, v21
; GFX12-TRUE16-NEXT:    v_cmp_eq_f32_e64 s2, 0, v12
; GFX12-TRUE16-NEXT:    v_cmp_eq_f32_e64 s3, 0, v17
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v4.l, v11.h, v0.l, vcc_lo
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v14.l, v3.h, v1.l, s0
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v19.l, v9.h, v2.l, s1
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v18.l, v15.h, v1.h, s2
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v16.l, v20.h, v0.h, s3
; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX12-TRUE16-NEXT:    v_dual_mov_b32 v0, v4 :: v_dual_mov_b32 v1, v14
; GFX12-TRUE16-NEXT:    v_dual_mov_b32 v2, v19 :: v_dual_mov_b32 v3, v18
; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX12-TRUE16-NEXT:    v_mov_b32_e32 v4, v16
; GFX12-TRUE16-NEXT:    s_setpc_b64 s[30:31]
;
; GFX12-FAKE16-LABEL: v_maximumnum_v16bf16:
; GFX12-FAKE16:       ; %bb.0:
; GFX12-FAKE16-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX12-FAKE16-NEXT:    s_wait_expcnt 0x0
; GFX12-FAKE16-NEXT:    s_wait_samplecnt 0x0
; GFX12-FAKE16-NEXT:    s_wait_bvhcnt 0x0
; GFX12-FAKE16-NEXT:    s_wait_kmcnt 0x0
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v21, 16, v14
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v16, 0xffff0000, v7
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v17, 16, v15
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v18, 16, v7
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v25, 16, v13
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v26, 16, v5
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v16, v16
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v19, 0xffff0000, v15
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v22, 16, v6
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v24, 0xffff0000, v5
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v28, 16, v12
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v16, v18, v17, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v19, v19
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v4
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v3
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v20, 0xffff0000, v6
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v18, 16, v16
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v17, v17, v16, vcc_lo
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v1
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v2
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v19, 16, v17
; GFX12-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v18, v19
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v18, v17, v16, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v20, v20
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v20, v22, v21 :: v_dual_and_b32 v19, 0xffff0000, v14
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v19, v19
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v19, v21, v20, vcc_lo
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v21, 16, v20
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v22, 16, v19
; GFX12-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v21, v22
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v21, v19, v20 :: v_dual_lshlrev_b32 v18, 16, v18
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-FAKE16-NEXT:    v_dual_max_num_f32 v18, v18, v18 :: v_dual_lshlrev_b32 v21, 16, v21
; GFX12-FAKE16-NEXT:    v_bfe_u32 v23, v18, 16, 1
; GFX12-FAKE16-NEXT:    v_or_b32_e32 v22, 0x400000, v18
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v18, v18
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX12-FAKE16-NEXT:    v_max_num_f32_e32 v21, v21, v21
; GFX12-FAKE16-NEXT:    v_add3_u32 v23, v23, v18, 0x7fff
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v18, v23, v22, vcc_lo
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v22, 0xffff0000, v13
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v24, v24
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v24, 16, v18
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v23, v26, v25, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v22, v22
; GFX12-FAKE16-NEXT:    v_bfe_u32 v26, v21, 16, 1
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v22, v25, v23, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v16
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v16, v24, v16 :: v_dual_lshlrev_b32 v25, 16, v23
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v17
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v16, v16, v17 :: v_dual_lshlrev_b32 v27, 16, v22
; GFX12-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v25, v27
; GFX12-FAKE16-NEXT:    v_add3_u32 v17, v26, v21, 0x7fff
; GFX12-FAKE16-NEXT:    v_or_b32_e32 v26, 0x400000, v21
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v27, 0xffff0000, v4
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v25, v22, v23, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v21, v21
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v21, 16, v25
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v17, v17, v26, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v27, v27
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v25, v29, v28 :: v_dual_and_b32 v18, 0xffff0000, v18
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX12-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v18
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v18, 16, v17
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v11
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v16, v24, v16 :: v_dual_lshlrev_b32 v27, 16, v25
; GFX12-FAKE16-NEXT:    v_dual_max_num_f32 v21, v21, v21 :: v_dual_and_b32 v26, 0xffff0000, v12
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v26, v26
; GFX12-FAKE16-NEXT:    v_bfe_u32 v26, v21, 16, 1
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v24, v28, v25, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v20
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v20, v18, v20, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v19
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v19, v20, v19 :: v_dual_lshlrev_b32 v28, 16, v24
; GFX12-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v27, v28
; GFX12-FAKE16-NEXT:    v_add3_u32 v20, v26, v21, 0x7fff
; GFX12-FAKE16-NEXT:    v_or_b32_e32 v26, 0x400000, v21
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v27, v24, v25 :: v_dual_and_b32 v28, 0xffff0000, v3
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v21, v21
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v17, 0xffff0000, v17
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v20, v20, v26 :: v_dual_lshlrev_b32 v21, 16, v27
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v28, v28
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v26, v30, v29 :: v_dual_and_b32 v27, 0xffff0000, v11
; GFX12-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v17
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v10
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v17, v18, v19 :: v_dual_lshlrev_b32 v28, 16, v26
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v27, v27
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v18, 16, v20
; GFX12-FAKE16-NEXT:    v_dual_max_num_f32 v19, v21, v21 :: v_dual_and_b32 v20, 0xffff0000, v20
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v21, v29, v26, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v23
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v23, v18, v23, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v22
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v22, v23, v22 :: v_dual_lshlrev_b32 v29, 16, v21
; GFX12-FAKE16-NEXT:    v_bfe_u32 v27, v19, 16, 1
; GFX12-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v28, v29
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
; GFX12-FAKE16-NEXT:    v_add3_u32 v23, v27, v19, 0x7fff
; GFX12-FAKE16-NEXT:    v_or_b32_e32 v27, 0x400000, v19
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v28, v21, v26, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v19, v19
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v19, v23, v27, vcc_lo
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_2)
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v23, 16, v28
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v29, 0xffff0000, v2
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v28, 0xffff0000, v10
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v29, v29
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v27, v31, v30, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v20
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v20, 16, v19
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v9
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v18, v18, v22, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v28, v28
; GFX12-FAKE16-NEXT:    v_max_num_f32_e32 v22, v23, v23
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v23, v30, v27, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v25
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_2)
; GFX12-FAKE16-NEXT:    v_bfe_u32 v28, v22, 16, 1
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v25, v20, v25, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v24
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v24, v25, v24, vcc_lo
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v30, 16, v23
; GFX12-FAKE16-NEXT:    v_add3_u32 v25, v28, v22, 0x7fff
; GFX12-FAKE16-NEXT:    v_or_b32_e32 v28, 0x400000, v22
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v29, 16, v27
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_4) | instid1(VALU_DEP_2)
; GFX12-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v29, v30
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v29, v23, v27 :: v_dual_and_b32 v30, 0xffff0000, v1
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v22, v22
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v22, v25, v28 :: v_dual_lshlrev_b32 v25, 16, v29
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v19, 0xffff0000, v19
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v30, v30
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v28, v32, v31, vcc_lo
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX12-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v19
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v0
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v29, 0xffff0000, v9
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v30, 16, v28
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v19, v20, v24 :: v_dual_max_num_f32 v24, v25, v25
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v29, v29
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v20, 16, v22
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v22, 0xffff0000, v22
; GFX12-FAKE16-NEXT:    v_or_b32_e32 v33, 0x400000, v24
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v25, v31, v28, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v26
; GFX12-FAKE16-NEXT:    v_bfe_u32 v29, v24, 16, 1
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_2)
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v26, v20, v26 :: v_dual_lshlrev_b32 v31, 16, v25
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v21
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v21, v26, v21, vcc_lo
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX12-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v30, v31
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v30, 0xffff0000, v0
; GFX12-FAKE16-NEXT:    v_add3_u32 v26, v29, v24, 0x7fff
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v8
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v29, v25, v28, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v30, v30
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_4) | instid1(VALU_DEP_3)
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v30, v32, v31 :: v_dual_lshlrev_b32 v29, 16, v29
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v34, 0xffff0000, v8
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v24, v24
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v24, v26, v33, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v34, v34
; GFX12-FAKE16-NEXT:    v_dual_max_num_f32 v26, v29, v29 :: v_dual_lshlrev_b32 v33, 16, v7
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v29, v31, v30, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v22
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v31, 16, v30
; GFX12-FAKE16-NEXT:    v_bfe_u32 v22, v26, 16, 1
; GFX12-FAKE16-NEXT:    v_or_b32_e32 v34, 0x400000, v26
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v20, v20, v21, vcc_lo
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v21, 16, v24
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v27
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v32, 16, v29
; GFX12-FAKE16-NEXT:    v_add3_u32 v22, v22, v26, 0x7fff
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v27, v21, v27 :: v_dual_and_b32 v24, 0xffff0000, v24
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX12-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v31, v32
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v31, v29, v30 :: v_dual_lshlrev_b32 v32, 16, v15
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v33, v33
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v7, v7, v15, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v26, v26
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v22, v22, v34, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v32, v32
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v15, v15, v7, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v23
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v23, v27, v23, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v24
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v27, 16, v22
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v21, v21, v23 :: v_dual_and_b32 v22, 0xffff0000, v22
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v28
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v26, 16, v31
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v31, 16, v15
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v23, v27, v28 :: v_dual_lshlrev_b32 v32, 16, v7
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX12-FAKE16-NEXT:    v_max_num_f32_e32 v26, v26, v26
; GFX12-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v32, v31
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX12-FAKE16-NEXT:    v_bfe_u32 v24, v26, 16, 1
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v28, v15, v7 :: v_dual_lshlrev_b32 v31, 16, v6
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v25
; GFX12-FAKE16-NEXT:    v_add3_u32 v24, v24, v26, 0x7fff
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_3)
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v23, v23, v25 :: v_dual_lshlrev_b32 v28, 16, v28
; GFX12-FAKE16-NEXT:    v_or_b32_e32 v25, 0x400000, v26
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v26, v26
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v26, 16, v14
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v24, v24, v25, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v31, v31
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v6, v6, v14, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v22
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
; GFX12-FAKE16-NEXT:    v_dual_max_num_f32 v25, v28, v28 :: v_dual_lshlrev_b32 v32, 16, v6
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v22, v27, v23, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v26, v26
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v23, 16, v24
; GFX12-FAKE16-NEXT:    v_bfe_u32 v27, v25, 16, 1
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v14, v14, v6, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v30
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
; GFX12-FAKE16-NEXT:    v_add3_u32 v27, v27, v25, 0x7fff
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v26, v23, v30, vcc_lo
; GFX12-FAKE16-NEXT:    v_or_b32_e32 v30, 0x400000, v25
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v28, 16, v5
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v28, v28
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v5, v5, v13, vcc_lo
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v31, 16, v14
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v25, v25
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v25, v27, v30, vcc_lo
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_1)
; GFX12-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v32, v31
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v27, v14, v6 :: v_dual_lshlrev_b32 v28, 16, v13
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v28, v28
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v28, 16, v25
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v25, 0xffff0000, v25
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v13, v13, v5, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v29
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v26, v26, v29, vcc_lo
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v24, 0xffff0000, v24
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v29, 16, v13
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
; GFX12-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v24
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v23, v23, v26, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v7
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v7, v28, v7 :: v_dual_lshlrev_b32 v30, 16, v5
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_2)
; GFX12-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v30, v29
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v27, 16, v27
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v26, v13, v5 :: v_dual_lshlrev_b32 v29, 16, v12
; GFX12-FAKE16-NEXT:    v_max_num_f32_e32 v24, v27, v27
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v15
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v7, v7, v15, vcc_lo
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-FAKE16-NEXT:    v_bfe_u32 v15, v24, 16, 1
; GFX12-FAKE16-NEXT:    v_add3_u32 v15, v15, v24, 0x7fff
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-FAKE16-NEXT:    v_dual_max_num_f32 v26, v26, v26 :: v_dual_lshlrev_b32 v27, 16, v4
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v27, v27
; GFX12-FAKE16-NEXT:    v_or_b32_e32 v27, 0x400000, v24
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v4, v4, v12, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v29, v29
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v29, 16, v4
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v12, v12, v4, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v24, v24
; GFX12-FAKE16-NEXT:    v_bfe_u32 v24, v26, 16, 1
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v15, v15, v27, vcc_lo
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v27, 16, v12
; GFX12-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v25
; GFX12-FAKE16-NEXT:    v_add3_u32 v24, v24, v26, 0x7fff
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v25, 16, v15
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v15, 0xffff0000, v15
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v7, v28, v7, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v29, v27
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v29, 16, v3
; GFX12-FAKE16-NEXT:    v_or_b32_e32 v28, 0x400000, v26
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v27, v12, v4, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v6
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v6, v25, v6, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v26, v26
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v26, 16, v27
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v24, v24, v28, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v29, v29
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v3, v3, v11, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v14
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v6, v6, v14 :: v_dual_lshlrev_b32 v29, 16, v3
; GFX12-FAKE16-NEXT:    v_dual_max_num_f32 v14, v26, v26 :: v_dual_lshlrev_b32 v27, 16, v11
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v26, 16, v24
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v24, 0xffff0000, v24
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v27, v27
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX12-FAKE16-NEXT:    v_bfe_u32 v27, v14, 16, 1
; GFX12-FAKE16-NEXT:    v_or_b32_e32 v30, 0x400000, v14
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v11, v11, v3, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v5
; GFX12-FAKE16-NEXT:    v_add3_u32 v27, v27, v14, 0x7fff
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_2)
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v5, v26, v5 :: v_dual_lshlrev_b32 v28, 16, v11
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v13
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v5, v5, v13, vcc_lo
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX12-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v29, v28
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v13, v11, v3, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v14, v14
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v14, v27, v30, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v15
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v27, 16, v0
; GFX12-FAKE16-NEXT:    v_perm_b32 v7, v16, v7, 0x5040100
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v15, 16, v14
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v6, v25, v6, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v24
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v24, 16, v2
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v25, 16, v10
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX12-FAKE16-NEXT:    v_perm_b32 v6, v17, v6, 0x5040100
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v5, v26, v5, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v4
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v4, v15, v4, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v24, v24
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v2, v10, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v12
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v4, v4, v12, vcc_lo
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v12, 0xffff0000, v14
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v25, v25
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
; GFX12-FAKE16-NEXT:    v_perm_b32 v5, v18, v5, 0x5040100
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v10, v10, v2 :: v_dual_lshlrev_b32 v25, 16, v1
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX12-FAKE16-NEXT:    v_dual_max_num_f32 v13, v13, v13 :: v_dual_lshlrev_b32 v26, 16, v10
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v25, v25
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v28, 16, v2
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v25, 16, v8
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX12-FAKE16-NEXT:    v_bfe_u32 v24, v13, 16, 1
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v1, v9, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v27, v27
; GFX12-FAKE16-NEXT:    v_add3_u32 v14, v24, v13, 0x7fff
; GFX12-FAKE16-NEXT:    v_or_b32_e32 v24, 0x400000, v13
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v0, v0, v8 :: v_dual_lshlrev_b32 v27, 16, v1
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v29, 16, v9
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v29, v29
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v9, v9, v1, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v28, v26
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v26, v10, v2, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v25, v25
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v8, v8, v0, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v13, v13
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v13, v14, v24, vcc_lo
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v14, 16, v8
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v25, 16, v9
; GFX12-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v27, v25
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v24, 16, v0
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v27, 16, v13
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v25, v9, v1, vcc_lo
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_3)
; GFX12-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v24, v14
; GFX12-FAKE16-NEXT:    v_max_num_f32_e32 v24, v26, v26
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v14, v8, v0, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v3
; GFX12-FAKE16-NEXT:    v_bfe_u32 v26, v24, 16, 1
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_2)
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v3, v27, v3 :: v_dual_lshlrev_b32 v14, 16, v14
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v11
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v3, v3, v11, vcc_lo
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_4)
; GFX12-FAKE16-NEXT:    v_max_num_f32_e32 v11, v14, v14
; GFX12-FAKE16-NEXT:    v_add3_u32 v14, v26, v24, 0x7fff
; GFX12-FAKE16-NEXT:    v_or_b32_e32 v26, 0x400000, v24
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v24, v24
; GFX12-FAKE16-NEXT:    v_bfe_u32 v29, v11, 16, 1
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v14, v14, v26, vcc_lo
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v14
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v25, 16, v25
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-FAKE16-NEXT:    v_max_num_f32_e32 v25, v25, v25
; GFX12-FAKE16-NEXT:    v_bfe_u32 v28, v25, 16, 1
; GFX12-FAKE16-NEXT:    v_or_b32_e32 v26, 0x400000, v25
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v25, v25
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX12-FAKE16-NEXT:    v_add3_u32 v24, v28, v25, 0x7fff
; GFX12-FAKE16-NEXT:    v_add3_u32 v28, v29, v11, 0x7fff
; GFX12-FAKE16-NEXT:    v_or_b32_e32 v29, 0x400000, v11
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v24, v24, v26, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v11, v11
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v11, v28, v29, vcc_lo
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v25, 16, v11
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v13, 0xffff0000, v13
; GFX12-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v13
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v13, 16, v24
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v3, v27, v3, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v2
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v30, v2, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v1
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v13, v1, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v10
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v2, v10, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v0
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v25, v0, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v9
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v1, v9, vcc_lo
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v9, 0xffff0000, v24
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v8
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v0, v8, vcc_lo
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v8, 0xffff0000, v14
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v10, 0xffff0000, v11
; GFX12-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v9
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v13, v1, vcc_lo
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX12-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v10
; GFX12-FAKE16-NEXT:    v_perm_b32 v1, v22, v1, 0x5040100
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v25, v0, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v8
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_4) | instid1(VALU_DEP_3)
; GFX12-FAKE16-NEXT:    v_perm_b32 v0, v23, v0, 0x5040100
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v30, v2, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v12
; GFX12-FAKE16-NEXT:    v_perm_b32 v3, v20, v3, 0x5040100
; GFX12-FAKE16-NEXT:    v_perm_b32 v2, v21, v2, 0x5040100
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v4, v15, v4, vcc_lo
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX12-FAKE16-NEXT:    v_perm_b32 v4, v19, v4, 0x5040100
; GFX12-FAKE16-NEXT:    s_setpc_b64 s[30:31]
  %result = call <16 x bfloat> @llvm.maximumnum.v16bf16(<16 x bfloat> %x, <16 x bfloat> %y)
  ret <16 x bfloat> %result
}

define <32 x bfloat> @v_maximumnum_v32bf16(<32 x bfloat> %x, <32 x bfloat> %y) {
; GFX7-LABEL: v_maximumnum_v32bf16:
; GFX7:       ; %bb.0:
; GFX7-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; GFX7-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:128
; GFX7-NEXT:    v_mul_f32_e32 v30, 1.0, v30
; GFX7-NEXT:    v_and_b32_e32 v30, 0xffff0000, v30
; GFX7-NEXT:    v_mul_f32_e32 v29, 1.0, v29
; GFX7-NEXT:    v_and_b32_e32 v29, 0xffff0000, v29
; GFX7-NEXT:    v_mul_f32_e32 v28, 1.0, v28
; GFX7-NEXT:    v_and_b32_e32 v28, 0xffff0000, v28
; GFX7-NEXT:    v_mul_f32_e32 v27, 1.0, v27
; GFX7-NEXT:    v_and_b32_e32 v27, 0xffff0000, v27
; GFX7-NEXT:    v_mul_f32_e32 v26, 1.0, v26
; GFX7-NEXT:    v_and_b32_e32 v26, 0xffff0000, v26
; GFX7-NEXT:    v_mul_f32_e32 v25, 1.0, v25
; GFX7-NEXT:    v_and_b32_e32 v25, 0xffff0000, v25
; GFX7-NEXT:    v_mul_f32_e32 v24, 1.0, v24
; GFX7-NEXT:    v_and_b32_e32 v24, 0xffff0000, v24
; GFX7-NEXT:    v_mul_f32_e32 v23, 1.0, v23
; GFX7-NEXT:    v_and_b32_e32 v23, 0xffff0000, v23
; GFX7-NEXT:    v_mul_f32_e32 v22, 1.0, v22
; GFX7-NEXT:    v_and_b32_e32 v22, 0xffff0000, v22
; GFX7-NEXT:    v_mul_f32_e32 v21, 1.0, v21
; GFX7-NEXT:    v_and_b32_e32 v21, 0xffff0000, v21
; GFX7-NEXT:    v_mul_f32_e32 v20, 1.0, v20
; GFX7-NEXT:    v_and_b32_e32 v20, 0xffff0000, v20
; GFX7-NEXT:    v_mul_f32_e32 v19, 1.0, v19
; GFX7-NEXT:    v_and_b32_e32 v19, 0xffff0000, v19
; GFX7-NEXT:    v_mul_f32_e32 v18, 1.0, v18
; GFX7-NEXT:    v_and_b32_e32 v18, 0xffff0000, v18
; GFX7-NEXT:    v_mul_f32_e32 v17, 1.0, v17
; GFX7-NEXT:    v_and_b32_e32 v17, 0xffff0000, v17
; GFX7-NEXT:    v_mul_f32_e32 v16, 1.0, v16
; GFX7-NEXT:    v_and_b32_e32 v16, 0xffff0000, v16
; GFX7-NEXT:    v_mul_f32_e32 v15, 1.0, v15
; GFX7-NEXT:    v_and_b32_e32 v15, 0xffff0000, v15
; GFX7-NEXT:    v_mul_f32_e32 v14, 1.0, v14
; GFX7-NEXT:    v_and_b32_e32 v14, 0xffff0000, v14
; GFX7-NEXT:    v_mul_f32_e32 v13, 1.0, v13
; GFX7-NEXT:    v_and_b32_e32 v13, 0xffff0000, v13
; GFX7-NEXT:    v_mul_f32_e32 v12, 1.0, v12
; GFX7-NEXT:    v_and_b32_e32 v12, 0xffff0000, v12
; GFX7-NEXT:    v_mul_f32_e32 v11, 1.0, v11
; GFX7-NEXT:    v_and_b32_e32 v11, 0xffff0000, v11
; GFX7-NEXT:    v_mul_f32_e32 v10, 1.0, v10
; GFX7-NEXT:    v_and_b32_e32 v10, 0xffff0000, v10
; GFX7-NEXT:    v_mul_f32_e32 v9, 1.0, v9
; GFX7-NEXT:    v_and_b32_e32 v9, 0xffff0000, v9
; GFX7-NEXT:    v_mul_f32_e32 v8, 1.0, v8
; GFX7-NEXT:    v_and_b32_e32 v8, 0xffff0000, v8
; GFX7-NEXT:    v_mul_f32_e32 v7, 1.0, v7
; GFX7-NEXT:    v_and_b32_e32 v7, 0xffff0000, v7
; GFX7-NEXT:    v_mul_f32_e32 v6, 1.0, v6
; GFX7-NEXT:    v_and_b32_e32 v6, 0xffff0000, v6
; GFX7-NEXT:    v_mul_f32_e32 v5, 1.0, v5
; GFX7-NEXT:    v_and_b32_e32 v5, 0xffff0000, v5
; GFX7-NEXT:    v_mul_f32_e32 v4, 1.0, v4
; GFX7-NEXT:    v_and_b32_e32 v4, 0xffff0000, v4
; GFX7-NEXT:    v_mul_f32_e32 v3, 1.0, v3
; GFX7-NEXT:    v_and_b32_e32 v3, 0xffff0000, v3
; GFX7-NEXT:    v_mul_f32_e32 v2, 1.0, v2
; GFX7-NEXT:    v_and_b32_e32 v2, 0xffff0000, v2
; GFX7-NEXT:    v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT:    v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT:    v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT:    s_waitcnt vmcnt(1)
; GFX7-NEXT:    v_mul_f32_e32 v31, 1.0, v31
; GFX7-NEXT:    s_waitcnt vmcnt(0)
; GFX7-NEXT:    v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT:    v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT:    v_and_b32_e32 v31, 0xffff0000, v31
; GFX7-NEXT:    v_max_f32_e32 v31, v31, v32
; GFX7-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:124
; GFX7-NEXT:    v_and_b32_e32 v31, 0xffff0000, v31
; GFX7-NEXT:    s_waitcnt vmcnt(0)
; GFX7-NEXT:    v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT:    v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT:    v_max_f32_e32 v30, v30, v32
; GFX7-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:120
; GFX7-NEXT:    v_and_b32_e32 v30, 0xffff0000, v30
; GFX7-NEXT:    s_waitcnt vmcnt(0)
; GFX7-NEXT:    v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT:    v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT:    v_max_f32_e32 v29, v29, v32
; GFX7-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:116
; GFX7-NEXT:    v_and_b32_e32 v29, 0xffff0000, v29
; GFX7-NEXT:    s_waitcnt vmcnt(0)
; GFX7-NEXT:    v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT:    v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT:    v_max_f32_e32 v28, v28, v32
; GFX7-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:112
; GFX7-NEXT:    v_and_b32_e32 v28, 0xffff0000, v28
; GFX7-NEXT:    s_waitcnt vmcnt(0)
; GFX7-NEXT:    v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT:    v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT:    v_max_f32_e32 v27, v27, v32
; GFX7-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:108
; GFX7-NEXT:    v_and_b32_e32 v27, 0xffff0000, v27
; GFX7-NEXT:    s_waitcnt vmcnt(0)
; GFX7-NEXT:    v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT:    v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT:    v_max_f32_e32 v26, v26, v32
; GFX7-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:104
; GFX7-NEXT:    v_and_b32_e32 v26, 0xffff0000, v26
; GFX7-NEXT:    s_waitcnt vmcnt(0)
; GFX7-NEXT:    v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT:    v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT:    v_max_f32_e32 v25, v25, v32
; GFX7-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:100
; GFX7-NEXT:    v_and_b32_e32 v25, 0xffff0000, v25
; GFX7-NEXT:    s_waitcnt vmcnt(0)
; GFX7-NEXT:    v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT:    v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT:    v_max_f32_e32 v24, v24, v32
; GFX7-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:96
; GFX7-NEXT:    v_and_b32_e32 v24, 0xffff0000, v24
; GFX7-NEXT:    s_waitcnt vmcnt(0)
; GFX7-NEXT:    v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT:    v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT:    v_max_f32_e32 v23, v23, v32
; GFX7-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:92
; GFX7-NEXT:    v_and_b32_e32 v23, 0xffff0000, v23
; GFX7-NEXT:    s_waitcnt vmcnt(0)
; GFX7-NEXT:    v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT:    v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT:    v_max_f32_e32 v22, v22, v32
; GFX7-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:88
; GFX7-NEXT:    v_and_b32_e32 v22, 0xffff0000, v22
; GFX7-NEXT:    s_waitcnt vmcnt(0)
; GFX7-NEXT:    v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT:    v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT:    v_max_f32_e32 v21, v21, v32
; GFX7-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:84
; GFX7-NEXT:    v_and_b32_e32 v21, 0xffff0000, v21
; GFX7-NEXT:    s_waitcnt vmcnt(0)
; GFX7-NEXT:    v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT:    v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT:    v_max_f32_e32 v20, v20, v32
; GFX7-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:80
; GFX7-NEXT:    v_and_b32_e32 v20, 0xffff0000, v20
; GFX7-NEXT:    s_waitcnt vmcnt(0)
; GFX7-NEXT:    v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT:    v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT:    v_max_f32_e32 v19, v19, v32
; GFX7-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:76
; GFX7-NEXT:    v_and_b32_e32 v19, 0xffff0000, v19
; GFX7-NEXT:    s_waitcnt vmcnt(0)
; GFX7-NEXT:    v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT:    v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT:    v_max_f32_e32 v18, v18, v32
; GFX7-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:72
; GFX7-NEXT:    v_and_b32_e32 v18, 0xffff0000, v18
; GFX7-NEXT:    s_waitcnt vmcnt(0)
; GFX7-NEXT:    v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT:    v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT:    v_max_f32_e32 v17, v17, v32
; GFX7-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:68
; GFX7-NEXT:    v_and_b32_e32 v17, 0xffff0000, v17
; GFX7-NEXT:    s_waitcnt vmcnt(0)
; GFX7-NEXT:    v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT:    v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT:    v_max_f32_e32 v16, v16, v32
; GFX7-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:64
; GFX7-NEXT:    v_and_b32_e32 v16, 0xffff0000, v16
; GFX7-NEXT:    s_waitcnt vmcnt(0)
; GFX7-NEXT:    v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT:    v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT:    v_max_f32_e32 v15, v15, v32
; GFX7-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:60
; GFX7-NEXT:    v_and_b32_e32 v15, 0xffff0000, v15
; GFX7-NEXT:    s_waitcnt vmcnt(0)
; GFX7-NEXT:    v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT:    v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT:    v_max_f32_e32 v14, v14, v32
; GFX7-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:56
; GFX7-NEXT:    v_and_b32_e32 v14, 0xffff0000, v14
; GFX7-NEXT:    s_waitcnt vmcnt(0)
; GFX7-NEXT:    v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT:    v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT:    v_max_f32_e32 v13, v13, v32
; GFX7-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:52
; GFX7-NEXT:    v_and_b32_e32 v13, 0xffff0000, v13
; GFX7-NEXT:    s_waitcnt vmcnt(0)
; GFX7-NEXT:    v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT:    v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT:    v_max_f32_e32 v12, v12, v32
; GFX7-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:48
; GFX7-NEXT:    v_and_b32_e32 v12, 0xffff0000, v12
; GFX7-NEXT:    s_waitcnt vmcnt(0)
; GFX7-NEXT:    v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT:    v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT:    v_max_f32_e32 v11, v11, v32
; GFX7-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:44
; GFX7-NEXT:    v_and_b32_e32 v11, 0xffff0000, v11
; GFX7-NEXT:    s_waitcnt vmcnt(0)
; GFX7-NEXT:    v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT:    v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT:    v_max_f32_e32 v10, v10, v32
; GFX7-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:40
; GFX7-NEXT:    v_and_b32_e32 v10, 0xffff0000, v10
; GFX7-NEXT:    s_waitcnt vmcnt(0)
; GFX7-NEXT:    v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT:    v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT:    v_max_f32_e32 v9, v9, v32
; GFX7-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:36
; GFX7-NEXT:    v_and_b32_e32 v9, 0xffff0000, v9
; GFX7-NEXT:    s_waitcnt vmcnt(0)
; GFX7-NEXT:    v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT:    v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT:    v_max_f32_e32 v8, v8, v32
; GFX7-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:32
; GFX7-NEXT:    v_and_b32_e32 v8, 0xffff0000, v8
; GFX7-NEXT:    s_waitcnt vmcnt(0)
; GFX7-NEXT:    v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT:    v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT:    v_max_f32_e32 v7, v7, v32
; GFX7-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:28
; GFX7-NEXT:    v_and_b32_e32 v7, 0xffff0000, v7
; GFX7-NEXT:    s_waitcnt vmcnt(0)
; GFX7-NEXT:    v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT:    v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT:    v_max_f32_e32 v6, v6, v32
; GFX7-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:24
; GFX7-NEXT:    v_and_b32_e32 v6, 0xffff0000, v6
; GFX7-NEXT:    s_waitcnt vmcnt(0)
; GFX7-NEXT:    v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT:    v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT:    v_max_f32_e32 v5, v5, v32
; GFX7-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:20
; GFX7-NEXT:    v_and_b32_e32 v5, 0xffff0000, v5
; GFX7-NEXT:    s_waitcnt vmcnt(0)
; GFX7-NEXT:    v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT:    v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT:    v_max_f32_e32 v4, v4, v32
; GFX7-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:16
; GFX7-NEXT:    v_and_b32_e32 v4, 0xffff0000, v4
; GFX7-NEXT:    s_waitcnt vmcnt(0)
; GFX7-NEXT:    v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT:    v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT:    v_max_f32_e32 v3, v3, v32
; GFX7-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:12
; GFX7-NEXT:    v_and_b32_e32 v3, 0xffff0000, v3
; GFX7-NEXT:    s_waitcnt vmcnt(0)
; GFX7-NEXT:    v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT:    v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT:    v_max_f32_e32 v2, v2, v32
; GFX7-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:8
; GFX7-NEXT:    v_and_b32_e32 v2, 0xffff0000, v2
; GFX7-NEXT:    s_waitcnt vmcnt(0)
; GFX7-NEXT:    v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT:    v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT:    v_max_f32_e32 v1, v1, v32
; GFX7-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:4
; GFX7-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT:    s_waitcnt vmcnt(0)
; GFX7-NEXT:    v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT:    v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT:    v_max_f32_e32 v0, v0, v32
; GFX7-NEXT:    v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT:    s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_maximumnum_v32bf16:
; GFX8:       ; %bb.0:
; GFX8-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT:    buffer_load_dword v55, off, s[0:3], s32
; GFX8-NEXT:    v_and_b32_e32 v31, 0xffff0000, v14
; GFX8-NEXT:    v_lshrrev_b32_e32 v34, 16, v30
; GFX8-NEXT:    v_lshrrev_b32_e32 v35, 16, v14
; GFX8-NEXT:    v_and_b32_e32 v36, 0xffff0000, v30
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v31, v31
; GFX8-NEXT:    v_cndmask_b32_e32 v31, v35, v34, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v36, v36
; GFX8-NEXT:    v_cndmask_b32_e32 v34, v34, v31, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v35, 16, v31
; GFX8-NEXT:    v_lshlrev_b32_e32 v36, 16, v34
; GFX8-NEXT:    v_cmp_gt_f32_e32 vcc, v35, v36
; GFX8-NEXT:    v_cndmask_b32_e32 v35, v34, v31, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v35, 16, v35
; GFX8-NEXT:    v_mul_f32_e32 v35, 1.0, v35
; GFX8-NEXT:    v_bfe_u32 v36, v35, 16, 1
; GFX8-NEXT:    s_movk_i32 s4, 0x7fff
; GFX8-NEXT:    v_add_u32_e32 v36, vcc, v36, v35
; GFX8-NEXT:    v_add_u32_e32 v36, vcc, s4, v36
; GFX8-NEXT:    v_or_b32_e32 v39, 0x400000, v35
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v35, v35
; GFX8-NEXT:    v_cndmask_b32_e32 v35, v36, v39, vcc
; GFX8-NEXT:    v_lshrrev_b32_e32 v36, 16, v35
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v31
; GFX8-NEXT:    v_and_b32_e32 v35, 0xffff0000, v35
; GFX8-NEXT:    v_cndmask_b32_e32 v31, v36, v31, vcc
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v34
; GFX8-NEXT:    v_and_b32_e32 v32, 0xffff0000, v15
; GFX8-NEXT:    v_cndmask_b32_e32 v31, v31, v34, vcc
; GFX8-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v35
; GFX8-NEXT:    v_lshrrev_b32_e32 v33, 16, v15
; GFX8-NEXT:    v_cndmask_b32_e32 v31, v36, v31, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX8-NEXT:    v_and_b32_e32 v37, 0xffff0000, v13
; GFX8-NEXT:    v_lshrrev_b32_e32 v38, 16, v29
; GFX8-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
; GFX8-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
; GFX8-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
; GFX8-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
; GFX8-NEXT:    s_waitcnt vmcnt(4)
; GFX8-NEXT:    v_lshrrev_b32_e32 v34, 16, v55
; GFX8-NEXT:    v_and_b32_e32 v35, 0xffff0000, v55
; GFX8-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v35, v35
; GFX8-NEXT:    v_cndmask_b32_e32 v34, v34, v32, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v33, 16, v32
; GFX8-NEXT:    v_lshlrev_b32_e32 v35, 16, v34
; GFX8-NEXT:    v_cmp_gt_f32_e32 vcc, v33, v35
; GFX8-NEXT:    v_cndmask_b32_e32 v33, v34, v32, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v33, 16, v33
; GFX8-NEXT:    v_mul_f32_e32 v33, 1.0, v33
; GFX8-NEXT:    v_bfe_u32 v35, v33, 16, 1
; GFX8-NEXT:    v_add_u32_e32 v35, vcc, v35, v33
; GFX8-NEXT:    v_add_u32_e32 v35, vcc, s4, v35
; GFX8-NEXT:    v_or_b32_e32 v36, 0x400000, v33
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v33, v33
; GFX8-NEXT:    v_cndmask_b32_e32 v33, v35, v36, vcc
; GFX8-NEXT:    v_lshrrev_b32_e32 v35, 16, v33
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v32
; GFX8-NEXT:    v_and_b32_e32 v33, 0xffff0000, v33
; GFX8-NEXT:    v_cndmask_b32_e32 v32, v35, v32, vcc
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v34
; GFX8-NEXT:    v_cndmask_b32_e32 v32, v32, v34, vcc
; GFX8-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v33
; GFX8-NEXT:    v_cndmask_b32_e32 v32, v35, v32, vcc
; GFX8-NEXT:    v_lshrrev_b32_e32 v33, 16, v13
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v37, v37
; GFX8-NEXT:    v_and_b32_e32 v35, 0xffff0000, v29
; GFX8-NEXT:    v_cndmask_b32_e32 v33, v33, v38, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v35, v35
; GFX8-NEXT:    v_cndmask_b32_e32 v35, v38, v33, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v34, 16, v33
; GFX8-NEXT:    v_lshlrev_b32_e32 v36, 16, v35
; GFX8-NEXT:    v_cmp_gt_f32_e32 vcc, v34, v36
; GFX8-NEXT:    v_cndmask_b32_e32 v34, v35, v33, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v34, 16, v34
; GFX8-NEXT:    v_mul_f32_e32 v34, 1.0, v34
; GFX8-NEXT:    v_bfe_u32 v36, v34, 16, 1
; GFX8-NEXT:    v_add_u32_e32 v36, vcc, v36, v34
; GFX8-NEXT:    v_add_u32_e32 v36, vcc, s4, v36
; GFX8-NEXT:    v_or_b32_e32 v37, 0x400000, v34
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v34, v34
; GFX8-NEXT:    v_cndmask_b32_e32 v34, v36, v37, vcc
; GFX8-NEXT:    v_lshrrev_b32_e32 v36, 16, v34
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v33
; GFX8-NEXT:    v_cndmask_b32_e32 v33, v36, v33, vcc
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v35
; GFX8-NEXT:    v_and_b32_e32 v34, 0xffff0000, v34
; GFX8-NEXT:    v_cndmask_b32_e32 v33, v33, v35, vcc
; GFX8-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v34
; GFX8-NEXT:    v_and_b32_e32 v34, 0xffff0000, v12
; GFX8-NEXT:    v_cndmask_b32_e32 v33, v36, v33, vcc
; GFX8-NEXT:    v_lshrrev_b32_e32 v35, 16, v28
; GFX8-NEXT:    v_lshrrev_b32_e32 v36, 16, v12
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v34, v34
; GFX8-NEXT:    v_and_b32_e32 v37, 0xffff0000, v28
; GFX8-NEXT:    v_cndmask_b32_e32 v34, v36, v35, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v37, v37
; GFX8-NEXT:    v_cndmask_b32_e32 v35, v35, v34, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v36, 16, v34
; GFX8-NEXT:    v_lshlrev_b32_e32 v37, 16, v35
; GFX8-NEXT:    v_cmp_gt_f32_e32 vcc, v36, v37
; GFX8-NEXT:    v_cndmask_b32_e32 v36, v35, v34, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v36, 16, v36
; GFX8-NEXT:    v_mul_f32_e32 v36, 1.0, v36
; GFX8-NEXT:    v_bfe_u32 v37, v36, 16, 1
; GFX8-NEXT:    v_add_u32_e32 v37, vcc, v37, v36
; GFX8-NEXT:    v_add_u32_e32 v37, vcc, s4, v37
; GFX8-NEXT:    v_or_b32_e32 v38, 0x400000, v36
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v36, v36
; GFX8-NEXT:    v_cndmask_b32_e32 v36, v37, v38, vcc
; GFX8-NEXT:    v_lshrrev_b32_e32 v37, 16, v36
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v34
; GFX8-NEXT:    v_cndmask_b32_e32 v34, v37, v34, vcc
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v35
; GFX8-NEXT:    v_cndmask_b32_e32 v34, v34, v35, vcc
; GFX8-NEXT:    v_and_b32_e32 v35, 0xffff0000, v36
; GFX8-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v35
; GFX8-NEXT:    v_and_b32_e32 v35, 0xffff0000, v11
; GFX8-NEXT:    v_cndmask_b32_e32 v34, v37, v34, vcc
; GFX8-NEXT:    v_lshrrev_b32_e32 v36, 16, v27
; GFX8-NEXT:    v_lshrrev_b32_e32 v37, 16, v11
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v35, v35
; GFX8-NEXT:    v_and_b32_e32 v38, 0xffff0000, v27
; GFX8-NEXT:    v_cndmask_b32_e32 v35, v37, v36, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v38, v38
; GFX8-NEXT:    v_cndmask_b32_e32 v36, v36, v35, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v37, 16, v35
; GFX8-NEXT:    v_lshlrev_b32_e32 v38, 16, v36
; GFX8-NEXT:    v_cmp_gt_f32_e32 vcc, v37, v38
; GFX8-NEXT:    v_cndmask_b32_e32 v37, v36, v35, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v37, 16, v37
; GFX8-NEXT:    v_mul_f32_e32 v37, 1.0, v37
; GFX8-NEXT:    v_bfe_u32 v38, v37, 16, 1
; GFX8-NEXT:    v_add_u32_e32 v38, vcc, v38, v37
; GFX8-NEXT:    v_add_u32_e32 v38, vcc, s4, v38
; GFX8-NEXT:    v_or_b32_e32 v39, 0x400000, v37
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v37, v37
; GFX8-NEXT:    v_cndmask_b32_e32 v37, v38, v39, vcc
; GFX8-NEXT:    v_lshrrev_b32_e32 v38, 16, v37
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v35
; GFX8-NEXT:    v_cndmask_b32_e32 v35, v38, v35, vcc
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v36
; GFX8-NEXT:    v_cndmask_b32_e32 v35, v35, v36, vcc
; GFX8-NEXT:    v_and_b32_e32 v36, 0xffff0000, v37
; GFX8-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v36
; GFX8-NEXT:    v_and_b32_e32 v36, 0xffff0000, v10
; GFX8-NEXT:    v_cndmask_b32_e32 v35, v38, v35, vcc
; GFX8-NEXT:    v_lshrrev_b32_e32 v37, 16, v26
; GFX8-NEXT:    v_lshrrev_b32_e32 v38, 16, v10
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v36, v36
; GFX8-NEXT:    v_and_b32_e32 v39, 0xffff0000, v26
; GFX8-NEXT:    v_cndmask_b32_e32 v36, v38, v37, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v39, v39
; GFX8-NEXT:    v_cndmask_b32_e32 v37, v37, v36, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v38, 16, v36
; GFX8-NEXT:    v_lshlrev_b32_e32 v39, 16, v37
; GFX8-NEXT:    v_cmp_gt_f32_e32 vcc, v38, v39
; GFX8-NEXT:    v_cndmask_b32_e32 v38, v37, v36, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v38, 16, v38
; GFX8-NEXT:    v_mul_f32_e32 v38, 1.0, v38
; GFX8-NEXT:    v_bfe_u32 v39, v38, 16, 1
; GFX8-NEXT:    v_add_u32_e32 v39, vcc, v39, v38
; GFX8-NEXT:    v_add_u32_e32 v39, vcc, s4, v39
; GFX8-NEXT:    v_or_b32_e32 v48, 0x400000, v38
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v38, v38
; GFX8-NEXT:    v_cndmask_b32_e32 v38, v39, v48, vcc
; GFX8-NEXT:    v_lshrrev_b32_e32 v39, 16, v38
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v36
; GFX8-NEXT:    v_cndmask_b32_e32 v36, v39, v36, vcc
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v37
; GFX8-NEXT:    v_cndmask_b32_e32 v36, v36, v37, vcc
; GFX8-NEXT:    v_and_b32_e32 v37, 0xffff0000, v38
; GFX8-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v37
; GFX8-NEXT:    v_and_b32_e32 v37, 0xffff0000, v9
; GFX8-NEXT:    v_cndmask_b32_e32 v36, v39, v36, vcc
; GFX8-NEXT:    v_lshrrev_b32_e32 v38, 16, v25
; GFX8-NEXT:    v_lshrrev_b32_e32 v39, 16, v9
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v37, v37
; GFX8-NEXT:    v_and_b32_e32 v48, 0xffff0000, v25
; GFX8-NEXT:    v_cndmask_b32_e32 v37, v39, v38, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v48, v48
; GFX8-NEXT:    v_cndmask_b32_e32 v38, v38, v37, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v39, 16, v37
; GFX8-NEXT:    v_lshlrev_b32_e32 v48, 16, v38
; GFX8-NEXT:    v_cmp_gt_f32_e32 vcc, v39, v48
; GFX8-NEXT:    v_cndmask_b32_e32 v39, v38, v37, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v39, 16, v39
; GFX8-NEXT:    v_mul_f32_e32 v39, 1.0, v39
; GFX8-NEXT:    v_bfe_u32 v48, v39, 16, 1
; GFX8-NEXT:    v_add_u32_e32 v48, vcc, v48, v39
; GFX8-NEXT:    v_add_u32_e32 v48, vcc, s4, v48
; GFX8-NEXT:    v_or_b32_e32 v49, 0x400000, v39
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v39, v39
; GFX8-NEXT:    v_cndmask_b32_e32 v39, v48, v49, vcc
; GFX8-NEXT:    v_lshrrev_b32_e32 v48, 16, v39
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v37
; GFX8-NEXT:    v_cndmask_b32_e32 v37, v48, v37, vcc
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v38
; GFX8-NEXT:    v_cndmask_b32_e32 v37, v37, v38, vcc
; GFX8-NEXT:    v_and_b32_e32 v38, 0xffff0000, v39
; GFX8-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v38
; GFX8-NEXT:    v_and_b32_e32 v38, 0xffff0000, v8
; GFX8-NEXT:    v_cndmask_b32_e32 v37, v48, v37, vcc
; GFX8-NEXT:    v_lshrrev_b32_e32 v39, 16, v24
; GFX8-NEXT:    v_lshrrev_b32_e32 v48, 16, v8
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v38, v38
; GFX8-NEXT:    v_and_b32_e32 v49, 0xffff0000, v24
; GFX8-NEXT:    v_cndmask_b32_e32 v38, v48, v39, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v49, v49
; GFX8-NEXT:    v_cndmask_b32_e32 v39, v39, v38, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v48, 16, v38
; GFX8-NEXT:    v_lshlrev_b32_e32 v49, 16, v39
; GFX8-NEXT:    v_cmp_gt_f32_e32 vcc, v48, v49
; GFX8-NEXT:    v_cndmask_b32_e32 v48, v39, v38, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v48, 16, v48
; GFX8-NEXT:    v_mul_f32_e32 v48, 1.0, v48
; GFX8-NEXT:    v_bfe_u32 v49, v48, 16, 1
; GFX8-NEXT:    v_add_u32_e32 v49, vcc, v49, v48
; GFX8-NEXT:    v_add_u32_e32 v49, vcc, s4, v49
; GFX8-NEXT:    v_or_b32_e32 v50, 0x400000, v48
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v48, v48
; GFX8-NEXT:    v_cndmask_b32_e32 v48, v49, v50, vcc
; GFX8-NEXT:    v_lshrrev_b32_e32 v49, 16, v48
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v38
; GFX8-NEXT:    v_cndmask_b32_e32 v38, v49, v38, vcc
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v39
; GFX8-NEXT:    v_cndmask_b32_e32 v38, v38, v39, vcc
; GFX8-NEXT:    v_and_b32_e32 v39, 0xffff0000, v48
; GFX8-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v39
; GFX8-NEXT:    v_and_b32_e32 v39, 0xffff0000, v7
; GFX8-NEXT:    v_cndmask_b32_e32 v38, v49, v38, vcc
; GFX8-NEXT:    v_lshrrev_b32_e32 v48, 16, v23
; GFX8-NEXT:    v_lshrrev_b32_e32 v49, 16, v7
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v39, v39
; GFX8-NEXT:    v_and_b32_e32 v50, 0xffff0000, v23
; GFX8-NEXT:    v_cndmask_b32_e32 v39, v49, v48, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v50, v50
; GFX8-NEXT:    v_cndmask_b32_e32 v48, v48, v39, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v49, 16, v39
; GFX8-NEXT:    v_lshlrev_b32_e32 v50, 16, v48
; GFX8-NEXT:    v_cmp_gt_f32_e32 vcc, v49, v50
; GFX8-NEXT:    v_cndmask_b32_e32 v49, v48, v39, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v49, 16, v49
; GFX8-NEXT:    v_mul_f32_e32 v49, 1.0, v49
; GFX8-NEXT:    v_bfe_u32 v50, v49, 16, 1
; GFX8-NEXT:    v_add_u32_e32 v50, vcc, v50, v49
; GFX8-NEXT:    v_add_u32_e32 v50, vcc, s4, v50
; GFX8-NEXT:    v_or_b32_e32 v51, 0x400000, v49
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v49, v49
; GFX8-NEXT:    v_cndmask_b32_e32 v49, v50, v51, vcc
; GFX8-NEXT:    v_lshrrev_b32_e32 v50, 16, v49
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v39
; GFX8-NEXT:    v_cndmask_b32_e32 v39, v50, v39, vcc
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v48
; GFX8-NEXT:    v_cndmask_b32_e32 v39, v39, v48, vcc
; GFX8-NEXT:    v_and_b32_e32 v48, 0xffff0000, v49
; GFX8-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v48
; GFX8-NEXT:    v_and_b32_e32 v48, 0xffff0000, v6
; GFX8-NEXT:    v_cndmask_b32_e32 v39, v50, v39, vcc
; GFX8-NEXT:    v_lshrrev_b32_e32 v49, 16, v22
; GFX8-NEXT:    v_lshrrev_b32_e32 v50, 16, v6
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v48, v48
; GFX8-NEXT:    v_and_b32_e32 v51, 0xffff0000, v22
; GFX8-NEXT:    v_cndmask_b32_e32 v48, v50, v49, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v51, v51
; GFX8-NEXT:    v_cndmask_b32_e32 v49, v49, v48, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v50, 16, v48
; GFX8-NEXT:    v_lshlrev_b32_e32 v51, 16, v49
; GFX8-NEXT:    v_cmp_gt_f32_e32 vcc, v50, v51
; GFX8-NEXT:    v_cndmask_b32_e32 v50, v49, v48, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v50, 16, v50
; GFX8-NEXT:    v_mul_f32_e32 v50, 1.0, v50
; GFX8-NEXT:    v_bfe_u32 v51, v50, 16, 1
; GFX8-NEXT:    v_add_u32_e32 v51, vcc, v51, v50
; GFX8-NEXT:    v_add_u32_e32 v51, vcc, s4, v51
; GFX8-NEXT:    v_or_b32_e32 v52, 0x400000, v50
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v50, v50
; GFX8-NEXT:    v_cndmask_b32_e32 v50, v51, v52, vcc
; GFX8-NEXT:    v_lshrrev_b32_e32 v51, 16, v50
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v48
; GFX8-NEXT:    v_cndmask_b32_e32 v48, v51, v48, vcc
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v49
; GFX8-NEXT:    v_cndmask_b32_e32 v48, v48, v49, vcc
; GFX8-NEXT:    v_and_b32_e32 v49, 0xffff0000, v50
; GFX8-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v49
; GFX8-NEXT:    v_and_b32_e32 v49, 0xffff0000, v5
; GFX8-NEXT:    v_cndmask_b32_e32 v48, v51, v48, vcc
; GFX8-NEXT:    v_lshrrev_b32_e32 v50, 16, v21
; GFX8-NEXT:    v_lshrrev_b32_e32 v51, 16, v5
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v49, v49
; GFX8-NEXT:    v_and_b32_e32 v52, 0xffff0000, v21
; GFX8-NEXT:    v_cndmask_b32_e32 v49, v51, v50, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v52, v52
; GFX8-NEXT:    v_cndmask_b32_e32 v50, v50, v49, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v51, 16, v49
; GFX8-NEXT:    v_lshlrev_b32_e32 v52, 16, v50
; GFX8-NEXT:    v_cmp_gt_f32_e32 vcc, v51, v52
; GFX8-NEXT:    v_cndmask_b32_e32 v51, v50, v49, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v51, 16, v51
; GFX8-NEXT:    v_mul_f32_e32 v51, 1.0, v51
; GFX8-NEXT:    v_bfe_u32 v52, v51, 16, 1
; GFX8-NEXT:    v_add_u32_e32 v52, vcc, v52, v51
; GFX8-NEXT:    v_add_u32_e32 v52, vcc, s4, v52
; GFX8-NEXT:    v_or_b32_e32 v53, 0x400000, v51
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v51, v51
; GFX8-NEXT:    v_cndmask_b32_e32 v51, v52, v53, vcc
; GFX8-NEXT:    v_lshrrev_b32_e32 v52, 16, v51
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v49
; GFX8-NEXT:    v_cndmask_b32_e32 v49, v52, v49, vcc
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v50
; GFX8-NEXT:    v_cndmask_b32_e32 v49, v49, v50, vcc
; GFX8-NEXT:    v_and_b32_e32 v50, 0xffff0000, v51
; GFX8-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v50
; GFX8-NEXT:    v_and_b32_e32 v50, 0xffff0000, v4
; GFX8-NEXT:    v_cndmask_b32_e32 v49, v52, v49, vcc
; GFX8-NEXT:    v_lshrrev_b32_e32 v51, 16, v20
; GFX8-NEXT:    v_lshrrev_b32_e32 v52, 16, v4
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v50, v50
; GFX8-NEXT:    v_and_b32_e32 v53, 0xffff0000, v20
; GFX8-NEXT:    v_cndmask_b32_e32 v50, v52, v51, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v53, v53
; GFX8-NEXT:    v_cndmask_b32_e32 v51, v51, v50, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v52, 16, v50
; GFX8-NEXT:    v_lshlrev_b32_e32 v53, 16, v51
; GFX8-NEXT:    v_cmp_gt_f32_e32 vcc, v52, v53
; GFX8-NEXT:    v_cndmask_b32_e32 v52, v51, v50, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v52, 16, v52
; GFX8-NEXT:    v_mul_f32_e32 v52, 1.0, v52
; GFX8-NEXT:    v_bfe_u32 v53, v52, 16, 1
; GFX8-NEXT:    v_add_u32_e32 v53, vcc, v53, v52
; GFX8-NEXT:    v_add_u32_e32 v53, vcc, s4, v53
; GFX8-NEXT:    v_or_b32_e32 v54, 0x400000, v52
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v52, v52
; GFX8-NEXT:    v_cndmask_b32_e32 v52, v53, v54, vcc
; GFX8-NEXT:    v_lshrrev_b32_e32 v53, 16, v52
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v50
; GFX8-NEXT:    v_cndmask_b32_e32 v50, v53, v50, vcc
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v51
; GFX8-NEXT:    v_cndmask_b32_e32 v50, v50, v51, vcc
; GFX8-NEXT:    v_and_b32_e32 v51, 0xffff0000, v52
; GFX8-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v51
; GFX8-NEXT:    v_and_b32_e32 v51, 0xffff0000, v3
; GFX8-NEXT:    v_cndmask_b32_e32 v50, v53, v50, vcc
; GFX8-NEXT:    v_lshrrev_b32_e32 v52, 16, v19
; GFX8-NEXT:    v_lshrrev_b32_e32 v53, 16, v3
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v51, v51
; GFX8-NEXT:    v_and_b32_e32 v54, 0xffff0000, v19
; GFX8-NEXT:    v_cndmask_b32_e32 v51, v53, v52, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v54, v54
; GFX8-NEXT:    v_cndmask_b32_e32 v52, v52, v51, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v53, 16, v51
; GFX8-NEXT:    v_lshlrev_b32_e32 v54, 16, v52
; GFX8-NEXT:    v_cmp_gt_f32_e32 vcc, v53, v54
; GFX8-NEXT:    v_cndmask_b32_e32 v53, v52, v51, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v53, 16, v53
; GFX8-NEXT:    v_mul_f32_e32 v53, 1.0, v53
; GFX8-NEXT:    v_bfe_u32 v54, v53, 16, 1
; GFX8-NEXT:    v_add_u32_e32 v54, vcc, v54, v53
; GFX8-NEXT:    v_add_u32_e32 v54, vcc, s4, v54
; GFX8-NEXT:    v_or_b32_e32 v40, 0x400000, v53
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v53, v53
; GFX8-NEXT:    v_cndmask_b32_e32 v53, v54, v40, vcc
; GFX8-NEXT:    v_lshrrev_b32_e32 v54, 16, v53
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v51
; GFX8-NEXT:    v_cndmask_b32_e32 v51, v54, v51, vcc
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v52
; GFX8-NEXT:    v_cndmask_b32_e32 v51, v51, v52, vcc
; GFX8-NEXT:    v_and_b32_e32 v52, 0xffff0000, v53
; GFX8-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v52
; GFX8-NEXT:    v_and_b32_e32 v52, 0xffff0000, v2
; GFX8-NEXT:    v_cndmask_b32_e32 v51, v54, v51, vcc
; GFX8-NEXT:    v_lshrrev_b32_e32 v53, 16, v18
; GFX8-NEXT:    v_lshrrev_b32_e32 v54, 16, v2
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v52, v52
; GFX8-NEXT:    v_and_b32_e32 v40, 0xffff0000, v18
; GFX8-NEXT:    v_cndmask_b32_e32 v52, v54, v53, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v40, v40
; GFX8-NEXT:    v_cndmask_b32_e32 v53, v53, v52, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v54, 16, v52
; GFX8-NEXT:    v_lshlrev_b32_e32 v40, 16, v53
; GFX8-NEXT:    v_cmp_gt_f32_e32 vcc, v54, v40
; GFX8-NEXT:    v_cndmask_b32_e32 v54, v53, v52, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v54, 16, v54
; GFX8-NEXT:    v_mul_f32_e32 v54, 1.0, v54
; GFX8-NEXT:    v_bfe_u32 v40, v54, 16, 1
; GFX8-NEXT:    v_add_u32_e32 v40, vcc, v40, v54
; GFX8-NEXT:    v_add_u32_e32 v40, vcc, s4, v40
; GFX8-NEXT:    v_or_b32_e32 v41, 0x400000, v54
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v54, v54
; GFX8-NEXT:    v_cndmask_b32_e32 v54, v40, v41, vcc
; GFX8-NEXT:    v_lshrrev_b32_e32 v40, 16, v54
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v52
; GFX8-NEXT:    v_cndmask_b32_e32 v52, v40, v52, vcc
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v53
; GFX8-NEXT:    v_cndmask_b32_e32 v52, v52, v53, vcc
; GFX8-NEXT:    v_and_b32_e32 v53, 0xffff0000, v54
; GFX8-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v53
; GFX8-NEXT:    v_and_b32_e32 v53, 0xffff0000, v1
; GFX8-NEXT:    v_cndmask_b32_e32 v52, v40, v52, vcc
; GFX8-NEXT:    v_lshrrev_b32_e32 v54, 16, v17
; GFX8-NEXT:    v_lshrrev_b32_e32 v40, 16, v1
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v53, v53
; GFX8-NEXT:    v_and_b32_e32 v41, 0xffff0000, v17
; GFX8-NEXT:    v_cndmask_b32_e32 v53, v40, v54, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v41, v41
; GFX8-NEXT:    v_cndmask_b32_e32 v54, v54, v53, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v40, 16, v53
; GFX8-NEXT:    v_lshlrev_b32_e32 v41, 16, v54
; GFX8-NEXT:    v_cmp_gt_f32_e32 vcc, v40, v41
; GFX8-NEXT:    v_cndmask_b32_e32 v40, v54, v53, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v40, 16, v40
; GFX8-NEXT:    v_mul_f32_e32 v40, 1.0, v40
; GFX8-NEXT:    v_bfe_u32 v41, v40, 16, 1
; GFX8-NEXT:    v_add_u32_e32 v41, vcc, v41, v40
; GFX8-NEXT:    v_add_u32_e32 v41, vcc, s4, v41
; GFX8-NEXT:    v_or_b32_e32 v42, 0x400000, v40
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v40, v40
; GFX8-NEXT:    v_cndmask_b32_e32 v40, v41, v42, vcc
; GFX8-NEXT:    v_lshrrev_b32_e32 v41, 16, v40
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v53
; GFX8-NEXT:    v_cndmask_b32_e32 v53, v41, v53, vcc
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v54
; GFX8-NEXT:    v_cndmask_b32_e32 v53, v53, v54, vcc
; GFX8-NEXT:    v_and_b32_e32 v54, 0xffff0000, v40
; GFX8-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v54
; GFX8-NEXT:    v_and_b32_e32 v54, 0xffff0000, v0
; GFX8-NEXT:    v_cndmask_b32_e32 v53, v41, v53, vcc
; GFX8-NEXT:    v_lshrrev_b32_e32 v40, 16, v16
; GFX8-NEXT:    v_lshrrev_b32_e32 v41, 16, v0
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v54, v54
; GFX8-NEXT:    v_and_b32_e32 v42, 0xffff0000, v16
; GFX8-NEXT:    v_cndmask_b32_e32 v54, v41, v40, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v42, v42
; GFX8-NEXT:    v_cndmask_b32_e32 v40, v40, v54, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v41, 16, v54
; GFX8-NEXT:    v_lshlrev_b32_e32 v42, 16, v40
; GFX8-NEXT:    v_cmp_gt_f32_e32 vcc, v41, v42
; GFX8-NEXT:    v_cndmask_b32_e32 v41, v40, v54, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v41, 16, v41
; GFX8-NEXT:    v_mul_f32_e32 v41, 1.0, v41
; GFX8-NEXT:    v_bfe_u32 v42, v41, 16, 1
; GFX8-NEXT:    v_add_u32_e32 v42, vcc, v42, v41
; GFX8-NEXT:    v_add_u32_e32 v42, vcc, s4, v42
; GFX8-NEXT:    v_or_b32_e32 v43, 0x400000, v41
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v41, v41
; GFX8-NEXT:    v_cndmask_b32_e32 v41, v42, v43, vcc
; GFX8-NEXT:    v_lshrrev_b32_e32 v42, 16, v41
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v54
; GFX8-NEXT:    v_cndmask_b32_e32 v54, v42, v54, vcc
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v40
; GFX8-NEXT:    v_cndmask_b32_e32 v54, v54, v40, vcc
; GFX8-NEXT:    v_and_b32_e32 v40, 0xffff0000, v41
; GFX8-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v40
; GFX8-NEXT:    v_lshlrev_b32_e32 v40, 16, v15
; GFX8-NEXT:    v_cndmask_b32_e32 v54, v42, v54, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v40, v40
; GFX8-NEXT:    v_lshlrev_b32_e32 v40, 16, v55
; GFX8-NEXT:    v_cndmask_b32_e32 v15, v15, v55, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v40, v40
; GFX8-NEXT:    v_cndmask_b32_e32 v55, v55, v15, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v40, 16, v55
; GFX8-NEXT:    v_lshlrev_b32_e32 v41, 16, v15
; GFX8-NEXT:    v_cmp_gt_f32_e32 vcc, v41, v40
; GFX8-NEXT:    v_cndmask_b32_e32 v40, v55, v15, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v40, 16, v40
; GFX8-NEXT:    v_mul_f32_e32 v40, 1.0, v40
; GFX8-NEXT:    v_bfe_u32 v41, v40, 16, 1
; GFX8-NEXT:    v_add_u32_e32 v41, vcc, v41, v40
; GFX8-NEXT:    v_add_u32_e32 v41, vcc, s4, v41
; GFX8-NEXT:    v_or_b32_e32 v42, 0x400000, v40
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v40, v40
; GFX8-NEXT:    v_cndmask_b32_e32 v40, v41, v42, vcc
; GFX8-NEXT:    v_lshrrev_b32_e32 v41, 16, v40
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v15
; GFX8-NEXT:    v_cndmask_b32_e32 v15, v41, v15, vcc
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v55
; GFX8-NEXT:    v_cndmask_b32_e32 v15, v15, v55, vcc
; GFX8-NEXT:    v_and_b32_e32 v55, 0xffff0000, v40
; GFX8-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v55
; GFX8-NEXT:    v_lshlrev_b32_e32 v55, 16, v14
; GFX8-NEXT:    v_cndmask_b32_e32 v15, v41, v15, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v55, v55
; GFX8-NEXT:    v_lshlrev_b32_e32 v55, 16, v30
; GFX8-NEXT:    v_cndmask_b32_e32 v14, v14, v30, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v55, v55
; GFX8-NEXT:    v_cndmask_b32_e32 v30, v30, v14, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v55, 16, v30
; GFX8-NEXT:    v_lshlrev_b32_e32 v40, 16, v14
; GFX8-NEXT:    v_cmp_gt_f32_e32 vcc, v40, v55
; GFX8-NEXT:    v_cndmask_b32_e32 v55, v30, v14, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v55, 16, v55
; GFX8-NEXT:    v_mul_f32_e32 v55, 1.0, v55
; GFX8-NEXT:    v_bfe_u32 v40, v55, 16, 1
; GFX8-NEXT:    v_add_u32_e32 v40, vcc, v40, v55
; GFX8-NEXT:    v_add_u32_e32 v40, vcc, s4, v40
; GFX8-NEXT:    v_or_b32_e32 v41, 0x400000, v55
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v55, v55
; GFX8-NEXT:    v_cndmask_b32_e32 v55, v40, v41, vcc
; GFX8-NEXT:    v_lshrrev_b32_e32 v40, 16, v55
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v14
; GFX8-NEXT:    v_cndmask_b32_e32 v14, v40, v14, vcc
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v30
; GFX8-NEXT:    v_cndmask_b32_e32 v14, v14, v30, vcc
; GFX8-NEXT:    v_and_b32_e32 v30, 0xffff0000, v55
; GFX8-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v30
; GFX8-NEXT:    v_lshlrev_b32_e32 v30, 16, v13
; GFX8-NEXT:    v_cndmask_b32_e32 v14, v40, v14, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v30, v30
; GFX8-NEXT:    v_lshlrev_b32_e32 v30, 16, v29
; GFX8-NEXT:    v_cndmask_b32_e32 v13, v13, v29, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v30, v30
; GFX8-NEXT:    v_cndmask_b32_e32 v29, v29, v13, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v30, 16, v29
; GFX8-NEXT:    v_lshlrev_b32_e32 v55, 16, v13
; GFX8-NEXT:    v_cmp_gt_f32_e32 vcc, v55, v30
; GFX8-NEXT:    v_cndmask_b32_e32 v30, v29, v13, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v30, 16, v30
; GFX8-NEXT:    v_mul_f32_e32 v30, 1.0, v30
; GFX8-NEXT:    v_bfe_u32 v55, v30, 16, 1
; GFX8-NEXT:    v_add_u32_e32 v55, vcc, v55, v30
; GFX8-NEXT:    v_add_u32_e32 v55, vcc, s4, v55
; GFX8-NEXT:    v_or_b32_e32 v40, 0x400000, v30
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v30, v30
; GFX8-NEXT:    v_cndmask_b32_e32 v30, v55, v40, vcc
; GFX8-NEXT:    v_lshrrev_b32_e32 v55, 16, v30
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v13
; GFX8-NEXT:    v_cndmask_b32_e32 v13, v55, v13, vcc
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v29
; GFX8-NEXT:    v_cndmask_b32_e32 v13, v13, v29, vcc
; GFX8-NEXT:    v_and_b32_e32 v29, 0xffff0000, v30
; GFX8-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v29
; GFX8-NEXT:    v_lshlrev_b32_e32 v29, 16, v12
; GFX8-NEXT:    v_cndmask_b32_e32 v13, v55, v13, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v29, v29
; GFX8-NEXT:    v_lshlrev_b32_e32 v29, 16, v28
; GFX8-NEXT:    v_cndmask_b32_e32 v12, v12, v28, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v29, v29
; GFX8-NEXT:    v_cndmask_b32_e32 v28, v28, v12, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v29, 16, v28
; GFX8-NEXT:    v_lshlrev_b32_e32 v30, 16, v12
; GFX8-NEXT:    v_cmp_gt_f32_e32 vcc, v30, v29
; GFX8-NEXT:    v_cndmask_b32_e32 v29, v28, v12, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v29, 16, v29
; GFX8-NEXT:    v_mul_f32_e32 v29, 1.0, v29
; GFX8-NEXT:    v_bfe_u32 v30, v29, 16, 1
; GFX8-NEXT:    v_add_u32_e32 v30, vcc, v30, v29
; GFX8-NEXT:    v_add_u32_e32 v30, vcc, s4, v30
; GFX8-NEXT:    v_or_b32_e32 v55, 0x400000, v29
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v29, v29
; GFX8-NEXT:    v_cndmask_b32_e32 v29, v30, v55, vcc
; GFX8-NEXT:    v_lshrrev_b32_e32 v30, 16, v29
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v12
; GFX8-NEXT:    v_cndmask_b32_e32 v12, v30, v12, vcc
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v28
; GFX8-NEXT:    v_cndmask_b32_e32 v12, v12, v28, vcc
; GFX8-NEXT:    v_and_b32_e32 v28, 0xffff0000, v29
; GFX8-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v28
; GFX8-NEXT:    v_lshlrev_b32_e32 v28, 16, v11
; GFX8-NEXT:    v_cndmask_b32_e32 v12, v30, v12, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v28, v28
; GFX8-NEXT:    v_lshlrev_b32_e32 v28, 16, v27
; GFX8-NEXT:    v_cndmask_b32_e32 v11, v11, v27, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v28, v28
; GFX8-NEXT:    v_cndmask_b32_e32 v27, v27, v11, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v28, 16, v27
; GFX8-NEXT:    v_lshlrev_b32_e32 v29, 16, v11
; GFX8-NEXT:    v_cmp_gt_f32_e32 vcc, v29, v28
; GFX8-NEXT:    v_cndmask_b32_e32 v28, v27, v11, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
; GFX8-NEXT:    v_mul_f32_e32 v28, 1.0, v28
; GFX8-NEXT:    v_bfe_u32 v29, v28, 16, 1
; GFX8-NEXT:    v_add_u32_e32 v29, vcc, v29, v28
; GFX8-NEXT:    v_add_u32_e32 v29, vcc, s4, v29
; GFX8-NEXT:    v_or_b32_e32 v30, 0x400000, v28
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v28, v28
; GFX8-NEXT:    v_cndmask_b32_e32 v28, v29, v30, vcc
; GFX8-NEXT:    v_lshrrev_b32_e32 v29, 16, v28
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v11
; GFX8-NEXT:    v_cndmask_b32_e32 v11, v29, v11, vcc
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v27
; GFX8-NEXT:    v_cndmask_b32_e32 v11, v11, v27, vcc
; GFX8-NEXT:    v_and_b32_e32 v27, 0xffff0000, v28
; GFX8-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v27
; GFX8-NEXT:    v_lshlrev_b32_e32 v27, 16, v10
; GFX8-NEXT:    v_cndmask_b32_e32 v11, v29, v11, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v27, v27
; GFX8-NEXT:    v_lshlrev_b32_e32 v27, 16, v26
; GFX8-NEXT:    v_cndmask_b32_e32 v10, v10, v26, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v27, v27
; GFX8-NEXT:    v_cndmask_b32_e32 v26, v26, v10, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v27, 16, v26
; GFX8-NEXT:    v_lshlrev_b32_e32 v28, 16, v10
; GFX8-NEXT:    v_cmp_gt_f32_e32 vcc, v28, v27
; GFX8-NEXT:    v_cndmask_b32_e32 v27, v26, v10, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v27, 16, v27
; GFX8-NEXT:    v_mul_f32_e32 v27, 1.0, v27
; GFX8-NEXT:    v_bfe_u32 v28, v27, 16, 1
; GFX8-NEXT:    v_add_u32_e32 v28, vcc, v28, v27
; GFX8-NEXT:    v_add_u32_e32 v28, vcc, s4, v28
; GFX8-NEXT:    v_or_b32_e32 v29, 0x400000, v27
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v27, v27
; GFX8-NEXT:    v_cndmask_b32_e32 v27, v28, v29, vcc
; GFX8-NEXT:    v_lshrrev_b32_e32 v28, 16, v27
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v10
; GFX8-NEXT:    v_cndmask_b32_e32 v10, v28, v10, vcc
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v26
; GFX8-NEXT:    v_cndmask_b32_e32 v10, v10, v26, vcc
; GFX8-NEXT:    v_and_b32_e32 v26, 0xffff0000, v27
; GFX8-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v26
; GFX8-NEXT:    v_lshlrev_b32_e32 v26, 16, v9
; GFX8-NEXT:    v_cndmask_b32_e32 v10, v28, v10, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v26, v26
; GFX8-NEXT:    v_lshlrev_b32_e32 v26, 16, v25
; GFX8-NEXT:    v_cndmask_b32_e32 v9, v9, v25, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v26, v26
; GFX8-NEXT:    v_cndmask_b32_e32 v25, v25, v9, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v26, 16, v25
; GFX8-NEXT:    v_lshlrev_b32_e32 v27, 16, v9
; GFX8-NEXT:    v_cmp_gt_f32_e32 vcc, v27, v26
; GFX8-NEXT:    v_cndmask_b32_e32 v26, v25, v9, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
; GFX8-NEXT:    v_mul_f32_e32 v26, 1.0, v26
; GFX8-NEXT:    v_bfe_u32 v27, v26, 16, 1
; GFX8-NEXT:    v_add_u32_e32 v27, vcc, v27, v26
; GFX8-NEXT:    v_add_u32_e32 v27, vcc, s4, v27
; GFX8-NEXT:    v_or_b32_e32 v28, 0x400000, v26
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v26, v26
; GFX8-NEXT:    v_cndmask_b32_e32 v26, v27, v28, vcc
; GFX8-NEXT:    v_lshrrev_b32_e32 v27, 16, v26
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v9
; GFX8-NEXT:    v_cndmask_b32_e32 v9, v27, v9, vcc
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v25
; GFX8-NEXT:    v_cndmask_b32_e32 v9, v9, v25, vcc
; GFX8-NEXT:    v_and_b32_e32 v25, 0xffff0000, v26
; GFX8-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v25
; GFX8-NEXT:    v_lshlrev_b32_e32 v25, 16, v8
; GFX8-NEXT:    v_cndmask_b32_e32 v9, v27, v9, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v25, v25
; GFX8-NEXT:    v_lshlrev_b32_e32 v25, 16, v24
; GFX8-NEXT:    v_cndmask_b32_e32 v8, v8, v24, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v25, v25
; GFX8-NEXT:    v_cndmask_b32_e32 v24, v24, v8, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v25, 16, v24
; GFX8-NEXT:    v_lshlrev_b32_e32 v26, 16, v8
; GFX8-NEXT:    v_cmp_gt_f32_e32 vcc, v26, v25
; GFX8-NEXT:    v_cndmask_b32_e32 v25, v24, v8, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v25, 16, v25
; GFX8-NEXT:    v_mul_f32_e32 v25, 1.0, v25
; GFX8-NEXT:    v_bfe_u32 v26, v25, 16, 1
; GFX8-NEXT:    v_add_u32_e32 v26, vcc, v26, v25
; GFX8-NEXT:    v_add_u32_e32 v26, vcc, s4, v26
; GFX8-NEXT:    v_or_b32_e32 v27, 0x400000, v25
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v25, v25
; GFX8-NEXT:    v_cndmask_b32_e32 v25, v26, v27, vcc
; GFX8-NEXT:    v_lshrrev_b32_e32 v26, 16, v25
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v8
; GFX8-NEXT:    v_cndmask_b32_e32 v8, v26, v8, vcc
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v24
; GFX8-NEXT:    v_cndmask_b32_e32 v8, v8, v24, vcc
; GFX8-NEXT:    v_and_b32_e32 v24, 0xffff0000, v25
; GFX8-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v24
; GFX8-NEXT:    v_lshlrev_b32_e32 v24, 16, v7
; GFX8-NEXT:    v_cndmask_b32_e32 v8, v26, v8, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v24, v24
; GFX8-NEXT:    v_lshlrev_b32_e32 v24, 16, v23
; GFX8-NEXT:    v_cndmask_b32_e32 v7, v7, v23, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v24, v24
; GFX8-NEXT:    v_cndmask_b32_e32 v23, v23, v7, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v24, 16, v23
; GFX8-NEXT:    v_lshlrev_b32_e32 v25, 16, v7
; GFX8-NEXT:    v_cmp_gt_f32_e32 vcc, v25, v24
; GFX8-NEXT:    v_cndmask_b32_e32 v24, v23, v7, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
; GFX8-NEXT:    v_mul_f32_e32 v24, 1.0, v24
; GFX8-NEXT:    v_bfe_u32 v25, v24, 16, 1
; GFX8-NEXT:    v_add_u32_e32 v25, vcc, v25, v24
; GFX8-NEXT:    v_add_u32_e32 v25, vcc, s4, v25
; GFX8-NEXT:    v_or_b32_e32 v26, 0x400000, v24
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v24, v24
; GFX8-NEXT:    v_cndmask_b32_e32 v24, v25, v26, vcc
; GFX8-NEXT:    v_lshrrev_b32_e32 v25, 16, v24
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v7
; GFX8-NEXT:    v_cndmask_b32_e32 v7, v25, v7, vcc
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v23
; GFX8-NEXT:    v_cndmask_b32_e32 v7, v7, v23, vcc
; GFX8-NEXT:    v_and_b32_e32 v23, 0xffff0000, v24
; GFX8-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v23
; GFX8-NEXT:    v_lshlrev_b32_e32 v23, 16, v6
; GFX8-NEXT:    v_cndmask_b32_e32 v7, v25, v7, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v23, v23
; GFX8-NEXT:    v_lshlrev_b32_e32 v23, 16, v22
; GFX8-NEXT:    v_cndmask_b32_e32 v6, v6, v22, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v23, v23
; GFX8-NEXT:    v_cndmask_b32_e32 v22, v22, v6, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v23, 16, v22
; GFX8-NEXT:    v_lshlrev_b32_e32 v24, 16, v6
; GFX8-NEXT:    v_cmp_gt_f32_e32 vcc, v24, v23
; GFX8-NEXT:    v_cndmask_b32_e32 v23, v22, v6, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; GFX8-NEXT:    v_mul_f32_e32 v23, 1.0, v23
; GFX8-NEXT:    v_bfe_u32 v24, v23, 16, 1
; GFX8-NEXT:    v_add_u32_e32 v24, vcc, v24, v23
; GFX8-NEXT:    v_add_u32_e32 v24, vcc, s4, v24
; GFX8-NEXT:    v_or_b32_e32 v25, 0x400000, v23
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v23, v23
; GFX8-NEXT:    v_cndmask_b32_e32 v23, v24, v25, vcc
; GFX8-NEXT:    v_lshrrev_b32_e32 v24, 16, v23
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v6
; GFX8-NEXT:    v_cndmask_b32_e32 v6, v24, v6, vcc
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v22
; GFX8-NEXT:    v_cndmask_b32_e32 v6, v6, v22, vcc
; GFX8-NEXT:    v_and_b32_e32 v22, 0xffff0000, v23
; GFX8-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v22
; GFX8-NEXT:    v_lshlrev_b32_e32 v22, 16, v5
; GFX8-NEXT:    v_cndmask_b32_e32 v6, v24, v6, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v22, v22
; GFX8-NEXT:    v_lshlrev_b32_e32 v22, 16, v21
; GFX8-NEXT:    v_cndmask_b32_e32 v5, v5, v21, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v22, v22
; GFX8-NEXT:    v_cndmask_b32_e32 v21, v21, v5, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v22, 16, v21
; GFX8-NEXT:    v_lshlrev_b32_e32 v23, 16, v5
; GFX8-NEXT:    v_cmp_gt_f32_e32 vcc, v23, v22
; GFX8-NEXT:    v_cndmask_b32_e32 v22, v21, v5, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v22, 16, v22
; GFX8-NEXT:    v_mul_f32_e32 v22, 1.0, v22
; GFX8-NEXT:    v_bfe_u32 v23, v22, 16, 1
; GFX8-NEXT:    v_add_u32_e32 v23, vcc, v23, v22
; GFX8-NEXT:    v_add_u32_e32 v23, vcc, s4, v23
; GFX8-NEXT:    v_or_b32_e32 v24, 0x400000, v22
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v22, v22
; GFX8-NEXT:    v_cndmask_b32_e32 v22, v23, v24, vcc
; GFX8-NEXT:    v_lshrrev_b32_e32 v23, 16, v22
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v5
; GFX8-NEXT:    v_cndmask_b32_e32 v5, v23, v5, vcc
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v21
; GFX8-NEXT:    v_cndmask_b32_e32 v5, v5, v21, vcc
; GFX8-NEXT:    v_and_b32_e32 v21, 0xffff0000, v22
; GFX8-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v21
; GFX8-NEXT:    v_lshlrev_b32_e32 v21, 16, v4
; GFX8-NEXT:    v_cndmask_b32_e32 v5, v23, v5, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v21, v21
; GFX8-NEXT:    v_lshlrev_b32_e32 v21, 16, v20
; GFX8-NEXT:    v_cndmask_b32_e32 v4, v4, v20, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v21, v21
; GFX8-NEXT:    v_cndmask_b32_e32 v20, v20, v4, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v21, 16, v20
; GFX8-NEXT:    v_lshlrev_b32_e32 v22, 16, v4
; GFX8-NEXT:    v_cmp_gt_f32_e32 vcc, v22, v21
; GFX8-NEXT:    v_cndmask_b32_e32 v21, v20, v4, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
; GFX8-NEXT:    v_mul_f32_e32 v21, 1.0, v21
; GFX8-NEXT:    v_bfe_u32 v22, v21, 16, 1
; GFX8-NEXT:    v_add_u32_e32 v22, vcc, v22, v21
; GFX8-NEXT:    v_add_u32_e32 v22, vcc, s4, v22
; GFX8-NEXT:    v_or_b32_e32 v23, 0x400000, v21
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v21, v21
; GFX8-NEXT:    v_cndmask_b32_e32 v21, v22, v23, vcc
; GFX8-NEXT:    v_lshrrev_b32_e32 v22, 16, v21
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v4
; GFX8-NEXT:    v_cndmask_b32_e32 v4, v22, v4, vcc
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v20
; GFX8-NEXT:    v_cndmask_b32_e32 v4, v4, v20, vcc
; GFX8-NEXT:    v_and_b32_e32 v20, 0xffff0000, v21
; GFX8-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v20
; GFX8-NEXT:    v_lshlrev_b32_e32 v20, 16, v3
; GFX8-NEXT:    v_cndmask_b32_e32 v4, v22, v4, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v20, v20
; GFX8-NEXT:    v_lshlrev_b32_e32 v20, 16, v19
; GFX8-NEXT:    v_cndmask_b32_e32 v3, v3, v19, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v20, v20
; GFX8-NEXT:    v_cndmask_b32_e32 v19, v19, v3, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v20, 16, v19
; GFX8-NEXT:    v_lshlrev_b32_e32 v21, 16, v3
; GFX8-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
; GFX8-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
; GFX8-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
; GFX8-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
; GFX8-NEXT:    v_cmp_gt_f32_e32 vcc, v21, v20
; GFX8-NEXT:    v_cndmask_b32_e32 v20, v19, v3, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
; GFX8-NEXT:    v_mul_f32_e32 v20, 1.0, v20
; GFX8-NEXT:    v_bfe_u32 v21, v20, 16, 1
; GFX8-NEXT:    v_add_u32_e32 v21, vcc, v21, v20
; GFX8-NEXT:    v_add_u32_e32 v21, vcc, s4, v21
; GFX8-NEXT:    v_or_b32_e32 v22, 0x400000, v20
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v20, v20
; GFX8-NEXT:    v_cndmask_b32_e32 v20, v21, v22, vcc
; GFX8-NEXT:    v_lshrrev_b32_e32 v21, 16, v20
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v3
; GFX8-NEXT:    v_cndmask_b32_e32 v3, v21, v3, vcc
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v19
; GFX8-NEXT:    v_cndmask_b32_e32 v3, v3, v19, vcc
; GFX8-NEXT:    v_and_b32_e32 v19, 0xffff0000, v20
; GFX8-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v19
; GFX8-NEXT:    v_lshlrev_b32_e32 v19, 16, v2
; GFX8-NEXT:    v_cndmask_b32_e32 v3, v21, v3, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v19, v19
; GFX8-NEXT:    v_lshlrev_b32_e32 v19, 16, v18
; GFX8-NEXT:    v_cndmask_b32_e32 v2, v2, v18, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v19, v19
; GFX8-NEXT:    v_cndmask_b32_e32 v18, v18, v2, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v19, 16, v18
; GFX8-NEXT:    v_lshlrev_b32_e32 v20, 16, v2
; GFX8-NEXT:    v_cmp_gt_f32_e32 vcc, v20, v19
; GFX8-NEXT:    v_cndmask_b32_e32 v19, v18, v2, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
; GFX8-NEXT:    v_mul_f32_e32 v19, 1.0, v19
; GFX8-NEXT:    v_bfe_u32 v20, v19, 16, 1
; GFX8-NEXT:    v_add_u32_e32 v20, vcc, v20, v19
; GFX8-NEXT:    v_add_u32_e32 v20, vcc, s4, v20
; GFX8-NEXT:    v_or_b32_e32 v21, 0x400000, v19
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v19, v19
; GFX8-NEXT:    v_cndmask_b32_e32 v19, v20, v21, vcc
; GFX8-NEXT:    v_lshrrev_b32_e32 v20, 16, v19
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v2
; GFX8-NEXT:    v_cndmask_b32_e32 v2, v20, v2, vcc
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v18
; GFX8-NEXT:    v_cndmask_b32_e32 v2, v2, v18, vcc
; GFX8-NEXT:    v_and_b32_e32 v18, 0xffff0000, v19
; GFX8-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v18
; GFX8-NEXT:    v_lshlrev_b32_e32 v18, 16, v1
; GFX8-NEXT:    v_cndmask_b32_e32 v2, v20, v2, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; GFX8-NEXT:    v_lshlrev_b32_e32 v18, 16, v17
; GFX8-NEXT:    v_cndmask_b32_e32 v1, v1, v17, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; GFX8-NEXT:    v_cndmask_b32_e32 v17, v17, v1, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v18, 16, v17
; GFX8-NEXT:    v_lshlrev_b32_e32 v19, 16, v1
; GFX8-NEXT:    v_cmp_gt_f32_e32 vcc, v19, v18
; GFX8-NEXT:    v_cndmask_b32_e32 v18, v17, v1, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
; GFX8-NEXT:    v_mul_f32_e32 v18, 1.0, v18
; GFX8-NEXT:    v_bfe_u32 v19, v18, 16, 1
; GFX8-NEXT:    v_add_u32_e32 v19, vcc, v19, v18
; GFX8-NEXT:    v_add_u32_e32 v19, vcc, s4, v19
; GFX8-NEXT:    v_or_b32_e32 v20, 0x400000, v18
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; GFX8-NEXT:    v_cndmask_b32_e32 v18, v19, v20, vcc
; GFX8-NEXT:    v_lshrrev_b32_e32 v19, 16, v18
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v1
; GFX8-NEXT:    v_cndmask_b32_e32 v1, v19, v1, vcc
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v17
; GFX8-NEXT:    v_cndmask_b32_e32 v1, v1, v17, vcc
; GFX8-NEXT:    v_and_b32_e32 v17, 0xffff0000, v18
; GFX8-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v17
; GFX8-NEXT:    v_lshlrev_b32_e32 v17, 16, v0
; GFX8-NEXT:    v_cndmask_b32_e32 v1, v19, v1, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX8-NEXT:    v_lshlrev_b32_e32 v17, 16, v16
; GFX8-NEXT:    v_cndmask_b32_e32 v0, v0, v16, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX8-NEXT:    v_cndmask_b32_e32 v16, v16, v0, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v17, 16, v16
; GFX8-NEXT:    v_lshlrev_b32_e32 v18, 16, v0
; GFX8-NEXT:    v_cmp_gt_f32_e32 vcc, v18, v17
; GFX8-NEXT:    v_cndmask_b32_e32 v17, v16, v0, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
; GFX8-NEXT:    v_mul_f32_e32 v17, 1.0, v17
; GFX8-NEXT:    v_bfe_u32 v18, v17, 16, 1
; GFX8-NEXT:    v_add_u32_e32 v18, vcc, v18, v17
; GFX8-NEXT:    v_add_u32_e32 v18, vcc, s4, v18
; GFX8-NEXT:    v_or_b32_e32 v19, 0x400000, v17
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX8-NEXT:    v_cndmask_b32_e32 v17, v18, v19, vcc
; GFX8-NEXT:    v_lshrrev_b32_e32 v18, 16, v17
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v0
; GFX8-NEXT:    v_cndmask_b32_e32 v0, v18, v0, vcc
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v16
; GFX8-NEXT:    v_cndmask_b32_e32 v0, v0, v16, vcc
; GFX8-NEXT:    v_and_b32_e32 v16, 0xffff0000, v17
; GFX8-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v16
; GFX8-NEXT:    v_cndmask_b32_e32 v0, v18, v0, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v16, 16, v54
; GFX8-NEXT:    v_or_b32_sdwa v0, v0, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX8-NEXT:    v_lshlrev_b32_e32 v16, 16, v53
; GFX8-NEXT:    v_or_b32_sdwa v1, v1, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX8-NEXT:    v_lshlrev_b32_e32 v16, 16, v52
; GFX8-NEXT:    v_or_b32_sdwa v2, v2, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX8-NEXT:    v_lshlrev_b32_e32 v16, 16, v51
; GFX8-NEXT:    v_or_b32_sdwa v3, v3, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX8-NEXT:    v_lshlrev_b32_e32 v16, 16, v50
; GFX8-NEXT:    v_or_b32_sdwa v4, v4, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX8-NEXT:    v_lshlrev_b32_e32 v16, 16, v49
; GFX8-NEXT:    v_or_b32_sdwa v5, v5, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX8-NEXT:    v_lshlrev_b32_e32 v16, 16, v48
; GFX8-NEXT:    v_or_b32_sdwa v6, v6, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX8-NEXT:    v_lshlrev_b32_e32 v16, 16, v39
; GFX8-NEXT:    v_or_b32_sdwa v7, v7, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX8-NEXT:    v_lshlrev_b32_e32 v16, 16, v38
; GFX8-NEXT:    v_or_b32_sdwa v8, v8, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX8-NEXT:    v_lshlrev_b32_e32 v16, 16, v37
; GFX8-NEXT:    v_or_b32_sdwa v9, v9, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX8-NEXT:    v_lshlrev_b32_e32 v16, 16, v36
; GFX8-NEXT:    v_or_b32_sdwa v10, v10, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX8-NEXT:    v_lshlrev_b32_e32 v16, 16, v35
; GFX8-NEXT:    v_or_b32_sdwa v11, v11, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX8-NEXT:    v_lshlrev_b32_e32 v16, 16, v34
; GFX8-NEXT:    v_or_b32_sdwa v12, v12, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX8-NEXT:    v_lshlrev_b32_e32 v16, 16, v33
; GFX8-NEXT:    v_or_b32_sdwa v13, v13, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX8-NEXT:    v_lshlrev_b32_e32 v16, 16, v31
; GFX8-NEXT:    v_or_b32_sdwa v14, v14, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX8-NEXT:    v_lshlrev_b32_e32 v16, 16, v32
; GFX8-NEXT:    v_or_b32_sdwa v15, v15, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX8-NEXT:    s_waitcnt vmcnt(0)
; GFX8-NEXT:    s_setpc_b64 s[30:31]
;
; GFX900-LABEL: v_maximumnum_v32bf16:
; GFX900:       ; %bb.0:
; GFX900-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX900-NEXT:    buffer_load_dword v55, off, s[0:3], s32
; GFX900-NEXT:    v_and_b32_e32 v31, 0xffff0000, v14
; GFX900-NEXT:    v_lshrrev_b32_e32 v34, 16, v30
; GFX900-NEXT:    v_lshrrev_b32_e32 v35, 16, v14
; GFX900-NEXT:    v_and_b32_e32 v37, 0xffff0000, v13
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v31, v31
; GFX900-NEXT:    v_and_b32_e32 v36, 0xffff0000, v30
; GFX900-NEXT:    v_lshrrev_b32_e32 v38, 16, v29
; GFX900-NEXT:    v_lshrrev_b32_e32 v39, 16, v13
; GFX900-NEXT:    v_cndmask_b32_e32 v31, v35, v34, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v37, v37
; GFX900-NEXT:    v_cndmask_b32_e32 v35, v39, v38, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v36, v36
; GFX900-NEXT:    v_cndmask_b32_e32 v34, v34, v31, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v37, 16, v31
; GFX900-NEXT:    v_lshlrev_b32_e32 v39, 16, v34
; GFX900-NEXT:    v_cmp_gt_f32_e32 vcc, v37, v39
; GFX900-NEXT:    v_cndmask_b32_e32 v37, v34, v31, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v37, 16, v37
; GFX900-NEXT:    v_max_f32_e32 v37, v37, v37
; GFX900-NEXT:    s_movk_i32 s4, 0x7fff
; GFX900-NEXT:    v_bfe_u32 v39, v37, 16, 1
; GFX900-NEXT:    v_or_b32_e32 v49, 0x400000, v37
; GFX900-NEXT:    v_add3_u32 v39, v39, v37, s4
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v37, v37
; GFX900-NEXT:    v_cndmask_b32_e32 v37, v39, v49, vcc
; GFX900-NEXT:    v_lshrrev_b32_e32 v39, 16, v37
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v31
; GFX900-NEXT:    v_and_b32_e32 v37, 0xffff0000, v37
; GFX900-NEXT:    v_cndmask_b32_e32 v31, v39, v31, vcc
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v34
; GFX900-NEXT:    v_and_b32_e32 v32, 0xffff0000, v15
; GFX900-NEXT:    v_cndmask_b32_e32 v31, v31, v34, vcc
; GFX900-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v37
; GFX900-NEXT:    v_lshrrev_b32_e32 v33, 16, v15
; GFX900-NEXT:    v_cndmask_b32_e32 v31, v39, v31, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX900-NEXT:    v_and_b32_e32 v48, 0xffff0000, v29
; GFX900-NEXT:    v_lshlrev_b32_e32 v36, 16, v35
; GFX900-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
; GFX900-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
; GFX900-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
; GFX900-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
; GFX900-NEXT:    s_waitcnt vmcnt(4)
; GFX900-NEXT:    v_lshrrev_b32_e32 v34, 16, v55
; GFX900-NEXT:    v_and_b32_e32 v37, 0xffff0000, v55
; GFX900-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v37, v37
; GFX900-NEXT:    v_cndmask_b32_e32 v34, v34, v32, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v33, 16, v32
; GFX900-NEXT:    v_lshlrev_b32_e32 v37, 16, v34
; GFX900-NEXT:    v_cmp_gt_f32_e32 vcc, v33, v37
; GFX900-NEXT:    v_cndmask_b32_e32 v33, v34, v32, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v33, 16, v33
; GFX900-NEXT:    v_max_f32_e32 v33, v33, v33
; GFX900-NEXT:    v_bfe_u32 v37, v33, 16, 1
; GFX900-NEXT:    v_or_b32_e32 v39, 0x400000, v33
; GFX900-NEXT:    v_add3_u32 v37, v37, v33, s4
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v33, v33
; GFX900-NEXT:    v_cndmask_b32_e32 v33, v37, v39, vcc
; GFX900-NEXT:    v_lshrrev_b32_e32 v37, 16, v33
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v32
; GFX900-NEXT:    v_and_b32_e32 v33, 0xffff0000, v33
; GFX900-NEXT:    v_cndmask_b32_e32 v32, v37, v32, vcc
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v34
; GFX900-NEXT:    v_cndmask_b32_e32 v32, v32, v34, vcc
; GFX900-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v33
; GFX900-NEXT:    v_cndmask_b32_e32 v32, v37, v32, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v48, v48
; GFX900-NEXT:    v_cndmask_b32_e32 v33, v38, v35, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v34, 16, v33
; GFX900-NEXT:    v_cmp_gt_f32_e32 vcc, v36, v34
; GFX900-NEXT:    v_cndmask_b32_e32 v34, v33, v35, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v34, 16, v34
; GFX900-NEXT:    v_max_f32_e32 v34, v34, v34
; GFX900-NEXT:    v_bfe_u32 v36, v34, 16, 1
; GFX900-NEXT:    v_add3_u32 v36, v36, v34, s4
; GFX900-NEXT:    v_or_b32_e32 v37, 0x400000, v34
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v34, v34
; GFX900-NEXT:    v_cndmask_b32_e32 v34, v36, v37, vcc
; GFX900-NEXT:    v_lshrrev_b32_e32 v36, 16, v34
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v35
; GFX900-NEXT:    v_cndmask_b32_e32 v35, v36, v35, vcc
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v33
; GFX900-NEXT:    v_and_b32_e32 v34, 0xffff0000, v34
; GFX900-NEXT:    v_cndmask_b32_e32 v33, v35, v33, vcc
; GFX900-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v34
; GFX900-NEXT:    v_and_b32_e32 v34, 0xffff0000, v12
; GFX900-NEXT:    v_cndmask_b32_e32 v33, v36, v33, vcc
; GFX900-NEXT:    v_lshrrev_b32_e32 v35, 16, v28
; GFX900-NEXT:    v_lshrrev_b32_e32 v36, 16, v12
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v34, v34
; GFX900-NEXT:    v_and_b32_e32 v37, 0xffff0000, v28
; GFX900-NEXT:    v_cndmask_b32_e32 v34, v36, v35, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v37, v37
; GFX900-NEXT:    v_cndmask_b32_e32 v35, v35, v34, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v36, 16, v34
; GFX900-NEXT:    v_lshlrev_b32_e32 v37, 16, v35
; GFX900-NEXT:    v_cmp_gt_f32_e32 vcc, v36, v37
; GFX900-NEXT:    v_cndmask_b32_e32 v36, v35, v34, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v36, 16, v36
; GFX900-NEXT:    v_max_f32_e32 v36, v36, v36
; GFX900-NEXT:    v_bfe_u32 v37, v36, 16, 1
; GFX900-NEXT:    v_add3_u32 v37, v37, v36, s4
; GFX900-NEXT:    v_or_b32_e32 v38, 0x400000, v36
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v36, v36
; GFX900-NEXT:    v_cndmask_b32_e32 v36, v37, v38, vcc
; GFX900-NEXT:    v_lshrrev_b32_e32 v37, 16, v36
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v34
; GFX900-NEXT:    v_cndmask_b32_e32 v34, v37, v34, vcc
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v35
; GFX900-NEXT:    v_cndmask_b32_e32 v34, v34, v35, vcc
; GFX900-NEXT:    v_and_b32_e32 v35, 0xffff0000, v36
; GFX900-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v35
; GFX900-NEXT:    v_and_b32_e32 v35, 0xffff0000, v11
; GFX900-NEXT:    v_cndmask_b32_e32 v34, v37, v34, vcc
; GFX900-NEXT:    v_lshrrev_b32_e32 v36, 16, v27
; GFX900-NEXT:    v_lshrrev_b32_e32 v37, 16, v11
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v35, v35
; GFX900-NEXT:    v_and_b32_e32 v38, 0xffff0000, v27
; GFX900-NEXT:    v_cndmask_b32_e32 v35, v37, v36, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v38, v38
; GFX900-NEXT:    v_cndmask_b32_e32 v36, v36, v35, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v37, 16, v35
; GFX900-NEXT:    v_lshlrev_b32_e32 v38, 16, v36
; GFX900-NEXT:    v_cmp_gt_f32_e32 vcc, v37, v38
; GFX900-NEXT:    v_cndmask_b32_e32 v37, v36, v35, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v37, 16, v37
; GFX900-NEXT:    v_max_f32_e32 v37, v37, v37
; GFX900-NEXT:    v_bfe_u32 v38, v37, 16, 1
; GFX900-NEXT:    v_add3_u32 v38, v38, v37, s4
; GFX900-NEXT:    v_or_b32_e32 v39, 0x400000, v37
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v37, v37
; GFX900-NEXT:    v_cndmask_b32_e32 v37, v38, v39, vcc
; GFX900-NEXT:    v_lshrrev_b32_e32 v38, 16, v37
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v35
; GFX900-NEXT:    v_cndmask_b32_e32 v35, v38, v35, vcc
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v36
; GFX900-NEXT:    v_cndmask_b32_e32 v35, v35, v36, vcc
; GFX900-NEXT:    v_and_b32_e32 v36, 0xffff0000, v37
; GFX900-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v36
; GFX900-NEXT:    v_and_b32_e32 v36, 0xffff0000, v10
; GFX900-NEXT:    v_cndmask_b32_e32 v35, v38, v35, vcc
; GFX900-NEXT:    v_lshrrev_b32_e32 v37, 16, v26
; GFX900-NEXT:    v_lshrrev_b32_e32 v38, 16, v10
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v36, v36
; GFX900-NEXT:    v_and_b32_e32 v39, 0xffff0000, v26
; GFX900-NEXT:    v_cndmask_b32_e32 v36, v38, v37, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v39, v39
; GFX900-NEXT:    v_cndmask_b32_e32 v37, v37, v36, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v38, 16, v36
; GFX900-NEXT:    v_lshlrev_b32_e32 v39, 16, v37
; GFX900-NEXT:    v_cmp_gt_f32_e32 vcc, v38, v39
; GFX900-NEXT:    v_cndmask_b32_e32 v38, v37, v36, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v38, 16, v38
; GFX900-NEXT:    v_max_f32_e32 v38, v38, v38
; GFX900-NEXT:    v_bfe_u32 v39, v38, 16, 1
; GFX900-NEXT:    v_add3_u32 v39, v39, v38, s4
; GFX900-NEXT:    v_or_b32_e32 v48, 0x400000, v38
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v38, v38
; GFX900-NEXT:    v_cndmask_b32_e32 v38, v39, v48, vcc
; GFX900-NEXT:    v_lshrrev_b32_e32 v39, 16, v38
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v36
; GFX900-NEXT:    v_cndmask_b32_e32 v36, v39, v36, vcc
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v37
; GFX900-NEXT:    v_cndmask_b32_e32 v36, v36, v37, vcc
; GFX900-NEXT:    v_and_b32_e32 v37, 0xffff0000, v38
; GFX900-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v37
; GFX900-NEXT:    v_and_b32_e32 v37, 0xffff0000, v9
; GFX900-NEXT:    v_cndmask_b32_e32 v36, v39, v36, vcc
; GFX900-NEXT:    v_lshrrev_b32_e32 v38, 16, v25
; GFX900-NEXT:    v_lshrrev_b32_e32 v39, 16, v9
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v37, v37
; GFX900-NEXT:    v_and_b32_e32 v48, 0xffff0000, v25
; GFX900-NEXT:    v_cndmask_b32_e32 v37, v39, v38, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v48, v48
; GFX900-NEXT:    v_cndmask_b32_e32 v38, v38, v37, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v39, 16, v37
; GFX900-NEXT:    v_lshlrev_b32_e32 v48, 16, v38
; GFX900-NEXT:    v_cmp_gt_f32_e32 vcc, v39, v48
; GFX900-NEXT:    v_cndmask_b32_e32 v39, v38, v37, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v39, 16, v39
; GFX900-NEXT:    v_max_f32_e32 v39, v39, v39
; GFX900-NEXT:    v_bfe_u32 v48, v39, 16, 1
; GFX900-NEXT:    v_add3_u32 v48, v48, v39, s4
; GFX900-NEXT:    v_or_b32_e32 v49, 0x400000, v39
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v39, v39
; GFX900-NEXT:    v_cndmask_b32_e32 v39, v48, v49, vcc
; GFX900-NEXT:    v_lshrrev_b32_e32 v48, 16, v39
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v37
; GFX900-NEXT:    v_cndmask_b32_e32 v37, v48, v37, vcc
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v38
; GFX900-NEXT:    v_cndmask_b32_e32 v37, v37, v38, vcc
; GFX900-NEXT:    v_and_b32_e32 v38, 0xffff0000, v39
; GFX900-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v38
; GFX900-NEXT:    v_and_b32_e32 v38, 0xffff0000, v8
; GFX900-NEXT:    v_cndmask_b32_e32 v37, v48, v37, vcc
; GFX900-NEXT:    v_lshrrev_b32_e32 v39, 16, v24
; GFX900-NEXT:    v_lshrrev_b32_e32 v48, 16, v8
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v38, v38
; GFX900-NEXT:    v_and_b32_e32 v49, 0xffff0000, v24
; GFX900-NEXT:    v_cndmask_b32_e32 v38, v48, v39, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v49, v49
; GFX900-NEXT:    v_cndmask_b32_e32 v39, v39, v38, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v48, 16, v38
; GFX900-NEXT:    v_lshlrev_b32_e32 v49, 16, v39
; GFX900-NEXT:    v_cmp_gt_f32_e32 vcc, v48, v49
; GFX900-NEXT:    v_cndmask_b32_e32 v48, v39, v38, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v48, 16, v48
; GFX900-NEXT:    v_max_f32_e32 v48, v48, v48
; GFX900-NEXT:    v_bfe_u32 v49, v48, 16, 1
; GFX900-NEXT:    v_add3_u32 v49, v49, v48, s4
; GFX900-NEXT:    v_or_b32_e32 v50, 0x400000, v48
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v48, v48
; GFX900-NEXT:    v_cndmask_b32_e32 v48, v49, v50, vcc
; GFX900-NEXT:    v_lshrrev_b32_e32 v49, 16, v48
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v38
; GFX900-NEXT:    v_cndmask_b32_e32 v38, v49, v38, vcc
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v39
; GFX900-NEXT:    v_cndmask_b32_e32 v38, v38, v39, vcc
; GFX900-NEXT:    v_and_b32_e32 v39, 0xffff0000, v48
; GFX900-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v39
; GFX900-NEXT:    v_and_b32_e32 v39, 0xffff0000, v7
; GFX900-NEXT:    v_cndmask_b32_e32 v38, v49, v38, vcc
; GFX900-NEXT:    v_lshrrev_b32_e32 v48, 16, v23
; GFX900-NEXT:    v_lshrrev_b32_e32 v49, 16, v7
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v39, v39
; GFX900-NEXT:    v_and_b32_e32 v50, 0xffff0000, v23
; GFX900-NEXT:    v_cndmask_b32_e32 v39, v49, v48, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v50, v50
; GFX900-NEXT:    v_cndmask_b32_e32 v48, v48, v39, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v49, 16, v39
; GFX900-NEXT:    v_lshlrev_b32_e32 v50, 16, v48
; GFX900-NEXT:    v_cmp_gt_f32_e32 vcc, v49, v50
; GFX900-NEXT:    v_cndmask_b32_e32 v49, v48, v39, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v49, 16, v49
; GFX900-NEXT:    v_max_f32_e32 v49, v49, v49
; GFX900-NEXT:    v_bfe_u32 v50, v49, 16, 1
; GFX900-NEXT:    v_add3_u32 v50, v50, v49, s4
; GFX900-NEXT:    v_or_b32_e32 v51, 0x400000, v49
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v49, v49
; GFX900-NEXT:    v_cndmask_b32_e32 v49, v50, v51, vcc
; GFX900-NEXT:    v_lshrrev_b32_e32 v50, 16, v49
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v39
; GFX900-NEXT:    v_cndmask_b32_e32 v39, v50, v39, vcc
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v48
; GFX900-NEXT:    v_cndmask_b32_e32 v39, v39, v48, vcc
; GFX900-NEXT:    v_and_b32_e32 v48, 0xffff0000, v49
; GFX900-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v48
; GFX900-NEXT:    v_and_b32_e32 v48, 0xffff0000, v6
; GFX900-NEXT:    v_cndmask_b32_e32 v39, v50, v39, vcc
; GFX900-NEXT:    v_lshrrev_b32_e32 v49, 16, v22
; GFX900-NEXT:    v_lshrrev_b32_e32 v50, 16, v6
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v48, v48
; GFX900-NEXT:    v_and_b32_e32 v51, 0xffff0000, v22
; GFX900-NEXT:    v_cndmask_b32_e32 v48, v50, v49, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v51, v51
; GFX900-NEXT:    v_cndmask_b32_e32 v49, v49, v48, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v50, 16, v48
; GFX900-NEXT:    v_lshlrev_b32_e32 v51, 16, v49
; GFX900-NEXT:    v_cmp_gt_f32_e32 vcc, v50, v51
; GFX900-NEXT:    v_cndmask_b32_e32 v50, v49, v48, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v50, 16, v50
; GFX900-NEXT:    v_max_f32_e32 v50, v50, v50
; GFX900-NEXT:    v_bfe_u32 v51, v50, 16, 1
; GFX900-NEXT:    v_add3_u32 v51, v51, v50, s4
; GFX900-NEXT:    v_or_b32_e32 v52, 0x400000, v50
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v50, v50
; GFX900-NEXT:    v_cndmask_b32_e32 v50, v51, v52, vcc
; GFX900-NEXT:    v_lshrrev_b32_e32 v51, 16, v50
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v48
; GFX900-NEXT:    v_cndmask_b32_e32 v48, v51, v48, vcc
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v49
; GFX900-NEXT:    v_cndmask_b32_e32 v48, v48, v49, vcc
; GFX900-NEXT:    v_and_b32_e32 v49, 0xffff0000, v50
; GFX900-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v49
; GFX900-NEXT:    v_and_b32_e32 v49, 0xffff0000, v5
; GFX900-NEXT:    v_cndmask_b32_e32 v48, v51, v48, vcc
; GFX900-NEXT:    v_lshrrev_b32_e32 v50, 16, v21
; GFX900-NEXT:    v_lshrrev_b32_e32 v51, 16, v5
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v49, v49
; GFX900-NEXT:    v_and_b32_e32 v52, 0xffff0000, v21
; GFX900-NEXT:    v_cndmask_b32_e32 v49, v51, v50, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v52, v52
; GFX900-NEXT:    v_cndmask_b32_e32 v50, v50, v49, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v51, 16, v49
; GFX900-NEXT:    v_lshlrev_b32_e32 v52, 16, v50
; GFX900-NEXT:    v_cmp_gt_f32_e32 vcc, v51, v52
; GFX900-NEXT:    v_cndmask_b32_e32 v51, v50, v49, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v51, 16, v51
; GFX900-NEXT:    v_max_f32_e32 v51, v51, v51
; GFX900-NEXT:    v_bfe_u32 v52, v51, 16, 1
; GFX900-NEXT:    v_add3_u32 v52, v52, v51, s4
; GFX900-NEXT:    v_or_b32_e32 v53, 0x400000, v51
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v51, v51
; GFX900-NEXT:    v_cndmask_b32_e32 v51, v52, v53, vcc
; GFX900-NEXT:    v_lshrrev_b32_e32 v52, 16, v51
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v49
; GFX900-NEXT:    v_cndmask_b32_e32 v49, v52, v49, vcc
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v50
; GFX900-NEXT:    v_cndmask_b32_e32 v49, v49, v50, vcc
; GFX900-NEXT:    v_and_b32_e32 v50, 0xffff0000, v51
; GFX900-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v50
; GFX900-NEXT:    v_and_b32_e32 v50, 0xffff0000, v4
; GFX900-NEXT:    v_cndmask_b32_e32 v49, v52, v49, vcc
; GFX900-NEXT:    v_lshrrev_b32_e32 v51, 16, v20
; GFX900-NEXT:    v_lshrrev_b32_e32 v52, 16, v4
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v50, v50
; GFX900-NEXT:    v_and_b32_e32 v53, 0xffff0000, v20
; GFX900-NEXT:    v_cndmask_b32_e32 v50, v52, v51, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v53, v53
; GFX900-NEXT:    v_cndmask_b32_e32 v51, v51, v50, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v52, 16, v50
; GFX900-NEXT:    v_lshlrev_b32_e32 v53, 16, v51
; GFX900-NEXT:    v_cmp_gt_f32_e32 vcc, v52, v53
; GFX900-NEXT:    v_cndmask_b32_e32 v52, v51, v50, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v52, 16, v52
; GFX900-NEXT:    v_max_f32_e32 v52, v52, v52
; GFX900-NEXT:    v_bfe_u32 v53, v52, 16, 1
; GFX900-NEXT:    v_add3_u32 v53, v53, v52, s4
; GFX900-NEXT:    v_or_b32_e32 v54, 0x400000, v52
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v52, v52
; GFX900-NEXT:    v_cndmask_b32_e32 v52, v53, v54, vcc
; GFX900-NEXT:    v_lshrrev_b32_e32 v53, 16, v52
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v50
; GFX900-NEXT:    v_cndmask_b32_e32 v50, v53, v50, vcc
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v51
; GFX900-NEXT:    v_cndmask_b32_e32 v50, v50, v51, vcc
; GFX900-NEXT:    v_and_b32_e32 v51, 0xffff0000, v52
; GFX900-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v51
; GFX900-NEXT:    v_and_b32_e32 v51, 0xffff0000, v3
; GFX900-NEXT:    v_cndmask_b32_e32 v50, v53, v50, vcc
; GFX900-NEXT:    v_lshrrev_b32_e32 v52, 16, v19
; GFX900-NEXT:    v_lshrrev_b32_e32 v53, 16, v3
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v51, v51
; GFX900-NEXT:    v_and_b32_e32 v54, 0xffff0000, v19
; GFX900-NEXT:    v_cndmask_b32_e32 v51, v53, v52, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v54, v54
; GFX900-NEXT:    v_cndmask_b32_e32 v52, v52, v51, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v53, 16, v51
; GFX900-NEXT:    v_lshlrev_b32_e32 v54, 16, v52
; GFX900-NEXT:    v_cmp_gt_f32_e32 vcc, v53, v54
; GFX900-NEXT:    v_cndmask_b32_e32 v53, v52, v51, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v53, 16, v53
; GFX900-NEXT:    v_max_f32_e32 v53, v53, v53
; GFX900-NEXT:    v_bfe_u32 v54, v53, 16, 1
; GFX900-NEXT:    v_add3_u32 v54, v54, v53, s4
; GFX900-NEXT:    v_or_b32_e32 v40, 0x400000, v53
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v53, v53
; GFX900-NEXT:    v_cndmask_b32_e32 v53, v54, v40, vcc
; GFX900-NEXT:    v_lshrrev_b32_e32 v54, 16, v53
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v51
; GFX900-NEXT:    v_cndmask_b32_e32 v51, v54, v51, vcc
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v52
; GFX900-NEXT:    v_cndmask_b32_e32 v51, v51, v52, vcc
; GFX900-NEXT:    v_and_b32_e32 v52, 0xffff0000, v53
; GFX900-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v52
; GFX900-NEXT:    v_and_b32_e32 v52, 0xffff0000, v2
; GFX900-NEXT:    v_cndmask_b32_e32 v51, v54, v51, vcc
; GFX900-NEXT:    v_lshrrev_b32_e32 v53, 16, v18
; GFX900-NEXT:    v_lshrrev_b32_e32 v54, 16, v2
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v52, v52
; GFX900-NEXT:    v_and_b32_e32 v40, 0xffff0000, v18
; GFX900-NEXT:    v_cndmask_b32_e32 v52, v54, v53, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v40, v40
; GFX900-NEXT:    v_cndmask_b32_e32 v53, v53, v52, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v54, 16, v52
; GFX900-NEXT:    v_lshlrev_b32_e32 v40, 16, v53
; GFX900-NEXT:    v_cmp_gt_f32_e32 vcc, v54, v40
; GFX900-NEXT:    v_cndmask_b32_e32 v54, v53, v52, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v54, 16, v54
; GFX900-NEXT:    v_max_f32_e32 v54, v54, v54
; GFX900-NEXT:    v_bfe_u32 v40, v54, 16, 1
; GFX900-NEXT:    v_add3_u32 v40, v40, v54, s4
; GFX900-NEXT:    v_or_b32_e32 v41, 0x400000, v54
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v54, v54
; GFX900-NEXT:    v_cndmask_b32_e32 v54, v40, v41, vcc
; GFX900-NEXT:    v_lshrrev_b32_e32 v40, 16, v54
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v52
; GFX900-NEXT:    v_cndmask_b32_e32 v52, v40, v52, vcc
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v53
; GFX900-NEXT:    v_cndmask_b32_e32 v52, v52, v53, vcc
; GFX900-NEXT:    v_and_b32_e32 v53, 0xffff0000, v54
; GFX900-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v53
; GFX900-NEXT:    v_and_b32_e32 v53, 0xffff0000, v1
; GFX900-NEXT:    v_cndmask_b32_e32 v52, v40, v52, vcc
; GFX900-NEXT:    v_lshrrev_b32_e32 v54, 16, v17
; GFX900-NEXT:    v_lshrrev_b32_e32 v40, 16, v1
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v53, v53
; GFX900-NEXT:    v_and_b32_e32 v41, 0xffff0000, v17
; GFX900-NEXT:    v_cndmask_b32_e32 v53, v40, v54, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v41, v41
; GFX900-NEXT:    v_cndmask_b32_e32 v54, v54, v53, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v40, 16, v53
; GFX900-NEXT:    v_lshlrev_b32_e32 v41, 16, v54
; GFX900-NEXT:    v_cmp_gt_f32_e32 vcc, v40, v41
; GFX900-NEXT:    v_cndmask_b32_e32 v40, v54, v53, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v40, 16, v40
; GFX900-NEXT:    v_max_f32_e32 v40, v40, v40
; GFX900-NEXT:    v_bfe_u32 v41, v40, 16, 1
; GFX900-NEXT:    v_add3_u32 v41, v41, v40, s4
; GFX900-NEXT:    v_or_b32_e32 v42, 0x400000, v40
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v40, v40
; GFX900-NEXT:    v_cndmask_b32_e32 v40, v41, v42, vcc
; GFX900-NEXT:    v_lshrrev_b32_e32 v41, 16, v40
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v53
; GFX900-NEXT:    v_cndmask_b32_e32 v53, v41, v53, vcc
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v54
; GFX900-NEXT:    v_cndmask_b32_e32 v53, v53, v54, vcc
; GFX900-NEXT:    v_and_b32_e32 v54, 0xffff0000, v40
; GFX900-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v54
; GFX900-NEXT:    v_and_b32_e32 v54, 0xffff0000, v0
; GFX900-NEXT:    v_cndmask_b32_e32 v53, v41, v53, vcc
; GFX900-NEXT:    v_lshrrev_b32_e32 v40, 16, v16
; GFX900-NEXT:    v_lshrrev_b32_e32 v41, 16, v0
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v54, v54
; GFX900-NEXT:    v_and_b32_e32 v42, 0xffff0000, v16
; GFX900-NEXT:    v_cndmask_b32_e32 v54, v41, v40, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v42, v42
; GFX900-NEXT:    v_cndmask_b32_e32 v40, v40, v54, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v41, 16, v54
; GFX900-NEXT:    v_lshlrev_b32_e32 v42, 16, v40
; GFX900-NEXT:    v_cmp_gt_f32_e32 vcc, v41, v42
; GFX900-NEXT:    v_cndmask_b32_e32 v41, v40, v54, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v41, 16, v41
; GFX900-NEXT:    v_max_f32_e32 v41, v41, v41
; GFX900-NEXT:    v_bfe_u32 v42, v41, 16, 1
; GFX900-NEXT:    v_add3_u32 v42, v42, v41, s4
; GFX900-NEXT:    v_or_b32_e32 v43, 0x400000, v41
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v41, v41
; GFX900-NEXT:    v_cndmask_b32_e32 v41, v42, v43, vcc
; GFX900-NEXT:    v_lshrrev_b32_e32 v42, 16, v41
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v54
; GFX900-NEXT:    v_cndmask_b32_e32 v54, v42, v54, vcc
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v40
; GFX900-NEXT:    v_cndmask_b32_e32 v54, v54, v40, vcc
; GFX900-NEXT:    v_and_b32_e32 v40, 0xffff0000, v41
; GFX900-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v40
; GFX900-NEXT:    v_lshlrev_b32_e32 v40, 16, v15
; GFX900-NEXT:    v_cndmask_b32_e32 v54, v42, v54, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v40, v40
; GFX900-NEXT:    v_lshlrev_b32_e32 v40, 16, v55
; GFX900-NEXT:    v_cndmask_b32_e32 v15, v15, v55, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v40, v40
; GFX900-NEXT:    v_cndmask_b32_e32 v55, v55, v15, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v40, 16, v55
; GFX900-NEXT:    v_lshlrev_b32_e32 v41, 16, v15
; GFX900-NEXT:    v_cmp_gt_f32_e32 vcc, v41, v40
; GFX900-NEXT:    v_cndmask_b32_e32 v40, v55, v15, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v40, 16, v40
; GFX900-NEXT:    v_max_f32_e32 v40, v40, v40
; GFX900-NEXT:    v_bfe_u32 v41, v40, 16, 1
; GFX900-NEXT:    v_add3_u32 v41, v41, v40, s4
; GFX900-NEXT:    v_or_b32_e32 v42, 0x400000, v40
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v40, v40
; GFX900-NEXT:    v_cndmask_b32_e32 v40, v41, v42, vcc
; GFX900-NEXT:    v_lshrrev_b32_e32 v41, 16, v40
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v15
; GFX900-NEXT:    v_cndmask_b32_e32 v15, v41, v15, vcc
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v55
; GFX900-NEXT:    v_cndmask_b32_e32 v15, v15, v55, vcc
; GFX900-NEXT:    v_and_b32_e32 v55, 0xffff0000, v40
; GFX900-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v55
; GFX900-NEXT:    v_lshlrev_b32_e32 v55, 16, v14
; GFX900-NEXT:    v_cndmask_b32_e32 v15, v41, v15, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v55, v55
; GFX900-NEXT:    v_lshlrev_b32_e32 v55, 16, v30
; GFX900-NEXT:    v_cndmask_b32_e32 v14, v14, v30, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v55, v55
; GFX900-NEXT:    v_cndmask_b32_e32 v30, v30, v14, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v55, 16, v30
; GFX900-NEXT:    v_lshlrev_b32_e32 v40, 16, v14
; GFX900-NEXT:    v_cmp_gt_f32_e32 vcc, v40, v55
; GFX900-NEXT:    v_cndmask_b32_e32 v55, v30, v14, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v55, 16, v55
; GFX900-NEXT:    v_max_f32_e32 v55, v55, v55
; GFX900-NEXT:    v_bfe_u32 v40, v55, 16, 1
; GFX900-NEXT:    v_add3_u32 v40, v40, v55, s4
; GFX900-NEXT:    v_or_b32_e32 v41, 0x400000, v55
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v55, v55
; GFX900-NEXT:    v_cndmask_b32_e32 v55, v40, v41, vcc
; GFX900-NEXT:    v_lshrrev_b32_e32 v40, 16, v55
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v14
; GFX900-NEXT:    v_cndmask_b32_e32 v14, v40, v14, vcc
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v30
; GFX900-NEXT:    v_cndmask_b32_e32 v14, v14, v30, vcc
; GFX900-NEXT:    v_and_b32_e32 v30, 0xffff0000, v55
; GFX900-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v30
; GFX900-NEXT:    v_lshlrev_b32_e32 v30, 16, v13
; GFX900-NEXT:    v_cndmask_b32_e32 v14, v40, v14, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v30, v30
; GFX900-NEXT:    v_lshlrev_b32_e32 v30, 16, v29
; GFX900-NEXT:    v_cndmask_b32_e32 v13, v13, v29, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v30, v30
; GFX900-NEXT:    v_cndmask_b32_e32 v29, v29, v13, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v30, 16, v29
; GFX900-NEXT:    v_lshlrev_b32_e32 v55, 16, v13
; GFX900-NEXT:    v_cmp_gt_f32_e32 vcc, v55, v30
; GFX900-NEXT:    v_cndmask_b32_e32 v30, v29, v13, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v30, 16, v30
; GFX900-NEXT:    v_max_f32_e32 v30, v30, v30
; GFX900-NEXT:    v_bfe_u32 v55, v30, 16, 1
; GFX900-NEXT:    v_add3_u32 v55, v55, v30, s4
; GFX900-NEXT:    v_or_b32_e32 v40, 0x400000, v30
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v30, v30
; GFX900-NEXT:    v_cndmask_b32_e32 v30, v55, v40, vcc
; GFX900-NEXT:    v_lshrrev_b32_e32 v55, 16, v30
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v13
; GFX900-NEXT:    v_cndmask_b32_e32 v13, v55, v13, vcc
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v29
; GFX900-NEXT:    v_cndmask_b32_e32 v13, v13, v29, vcc
; GFX900-NEXT:    v_and_b32_e32 v29, 0xffff0000, v30
; GFX900-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v29
; GFX900-NEXT:    v_lshlrev_b32_e32 v29, 16, v12
; GFX900-NEXT:    v_cndmask_b32_e32 v13, v55, v13, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v29, v29
; GFX900-NEXT:    v_lshlrev_b32_e32 v29, 16, v28
; GFX900-NEXT:    v_cndmask_b32_e32 v12, v12, v28, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v29, v29
; GFX900-NEXT:    v_cndmask_b32_e32 v28, v28, v12, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v29, 16, v28
; GFX900-NEXT:    v_lshlrev_b32_e32 v30, 16, v12
; GFX900-NEXT:    v_cmp_gt_f32_e32 vcc, v30, v29
; GFX900-NEXT:    v_cndmask_b32_e32 v29, v28, v12, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v29, 16, v29
; GFX900-NEXT:    v_max_f32_e32 v29, v29, v29
; GFX900-NEXT:    v_bfe_u32 v30, v29, 16, 1
; GFX900-NEXT:    v_add3_u32 v30, v30, v29, s4
; GFX900-NEXT:    v_or_b32_e32 v55, 0x400000, v29
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v29, v29
; GFX900-NEXT:    v_cndmask_b32_e32 v29, v30, v55, vcc
; GFX900-NEXT:    v_lshrrev_b32_e32 v30, 16, v29
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v12
; GFX900-NEXT:    v_cndmask_b32_e32 v12, v30, v12, vcc
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v28
; GFX900-NEXT:    v_cndmask_b32_e32 v12, v12, v28, vcc
; GFX900-NEXT:    v_and_b32_e32 v28, 0xffff0000, v29
; GFX900-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v28
; GFX900-NEXT:    v_lshlrev_b32_e32 v28, 16, v11
; GFX900-NEXT:    v_cndmask_b32_e32 v12, v30, v12, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v28, v28
; GFX900-NEXT:    v_lshlrev_b32_e32 v28, 16, v27
; GFX900-NEXT:    v_cndmask_b32_e32 v11, v11, v27, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v28, v28
; GFX900-NEXT:    v_cndmask_b32_e32 v27, v27, v11, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v28, 16, v27
; GFX900-NEXT:    v_lshlrev_b32_e32 v29, 16, v11
; GFX900-NEXT:    v_cmp_gt_f32_e32 vcc, v29, v28
; GFX900-NEXT:    v_cndmask_b32_e32 v28, v27, v11, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
; GFX900-NEXT:    v_max_f32_e32 v28, v28, v28
; GFX900-NEXT:    v_bfe_u32 v29, v28, 16, 1
; GFX900-NEXT:    v_add3_u32 v29, v29, v28, s4
; GFX900-NEXT:    v_or_b32_e32 v30, 0x400000, v28
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v28, v28
; GFX900-NEXT:    v_cndmask_b32_e32 v28, v29, v30, vcc
; GFX900-NEXT:    v_lshrrev_b32_e32 v29, 16, v28
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v11
; GFX900-NEXT:    v_cndmask_b32_e32 v11, v29, v11, vcc
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v27
; GFX900-NEXT:    v_cndmask_b32_e32 v11, v11, v27, vcc
; GFX900-NEXT:    v_and_b32_e32 v27, 0xffff0000, v28
; GFX900-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v27
; GFX900-NEXT:    v_lshlrev_b32_e32 v27, 16, v10
; GFX900-NEXT:    v_cndmask_b32_e32 v11, v29, v11, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v27, v27
; GFX900-NEXT:    v_lshlrev_b32_e32 v27, 16, v26
; GFX900-NEXT:    v_cndmask_b32_e32 v10, v10, v26, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v27, v27
; GFX900-NEXT:    v_cndmask_b32_e32 v26, v26, v10, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v27, 16, v26
; GFX900-NEXT:    v_lshlrev_b32_e32 v28, 16, v10
; GFX900-NEXT:    v_cmp_gt_f32_e32 vcc, v28, v27
; GFX900-NEXT:    v_cndmask_b32_e32 v27, v26, v10, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v27, 16, v27
; GFX900-NEXT:    v_max_f32_e32 v27, v27, v27
; GFX900-NEXT:    v_bfe_u32 v28, v27, 16, 1
; GFX900-NEXT:    v_add3_u32 v28, v28, v27, s4
; GFX900-NEXT:    v_or_b32_e32 v29, 0x400000, v27
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v27, v27
; GFX900-NEXT:    v_cndmask_b32_e32 v27, v28, v29, vcc
; GFX900-NEXT:    v_lshrrev_b32_e32 v28, 16, v27
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v10
; GFX900-NEXT:    v_cndmask_b32_e32 v10, v28, v10, vcc
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v26
; GFX900-NEXT:    v_cndmask_b32_e32 v10, v10, v26, vcc
; GFX900-NEXT:    v_and_b32_e32 v26, 0xffff0000, v27
; GFX900-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v26
; GFX900-NEXT:    v_lshlrev_b32_e32 v26, 16, v9
; GFX900-NEXT:    v_cndmask_b32_e32 v10, v28, v10, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v26, v26
; GFX900-NEXT:    v_lshlrev_b32_e32 v26, 16, v25
; GFX900-NEXT:    v_cndmask_b32_e32 v9, v9, v25, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v26, v26
; GFX900-NEXT:    v_cndmask_b32_e32 v25, v25, v9, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v26, 16, v25
; GFX900-NEXT:    v_lshlrev_b32_e32 v27, 16, v9
; GFX900-NEXT:    v_cmp_gt_f32_e32 vcc, v27, v26
; GFX900-NEXT:    v_cndmask_b32_e32 v26, v25, v9, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
; GFX900-NEXT:    v_max_f32_e32 v26, v26, v26
; GFX900-NEXT:    v_bfe_u32 v27, v26, 16, 1
; GFX900-NEXT:    v_add3_u32 v27, v27, v26, s4
; GFX900-NEXT:    v_or_b32_e32 v28, 0x400000, v26
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v26, v26
; GFX900-NEXT:    v_cndmask_b32_e32 v26, v27, v28, vcc
; GFX900-NEXT:    v_lshrrev_b32_e32 v27, 16, v26
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v9
; GFX900-NEXT:    v_cndmask_b32_e32 v9, v27, v9, vcc
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v25
; GFX900-NEXT:    v_cndmask_b32_e32 v9, v9, v25, vcc
; GFX900-NEXT:    v_and_b32_e32 v25, 0xffff0000, v26
; GFX900-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v25
; GFX900-NEXT:    v_lshlrev_b32_e32 v25, 16, v8
; GFX900-NEXT:    v_cndmask_b32_e32 v9, v27, v9, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v25, v25
; GFX900-NEXT:    v_lshlrev_b32_e32 v25, 16, v24
; GFX900-NEXT:    v_cndmask_b32_e32 v8, v8, v24, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v25, v25
; GFX900-NEXT:    v_cndmask_b32_e32 v24, v24, v8, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v25, 16, v24
; GFX900-NEXT:    v_lshlrev_b32_e32 v26, 16, v8
; GFX900-NEXT:    v_cmp_gt_f32_e32 vcc, v26, v25
; GFX900-NEXT:    v_cndmask_b32_e32 v25, v24, v8, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v25, 16, v25
; GFX900-NEXT:    v_max_f32_e32 v25, v25, v25
; GFX900-NEXT:    v_bfe_u32 v26, v25, 16, 1
; GFX900-NEXT:    v_add3_u32 v26, v26, v25, s4
; GFX900-NEXT:    v_or_b32_e32 v27, 0x400000, v25
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v25, v25
; GFX900-NEXT:    v_cndmask_b32_e32 v25, v26, v27, vcc
; GFX900-NEXT:    v_lshrrev_b32_e32 v26, 16, v25
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v8
; GFX900-NEXT:    v_cndmask_b32_e32 v8, v26, v8, vcc
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v24
; GFX900-NEXT:    v_cndmask_b32_e32 v8, v8, v24, vcc
; GFX900-NEXT:    v_and_b32_e32 v24, 0xffff0000, v25
; GFX900-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v24
; GFX900-NEXT:    v_lshlrev_b32_e32 v24, 16, v7
; GFX900-NEXT:    v_cndmask_b32_e32 v8, v26, v8, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v24, v24
; GFX900-NEXT:    v_lshlrev_b32_e32 v24, 16, v23
; GFX900-NEXT:    v_cndmask_b32_e32 v7, v7, v23, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v24, v24
; GFX900-NEXT:    v_cndmask_b32_e32 v23, v23, v7, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v24, 16, v23
; GFX900-NEXT:    v_lshlrev_b32_e32 v25, 16, v7
; GFX900-NEXT:    v_cmp_gt_f32_e32 vcc, v25, v24
; GFX900-NEXT:    v_cndmask_b32_e32 v24, v23, v7, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
; GFX900-NEXT:    v_max_f32_e32 v24, v24, v24
; GFX900-NEXT:    v_bfe_u32 v25, v24, 16, 1
; GFX900-NEXT:    v_add3_u32 v25, v25, v24, s4
; GFX900-NEXT:    v_or_b32_e32 v26, 0x400000, v24
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v24, v24
; GFX900-NEXT:    v_cndmask_b32_e32 v24, v25, v26, vcc
; GFX900-NEXT:    v_lshrrev_b32_e32 v25, 16, v24
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v7
; GFX900-NEXT:    v_cndmask_b32_e32 v7, v25, v7, vcc
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v23
; GFX900-NEXT:    v_cndmask_b32_e32 v7, v7, v23, vcc
; GFX900-NEXT:    v_and_b32_e32 v23, 0xffff0000, v24
; GFX900-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v23
; GFX900-NEXT:    v_lshlrev_b32_e32 v23, 16, v6
; GFX900-NEXT:    v_cndmask_b32_e32 v7, v25, v7, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v23, v23
; GFX900-NEXT:    v_lshlrev_b32_e32 v23, 16, v22
; GFX900-NEXT:    v_cndmask_b32_e32 v6, v6, v22, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v23, v23
; GFX900-NEXT:    v_cndmask_b32_e32 v22, v22, v6, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v23, 16, v22
; GFX900-NEXT:    v_lshlrev_b32_e32 v24, 16, v6
; GFX900-NEXT:    v_cmp_gt_f32_e32 vcc, v24, v23
; GFX900-NEXT:    v_cndmask_b32_e32 v23, v22, v6, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; GFX900-NEXT:    v_max_f32_e32 v23, v23, v23
; GFX900-NEXT:    v_bfe_u32 v24, v23, 16, 1
; GFX900-NEXT:    v_add3_u32 v24, v24, v23, s4
; GFX900-NEXT:    v_or_b32_e32 v25, 0x400000, v23
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v23, v23
; GFX900-NEXT:    v_cndmask_b32_e32 v23, v24, v25, vcc
; GFX900-NEXT:    v_lshrrev_b32_e32 v24, 16, v23
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v6
; GFX900-NEXT:    v_cndmask_b32_e32 v6, v24, v6, vcc
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v22
; GFX900-NEXT:    v_cndmask_b32_e32 v6, v6, v22, vcc
; GFX900-NEXT:    v_and_b32_e32 v22, 0xffff0000, v23
; GFX900-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v22
; GFX900-NEXT:    v_lshlrev_b32_e32 v22, 16, v5
; GFX900-NEXT:    v_cndmask_b32_e32 v6, v24, v6, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v22, v22
; GFX900-NEXT:    v_lshlrev_b32_e32 v22, 16, v21
; GFX900-NEXT:    v_cndmask_b32_e32 v5, v5, v21, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v22, v22
; GFX900-NEXT:    v_cndmask_b32_e32 v21, v21, v5, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v22, 16, v21
; GFX900-NEXT:    v_lshlrev_b32_e32 v23, 16, v5
; GFX900-NEXT:    v_cmp_gt_f32_e32 vcc, v23, v22
; GFX900-NEXT:    v_cndmask_b32_e32 v22, v21, v5, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v22, 16, v22
; GFX900-NEXT:    v_max_f32_e32 v22, v22, v22
; GFX900-NEXT:    v_bfe_u32 v23, v22, 16, 1
; GFX900-NEXT:    v_add3_u32 v23, v23, v22, s4
; GFX900-NEXT:    v_or_b32_e32 v24, 0x400000, v22
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v22, v22
; GFX900-NEXT:    v_cndmask_b32_e32 v22, v23, v24, vcc
; GFX900-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
; GFX900-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
; GFX900-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
; GFX900-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
; GFX900-NEXT:    v_lshrrev_b32_e32 v23, 16, v22
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v5
; GFX900-NEXT:    v_cndmask_b32_e32 v5, v23, v5, vcc
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v21
; GFX900-NEXT:    v_cndmask_b32_e32 v5, v5, v21, vcc
; GFX900-NEXT:    v_and_b32_e32 v21, 0xffff0000, v22
; GFX900-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v21
; GFX900-NEXT:    v_lshlrev_b32_e32 v21, 16, v4
; GFX900-NEXT:    v_cndmask_b32_e32 v5, v23, v5, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v21, v21
; GFX900-NEXT:    v_lshlrev_b32_e32 v21, 16, v20
; GFX900-NEXT:    v_cndmask_b32_e32 v4, v4, v20, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v21, v21
; GFX900-NEXT:    v_cndmask_b32_e32 v20, v20, v4, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v21, 16, v20
; GFX900-NEXT:    v_lshlrev_b32_e32 v22, 16, v4
; GFX900-NEXT:    v_cmp_gt_f32_e32 vcc, v22, v21
; GFX900-NEXT:    v_cndmask_b32_e32 v21, v20, v4, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
; GFX900-NEXT:    v_max_f32_e32 v21, v21, v21
; GFX900-NEXT:    v_bfe_u32 v22, v21, 16, 1
; GFX900-NEXT:    v_add3_u32 v22, v22, v21, s4
; GFX900-NEXT:    v_or_b32_e32 v23, 0x400000, v21
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v21, v21
; GFX900-NEXT:    v_cndmask_b32_e32 v21, v22, v23, vcc
; GFX900-NEXT:    v_lshrrev_b32_e32 v22, 16, v21
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v4
; GFX900-NEXT:    v_cndmask_b32_e32 v4, v22, v4, vcc
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v20
; GFX900-NEXT:    v_cndmask_b32_e32 v4, v4, v20, vcc
; GFX900-NEXT:    v_and_b32_e32 v20, 0xffff0000, v21
; GFX900-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v20
; GFX900-NEXT:    v_lshlrev_b32_e32 v20, 16, v3
; GFX900-NEXT:    v_cndmask_b32_e32 v4, v22, v4, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v20, v20
; GFX900-NEXT:    v_lshlrev_b32_e32 v20, 16, v19
; GFX900-NEXT:    v_cndmask_b32_e32 v3, v3, v19, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v20, v20
; GFX900-NEXT:    v_cndmask_b32_e32 v19, v19, v3, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v20, 16, v19
; GFX900-NEXT:    v_lshlrev_b32_e32 v21, 16, v3
; GFX900-NEXT:    v_cmp_gt_f32_e32 vcc, v21, v20
; GFX900-NEXT:    v_cndmask_b32_e32 v20, v19, v3, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
; GFX900-NEXT:    v_max_f32_e32 v20, v20, v20
; GFX900-NEXT:    v_bfe_u32 v21, v20, 16, 1
; GFX900-NEXT:    v_add3_u32 v21, v21, v20, s4
; GFX900-NEXT:    v_or_b32_e32 v22, 0x400000, v20
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v20, v20
; GFX900-NEXT:    v_cndmask_b32_e32 v20, v21, v22, vcc
; GFX900-NEXT:    v_lshrrev_b32_e32 v21, 16, v20
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v3
; GFX900-NEXT:    v_cndmask_b32_e32 v3, v21, v3, vcc
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v19
; GFX900-NEXT:    v_cndmask_b32_e32 v3, v3, v19, vcc
; GFX900-NEXT:    v_and_b32_e32 v19, 0xffff0000, v20
; GFX900-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v19
; GFX900-NEXT:    v_lshlrev_b32_e32 v19, 16, v2
; GFX900-NEXT:    v_cndmask_b32_e32 v3, v21, v3, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v19, v19
; GFX900-NEXT:    v_lshlrev_b32_e32 v19, 16, v18
; GFX900-NEXT:    v_cndmask_b32_e32 v2, v2, v18, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v19, v19
; GFX900-NEXT:    v_cndmask_b32_e32 v18, v18, v2, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v19, 16, v18
; GFX900-NEXT:    v_lshlrev_b32_e32 v20, 16, v2
; GFX900-NEXT:    v_cmp_gt_f32_e32 vcc, v20, v19
; GFX900-NEXT:    v_cndmask_b32_e32 v19, v18, v2, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
; GFX900-NEXT:    v_max_f32_e32 v19, v19, v19
; GFX900-NEXT:    v_bfe_u32 v20, v19, 16, 1
; GFX900-NEXT:    v_add3_u32 v20, v20, v19, s4
; GFX900-NEXT:    v_or_b32_e32 v21, 0x400000, v19
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v19, v19
; GFX900-NEXT:    v_cndmask_b32_e32 v19, v20, v21, vcc
; GFX900-NEXT:    v_lshrrev_b32_e32 v20, 16, v19
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v2
; GFX900-NEXT:    v_cndmask_b32_e32 v2, v20, v2, vcc
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v18
; GFX900-NEXT:    v_cndmask_b32_e32 v2, v2, v18, vcc
; GFX900-NEXT:    v_and_b32_e32 v18, 0xffff0000, v19
; GFX900-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v18
; GFX900-NEXT:    v_lshlrev_b32_e32 v18, 16, v1
; GFX900-NEXT:    v_cndmask_b32_e32 v2, v20, v2, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; GFX900-NEXT:    v_lshlrev_b32_e32 v18, 16, v17
; GFX900-NEXT:    v_cndmask_b32_e32 v1, v1, v17, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; GFX900-NEXT:    v_cndmask_b32_e32 v17, v17, v1, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v18, 16, v17
; GFX900-NEXT:    v_lshlrev_b32_e32 v19, 16, v1
; GFX900-NEXT:    v_cmp_gt_f32_e32 vcc, v19, v18
; GFX900-NEXT:    v_cndmask_b32_e32 v18, v17, v1, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
; GFX900-NEXT:    v_max_f32_e32 v18, v18, v18
; GFX900-NEXT:    v_bfe_u32 v19, v18, 16, 1
; GFX900-NEXT:    v_add3_u32 v19, v19, v18, s4
; GFX900-NEXT:    v_or_b32_e32 v20, 0x400000, v18
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; GFX900-NEXT:    v_cndmask_b32_e32 v18, v19, v20, vcc
; GFX900-NEXT:    v_lshrrev_b32_e32 v19, 16, v18
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v1
; GFX900-NEXT:    v_cndmask_b32_e32 v1, v19, v1, vcc
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v17
; GFX900-NEXT:    v_cndmask_b32_e32 v1, v1, v17, vcc
; GFX900-NEXT:    v_and_b32_e32 v17, 0xffff0000, v18
; GFX900-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v17
; GFX900-NEXT:    v_lshlrev_b32_e32 v17, 16, v0
; GFX900-NEXT:    v_cndmask_b32_e32 v1, v19, v1, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX900-NEXT:    v_lshlrev_b32_e32 v17, 16, v16
; GFX900-NEXT:    v_cndmask_b32_e32 v0, v0, v16, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX900-NEXT:    v_cndmask_b32_e32 v16, v16, v0, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v17, 16, v16
; GFX900-NEXT:    v_lshlrev_b32_e32 v18, 16, v0
; GFX900-NEXT:    v_cmp_gt_f32_e32 vcc, v18, v17
; GFX900-NEXT:    v_cndmask_b32_e32 v17, v16, v0, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
; GFX900-NEXT:    v_max_f32_e32 v17, v17, v17
; GFX900-NEXT:    v_bfe_u32 v18, v17, 16, 1
; GFX900-NEXT:    v_add3_u32 v18, v18, v17, s4
; GFX900-NEXT:    v_or_b32_e32 v19, 0x400000, v17
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX900-NEXT:    v_cndmask_b32_e32 v17, v18, v19, vcc
; GFX900-NEXT:    v_lshrrev_b32_e32 v18, 16, v17
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v0
; GFX900-NEXT:    v_cndmask_b32_e32 v0, v18, v0, vcc
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v16
; GFX900-NEXT:    v_cndmask_b32_e32 v0, v0, v16, vcc
; GFX900-NEXT:    v_and_b32_e32 v16, 0xffff0000, v17
; GFX900-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v16
; GFX900-NEXT:    v_cndmask_b32_e32 v0, v18, v0, vcc
; GFX900-NEXT:    s_mov_b32 s4, 0x5040100
; GFX900-NEXT:    v_perm_b32 v0, v54, v0, s4
; GFX900-NEXT:    v_perm_b32 v1, v53, v1, s4
; GFX900-NEXT:    v_perm_b32 v2, v52, v2, s4
; GFX900-NEXT:    v_perm_b32 v3, v51, v3, s4
; GFX900-NEXT:    v_perm_b32 v4, v50, v4, s4
; GFX900-NEXT:    v_perm_b32 v5, v49, v5, s4
; GFX900-NEXT:    v_perm_b32 v6, v48, v6, s4
; GFX900-NEXT:    v_perm_b32 v7, v39, v7, s4
; GFX900-NEXT:    v_perm_b32 v8, v38, v8, s4
; GFX900-NEXT:    v_perm_b32 v9, v37, v9, s4
; GFX900-NEXT:    v_perm_b32 v10, v36, v10, s4
; GFX900-NEXT:    v_perm_b32 v11, v35, v11, s4
; GFX900-NEXT:    v_perm_b32 v12, v34, v12, s4
; GFX900-NEXT:    v_perm_b32 v13, v33, v13, s4
; GFX900-NEXT:    v_perm_b32 v14, v31, v14, s4
; GFX900-NEXT:    v_perm_b32 v15, v32, v15, s4
; GFX900-NEXT:    s_waitcnt vmcnt(0)
; GFX900-NEXT:    s_setpc_b64 s[30:31]
;
; GFX950-LABEL: v_maximumnum_v32bf16:
; GFX950:       ; %bb.0:
; GFX950-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX950-NEXT:    scratch_load_dword v51, off, s32
; GFX950-NEXT:    v_and_b32_e32 v31, 0xffff0000, v14
; GFX950-NEXT:    v_lshrrev_b32_e32 v34, 16, v30
; GFX950-NEXT:    v_lshrrev_b32_e32 v35, 16, v14
; GFX950-NEXT:    v_and_b32_e32 v37, 0xffff0000, v13
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v31, v31
; GFX950-NEXT:    v_and_b32_e32 v36, 0xffff0000, v30
; GFX950-NEXT:    v_lshrrev_b32_e32 v38, 16, v29
; GFX950-NEXT:    v_lshrrev_b32_e32 v39, 16, v13
; GFX950-NEXT:    v_cndmask_b32_e32 v31, v35, v34, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v37, v37
; GFX950-NEXT:    v_and_b32_e32 v48, 0xffff0000, v29
; GFX950-NEXT:    v_lshlrev_b32_e32 v37, 16, v31
; GFX950-NEXT:    v_cndmask_b32_e32 v35, v39, v38, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v36, v36
; GFX950-NEXT:    v_lshlrev_b32_e32 v36, 16, v35
; GFX950-NEXT:    v_and_b32_e32 v32, 0xffff0000, v15
; GFX950-NEXT:    v_cndmask_b32_e32 v34, v34, v31, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v48, v48
; GFX950-NEXT:    v_lshlrev_b32_e32 v39, 16, v34
; GFX950-NEXT:    v_lshrrev_b32_e32 v33, 16, v15
; GFX950-NEXT:    v_cndmask_b32_e32 v38, v38, v35, vcc
; GFX950-NEXT:    v_cmp_gt_f32_e32 vcc, v37, v39
; GFX950-NEXT:    v_lshlrev_b32_e32 v48, 16, v38
; GFX950-NEXT:    v_and_b32_e32 v49, 0xffff0000, v24
; GFX950-NEXT:    v_cndmask_b32_e32 v37, v34, v31, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v37, 16, v37
; GFX950-NEXT:    v_cmp_gt_f32_e32 vcc, v36, v48
; GFX950-NEXT:    v_max_f32_e32 v37, v37, v37
; GFX950-NEXT:    v_cvt_pk_bf16_f32 v37, v37, s0
; GFX950-NEXT:    v_cndmask_b32_e32 v36, v38, v35, vcc
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v31
; GFX950-NEXT:    v_lshlrev_b32_e32 v39, 16, v37
; GFX950-NEXT:    v_lshlrev_b32_e32 v36, 16, v36
; GFX950-NEXT:    v_cndmask_b32_e32 v31, v37, v31, vcc
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v34
; GFX950-NEXT:    v_max_f32_e32 v36, v36, v36
; GFX950-NEXT:    v_cvt_pk_bf16_f32 v36, v36, s0
; GFX950-NEXT:    v_cndmask_b32_e32 v31, v31, v34, vcc
; GFX950-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v39
; GFX950-NEXT:    v_and_b32_e32 v39, 0xffff0000, v26
; GFX950-NEXT:    v_and_b32_e32 v48, 0xffff0000, v25
; GFX950-NEXT:    v_cndmask_b32_e32 v31, v37, v31, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX950-NEXT:    v_and_b32_e32 v50, 0xffff0000, v23
; GFX950-NEXT:    v_and_b32_e32 v52, 0xffff0000, v22
; GFX950-NEXT:    v_and_b32_e32 v53, 0xffff0000, v21
; GFX950-NEXT:    v_and_b32_e32 v54, 0xffff0000, v20
; GFX950-NEXT:    v_and_b32_e32 v55, 0xffff0000, v19
; GFX950-NEXT:    v_accvgpr_write_b32 a0, v40 ; Reload Reuse
; GFX950-NEXT:    v_and_b32_e32 v40, 0xffff0000, v18
; GFX950-NEXT:    v_accvgpr_write_b32 a1, v41 ; Reload Reuse
; GFX950-NEXT:    v_and_b32_e32 v41, 0xffff0000, v17
; GFX950-NEXT:    v_accvgpr_write_b32 a2, v42 ; Reload Reuse
; GFX950-NEXT:    v_and_b32_e32 v42, 0xffff0000, v16
; GFX950-NEXT:    s_waitcnt vmcnt(0)
; GFX950-NEXT:    v_lshrrev_b32_e32 v34, 16, v51
; GFX950-NEXT:    v_and_b32_e32 v37, 0xffff0000, v51
; GFX950-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v37, v37
; GFX950-NEXT:    v_lshlrev_b32_e32 v33, 16, v32
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v34, v34, v32, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v37, 16, v34
; GFX950-NEXT:    v_cmp_gt_f32_e32 vcc, v33, v37
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v33, v34, v32, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v33, 16, v33
; GFX950-NEXT:    v_max_f32_e32 v33, v33, v33
; GFX950-NEXT:    v_cvt_pk_bf16_f32 v33, v33, s0
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v32
; GFX950-NEXT:    v_lshlrev_b32_e32 v37, 16, v33
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v32, v33, v32, vcc
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v34
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v32, v32, v34, vcc
; GFX950-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v37
; GFX950-NEXT:    v_lshlrev_b32_e32 v34, 16, v36
; GFX950-NEXT:    v_and_b32_e32 v37, 0xffff0000, v28
; GFX950-NEXT:    v_cndmask_b32_e32 v32, v33, v32, vcc
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v35
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v33, v36, v35, vcc
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v38
; GFX950-NEXT:    v_lshrrev_b32_e32 v35, 16, v28
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v33, v33, v38, vcc
; GFX950-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v34
; GFX950-NEXT:    v_and_b32_e32 v34, 0xffff0000, v12
; GFX950-NEXT:    v_and_b32_e32 v38, 0xffff0000, v27
; GFX950-NEXT:    v_cndmask_b32_e32 v33, v36, v33, vcc
; GFX950-NEXT:    v_lshrrev_b32_e32 v36, 16, v12
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v34, v34
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v34, v36, v35, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v37, v37
; GFX950-NEXT:    v_lshlrev_b32_e32 v36, 16, v34
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v35, v35, v34, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v37, 16, v35
; GFX950-NEXT:    v_cmp_gt_f32_e32 vcc, v36, v37
; GFX950-NEXT:    v_lshrrev_b32_e32 v37, 16, v11
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v36, v35, v34, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v36, 16, v36
; GFX950-NEXT:    v_max_f32_e32 v36, v36, v36
; GFX950-NEXT:    v_cvt_pk_bf16_f32 v36, v36, s0
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v34
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v34, v36, v34, vcc
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v35
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v34, v34, v35, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v35, 16, v36
; GFX950-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v35
; GFX950-NEXT:    v_and_b32_e32 v35, 0xffff0000, v11
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v34, v36, v34, vcc
; GFX950-NEXT:    v_lshrrev_b32_e32 v36, 16, v27
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v35, v35
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v35, v37, v36, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v38, v38
; GFX950-NEXT:    v_lshlrev_b32_e32 v37, 16, v35
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v36, v36, v35, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v38, 16, v36
; GFX950-NEXT:    v_cmp_gt_f32_e32 vcc, v37, v38
; GFX950-NEXT:    v_lshrrev_b32_e32 v38, 16, v10
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v37, v36, v35, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v37, 16, v37
; GFX950-NEXT:    v_max_f32_e32 v37, v37, v37
; GFX950-NEXT:    v_cvt_pk_bf16_f32 v37, v37, s0
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v35
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v35, v37, v35, vcc
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v36
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v35, v35, v36, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v36, 16, v37
; GFX950-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v36
; GFX950-NEXT:    v_and_b32_e32 v36, 0xffff0000, v10
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v35, v37, v35, vcc
; GFX950-NEXT:    v_lshrrev_b32_e32 v37, 16, v26
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v36, v36
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v36, v38, v37, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v39, v39
; GFX950-NEXT:    v_lshlrev_b32_e32 v38, 16, v36
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v37, v37, v36, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v39, 16, v37
; GFX950-NEXT:    v_cmp_gt_f32_e32 vcc, v38, v39
; GFX950-NEXT:    v_lshrrev_b32_e32 v39, 16, v9
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v38, v37, v36, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v38, 16, v38
; GFX950-NEXT:    v_max_f32_e32 v38, v38, v38
; GFX950-NEXT:    v_cvt_pk_bf16_f32 v38, v38, s0
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v36
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v36, v38, v36, vcc
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v37
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v36, v36, v37, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v37, 16, v38
; GFX950-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v37
; GFX950-NEXT:    v_and_b32_e32 v37, 0xffff0000, v9
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v36, v38, v36, vcc
; GFX950-NEXT:    v_lshrrev_b32_e32 v38, 16, v25
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v37, v37
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v37, v39, v38, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v48, v48
; GFX950-NEXT:    v_lshlrev_b32_e32 v39, 16, v37
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v38, v38, v37, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v48, 16, v38
; GFX950-NEXT:    v_cmp_gt_f32_e32 vcc, v39, v48
; GFX950-NEXT:    v_lshrrev_b32_e32 v48, 16, v8
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v39, v38, v37, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v39, 16, v39
; GFX950-NEXT:    v_max_f32_e32 v39, v39, v39
; GFX950-NEXT:    v_cvt_pk_bf16_f32 v39, v39, s0
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v37
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v37, v39, v37, vcc
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v38
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v37, v37, v38, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v38, 16, v39
; GFX950-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v38
; GFX950-NEXT:    v_and_b32_e32 v38, 0xffff0000, v8
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v37, v39, v37, vcc
; GFX950-NEXT:    v_lshrrev_b32_e32 v39, 16, v24
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v38, v38
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v38, v48, v39, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v49, v49
; GFX950-NEXT:    v_lshlrev_b32_e32 v48, 16, v38
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v39, v39, v38, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v49, 16, v39
; GFX950-NEXT:    v_cmp_gt_f32_e32 vcc, v48, v49
; GFX950-NEXT:    v_lshrrev_b32_e32 v49, 16, v7
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v48, v39, v38, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v48, 16, v48
; GFX950-NEXT:    v_max_f32_e32 v48, v48, v48
; GFX950-NEXT:    v_cvt_pk_bf16_f32 v48, v48, s0
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v38
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v38, v48, v38, vcc
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v39
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v38, v38, v39, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v39, 16, v48
; GFX950-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v39
; GFX950-NEXT:    v_and_b32_e32 v39, 0xffff0000, v7
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v38, v48, v38, vcc
; GFX950-NEXT:    v_lshrrev_b32_e32 v48, 16, v23
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v39, v39
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v39, v49, v48, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v50, v50
; GFX950-NEXT:    v_lshlrev_b32_e32 v49, 16, v39
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v48, v48, v39, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v50, 16, v48
; GFX950-NEXT:    v_cmp_gt_f32_e32 vcc, v49, v50
; GFX950-NEXT:    v_lshrrev_b32_e32 v50, 16, v6
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v49, v48, v39, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v49, 16, v49
; GFX950-NEXT:    v_max_f32_e32 v49, v49, v49
; GFX950-NEXT:    v_cvt_pk_bf16_f32 v49, v49, s0
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v39
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v39, v49, v39, vcc
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v48
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v39, v39, v48, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v48, 16, v49
; GFX950-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v48
; GFX950-NEXT:    v_and_b32_e32 v48, 0xffff0000, v6
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v39, v49, v39, vcc
; GFX950-NEXT:    v_lshrrev_b32_e32 v49, 16, v22
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v48, v48
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v48, v50, v49, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v52, v52
; GFX950-NEXT:    v_lshlrev_b32_e32 v50, 16, v48
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v49, v49, v48, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v52, 16, v49
; GFX950-NEXT:    v_cmp_gt_f32_e32 vcc, v50, v52
; GFX950-NEXT:    v_lshrrev_b32_e32 v52, 16, v5
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v50, v49, v48, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v50, 16, v50
; GFX950-NEXT:    v_max_f32_e32 v50, v50, v50
; GFX950-NEXT:    v_cvt_pk_bf16_f32 v50, v50, s0
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v48
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v48, v50, v48, vcc
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v49
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v48, v48, v49, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v49, 16, v50
; GFX950-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v49
; GFX950-NEXT:    v_and_b32_e32 v49, 0xffff0000, v5
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v48, v50, v48, vcc
; GFX950-NEXT:    v_lshrrev_b32_e32 v50, 16, v21
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v49, v49
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v49, v52, v50, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v53, v53
; GFX950-NEXT:    v_lshlrev_b32_e32 v52, 16, v49
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v50, v50, v49, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v53, 16, v50
; GFX950-NEXT:    v_cmp_gt_f32_e32 vcc, v52, v53
; GFX950-NEXT:    v_lshrrev_b32_e32 v53, 16, v4
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v52, v50, v49, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v52, 16, v52
; GFX950-NEXT:    v_max_f32_e32 v52, v52, v52
; GFX950-NEXT:    v_cvt_pk_bf16_f32 v52, v52, s0
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v49
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v49, v52, v49, vcc
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v50
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v49, v49, v50, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v50, 16, v52
; GFX950-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v50
; GFX950-NEXT:    v_and_b32_e32 v50, 0xffff0000, v4
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v49, v52, v49, vcc
; GFX950-NEXT:    v_lshrrev_b32_e32 v52, 16, v20
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v50, v50
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v50, v53, v52, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v54, v54
; GFX950-NEXT:    v_lshlrev_b32_e32 v53, 16, v50
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v52, v52, v50, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v54, 16, v52
; GFX950-NEXT:    v_cmp_gt_f32_e32 vcc, v53, v54
; GFX950-NEXT:    v_lshrrev_b32_e32 v54, 16, v3
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v53, v52, v50, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v53, 16, v53
; GFX950-NEXT:    v_max_f32_e32 v53, v53, v53
; GFX950-NEXT:    v_cvt_pk_bf16_f32 v53, v53, s0
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v50
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v50, v53, v50, vcc
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v52
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v50, v50, v52, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v52, 16, v53
; GFX950-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v52
; GFX950-NEXT:    v_and_b32_e32 v52, 0xffff0000, v3
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v50, v53, v50, vcc
; GFX950-NEXT:    v_lshrrev_b32_e32 v53, 16, v19
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v52, v52
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v52, v54, v53, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v55, v55
; GFX950-NEXT:    v_lshlrev_b32_e32 v54, 16, v52
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v53, v53, v52, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v55, 16, v53
; GFX950-NEXT:    v_cmp_gt_f32_e32 vcc, v54, v55
; GFX950-NEXT:    v_lshrrev_b32_e32 v55, 16, v2
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v54, v53, v52, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v54, 16, v54
; GFX950-NEXT:    v_max_f32_e32 v54, v54, v54
; GFX950-NEXT:    v_cvt_pk_bf16_f32 v54, v54, s0
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v52
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v52, v54, v52, vcc
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v53
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v52, v52, v53, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v53, 16, v54
; GFX950-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v53
; GFX950-NEXT:    v_and_b32_e32 v53, 0xffff0000, v2
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v52, v54, v52, vcc
; GFX950-NEXT:    v_lshrrev_b32_e32 v54, 16, v18
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v53, v53
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v53, v55, v54, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v40, v40
; GFX950-NEXT:    v_lshlrev_b32_e32 v55, 16, v53
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v54, v54, v53, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v40, 16, v54
; GFX950-NEXT:    v_cmp_gt_f32_e32 vcc, v55, v40
; GFX950-NEXT:    v_lshrrev_b32_e32 v40, 16, v1
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v55, v54, v53, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v55, 16, v55
; GFX950-NEXT:    v_max_f32_e32 v55, v55, v55
; GFX950-NEXT:    v_cvt_pk_bf16_f32 v55, v55, s0
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v53
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v53, v55, v53, vcc
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v54
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v53, v53, v54, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v54, 16, v55
; GFX950-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v54
; GFX950-NEXT:    v_and_b32_e32 v54, 0xffff0000, v1
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v53, v55, v53, vcc
; GFX950-NEXT:    v_lshrrev_b32_e32 v55, 16, v17
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v54, v54
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v54, v40, v55, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v41, v41
; GFX950-NEXT:    v_lshlrev_b32_e32 v40, 16, v54
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v55, v55, v54, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v41, 16, v55
; GFX950-NEXT:    v_cmp_gt_f32_e32 vcc, v40, v41
; GFX950-NEXT:    v_lshrrev_b32_e32 v41, 16, v0
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v40, v55, v54, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v40, 16, v40
; GFX950-NEXT:    v_max_f32_e32 v40, v40, v40
; GFX950-NEXT:    v_cvt_pk_bf16_f32 v40, v40, s0
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v54
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v54, v40, v54, vcc
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v55
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v54, v54, v55, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v55, 16, v40
; GFX950-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v55
; GFX950-NEXT:    v_and_b32_e32 v55, 0xffff0000, v0
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v54, v40, v54, vcc
; GFX950-NEXT:    v_lshrrev_b32_e32 v40, 16, v16
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v55, v55
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v55, v41, v40, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v42, v42
; GFX950-NEXT:    v_lshlrev_b32_e32 v41, 16, v55
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v40, v40, v55, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v42, 16, v40
; GFX950-NEXT:    v_cmp_gt_f32_e32 vcc, v41, v42
; GFX950-NEXT:    v_accvgpr_read_b32 v42, a2 ; Reload Reuse
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v41, v40, v55, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v41, 16, v41
; GFX950-NEXT:    v_max_f32_e32 v41, v41, v41
; GFX950-NEXT:    v_cvt_pk_bf16_f32 v41, v41, s0
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v55
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v55, v41, v55, vcc
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v40
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v55, v55, v40, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v40, 16, v41
; GFX950-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v40
; GFX950-NEXT:    v_lshlrev_b32_e32 v40, 16, v15
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v55, v41, v55, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v40, v40
; GFX950-NEXT:    v_lshlrev_b32_e32 v40, 16, v51
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v15, v15, v51, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v40, v40
; GFX950-NEXT:    v_lshlrev_b32_e32 v41, 16, v15
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v51, v51, v15, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v40, 16, v51
; GFX950-NEXT:    v_cmp_gt_f32_e32 vcc, v41, v40
; GFX950-NEXT:    v_accvgpr_read_b32 v41, a1 ; Reload Reuse
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v40, v51, v15, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v40, 16, v40
; GFX950-NEXT:    v_max_f32_e32 v40, v40, v40
; GFX950-NEXT:    v_cvt_pk_bf16_f32 v40, v40, s0
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v15
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v15, v40, v15, vcc
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v51
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v15, v15, v51, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v51, 16, v40
; GFX950-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v51
; GFX950-NEXT:    v_lshlrev_b32_e32 v51, 16, v14
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v15, v40, v15, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v51, v51
; GFX950-NEXT:    v_lshlrev_b32_e32 v51, 16, v30
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v14, v14, v30, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v51, v51
; GFX950-NEXT:    v_lshlrev_b32_e32 v40, 16, v14
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v30, v30, v14, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v51, 16, v30
; GFX950-NEXT:    v_cmp_gt_f32_e32 vcc, v40, v51
; GFX950-NEXT:    v_accvgpr_read_b32 v40, a0 ; Reload Reuse
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v51, v30, v14, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v51, 16, v51
; GFX950-NEXT:    v_max_f32_e32 v51, v51, v51
; GFX950-NEXT:    v_cvt_pk_bf16_f32 v51, v51, s0
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v14
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v14, v51, v14, vcc
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v30
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v14, v14, v30, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v30, 16, v51
; GFX950-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v30
; GFX950-NEXT:    v_lshlrev_b32_e32 v30, 16, v13
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v14, v51, v14, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v30, v30
; GFX950-NEXT:    v_lshlrev_b32_e32 v30, 16, v29
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v13, v13, v29, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v30, v30
; GFX950-NEXT:    v_lshlrev_b32_e32 v51, 16, v13
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v29, v29, v13, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v30, 16, v29
; GFX950-NEXT:    v_cmp_gt_f32_e32 vcc, v51, v30
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v30, v29, v13, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v30, 16, v30
; GFX950-NEXT:    v_max_f32_e32 v30, v30, v30
; GFX950-NEXT:    v_cvt_pk_bf16_f32 v30, v30, s0
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v13
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v13, v30, v13, vcc
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v29
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v13, v13, v29, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v29, 16, v30
; GFX950-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v29
; GFX950-NEXT:    v_lshlrev_b32_e32 v29, 16, v12
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v13, v30, v13, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v29, v29
; GFX950-NEXT:    v_lshlrev_b32_e32 v29, 16, v28
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v12, v12, v28, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v29, v29
; GFX950-NEXT:    v_lshlrev_b32_e32 v30, 16, v12
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v28, v28, v12, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v29, 16, v28
; GFX950-NEXT:    v_cmp_gt_f32_e32 vcc, v30, v29
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v29, v28, v12, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v29, 16, v29
; GFX950-NEXT:    v_max_f32_e32 v29, v29, v29
; GFX950-NEXT:    v_cvt_pk_bf16_f32 v29, v29, s0
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v12
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v12, v29, v12, vcc
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v28
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v12, v12, v28, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v28, 16, v29
; GFX950-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v28
; GFX950-NEXT:    v_lshlrev_b32_e32 v28, 16, v11
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v12, v29, v12, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v28, v28
; GFX950-NEXT:    v_lshlrev_b32_e32 v28, 16, v27
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v11, v11, v27, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v28, v28
; GFX950-NEXT:    v_lshlrev_b32_e32 v29, 16, v11
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v27, v27, v11, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v28, 16, v27
; GFX950-NEXT:    v_cmp_gt_f32_e32 vcc, v29, v28
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v28, v27, v11, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
; GFX950-NEXT:    v_max_f32_e32 v28, v28, v28
; GFX950-NEXT:    v_cvt_pk_bf16_f32 v28, v28, s0
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v11
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v11, v28, v11, vcc
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v27
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v11, v11, v27, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v27, 16, v28
; GFX950-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v27
; GFX950-NEXT:    v_lshlrev_b32_e32 v27, 16, v10
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v11, v28, v11, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v27, v27
; GFX950-NEXT:    v_lshlrev_b32_e32 v27, 16, v26
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v10, v10, v26, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v27, v27
; GFX950-NEXT:    v_lshlrev_b32_e32 v28, 16, v10
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v26, v26, v10, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v27, 16, v26
; GFX950-NEXT:    v_cmp_gt_f32_e32 vcc, v28, v27
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v27, v26, v10, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v27, 16, v27
; GFX950-NEXT:    v_max_f32_e32 v27, v27, v27
; GFX950-NEXT:    v_cvt_pk_bf16_f32 v27, v27, s0
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v10
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v10, v27, v10, vcc
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v26
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v10, v10, v26, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v26, 16, v27
; GFX950-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v26
; GFX950-NEXT:    v_lshlrev_b32_e32 v26, 16, v9
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v10, v27, v10, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v26, v26
; GFX950-NEXT:    v_lshlrev_b32_e32 v26, 16, v25
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v9, v9, v25, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v26, v26
; GFX950-NEXT:    v_lshlrev_b32_e32 v27, 16, v9
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v25, v25, v9, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v26, 16, v25
; GFX950-NEXT:    v_cmp_gt_f32_e32 vcc, v27, v26
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v26, v25, v9, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
; GFX950-NEXT:    v_max_f32_e32 v26, v26, v26
; GFX950-NEXT:    v_cvt_pk_bf16_f32 v26, v26, s0
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v9
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v9, v26, v9, vcc
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v25
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v9, v9, v25, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v25, 16, v26
; GFX950-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v25
; GFX950-NEXT:    v_lshlrev_b32_e32 v25, 16, v8
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v9, v26, v9, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v25, v25
; GFX950-NEXT:    v_lshlrev_b32_e32 v25, 16, v24
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v8, v8, v24, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v25, v25
; GFX950-NEXT:    v_lshlrev_b32_e32 v26, 16, v8
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v24, v24, v8, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v25, 16, v24
; GFX950-NEXT:    v_cmp_gt_f32_e32 vcc, v26, v25
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v25, v24, v8, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v25, 16, v25
; GFX950-NEXT:    v_max_f32_e32 v25, v25, v25
; GFX950-NEXT:    v_cvt_pk_bf16_f32 v25, v25, s0
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v8
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v8, v25, v8, vcc
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v24
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v8, v8, v24, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v24, 16, v25
; GFX950-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v24
; GFX950-NEXT:    v_lshlrev_b32_e32 v24, 16, v7
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v8, v25, v8, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v24, v24
; GFX950-NEXT:    v_lshlrev_b32_e32 v24, 16, v23
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v7, v7, v23, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v24, v24
; GFX950-NEXT:    v_lshlrev_b32_e32 v25, 16, v7
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v23, v23, v7, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v24, 16, v23
; GFX950-NEXT:    v_cmp_gt_f32_e32 vcc, v25, v24
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v24, v23, v7, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
; GFX950-NEXT:    v_max_f32_e32 v24, v24, v24
; GFX950-NEXT:    v_cvt_pk_bf16_f32 v24, v24, s0
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v7
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v7, v24, v7, vcc
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v23
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v7, v7, v23, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v23, 16, v24
; GFX950-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v23
; GFX950-NEXT:    v_lshlrev_b32_e32 v23, 16, v6
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v7, v24, v7, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v23, v23
; GFX950-NEXT:    v_lshlrev_b32_e32 v23, 16, v22
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v6, v6, v22, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v23, v23
; GFX950-NEXT:    v_lshlrev_b32_e32 v24, 16, v6
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v22, v22, v6, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v23, 16, v22
; GFX950-NEXT:    v_cmp_gt_f32_e32 vcc, v24, v23
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v23, v22, v6, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; GFX950-NEXT:    v_max_f32_e32 v23, v23, v23
; GFX950-NEXT:    v_cvt_pk_bf16_f32 v23, v23, s0
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v6
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v6, v23, v6, vcc
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v22
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v6, v6, v22, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v22, 16, v23
; GFX950-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v22
; GFX950-NEXT:    v_lshlrev_b32_e32 v22, 16, v5
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v6, v23, v6, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v22, v22
; GFX950-NEXT:    v_lshlrev_b32_e32 v22, 16, v21
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v5, v5, v21, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v22, v22
; GFX950-NEXT:    v_lshlrev_b32_e32 v23, 16, v5
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v21, v21, v5, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v22, 16, v21
; GFX950-NEXT:    v_cmp_gt_f32_e32 vcc, v23, v22
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v22, v21, v5, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v22, 16, v22
; GFX950-NEXT:    v_max_f32_e32 v22, v22, v22
; GFX950-NEXT:    v_cvt_pk_bf16_f32 v22, v22, s0
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v5
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v5, v22, v5, vcc
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v21
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v5, v5, v21, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v21, 16, v22
; GFX950-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v21
; GFX950-NEXT:    v_lshlrev_b32_e32 v21, 16, v4
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v5, v22, v5, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v21, v21
; GFX950-NEXT:    v_lshlrev_b32_e32 v21, 16, v20
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v4, v4, v20, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v21, v21
; GFX950-NEXT:    v_lshlrev_b32_e32 v22, 16, v4
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v20, v20, v4, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v21, 16, v20
; GFX950-NEXT:    v_cmp_gt_f32_e32 vcc, v22, v21
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v21, v20, v4, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
; GFX950-NEXT:    v_max_f32_e32 v21, v21, v21
; GFX950-NEXT:    v_cvt_pk_bf16_f32 v21, v21, s0
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v4
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v4, v21, v4, vcc
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v20
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v4, v4, v20, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v20, 16, v21
; GFX950-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v20
; GFX950-NEXT:    v_lshlrev_b32_e32 v20, 16, v3
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v4, v21, v4, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v20, v20
; GFX950-NEXT:    v_lshlrev_b32_e32 v20, 16, v19
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v3, v3, v19, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v20, v20
; GFX950-NEXT:    v_lshlrev_b32_e32 v21, 16, v3
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v19, v19, v3, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v20, 16, v19
; GFX950-NEXT:    v_cmp_gt_f32_e32 vcc, v21, v20
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v20, v19, v3, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
; GFX950-NEXT:    v_max_f32_e32 v20, v20, v20
; GFX950-NEXT:    v_cvt_pk_bf16_f32 v20, v20, s0
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v3
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v3, v20, v3, vcc
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v19
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v3, v3, v19, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v19, 16, v20
; GFX950-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v19
; GFX950-NEXT:    v_lshlrev_b32_e32 v19, 16, v2
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v3, v20, v3, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v19, v19
; GFX950-NEXT:    v_lshlrev_b32_e32 v19, 16, v18
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v2, v2, v18, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v19, v19
; GFX950-NEXT:    v_lshlrev_b32_e32 v20, 16, v2
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v18, v18, v2, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v19, 16, v18
; GFX950-NEXT:    v_cmp_gt_f32_e32 vcc, v20, v19
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v19, v18, v2, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
; GFX950-NEXT:    v_max_f32_e32 v19, v19, v19
; GFX950-NEXT:    v_cvt_pk_bf16_f32 v19, v19, s0
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v2
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v2, v19, v2, vcc
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v18
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v2, v2, v18, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v18, 16, v19
; GFX950-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v18
; GFX950-NEXT:    v_lshlrev_b32_e32 v18, 16, v1
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v2, v19, v2, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; GFX950-NEXT:    v_lshlrev_b32_e32 v18, 16, v17
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v1, v1, v17, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; GFX950-NEXT:    v_lshlrev_b32_e32 v19, 16, v1
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v17, v17, v1, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v18, 16, v17
; GFX950-NEXT:    v_cmp_gt_f32_e32 vcc, v19, v18
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v18, v17, v1, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
; GFX950-NEXT:    v_max_f32_e32 v18, v18, v18
; GFX950-NEXT:    v_cvt_pk_bf16_f32 v18, v18, s0
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v1
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v1, v18, v1, vcc
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v17
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v1, v1, v17, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v17, 16, v18
; GFX950-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v17
; GFX950-NEXT:    v_lshlrev_b32_e32 v17, 16, v0
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v1, v18, v1, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX950-NEXT:    v_lshlrev_b32_e32 v17, 16, v16
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v0, v0, v16, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX950-NEXT:    v_lshlrev_b32_e32 v18, 16, v0
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v16, v16, v0, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v17, 16, v16
; GFX950-NEXT:    v_cmp_gt_f32_e32 vcc, v18, v17
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v17, v16, v0, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
; GFX950-NEXT:    v_max_f32_e32 v17, v17, v17
; GFX950-NEXT:    v_cvt_pk_bf16_f32 v17, v17, s0
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v0
; GFX950-NEXT:    s_mov_b32 s0, 0x5040100
; GFX950-NEXT:    v_perm_b32 v1, v54, v1, s0
; GFX950-NEXT:    v_cndmask_b32_e32 v0, v17, v0, vcc
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v16
; GFX950-NEXT:    v_perm_b32 v2, v53, v2, s0
; GFX950-NEXT:    v_perm_b32 v3, v52, v3, s0
; GFX950-NEXT:    v_cndmask_b32_e32 v0, v0, v16, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v16, 16, v17
; GFX950-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v16
; GFX950-NEXT:    v_perm_b32 v4, v50, v4, s0
; GFX950-NEXT:    v_perm_b32 v5, v49, v5, s0
; GFX950-NEXT:    v_cndmask_b32_e32 v0, v17, v0, vcc
; GFX950-NEXT:    v_perm_b32 v0, v55, v0, s0
; GFX950-NEXT:    v_perm_b32 v6, v48, v6, s0
; GFX950-NEXT:    v_perm_b32 v7, v39, v7, s0
; GFX950-NEXT:    v_perm_b32 v8, v38, v8, s0
; GFX950-NEXT:    v_perm_b32 v9, v37, v9, s0
; GFX950-NEXT:    v_perm_b32 v10, v36, v10, s0
; GFX950-NEXT:    v_perm_b32 v11, v35, v11, s0
; GFX950-NEXT:    v_perm_b32 v12, v34, v12, s0
; GFX950-NEXT:    v_perm_b32 v13, v33, v13, s0
; GFX950-NEXT:    v_perm_b32 v14, v31, v14, s0
; GFX950-NEXT:    v_perm_b32 v15, v32, v15, s0
; GFX950-NEXT:    s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_maximumnum_v32bf16:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT:    buffer_load_dword v53, off, s[0:3], s32
; GFX10-NEXT:    v_and_b32_e32 v31, 0xffff0000, v15
; GFX10-NEXT:    v_lshrrev_b32_e32 v32, 16, v15
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v31, v31
; GFX10-NEXT:    s_waitcnt vmcnt(0)
; GFX10-NEXT:    v_lshrrev_b32_e32 v31, 16, v53
; GFX10-NEXT:    v_and_b32_e32 v34, 0xffff0000, v53
; GFX10-NEXT:    v_cndmask_b32_e32 v32, v32, v31, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v34, v34
; GFX10-NEXT:    v_lshlrev_b32_e32 v33, 16, v32
; GFX10-NEXT:    v_cndmask_b32_e32 v31, v31, v32, vcc_lo
; GFX10-NEXT:    v_lshlrev_b32_e32 v34, 16, v31
; GFX10-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v33, v34
; GFX10-NEXT:    v_cndmask_b32_e32 v33, v31, v32, vcc_lo
; GFX10-NEXT:    v_lshlrev_b32_e32 v33, 16, v33
; GFX10-NEXT:    v_max_f32_e32 v33, v33, v33
; GFX10-NEXT:    v_bfe_u32 v34, v33, 16, 1
; GFX10-NEXT:    v_or_b32_e32 v35, 0x400000, v33
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v33, v33
; GFX10-NEXT:    v_add3_u32 v34, v34, v33, 0x7fff
; GFX10-NEXT:    v_cndmask_b32_e32 v33, v34, v35, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v32
; GFX10-NEXT:    v_and_b32_e32 v35, 0xffff0000, v30
; GFX10-NEXT:    v_lshrrev_b32_e32 v34, 16, v33
; GFX10-NEXT:    v_cndmask_b32_e32 v32, v34, v32, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v31
; GFX10-NEXT:    v_cndmask_b32_e32 v31, v32, v31, vcc_lo
; GFX10-NEXT:    v_and_b32_e32 v32, 0xffff0000, v33
; GFX10-NEXT:    v_lshrrev_b32_e32 v33, 16, v14
; GFX10-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v32
; GFX10-NEXT:    v_and_b32_e32 v32, 0xffff0000, v14
; GFX10-NEXT:    v_cndmask_b32_e32 v31, v34, v31, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v32, v32
; GFX10-NEXT:    v_lshrrev_b32_e32 v32, 16, v30
; GFX10-NEXT:    v_cndmask_b32_e32 v33, v33, v32, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
; GFX10-NEXT:    v_lshlrev_b32_e32 v34, 16, v33
; GFX10-NEXT:    v_cndmask_b32_e32 v32, v32, v33, vcc_lo
; GFX10-NEXT:    v_lshlrev_b32_e32 v35, 16, v32
; GFX10-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v34, v35
; GFX10-NEXT:    v_cndmask_b32_e32 v34, v32, v33, vcc_lo
; GFX10-NEXT:    v_lshlrev_b32_e32 v34, 16, v34
; GFX10-NEXT:    v_max_f32_e32 v34, v34, v34
; GFX10-NEXT:    v_bfe_u32 v35, v34, 16, 1
; GFX10-NEXT:    v_or_b32_e32 v36, 0x400000, v34
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v34, v34
; GFX10-NEXT:    v_add3_u32 v35, v35, v34, 0x7fff
; GFX10-NEXT:    v_cndmask_b32_e32 v34, v35, v36, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v33
; GFX10-NEXT:    v_and_b32_e32 v36, 0xffff0000, v29
; GFX10-NEXT:    v_lshrrev_b32_e32 v35, 16, v34
; GFX10-NEXT:    v_cndmask_b32_e32 v33, v35, v33, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v32
; GFX10-NEXT:    v_cndmask_b32_e32 v32, v33, v32, vcc_lo
; GFX10-NEXT:    v_and_b32_e32 v33, 0xffff0000, v34
; GFX10-NEXT:    v_lshrrev_b32_e32 v34, 16, v13
; GFX10-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v33
; GFX10-NEXT:    v_and_b32_e32 v33, 0xffff0000, v13
; GFX10-NEXT:    v_cndmask_b32_e32 v32, v35, v32, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v33, v33
; GFX10-NEXT:    v_lshrrev_b32_e32 v33, 16, v29
; GFX10-NEXT:    v_cndmask_b32_e32 v34, v34, v33, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v36, v36
; GFX10-NEXT:    v_lshlrev_b32_e32 v35, 16, v34
; GFX10-NEXT:    v_cndmask_b32_e32 v33, v33, v34, vcc_lo
; GFX10-NEXT:    v_lshlrev_b32_e32 v36, 16, v33
; GFX10-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v35, v36
; GFX10-NEXT:    v_cndmask_b32_e32 v35, v33, v34, vcc_lo
; GFX10-NEXT:    v_lshlrev_b32_e32 v35, 16, v35
; GFX10-NEXT:    v_max_f32_e32 v35, v35, v35
; GFX10-NEXT:    v_bfe_u32 v36, v35, 16, 1
; GFX10-NEXT:    v_or_b32_e32 v37, 0x400000, v35
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
; GFX10-NEXT:    v_add3_u32 v36, v36, v35, 0x7fff
; GFX10-NEXT:    v_cndmask_b32_e32 v35, v36, v37, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v34
; GFX10-NEXT:    v_and_b32_e32 v37, 0xffff0000, v28
; GFX10-NEXT:    v_lshrrev_b32_e32 v36, 16, v35
; GFX10-NEXT:    v_cndmask_b32_e32 v34, v36, v34, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v33
; GFX10-NEXT:    v_cndmask_b32_e32 v33, v34, v33, vcc_lo
; GFX10-NEXT:    v_and_b32_e32 v34, 0xffff0000, v35
; GFX10-NEXT:    v_lshrrev_b32_e32 v35, 16, v12
; GFX10-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v34
; GFX10-NEXT:    v_and_b32_e32 v34, 0xffff0000, v12
; GFX10-NEXT:    v_cndmask_b32_e32 v33, v36, v33, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v34, v34
; GFX10-NEXT:    v_lshrrev_b32_e32 v34, 16, v28
; GFX10-NEXT:    v_cndmask_b32_e32 v35, v35, v34, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v37, v37
; GFX10-NEXT:    v_lshlrev_b32_e32 v36, 16, v35
; GFX10-NEXT:    v_cndmask_b32_e32 v34, v34, v35, vcc_lo
; GFX10-NEXT:    v_lshlrev_b32_e32 v37, 16, v34
; GFX10-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v36, v37
; GFX10-NEXT:    v_cndmask_b32_e32 v36, v34, v35, vcc_lo
; GFX10-NEXT:    v_lshlrev_b32_e32 v36, 16, v36
; GFX10-NEXT:    v_max_f32_e32 v36, v36, v36
; GFX10-NEXT:    v_bfe_u32 v37, v36, 16, 1
; GFX10-NEXT:    v_or_b32_e32 v38, 0x400000, v36
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v36, v36
; GFX10-NEXT:    v_add3_u32 v37, v37, v36, 0x7fff
; GFX10-NEXT:    v_cndmask_b32_e32 v36, v37, v38, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v35
; GFX10-NEXT:    v_and_b32_e32 v38, 0xffff0000, v27
; GFX10-NEXT:    v_lshrrev_b32_e32 v37, 16, v36
; GFX10-NEXT:    v_cndmask_b32_e32 v35, v37, v35, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v34
; GFX10-NEXT:    v_cndmask_b32_e32 v34, v35, v34, vcc_lo
; GFX10-NEXT:    v_and_b32_e32 v35, 0xffff0000, v36
; GFX10-NEXT:    v_lshrrev_b32_e32 v36, 16, v11
; GFX10-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v35
; GFX10-NEXT:    v_and_b32_e32 v35, 0xffff0000, v11
; GFX10-NEXT:    v_cndmask_b32_e32 v34, v37, v34, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
; GFX10-NEXT:    v_lshrrev_b32_e32 v35, 16, v27
; GFX10-NEXT:    v_cndmask_b32_e32 v36, v36, v35, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v38, v38
; GFX10-NEXT:    v_lshlrev_b32_e32 v37, 16, v36
; GFX10-NEXT:    v_cndmask_b32_e32 v35, v35, v36, vcc_lo
; GFX10-NEXT:    v_lshlrev_b32_e32 v38, 16, v35
; GFX10-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v37, v38
; GFX10-NEXT:    v_cndmask_b32_e32 v37, v35, v36, vcc_lo
; GFX10-NEXT:    v_lshlrev_b32_e32 v37, 16, v37
; GFX10-NEXT:    v_max_f32_e32 v37, v37, v37
; GFX10-NEXT:    v_bfe_u32 v38, v37, 16, 1
; GFX10-NEXT:    v_or_b32_e32 v39, 0x400000, v37
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v37, v37
; GFX10-NEXT:    v_add3_u32 v38, v38, v37, 0x7fff
; GFX10-NEXT:    v_cndmask_b32_e32 v37, v38, v39, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v36
; GFX10-NEXT:    v_and_b32_e32 v39, 0xffff0000, v26
; GFX10-NEXT:    v_lshrrev_b32_e32 v38, 16, v37
; GFX10-NEXT:    v_cndmask_b32_e32 v36, v38, v36, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v35
; GFX10-NEXT:    v_cndmask_b32_e32 v35, v36, v35, vcc_lo
; GFX10-NEXT:    v_and_b32_e32 v36, 0xffff0000, v37
; GFX10-NEXT:    v_lshrrev_b32_e32 v37, 16, v10
; GFX10-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v36
; GFX10-NEXT:    v_and_b32_e32 v36, 0xffff0000, v10
; GFX10-NEXT:    v_cndmask_b32_e32 v35, v38, v35, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v36, v36
; GFX10-NEXT:    v_lshrrev_b32_e32 v36, 16, v26
; GFX10-NEXT:    v_cndmask_b32_e32 v37, v37, v36, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v39, v39
; GFX10-NEXT:    v_lshlrev_b32_e32 v38, 16, v37
; GFX10-NEXT:    v_cndmask_b32_e32 v36, v36, v37, vcc_lo
; GFX10-NEXT:    v_lshlrev_b32_e32 v39, 16, v36
; GFX10-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v38, v39
; GFX10-NEXT:    v_cndmask_b32_e32 v38, v36, v37, vcc_lo
; GFX10-NEXT:    v_lshlrev_b32_e32 v38, 16, v38
; GFX10-NEXT:    v_max_f32_e32 v38, v38, v38
; GFX10-NEXT:    v_bfe_u32 v39, v38, 16, 1
; GFX10-NEXT:    v_or_b32_e32 v48, 0x400000, v38
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v38, v38
; GFX10-NEXT:    v_add3_u32 v39, v39, v38, 0x7fff
; GFX10-NEXT:    v_cndmask_b32_e32 v38, v39, v48, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v37
; GFX10-NEXT:    v_and_b32_e32 v48, 0xffff0000, v25
; GFX10-NEXT:    v_lshrrev_b32_e32 v39, 16, v38
; GFX10-NEXT:    v_cndmask_b32_e32 v37, v39, v37, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v36
; GFX10-NEXT:    v_cndmask_b32_e32 v36, v37, v36, vcc_lo
; GFX10-NEXT:    v_and_b32_e32 v37, 0xffff0000, v38
; GFX10-NEXT:    v_lshrrev_b32_e32 v38, 16, v9
; GFX10-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v37
; GFX10-NEXT:    v_and_b32_e32 v37, 0xffff0000, v9
; GFX10-NEXT:    v_cndmask_b32_e32 v36, v39, v36, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v37, v37
; GFX10-NEXT:    v_lshrrev_b32_e32 v37, 16, v25
; GFX10-NEXT:    v_cndmask_b32_e32 v38, v38, v37, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v48, v48
; GFX10-NEXT:    v_lshlrev_b32_e32 v39, 16, v38
; GFX10-NEXT:    v_cndmask_b32_e32 v37, v37, v38, vcc_lo
; GFX10-NEXT:    v_lshlrev_b32_e32 v48, 16, v37
; GFX10-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v39, v48
; GFX10-NEXT:    v_cndmask_b32_e32 v39, v37, v38, vcc_lo
; GFX10-NEXT:    v_lshlrev_b32_e32 v39, 16, v39
; GFX10-NEXT:    v_max_f32_e32 v39, v39, v39
; GFX10-NEXT:    v_bfe_u32 v48, v39, 16, 1
; GFX10-NEXT:    v_or_b32_e32 v49, 0x400000, v39
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v39, v39
; GFX10-NEXT:    v_add3_u32 v48, v48, v39, 0x7fff
; GFX10-NEXT:    v_cndmask_b32_e32 v39, v48, v49, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v38
; GFX10-NEXT:    v_and_b32_e32 v49, 0xffff0000, v24
; GFX10-NEXT:    v_lshrrev_b32_e32 v48, 16, v39
; GFX10-NEXT:    v_cndmask_b32_e32 v38, v48, v38, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v37
; GFX10-NEXT:    v_cndmask_b32_e32 v37, v38, v37, vcc_lo
; GFX10-NEXT:    v_and_b32_e32 v38, 0xffff0000, v39
; GFX10-NEXT:    v_lshrrev_b32_e32 v39, 16, v8
; GFX10-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v38
; GFX10-NEXT:    v_and_b32_e32 v38, 0xffff0000, v8
; GFX10-NEXT:    v_cndmask_b32_e32 v37, v48, v37, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v38, v38
; GFX10-NEXT:    v_lshrrev_b32_e32 v38, 16, v24
; GFX10-NEXT:    v_cndmask_b32_e32 v39, v39, v38, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v49, v49
; GFX10-NEXT:    v_lshlrev_b32_e32 v48, 16, v39
; GFX10-NEXT:    v_cndmask_b32_e32 v38, v38, v39, vcc_lo
; GFX10-NEXT:    v_lshlrev_b32_e32 v49, 16, v38
; GFX10-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v48, v49
; GFX10-NEXT:    v_cndmask_b32_e32 v48, v38, v39, vcc_lo
; GFX10-NEXT:    v_lshlrev_b32_e32 v48, 16, v48
; GFX10-NEXT:    v_max_f32_e32 v48, v48, v48
; GFX10-NEXT:    v_bfe_u32 v49, v48, 16, 1
; GFX10-NEXT:    v_or_b32_e32 v50, 0x400000, v48
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v48, v48
; GFX10-NEXT:    v_add3_u32 v49, v49, v48, 0x7fff
; GFX10-NEXT:    v_cndmask_b32_e32 v48, v49, v50, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v39
; GFX10-NEXT:    v_and_b32_e32 v50, 0xffff0000, v23
; GFX10-NEXT:    v_lshrrev_b32_e32 v49, 16, v48
; GFX10-NEXT:    v_cndmask_b32_e32 v39, v49, v39, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v38
; GFX10-NEXT:    v_cndmask_b32_e32 v38, v39, v38, vcc_lo
; GFX10-NEXT:    v_and_b32_e32 v39, 0xffff0000, v48
; GFX10-NEXT:    v_lshrrev_b32_e32 v48, 16, v7
; GFX10-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v39
; GFX10-NEXT:    v_and_b32_e32 v39, 0xffff0000, v7
; GFX10-NEXT:    v_cndmask_b32_e32 v38, v49, v38, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v39, v39
; GFX10-NEXT:    v_lshrrev_b32_e32 v39, 16, v23
; GFX10-NEXT:    v_cndmask_b32_e32 v48, v48, v39, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v50, v50
; GFX10-NEXT:    v_lshlrev_b32_e32 v49, 16, v48
; GFX10-NEXT:    v_cndmask_b32_e32 v39, v39, v48, vcc_lo
; GFX10-NEXT:    v_lshlrev_b32_e32 v50, 16, v39
; GFX10-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v49, v50
; GFX10-NEXT:    v_cndmask_b32_e32 v49, v39, v48, vcc_lo
; GFX10-NEXT:    v_lshlrev_b32_e32 v49, 16, v49
; GFX10-NEXT:    v_max_f32_e32 v49, v49, v49
; GFX10-NEXT:    v_bfe_u32 v50, v49, 16, 1
; GFX10-NEXT:    v_or_b32_e32 v51, 0x400000, v49
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v49, v49
; GFX10-NEXT:    v_add3_u32 v50, v50, v49, 0x7fff
; GFX10-NEXT:    v_cndmask_b32_e32 v49, v50, v51, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v48
; GFX10-NEXT:    v_and_b32_e32 v51, 0xffff0000, v22
; GFX10-NEXT:    v_lshrrev_b32_e32 v50, 16, v49
; GFX10-NEXT:    v_cndmask_b32_e32 v48, v50, v48, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v39
; GFX10-NEXT:    v_cndmask_b32_e32 v39, v48, v39, vcc_lo
; GFX10-NEXT:    v_and_b32_e32 v48, 0xffff0000, v49
; GFX10-NEXT:    v_lshrrev_b32_e32 v49, 16, v6
; GFX10-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v48
; GFX10-NEXT:    v_and_b32_e32 v48, 0xffff0000, v6
; GFX10-NEXT:    v_cndmask_b32_e32 v39, v50, v39, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v48, v48
; GFX10-NEXT:    v_lshrrev_b32_e32 v48, 16, v22
; GFX10-NEXT:    v_cndmask_b32_e32 v49, v49, v48, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v51, v51
; GFX10-NEXT:    v_lshlrev_b32_e32 v50, 16, v49
; GFX10-NEXT:    v_cndmask_b32_e32 v48, v48, v49, vcc_lo
; GFX10-NEXT:    v_lshlrev_b32_e32 v51, 16, v48
; GFX10-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v50, v51
; GFX10-NEXT:    v_cndmask_b32_e32 v50, v48, v49, vcc_lo
; GFX10-NEXT:    v_lshlrev_b32_e32 v50, 16, v50
; GFX10-NEXT:    v_max_f32_e32 v50, v50, v50
; GFX10-NEXT:    v_bfe_u32 v51, v50, 16, 1
; GFX10-NEXT:    v_or_b32_e32 v52, 0x400000, v50
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v50, v50
; GFX10-NEXT:    v_add3_u32 v51, v51, v50, 0x7fff
; GFX10-NEXT:    v_cndmask_b32_e32 v50, v51, v52, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v49
; GFX10-NEXT:    v_and_b32_e32 v52, 0xffff0000, v21
; GFX10-NEXT:    v_lshrrev_b32_e32 v51, 16, v50
; GFX10-NEXT:    v_cndmask_b32_e32 v49, v51, v49, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v48
; GFX10-NEXT:    v_cndmask_b32_e32 v48, v49, v48, vcc_lo
; GFX10-NEXT:    v_and_b32_e32 v49, 0xffff0000, v50
; GFX10-NEXT:    v_lshrrev_b32_e32 v50, 16, v5
; GFX10-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v49
; GFX10-NEXT:    v_and_b32_e32 v49, 0xffff0000, v5
; GFX10-NEXT:    v_cndmask_b32_e32 v48, v51, v48, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v49, v49
; GFX10-NEXT:    v_lshrrev_b32_e32 v49, 16, v21
; GFX10-NEXT:    v_cndmask_b32_e32 v50, v50, v49, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v52, v52
; GFX10-NEXT:    v_lshlrev_b32_e32 v51, 16, v50
; GFX10-NEXT:    v_cndmask_b32_e32 v49, v49, v50, vcc_lo
; GFX10-NEXT:    v_lshlrev_b32_e32 v52, 16, v49
; GFX10-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v51, v52
; GFX10-NEXT:    v_cndmask_b32_e32 v51, v49, v50, vcc_lo
; GFX10-NEXT:    v_lshlrev_b32_e32 v51, 16, v51
; GFX10-NEXT:    v_max_f32_e32 v51, v51, v51
; GFX10-NEXT:    v_bfe_u32 v52, v51, 16, 1
; GFX10-NEXT:    v_or_b32_e32 v54, 0x400000, v51
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v51, v51
; GFX10-NEXT:    v_add3_u32 v52, v52, v51, 0x7fff
; GFX10-NEXT:    v_cndmask_b32_e32 v51, v52, v54, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v50
; GFX10-NEXT:    v_and_b32_e32 v54, 0xffff0000, v20
; GFX10-NEXT:    v_lshrrev_b32_e32 v52, 16, v51
; GFX10-NEXT:    v_cndmask_b32_e32 v50, v52, v50, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v49
; GFX10-NEXT:    v_cndmask_b32_e32 v49, v50, v49, vcc_lo
; GFX10-NEXT:    v_and_b32_e32 v50, 0xffff0000, v51
; GFX10-NEXT:    v_lshrrev_b32_e32 v51, 16, v4
; GFX10-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v50
; GFX10-NEXT:    v_and_b32_e32 v50, 0xffff0000, v4
; GFX10-NEXT:    v_cndmask_b32_e32 v49, v52, v49, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v50, v50
; GFX10-NEXT:    v_lshrrev_b32_e32 v50, 16, v20
; GFX10-NEXT:    v_cndmask_b32_e32 v51, v51, v50, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v54, v54
; GFX10-NEXT:    v_lshlrev_b32_e32 v52, 16, v51
; GFX10-NEXT:    v_cndmask_b32_e32 v50, v50, v51, vcc_lo
; GFX10-NEXT:    v_lshlrev_b32_e32 v54, 16, v50
; GFX10-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v52, v54
; GFX10-NEXT:    v_cndmask_b32_e32 v52, v50, v51, vcc_lo
; GFX10-NEXT:    v_lshlrev_b32_e32 v52, 16, v52
; GFX10-NEXT:    v_max_f32_e32 v52, v52, v52
; GFX10-NEXT:    v_bfe_u32 v54, v52, 16, 1
; GFX10-NEXT:    v_or_b32_e32 v55, 0x400000, v52
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v52, v52
; GFX10-NEXT:    v_add3_u32 v54, v54, v52, 0x7fff
; GFX10-NEXT:    v_cndmask_b32_e32 v52, v54, v55, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v51
; GFX10-NEXT:    v_and_b32_e32 v55, 0xffff0000, v19
; GFX10-NEXT:    v_lshrrev_b32_e32 v54, 16, v52
; GFX10-NEXT:    v_cndmask_b32_e32 v51, v54, v51, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v50
; GFX10-NEXT:    v_cndmask_b32_e32 v50, v51, v50, vcc_lo
; GFX10-NEXT:    v_and_b32_e32 v51, 0xffff0000, v52
; GFX10-NEXT:    v_lshrrev_b32_e32 v52, 16, v3
; GFX10-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v51
; GFX10-NEXT:    v_and_b32_e32 v51, 0xffff0000, v3
; GFX10-NEXT:    v_cndmask_b32_e32 v50, v54, v50, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v51, v51
; GFX10-NEXT:    v_lshrrev_b32_e32 v51, 16, v19
; GFX10-NEXT:    v_cndmask_b32_e32 v52, v52, v51, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v55, v55
; GFX10-NEXT:    v_lshlrev_b32_e32 v54, 16, v52
; GFX10-NEXT:    v_cndmask_b32_e32 v51, v51, v52, vcc_lo
; GFX10-NEXT:    v_lshlrev_b32_e32 v55, 16, v51
; GFX10-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v54, v55
; GFX10-NEXT:    v_cndmask_b32_e32 v54, v51, v52, vcc_lo
; GFX10-NEXT:    v_lshlrev_b32_e32 v54, 16, v54
; GFX10-NEXT:    v_max_f32_e32 v54, v54, v54
; GFX10-NEXT:    v_bfe_u32 v55, v54, 16, 1
; GFX10-NEXT:    v_or_b32_e32 v64, 0x400000, v54
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v54, v54
; GFX10-NEXT:    v_add3_u32 v55, v55, v54, 0x7fff
; GFX10-NEXT:    v_cndmask_b32_e32 v54, v55, v64, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v52
; GFX10-NEXT:    v_and_b32_e32 v64, 0xffff0000, v18
; GFX10-NEXT:    v_lshrrev_b32_e32 v55, 16, v54
; GFX10-NEXT:    v_cndmask_b32_e32 v52, v55, v52, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v51
; GFX10-NEXT:    v_cndmask_b32_e32 v51, v52, v51, vcc_lo
; GFX10-NEXT:    v_and_b32_e32 v52, 0xffff0000, v54
; GFX10-NEXT:    v_lshrrev_b32_e32 v54, 16, v2
; GFX10-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v52
; GFX10-NEXT:    v_and_b32_e32 v52, 0xffff0000, v2
; GFX10-NEXT:    v_cndmask_b32_e32 v51, v55, v51, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v52, v52
; GFX10-NEXT:    v_lshrrev_b32_e32 v52, 16, v18
; GFX10-NEXT:    v_cndmask_b32_e32 v54, v54, v52, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v64, v64
; GFX10-NEXT:    v_lshlrev_b32_e32 v55, 16, v54
; GFX10-NEXT:    v_cndmask_b32_e32 v52, v52, v54, vcc_lo
; GFX10-NEXT:    v_lshlrev_b32_e32 v64, 16, v52
; GFX10-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v55, v64
; GFX10-NEXT:    v_cndmask_b32_e32 v55, v52, v54, vcc_lo
; GFX10-NEXT:    v_lshlrev_b32_e32 v55, 16, v55
; GFX10-NEXT:    v_max_f32_e32 v55, v55, v55
; GFX10-NEXT:    v_bfe_u32 v64, v55, 16, 1
; GFX10-NEXT:    v_or_b32_e32 v65, 0x400000, v55
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v55, v55
; GFX10-NEXT:    v_add3_u32 v64, v64, v55, 0x7fff
; GFX10-NEXT:    v_cndmask_b32_e32 v55, v64, v65, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v54
; GFX10-NEXT:    v_and_b32_e32 v65, 0xffff0000, v17
; GFX10-NEXT:    v_lshrrev_b32_e32 v64, 16, v55
; GFX10-NEXT:    v_cndmask_b32_e32 v54, v64, v54, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v52
; GFX10-NEXT:    v_cndmask_b32_e32 v52, v54, v52, vcc_lo
; GFX10-NEXT:    v_and_b32_e32 v54, 0xffff0000, v55
; GFX10-NEXT:    v_lshrrev_b32_e32 v55, 16, v1
; GFX10-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v54
; GFX10-NEXT:    v_and_b32_e32 v54, 0xffff0000, v1
; GFX10-NEXT:    v_cndmask_b32_e32 v52, v64, v52, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v54, v54
; GFX10-NEXT:    v_lshrrev_b32_e32 v54, 16, v17
; GFX10-NEXT:    v_cndmask_b32_e32 v55, v55, v54, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v65, v65
; GFX10-NEXT:    v_lshlrev_b32_e32 v64, 16, v55
; GFX10-NEXT:    v_cndmask_b32_e32 v54, v54, v55, vcc_lo
; GFX10-NEXT:    v_lshlrev_b32_e32 v65, 16, v54
; GFX10-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v64, v65
; GFX10-NEXT:    v_cndmask_b32_e32 v64, v54, v55, vcc_lo
; GFX10-NEXT:    v_lshlrev_b32_e32 v64, 16, v64
; GFX10-NEXT:    v_max_f32_e32 v64, v64, v64
; GFX10-NEXT:    v_bfe_u32 v65, v64, 16, 1
; GFX10-NEXT:    v_or_b32_e32 v66, 0x400000, v64
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v64, v64
; GFX10-NEXT:    v_add3_u32 v65, v65, v64, 0x7fff
; GFX10-NEXT:    v_cndmask_b32_e32 v64, v65, v66, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v55
; GFX10-NEXT:    v_and_b32_e32 v66, 0xffff0000, v16
; GFX10-NEXT:    v_lshrrev_b32_e32 v65, 16, v64
; GFX10-NEXT:    v_cndmask_b32_e32 v55, v65, v55, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v54
; GFX10-NEXT:    v_cndmask_b32_e32 v54, v55, v54, vcc_lo
; GFX10-NEXT:    v_and_b32_e32 v55, 0xffff0000, v64
; GFX10-NEXT:    v_lshrrev_b32_e32 v64, 16, v0
; GFX10-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v55
; GFX10-NEXT:    v_and_b32_e32 v55, 0xffff0000, v0
; GFX10-NEXT:    v_cndmask_b32_e32 v54, v65, v54, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v55, v55
; GFX10-NEXT:    v_lshrrev_b32_e32 v55, 16, v16
; GFX10-NEXT:    v_cndmask_b32_e32 v64, v64, v55, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v66, v66
; GFX10-NEXT:    v_lshlrev_b32_e32 v65, 16, v64
; GFX10-NEXT:    v_cndmask_b32_e32 v55, v55, v64, vcc_lo
; GFX10-NEXT:    v_lshlrev_b32_e32 v66, 16, v55
; GFX10-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v65, v66
; GFX10-NEXT:    v_cndmask_b32_e32 v65, v55, v64, vcc_lo
; GFX10-NEXT:    v_lshlrev_b32_e32 v65, 16, v65
; GFX10-NEXT:    v_max_f32_e32 v65, v65, v65
; GFX10-NEXT:    v_bfe_u32 v66, v65, 16, 1
; GFX10-NEXT:    v_or_b32_e32 v67, 0x400000, v65
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v65, v65
; GFX10-NEXT:    v_add3_u32 v66, v66, v65, 0x7fff
; GFX10-NEXT:    v_cndmask_b32_e32 v65, v66, v67, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v64
; GFX10-NEXT:    v_lshrrev_b32_e32 v66, 16, v65
; GFX10-NEXT:    v_cndmask_b32_e32 v64, v66, v64, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v55
; GFX10-NEXT:    v_cndmask_b32_e32 v55, v64, v55, vcc_lo
; GFX10-NEXT:    v_and_b32_e32 v64, 0xffff0000, v65
; GFX10-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v64
; GFX10-NEXT:    v_lshlrev_b32_e32 v64, 16, v15
; GFX10-NEXT:    v_cndmask_b32_e32 v55, v66, v55, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v64, v64
; GFX10-NEXT:    v_lshlrev_b32_e32 v64, 16, v53
; GFX10-NEXT:    v_cndmask_b32_e32 v15, v15, v53, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v64, v64
; GFX10-NEXT:    v_lshlrev_b32_e32 v65, 16, v15
; GFX10-NEXT:    v_cndmask_b32_e32 v53, v53, v15, vcc_lo
; GFX10-NEXT:    v_lshlrev_b32_e32 v64, 16, v53
; GFX10-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v65, v64
; GFX10-NEXT:    v_cndmask_b32_e32 v64, v53, v15, vcc_lo
; GFX10-NEXT:    v_lshlrev_b32_e32 v64, 16, v64
; GFX10-NEXT:    v_max_f32_e32 v64, v64, v64
; GFX10-NEXT:    v_bfe_u32 v65, v64, 16, 1
; GFX10-NEXT:    v_or_b32_e32 v66, 0x400000, v64
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v64, v64
; GFX10-NEXT:    v_add3_u32 v65, v65, v64, 0x7fff
; GFX10-NEXT:    v_cndmask_b32_e32 v64, v65, v66, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v15
; GFX10-NEXT:    v_lshrrev_b32_e32 v65, 16, v64
; GFX10-NEXT:    v_cndmask_b32_e32 v15, v65, v15, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v53
; GFX10-NEXT:    v_cndmask_b32_e32 v15, v15, v53, vcc_lo
; GFX10-NEXT:    v_and_b32_e32 v53, 0xffff0000, v64
; GFX10-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v53
; GFX10-NEXT:    v_lshlrev_b32_e32 v53, 16, v14
; GFX10-NEXT:    v_cndmask_b32_e32 v15, v65, v15, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v53, v53
; GFX10-NEXT:    v_lshlrev_b32_e32 v53, 16, v30
; GFX10-NEXT:    v_perm_b32 v15, v31, v15, 0x5040100
; GFX10-NEXT:    v_cndmask_b32_e32 v14, v14, v30, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v53, v53
; GFX10-NEXT:    v_lshlrev_b32_e32 v64, 16, v14
; GFX10-NEXT:    v_cndmask_b32_e32 v30, v30, v14, vcc_lo
; GFX10-NEXT:    v_lshlrev_b32_e32 v53, 16, v30
; GFX10-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v64, v53
; GFX10-NEXT:    v_cndmask_b32_e32 v53, v30, v14, vcc_lo
; GFX10-NEXT:    v_lshlrev_b32_e32 v53, 16, v53
; GFX10-NEXT:    v_max_f32_e32 v53, v53, v53
; GFX10-NEXT:    v_bfe_u32 v64, v53, 16, 1
; GFX10-NEXT:    v_or_b32_e32 v65, 0x400000, v53
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v53, v53
; GFX10-NEXT:    v_add3_u32 v64, v64, v53, 0x7fff
; GFX10-NEXT:    v_cndmask_b32_e32 v53, v64, v65, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v14
; GFX10-NEXT:    v_lshrrev_b32_e32 v64, 16, v53
; GFX10-NEXT:    v_cndmask_b32_e32 v14, v64, v14, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v30
; GFX10-NEXT:    v_cndmask_b32_e32 v14, v14, v30, vcc_lo
; GFX10-NEXT:    v_and_b32_e32 v30, 0xffff0000, v53
; GFX10-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v30
; GFX10-NEXT:    v_lshlrev_b32_e32 v30, 16, v13
; GFX10-NEXT:    v_cndmask_b32_e32 v14, v64, v14, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v30, v30
; GFX10-NEXT:    v_lshlrev_b32_e32 v30, 16, v29
; GFX10-NEXT:    v_perm_b32 v14, v32, v14, 0x5040100
; GFX10-NEXT:    v_cndmask_b32_e32 v13, v13, v29, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v30, v30
; GFX10-NEXT:    v_lshlrev_b32_e32 v53, 16, v13
; GFX10-NEXT:    v_cndmask_b32_e32 v29, v29, v13, vcc_lo
; GFX10-NEXT:    v_lshlrev_b32_e32 v30, 16, v29
; GFX10-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v53, v30
; GFX10-NEXT:    v_cndmask_b32_e32 v30, v29, v13, vcc_lo
; GFX10-NEXT:    v_lshlrev_b32_e32 v30, 16, v30
; GFX10-NEXT:    v_max_f32_e32 v30, v30, v30
; GFX10-NEXT:    v_bfe_u32 v53, v30, 16, 1
; GFX10-NEXT:    v_or_b32_e32 v64, 0x400000, v30
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v30, v30
; GFX10-NEXT:    v_add3_u32 v53, v53, v30, 0x7fff
; GFX10-NEXT:    v_cndmask_b32_e32 v30, v53, v64, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v13
; GFX10-NEXT:    v_lshrrev_b32_e32 v53, 16, v30
; GFX10-NEXT:    v_cndmask_b32_e32 v13, v53, v13, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v29
; GFX10-NEXT:    v_cndmask_b32_e32 v13, v13, v29, vcc_lo
; GFX10-NEXT:    v_and_b32_e32 v29, 0xffff0000, v30
; GFX10-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v29
; GFX10-NEXT:    v_lshlrev_b32_e32 v29, 16, v12
; GFX10-NEXT:    v_cndmask_b32_e32 v13, v53, v13, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v29, v29
; GFX10-NEXT:    v_lshlrev_b32_e32 v29, 16, v28
; GFX10-NEXT:    v_perm_b32 v13, v33, v13, 0x5040100
; GFX10-NEXT:    v_cndmask_b32_e32 v12, v12, v28, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v29, v29
; GFX10-NEXT:    v_lshlrev_b32_e32 v30, 16, v12
; GFX10-NEXT:    v_cndmask_b32_e32 v28, v28, v12, vcc_lo
; GFX10-NEXT:    v_lshlrev_b32_e32 v29, 16, v28
; GFX10-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v30, v29
; GFX10-NEXT:    v_cndmask_b32_e32 v29, v28, v12, vcc_lo
; GFX10-NEXT:    v_lshlrev_b32_e32 v29, 16, v29
; GFX10-NEXT:    v_max_f32_e32 v29, v29, v29
; GFX10-NEXT:    v_bfe_u32 v30, v29, 16, 1
; GFX10-NEXT:    v_or_b32_e32 v53, 0x400000, v29
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v29, v29
; GFX10-NEXT:    v_add3_u32 v30, v30, v29, 0x7fff
; GFX10-NEXT:    v_cndmask_b32_e32 v29, v30, v53, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v12
; GFX10-NEXT:    v_lshrrev_b32_e32 v30, 16, v29
; GFX10-NEXT:    v_cndmask_b32_e32 v12, v30, v12, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v28
; GFX10-NEXT:    v_cndmask_b32_e32 v12, v12, v28, vcc_lo
; GFX10-NEXT:    v_and_b32_e32 v28, 0xffff0000, v29
; GFX10-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v28
; GFX10-NEXT:    v_lshlrev_b32_e32 v28, 16, v11
; GFX10-NEXT:    v_cndmask_b32_e32 v12, v30, v12, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v28, v28
; GFX10-NEXT:    v_lshlrev_b32_e32 v28, 16, v27
; GFX10-NEXT:    v_perm_b32 v12, v34, v12, 0x5040100
; GFX10-NEXT:    v_cndmask_b32_e32 v11, v11, v27, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v28, v28
; GFX10-NEXT:    v_lshlrev_b32_e32 v29, 16, v11
; GFX10-NEXT:    v_cndmask_b32_e32 v27, v27, v11, vcc_lo
; GFX10-NEXT:    v_lshlrev_b32_e32 v28, 16, v27
; GFX10-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v29, v28
; GFX10-NEXT:    v_cndmask_b32_e32 v28, v27, v11, vcc_lo
; GFX10-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
; GFX10-NEXT:    v_max_f32_e32 v28, v28, v28
; GFX10-NEXT:    v_bfe_u32 v29, v28, 16, 1
; GFX10-NEXT:    v_or_b32_e32 v30, 0x400000, v28
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v28, v28
; GFX10-NEXT:    v_add3_u32 v29, v29, v28, 0x7fff
; GFX10-NEXT:    v_cndmask_b32_e32 v28, v29, v30, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v11
; GFX10-NEXT:    v_lshrrev_b32_e32 v29, 16, v28
; GFX10-NEXT:    v_cndmask_b32_e32 v11, v29, v11, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v27
; GFX10-NEXT:    v_cndmask_b32_e32 v11, v11, v27, vcc_lo
; GFX10-NEXT:    v_and_b32_e32 v27, 0xffff0000, v28
; GFX10-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v27
; GFX10-NEXT:    v_lshlrev_b32_e32 v27, 16, v10
; GFX10-NEXT:    v_cndmask_b32_e32 v11, v29, v11, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v27, v27
; GFX10-NEXT:    v_lshlrev_b32_e32 v27, 16, v26
; GFX10-NEXT:    v_perm_b32 v11, v35, v11, 0x5040100
; GFX10-NEXT:    v_cndmask_b32_e32 v10, v10, v26, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v27, v27
; GFX10-NEXT:    v_lshlrev_b32_e32 v28, 16, v10
; GFX10-NEXT:    v_cndmask_b32_e32 v26, v26, v10, vcc_lo
; GFX10-NEXT:    v_lshlrev_b32_e32 v27, 16, v26
; GFX10-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v28, v27
; GFX10-NEXT:    v_cndmask_b32_e32 v27, v26, v10, vcc_lo
; GFX10-NEXT:    v_lshlrev_b32_e32 v27, 16, v27
; GFX10-NEXT:    v_max_f32_e32 v27, v27, v27
; GFX10-NEXT:    v_bfe_u32 v28, v27, 16, 1
; GFX10-NEXT:    v_or_b32_e32 v29, 0x400000, v27
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v27, v27
; GFX10-NEXT:    v_add3_u32 v28, v28, v27, 0x7fff
; GFX10-NEXT:    v_cndmask_b32_e32 v27, v28, v29, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v10
; GFX10-NEXT:    v_lshrrev_b32_e32 v28, 16, v27
; GFX10-NEXT:    v_cndmask_b32_e32 v10, v28, v10, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v26
; GFX10-NEXT:    v_cndmask_b32_e32 v10, v10, v26, vcc_lo
; GFX10-NEXT:    v_and_b32_e32 v26, 0xffff0000, v27
; GFX10-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v26
; GFX10-NEXT:    v_lshlrev_b32_e32 v26, 16, v9
; GFX10-NEXT:    v_cndmask_b32_e32 v10, v28, v10, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v26, v26
; GFX10-NEXT:    v_lshlrev_b32_e32 v26, 16, v25
; GFX10-NEXT:    v_perm_b32 v10, v36, v10, 0x5040100
; GFX10-NEXT:    v_cndmask_b32_e32 v9, v9, v25, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v26, v26
; GFX10-NEXT:    v_lshlrev_b32_e32 v27, 16, v9
; GFX10-NEXT:    v_cndmask_b32_e32 v25, v25, v9, vcc_lo
; GFX10-NEXT:    v_lshlrev_b32_e32 v26, 16, v25
; GFX10-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v27, v26
; GFX10-NEXT:    v_cndmask_b32_e32 v26, v25, v9, vcc_lo
; GFX10-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
; GFX10-NEXT:    v_max_f32_e32 v26, v26, v26
; GFX10-NEXT:    v_bfe_u32 v27, v26, 16, 1
; GFX10-NEXT:    v_or_b32_e32 v28, 0x400000, v26
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v26, v26
; GFX10-NEXT:    v_add3_u32 v27, v27, v26, 0x7fff
; GFX10-NEXT:    v_cndmask_b32_e32 v26, v27, v28, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v9
; GFX10-NEXT:    v_lshrrev_b32_e32 v27, 16, v26
; GFX10-NEXT:    v_cndmask_b32_e32 v9, v27, v9, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v25
; GFX10-NEXT:    v_cndmask_b32_e32 v9, v9, v25, vcc_lo
; GFX10-NEXT:    v_and_b32_e32 v25, 0xffff0000, v26
; GFX10-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v25
; GFX10-NEXT:    v_lshlrev_b32_e32 v25, 16, v8
; GFX10-NEXT:    v_cndmask_b32_e32 v9, v27, v9, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v25, v25
; GFX10-NEXT:    v_lshlrev_b32_e32 v25, 16, v24
; GFX10-NEXT:    v_perm_b32 v9, v37, v9, 0x5040100
; GFX10-NEXT:    v_cndmask_b32_e32 v8, v8, v24, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v25, v25
; GFX10-NEXT:    v_lshlrev_b32_e32 v26, 16, v8
; GFX10-NEXT:    v_cndmask_b32_e32 v24, v24, v8, vcc_lo
; GFX10-NEXT:    v_lshlrev_b32_e32 v25, 16, v24
; GFX10-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v26, v25
; GFX10-NEXT:    v_cndmask_b32_e32 v25, v24, v8, vcc_lo
; GFX10-NEXT:    v_lshlrev_b32_e32 v25, 16, v25
; GFX10-NEXT:    v_max_f32_e32 v25, v25, v25
; GFX10-NEXT:    v_bfe_u32 v26, v25, 16, 1
; GFX10-NEXT:    v_or_b32_e32 v27, 0x400000, v25
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v25, v25
; GFX10-NEXT:    v_add3_u32 v26, v26, v25, 0x7fff
; GFX10-NEXT:    v_cndmask_b32_e32 v25, v26, v27, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v8
; GFX10-NEXT:    v_lshrrev_b32_e32 v26, 16, v25
; GFX10-NEXT:    v_cndmask_b32_e32 v8, v26, v8, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v24
; GFX10-NEXT:    v_cndmask_b32_e32 v8, v8, v24, vcc_lo
; GFX10-NEXT:    v_and_b32_e32 v24, 0xffff0000, v25
; GFX10-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v24
; GFX10-NEXT:    v_lshlrev_b32_e32 v24, 16, v7
; GFX10-NEXT:    v_cndmask_b32_e32 v8, v26, v8, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v24, v24
; GFX10-NEXT:    v_lshlrev_b32_e32 v24, 16, v23
; GFX10-NEXT:    v_perm_b32 v8, v38, v8, 0x5040100
; GFX10-NEXT:    v_cndmask_b32_e32 v7, v7, v23, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v24, v24
; GFX10-NEXT:    v_lshlrev_b32_e32 v25, 16, v7
; GFX10-NEXT:    v_cndmask_b32_e32 v23, v23, v7, vcc_lo
; GFX10-NEXT:    v_lshlrev_b32_e32 v24, 16, v23
; GFX10-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v25, v24
; GFX10-NEXT:    v_cndmask_b32_e32 v24, v23, v7, vcc_lo
; GFX10-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
; GFX10-NEXT:    v_max_f32_e32 v24, v24, v24
; GFX10-NEXT:    v_bfe_u32 v25, v24, 16, 1
; GFX10-NEXT:    v_or_b32_e32 v26, 0x400000, v24
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v24, v24
; GFX10-NEXT:    v_add3_u32 v25, v25, v24, 0x7fff
; GFX10-NEXT:    v_cndmask_b32_e32 v24, v25, v26, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v7
; GFX10-NEXT:    v_lshrrev_b32_e32 v25, 16, v24
; GFX10-NEXT:    v_cndmask_b32_e32 v7, v25, v7, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v23
; GFX10-NEXT:    v_cndmask_b32_e32 v7, v7, v23, vcc_lo
; GFX10-NEXT:    v_and_b32_e32 v23, 0xffff0000, v24
; GFX10-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v23
; GFX10-NEXT:    v_lshlrev_b32_e32 v23, 16, v6
; GFX10-NEXT:    v_cndmask_b32_e32 v7, v25, v7, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v23, v23
; GFX10-NEXT:    v_lshlrev_b32_e32 v23, 16, v22
; GFX10-NEXT:    v_perm_b32 v7, v39, v7, 0x5040100
; GFX10-NEXT:    v_cndmask_b32_e32 v6, v6, v22, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v23, v23
; GFX10-NEXT:    v_lshlrev_b32_e32 v24, 16, v6
; GFX10-NEXT:    v_cndmask_b32_e32 v22, v22, v6, vcc_lo
; GFX10-NEXT:    v_lshlrev_b32_e32 v23, 16, v22
; GFX10-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v24, v23
; GFX10-NEXT:    v_cndmask_b32_e32 v23, v22, v6, vcc_lo
; GFX10-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; GFX10-NEXT:    v_max_f32_e32 v23, v23, v23
; GFX10-NEXT:    v_bfe_u32 v24, v23, 16, 1
; GFX10-NEXT:    v_or_b32_e32 v25, 0x400000, v23
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v23, v23
; GFX10-NEXT:    v_add3_u32 v24, v24, v23, 0x7fff
; GFX10-NEXT:    v_cndmask_b32_e32 v23, v24, v25, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v6
; GFX10-NEXT:    v_lshrrev_b32_e32 v24, 16, v23
; GFX10-NEXT:    v_cndmask_b32_e32 v6, v24, v6, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v22
; GFX10-NEXT:    v_cndmask_b32_e32 v6, v6, v22, vcc_lo
; GFX10-NEXT:    v_and_b32_e32 v22, 0xffff0000, v23
; GFX10-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v22
; GFX10-NEXT:    v_lshlrev_b32_e32 v22, 16, v5
; GFX10-NEXT:    v_cndmask_b32_e32 v6, v24, v6, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v22, v22
; GFX10-NEXT:    v_lshlrev_b32_e32 v22, 16, v21
; GFX10-NEXT:    v_perm_b32 v6, v48, v6, 0x5040100
; GFX10-NEXT:    v_cndmask_b32_e32 v5, v5, v21, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v22, v22
; GFX10-NEXT:    v_lshlrev_b32_e32 v23, 16, v5
; GFX10-NEXT:    v_cndmask_b32_e32 v21, v21, v5, vcc_lo
; GFX10-NEXT:    v_lshlrev_b32_e32 v22, 16, v21
; GFX10-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v23, v22
; GFX10-NEXT:    v_cndmask_b32_e32 v22, v21, v5, vcc_lo
; GFX10-NEXT:    v_lshlrev_b32_e32 v22, 16, v22
; GFX10-NEXT:    v_max_f32_e32 v22, v22, v22
; GFX10-NEXT:    v_bfe_u32 v23, v22, 16, 1
; GFX10-NEXT:    v_or_b32_e32 v24, 0x400000, v22
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v22, v22
; GFX10-NEXT:    v_add3_u32 v23, v23, v22, 0x7fff
; GFX10-NEXT:    v_cndmask_b32_e32 v22, v23, v24, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v5
; GFX10-NEXT:    v_lshrrev_b32_e32 v23, 16, v22
; GFX10-NEXT:    v_cndmask_b32_e32 v5, v23, v5, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v21
; GFX10-NEXT:    v_cndmask_b32_e32 v5, v5, v21, vcc_lo
; GFX10-NEXT:    v_and_b32_e32 v21, 0xffff0000, v22
; GFX10-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v21
; GFX10-NEXT:    v_lshlrev_b32_e32 v21, 16, v4
; GFX10-NEXT:    v_cndmask_b32_e32 v5, v23, v5, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v21, v21
; GFX10-NEXT:    v_lshlrev_b32_e32 v21, 16, v20
; GFX10-NEXT:    v_perm_b32 v5, v49, v5, 0x5040100
; GFX10-NEXT:    v_cndmask_b32_e32 v4, v4, v20, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v21, v21
; GFX10-NEXT:    v_lshlrev_b32_e32 v22, 16, v4
; GFX10-NEXT:    v_cndmask_b32_e32 v20, v20, v4, vcc_lo
; GFX10-NEXT:    v_lshlrev_b32_e32 v21, 16, v20
; GFX10-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v22, v21
; GFX10-NEXT:    v_cndmask_b32_e32 v21, v20, v4, vcc_lo
; GFX10-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
; GFX10-NEXT:    v_max_f32_e32 v21, v21, v21
; GFX10-NEXT:    v_bfe_u32 v22, v21, 16, 1
; GFX10-NEXT:    v_or_b32_e32 v23, 0x400000, v21
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v21, v21
; GFX10-NEXT:    v_add3_u32 v22, v22, v21, 0x7fff
; GFX10-NEXT:    v_cndmask_b32_e32 v21, v22, v23, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v4
; GFX10-NEXT:    v_lshrrev_b32_e32 v22, 16, v21
; GFX10-NEXT:    v_cndmask_b32_e32 v4, v22, v4, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v20
; GFX10-NEXT:    v_cndmask_b32_e32 v4, v4, v20, vcc_lo
; GFX10-NEXT:    v_and_b32_e32 v20, 0xffff0000, v21
; GFX10-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v20
; GFX10-NEXT:    v_lshlrev_b32_e32 v20, 16, v3
; GFX10-NEXT:    v_cndmask_b32_e32 v4, v22, v4, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v20, v20
; GFX10-NEXT:    v_lshlrev_b32_e32 v20, 16, v19
; GFX10-NEXT:    v_perm_b32 v4, v50, v4, 0x5040100
; GFX10-NEXT:    v_cndmask_b32_e32 v3, v3, v19, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v20, v20
; GFX10-NEXT:    v_lshlrev_b32_e32 v21, 16, v3
; GFX10-NEXT:    v_cndmask_b32_e32 v19, v19, v3, vcc_lo
; GFX10-NEXT:    v_lshlrev_b32_e32 v20, 16, v19
; GFX10-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v21, v20
; GFX10-NEXT:    v_cndmask_b32_e32 v20, v19, v3, vcc_lo
; GFX10-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
; GFX10-NEXT:    v_max_f32_e32 v20, v20, v20
; GFX10-NEXT:    v_bfe_u32 v21, v20, 16, 1
; GFX10-NEXT:    v_or_b32_e32 v22, 0x400000, v20
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v20, v20
; GFX10-NEXT:    v_add3_u32 v21, v21, v20, 0x7fff
; GFX10-NEXT:    v_cndmask_b32_e32 v20, v21, v22, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v3
; GFX10-NEXT:    v_lshrrev_b32_e32 v21, 16, v20
; GFX10-NEXT:    v_cndmask_b32_e32 v3, v21, v3, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v19
; GFX10-NEXT:    v_cndmask_b32_e32 v3, v3, v19, vcc_lo
; GFX10-NEXT:    v_and_b32_e32 v19, 0xffff0000, v20
; GFX10-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v19
; GFX10-NEXT:    v_lshlrev_b32_e32 v19, 16, v2
; GFX10-NEXT:    v_cndmask_b32_e32 v3, v21, v3, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v19, v19
; GFX10-NEXT:    v_lshlrev_b32_e32 v19, 16, v18
; GFX10-NEXT:    v_perm_b32 v3, v51, v3, 0x5040100
; GFX10-NEXT:    v_cndmask_b32_e32 v2, v2, v18, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v19, v19
; GFX10-NEXT:    v_lshlrev_b32_e32 v20, 16, v2
; GFX10-NEXT:    v_cndmask_b32_e32 v18, v18, v2, vcc_lo
; GFX10-NEXT:    v_lshlrev_b32_e32 v19, 16, v18
; GFX10-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v20, v19
; GFX10-NEXT:    v_cndmask_b32_e32 v19, v18, v2, vcc_lo
; GFX10-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
; GFX10-NEXT:    v_max_f32_e32 v19, v19, v19
; GFX10-NEXT:    v_bfe_u32 v20, v19, 16, 1
; GFX10-NEXT:    v_or_b32_e32 v21, 0x400000, v19
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v19, v19
; GFX10-NEXT:    v_add3_u32 v20, v20, v19, 0x7fff
; GFX10-NEXT:    v_cndmask_b32_e32 v19, v20, v21, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v2
; GFX10-NEXT:    v_lshrrev_b32_e32 v20, 16, v19
; GFX10-NEXT:    v_cndmask_b32_e32 v2, v20, v2, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v18
; GFX10-NEXT:    v_cndmask_b32_e32 v2, v2, v18, vcc_lo
; GFX10-NEXT:    v_and_b32_e32 v18, 0xffff0000, v19
; GFX10-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v18
; GFX10-NEXT:    v_lshlrev_b32_e32 v18, 16, v1
; GFX10-NEXT:    v_cndmask_b32_e32 v2, v20, v2, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v18, v18
; GFX10-NEXT:    v_lshlrev_b32_e32 v18, 16, v17
; GFX10-NEXT:    v_perm_b32 v2, v52, v2, 0x5040100
; GFX10-NEXT:    v_cndmask_b32_e32 v1, v1, v17, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v18, v18
; GFX10-NEXT:    v_lshlrev_b32_e32 v19, 16, v1
; GFX10-NEXT:    v_cndmask_b32_e32 v17, v17, v1, vcc_lo
; GFX10-NEXT:    v_lshlrev_b32_e32 v18, 16, v17
; GFX10-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v19, v18
; GFX10-NEXT:    v_cndmask_b32_e32 v18, v17, v1, vcc_lo
; GFX10-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
; GFX10-NEXT:    v_max_f32_e32 v18, v18, v18
; GFX10-NEXT:    v_bfe_u32 v19, v18, 16, 1
; GFX10-NEXT:    v_or_b32_e32 v20, 0x400000, v18
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v18, v18
; GFX10-NEXT:    v_add3_u32 v19, v19, v18, 0x7fff
; GFX10-NEXT:    v_cndmask_b32_e32 v18, v19, v20, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v1
; GFX10-NEXT:    v_lshrrev_b32_e32 v19, 16, v18
; GFX10-NEXT:    v_cndmask_b32_e32 v1, v19, v1, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v17
; GFX10-NEXT:    v_cndmask_b32_e32 v1, v1, v17, vcc_lo
; GFX10-NEXT:    v_and_b32_e32 v17, 0xffff0000, v18
; GFX10-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v17
; GFX10-NEXT:    v_lshlrev_b32_e32 v17, 16, v0
; GFX10-NEXT:    v_cndmask_b32_e32 v1, v19, v1, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v17, v17
; GFX10-NEXT:    v_lshlrev_b32_e32 v17, 16, v16
; GFX10-NEXT:    v_perm_b32 v1, v54, v1, 0x5040100
; GFX10-NEXT:    v_cndmask_b32_e32 v0, v0, v16, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v17, v17
; GFX10-NEXT:    v_lshlrev_b32_e32 v18, 16, v0
; GFX10-NEXT:    v_cndmask_b32_e32 v16, v16, v0, vcc_lo
; GFX10-NEXT:    v_lshlrev_b32_e32 v17, 16, v16
; GFX10-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v18, v17
; GFX10-NEXT:    v_cndmask_b32_e32 v17, v16, v0, vcc_lo
; GFX10-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
; GFX10-NEXT:    v_max_f32_e32 v17, v17, v17
; GFX10-NEXT:    v_bfe_u32 v18, v17, 16, 1
; GFX10-NEXT:    v_or_b32_e32 v19, 0x400000, v17
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v17, v17
; GFX10-NEXT:    v_add3_u32 v18, v18, v17, 0x7fff
; GFX10-NEXT:    v_cndmask_b32_e32 v17, v18, v19, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v0
; GFX10-NEXT:    v_lshrrev_b32_e32 v18, 16, v17
; GFX10-NEXT:    v_cndmask_b32_e32 v0, v18, v0, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v16
; GFX10-NEXT:    v_cndmask_b32_e32 v0, v0, v16, vcc_lo
; GFX10-NEXT:    v_and_b32_e32 v16, 0xffff0000, v17
; GFX10-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v16
; GFX10-NEXT:    v_cndmask_b32_e32 v0, v18, v0, vcc_lo
; GFX10-NEXT:    v_perm_b32 v0, v55, v0, 0x5040100
; GFX10-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-TRUE16-LABEL: v_maximumnum_v32bf16:
; GFX11-TRUE16:       ; %bb.0:
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT:    scratch_load_b32 v51, off, s32
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v81, 0xffff0000, v2
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v52, 0xffff0000, v25
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v82, 0xffff0000, v18
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v69, 0xffff0000, v4
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v70, 0xffff0000, v20
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s24, v81, v81
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s11, v52, v52
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s25, v82, v82
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s20, v69, v69
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v80, 0xffff0000, v19
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v52.l, v2.h, v18.h, s24
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s21, v70, v70
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v50, v15 :: v_dual_mov_b32 v49, v14
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v32, 0xffff0000, v13
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v82.l, v18.h, v52.l, s25
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v112.l, v52.l
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v38, 0xffff0000, v10
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s23, v80, v80
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v15, 0xffff0000, v49
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v134.l, v82.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v112, 16, v112
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v33, 0xffff0000, v29
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s2, v32, v32
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s8, v38, v38
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v134, 16, v134
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v55, 0xffff0000, v7
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v66, 0xffff0000, v22
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v67, 0xffff0000, v5
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s0, v15, v15
; GFX11-TRUE16-NEXT:    v_cmp_gt_f32_e64 s44, v112, v134
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s3, v33, v33
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v15.l, v13.h, v29.h, s2
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v33.l, v10.h, v26.h, s8
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v68, 0xffff0000, v21
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v112.l, v82.l, v52.l, s44
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v71, 0xffff0000, v3
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s14, v55, v55
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s17, v66, v66
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s18, v67, v67
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v112, 16, v112
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v39, 0xffff0000, v26
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v55.l, v29.h, v15.l, s3
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v48, 0xffff0000, v9
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v83, 0xffff0000, v1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s19, v68, v68
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s9, v39, v39
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v39.l, v4.h, v20.h, s20
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s22, v71, v71
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v38.l, v5.h, v21.h, s18
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v85.l, v15.l
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v66.l, v26.h, v33.l, s9
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v80.l, v20.h, v39.l, s21
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v102.l, v39.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v96.l, v33.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v115.l, v55.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v118.l, v66.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v132.l, v80.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v102, 16, v102
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s10, v48, v48
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s26, v83, v83
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v48.l, v3.h, v19.h, s22
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v132, 16, v132
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v71.l, v21.h, v38.l, s19
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v85, 16, v85
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v96, 16, v96
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v115, 16, v115
; GFX11-TRUE16-NEXT:    v_cmp_gt_f32_e64 s42, v102, v132
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v118, 16, v118
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v101.l, v38.l
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v81.l, v19.h, v48.l, s23
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v131.l, v71.l
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v102.l, v80.l, v39.l, s42
; GFX11-TRUE16-NEXT:    v_cmp_gt_f32_e64 s23, v85, v115
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v101, 16, v101
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v34, 0xffff0000, v12
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v131, 16, v131
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v102, 16, v102
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v53, 0xffff0000, v8
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v85.l, v55.l, v15.l, s23
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v54, 0xffff0000, v24
; GFX11-TRUE16-NEXT:    v_cmp_gt_f32_e64 s41, v101, v131
; GFX11-TRUE16-NEXT:    v_max_f32_e32 v102, v102, v102
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s12, v53, v53
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v53.l, v1.h, v17.h, s26
; GFX11-TRUE16-NEXT:    v_cmp_gt_f32_e64 s26, v96, v118
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v85, 16, v85
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xffff0000, v50
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v101.l, v71.l, v38.l, s41
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v35, 0xffff0000, v28
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v96.l, v66.l, v33.l, s26
; GFX11-TRUE16-NEXT:    v_max_f32_e32 v85, v85, v85
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v14, v14
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v14.l, v49.h, v30.h, s0
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v84, 0xffff0000, v17
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v96, 16, v96
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v31, 0xffff0000, v30
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s4, v34, v34
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s13, v54, v54
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_dual_max_f32 v96, v96, v96 :: v_dual_lshlrev_b32 v101, 16, v101
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s1, v31, v31
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v36, 0xffff0000, v11
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v37, 0xffff0000, v27
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v64, 0xffff0000, v23
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s5, v35, v35
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v54.l, v30.h, v14.l, s1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s27, v84, v84
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v31.l, v12.h, v28.h, s4
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v84.l, v14.l
; GFX11-TRUE16-NEXT:    v_max_f32_e32 v101, v101, v101
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s6, v36, v36
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v114.l, v54.l
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v65, 0xffff0000, v6
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s7, v37, v37
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s15, v64, v64
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v32.l, v11.h, v27.h, s6
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v64.l, v28.h, v31.l, s5
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v84, 16, v84
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v114, 16, v114
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s16, v65, v65
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v86.l, v31.l
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v65.l, v27.h, v32.l, s7
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v116.l, v64.l
; GFX11-TRUE16-NEXT:    v_cmp_gt_f32_e64 s22, v84, v114
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v34.l, v9.h, v25.h, s10
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v87.l, v32.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v86, 16, v86
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v117.l, v65.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v116, 16, v116
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v84.l, v54.l, v14.l, s22
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v35.l, v8.h, v24.h, s12
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v67.l, v25.h, v34.l, s11
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v87, 16, v87
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v117, 16, v117
; GFX11-TRUE16-NEXT:    v_cmp_gt_f32_e64 s24, v86, v116
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v84, 16, v84
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v36.l, v7.h, v23.h, s14
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v97.l, v34.l
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v68.l, v24.h, v35.l, s13
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v119.l, v67.l
; GFX11-TRUE16-NEXT:    v_cmp_gt_f32_e64 s25, v87, v117
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v86.l, v64.l, v31.l, s24
; GFX11-TRUE16-NEXT:    v_max_f32_e32 v84, v84, v84
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v37.l, v6.h, v22.h, s16
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v98.l, v35.l
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v69.l, v23.h, v36.l, s15
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v97, 16, v97
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v128.l, v68.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v119, 16, v119
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v87.l, v65.l, v32.l, s25
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v86, 16, v86
; GFX11-TRUE16-NEXT:    v_bfe_u32 v114, v84, 16, 1
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v99.l, v36.l
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v70.l, v22.h, v37.l, s17
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v83.l, v17.h, v53.l, s27
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v98, 16, v98
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v129.l, v69.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v128, 16, v128
; GFX11-TRUE16-NEXT:    v_cmp_gt_f32_e64 s27, v97, v119
; GFX11-TRUE16-NEXT:    v_dual_max_f32 v86, v86, v86 :: v_dual_lshlrev_b32 v87, 16, v87
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v115, 0x400000, v84
; GFX11-TRUE16-NEXT:    v_bfe_u32 v116, v85, 16, 1
; GFX11-TRUE16-NEXT:    v_add3_u32 v114, v114, v84, 0x7fff
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s22, v84, v84
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v100.l, v37.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v99, 16, v99
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v130.l, v70.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v129, 16, v129
; GFX11-TRUE16-NEXT:    v_cmp_gt_f32_e64 s28, v98, v128
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v97.l, v67.l, v34.l, s27
; GFX11-TRUE16-NEXT:    v_max_f32_e32 v87, v87, v87
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v117, 0x400000, v85
; GFX11-TRUE16-NEXT:    v_bfe_u32 v118, v86, 16, 1
; GFX11-TRUE16-NEXT:    v_add3_u32 v116, v116, v85, 0x7fff
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e64 v114, v114, v115, s22
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s22, v85, v85
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v100, 16, v100
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v130, 16, v130
; GFX11-TRUE16-NEXT:    v_cmp_gt_f32_e64 s29, v99, v129
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v98.l, v68.l, v35.l, s28
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v97, 16, v97
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v119, 0x400000, v86
; GFX11-TRUE16-NEXT:    v_bfe_u32 v128, v87, 16, 1
; GFX11-TRUE16-NEXT:    v_add3_u32 v118, v118, v86, 0x7fff
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e64 v116, v116, v117, s22
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s22, v86, v86
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v103.l, v48.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v133.l, v81.l
; GFX11-TRUE16-NEXT:    v_cmp_gt_f32_e64 s40, v100, v130
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v99.l, v69.l, v36.l, s29
; GFX11-TRUE16-NEXT:    v_dual_max_f32 v97, v97, v97 :: v_dual_lshlrev_b32 v98, 16, v98
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v129, 0x400000, v87
; GFX11-TRUE16-NEXT:    v_bfe_u32 v130, v96, 16, 1
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e64 v118, v118, v119, s22
; GFX11-TRUE16-NEXT:    v_add3_u32 v128, v128, v87, 0x7fff
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s22, v87, v87
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v103, 16, v103
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v133, 16, v133
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v100.l, v70.l, v37.l, s40
; GFX11-TRUE16-NEXT:    v_dual_max_f32 v98, v98, v98 :: v_dual_lshlrev_b32 v99, 16, v99
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v131, 0x400000, v96
; GFX11-TRUE16-NEXT:    v_bfe_u32 v132, v97, 16, 1
; GFX11-TRUE16-NEXT:    v_add3_u32 v130, v130, v96, 0x7fff
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e64 v128, v128, v129, s22
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s22, v96, v96
; GFX11-TRUE16-NEXT:    v_cmp_gt_f32_e64 s43, v103, v133
; GFX11-TRUE16-NEXT:    v_dual_max_f32 v99, v99, v99 :: v_dual_lshlrev_b32 v100, 16, v100
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v133, 0x400000, v97
; GFX11-TRUE16-NEXT:    v_bfe_u32 v134, v98, 16, 1
; GFX11-TRUE16-NEXT:    v_add3_u32 v132, v132, v97, 0x7fff
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e64 v96, v130, v131, s22
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s22, v97, v97
; GFX11-TRUE16-NEXT:    v_max_f32_e32 v100, v100, v100
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v144, 0x400000, v98
; GFX11-TRUE16-NEXT:    v_bfe_u32 v145, v99, 16, 1
; GFX11-TRUE16-NEXT:    v_add3_u32 v134, v134, v98, 0x7fff
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e64 v97, v132, v133, s22
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s22, v98, v98
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v146, 0x400000, v99
; GFX11-TRUE16-NEXT:    v_bfe_u32 v147, v100, 16, 1
; GFX11-TRUE16-NEXT:    v_add3_u32 v145, v145, v99, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v84, 0x400000, v100
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e64 v98, v134, v144, s22
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s22, v99, v99
; GFX11-TRUE16-NEXT:    v_bfe_u32 v115, v101, 16, 1
; GFX11-TRUE16-NEXT:    v_add3_u32 v147, v147, v100, 0x7fff
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s1, 0, v15.l
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s5, 0, v31.l
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e64 v99, v145, v146, s22
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s22, v100, v100
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v85, 0x400000, v101
; GFX11-TRUE16-NEXT:    v_add3_u32 v115, v115, v101, 0x7fff
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s2, 0, v32.l
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v14.h, v116.h, v15.l, s1
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e64 v84, v147, v84, s22
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s22, v101, v101
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v15.l, v118.h, v31.l, s5
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s7, 0, v34.l
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v15.h, v128.h, v32.l, s2
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s6, 0, v33.l
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s3, 0, v35.l
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s8, 0, v36.l
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v32.h, v97.h, v34.l, s7
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s4, 0, v37.l
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v31.h, v96.h, v33.l, s6
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v33.l, v98.h, v35.l, s3
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v33.h, v99.h, v36.l, s8
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s0, 0, v14.l
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s9, 0, v38.l
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e64 v85, v115, v85, s22
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s11, 0, v54.l
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s12, 0, v55.l
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v14.l, v114.h, v14.l, s0
; GFX11-TRUE16-NEXT:    v_bfe_u32 v117, v102, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s13, 0, v64.l
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v86, 0x400000, v102
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s22, v102, v102
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v14.l, v14.l, v54.l, s11
; GFX11-TRUE16-NEXT:    v_add3_u32 v117, v117, v102, 0x7fff
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s10, 0, v39.l
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v100, 0xffff0000, v114
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v15.l, v15.l, v64.l, s13
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s14, 0, v65.l
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e64 v86, v117, v86, s22
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v115, 0xffff0000, v128
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v113.l, v53.l
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s15, 0, v66.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v135.l, v83.l
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v103.l, v81.l, v48.l, s43
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v102, 0xffff0000, v116
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v15.h, v15.h, v65.l, s14
; GFX11-TRUE16-NEXT:    v_cmp_eq_f32_e64 s2, 0, v115
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s20, 0, v71.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v113, 16, v113
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v103, 16, v103
; GFX11-TRUE16-NEXT:    v_cmp_eq_f32_e64 s0, 0, v102
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s16, 0, v67.l
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s17, 0, v68.l
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v131, 0xffff0000, v98
; GFX11-TRUE16-NEXT:    v_max_f32_e32 v103, v103, v103
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s19, 0, v70.l
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v132, 0xffff0000, v84
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v33.l, v33.l, v68.l, s17
; GFX11-TRUE16-NEXT:    v_cmp_eq_f32_e64 s5, 0, v131
; GFX11-TRUE16-NEXT:    v_bfe_u32 v119, v103, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v87, 0x400000, v103
; GFX11-TRUE16-NEXT:    v_cmp_eq_f32_e64 s7, 0, v132
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v117, 0xffff0000, v96
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s21, 0, v80.l
; GFX11-TRUE16-NEXT:    v_add3_u32 v119, v119, v103, 0x7fff
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v129, 0xffff0000, v97
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s18, 0, v69.l
; GFX11-TRUE16-NEXT:    v_cmp_eq_f32_e64 s3, 0, v117
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v133, 0xffff0000, v85
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT:    v_cmp_eq_f32_e64 s8, 0, v133
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v101, 0xffff0000, v51
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v31.l, v50.h, v51.h, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v101, v101
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.l, v31.l
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v101, 0xffff0000, v118
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v32.l, v51.h, v31.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v35, 16, v34
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v34.l, v84.h, v37.l, s4
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v34.h, v85.h, v38.l, s9
; GFX11-TRUE16-NEXT:    v_cmp_eq_f32_e64 s1, 0, v101
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v130.l, v32.l
; GFX11-TRUE16-NEXT:    v_cmp_eq_f32_e64 s4, 0, v129
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v34.l, v34.l, v70.l, s19
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v36, 16, v130
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v130, 0xffff0000, v99
; GFX11-TRUE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v35, v36
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v35.h, v14.h, v55.l, s12
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v35.l, v86.h, v39.l, s10
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v39, 0xffff0000, v86
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s10, 0, v31.l
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v36.l, v32.l, v31.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v100
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v36.h, v118.h, v15.l, s1
; GFX11-TRUE16-NEXT:    v_cmp_eq_f32_e64 s9, 0, v39
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v35.l, v35.l, v80.l, s21
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v37.l, v36.l
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v14.h, v114.h, v14.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v32.l
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v36.l, v31.h, v66.l, s15
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v31.h, v128.h, v15.h, s2
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v38, 16, v37
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v37.l, v32.h, v67.l, s16
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v48.h, v86.h, v35.l, s9
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v32.h, v96.h, v36.l, s3
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v37.h, v33.h, v69.l, s18
; GFX11-TRUE16-NEXT:    v_max_f32_e32 v54, v38, v38
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v38.l, v34.h, v71.l, s20
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v34.h, v116.h, v35.h, s0
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v35.h, v98.h, v33.l, s5
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v38.h, v84.h, v34.l, s7
; GFX11-TRUE16-NEXT:    v_bfe_u32 v55, v54, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v64, 0x400000, v54
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s11, v54, v54
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v33.h, v97.h, v37.l, s4
; GFX11-TRUE16-NEXT:    v_cmp_eq_f32_e64 s6, 0, v130
; GFX11-TRUE16-NEXT:    v_add3_u32 v55, v55, v54, 0x7fff
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v37.h, v99.h, v37.h, s6
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e64 v54, v55, v64, s11
; GFX11-TRUE16-NEXT:    v_max_f32_e32 v64, v112, v112
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v14.l, v54.h, v31.l, s10
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v39, 0xffff0000, v54
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v65, v64, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v66, 0x400000, v64
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s1, v64, v64
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v14.l, v14.l, v32.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v39
; GFX11-TRUE16-NEXT:    v_add3_u32 v65, v65, v64, 0x7fff
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v39.h, v85.h, v38.l, s8
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v15.h, v54.h, v14.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v54, 16, v135
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v103, v103
; GFX11-TRUE16-NEXT:    v_cmp_gt_f32_e64 s0, v113, v54
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v55, v119, v87 :: v_dual_and_b32 v54, 0xffff0000, v0
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v48.l
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v15.l, v83.l, v53.l, s0
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s0, v54, v54
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v54, 0xffff0000, v16
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v14.l, v55.h, v48.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v81.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v67.l, v15.l
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v15.l, v0.h, v16.h, s0
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s0, v54, v54
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e64 v54, v65, v66, s1
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s1, 0, v52.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v64, 16, v67
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v65, 0xffff0000, v55
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v31.l, v16.h, v15.l, s0
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v14.l, v14.l, v81.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v32.l, v54.h, v52.l, s1
; GFX11-TRUE16-NEXT:    v_max_f32_e32 v64, v64, v64
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v52.l, v15.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v66.l, v31.l
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v82.l
; GFX11-TRUE16-NEXT:    v_cmp_eq_f32_e64 s0, 0, v65
; GFX11-TRUE16-NEXT:    v_bfe_u32 v67, v64, 16, 1
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v52, 16, v52
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v65, 16, v66
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v32.l, v32.l, v82.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s1, v64, v64
; GFX11-TRUE16-NEXT:    v_add3_u32 v66, v67, v64, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v67, 0x400000, v64
; GFX11-TRUE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v52, v65
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v52, 16, v50
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v68, 0xffff0000, v54
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v50.h, v55.h, v14.l, s0
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e64 v64, v66, v67, s1
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s1, 0, v53.l
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v33.l, v31.l, v15.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v52, v52
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v52, 16, v51
; GFX11-TRUE16-NEXT:    v_cmp_eq_f32_e64 s2, 0, v68
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v34.l, v64.h, v53.l, s1
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v53.l, v33.l
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v33.l, v50.l, v51.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v52, v52
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s1, 0, v83.l
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v51.h, v54.h, v32.l, s2
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v52, 16, v53
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v53, 16, v49
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v14.l, v51.l, v33.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v65.l, v33.l
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v32.l, v34.l, v83.l, s1
; GFX11-TRUE16-NEXT:    v_max_f32_e32 v52, v52, v52
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v53, v53
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v55.l, v14.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v53, 16, v30
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v54, 0xffff0000, v64
; GFX11-TRUE16-NEXT:    v_bfe_u32 v66, v52, 16, 1
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v65, 16, v65
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v55, 16, v55
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v34.l, v49.l, v30.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v53, v53
; GFX11-TRUE16-NEXT:    v_cmp_eq_f32_e64 s0, 0, v54
; GFX11-TRUE16-NEXT:    v_add3_u32 v49, v66, v52, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v53, 0x400000, v52
; GFX11-TRUE16-NEXT:    v_cmp_gt_f32_e64 s1, v65, v55
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v30.l, v30.l, v34.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v52, v52
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v54, 16, v13
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v30.h, v64.h, v32.l, s0
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v32.l, v14.l, v33.l, s1
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v52.l, v30.l
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v49, v49, v53, vcc_lo
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v53.l, v34.l
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s0, v54, v54
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v54, 16, v29
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v15.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v55.l, v32.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v52, 16, v52
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v53, 16, v53
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v13.l, v13.l, v29.l, s0
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s0, v54, v54
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v13.h, v49.h, v15.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v55, 16, v55
; GFX11-TRUE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v53, v52
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s1, 0, v31.l
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v29.l, v29.l, v13.l, s0
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_dual_max_f32 v52, v55, v55 :: v_dual_and_b32 v53, 0xffff0000, v49
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v15.l, v30.l, v34.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v64.l, v13.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v55.l, v29.l
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v13.h, v13.h, v31.l, s1
; GFX11-TRUE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v53
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v53.l, v15.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v64, 16, v64
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v55, 16, v55
; GFX11-TRUE16-NEXT:    v_bfe_u32 v54, v52, 16, 1
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v13.h, v49.h, v13.h, vcc_lo
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v53, 16, v53
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v49, 16, v12
; GFX11-TRUE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v64, v55
; GFX11-TRUE16-NEXT:    v_add3_u32 v54, v54, v52, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v65, 0x400000, v52
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s0, v52, v52
; GFX11-TRUE16-NEXT:    v_max_f32_e32 v53, v53, v53
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v15.l, v29.l, v13.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v49, v49
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v49, 16, v28
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e64 v52, v54, v65, s0
; GFX11-TRUE16-NEXT:    v_bfe_u32 v54, v53, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s0, 0, v33.l
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v12.l, v12.l, v28.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v49, v49
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v55, 0x400000, v53
; GFX11-TRUE16-NEXT:    v_add3_u32 v54, v54, v53, 0x7fff
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v64.l, v15.l
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s1, v53, v53
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v28.l, v28.l, v12.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v12.h, v52.h, v33.l, s0
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s0, 0, v14.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v53, 16, v64
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e64 v49, v54, v55, s1
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s1, 0, v34.l
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v54, 0xffff0000, v52
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v55.l, v28.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v64.l, v12.l
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v12.h, v12.h, v14.l, s0
; GFX11-TRUE16-NEXT:    v_max_f32_e32 v53, v53, v53
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v14.l, v49.h, v34.l, s1
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v30.l
; GFX11-TRUE16-NEXT:    v_cmp_eq_f32_e64 s0, 0, v54
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v54, 16, v55
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v55, 16, v64
; GFX11-TRUE16-NEXT:    v_bfe_u32 v65, v53, 16, 1
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v14.l, v14.l, v30.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v66, 0xffff0000, v49
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s1, v53, v53
; GFX11-TRUE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v55, v54
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v54, 16, v11
; GFX11-TRUE16-NEXT:    v_add3_u32 v64, v65, v53, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v65, 0x400000, v53
; GFX11-TRUE16-NEXT:    v_cmp_eq_f32_e64 s2, 0, v66
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v15.l, v28.l, v12.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v54, v54
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v54, 16, v27
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e64 v53, v64, v65, s1
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s1, 0, v13.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v55.l, v15.l
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v11.l, v11.l, v27.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v54, v54
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v14.l, v49.h, v14.l, s2
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v11.h, v53.h, v13.l, s1
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s1, 0, v29.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v49, 16, v55
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v13.l, v27.l, v11.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v27, 16, v10
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v54.l, v11.l
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v11.h, v11.h, v29.l, s1
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v29, 0xffff0000, v53
; GFX11-TRUE16-NEXT:    v_max_f32_e32 v49, v49, v49
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v52.l, v13.l
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v27, v27
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v27, 16, v26
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v15.l, v52.h, v12.h, s0
; GFX11-TRUE16-NEXT:    v_bfe_u32 v55, v49, 16, 1
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v52, 16, v52
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v54, 16, v54
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v10.l, v10.l, v26.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v27, v27
; GFX11-TRUE16-NEXT:    v_cmp_eq_f32_e64 s0, 0, v29
; GFX11-TRUE16-NEXT:    v_add3_u32 v27, v55, v49, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v29, 0x400000, v49
; GFX11-TRUE16-NEXT:    v_cmp_gt_f32_e64 s1, v54, v52
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v26.l, v26.l, v10.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v34.l, v53.h, v11.h, s0
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v49, v49
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v53, 16, v9
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v52.l, v10.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v49.l, v26.l
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v29, v27, v29, vcc_lo
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v27.l, v13.l, v11.l, s1
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v12.l
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s0, v53, v53
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v53, 16, v25
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v49, 16, v49
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v52, 16, v52
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v9.h, v29.h, v12.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 16, v27
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v9.l, v9.l, v25.l, s0
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s0, v53, v53
; GFX11-TRUE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v52, v49
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s1, 0, v28.l
; GFX11-TRUE16-NEXT:    v_max_f32_e32 v27, v12, v12
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v49, 0xffff0000, v29
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v25.l, v25.l, v9.l, s0
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v12.l, v26.l, v10.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v9.h, v9.h, v28.l, s1
; GFX11-TRUE16-NEXT:    v_bfe_u32 v28, v27, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v49
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v49.l, v25.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v52.l, v9.l
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v53, 0x400000, v27
; GFX11-TRUE16-NEXT:    v_add3_u32 v28, v28, v27, 0x7fff
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v49, 16, v49
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v52, 16, v52
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s0, v27, v27
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v36.l, v29.h, v9.h, vcc_lo
; GFX11-TRUE16-NEXT:    v_max_f32_e32 v12, v12, v12
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v52, v49
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e64 v27, v28, v53, s0
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s0, 0, v11.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v28, 16, v8
; GFX11-TRUE16-NEXT:    v_bfe_u32 v29, v12, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v49, 0x400000, v12
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s1, v12, v12
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v8.h, v27.h, v11.l, s0
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v11.l, v25.l, v9.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v28, v28
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v28, 16, v24
; GFX11-TRUE16-NEXT:    v_add3_u32 v29, v29, v12, 0x7fff
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s0, 0, v13.l
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v8.l, v8.l, v24.l, vcc_lo
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v28, v28
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e64 v12, v29, v49, s1
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s1, 0, v10.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v28, 16, v11
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v29.l, v8.l
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v11.l, v24.l, v8.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v24, 0xffff0000, v27
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v9.h, v12.h, v10.l, s1
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v26.l
; GFX11-TRUE16-NEXT:    v_max_f32_e32 v10, v28, v28
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v28.l, v11.l
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v8.h, v8.h, v13.l, s0
; GFX11-TRUE16-NEXT:    v_cmp_eq_f32_e64 s0, 0, v24
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v9.h, v9.h, v26.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v24, v10, 16, 1
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v26, 16, v28
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v28, 16, v29
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v29, 16, v7
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v52, 0x400000, v10
; GFX11-TRUE16-NEXT:    v_add3_u32 v24, v24, v10, 0x7fff
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s2, v10, v10
; GFX11-TRUE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v28, v26
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s1, v29, v29
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v26, 16, v23
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v49, 0xffff0000, v12
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e64 v24, v24, v52, s2
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s2, 0, v9.l
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v10.l, v11.l, v8.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v7.l, v7.l, v23.l, s1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v26, v26
; GFX11-TRUE16-NEXT:    v_cmp_eq_f32_e64 s1, 0, v49
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v7.h, v24.h, v9.l, s2
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s2, 0, v25.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v9.l, v23.l, v7.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.l, v7.l
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v31.l, v27.h, v8.h, s0
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v7.h, v7.h, v25.l, s2
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v25, 16, v6
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v12.l, v9.l
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v32.l, v12.h, v9.h, s1
; GFX11-TRUE16-NEXT:    v_dual_max_f32 v27, v10, v10 :: v_dual_and_b32 v26, 0xffff0000, v24
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v25, v25
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v10, 16, v12
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 16, v23
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v23, 16, v22
; GFX11-TRUE16-NEXT:    v_bfe_u32 v25, v27, 16, 1
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v6.l, v6.l, v22.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_eq_f32_e64 s0, 0, v26
; GFX11-TRUE16-NEXT:    v_cmp_gt_f32_e64 s1, v12, v10
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v23, v23
; GFX11-TRUE16-NEXT:    v_add3_u32 v23, v25, v27, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v25, 0x400000, v27
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v24.l, v6.l
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v10.l, v9.l, v7.l, s1
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v12.l, v22.l, v6.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v27, v27
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v33.l, v24.h, v7.h, s0
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s0, 0, v11.l
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v22, v23, v25, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v8.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.l, v12.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v25, 16, v5
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v5.h, v22.h, v8.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 16, v10
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v10, 16, v23
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v23, 16, v24
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v25, v25
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_dual_max_f32 v25, v8, v8 :: v_dual_lshlrev_b32 v24, 16, v21
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v5.h, v5.h, v11.l, s0
; GFX11-TRUE16-NEXT:    v_cmp_gt_f32_e64 s1, v23, v10
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v5.l, v5.l, v21.l, vcc_lo
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v24, v24
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v23, 0xffff0000, v22
; GFX11-TRUE16-NEXT:    v_bfe_u32 v24, v25, 16, 1
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v8.l, v12.l, v6.l, s1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s0, v25, v25
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v10.l, v21.l, v5.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v23
; GFX11-TRUE16-NEXT:    v_add3_u32 v11, v24, v25, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v21, 0x400000, v25
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v24.l, v5.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.l, v10.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v35.l, v22.h, v5.h, vcc_lo
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e64 v11, v11, v21, s0
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s0, 0, v7.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v21, 16, v23
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v23, 16, v24
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v24, 16, v4
; GFX11-TRUE16-NEXT:    v_max_f32_e32 v8, v8, v8
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v4.h, v11.h, v7.l, s0
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s1, 0, v9.l
; GFX11-TRUE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v23, v21
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s0, v24, v24
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v21, 16, v20
; GFX11-TRUE16-NEXT:    v_bfe_u32 v22, v8, 16, 1
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v4.h, v4.h, v9.l, s1
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v7.l, v10.l, v5.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v4.l, v4.l, v20.l, s0
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v21, v21
; GFX11-TRUE16-NEXT:    v_add3_u32 v9, v22, v8, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v21, 0x400000, v8
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v22.l, v7.l
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v7.l, v20.l, v4.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v20, 0xffff0000, v11
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v8, v9, v21 :: v_dual_lshlrev_b32 v9, 16, v22
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.l, v7.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v22.l, v4.l
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v6.l
; GFX11-TRUE16-NEXT:    v_cmp_eq_f32_e64 s0, 0, v20
; GFX11-TRUE16-NEXT:    v_max_f32_e32 v9, v9, v9
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v20, 16, v21
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v21, 16, v22
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v5.h, v8.h, v6.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v12.l
; GFX11-TRUE16-NEXT:    v_bfe_u32 v6, v9, 16, 1
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v37.l, v11.h, v4.h, s0
; GFX11-TRUE16-NEXT:    v_cmp_gt_f32_e64 s0, v21, v20
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v11, 16, v3
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v3.h, v5.h, v12.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_add3_u32 v12, v6, v9, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v20, 0x400000, v9
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v6.l, v7.l, v4.l, s0
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v11, v11
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v11, 16, v19
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s0, v9, v9
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v3.l, v3.l, v19.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v11, v11
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e64 v9, v12, v20, s0
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s0, 0, v5.l
; GFX11-TRUE16-NEXT:    v_dual_max_f32 v6, v6, v6 :: v_dual_and_b32 v11, 0xffff0000, v8
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 16, v2
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v4.h, v9.h, v5.l, s0
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s0, 0, v10.l
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v5.l, v19.l, v3.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v11
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xffff0000, v9
; GFX11-TRUE16-NEXT:    v_bfe_u32 v19, v6, 16, 1
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v4.h, v4.h, v10.l, s0
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v10.l, v5.l
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s1, v12, v12
; GFX11-TRUE16-NEXT:    v_cmp_eq_f32_e64 s0, 0, v11
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v11.l, v3.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 16, v18
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
; GFX11-TRUE16-NEXT:    v_add3_u32 v19, v19, v6, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v20, 0x400000, v6
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v2.l, v2.l, v18.l, s1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s1, v12, v12
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s3, v6, v6
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v39.l, v9.h, v4.h, s0
; GFX11-TRUE16-NEXT:    v_cmp_gt_f32_e64 s2, v11, v10
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 16, v1
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v6.l, v18.l, v2.l, s1
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e64 v11, v19, v20, s3
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s3, 0, v4.l
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v38.l, v8.h, v3.h, vcc_lo
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v10.l, v2.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 16, v0
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v2.h, v11.h, v4.l, s3
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, v6.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 16, v17
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s0, v12, v12
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 16, v16
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v1.l, v1.l, v17.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v8.l, v5.l, v3.l, s2
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v0.l, v0.l, v16.l, s0
; GFX11-TRUE16-NEXT:    v_cmp_gt_f32_e64 s1, v10, v4
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s0, v12, v12
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v4.l, v17.l, v1.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v10, 16, v8
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.l, v0.l
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v8.l, v6.l, v2.l, s1
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v9.l, v16.l, v0.l, s0
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v12.l, v4.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v16.l, v1.l
; GFX11-TRUE16-NEXT:    v_max_f32_e32 v10, v10, v10
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v17.l, v9.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
; GFX11-TRUE16-NEXT:    v_bfe_u32 v19, v10, 16, 1
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v7.l
; GFX11-TRUE16-NEXT:    v_cmp_gt_f32_e64 s0, v16, v12
; GFX11-TRUE16-NEXT:    v_add3_u32 v12, v19, v10, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v16, 0x400000, v10
; GFX11-TRUE16-NEXT:    v_cmp_gt_f32_e64 s1, v18, v17
; GFX11-TRUE16-NEXT:    v_max_f32_e32 v17, v8, v8
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v8.l, v4.l, v1.l, s0
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s0, v10, v10
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v0.h, v2.h, v7.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v10.l, v9.l, v0.l, s1
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v5.l
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e64 v12, v12, v16, s0
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s0, 0, v3.l
; GFX11-TRUE16-NEXT:    v_bfe_u32 v16, v17, 16, 1
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v1.h, v12.h, v3.l, s0
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v8
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_add3_u32 v7, v16, v17, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, 0x400000, v17
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s0, v17, v17
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v1.h, v1.h, v5.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_max_f32_e32 v3, v3, v3
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v6.l
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v16, 0xffff0000, v11
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e64 v7, v7, v8, s0
; GFX11-TRUE16-NEXT:    v_max_f32_e32 v8, v10, v10
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s0, 0, v2.l
; GFX11-TRUE16-NEXT:    v_bfe_u32 v10, v3, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v17, 0x400000, v3
; GFX11-TRUE16-NEXT:    v_cmp_eq_f32_e64 s3, 0, v16
; GFX11-TRUE16-NEXT:    v_bfe_u32 v5, v8, 16, 1
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v2.l, v7.h, v2.l, s0
; GFX11-TRUE16-NEXT:    v_add3_u32 v10, v10, v3, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v18, 0x400000, v8
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s0, v3, v3
; GFX11-TRUE16-NEXT:    v_add3_u32 v5, v5, v8, 0x7fff
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s1, v8, v8
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xffff0000, v12
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v2.l, v2.l, v6.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e64 v3, v10, v17, s0
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s0, 0, v1.l
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e64 v5, v5, v18, s1
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s1, 0, v0.l
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v4.l
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xffff0000, v3
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v1.l, v3.h, v1.l, s0
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s0, 0, v9.l
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v0.l, v5.h, v0.l, s1
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xffff0000, v5
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xffff0000, v7
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v1.l, v1.l, v4.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_eq_f32_e64 s2, 0, v8
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v0.l, v0.l, v9.l, s0
; GFX11-TRUE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v6
; GFX11-TRUE16-NEXT:    v_cmp_eq_f32_e64 s0, 0, v10
; GFX11-TRUE16-NEXT:    v_cmp_eq_f32_e64 s1, 0, v17
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v50.l, v12.h, v1.h, s2
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v48.l, v11.h, v0.h, s3
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v13.l, v5.h, v0.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v30.l, v3.h, v1.l, s0
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v51.l, v7.h, v2.l, s1
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v3, v50 :: v_dual_mov_b32 v4, v48
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, v13 :: v_dual_mov_b32 v1, v30
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v2, v51
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v5, v39 :: v_dual_mov_b32 v6, v38
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v7, v37 :: v_dual_mov_b32 v8, v35
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v9, v33 :: v_dual_mov_b32 v10, v32
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v11, v31 :: v_dual_mov_b32 v12, v36
; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v13, v34
; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-FAKE16-LABEL: v_maximumnum_v32bf16:
; GFX11-FAKE16:       ; %bb.0:
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v28
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v52, 16, v12
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v34, 0xffff0000, v14
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v30
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v14
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v71, 16, v25
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v80, 16, v9
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v37, 0xffff0000, v13
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v34, v34
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v31, 0xffff0000, v15
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v29
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v13
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v83, 16, v24
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v34, v35, v33, vcc_lo
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v84, 16, v8
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v49, 0xffff0000, v12
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v37, v37
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v48, 0xffff0000, v29
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v54, 0xffff0000, v11
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v55, 16, v27
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v64, 16, v11
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v35, v39, v38, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v49, v49
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v67, 16, v26
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v68, 16, v10
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v66, 0xffff0000, v10
; GFX11-FAKE16-NEXT:    scratch_load_b32 v50, off, s32
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v37, v52, v51, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v54, v54
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v36, 0xffff0000, v30
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v82, 0xffff0000, v8
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v87, 16, v23
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v96, 16, v7
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v39, v64, v55 :: v_dual_and_b32 v70, 0xffff0000, v9
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v66, v66
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v99, 16, v22
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v100, 16, v6
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v103, 16, v21
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v112, 16, v5
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v49, v68, v67, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v70, v70
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v86, 0xffff0000, v7
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v115, 16, v20
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v116, 16, v4
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v119, 16, v19
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v52, v80, v71, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v82, v82
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v98, 0xffff0000, v6
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v128, 16, v3
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v131, 16, v18
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v132, 16, v2
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v54, v84, v83, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v86, v86
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v53, 0xffff0000, v28
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v102, 0xffff0000, v5
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v135, 16, v17
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v144, 16, v1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v64, v96, v87, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v98, v98
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v69, 0xffff0000, v26
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v118, 0xffff0000, v3
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v15
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v66, v100, v99, vcc_lo
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v114, 0xffff0000, v4
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v102, v102
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v68, v112, v103 :: v_dual_and_b32 v81, 0xffff0000, v25
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v114, v114
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v85, 0xffff0000, v24
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v130, 0xffff0000, v2
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v70, v116, v115, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v118, v118
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v101, 0xffff0000, v22
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v134, 0xffff0000, v1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v80, v128, v119, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v130, v130
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v82, v132, v131 :: v_dual_and_b32 v113, 0xffff0000, v21
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v134, v134
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v84, v144, v135 :: v_dual_and_b32 v117, 0xffff0000, v20
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v36, v36
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v33, v33, v34, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v48, v48
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v65, 0xffff0000, v27
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v133, 0xffff0000, v18
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v114, 16, v33
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v38, v38, v35, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v53, v53
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v51, v51, v37, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v65, v65
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v97, 0xffff0000, v23
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v116, 16, v51
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v55, v55, v39, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v69, v69
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v48, 16, v37
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v67, v67, v49, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v81, v81
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v71, v71, v52 :: v_dual_lshlrev_b32 v118, 16, v67
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v85, v85
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v83, v83, v54, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v97, v97
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v87, v87, v64 :: v_dual_lshlrev_b32 v128, 16, v83
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v101, v101
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v97, v99, v66, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v113, v113
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v129, 0xffff0000, v19
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v99, v103, v68, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v117, v117
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v53, 16, v39
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v101, v115, v70, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v129, v129
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v86, 16, v34
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v115, 16, v38
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v65, 16, v49
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v129, 16, v87
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v103, v119, v80, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v133, v133
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v36, 16, v35
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v119, 16, v71
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v81, 16, v54
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v133, 16, v103
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v113, v131, v82, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v86, v114
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v69, 16, v52
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v131, 16, v99
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v134, 16, v113
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v86, v33, v34, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v36, v115
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v145, 0xffff0000, v17
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v86, 16, v86
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v36, v38, v35, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v48, v116
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_dual_max_f32 v86, v86, v86 :: v_dual_lshlrev_b32 v85, 16, v64
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v48, v51, v37, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v114, v86, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v115, 0x400000, v86
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v48, 16, v48
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v117, 16, v55
; GFX11-FAKE16-NEXT:    v_add3_u32 v114, v114, v86, 0x7fff
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_max_f32_e32 v48, v48, v48
; GFX11-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v53, v117
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v53, v55, v39, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v65, v118
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v96, 16, v66
; GFX11-FAKE16-NEXT:    v_bfe_u32 v118, v48, 16, 1
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v53, 16, v53
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v65, v67, v49, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v69, v119
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v100, 16, v70
; GFX11-FAKE16-NEXT:    v_dual_max_f32 v53, v53, v53 :: v_dual_lshlrev_b32 v102, 16, v80
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v119, 0x400000, v48
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v69, v71, v52, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v81, v128
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v98, 16, v68
; GFX11-FAKE16-NEXT:    v_add3_u32 v118, v118, v48, 0x7fff
; GFX11-FAKE16-NEXT:    v_bfe_u32 v128, v53, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v81, v83, v54, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v85, v129
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v112, 16, v82
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v129, 0x400000, v53
; GFX11-FAKE16-NEXT:    v_add3_u32 v128, v128, v53, 0x7fff
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v81, 16, v81
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v85, v87, v64, vcc_lo
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_max_f32_e32 v81, v81, v81
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v85, 16, v85
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v144, 0x400000, v81
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v130, 16, v97
; GFX11-FAKE16-NEXT:    v_max_f32_e32 v85, v85, v85
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v96, v130
; GFX11-FAKE16-NEXT:    v_bfe_u32 v146, v85, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v147, 0x400000, v85
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v96, v97, v66, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v98, v131
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_add3_u32 v146, v146, v85, 0x7fff
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v132, 16, v101
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v96, 16, v96
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v98, v99, v68, vcc_lo
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v100, v132
; GFX11-FAKE16-NEXT:    v_max_f32_e32 v96, v96, v96
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v98, 16, v98
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v100, v101, v70, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v102, v133
; GFX11-FAKE16-NEXT:    v_max_f32_e32 v98, v98, v98
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v100, 16, v100
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v102, v103, v80, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v112, v134
; GFX11-FAKE16-NEXT:    v_bfe_u32 v134, v81, 16, 1
; GFX11-FAKE16-NEXT:    v_max_f32_e32 v100, v100, v100
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v112, v113, v82, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v86, v86
; GFX11-FAKE16-NEXT:    v_bfe_u32 v86, v96, 16, 1
; GFX11-FAKE16-NEXT:    v_add3_u32 v134, v134, v81, 0x7fff
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v114, v114, v115, vcc_lo
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v115, 0x400000, v96
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_add3_u32 v86, v86, v96, 0x7fff
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v69, 16, v69
; GFX11-FAKE16-NEXT:    v_max_f32_e32 v69, v69, v69
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v132, v69, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v133, 0x400000, v69
; GFX11-FAKE16-NEXT:    v_add3_u32 v132, v132, v69, 0x7fff
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v36, 16, v36
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_max_f32_e32 v36, v36, v36
; GFX11-FAKE16-NEXT:    v_bfe_u32 v116, v36, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v117, 0x400000, v36
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v36, v36
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_add3_u32 v116, v116, v36, 0x7fff
; GFX11-FAKE16-NEXT:    v_bfe_u32 v36, v98, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v116, v116, v117, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v48, v48
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v65, 16, v65
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v117, 0x400000, v98
; GFX11-FAKE16-NEXT:    v_bfe_u32 v48, v100, 16, 1
; GFX11-FAKE16-NEXT:    v_add3_u32 v36, v36, v98, 0x7fff
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v118, v118, v119 :: v_dual_max_f32 v65, v65, v65
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v53, v53
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v119, 0x400000, v100
; GFX11-FAKE16-NEXT:    v_add3_u32 v48, v48, v100, 0x7fff
; GFX11-FAKE16-NEXT:    v_bfe_u32 v130, v65, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v131, 0x400000, v65
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v128, v128, v129, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v65, v65
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_add3_u32 v130, v130, v65, 0x7fff
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v65, v130, v131, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v69, v69
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v69, v132, v133, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v81, v81
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v81, v134, v144, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v85, v85
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_4) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v130, 16, v81
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v85, v146, v147, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v96, v96
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v146, 0xffff0000, v50
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v131, 16, v85
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v86, v86, v115, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v98, v98
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v115, 16, v118
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v98, 0xffff0000, v114
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v132, 16, v86
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v36, v36, v117, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v100, v100
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v100, 16, v116
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v117, 16, v128
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v133, 16, v36
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v96, v48, v119, vcc_lo
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v48, 16, v114
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v34
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v119, 16, v65
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v134, 0xffff0000, v36
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v114, 0xffff0000, v116
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v116, 0xffff0000, v118
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v34, v48, v34, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v35
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v118, 0xffff0000, v128
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v128, 16, v69
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v144, 16, v96
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v35, v100, v35, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v37
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v36, v115, v37, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v39
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v37, v117, v39, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v49
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v39, v119, v49, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v52
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v49, v128, v52, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v54
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v52, v130, v54 :: v_dual_and_b32 v65, 0xffff0000, v65
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v64
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v69, 0xffff0000, v69
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v54, 16, v50
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v64, v131, v64, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v31, v31
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v147, v32, v54, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v66
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v31, v132, v66, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v68
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v32, v133, v68 :: v_dual_and_b32 v81, 0xffff0000, v81
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v146, v146
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v54, v54, v147 :: v_dual_and_b32 v85, 0xffff0000, v85
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v70
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v68, 16, v54
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v66, v144, v70, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v33
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v102, 16, v102
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v70, 16, v0
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v33, v34, v33, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v38
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v34, 16, v147
; GFX11-FAKE16-NEXT:    v_max_f32_e32 v102, v102, v102
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v35, v35, v38, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v51
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v53, v102, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v129, 0x400000, v102
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v36, v36, v51, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v34, v68
; GFX11-FAKE16-NEXT:    v_add3_u32 v53, v53, v102, 0x7fff
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v68, 0xffff0000, v0
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v34, v54, v147, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v55
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v37, v37, v55 :: v_dual_lshlrev_b32 v34, 16, v34
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v67
; GFX11-FAKE16-NEXT:    v_max_f32_e32 v34, v34, v34
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v38, v39, v67, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v71
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v55, v34, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v39, v49, v71, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v83
; GFX11-FAKE16-NEXT:    v_add3_u32 v55, v55, v34, 0x7fff
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v49, v52, v83, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v87
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v51, v64, v87, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v97
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v52, v31, v97, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v99
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v64, v32, v99, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v98
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v31, v48, v33, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v114
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v48, 0x400000, v34
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v32, v100, v35, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v116
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v33, v115, v36 :: v_dual_and_b32 v86, 0xffff0000, v86
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v34, v34
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v48, v55, v48, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v118
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v55, 16, v48
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v34, v117, v37, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v65
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v65, 0xffff0000, v96
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v35, v119, v38, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v69
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v69, 16, v16
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v36, v128, v39, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v81
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v37, v130, v49, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e64 vcc_lo, 0, v147
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v39, v55, v147, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v85
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v38, v131, v51, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v54
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v51, 0xffff0000, v48
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v39, v39, v54, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v86
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v48, v132, v52, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v134
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v49, v133, v64, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v51
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v51, 16, v112
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v39, v55, v39, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v145, v145
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_max_f32_e32 v67, v51, v51
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v54, v135, v84, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v102, v102
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v52, v53, v129 :: v_dual_lshlrev_b32 v55, 16, v54
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v53, 16, v84
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v101
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v64, v66, v101, vcc_lo
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v53, v55
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v66, 16, v52
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v52, 0xffff0000, v52
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v53, v54, v84, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v80
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v53, 16, v53
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v55, v66, v80, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v65
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v65, 0x400000, v67
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_dual_max_f32 v53, v53, v53 :: v_dual_lshlrev_b32 v80, 16, v15
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v51, v144, v64, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v64, v67, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v103
; GFX11-FAKE16-NEXT:    v_add3_u32 v64, v64, v67, 0x7fff
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v55, v55, v103, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v67, v67
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v64, v64, v65, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v68, v68
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v65, 0xffff0000, v16
; GFX11-FAKE16-NEXT:    v_bfe_u32 v68, v53, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v67, v70, v69, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v52
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v52, v66, v55, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v65, v65
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v65, 16, v64
; GFX11-FAKE16-NEXT:    v_add3_u32 v66, v68, v53, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v68, 0x400000, v53
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v55, v69, v67, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v82
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v69, 16, v67
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v71, v65, v82 :: v_dual_lshlrev_b32 v70, 16, v55
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v53, v53
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v53, v66, v68, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v69, v70
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v70, 16, v53
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v66, v55, v67, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v80, v80
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v68, v15, v50 :: v_dual_and_b32 v53, 0xffff0000, v53
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v15, 16, v50
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v113
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v69, v71, v113 :: v_dual_lshlrev_b32 v80, 16, v68
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v15, v15
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v15, 0xffff0000, v64
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v50, v50, v68, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v84
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v64, v70, v84, vcc_lo
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v15
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v15, v65, v69, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v54
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v69, 16, v14
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v71, 16, v50
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v54, v64, v54, vcc_lo
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v80, v71
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v65, v50, v68 :: v_dual_lshlrev_b32 v80, 16, v30
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v69, v69
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v65, 16, v65
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v66, 16, v66
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v69, v14, v30, vcc_lo
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_dual_max_f32 v65, v65, v65 :: v_dual_max_f32 v66, v66, v66
; GFX11-FAKE16-NEXT:    v_bfe_u32 v64, v66, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v71, 0x400000, v66
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v66, v66
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_add3_u32 v64, v64, v66, 0x7fff
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v64, v64, v71 :: v_dual_lshlrev_b32 v71, 16, v13
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v80, v80
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v80, 0x400000, v65
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v30, v30, v69, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v53
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v53, 16, v64
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v14, v70, v54, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v67
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v70, 16, v69
; GFX11-FAKE16-NEXT:    v_bfe_u32 v54, v65, 16, 1
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v67, v53, v67 :: v_dual_lshlrev_b32 v66, 16, v30
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_add3_u32 v54, v54, v65, 0x7fff
; GFX11-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v70, v66
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v66, v30, v69, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v71, v71
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v71, v13, v29, vcc_lo
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v70, 16, v29
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v65, v65
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v54, v54, v80, vcc_lo
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v70, v70
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v65, v29, v71, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v55
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v29, v67, v55, vcc_lo
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v67, 16, v65
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v55, 0xffff0000, v64
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v13, 16, v66
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v64, 16, v54
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v55
; GFX11-FAKE16-NEXT:    v_max_f32_e32 v66, v13, v13
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v13, v53, v29 :: v_dual_lshlrev_b32 v70, 16, v71
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v68
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v29, v64, v68, vcc_lo
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v70, v67
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v67, 16, v12
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v55, v65, v71, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v50
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v55, 16, v55
; GFX11-FAKE16-NEXT:    v_bfe_u32 v53, v66, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v29, v29, v50, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v66, v66
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v54, 0xffff0000, v54
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_add3_u32 v50, v53, v66, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v53, 0x400000, v66
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v66, 16, v11
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v50, v50, v53 :: v_dual_max_f32 v53, v55, v55
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v67, v67
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v55, 16, v28
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v67, 0x400000, v53
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v12, v12, v28, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v54
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v54, 16, v50
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v29, v64, v29, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v55, v55
; GFX11-FAKE16-NEXT:    v_bfe_u32 v64, v53, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v28, v28, v12, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v69
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_add3_u32 v64, v64, v53, 0x7fff
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v55, v54, v69, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v66, v66
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v66, 16, v27
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v11, v11, v27, vcc_lo
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v69, 16, v12
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v53, v53
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v53, v64, v67 :: v_dual_lshlrev_b32 v68, 16, v28
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v69, v68
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v64, v28, v12 :: v_dual_lshlrev_b32 v67, 16, v11
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v66, v66
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v66, v27, v11, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v30
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v27, v55, v30, vcc_lo
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v53
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v71
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v50, 0xffff0000, v50
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v68, v30, v71 :: v_dual_lshlrev_b32 v55, 16, v64
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v64, 16, v66
; GFX11-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v67, v64
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v64, v66, v11, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v50
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v27, v54, v27, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v65
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v54, v68, v65 :: v_dual_max_f32 v55, v55, v55
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v68, 16, v26
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v50, v55, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v67, 0x400000, v55
; GFX11-FAKE16-NEXT:    v_add3_u32 v50, v50, v55, 0x7fff
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v65, 16, v10
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_4) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v65, v65
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v10, v10, v26, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v55, v55
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v50, v50, v67, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v68, v68
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v65, v26, v10 :: v_dual_lshlrev_b32 v64, 16, v64
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_dual_max_f32 v64, v64, v64 :: v_dual_and_b32 v53, 0xffff0000, v53
; GFX11-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v53
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v55, v64, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v26, v30, v54, vcc_lo
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v50
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v12
; GFX11-FAKE16-NEXT:    v_add3_u32 v53, v55, v64, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v54, 0x400000, v64
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v55, 16, v65
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v12, v30, v12 :: v_dual_lshlrev_b32 v67, 16, v10
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v64, v64
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v68, 16, v9
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v50, 0xffff0000, v50
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v53, v53, v54, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v67, v55
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v54, v65, v10 :: v_dual_lshlrev_b32 v55, 16, v25
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v68, v68
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v9, v9, v25, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v28
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v12, v12, v28, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v55, v55
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v28, 16, v53
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v53, 0xffff0000, v53
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v25, v25, v9, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v11
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v55, 16, v25
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v11, v28, v11, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v50
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v12, v30, v12, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v66
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v64, 16, v9
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_perm_b32 v12, v33, v12, 0x5040100
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v11, v11, v66, vcc_lo
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v54, 16, v54
; GFX11-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v64, v55
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v55, 16, v8
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v66, 16, v24
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_max_f32_e32 v54, v54, v54
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v50, v25, v9, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v55, v55
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v30, v54, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v64, 0x400000, v54
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v8, v8, v24, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v54, v54
; GFX11-FAKE16-NEXT:    v_add3_u32 v30, v30, v54, 0x7fff
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_4) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v30, v30, v64, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v66, v66
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v64, 16, v7
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v24, v24, v8, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v53
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v11, v28, v11 :: v_dual_lshlrev_b32 v54, 16, v24
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v28, 16, v30
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v10
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v55, 16, v8
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_perm_b32 v11, v34, v11, 0x5040100
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v10, v28, v10, vcc_lo
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v55, v54
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v54, v24, v8 :: v_dual_lshlrev_b32 v55, 16, v23
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v64, v64
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v7, v7, v23 :: v_dual_lshlrev_b32 v50, 16, v50
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v64, 16, v7
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_max_f32_e32 v50, v50, v50
; GFX11-FAKE16-NEXT:    v_bfe_u32 v53, v50, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v66, 0x400000, v50
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v50, v50
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_add3_u32 v53, v53, v50, 0x7fff
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v50, v53, v66, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v55, v55
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v53, 16, v54
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v30, 0xffff0000, v30
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v54, 16, v50
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v23, v23, v7, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v65
; GFX11-FAKE16-NEXT:    v_dual_max_f32 v53, v53, v53 :: v_dual_and_b32 v50, 0xffff0000, v50
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v10, v10, v65 :: v_dual_lshlrev_b32 v55, 16, v23
; GFX11-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v30
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v10, v28, v10, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v9
; GFX11-FAKE16-NEXT:    v_bfe_u32 v28, v53, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v9, v54, v9, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v64, v55
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v30, v23, v7 :: v_dual_lshlrev_b32 v55, 16, v6
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v25
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v9, v9, v25 :: v_dual_lshlrev_b32 v30, 16, v30
; GFX11-FAKE16-NEXT:    v_add3_u32 v25, v28, v53, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v28, 0x400000, v53
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v53, v53
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v25, v25, v28, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v55, v55
; GFX11-FAKE16-NEXT:    v_max_f32_e32 v28, v30, v30
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v30, 16, v22
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v6, v6, v22, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v50
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v50, 16, v25
; GFX11-FAKE16-NEXT:    v_bfe_u32 v53, v28, 16, 1
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v9, v54, v9 :: v_dual_lshlrev_b32 v64, 16, v6
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v30, v30
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v30, 16, v5
; GFX11-FAKE16-NEXT:    v_add3_u32 v53, v53, v28, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v54, 0x400000, v28
; GFX11-FAKE16-NEXT:    v_perm_b32 v9, v36, v9, 0x5040100
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v22, v22, v6, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v8
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v8, v50, v8 :: v_dual_lshlrev_b32 v55, 16, v22
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v30, v30
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v30, 16, v21
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v5, v5, v21, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v28, v28
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v28, v53, v54, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v64, v55
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v53, v22, v6 :: v_dual_lshlrev_b32 v54, 16, v5
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v30, v30
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v30, 16, v53
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v21, v21, v5, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v24
; GFX11-FAKE16-NEXT:    v_perm_b32 v10, v35, v10, 0x5040100
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v8, v8, v24 :: v_dual_lshlrev_b32 v53, 16, v21
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v24, 0xffff0000, v25
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v25, 16, v28
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v28, 0xffff0000, v28
; GFX11-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v24
; GFX11-FAKE16-NEXT:    v_max_f32_e32 v24, v30, v30
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v8, v50, v8, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v7
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v7, v25, v7 :: v_dual_lshlrev_b32 v50, 16, v4
; GFX11-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v54, v53
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v30, v21, v5 :: v_dual_lshlrev_b32 v53, 16, v20
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v23
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v7, v7, v23 :: v_dual_lshlrev_b32 v30, 16, v30
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v50, v50
; GFX11-FAKE16-NEXT:    v_bfe_u32 v23, v24, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v50, 0x400000, v24
; GFX11-FAKE16-NEXT:    v_max_f32_e32 v30, v30, v30
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v4, v4, v20, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v53, v53
; GFX11-FAKE16-NEXT:    v_add3_u32 v23, v23, v24, 0x7fff
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v53, 16, v4
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v20, v20, v4, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v24, v24
; GFX11-FAKE16-NEXT:    v_bfe_u32 v24, v30, 16, 1
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v23, v23, v50 :: v_dual_lshlrev_b32 v50, 16, v20
; GFX11-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v28
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_add3_u32 v24, v24, v30, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v28, 0x400000, v30
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v7, v25, v7, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v53, v50
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v25, 16, v23
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v53, 16, v3
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v23, 0xffff0000, v23
; GFX11-FAKE16-NEXT:    v_perm_b32 v7, v38, v7, 0x5040100
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v50, v20, v4, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v6
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v6, v25, v6, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v30, v30
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v30, 16, v19
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v24, v24, v28, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v53, v53
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v3, v3, v19 :: v_dual_lshlrev_b32 v28, 16, v50
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v22
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v6, v6, v22 :: v_dual_lshlrev_b32 v53, 16, v3
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v30, v30
; GFX11-FAKE16-NEXT:    v_max_f32_e32 v22, v28, v28
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v28, 16, v24
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v19, v19, v3 :: v_dual_and_b32 v24, 0xffff0000, v24
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v5
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v30, v22, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v54, 0x400000, v22
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v5, v28, v5 :: v_dual_lshlrev_b32 v50, 16, v19
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v21
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_add3_u32 v30, v30, v22, 0x7fff
; GFX11-FAKE16-NEXT:    v_perm_b32 v8, v37, v8, 0x5040100
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v5, v5, v21, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v53, v50
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v53, 16, v17
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v21, v19, v3, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v22, v22
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v22, v30, v54 :: v_dual_lshlrev_b32 v21, 16, v21
; GFX11-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v23
; GFX11-FAKE16-NEXT:    v_dual_max_f32 v21, v21, v21 :: v_dual_lshlrev_b32 v30, 16, v0
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v23, 16, v22
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v6, v25, v6, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v24
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v24, 16, v2
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v25, 16, v18
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v5, v28, v5, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v4
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_perm_b32 v5, v49, v5, 0x5040100
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v4, v23, v4, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v24, v24
; GFX11-FAKE16-NEXT:    v_bfe_u32 v24, v21, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v2, v18, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v20
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v50, 16, v2
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v4, v4, v20, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v25, v25
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v25, 16, v1
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v20, 0xffff0000, v22
; GFX11-FAKE16-NEXT:    v_add3_u32 v22, v24, v21, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v24, 0x400000, v21
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v18, v18, v2, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v25, v25
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v25, 16, v16
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v1, v1, v17 :: v_dual_lshlrev_b32 v28, 16, v18
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v30, v30
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v30, 16, v1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v0, v16, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v53, v53
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v17, v17, v1, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v50, v28
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v28, v18, v2, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v25, v25
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v25, 16, v17
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v16, v16, v0, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v21, v21
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v21, v22, v24, vcc_lo
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v22, 16, v16
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v24, 16, v0
; GFX11-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v30, v25
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v21
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v21, 0xffff0000, v21
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v25, v17, v1, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v24, v22
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_dual_max_f32 v24, v28, v28 :: v_dual_lshlrev_b32 v25, 16, v25
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v22, v16, v0, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v3
; GFX11-FAKE16-NEXT:    v_bfe_u32 v28, v24, 16, 1
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_dual_max_f32 v25, v25, v25 :: v_dual_lshlrev_b32 v22, 16, v22
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v3, v30, v3, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v19
; GFX11-FAKE16-NEXT:    v_bfe_u32 v50, v25, 16, 1
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v3, v3, v19, vcc_lo
; GFX11-FAKE16-NEXT:    v_max_f32_e32 v19, v22, v22
; GFX11-FAKE16-NEXT:    v_add3_u32 v22, v28, v24, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v28, 0x400000, v24
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v24, v24
; GFX11-FAKE16-NEXT:    v_add3_u32 v24, v50, v25, 0x7fff
; GFX11-FAKE16-NEXT:    v_bfe_u32 v53, v19, 16, 1
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v22, v22, v28, vcc_lo
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v28, 0x400000, v25
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v25, v25
; GFX11-FAKE16-NEXT:    v_add3_u32 v50, v53, v19, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v53, 0x400000, v19
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v54, 16, v22
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v24, v24, v28, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v19, v19
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v19, v50, v53, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v21
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v21, 16, v24
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v25, 16, v19
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v3, v30, v3, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v2
; GFX11-FAKE16-NEXT:    v_perm_b32 v6, v48, v6, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v3, v52, v3, 0x5040100
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v54, v2, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v21, v1, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v18
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v2, v18, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v0
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v18, 0xffff0000, v19
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v25, v0, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v17
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v1, v17, vcc_lo
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v17, 0xffff0000, v24
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v16
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v0, v16, vcc_lo
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v17
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v1, v21, v1 :: v_dual_and_b32 v16, 0xffff0000, v22
; GFX11-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v18
; GFX11-FAKE16-NEXT:    v_perm_b32 v1, v14, v1, 0x5040100
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v25, v0, vcc_lo
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v16
; GFX11-FAKE16-NEXT:    v_perm_b32 v14, v31, v27, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v0, v13, v0, 0x5040100
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v54, v2, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v20
; GFX11-FAKE16-NEXT:    v_perm_b32 v13, v32, v26, 0x5040100
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_perm_b32 v2, v15, v2, 0x5040100
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v4, v23, v4, vcc_lo
; GFX11-FAKE16-NEXT:    v_perm_b32 v15, v39, v29, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v4, v51, v4, 0x5040100
; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
;
; GFX12-TRUE16-LABEL: v_maximumnum_v32bf16:
; GFX12-TRUE16:       ; %bb.0:
; GFX12-TRUE16-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX12-TRUE16-NEXT:    s_wait_expcnt 0x0
; GFX12-TRUE16-NEXT:    s_wait_samplecnt 0x0
; GFX12-TRUE16-NEXT:    s_wait_bvhcnt 0x0
; GFX12-TRUE16-NEXT:    s_wait_kmcnt 0x0
; GFX12-TRUE16-NEXT:    scratch_load_b32 v51, off, s32
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v81, 0xffff0000, v2
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v52, 0xffff0000, v25
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v82, 0xffff0000, v18
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v69, 0xffff0000, v4
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v70, 0xffff0000, v20
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s24, v81, v81
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s11, v52, v52
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s25, v82, v82
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s20, v69, v69
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v80, 0xffff0000, v19
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v52.l, v2.h, v18.h, s24
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s21, v70, v70
; GFX12-TRUE16-NEXT:    v_dual_mov_b32 v50, v15 :: v_dual_mov_b32 v49, v14
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v32, 0xffff0000, v13
; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v82.l, v18.h, v52.l, s25
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v112.l, v52.l
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v38, 0xffff0000, v10
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s23, v80, v80
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v15, 0xffff0000, v49
; GFX12-TRUE16-NEXT:    v_mov_b16_e64 v134.l, v82.l
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v112, 16, v112
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v33, 0xffff0000, v29
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s2, v32, v32
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s8, v38, v38
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v134, 16, v134
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v55, 0xffff0000, v7
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v66, 0xffff0000, v22
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v67, 0xffff0000, v5
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s0, v15, v15
; GFX12-TRUE16-NEXT:    v_cmp_gt_f32_e64 s44, v112, v134
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s3, v33, v33
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v15.l, v13.h, v29.h, s2
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v33.l, v10.h, v26.h, s8
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v68, 0xffff0000, v21
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v112.l, v82.l, v52.l, s44
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v71, 0xffff0000, v3
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s14, v55, v55
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s17, v66, v66
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s18, v67, v67
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v112, 16, v112
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v39, 0xffff0000, v26
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v55.l, v29.h, v15.l, s3
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v48, 0xffff0000, v9
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v83, 0xffff0000, v1
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s19, v68, v68
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s9, v39, v39
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v39.l, v4.h, v20.h, s20
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s22, v71, v71
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v38.l, v5.h, v21.h, s18
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v85.l, v15.l
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v66.l, v26.h, v33.l, s9
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v80.l, v20.h, v39.l, s21
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v102.l, v39.l
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v96.l, v33.l
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v115.l, v55.l
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v118.l, v66.l
; GFX12-TRUE16-NEXT:    v_mov_b16_e64 v132.l, v80.l
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v102, 16, v102
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s10, v48, v48
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s26, v83, v83
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v48.l, v3.h, v19.h, s22
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v132, 16, v132
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v71.l, v21.h, v38.l, s19
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v85, 16, v85
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v96, 16, v96
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v115, 16, v115
; GFX12-TRUE16-NEXT:    v_cmp_gt_f32_e64 s42, v102, v132
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v118, 16, v118
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v101.l, v38.l
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v81.l, v19.h, v48.l, s23
; GFX12-TRUE16-NEXT:    v_mov_b16_e64 v131.l, v71.l
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v102.l, v80.l, v39.l, s42
; GFX12-TRUE16-NEXT:    v_cmp_gt_f32_e64 s23, v85, v115
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v101, 16, v101
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v34, 0xffff0000, v12
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v131, 16, v131
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v102, 16, v102
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v53, 0xffff0000, v8
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v85.l, v55.l, v15.l, s23
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v54, 0xffff0000, v24
; GFX12-TRUE16-NEXT:    v_cmp_gt_f32_e64 s41, v101, v131
; GFX12-TRUE16-NEXT:    v_max_num_f32_e32 v102, v102, v102
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s12, v53, v53
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v53.l, v1.h, v17.h, s26
; GFX12-TRUE16-NEXT:    v_cmp_gt_f32_e64 s26, v96, v118
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v85, 16, v85
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v14, 0xffff0000, v50
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v101.l, v71.l, v38.l, s41
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v35, 0xffff0000, v28
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v96.l, v66.l, v33.l, s26
; GFX12-TRUE16-NEXT:    v_max_num_f32_e32 v85, v85, v85
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v14, v14
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v14.l, v49.h, v30.h, s0
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v84, 0xffff0000, v17
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v96, 16, v96
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v31, 0xffff0000, v30
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s4, v34, v34
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s13, v54, v54
; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX12-TRUE16-NEXT:    v_dual_max_num_f32 v96, v96, v96 :: v_dual_lshlrev_b32 v101, 16, v101
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s1, v31, v31
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v36, 0xffff0000, v11
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v37, 0xffff0000, v27
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v64, 0xffff0000, v23
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s5, v35, v35
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v54.l, v30.h, v14.l, s1
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s27, v84, v84
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v31.l, v12.h, v28.h, s4
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v84.l, v14.l
; GFX12-TRUE16-NEXT:    v_max_num_f32_e32 v101, v101, v101
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s6, v36, v36
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v114.l, v54.l
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v65, 0xffff0000, v6
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s7, v37, v37
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s15, v64, v64
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v32.l, v11.h, v27.h, s6
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v64.l, v28.h, v31.l, s5
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v84, 16, v84
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v114, 16, v114
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s16, v65, v65
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v86.l, v31.l
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v65.l, v27.h, v32.l, s7
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v116.l, v64.l
; GFX12-TRUE16-NEXT:    v_cmp_gt_f32_e64 s22, v84, v114
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v34.l, v9.h, v25.h, s10
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v87.l, v32.l
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v86, 16, v86
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v117.l, v65.l
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v116, 16, v116
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v84.l, v54.l, v14.l, s22
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v35.l, v8.h, v24.h, s12
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v67.l, v25.h, v34.l, s11
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v87, 16, v87
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v117, 16, v117
; GFX12-TRUE16-NEXT:    v_cmp_gt_f32_e64 s24, v86, v116
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v84, 16, v84
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v36.l, v7.h, v23.h, s14
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v97.l, v34.l
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v68.l, v24.h, v35.l, s13
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v119.l, v67.l
; GFX12-TRUE16-NEXT:    v_cmp_gt_f32_e64 s25, v87, v117
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v86.l, v64.l, v31.l, s24
; GFX12-TRUE16-NEXT:    v_max_num_f32_e32 v84, v84, v84
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v37.l, v6.h, v22.h, s16
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v98.l, v35.l
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v69.l, v23.h, v36.l, s15
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v97, 16, v97
; GFX12-TRUE16-NEXT:    v_mov_b16_e64 v128.l, v68.l
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v119, 16, v119
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v87.l, v65.l, v32.l, s25
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v86, 16, v86
; GFX12-TRUE16-NEXT:    v_bfe_u32 v114, v84, 16, 1
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v99.l, v36.l
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v70.l, v22.h, v37.l, s17
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v83.l, v17.h, v53.l, s27
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v98, 16, v98
; GFX12-TRUE16-NEXT:    v_mov_b16_e64 v129.l, v69.l
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v128, 16, v128
; GFX12-TRUE16-NEXT:    v_cmp_gt_f32_e64 s27, v97, v119
; GFX12-TRUE16-NEXT:    v_dual_max_num_f32 v86, v86, v86 :: v_dual_lshlrev_b32 v87, 16, v87
; GFX12-TRUE16-NEXT:    v_or_b32_e32 v115, 0x400000, v84
; GFX12-TRUE16-NEXT:    v_bfe_u32 v116, v85, 16, 1
; GFX12-TRUE16-NEXT:    v_add3_u32 v114, v114, v84, 0x7fff
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s22, v84, v84
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v100.l, v37.l
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v99, 16, v99
; GFX12-TRUE16-NEXT:    v_mov_b16_e64 v130.l, v70.l
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v129, 16, v129
; GFX12-TRUE16-NEXT:    v_cmp_gt_f32_e64 s28, v98, v128
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v97.l, v67.l, v34.l, s27
; GFX12-TRUE16-NEXT:    v_max_num_f32_e32 v87, v87, v87
; GFX12-TRUE16-NEXT:    v_or_b32_e32 v117, 0x400000, v85
; GFX12-TRUE16-NEXT:    v_bfe_u32 v118, v86, 16, 1
; GFX12-TRUE16-NEXT:    v_add3_u32 v116, v116, v85, 0x7fff
; GFX12-TRUE16-NEXT:    v_cndmask_b32_e64 v114, v114, v115, s22
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s22, v85, v85
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v100, 16, v100
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v130, 16, v130
; GFX12-TRUE16-NEXT:    v_cmp_gt_f32_e64 s29, v99, v129
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v98.l, v68.l, v35.l, s28
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v97, 16, v97
; GFX12-TRUE16-NEXT:    v_or_b32_e32 v119, 0x400000, v86
; GFX12-TRUE16-NEXT:    v_bfe_u32 v128, v87, 16, 1
; GFX12-TRUE16-NEXT:    v_add3_u32 v118, v118, v86, 0x7fff
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b32_e64 v116, v116, v117, s22
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s22, v86, v86
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v103.l, v48.l
; GFX12-TRUE16-NEXT:    v_mov_b16_e64 v133.l, v81.l
; GFX12-TRUE16-NEXT:    v_cmp_gt_f32_e64 s40, v100, v130
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v99.l, v69.l, v36.l, s29
; GFX12-TRUE16-NEXT:    v_dual_max_num_f32 v97, v97, v97 :: v_dual_lshlrev_b32 v98, 16, v98
; GFX12-TRUE16-NEXT:    v_or_b32_e32 v129, 0x400000, v87
; GFX12-TRUE16-NEXT:    v_bfe_u32 v130, v96, 16, 1
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b32_e64 v118, v118, v119, s22
; GFX12-TRUE16-NEXT:    v_add3_u32 v128, v128, v87, 0x7fff
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s22, v87, v87
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v103, 16, v103
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v133, 16, v133
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v100.l, v70.l, v37.l, s40
; GFX12-TRUE16-NEXT:    v_dual_max_num_f32 v98, v98, v98 :: v_dual_lshlrev_b32 v99, 16, v99
; GFX12-TRUE16-NEXT:    v_or_b32_e32 v131, 0x400000, v96
; GFX12-TRUE16-NEXT:    v_bfe_u32 v132, v97, 16, 1
; GFX12-TRUE16-NEXT:    v_add3_u32 v130, v130, v96, 0x7fff
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b32_e64 v128, v128, v129, s22
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s22, v96, v96
; GFX12-TRUE16-NEXT:    v_cmp_gt_f32_e64 s43, v103, v133
; GFX12-TRUE16-NEXT:    v_dual_max_num_f32 v99, v99, v99 :: v_dual_lshlrev_b32 v100, 16, v100
; GFX12-TRUE16-NEXT:    v_or_b32_e32 v133, 0x400000, v97
; GFX12-TRUE16-NEXT:    v_bfe_u32 v134, v98, 16, 1
; GFX12-TRUE16-NEXT:    v_add3_u32 v132, v132, v97, 0x7fff
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b32_e64 v96, v130, v131, s22
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s22, v97, v97
; GFX12-TRUE16-NEXT:    v_max_num_f32_e32 v100, v100, v100
; GFX12-TRUE16-NEXT:    v_or_b32_e32 v144, 0x400000, v98
; GFX12-TRUE16-NEXT:    v_bfe_u32 v145, v99, 16, 1
; GFX12-TRUE16-NEXT:    v_add3_u32 v134, v134, v98, 0x7fff
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b32_e64 v97, v132, v133, s22
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s22, v98, v98
; GFX12-TRUE16-NEXT:    v_or_b32_e32 v146, 0x400000, v99
; GFX12-TRUE16-NEXT:    v_bfe_u32 v147, v100, 16, 1
; GFX12-TRUE16-NEXT:    v_add3_u32 v145, v145, v99, 0x7fff
; GFX12-TRUE16-NEXT:    v_or_b32_e32 v84, 0x400000, v100
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b32_e64 v98, v134, v144, s22
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s22, v99, v99
; GFX12-TRUE16-NEXT:    v_bfe_u32 v115, v101, 16, 1
; GFX12-TRUE16-NEXT:    v_add3_u32 v147, v147, v100, 0x7fff
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s1, 0, v15.l
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s5, 0, v31.l
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b32_e64 v99, v145, v146, s22
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s22, v100, v100
; GFX12-TRUE16-NEXT:    v_or_b32_e32 v85, 0x400000, v101
; GFX12-TRUE16-NEXT:    v_add3_u32 v115, v115, v101, 0x7fff
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s2, 0, v32.l
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v14.h, v116.h, v15.l, s1
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b32_e64 v84, v147, v84, s22
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s22, v101, v101
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v15.l, v118.h, v31.l, s5
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s7, 0, v34.l
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v15.h, v128.h, v32.l, s2
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s6, 0, v33.l
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s3, 0, v35.l
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s8, 0, v36.l
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v32.h, v97.h, v34.l, s7
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s4, 0, v37.l
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v31.h, v96.h, v33.l, s6
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v33.l, v98.h, v35.l, s3
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v33.h, v99.h, v36.l, s8
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s0, 0, v14.l
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s9, 0, v38.l
; GFX12-TRUE16-NEXT:    v_cndmask_b32_e64 v85, v115, v85, s22
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s11, 0, v54.l
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s12, 0, v55.l
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v14.l, v114.h, v14.l, s0
; GFX12-TRUE16-NEXT:    v_bfe_u32 v117, v102, 16, 1
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s13, 0, v64.l
; GFX12-TRUE16-NEXT:    v_or_b32_e32 v86, 0x400000, v102
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s22, v102, v102
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v14.l, v14.l, v54.l, s11
; GFX12-TRUE16-NEXT:    v_add3_u32 v117, v117, v102, 0x7fff
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s10, 0, v39.l
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v100, 0xffff0000, v114
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v15.l, v15.l, v64.l, s13
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s14, 0, v65.l
; GFX12-TRUE16-NEXT:    v_cndmask_b32_e64 v86, v117, v86, s22
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v115, 0xffff0000, v128
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v113.l, v53.l
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s15, 0, v66.l
; GFX12-TRUE16-NEXT:    v_mov_b16_e64 v135.l, v83.l
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v103.l, v81.l, v48.l, s43
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v102, 0xffff0000, v116
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v15.h, v15.h, v65.l, s14
; GFX12-TRUE16-NEXT:    v_cmp_eq_f32_e64 s2, 0, v115
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s20, 0, v71.l
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v113, 16, v113
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v103, 16, v103
; GFX12-TRUE16-NEXT:    v_cmp_eq_f32_e64 s0, 0, v102
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s16, 0, v67.l
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s17, 0, v68.l
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v131, 0xffff0000, v98
; GFX12-TRUE16-NEXT:    v_max_num_f32_e32 v103, v103, v103
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s19, 0, v70.l
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v132, 0xffff0000, v84
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v33.l, v33.l, v68.l, s17
; GFX12-TRUE16-NEXT:    v_cmp_eq_f32_e64 s5, 0, v131
; GFX12-TRUE16-NEXT:    v_bfe_u32 v119, v103, 16, 1
; GFX12-TRUE16-NEXT:    v_or_b32_e32 v87, 0x400000, v103
; GFX12-TRUE16-NEXT:    v_cmp_eq_f32_e64 s7, 0, v132
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v117, 0xffff0000, v96
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s21, 0, v80.l
; GFX12-TRUE16-NEXT:    v_add3_u32 v119, v119, v103, 0x7fff
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v129, 0xffff0000, v97
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s18, 0, v69.l
; GFX12-TRUE16-NEXT:    v_cmp_eq_f32_e64 s3, 0, v117
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v133, 0xffff0000, v85
; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_4) | instid1(VALU_DEP_2)
; GFX12-TRUE16-NEXT:    v_cmp_eq_f32_e64 s8, 0, v133
; GFX12-TRUE16-NEXT:    s_wait_loadcnt 0x0
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v101, 0xffff0000, v51
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v31.l, v50.h, v51.h, vcc_lo
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v101, v101
; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v34.l, v31.l
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v101, 0xffff0000, v118
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v32.l, v51.h, v31.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v35, 16, v34
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v34.l, v84.h, v37.l, s4
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v34.h, v85.h, v38.l, s9
; GFX12-TRUE16-NEXT:    v_cmp_eq_f32_e64 s1, 0, v101
; GFX12-TRUE16-NEXT:    v_mov_b16_e64 v130.l, v32.l
; GFX12-TRUE16-NEXT:    v_cmp_eq_f32_e64 s4, 0, v129
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v34.l, v34.l, v70.l, s19
; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v36, 16, v130
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v130, 0xffff0000, v99
; GFX12-TRUE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v35, v36
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v35.h, v14.h, v55.l, s12
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v35.l, v86.h, v39.l, s10
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v39, 0xffff0000, v86
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s10, 0, v31.l
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v36.l, v32.l, v31.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v100
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v36.h, v118.h, v15.l, s1
; GFX12-TRUE16-NEXT:    v_cmp_eq_f32_e64 s9, 0, v39
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v35.l, v35.l, v80.l, s21
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v37.l, v36.l
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v14.h, v114.h, v14.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v32.l
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v36.l, v31.h, v66.l, s15
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v31.h, v128.h, v15.h, s2
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v38, 16, v37
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v37.l, v32.h, v67.l, s16
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v48.h, v86.h, v35.l, s9
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v32.h, v96.h, v36.l, s3
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v37.h, v33.h, v69.l, s18
; GFX12-TRUE16-NEXT:    v_max_num_f32_e32 v54, v38, v38
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v38.l, v34.h, v71.l, s20
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v34.h, v116.h, v35.h, s0
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v35.h, v98.h, v33.l, s5
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v38.h, v84.h, v34.l, s7
; GFX12-TRUE16-NEXT:    v_bfe_u32 v55, v54, 16, 1
; GFX12-TRUE16-NEXT:    v_or_b32_e32 v64, 0x400000, v54
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s11, v54, v54
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v33.h, v97.h, v37.l, s4
; GFX12-TRUE16-NEXT:    v_cmp_eq_f32_e64 s6, 0, v130
; GFX12-TRUE16-NEXT:    v_add3_u32 v55, v55, v54, 0x7fff
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v37.h, v99.h, v37.h, s6
; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX12-TRUE16-NEXT:    v_cndmask_b32_e64 v54, v55, v64, s11
; GFX12-TRUE16-NEXT:    v_max_num_f32_e32 v64, v112, v112
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v14.l, v54.h, v31.l, s10
; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v39, 0xffff0000, v54
; GFX12-TRUE16-NEXT:    v_bfe_u32 v65, v64, 16, 1
; GFX12-TRUE16-NEXT:    v_or_b32_e32 v66, 0x400000, v64
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s1, v64, v64
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v14.l, v14.l, v32.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v39
; GFX12-TRUE16-NEXT:    v_add3_u32 v65, v65, v64, 0x7fff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v39.h, v85.h, v38.l, s8
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v15.h, v54.h, v14.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v54, 16, v135
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v103, v103
; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
; GFX12-TRUE16-NEXT:    v_cmp_gt_f32_e64 s0, v113, v54
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_dual_cndmask_b32 v55, v119, v87 :: v_dual_and_b32 v54, 0xffff0000, v0
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v48.l
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v15.l, v83.l, v53.l, s0
; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s0, v54, v54
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v54, 0xffff0000, v16
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v14.l, v55.h, v48.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v81.l
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v67.l, v15.l
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v15.l, v0.h, v16.h, s0
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s0, v54, v54
; GFX12-TRUE16-NEXT:    v_cndmask_b32_e64 v54, v65, v66, s1
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s1, 0, v52.l
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v64, 16, v67
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v65, 0xffff0000, v55
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v31.l, v16.h, v15.l, s0
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v14.l, v14.l, v81.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v32.l, v54.h, v52.l, s1
; GFX12-TRUE16-NEXT:    v_max_num_f32_e32 v64, v64, v64
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v52.l, v15.l
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v66.l, v31.l
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v82.l
; GFX12-TRUE16-NEXT:    v_cmp_eq_f32_e64 s0, 0, v65
; GFX12-TRUE16-NEXT:    v_bfe_u32 v67, v64, 16, 1
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v52, 16, v52
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v65, 16, v66
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v32.l, v32.l, v82.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s1, v64, v64
; GFX12-TRUE16-NEXT:    v_add3_u32 v66, v67, v64, 0x7fff
; GFX12-TRUE16-NEXT:    v_or_b32_e32 v67, 0x400000, v64
; GFX12-TRUE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v52, v65
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v52, 16, v50
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v68, 0xffff0000, v54
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v50.h, v55.h, v14.l, s0
; GFX12-TRUE16-NEXT:    v_cndmask_b32_e64 v64, v66, v67, s1
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s1, 0, v53.l
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v33.l, v31.l, v15.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v52, v52
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v52, 16, v51
; GFX12-TRUE16-NEXT:    v_cmp_eq_f32_e64 s2, 0, v68
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v34.l, v64.h, v53.l, s1
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v53.l, v33.l
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v33.l, v50.l, v51.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v52, v52
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s1, 0, v83.l
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v51.h, v54.h, v32.l, s2
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v52, 16, v53
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v53, 16, v49
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v14.l, v51.l, v33.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v65.l, v33.l
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v32.l, v34.l, v83.l, s1
; GFX12-TRUE16-NEXT:    v_max_num_f32_e32 v52, v52, v52
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v53, v53
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v55.l, v14.l
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v53, 16, v30
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v54, 0xffff0000, v64
; GFX12-TRUE16-NEXT:    v_bfe_u32 v66, v52, 16, 1
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v65, 16, v65
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v55, 16, v55
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v34.l, v49.l, v30.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v53, v53
; GFX12-TRUE16-NEXT:    v_cmp_eq_f32_e64 s0, 0, v54
; GFX12-TRUE16-NEXT:    v_add3_u32 v49, v66, v52, 0x7fff
; GFX12-TRUE16-NEXT:    v_or_b32_e32 v53, 0x400000, v52
; GFX12-TRUE16-NEXT:    v_cmp_gt_f32_e64 s1, v65, v55
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v30.l, v30.l, v34.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v52, v52
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v54, 16, v13
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v30.h, v64.h, v32.l, s0
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v32.l, v14.l, v33.l, s1
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v52.l, v30.l
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b32_e32 v49, v49, v53, vcc_lo
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v53.l, v34.l
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s0, v54, v54
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v54, 16, v29
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v15.l
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v55.l, v32.l
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v52, 16, v52
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v53, 16, v53
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v13.l, v13.l, v29.l, s0
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s0, v54, v54
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v13.h, v49.h, v15.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v55, 16, v55
; GFX12-TRUE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v53, v52
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s1, 0, v31.l
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v29.l, v29.l, v13.l, s0
; GFX12-TRUE16-NEXT:    v_dual_max_num_f32 v52, v55, v55 :: v_dual_and_b32 v53, 0xffff0000, v49
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v15.l, v30.l, v34.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v64.l, v13.l
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v55.l, v29.l
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v13.h, v13.h, v31.l, s1
; GFX12-TRUE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v53
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v53.l, v15.l
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v64, 16, v64
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v55, 16, v55
; GFX12-TRUE16-NEXT:    v_bfe_u32 v54, v52, 16, 1
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v13.h, v49.h, v13.h, vcc_lo
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v53, 16, v53
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v49, 16, v12
; GFX12-TRUE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v64, v55
; GFX12-TRUE16-NEXT:    v_add3_u32 v54, v54, v52, 0x7fff
; GFX12-TRUE16-NEXT:    v_or_b32_e32 v65, 0x400000, v52
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s0, v52, v52
; GFX12-TRUE16-NEXT:    v_max_num_f32_e32 v53, v53, v53
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v15.l, v29.l, v13.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v49, v49
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v49, 16, v28
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b32_e64 v52, v54, v65, s0
; GFX12-TRUE16-NEXT:    v_bfe_u32 v54, v53, 16, 1
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s0, 0, v33.l
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v12.l, v12.l, v28.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v49, v49
; GFX12-TRUE16-NEXT:    v_or_b32_e32 v55, 0x400000, v53
; GFX12-TRUE16-NEXT:    v_add3_u32 v54, v54, v53, 0x7fff
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v64.l, v15.l
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s1, v53, v53
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v28.l, v28.l, v12.l, vcc_lo
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v12.h, v52.h, v33.l, s0
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s0, 0, v14.l
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v53, 16, v64
; GFX12-TRUE16-NEXT:    v_cndmask_b32_e64 v49, v54, v55, s1
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s1, 0, v34.l
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v54, 0xffff0000, v52
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v55.l, v28.l
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v64.l, v12.l
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v12.h, v12.h, v14.l, s0
; GFX12-TRUE16-NEXT:    v_max_num_f32_e32 v53, v53, v53
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v14.l, v49.h, v34.l, s1
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v30.l
; GFX12-TRUE16-NEXT:    v_cmp_eq_f32_e64 s0, 0, v54
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v54, 16, v55
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v55, 16, v64
; GFX12-TRUE16-NEXT:    v_bfe_u32 v65, v53, 16, 1
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v14.l, v14.l, v30.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v66, 0xffff0000, v49
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s1, v53, v53
; GFX12-TRUE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v55, v54
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v54, 16, v11
; GFX12-TRUE16-NEXT:    v_add3_u32 v64, v65, v53, 0x7fff
; GFX12-TRUE16-NEXT:    v_or_b32_e32 v65, 0x400000, v53
; GFX12-TRUE16-NEXT:    v_cmp_eq_f32_e64 s2, 0, v66
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v15.l, v28.l, v12.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v54, v54
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v54, 16, v27
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b32_e64 v53, v64, v65, s1
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s1, 0, v13.l
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v55.l, v15.l
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v11.l, v11.l, v27.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v54, v54
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v14.l, v49.h, v14.l, s2
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v11.h, v53.h, v13.l, s1
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s1, 0, v29.l
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v49, 16, v55
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v13.l, v27.l, v11.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v27, 16, v10
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v54.l, v11.l
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v11.h, v11.h, v29.l, s1
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v29, 0xffff0000, v53
; GFX12-TRUE16-NEXT:    v_max_num_f32_e32 v49, v49, v49
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v52.l, v13.l
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v27, v27
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v27, 16, v26
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v15.l, v52.h, v12.h, s0
; GFX12-TRUE16-NEXT:    v_bfe_u32 v55, v49, 16, 1
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v52, 16, v52
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v54, 16, v54
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v10.l, v10.l, v26.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v27, v27
; GFX12-TRUE16-NEXT:    v_cmp_eq_f32_e64 s0, 0, v29
; GFX12-TRUE16-NEXT:    v_add3_u32 v27, v55, v49, 0x7fff
; GFX12-TRUE16-NEXT:    v_or_b32_e32 v29, 0x400000, v49
; GFX12-TRUE16-NEXT:    v_cmp_gt_f32_e64 s1, v54, v52
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v26.l, v26.l, v10.l, vcc_lo
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v34.l, v53.h, v11.h, s0
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v49, v49
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v53, 16, v9
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v52.l, v10.l
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v49.l, v26.l
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b32_e32 v29, v27, v29, vcc_lo
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v27.l, v13.l, v11.l, s1
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v12.l
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s0, v53, v53
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v53, 16, v25
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v49, 16, v49
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v52, 16, v52
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v9.h, v29.h, v12.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 16, v27
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v9.l, v9.l, v25.l, s0
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s0, v53, v53
; GFX12-TRUE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v52, v49
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s1, 0, v28.l
; GFX12-TRUE16-NEXT:    v_max_num_f32_e32 v27, v12, v12
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v49, 0xffff0000, v29
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v25.l, v25.l, v9.l, s0
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v12.l, v26.l, v10.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v9.h, v9.h, v28.l, s1
; GFX12-TRUE16-NEXT:    v_bfe_u32 v28, v27, 16, 1
; GFX12-TRUE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v49
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v49.l, v25.l
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v52.l, v9.l
; GFX12-TRUE16-NEXT:    v_or_b32_e32 v53, 0x400000, v27
; GFX12-TRUE16-NEXT:    v_add3_u32 v28, v28, v27, 0x7fff
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v49, 16, v49
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v52, 16, v52
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s0, v27, v27
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v36.l, v29.h, v9.h, vcc_lo
; GFX12-TRUE16-NEXT:    v_max_num_f32_e32 v12, v12, v12
; GFX12-TRUE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v52, v49
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b32_e64 v27, v28, v53, s0
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s0, 0, v11.l
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v28, 16, v8
; GFX12-TRUE16-NEXT:    v_bfe_u32 v29, v12, 16, 1
; GFX12-TRUE16-NEXT:    v_or_b32_e32 v49, 0x400000, v12
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s1, v12, v12
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v8.h, v27.h, v11.l, s0
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v11.l, v25.l, v9.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v28, v28
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v28, 16, v24
; GFX12-TRUE16-NEXT:    v_add3_u32 v29, v29, v12, 0x7fff
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s0, 0, v13.l
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v8.l, v8.l, v24.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v28, v28
; GFX12-TRUE16-NEXT:    v_cndmask_b32_e64 v12, v29, v49, s1
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s1, 0, v10.l
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v28, 16, v11
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v29.l, v8.l
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v11.l, v24.l, v8.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v24, 0xffff0000, v27
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v9.h, v12.h, v10.l, s1
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v26.l
; GFX12-TRUE16-NEXT:    v_max_num_f32_e32 v10, v28, v28
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v28.l, v11.l
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v8.h, v8.h, v13.l, s0
; GFX12-TRUE16-NEXT:    v_cmp_eq_f32_e64 s0, 0, v24
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v9.h, v9.h, v26.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_bfe_u32 v24, v10, 16, 1
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v26, 16, v28
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v28, 16, v29
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v29, 16, v7
; GFX12-TRUE16-NEXT:    v_or_b32_e32 v52, 0x400000, v10
; GFX12-TRUE16-NEXT:    v_add3_u32 v24, v24, v10, 0x7fff
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s2, v10, v10
; GFX12-TRUE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v28, v26
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s1, v29, v29
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v26, 16, v23
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v49, 0xffff0000, v12
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b32_e64 v24, v24, v52, s2
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s2, 0, v9.l
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v10.l, v11.l, v8.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v7.l, v7.l, v23.l, s1
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v26, v26
; GFX12-TRUE16-NEXT:    v_cmp_eq_f32_e64 s1, 0, v49
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v7.h, v24.h, v9.l, s2
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s2, 0, v25.l
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v9.l, v23.l, v7.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v23.l, v7.l
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v31.l, v27.h, v8.h, s0
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v7.h, v7.h, v25.l, s2
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v25, 16, v6
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v12.l, v9.l
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v32.l, v12.h, v9.h, s1
; GFX12-TRUE16-NEXT:    v_dual_max_num_f32 v27, v10, v10 :: v_dual_and_b32 v26, 0xffff0000, v24
; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v25, v25
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v10, 16, v12
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 16, v23
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v23, 16, v22
; GFX12-TRUE16-NEXT:    v_bfe_u32 v25, v27, 16, 1
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v6.l, v6.l, v22.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_cmp_eq_f32_e64 s0, 0, v26
; GFX12-TRUE16-NEXT:    v_cmp_gt_f32_e64 s1, v12, v10
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v23, v23
; GFX12-TRUE16-NEXT:    v_add3_u32 v23, v25, v27, 0x7fff
; GFX12-TRUE16-NEXT:    v_or_b32_e32 v25, 0x400000, v27
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v24.l, v6.l
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v10.l, v9.l, v7.l, s1
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v12.l, v22.l, v6.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v27, v27
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v33.l, v24.h, v7.h, s0
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s0, 0, v11.l
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b32_e32 v22, v23, v25, vcc_lo
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v8.l
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v23.l, v12.l
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v25, 16, v5
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v5.h, v22.h, v8.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 16, v10
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v10, 16, v23
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v23, 16, v24
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v25, v25
; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX12-TRUE16-NEXT:    v_dual_max_num_f32 v25, v8, v8 :: v_dual_lshlrev_b32 v24, 16, v21
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v5.h, v5.h, v11.l, s0
; GFX12-TRUE16-NEXT:    v_cmp_gt_f32_e64 s1, v23, v10
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v5.l, v5.l, v21.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v24, v24
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v23, 0xffff0000, v22
; GFX12-TRUE16-NEXT:    v_bfe_u32 v24, v25, 16, 1
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v8.l, v12.l, v6.l, s1
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s0, v25, v25
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v10.l, v21.l, v5.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v23
; GFX12-TRUE16-NEXT:    v_add3_u32 v11, v24, v25, 0x7fff
; GFX12-TRUE16-NEXT:    v_or_b32_e32 v21, 0x400000, v25
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v24.l, v5.l
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v23.l, v10.l
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v35.l, v22.h, v5.h, vcc_lo
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b32_e64 v11, v11, v21, s0
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s0, 0, v7.l
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v21, 16, v23
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v23, 16, v24
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v24, 16, v4
; GFX12-TRUE16-NEXT:    v_max_num_f32_e32 v8, v8, v8
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v4.h, v11.h, v7.l, s0
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s1, 0, v9.l
; GFX12-TRUE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v23, v21
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s0, v24, v24
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v21, 16, v20
; GFX12-TRUE16-NEXT:    v_bfe_u32 v22, v8, 16, 1
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v4.h, v4.h, v9.l, s1
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v7.l, v10.l, v5.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v4.l, v4.l, v20.l, s0
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v21, v21
; GFX12-TRUE16-NEXT:    v_add3_u32 v9, v22, v8, 0x7fff
; GFX12-TRUE16-NEXT:    v_or_b32_e32 v21, 0x400000, v8
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v22.l, v7.l
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v7.l, v20.l, v4.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v20, 0xffff0000, v11
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_dual_cndmask_b32 v8, v9, v21 :: v_dual_lshlrev_b32 v9, 16, v22
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v21.l, v7.l
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v22.l, v4.l
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v6.l
; GFX12-TRUE16-NEXT:    v_cmp_eq_f32_e64 s0, 0, v20
; GFX12-TRUE16-NEXT:    v_max_num_f32_e32 v9, v9, v9
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v20, 16, v21
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v21, 16, v22
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v5.h, v8.h, v6.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v12.l
; GFX12-TRUE16-NEXT:    v_bfe_u32 v6, v9, 16, 1
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v37.l, v11.h, v4.h, s0
; GFX12-TRUE16-NEXT:    v_cmp_gt_f32_e64 s0, v21, v20
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v11, 16, v3
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v3.h, v5.h, v12.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_add3_u32 v12, v6, v9, 0x7fff
; GFX12-TRUE16-NEXT:    v_or_b32_e32 v20, 0x400000, v9
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v6.l, v7.l, v4.l, s0
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v11, v11
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v11, 16, v19
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s0, v9, v9
; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v3.l, v3.l, v19.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v11, v11
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b32_e64 v9, v12, v20, s0
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s0, 0, v5.l
; GFX12-TRUE16-NEXT:    v_dual_max_num_f32 v6, v6, v6 :: v_dual_and_b32 v11, 0xffff0000, v8
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 16, v2
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v4.h, v9.h, v5.l, s0
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s0, 0, v10.l
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v5.l, v19.l, v3.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v11
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v11, 0xffff0000, v9
; GFX12-TRUE16-NEXT:    v_bfe_u32 v19, v6, 16, 1
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v4.h, v4.h, v10.l, s0
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v10.l, v5.l
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s1, v12, v12
; GFX12-TRUE16-NEXT:    v_cmp_eq_f32_e64 s0, 0, v11
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v11.l, v3.l
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 16, v18
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
; GFX12-TRUE16-NEXT:    v_add3_u32 v19, v19, v6, 0x7fff
; GFX12-TRUE16-NEXT:    v_or_b32_e32 v20, 0x400000, v6
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v2.l, v2.l, v18.l, s1
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s1, v12, v12
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s3, v6, v6
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v39.l, v9.h, v4.h, s0
; GFX12-TRUE16-NEXT:    v_cmp_gt_f32_e64 s2, v11, v10
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 16, v1
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v6.l, v18.l, v2.l, s1
; GFX12-TRUE16-NEXT:    v_cndmask_b32_e64 v11, v19, v20, s3
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s3, 0, v4.l
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v38.l, v8.h, v3.h, vcc_lo
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v10.l, v2.l
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 16, v0
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v2.h, v11.h, v4.l, s3
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v4.l, v6.l
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 16, v17
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s0, v12, v12
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 16, v16
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v1.l, v1.l, v17.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v8.l, v5.l, v3.l, s2
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v0.l, v0.l, v16.l, s0
; GFX12-TRUE16-NEXT:    v_cmp_gt_f32_e64 s1, v10, v4
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s0, v12, v12
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v4.l, v17.l, v1.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v10, 16, v8
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v18.l, v0.l
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v8.l, v6.l, v2.l, s1
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v9.l, v16.l, v0.l, s0
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v12.l, v4.l
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v16.l, v1.l
; GFX12-TRUE16-NEXT:    v_max_num_f32_e32 v10, v10, v10
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v17.l, v9.l
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
; GFX12-TRUE16-NEXT:    v_bfe_u32 v19, v10, 16, 1
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v7.l
; GFX12-TRUE16-NEXT:    v_cmp_gt_f32_e64 s0, v16, v12
; GFX12-TRUE16-NEXT:    v_add3_u32 v12, v19, v10, 0x7fff
; GFX12-TRUE16-NEXT:    v_or_b32_e32 v16, 0x400000, v10
; GFX12-TRUE16-NEXT:    v_cmp_gt_f32_e64 s1, v18, v17
; GFX12-TRUE16-NEXT:    v_max_num_f32_e32 v17, v8, v8
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v8.l, v4.l, v1.l, s0
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s0, v10, v10
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v0.h, v2.h, v7.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v10.l, v9.l, v0.l, s1
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v5.l
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b32_e64 v12, v12, v16, s0
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s0, 0, v3.l
; GFX12-TRUE16-NEXT:    v_bfe_u32 v16, v17, 16, 1
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v1.h, v12.h, v3.l, s0
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v8
; GFX12-TRUE16-NEXT:    v_add3_u32 v7, v16, v17, 0x7fff
; GFX12-TRUE16-NEXT:    v_or_b32_e32 v8, 0x400000, v17
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s0, v17, v17
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v1.h, v1.h, v5.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_max_num_f32_e32 v3, v3, v3
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v6.l
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v16, 0xffff0000, v11
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b32_e64 v7, v7, v8, s0
; GFX12-TRUE16-NEXT:    v_max_num_f32_e32 v8, v10, v10
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s0, 0, v2.l
; GFX12-TRUE16-NEXT:    v_bfe_u32 v10, v3, 16, 1
; GFX12-TRUE16-NEXT:    v_or_b32_e32 v17, 0x400000, v3
; GFX12-TRUE16-NEXT:    v_cmp_eq_f32_e64 s3, 0, v16
; GFX12-TRUE16-NEXT:    v_bfe_u32 v5, v8, 16, 1
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v2.l, v7.h, v2.l, s0
; GFX12-TRUE16-NEXT:    v_add3_u32 v10, v10, v3, 0x7fff
; GFX12-TRUE16-NEXT:    v_or_b32_e32 v18, 0x400000, v8
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s0, v3, v3
; GFX12-TRUE16-NEXT:    v_add3_u32 v5, v5, v8, 0x7fff
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s1, v8, v8
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v8, 0xffff0000, v12
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v2.l, v2.l, v6.l, vcc_lo
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b32_e64 v3, v10, v17, s0
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s0, 0, v1.l
; GFX12-TRUE16-NEXT:    v_cndmask_b32_e64 v5, v5, v18, s1
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s1, 0, v0.l
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v4.l
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v10, 0xffff0000, v3
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v1.l, v3.h, v1.l, s0
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s0, 0, v9.l
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v0.l, v5.h, v0.l, s1
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v6, 0xffff0000, v5
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v17, 0xffff0000, v7
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v1.l, v1.l, v4.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_cmp_eq_f32_e64 s2, 0, v8
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v0.l, v0.l, v9.l, s0
; GFX12-TRUE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v6
; GFX12-TRUE16-NEXT:    v_cmp_eq_f32_e64 s0, 0, v10
; GFX12-TRUE16-NEXT:    v_cmp_eq_f32_e64 s1, 0, v17
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v50.l, v12.h, v1.h, s2
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v48.l, v11.h, v0.h, s3
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v13.l, v5.h, v0.l, vcc_lo
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v30.l, v3.h, v1.l, s0
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v51.l, v7.h, v2.l, s1
; GFX12-TRUE16-NEXT:    v_dual_mov_b32 v3, v50 :: v_dual_mov_b32 v4, v48
; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX12-TRUE16-NEXT:    v_dual_mov_b32 v0, v13 :: v_dual_mov_b32 v1, v30
; GFX12-TRUE16-NEXT:    v_dual_mov_b32 v2, v51 :: v_dual_mov_b32 v5, v39
; GFX12-TRUE16-NEXT:    v_dual_mov_b32 v6, v38 :: v_dual_mov_b32 v7, v37
; GFX12-TRUE16-NEXT:    v_dual_mov_b32 v8, v35 :: v_dual_mov_b32 v9, v33
; GFX12-TRUE16-NEXT:    v_dual_mov_b32 v10, v32 :: v_dual_mov_b32 v11, v31
; GFX12-TRUE16-NEXT:    v_dual_mov_b32 v12, v36 :: v_dual_mov_b32 v13, v34
; GFX12-TRUE16-NEXT:    s_setpc_b64 s[30:31]
;
; GFX12-FAKE16-LABEL: v_maximumnum_v32bf16:
; GFX12-FAKE16:       ; %bb.0:
; GFX12-FAKE16-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX12-FAKE16-NEXT:    s_wait_expcnt 0x0
; GFX12-FAKE16-NEXT:    s_wait_samplecnt 0x0
; GFX12-FAKE16-NEXT:    s_wait_bvhcnt 0x0
; GFX12-FAKE16-NEXT:    s_wait_kmcnt 0x0
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v28
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v52, 16, v12
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v34, 0xffff0000, v14
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v30
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v14
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v71, 16, v25
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v80, 16, v9
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v37, 0xffff0000, v13
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v34, v34
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v31, 0xffff0000, v15
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v29
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v13
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v83, 16, v24
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v34, v35, v33, vcc_lo
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v84, 16, v8
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v49, 0xffff0000, v12
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v37, v37
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v48, 0xffff0000, v29
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v54, 0xffff0000, v11
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v55, 16, v27
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v64, 16, v11
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v35, v39, v38, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v49, v49
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v67, 16, v26
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v68, 16, v10
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v66, 0xffff0000, v10
; GFX12-FAKE16-NEXT:    scratch_load_b32 v50, off, s32
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v37, v52, v51, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v54, v54
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v36, 0xffff0000, v30
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v82, 0xffff0000, v8
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v87, 16, v23
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v96, 16, v7
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v39, v64, v55 :: v_dual_and_b32 v70, 0xffff0000, v9
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v66, v66
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v99, 16, v22
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v100, 16, v6
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v103, 16, v21
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v112, 16, v5
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v49, v68, v67, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v70, v70
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v86, 0xffff0000, v7
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v115, 16, v20
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v116, 16, v4
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v119, 16, v19
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v52, v80, v71, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v82, v82
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v98, 0xffff0000, v6
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v128, 16, v3
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v131, 16, v18
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v132, 16, v2
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v54, v84, v83, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v86, v86
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v53, 0xffff0000, v28
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v102, 0xffff0000, v5
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v135, 16, v17
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v144, 16, v1
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v64, v96, v87, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v98, v98
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v69, 0xffff0000, v26
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v118, 0xffff0000, v3
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v15
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v66, v100, v99, vcc_lo
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v114, 0xffff0000, v4
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v102, v102
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v68, v112, v103 :: v_dual_and_b32 v81, 0xffff0000, v25
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v114, v114
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v85, 0xffff0000, v24
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v130, 0xffff0000, v2
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v70, v116, v115, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v118, v118
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v101, 0xffff0000, v22
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v134, 0xffff0000, v1
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v80, v128, v119, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v130, v130
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v82, v132, v131 :: v_dual_and_b32 v113, 0xffff0000, v21
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v134, v134
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v84, v144, v135 :: v_dual_and_b32 v117, 0xffff0000, v20
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v36, v36
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v33, v33, v34, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v48, v48
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v65, 0xffff0000, v27
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v133, 0xffff0000, v18
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v114, 16, v33
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v38, v38, v35, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v53, v53
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v51, v51, v37, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v65, v65
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v97, 0xffff0000, v23
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v116, 16, v51
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v55, v55, v39, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v69, v69
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v48, 16, v37
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v67, v67, v49, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v81, v81
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v71, v71, v52 :: v_dual_lshlrev_b32 v118, 16, v67
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v85, v85
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v83, v83, v54, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v97, v97
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v87, v87, v64 :: v_dual_lshlrev_b32 v128, 16, v83
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v101, v101
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v97, v99, v66, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v113, v113
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v129, 0xffff0000, v19
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v99, v103, v68, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v117, v117
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v53, 16, v39
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v101, v115, v70, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v129, v129
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v86, 16, v34
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v115, 16, v38
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v65, 16, v49
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v129, 16, v87
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v103, v119, v80, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v133, v133
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v36, 16, v35
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v119, 16, v71
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v81, 16, v54
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v133, 16, v103
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v113, v131, v82, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v86, v114
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v69, 16, v52
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v131, 16, v99
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_3)
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v134, 16, v113
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v86, v33, v34, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v36, v115
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v145, 0xffff0000, v17
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v86, 16, v86
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v36, v38, v35, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v48, v116
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_2)
; GFX12-FAKE16-NEXT:    v_dual_max_num_f32 v86, v86, v86 :: v_dual_lshlrev_b32 v85, 16, v64
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v48, v51, v37, vcc_lo
; GFX12-FAKE16-NEXT:    v_bfe_u32 v114, v86, 16, 1
; GFX12-FAKE16-NEXT:    v_or_b32_e32 v115, 0x400000, v86
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v48, 16, v48
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v117, 16, v55
; GFX12-FAKE16-NEXT:    v_add3_u32 v114, v114, v86, 0x7fff
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX12-FAKE16-NEXT:    v_max_num_f32_e32 v48, v48, v48
; GFX12-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v53, v117
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v53, v55, v39, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v65, v118
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v96, 16, v66
; GFX12-FAKE16-NEXT:    v_bfe_u32 v118, v48, 16, 1
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v53, 16, v53
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v65, v67, v49, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v69, v119
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v100, 16, v70
; GFX12-FAKE16-NEXT:    v_dual_max_num_f32 v53, v53, v53 :: v_dual_lshlrev_b32 v102, 16, v80
; GFX12-FAKE16-NEXT:    v_or_b32_e32 v119, 0x400000, v48
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v69, v71, v52, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v81, v128
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v98, 16, v68
; GFX12-FAKE16-NEXT:    v_add3_u32 v118, v118, v48, 0x7fff
; GFX12-FAKE16-NEXT:    v_bfe_u32 v128, v53, 16, 1
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v81, v83, v54, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v85, v129
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v112, 16, v82
; GFX12-FAKE16-NEXT:    v_or_b32_e32 v129, 0x400000, v53
; GFX12-FAKE16-NEXT:    v_add3_u32 v128, v128, v53, 0x7fff
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v81, 16, v81
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v85, v87, v64, vcc_lo
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX12-FAKE16-NEXT:    v_max_num_f32_e32 v81, v81, v81
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v85, 16, v85
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX12-FAKE16-NEXT:    v_or_b32_e32 v144, 0x400000, v81
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v130, 16, v97
; GFX12-FAKE16-NEXT:    v_max_num_f32_e32 v85, v85, v85
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX12-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v96, v130
; GFX12-FAKE16-NEXT:    v_bfe_u32 v146, v85, 16, 1
; GFX12-FAKE16-NEXT:    v_or_b32_e32 v147, 0x400000, v85
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v96, v97, v66, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v98, v131
; GFX12-FAKE16-NEXT:    v_add3_u32 v146, v146, v85, 0x7fff
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v132, 16, v101
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v96, 16, v96
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v98, v99, v68, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v100, v132
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX12-FAKE16-NEXT:    v_max_num_f32_e32 v96, v96, v96
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v98, 16, v98
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v100, v101, v70, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v102, v133
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX12-FAKE16-NEXT:    v_max_num_f32_e32 v98, v98, v98
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v100, 16, v100
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v102, v103, v80, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v112, v134
; GFX12-FAKE16-NEXT:    v_bfe_u32 v134, v81, 16, 1
; GFX12-FAKE16-NEXT:    v_max_num_f32_e32 v100, v100, v100
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v112, v113, v82, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v86, v86
; GFX12-FAKE16-NEXT:    v_bfe_u32 v86, v96, 16, 1
; GFX12-FAKE16-NEXT:    v_add3_u32 v134, v134, v81, 0x7fff
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v114, v114, v115, vcc_lo
; GFX12-FAKE16-NEXT:    v_or_b32_e32 v115, 0x400000, v96
; GFX12-FAKE16-NEXT:    v_add3_u32 v86, v86, v96, 0x7fff
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v69, 16, v69
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-FAKE16-NEXT:    v_max_num_f32_e32 v69, v69, v69
; GFX12-FAKE16-NEXT:    v_bfe_u32 v132, v69, 16, 1
; GFX12-FAKE16-NEXT:    v_or_b32_e32 v133, 0x400000, v69
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX12-FAKE16-NEXT:    v_add3_u32 v132, v132, v69, 0x7fff
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v36, 16, v36
; GFX12-FAKE16-NEXT:    v_max_num_f32_e32 v36, v36, v36
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX12-FAKE16-NEXT:    v_bfe_u32 v116, v36, 16, 1
; GFX12-FAKE16-NEXT:    v_or_b32_e32 v117, 0x400000, v36
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v36, v36
; GFX12-FAKE16-NEXT:    v_add3_u32 v116, v116, v36, 0x7fff
; GFX12-FAKE16-NEXT:    v_bfe_u32 v36, v98, 16, 1
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v116, v116, v117, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v48, v48
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v65, 16, v65
; GFX12-FAKE16-NEXT:    v_or_b32_e32 v117, 0x400000, v98
; GFX12-FAKE16-NEXT:    v_bfe_u32 v48, v100, 16, 1
; GFX12-FAKE16-NEXT:    v_add3_u32 v36, v36, v98, 0x7fff
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v118, v118, v119 :: v_dual_max_num_f32 v65, v65, v65
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v53, v53
; GFX12-FAKE16-NEXT:    v_or_b32_e32 v119, 0x400000, v100
; GFX12-FAKE16-NEXT:    v_add3_u32 v48, v48, v100, 0x7fff
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX12-FAKE16-NEXT:    v_bfe_u32 v130, v65, 16, 1
; GFX12-FAKE16-NEXT:    v_or_b32_e32 v131, 0x400000, v65
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v128, v128, v129, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v65, v65
; GFX12-FAKE16-NEXT:    v_add3_u32 v130, v130, v65, 0x7fff
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v65, v130, v131, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v69, v69
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v69, v132, v133, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v81, v81
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v81, v134, v144, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v85, v85
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v130, 16, v81
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v85, v146, v147, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v96, v96
; GFX12-FAKE16-NEXT:    s_wait_loadcnt 0x0
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v146, 0xffff0000, v50
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v131, 16, v85
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v86, v86, v115, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v98, v98
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v115, 16, v118
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v98, 0xffff0000, v114
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v132, 16, v86
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v36, v36, v117, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v100, v100
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v100, 16, v116
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v117, 16, v128
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v133, 16, v36
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v96, v48, v119, vcc_lo
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v48, 16, v114
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v34
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v119, 16, v65
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v134, 0xffff0000, v36
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v114, 0xffff0000, v116
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v116, 0xffff0000, v118
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v34, v48, v34, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v35
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v118, 0xffff0000, v128
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v128, 16, v69
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v144, 16, v96
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v35, v100, v35, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v37
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v36, v115, v37, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v39
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v37, v117, v39, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v49
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v39, v119, v49, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v52
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v49, v128, v52, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v54
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v52, v130, v54 :: v_dual_and_b32 v65, 0xffff0000, v65
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v64
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v69, 0xffff0000, v69
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v54, 16, v50
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v64, v131, v64, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v31, v31
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v147, v32, v54, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v66
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v31, v132, v66, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v68
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v32, v133, v68 :: v_dual_and_b32 v81, 0xffff0000, v81
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v146, v146
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v54, v54, v147 :: v_dual_and_b32 v85, 0xffff0000, v85
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v70
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v68, 16, v54
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v66, v144, v70, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v33
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v102, 16, v102
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v70, 16, v0
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v33, v34, v33, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v38
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v34, 16, v147
; GFX12-FAKE16-NEXT:    v_max_num_f32_e32 v102, v102, v102
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v35, v35, v38, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v51
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX12-FAKE16-NEXT:    v_bfe_u32 v53, v102, 16, 1
; GFX12-FAKE16-NEXT:    v_or_b32_e32 v129, 0x400000, v102
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v36, v36, v51, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v34, v68
; GFX12-FAKE16-NEXT:    v_add3_u32 v53, v53, v102, 0x7fff
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v68, 0xffff0000, v0
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v34, v54, v147, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v55
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v37, v37, v55 :: v_dual_lshlrev_b32 v34, 16, v34
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v67
; GFX12-FAKE16-NEXT:    v_max_num_f32_e32 v34, v34, v34
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v38, v39, v67, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v71
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX12-FAKE16-NEXT:    v_bfe_u32 v55, v34, 16, 1
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v39, v49, v71, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v83
; GFX12-FAKE16-NEXT:    v_add3_u32 v55, v55, v34, 0x7fff
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v49, v52, v83, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v87
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v51, v64, v87, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v97
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v52, v31, v97, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v99
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v64, v32, v99, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v98
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v31, v48, v33, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v114
; GFX12-FAKE16-NEXT:    v_or_b32_e32 v48, 0x400000, v34
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v32, v100, v35, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v116
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v33, v115, v36 :: v_dual_and_b32 v86, 0xffff0000, v86
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v34, v34
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v48, v55, v48, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v118
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v55, 16, v48
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v34, v117, v37, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v65
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v65, 0xffff0000, v96
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v35, v119, v38, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v69
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v69, 16, v16
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v36, v128, v39, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v81
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v37, v130, v49, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e64 vcc_lo, 0, v147
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v39, v55, v147, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v85
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v38, v131, v51, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v54
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v51, 0xffff0000, v48
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v39, v39, v54, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v86
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v48, v132, v52, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v134
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v49, v133, v64, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v51
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v51, 16, v112
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v39, v55, v39, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v145, v145
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_2)
; GFX12-FAKE16-NEXT:    v_max_num_f32_e32 v67, v51, v51
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v54, v135, v84, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v102, v102
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v52, v53, v129 :: v_dual_lshlrev_b32 v55, 16, v54
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v53, 16, v84
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v101
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v64, v66, v101, vcc_lo
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX12-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v53, v55
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v66, 16, v52
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v52, 0xffff0000, v52
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v53, v54, v84, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v80
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v53, 16, v53
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v55, v66, v80, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v65
; GFX12-FAKE16-NEXT:    v_or_b32_e32 v65, 0x400000, v67
; GFX12-FAKE16-NEXT:    v_dual_max_num_f32 v53, v53, v53 :: v_dual_lshlrev_b32 v80, 16, v15
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v51, v144, v64, vcc_lo
; GFX12-FAKE16-NEXT:    v_bfe_u32 v64, v67, 16, 1
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v103
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
; GFX12-FAKE16-NEXT:    v_add3_u32 v64, v64, v67, 0x7fff
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v55, v55, v103, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v67, v67
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v64, v64, v65, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v68, v68
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v65, 0xffff0000, v16
; GFX12-FAKE16-NEXT:    v_bfe_u32 v68, v53, 16, 1
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v67, v70, v69, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v52
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v52, v66, v55, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v65, v65
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v65, 16, v64
; GFX12-FAKE16-NEXT:    v_add3_u32 v66, v68, v53, 0x7fff
; GFX12-FAKE16-NEXT:    v_or_b32_e32 v68, 0x400000, v53
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v55, v69, v67, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v82
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v69, 16, v67
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v71, v65, v82 :: v_dual_lshlrev_b32 v70, 16, v55
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v53, v53
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v53, v66, v68, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v69, v70
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v70, 16, v53
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v66, v55, v67, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v80, v80
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v68, v15, v50 :: v_dual_and_b32 v53, 0xffff0000, v53
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v15, 16, v50
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v113
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v69, v71, v113 :: v_dual_lshlrev_b32 v80, 16, v68
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v15, v15
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v15, 0xffff0000, v64
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v50, v50, v68, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v84
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v64, v70, v84, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v15
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v15, v65, v69, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v54
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v69, 16, v14
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v71, 16, v50
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v54, v64, v54, vcc_lo
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_2)
; GFX12-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v80, v71
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v65, v50, v68 :: v_dual_lshlrev_b32 v80, 16, v30
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v69, v69
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v65, 16, v65
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v66, 16, v66
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v69, v14, v30, vcc_lo
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-FAKE16-NEXT:    v_dual_max_num_f32 v65, v65, v65 :: v_dual_max_num_f32 v66, v66, v66
; GFX12-FAKE16-NEXT:    v_bfe_u32 v64, v66, 16, 1
; GFX12-FAKE16-NEXT:    v_or_b32_e32 v71, 0x400000, v66
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v66, v66
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX12-FAKE16-NEXT:    v_add3_u32 v64, v64, v66, 0x7fff
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v64, v64, v71 :: v_dual_lshlrev_b32 v71, 16, v13
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v80, v80
; GFX12-FAKE16-NEXT:    v_or_b32_e32 v80, 0x400000, v65
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v30, v30, v69, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v53
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v53, 16, v64
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v14, v70, v54, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v67
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v70, 16, v69
; GFX12-FAKE16-NEXT:    v_bfe_u32 v54, v65, 16, 1
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v67, v53, v67 :: v_dual_lshlrev_b32 v66, 16, v30
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX12-FAKE16-NEXT:    v_add3_u32 v54, v54, v65, 0x7fff
; GFX12-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v70, v66
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v66, v30, v69, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v71, v71
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v71, v13, v29, vcc_lo
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v70, 16, v29
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v65, v65
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v54, v54, v80, vcc_lo
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v70, v70
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v65, v29, v71, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v55
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v29, v67, v55, vcc_lo
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v67, 16, v65
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v55, 0xffff0000, v64
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v13, 16, v66
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v64, 16, v54
; GFX12-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v55
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX12-FAKE16-NEXT:    v_max_num_f32_e32 v66, v13, v13
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v13, v53, v29 :: v_dual_lshlrev_b32 v70, 16, v71
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v68
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v29, v64, v68, vcc_lo
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_2)
; GFX12-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v70, v67
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v67, 16, v12
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v55, v65, v71, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v50
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v55, 16, v55
; GFX12-FAKE16-NEXT:    v_bfe_u32 v53, v66, 16, 1
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v29, v29, v50, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v66, v66
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v54, 0xffff0000, v54
; GFX12-FAKE16-NEXT:    v_add3_u32 v50, v53, v66, 0x7fff
; GFX12-FAKE16-NEXT:    v_or_b32_e32 v53, 0x400000, v66
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v66, 16, v11
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v50, v50, v53 :: v_dual_max_num_f32 v53, v55, v55
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v67, v67
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v55, 16, v28
; GFX12-FAKE16-NEXT:    v_or_b32_e32 v67, 0x400000, v53
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v12, v12, v28, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v54
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v54, 16, v50
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v29, v64, v29, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v55, v55
; GFX12-FAKE16-NEXT:    v_bfe_u32 v64, v53, 16, 1
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v28, v28, v12, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v69
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX12-FAKE16-NEXT:    v_add3_u32 v64, v64, v53, 0x7fff
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v55, v54, v69, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v66, v66
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v66, 16, v27
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v11, v11, v27, vcc_lo
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v69, 16, v12
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v53, v53
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v53, v64, v67 :: v_dual_lshlrev_b32 v68, 16, v28
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX12-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v69, v68
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v64, v28, v12 :: v_dual_lshlrev_b32 v67, 16, v11
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v66, v66
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v66, v27, v11, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v30
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v27, v55, v30, vcc_lo
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v53
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v71
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v50, 0xffff0000, v50
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v68, v30, v71 :: v_dual_lshlrev_b32 v55, 16, v64
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v64, 16, v66
; GFX12-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v67, v64
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v64, v66, v11, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v50
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v27, v54, v27, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v65
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v54, v68, v65 :: v_dual_max_num_f32 v55, v55, v55
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v68, 16, v26
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX12-FAKE16-NEXT:    v_bfe_u32 v50, v55, 16, 1
; GFX12-FAKE16-NEXT:    v_or_b32_e32 v67, 0x400000, v55
; GFX12-FAKE16-NEXT:    v_add3_u32 v50, v50, v55, 0x7fff
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v65, 16, v10
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v65, v65
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v10, v10, v26, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v55, v55
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v50, v50, v67, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v68, v68
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v65, v26, v10 :: v_dual_lshlrev_b32 v64, 16, v64
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-FAKE16-NEXT:    v_dual_max_num_f32 v64, v64, v64 :: v_dual_and_b32 v53, 0xffff0000, v53
; GFX12-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v53
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
; GFX12-FAKE16-NEXT:    v_bfe_u32 v55, v64, 16, 1
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v26, v30, v54, vcc_lo
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v50
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v12
; GFX12-FAKE16-NEXT:    v_add3_u32 v53, v55, v64, 0x7fff
; GFX12-FAKE16-NEXT:    v_or_b32_e32 v54, 0x400000, v64
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v55, 16, v65
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v12, v30, v12 :: v_dual_lshlrev_b32 v67, 16, v10
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v64, v64
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v68, 16, v9
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v50, 0xffff0000, v50
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v53, v53, v54, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v67, v55
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v54, v65, v10 :: v_dual_lshlrev_b32 v55, 16, v25
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v68, v68
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v9, v9, v25, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v28
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v12, v12, v28, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v55, v55
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v28, 16, v53
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v53, 0xffff0000, v53
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v25, v25, v9, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v11
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v55, 16, v25
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v11, v28, v11, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v50
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v12, v30, v12, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v66
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v64, 16, v9
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX12-FAKE16-NEXT:    v_perm_b32 v12, v33, v12, 0x5040100
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v11, v11, v66, vcc_lo
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v54, 16, v54
; GFX12-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v64, v55
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v55, 16, v8
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v66, 16, v24
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX12-FAKE16-NEXT:    v_max_num_f32_e32 v54, v54, v54
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v50, v25, v9, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v55, v55
; GFX12-FAKE16-NEXT:    v_bfe_u32 v30, v54, 16, 1
; GFX12-FAKE16-NEXT:    v_or_b32_e32 v64, 0x400000, v54
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v8, v8, v24, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v54, v54
; GFX12-FAKE16-NEXT:    v_add3_u32 v30, v30, v54, 0x7fff
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v30, v30, v64, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v66, v66
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v64, 16, v7
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v24, v24, v8, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v53
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_4)
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v11, v28, v11 :: v_dual_lshlrev_b32 v54, 16, v24
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v28, 16, v30
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v10
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v55, 16, v8
; GFX12-FAKE16-NEXT:    v_perm_b32 v11, v34, v11, 0x5040100
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v10, v28, v10, vcc_lo
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX12-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v55, v54
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v54, v24, v8 :: v_dual_lshlrev_b32 v55, 16, v23
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v64, v64
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v7, v7, v23 :: v_dual_lshlrev_b32 v50, 16, v50
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v64, 16, v7
; GFX12-FAKE16-NEXT:    v_max_num_f32_e32 v50, v50, v50
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX12-FAKE16-NEXT:    v_bfe_u32 v53, v50, 16, 1
; GFX12-FAKE16-NEXT:    v_or_b32_e32 v66, 0x400000, v50
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v50, v50
; GFX12-FAKE16-NEXT:    v_add3_u32 v53, v53, v50, 0x7fff
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_4)
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v50, v53, v66, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v55, v55
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v53, 16, v54
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v30, 0xffff0000, v30
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v54, 16, v50
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v23, v23, v7, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v65
; GFX12-FAKE16-NEXT:    v_dual_max_num_f32 v53, v53, v53 :: v_dual_and_b32 v50, 0xffff0000, v50
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_2)
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v10, v10, v65 :: v_dual_lshlrev_b32 v55, 16, v23
; GFX12-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v30
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v10, v28, v10, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v9
; GFX12-FAKE16-NEXT:    v_bfe_u32 v28, v53, 16, 1
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v9, v54, v9, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v64, v55
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v30, v23, v7 :: v_dual_lshlrev_b32 v55, 16, v6
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v25
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_4) | instid1(VALU_DEP_2)
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v9, v9, v25 :: v_dual_lshlrev_b32 v30, 16, v30
; GFX12-FAKE16-NEXT:    v_add3_u32 v25, v28, v53, 0x7fff
; GFX12-FAKE16-NEXT:    v_or_b32_e32 v28, 0x400000, v53
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v53, v53
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v25, v25, v28, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v55, v55
; GFX12-FAKE16-NEXT:    v_max_num_f32_e32 v28, v30, v30
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v30, 16, v22
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v6, v6, v22, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v50
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v50, 16, v25
; GFX12-FAKE16-NEXT:    v_bfe_u32 v53, v28, 16, 1
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v9, v54, v9 :: v_dual_lshlrev_b32 v64, 16, v6
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v30, v30
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v30, 16, v5
; GFX12-FAKE16-NEXT:    v_add3_u32 v53, v53, v28, 0x7fff
; GFX12-FAKE16-NEXT:    v_or_b32_e32 v54, 0x400000, v28
; GFX12-FAKE16-NEXT:    v_perm_b32 v9, v36, v9, 0x5040100
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v22, v22, v6, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v8
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v8, v50, v8 :: v_dual_lshlrev_b32 v55, 16, v22
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v30, v30
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v30, 16, v21
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v5, v5, v21, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v28, v28
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v28, v53, v54, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v64, v55
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v53, v22, v6 :: v_dual_lshlrev_b32 v54, 16, v5
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v30, v30
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v30, 16, v53
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v21, v21, v5, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v24
; GFX12-FAKE16-NEXT:    v_perm_b32 v10, v35, v10, 0x5040100
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v8, v8, v24 :: v_dual_lshlrev_b32 v53, 16, v21
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v24, 0xffff0000, v25
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v25, 16, v28
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v28, 0xffff0000, v28
; GFX12-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v24
; GFX12-FAKE16-NEXT:    v_max_num_f32_e32 v24, v30, v30
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v8, v50, v8, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v7
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v7, v25, v7 :: v_dual_lshlrev_b32 v50, 16, v4
; GFX12-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v54, v53
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v30, v21, v5 :: v_dual_lshlrev_b32 v53, 16, v20
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v23
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_4)
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v7, v7, v23 :: v_dual_lshlrev_b32 v30, 16, v30
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v50, v50
; GFX12-FAKE16-NEXT:    v_bfe_u32 v23, v24, 16, 1
; GFX12-FAKE16-NEXT:    v_or_b32_e32 v50, 0x400000, v24
; GFX12-FAKE16-NEXT:    v_max_num_f32_e32 v30, v30, v30
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v4, v4, v20, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v53, v53
; GFX12-FAKE16-NEXT:    v_add3_u32 v23, v23, v24, 0x7fff
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v53, 16, v4
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v20, v20, v4, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v24, v24
; GFX12-FAKE16-NEXT:    v_bfe_u32 v24, v30, 16, 1
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v23, v23, v50 :: v_dual_lshlrev_b32 v50, 16, v20
; GFX12-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v28
; GFX12-FAKE16-NEXT:    v_add3_u32 v24, v24, v30, 0x7fff
; GFX12-FAKE16-NEXT:    v_or_b32_e32 v28, 0x400000, v30
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v7, v25, v7, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v53, v50
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v25, 16, v23
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v53, 16, v3
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v23, 0xffff0000, v23
; GFX12-FAKE16-NEXT:    v_perm_b32 v7, v38, v7, 0x5040100
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v50, v20, v4, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v6
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v6, v25, v6, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v30, v30
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v30, 16, v19
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v24, v24, v28, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v53, v53
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v3, v3, v19 :: v_dual_lshlrev_b32 v28, 16, v50
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v22
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v6, v6, v22 :: v_dual_lshlrev_b32 v53, 16, v3
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v30, v30
; GFX12-FAKE16-NEXT:    v_max_num_f32_e32 v22, v28, v28
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v28, 16, v24
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v19, v19, v3 :: v_dual_and_b32 v24, 0xffff0000, v24
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v5
; GFX12-FAKE16-NEXT:    v_bfe_u32 v30, v22, 16, 1
; GFX12-FAKE16-NEXT:    v_or_b32_e32 v54, 0x400000, v22
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v5, v28, v5 :: v_dual_lshlrev_b32 v50, 16, v19
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v21
; GFX12-FAKE16-NEXT:    v_add3_u32 v30, v30, v22, 0x7fff
; GFX12-FAKE16-NEXT:    v_perm_b32 v8, v37, v8, 0x5040100
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v5, v5, v21, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v53, v50
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v53, 16, v17
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v21, v19, v3, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v22, v22
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v22, v30, v54 :: v_dual_lshlrev_b32 v21, 16, v21
; GFX12-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v23
; GFX12-FAKE16-NEXT:    v_dual_max_num_f32 v21, v21, v21 :: v_dual_lshlrev_b32 v30, 16, v0
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v23, 16, v22
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v6, v25, v6, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v24
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v24, 16, v2
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v25, 16, v18
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v5, v28, v5, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v4
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
; GFX12-FAKE16-NEXT:    v_perm_b32 v5, v49, v5, 0x5040100
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v4, v23, v4, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v24, v24
; GFX12-FAKE16-NEXT:    v_bfe_u32 v24, v21, 16, 1
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v2, v18, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v20
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v50, 16, v2
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v4, v4, v20, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v25, v25
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v25, 16, v1
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v20, 0xffff0000, v22
; GFX12-FAKE16-NEXT:    v_add3_u32 v22, v24, v21, 0x7fff
; GFX12-FAKE16-NEXT:    v_or_b32_e32 v24, 0x400000, v21
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v18, v18, v2, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v25, v25
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v25, 16, v16
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v1, v1, v17 :: v_dual_lshlrev_b32 v28, 16, v18
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v30, v30
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v30, 16, v1
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v0, v16, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v53, v53
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v17, v17, v1, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v50, v28
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v28, v18, v2, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v25, v25
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v25, 16, v17
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v16, v16, v0, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v21, v21
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v21, v22, v24, vcc_lo
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v22, 16, v16
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v24, 16, v0
; GFX12-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v30, v25
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v21
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v21, 0xffff0000, v21
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v25, v17, v1, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v24, v22
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX12-FAKE16-NEXT:    v_dual_max_num_f32 v24, v28, v28 :: v_dual_lshlrev_b32 v25, 16, v25
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v22, v16, v0, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v3
; GFX12-FAKE16-NEXT:    v_bfe_u32 v28, v24, 16, 1
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX12-FAKE16-NEXT:    v_dual_max_num_f32 v25, v25, v25 :: v_dual_lshlrev_b32 v22, 16, v22
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v3, v30, v3, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v19
; GFX12-FAKE16-NEXT:    v_bfe_u32 v50, v25, 16, 1
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v3, v3, v19, vcc_lo
; GFX12-FAKE16-NEXT:    v_max_num_f32_e32 v19, v22, v22
; GFX12-FAKE16-NEXT:    v_add3_u32 v22, v28, v24, 0x7fff
; GFX12-FAKE16-NEXT:    v_or_b32_e32 v28, 0x400000, v24
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v24, v24
; GFX12-FAKE16-NEXT:    v_add3_u32 v24, v50, v25, 0x7fff
; GFX12-FAKE16-NEXT:    v_bfe_u32 v53, v19, 16, 1
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v22, v22, v28, vcc_lo
; GFX12-FAKE16-NEXT:    v_or_b32_e32 v28, 0x400000, v25
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v25, v25
; GFX12-FAKE16-NEXT:    v_add3_u32 v50, v53, v19, 0x7fff
; GFX12-FAKE16-NEXT:    v_or_b32_e32 v53, 0x400000, v19
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v54, 16, v22
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v24, v24, v28, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v19, v19
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v19, v50, v53, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v21
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v21, 16, v24
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_3)
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v25, 16, v19
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v3, v30, v3, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v2
; GFX12-FAKE16-NEXT:    v_perm_b32 v6, v48, v6, 0x5040100
; GFX12-FAKE16-NEXT:    v_perm_b32 v3, v52, v3, 0x5040100
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v54, v2, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v1
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v21, v1, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v18
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v2, v18, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v0
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v18, 0xffff0000, v19
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v25, v0, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v17
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v1, v17, vcc_lo
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v17, 0xffff0000, v24
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v16
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v0, v16, vcc_lo
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_2)
; GFX12-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v17
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v1, v21, v1 :: v_dual_and_b32 v16, 0xffff0000, v22
; GFX12-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v18
; GFX12-FAKE16-NEXT:    v_perm_b32 v1, v14, v1, 0x5040100
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v25, v0, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v16
; GFX12-FAKE16-NEXT:    v_perm_b32 v14, v31, v27, 0x5040100
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_3)
; GFX12-FAKE16-NEXT:    v_perm_b32 v0, v13, v0, 0x5040100
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v54, v2, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v20
; GFX12-FAKE16-NEXT:    v_perm_b32 v13, v32, v26, 0x5040100
; GFX12-FAKE16-NEXT:    v_perm_b32 v2, v15, v2, 0x5040100
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v4, v23, v4, vcc_lo
; GFX12-FAKE16-NEXT:    v_perm_b32 v15, v39, v29, 0x5040100
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
; GFX12-FAKE16-NEXT:    v_perm_b32 v4, v51, v4, 0x5040100
; GFX12-FAKE16-NEXT:    s_setpc_b64 s[30:31]
  %result = call <32 x bfloat> @llvm.maximumnum.v32bf16(<32 x bfloat> %x, <32 x bfloat> %y)
  ret <32 x bfloat> %result
}

define bfloat @v_maximumnum_bf16_no_ieee(bfloat %x, bfloat %y) #0 {
; GFX7-LABEL: v_maximumnum_bf16_no_ieee:
; GFX7:       ; %bb.0:
; GFX7-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT:    v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT:    v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT:    v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT:    v_max_f32_e32 v0, v0, v1
; GFX7-NEXT:    v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT:    s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_maximumnum_bf16_no_ieee:
; GFX8:       ; %bb.0:
; GFX8-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT:    v_lshlrev_b32_e32 v2, 16, v0
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v2, v2
; GFX8-NEXT:    v_lshlrev_b32_e32 v2, 16, v1
; GFX8-NEXT:    v_cndmask_b32_e32 v0, v0, v1, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v2, v2
; GFX8-NEXT:    v_cndmask_b32_e32 v1, v1, v0, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v2, 16, v1
; GFX8-NEXT:    v_lshlrev_b32_e32 v3, 16, v0
; GFX8-NEXT:    v_cmp_gt_f32_e32 vcc, v3, v2
; GFX8-NEXT:    v_cndmask_b32_e32 v2, v1, v0, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; GFX8-NEXT:    v_mul_f32_e32 v2, 1.0, v2
; GFX8-NEXT:    v_bfe_u32 v3, v2, 16, 1
; GFX8-NEXT:    v_add_u32_e32 v3, vcc, v3, v2
; GFX8-NEXT:    s_movk_i32 s4, 0x7fff
; GFX8-NEXT:    v_add_u32_e32 v3, vcc, s4, v3
; GFX8-NEXT:    v_or_b32_e32 v4, 0x400000, v2
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v2, v2
; GFX8-NEXT:    v_cndmask_b32_e32 v2, v3, v4, vcc
; GFX8-NEXT:    v_lshrrev_b32_e32 v3, 16, v2
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v0
; GFX8-NEXT:    v_cndmask_b32_e32 v0, v3, v0, vcc
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v1
; GFX8-NEXT:    v_cndmask_b32_e32 v0, v0, v1, vcc
; GFX8-NEXT:    v_and_b32_e32 v1, 0xffff0000, v2
; GFX8-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v1
; GFX8-NEXT:    v_cndmask_b32_e32 v0, v3, v0, vcc
; GFX8-NEXT:    s_setpc_b64 s[30:31]
;
; GFX900-LABEL: v_maximumnum_bf16_no_ieee:
; GFX900:       ; %bb.0:
; GFX900-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX900-NEXT:    v_lshlrev_b32_e32 v2, 16, v0
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v2, v2
; GFX900-NEXT:    v_lshlrev_b32_e32 v3, 16, v1
; GFX900-NEXT:    v_cndmask_b32_e32 v0, v0, v1, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; GFX900-NEXT:    v_cndmask_b32_e32 v1, v1, v0, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v2, 16, v0
; GFX900-NEXT:    v_lshlrev_b32_e32 v3, 16, v1
; GFX900-NEXT:    v_cmp_gt_f32_e32 vcc, v2, v3
; GFX900-NEXT:    v_cndmask_b32_e32 v2, v1, v0, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; GFX900-NEXT:    v_max_f32_e32 v2, v2, v2
; GFX900-NEXT:    v_bfe_u32 v3, v2, 16, 1
; GFX900-NEXT:    s_movk_i32 s4, 0x7fff
; GFX900-NEXT:    v_add3_u32 v3, v3, v2, s4
; GFX900-NEXT:    v_or_b32_e32 v4, 0x400000, v2
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v2, v2
; GFX900-NEXT:    v_cndmask_b32_e32 v2, v3, v4, vcc
; GFX900-NEXT:    v_lshrrev_b32_e32 v3, 16, v2
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v0
; GFX900-NEXT:    v_cndmask_b32_e32 v0, v3, v0, vcc
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v1
; GFX900-NEXT:    v_cndmask_b32_e32 v0, v0, v1, vcc
; GFX900-NEXT:    v_and_b32_e32 v1, 0xffff0000, v2
; GFX900-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v1
; GFX900-NEXT:    v_cndmask_b32_e32 v0, v3, v0, vcc
; GFX900-NEXT:    s_setpc_b64 s[30:31]
;
; GFX950-LABEL: v_maximumnum_bf16_no_ieee:
; GFX950:       ; %bb.0:
; GFX950-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX950-NEXT:    v_lshlrev_b32_e32 v2, 16, v0
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v2, v2
; GFX950-NEXT:    v_lshlrev_b32_e32 v3, 16, v1
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v0, v0, v1, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; GFX950-NEXT:    v_lshlrev_b32_e32 v2, 16, v0
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v1, v1, v0, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v3, 16, v1
; GFX950-NEXT:    v_cmp_gt_f32_e32 vcc, v2, v3
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v2, v1, v0, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; GFX950-NEXT:    v_max_f32_e32 v2, v2, v2
; GFX950-NEXT:    v_cvt_pk_bf16_f32 v2, v2, s0
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v0
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v0, v2, v0, vcc
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v1
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v0, v0, v1, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v1, 16, v2
; GFX950-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v1
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v0, v2, v0, vcc
; GFX950-NEXT:    s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_maximumnum_bf16_no_ieee:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT:    v_lshlrev_b32_e32 v2, 16, v0
; GFX10-NEXT:    v_lshlrev_b32_e32 v3, 16, v1
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX10-NEXT:    v_cndmask_b32_e32 v0, v0, v1, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX10-NEXT:    v_lshlrev_b32_e32 v2, 16, v0
; GFX10-NEXT:    v_cndmask_b32_e32 v1, v1, v0, vcc_lo
; GFX10-NEXT:    v_lshlrev_b32_e32 v3, 16, v1
; GFX10-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v2, v3
; GFX10-NEXT:    v_cndmask_b32_e32 v2, v1, v0, vcc_lo
; GFX10-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; GFX10-NEXT:    v_max_f32_e32 v2, v2, v2
; GFX10-NEXT:    v_bfe_u32 v3, v2, 16, 1
; GFX10-NEXT:    v_or_b32_e32 v4, 0x400000, v2
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX10-NEXT:    v_add3_u32 v3, v3, v2, 0x7fff
; GFX10-NEXT:    v_cndmask_b32_e32 v2, v3, v4, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v0
; GFX10-NEXT:    v_lshrrev_b32_e32 v3, 16, v2
; GFX10-NEXT:    v_and_b32_e32 v2, 0xffff0000, v2
; GFX10-NEXT:    v_cndmask_b32_e32 v0, v3, v0, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v1
; GFX10-NEXT:    v_cndmask_b32_e32 v0, v0, v1, vcc_lo
; GFX10-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v2
; GFX10-NEXT:    v_cndmask_b32_e32 v0, v3, v0, vcc_lo
; GFX10-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-TRUE16-LABEL: v_maximumnum_bf16_no_ieee:
; GFX11-TRUE16:       ; %bb.0:
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v0
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.h, v1.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s0, v1, v1
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v0.l, v0.l, v0.h, vcc_lo
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v1.l, v0.h, v0.l, s0
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, v0.l
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v3.l, v1.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; GFX11-TRUE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v2, v3
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v2.l, v1.l, v0.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v0.l
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; GFX11-TRUE16-NEXT:    v_max_f32_e32 v2, v2, v2
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v3, v2, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, 0x400000, v2
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s0, v2, v2
; GFX11-TRUE16-NEXT:    v_add3_u32 v3, v3, v2, 0x7fff
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e64 v2, v3, v4, s0
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s0, 0, v1.l
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v0.l, v2.h, v0.l, vcc_lo
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff0000, v2
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v0.l, v0.l, v1.l, s0
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v3
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v0.l, v2.h, v0.l, vcc_lo
; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-FAKE16-LABEL: v_maximumnum_bf16_no_ieee:
; GFX11-FAKE16:       ; %bb.0:
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v0
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v0, v1, vcc_lo
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v1
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v0
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v1, v0, vcc_lo
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v1
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v2, v3
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v1, v0, vcc_lo
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_max_f32_e32 v2, v2, v2
; GFX11-FAKE16-NEXT:    v_bfe_u32 v3, v2, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, 0x400000, v2
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_add3_u32 v3, v3, v2, 0x7fff
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v3, v4, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v0
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v3, 16, v2
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xffff0000, v2
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v3, v0, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v1
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v0, v1, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v2
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v3, v0, vcc_lo
; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
;
; GFX12-TRUE16-LABEL: v_maximumnum_bf16_no_ieee:
; GFX12-TRUE16:       ; %bb.0:
; GFX12-TRUE16-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX12-TRUE16-NEXT:    s_wait_expcnt 0x0
; GFX12-TRUE16-NEXT:    s_wait_samplecnt 0x0
; GFX12-TRUE16-NEXT:    s_wait_bvhcnt 0x0
; GFX12-TRUE16-NEXT:    s_wait_kmcnt 0x0
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v0
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v0.h, v1.l
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s0, v1, v1
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v0.l, v0.l, v0.h, vcc_lo
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v1.l, v0.h, v0.l, s0
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v2.l, v0.l
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v3.l, v1.l
; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_2)
; GFX12-TRUE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v2, v3
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v2.l, v1.l, v0.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v0.l
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT:    v_max_num_f32_e32 v2, v2, v2
; GFX12-TRUE16-NEXT:    v_bfe_u32 v3, v2, 16, 1
; GFX12-TRUE16-NEXT:    v_or_b32_e32 v4, 0x400000, v2
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s0, v2, v2
; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT:    v_add3_u32 v3, v3, v2, 0x7fff
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b32_e64 v2, v3, v4, s0
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s0, 0, v1.l
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v0.l, v2.h, v0.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff0000, v2
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v0.l, v0.l, v1.l, s0
; GFX12-TRUE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v3
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v0.l, v2.h, v0.l, vcc_lo
; GFX12-TRUE16-NEXT:    s_setpc_b64 s[30:31]
;
; GFX12-FAKE16-LABEL: v_maximumnum_bf16_no_ieee:
; GFX12-FAKE16:       ; %bb.0:
; GFX12-FAKE16-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX12-FAKE16-NEXT:    s_wait_expcnt 0x0
; GFX12-FAKE16-NEXT:    s_wait_samplecnt 0x0
; GFX12-FAKE16-NEXT:    s_wait_bvhcnt 0x0
; GFX12-FAKE16-NEXT:    s_wait_kmcnt 0x0
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v0
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_2)
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v0, v1, vcc_lo
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v1
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v0
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1)
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v1, v0, vcc_lo
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v1
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1)
; GFX12-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v2, v3
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v1, v0, vcc_lo
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-FAKE16-NEXT:    v_max_num_f32_e32 v2, v2, v2
; GFX12-FAKE16-NEXT:    v_bfe_u32 v3, v2, 16, 1
; GFX12-FAKE16-NEXT:    v_or_b32_e32 v4, 0x400000, v2
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX12-FAKE16-NEXT:    v_add3_u32 v3, v3, v2, 0x7fff
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v3, v4, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v0
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_2)
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v3, 16, v2
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v2, 0xffff0000, v2
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v3, v0, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v1
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_2)
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v0, v1, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v2
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v3, v0, vcc_lo
; GFX12-FAKE16-NEXT:    s_setpc_b64 s[30:31]
  %result = call bfloat @llvm.maximumnum.bf16(bfloat %x, bfloat %y)
  ret bfloat %result
}

define <2 x bfloat> @v_maximumnum_v2bf16_no_ieee(<2 x bfloat> %x, <2 x bfloat> %y) #0 {
; GFX7-LABEL: v_maximumnum_v2bf16_no_ieee:
; GFX7:       ; %bb.0:
; GFX7-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT:    v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT:    v_mul_f32_e32 v2, 1.0, v2
; GFX7-NEXT:    v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT:    v_mul_f32_e32 v3, 1.0, v3
; GFX7-NEXT:    v_and_b32_e32 v3, 0xffff0000, v3
; GFX7-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT:    v_and_b32_e32 v2, 0xffff0000, v2
; GFX7-NEXT:    v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT:    v_max_f32_e32 v1, v1, v3
; GFX7-NEXT:    v_max_f32_e32 v0, v0, v2
; GFX7-NEXT:    v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT:    s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_maximumnum_v2bf16_no_ieee:
; GFX8:       ; %bb.0:
; GFX8-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT:    v_and_b32_e32 v2, 0xffff0000, v0
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v2, v2
; GFX8-NEXT:    v_lshrrev_b32_e32 v2, 16, v1
; GFX8-NEXT:    v_and_b32_e32 v5, 0xffff0000, v1
; GFX8-NEXT:    v_cndmask_b32_sdwa v3, v0, v2, vcc dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; GFX8-NEXT:    v_cndmask_b32_e32 v2, v2, v3, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v4, 16, v3
; GFX8-NEXT:    v_lshlrev_b32_e32 v5, 16, v2
; GFX8-NEXT:    v_cmp_gt_f32_e32 vcc, v4, v5
; GFX8-NEXT:    v_cndmask_b32_e32 v4, v2, v3, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; GFX8-NEXT:    v_mul_f32_e32 v4, 1.0, v4
; GFX8-NEXT:    v_bfe_u32 v5, v4, 16, 1
; GFX8-NEXT:    v_add_u32_e32 v5, vcc, v5, v4
; GFX8-NEXT:    s_movk_i32 s4, 0x7fff
; GFX8-NEXT:    v_add_u32_e32 v5, vcc, s4, v5
; GFX8-NEXT:    v_or_b32_e32 v6, 0x400000, v4
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v4, v4
; GFX8-NEXT:    v_cndmask_b32_e32 v4, v5, v6, vcc
; GFX8-NEXT:    v_lshrrev_b32_e32 v5, 16, v4
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v3
; GFX8-NEXT:    v_cndmask_b32_e32 v3, v5, v3, vcc
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v2
; GFX8-NEXT:    v_cndmask_b32_e32 v2, v3, v2, vcc
; GFX8-NEXT:    v_and_b32_e32 v3, 0xffff0000, v4
; GFX8-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v3
; GFX8-NEXT:    v_lshlrev_b32_e32 v3, 16, v0
; GFX8-NEXT:    v_cndmask_b32_e32 v2, v5, v2, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; GFX8-NEXT:    v_lshlrev_b32_e32 v3, 16, v1
; GFX8-NEXT:    v_cndmask_b32_e32 v0, v0, v1, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; GFX8-NEXT:    v_cndmask_b32_e32 v1, v1, v0, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v3, 16, v1
; GFX8-NEXT:    v_lshlrev_b32_e32 v4, 16, v0
; GFX8-NEXT:    v_cmp_gt_f32_e32 vcc, v4, v3
; GFX8-NEXT:    v_cndmask_b32_e32 v3, v1, v0, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; GFX8-NEXT:    v_mul_f32_e32 v3, 1.0, v3
; GFX8-NEXT:    v_bfe_u32 v4, v3, 16, 1
; GFX8-NEXT:    v_add_u32_e32 v4, vcc, v4, v3
; GFX8-NEXT:    v_add_u32_e32 v4, vcc, s4, v4
; GFX8-NEXT:    v_or_b32_e32 v5, 0x400000, v3
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; GFX8-NEXT:    v_cndmask_b32_e32 v3, v4, v5, vcc
; GFX8-NEXT:    v_lshrrev_b32_e32 v4, 16, v3
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v0
; GFX8-NEXT:    v_cndmask_b32_e32 v0, v4, v0, vcc
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v1
; GFX8-NEXT:    v_cndmask_b32_e32 v0, v0, v1, vcc
; GFX8-NEXT:    v_and_b32_e32 v1, 0xffff0000, v3
; GFX8-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v1
; GFX8-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; GFX8-NEXT:    v_cndmask_b32_e32 v0, v4, v0, vcc
; GFX8-NEXT:    v_or_b32_sdwa v0, v0, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX8-NEXT:    s_setpc_b64 s[30:31]
;
; GFX900-LABEL: v_maximumnum_v2bf16_no_ieee:
; GFX900:       ; %bb.0:
; GFX900-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX900-NEXT:    v_and_b32_e32 v2, 0xffff0000, v0
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v2, v2
; GFX900-NEXT:    v_lshrrev_b32_e32 v2, 16, v1
; GFX900-NEXT:    v_and_b32_e32 v5, 0xffff0000, v1
; GFX900-NEXT:    v_cndmask_b32_sdwa v3, v0, v2, vcc dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; GFX900-NEXT:    v_cndmask_b32_e32 v2, v2, v3, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v4, 16, v3
; GFX900-NEXT:    v_lshlrev_b32_e32 v5, 16, v2
; GFX900-NEXT:    v_cmp_gt_f32_e32 vcc, v4, v5
; GFX900-NEXT:    v_cndmask_b32_e32 v4, v2, v3, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; GFX900-NEXT:    v_max_f32_e32 v4, v4, v4
; GFX900-NEXT:    v_bfe_u32 v5, v4, 16, 1
; GFX900-NEXT:    s_movk_i32 s4, 0x7fff
; GFX900-NEXT:    v_add3_u32 v5, v5, v4, s4
; GFX900-NEXT:    v_or_b32_e32 v6, 0x400000, v4
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v4, v4
; GFX900-NEXT:    v_cndmask_b32_e32 v4, v5, v6, vcc
; GFX900-NEXT:    v_lshrrev_b32_e32 v5, 16, v4
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v3
; GFX900-NEXT:    v_cndmask_b32_e32 v3, v5, v3, vcc
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v2
; GFX900-NEXT:    v_cndmask_b32_e32 v2, v3, v2, vcc
; GFX900-NEXT:    v_and_b32_e32 v3, 0xffff0000, v4
; GFX900-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v3
; GFX900-NEXT:    v_lshlrev_b32_e32 v3, 16, v0
; GFX900-NEXT:    v_cndmask_b32_e32 v2, v5, v2, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; GFX900-NEXT:    v_lshlrev_b32_e32 v3, 16, v1
; GFX900-NEXT:    v_cndmask_b32_e32 v0, v0, v1, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; GFX900-NEXT:    v_cndmask_b32_e32 v1, v1, v0, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v3, 16, v1
; GFX900-NEXT:    v_lshlrev_b32_e32 v4, 16, v0
; GFX900-NEXT:    v_cmp_gt_f32_e32 vcc, v4, v3
; GFX900-NEXT:    v_cndmask_b32_e32 v3, v1, v0, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; GFX900-NEXT:    v_max_f32_e32 v3, v3, v3
; GFX900-NEXT:    v_bfe_u32 v4, v3, 16, 1
; GFX900-NEXT:    v_add3_u32 v4, v4, v3, s4
; GFX900-NEXT:    v_or_b32_e32 v5, 0x400000, v3
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; GFX900-NEXT:    v_cndmask_b32_e32 v3, v4, v5, vcc
; GFX900-NEXT:    v_lshrrev_b32_e32 v4, 16, v3
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v0
; GFX900-NEXT:    v_cndmask_b32_e32 v0, v4, v0, vcc
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v1
; GFX900-NEXT:    v_cndmask_b32_e32 v0, v0, v1, vcc
; GFX900-NEXT:    v_and_b32_e32 v1, 0xffff0000, v3
; GFX900-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v1
; GFX900-NEXT:    v_cndmask_b32_e32 v0, v4, v0, vcc
; GFX900-NEXT:    s_mov_b32 s4, 0x5040100
; GFX900-NEXT:    v_perm_b32 v0, v2, v0, s4
; GFX900-NEXT:    s_setpc_b64 s[30:31]
;
; GFX950-LABEL: v_maximumnum_v2bf16_no_ieee:
; GFX950:       ; %bb.0:
; GFX950-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX950-NEXT:    v_and_b32_e32 v2, 0xffff0000, v0
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v2, v2
; GFX950-NEXT:    v_lshrrev_b32_e32 v2, 16, v1
; GFX950-NEXT:    v_and_b32_e32 v5, 0xffff0000, v1
; GFX950-NEXT:    v_cndmask_b32_sdwa v3, v0, v2, vcc dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; GFX950-NEXT:    v_lshlrev_b32_e32 v4, 16, v3
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v2, v2, v3, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v5, 16, v2
; GFX950-NEXT:    v_cmp_gt_f32_e32 vcc, v4, v5
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v4, v2, v3, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; GFX950-NEXT:    v_max_f32_e32 v4, v4, v4
; GFX950-NEXT:    v_cvt_pk_bf16_f32 v4, v4, s0
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v3
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v3, v4, v3, vcc
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v2
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v2, v3, v2, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v3, 16, v4
; GFX950-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v3
; GFX950-NEXT:    v_lshlrev_b32_e32 v3, 16, v0
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v2, v4, v2, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; GFX950-NEXT:    v_lshlrev_b32_e32 v3, 16, v1
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v0, v0, v1, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; GFX950-NEXT:    v_lshlrev_b32_e32 v4, 16, v0
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v1, v1, v0, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v3, 16, v1
; GFX950-NEXT:    v_cmp_gt_f32_e32 vcc, v4, v3
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v3, v1, v0, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; GFX950-NEXT:    v_max_f32_e32 v3, v3, v3
; GFX950-NEXT:    v_cvt_pk_bf16_f32 v3, v3, s0
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v0
; GFX950-NEXT:    s_mov_b32 s0, 0x5040100
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v0, v3, v0, vcc
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v1
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v0, v0, v1, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v1, 16, v3
; GFX950-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v1
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v0, v3, v0, vcc
; GFX950-NEXT:    v_perm_b32 v0, v2, v0, s0
; GFX950-NEXT:    s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_maximumnum_v2bf16_no_ieee:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT:    v_and_b32_e32 v2, 0xffff0000, v0
; GFX10-NEXT:    v_lshlrev_b32_e32 v4, 16, v0
; GFX10-NEXT:    v_and_b32_e32 v3, 0xffff0000, v1
; GFX10-NEXT:    v_lshlrev_b32_e32 v5, 16, v1
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX10-NEXT:    v_lshrrev_b32_e32 v2, 16, v1
; GFX10-NEXT:    v_cndmask_b32_sdwa v6, v0, v2, vcc_lo dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX10-NEXT:    v_cndmask_b32_e32 v0, v0, v1, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX10-NEXT:    v_lshlrev_b32_e32 v3, 16, v6
; GFX10-NEXT:    v_cndmask_b32_e32 v2, v2, v6, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX10-NEXT:    v_lshlrev_b32_e32 v5, 16, v0
; GFX10-NEXT:    v_lshlrev_b32_e32 v4, 16, v2
; GFX10-NEXT:    v_cndmask_b32_e32 v1, v1, v0, vcc_lo
; GFX10-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v3, v4
; GFX10-NEXT:    v_lshlrev_b32_e32 v7, 16, v1
; GFX10-NEXT:    v_cndmask_b32_e32 v3, v2, v6, vcc_lo
; GFX10-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v5, v7
; GFX10-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; GFX10-NEXT:    v_cndmask_b32_e32 v4, v1, v0, vcc_lo
; GFX10-NEXT:    v_max_f32_e32 v3, v3, v3
; GFX10-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; GFX10-NEXT:    v_bfe_u32 v5, v3, 16, 1
; GFX10-NEXT:    v_max_f32_e32 v4, v4, v4
; GFX10-NEXT:    v_or_b32_e32 v8, 0x400000, v3
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX10-NEXT:    v_add3_u32 v5, v5, v3, 0x7fff
; GFX10-NEXT:    v_bfe_u32 v7, v4, 16, 1
; GFX10-NEXT:    v_or_b32_e32 v9, 0x400000, v4
; GFX10-NEXT:    v_cndmask_b32_e32 v3, v5, v8, vcc_lo
; GFX10-NEXT:    v_add3_u32 v7, v7, v4, 0x7fff
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX10-NEXT:    v_lshrrev_b32_e32 v5, 16, v3
; GFX10-NEXT:    v_and_b32_e32 v3, 0xffff0000, v3
; GFX10-NEXT:    v_cndmask_b32_e32 v4, v7, v9, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v6
; GFX10-NEXT:    v_lshrrev_b32_e32 v7, 16, v4
; GFX10-NEXT:    v_cndmask_b32_e32 v6, v5, v6, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v0
; GFX10-NEXT:    v_and_b32_e32 v4, 0xffff0000, v4
; GFX10-NEXT:    v_cndmask_b32_e32 v0, v7, v0, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v2
; GFX10-NEXT:    v_cndmask_b32_e32 v2, v6, v2, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v1
; GFX10-NEXT:    v_cndmask_b32_e32 v0, v0, v1, vcc_lo
; GFX10-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v3
; GFX10-NEXT:    v_cndmask_b32_e32 v1, v5, v2, vcc_lo
; GFX10-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v4
; GFX10-NEXT:    v_cndmask_b32_e32 v0, v7, v0, vcc_lo
; GFX10-NEXT:    v_perm_b32 v0, v1, v0, 0x5040100
; GFX10-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-TRUE16-LABEL: v_maximumnum_v2bf16_no_ieee:
; GFX11-TRUE16:       ; %bb.0:
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff0000, v0
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v0
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xffff0000, v1
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 16, v1
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s0, v3, v3
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s1, v4, v4
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v2.l, v0.h, v1.h, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v0.l, v0.l, v1.l, s0
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v3.l, v1.h, v2.l, s1
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, v2.l
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v1.l, v1.l, v0.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, v0.l
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v5.l, v3.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.l, v1.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s2, 0, v1.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; GFX11-TRUE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v4, v5
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_cmp_gt_f32_e64 s0, v6, v7
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v4.l, v3.l, v2.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v2.l
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v5.l, v1.l, v0.l, s0
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s0, 0, v0.l
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; GFX11-TRUE16-NEXT:    v_dual_max_f32 v4, v4, v4 :: v_dual_lshlrev_b32 v5, 16, v5
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT:    v_max_f32_e32 v5, v5, v5
; GFX11-TRUE16-NEXT:    v_bfe_u32 v6, v4, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, 0x400000, v4
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v7, v5, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s1, v4, v4
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, 0x400000, v5
; GFX11-TRUE16-NEXT:    v_add3_u32 v6, v6, v4, 0x7fff
; GFX11-TRUE16-NEXT:    v_add3_u32 v7, v7, v5, 0x7fff
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e64 v4, v6, v8, s1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s1, v5, v5
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v0.h, v4.h, v2.l, vcc_lo
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e64 v5, v7, v9, s1
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s1, 0, v3.l
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff0000, v4
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v0.l, v5.h, v0.l, s0
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xffff0000, v5
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v0.h, v0.h, v3.l, s1
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v2
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v0.l, v0.l, v1.l, s2
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_cmp_eq_f32_e64 s0, 0, v6
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v0.h, v4.h, v0.h, vcc_lo
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v0.l, v5.h, v0.l, s0
; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-FAKE16-LABEL: v_maximumnum_v2bf16_no_ieee:
; GFX11-FAKE16:       ; %bb.0:
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xffff0000, v0
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v3, 16, v1
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v4, 16, v0
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v2, v4, v3 :: v_dual_lshlrev_b32 v5, 16, v0
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v6, 0xffff0000, v1
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v4, 16, v2
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v0, v1, vcc_lo
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v3, v3, v2, vcc_lo
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v7, 16, v1
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v5, 16, v3
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v6, 16, v0
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v1, v0, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v4, v5
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v4, v3, v2, vcc_lo
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; GFX11-FAKE16-NEXT:    v_dual_max_f32 v4, v4, v4 :: v_dual_lshlrev_b32 v7, 16, v1
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v6, v7
; GFX11-FAKE16-NEXT:    v_bfe_u32 v6, v4, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v5, v1, v0, vcc_lo
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, 0x400000, v4
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_add3_u32 v6, v6, v4, 0x7fff
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v4, v6, v8 :: v_dual_lshlrev_b32 v5, 16, v5
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_max_f32_e32 v5, v5, v5
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v6, 16, v4
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v7, v5, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, 0x400000, v5
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xffff0000, v4
; GFX11-FAKE16-NEXT:    v_add3_u32 v7, v7, v5, 0x7fff
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v5, v7, v9, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v2
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v7, 16, v5
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v6, v2, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v0
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v0, v7, v0 :: v_dual_and_b32 v5, 0xffff0000, v5
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v3
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v2, v3, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v1
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v0, v1, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v4
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v6, v2, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v5
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v7, v0, vcc_lo
; GFX11-FAKE16-NEXT:    v_perm_b32 v0, v1, v0, 0x5040100
; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
;
; GFX12-TRUE16-LABEL: v_maximumnum_v2bf16_no_ieee:
; GFX12-TRUE16:       ; %bb.0:
; GFX12-TRUE16-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX12-TRUE16-NEXT:    s_wait_expcnt 0x0
; GFX12-TRUE16-NEXT:    s_wait_samplecnt 0x0
; GFX12-TRUE16-NEXT:    s_wait_bvhcnt 0x0
; GFX12-TRUE16-NEXT:    s_wait_kmcnt 0x0
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff0000, v0
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v0
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v4, 0xffff0000, v1
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 16, v1
; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s0, v3, v3
; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s1, v4, v4
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v2.l, v0.h, v1.h, vcc_lo
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v0.l, v0.l, v1.l, s0
; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v3.l, v1.h, v2.l, s1
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v4.l, v2.l
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v1.l, v1.l, v0.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v6.l, v0.l
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v5.l, v3.l
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v7.l, v1.l
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s2, 0, v1.l
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; GFX12-TRUE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v4, v5
; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
; GFX12-TRUE16-NEXT:    v_cmp_gt_f32_e64 s0, v6, v7
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v4.l, v3.l, v2.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v2.l
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v5.l, v1.l, v0.l, s0
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s0, 0, v0.l
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT:    v_dual_max_num_f32 v4, v4, v4 :: v_dual_lshlrev_b32 v5, 16, v5
; GFX12-TRUE16-NEXT:    v_max_num_f32_e32 v5, v5, v5
; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX12-TRUE16-NEXT:    v_bfe_u32 v6, v4, 16, 1
; GFX12-TRUE16-NEXT:    v_or_b32_e32 v8, 0x400000, v4
; GFX12-TRUE16-NEXT:    v_bfe_u32 v7, v5, 16, 1
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s1, v4, v4
; GFX12-TRUE16-NEXT:    v_or_b32_e32 v9, 0x400000, v5
; GFX12-TRUE16-NEXT:    v_add3_u32 v6, v6, v4, 0x7fff
; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX12-TRUE16-NEXT:    v_add3_u32 v7, v7, v5, 0x7fff
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b32_e64 v4, v6, v8, s1
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s1, v5, v5
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v0.h, v4.h, v2.l, vcc_lo
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b32_e64 v5, v7, v9, s1
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s1, 0, v3.l
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff0000, v4
; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v0.l, v5.h, v0.l, s0
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v6, 0xffff0000, v5
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v0.h, v0.h, v3.l, s1
; GFX12-TRUE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v2
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v0.l, v0.l, v1.l, s2
; GFX12-TRUE16-NEXT:    v_cmp_eq_f32_e64 s0, 0, v6
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v0.h, v4.h, v0.h, vcc_lo
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v0.l, v5.h, v0.l, s0
; GFX12-TRUE16-NEXT:    s_setpc_b64 s[30:31]
;
; GFX12-FAKE16-LABEL: v_maximumnum_v2bf16_no_ieee:
; GFX12-FAKE16:       ; %bb.0:
; GFX12-FAKE16-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX12-FAKE16-NEXT:    s_wait_expcnt 0x0
; GFX12-FAKE16-NEXT:    s_wait_samplecnt 0x0
; GFX12-FAKE16-NEXT:    s_wait_bvhcnt 0x0
; GFX12-FAKE16-NEXT:    s_wait_kmcnt 0x0
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v2, 0xffff0000, v0
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v3, 16, v1
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v4, 16, v0
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v2, v4, v3 :: v_dual_lshlrev_b32 v5, 16, v0
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v6, 0xffff0000, v1
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v4, 16, v2
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v0, v1, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v3, v3, v2, vcc_lo
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v7, 16, v1
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v5, 16, v3
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v6, 16, v0
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v1, v0, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v4, v5
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v4, v3, v2, vcc_lo
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; GFX12-FAKE16-NEXT:    v_dual_max_num_f32 v4, v4, v4 :: v_dual_lshlrev_b32 v7, 16, v1
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX12-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v6, v7
; GFX12-FAKE16-NEXT:    v_bfe_u32 v6, v4, 16, 1
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v5, v1, v0, vcc_lo
; GFX12-FAKE16-NEXT:    v_or_b32_e32 v8, 0x400000, v4
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX12-FAKE16-NEXT:    v_add3_u32 v6, v6, v4, 0x7fff
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v4, v6, v8 :: v_dual_lshlrev_b32 v5, 16, v5
; GFX12-FAKE16-NEXT:    v_max_num_f32_e32 v5, v5, v5
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v6, 16, v4
; GFX12-FAKE16-NEXT:    v_bfe_u32 v7, v5, 16, 1
; GFX12-FAKE16-NEXT:    v_or_b32_e32 v9, 0x400000, v5
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v4, 0xffff0000, v4
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX12-FAKE16-NEXT:    v_add3_u32 v7, v7, v5, 0x7fff
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v5, v7, v9, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v2
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v7, 16, v5
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v6, v2, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v0
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v0, v7, v0 :: v_dual_and_b32 v5, 0xffff0000, v5
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v3
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v2, v3, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v1
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v0, v1, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v4
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v6, v2, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v5
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v7, v0, vcc_lo
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX12-FAKE16-NEXT:    v_perm_b32 v0, v1, v0, 0x5040100
; GFX12-FAKE16-NEXT:    s_setpc_b64 s[30:31]
  %result = call <2 x bfloat> @llvm.maximumnum.v2bf16(<2 x bfloat> %x, <2 x bfloat> %y)
  ret <2 x bfloat> %result
}

define <3 x bfloat> @v_maximumnum_v3bf16_no_ieee(<3 x bfloat> %x, <3 x bfloat> %y) #0 {
; GFX7-LABEL: v_maximumnum_v3bf16_no_ieee:
; GFX7:       ; %bb.0:
; GFX7-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT:    v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT:    v_mul_f32_e32 v3, 1.0, v3
; GFX7-NEXT:    v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT:    v_mul_f32_e32 v4, 1.0, v4
; GFX7-NEXT:    v_mul_f32_e32 v2, 1.0, v2
; GFX7-NEXT:    v_mul_f32_e32 v5, 1.0, v5
; GFX7-NEXT:    v_and_b32_e32 v5, 0xffff0000, v5
; GFX7-NEXT:    v_and_b32_e32 v2, 0xffff0000, v2
; GFX7-NEXT:    v_and_b32_e32 v4, 0xffff0000, v4
; GFX7-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT:    v_and_b32_e32 v3, 0xffff0000, v3
; GFX7-NEXT:    v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT:    v_max_f32_e32 v2, v2, v5
; GFX7-NEXT:    v_max_f32_e32 v1, v1, v4
; GFX7-NEXT:    v_max_f32_e32 v0, v0, v3
; GFX7-NEXT:    v_and_b32_e32 v2, 0xffff0000, v2
; GFX7-NEXT:    v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT:    s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_maximumnum_v3bf16_no_ieee:
; GFX8:       ; %bb.0:
; GFX8-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT:    v_and_b32_e32 v4, 0xffff0000, v0
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v4, v4
; GFX8-NEXT:    v_lshrrev_b32_e32 v4, 16, v2
; GFX8-NEXT:    v_and_b32_e32 v7, 0xffff0000, v2
; GFX8-NEXT:    v_cndmask_b32_sdwa v5, v0, v4, vcc dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v7, v7
; GFX8-NEXT:    v_cndmask_b32_e32 v4, v4, v5, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v6, 16, v5
; GFX8-NEXT:    v_lshlrev_b32_e32 v7, 16, v4
; GFX8-NEXT:    v_cmp_gt_f32_e32 vcc, v6, v7
; GFX8-NEXT:    v_cndmask_b32_e32 v6, v4, v5, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; GFX8-NEXT:    v_mul_f32_e32 v6, 1.0, v6
; GFX8-NEXT:    v_bfe_u32 v7, v6, 16, 1
; GFX8-NEXT:    v_add_u32_e32 v7, vcc, v7, v6
; GFX8-NEXT:    s_movk_i32 s4, 0x7fff
; GFX8-NEXT:    v_add_u32_e32 v7, vcc, s4, v7
; GFX8-NEXT:    v_or_b32_e32 v8, 0x400000, v6
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v6, v6
; GFX8-NEXT:    v_cndmask_b32_e32 v6, v7, v8, vcc
; GFX8-NEXT:    v_lshrrev_b32_e32 v7, 16, v6
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v5
; GFX8-NEXT:    v_cndmask_b32_e32 v5, v7, v5, vcc
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v4
; GFX8-NEXT:    v_cndmask_b32_e32 v4, v5, v4, vcc
; GFX8-NEXT:    v_and_b32_e32 v5, 0xffff0000, v6
; GFX8-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v5
; GFX8-NEXT:    v_lshlrev_b32_e32 v5, 16, v1
; GFX8-NEXT:    v_cndmask_b32_e32 v4, v7, v4, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; GFX8-NEXT:    v_lshlrev_b32_e32 v5, 16, v3
; GFX8-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; GFX8-NEXT:    v_cndmask_b32_e32 v3, v3, v1, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v5, 16, v3
; GFX8-NEXT:    v_lshlrev_b32_e32 v6, 16, v1
; GFX8-NEXT:    v_cmp_gt_f32_e32 vcc, v6, v5
; GFX8-NEXT:    v_cndmask_b32_e32 v5, v3, v1, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; GFX8-NEXT:    v_mul_f32_e32 v5, 1.0, v5
; GFX8-NEXT:    v_bfe_u32 v6, v5, 16, 1
; GFX8-NEXT:    v_add_u32_e32 v6, vcc, v6, v5
; GFX8-NEXT:    v_add_u32_e32 v6, vcc, s4, v6
; GFX8-NEXT:    v_or_b32_e32 v7, 0x400000, v5
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; GFX8-NEXT:    v_cndmask_b32_e32 v5, v6, v7, vcc
; GFX8-NEXT:    v_lshrrev_b32_e32 v6, 16, v5
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v1
; GFX8-NEXT:    v_cndmask_b32_e32 v1, v6, v1, vcc
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v3
; GFX8-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc
; GFX8-NEXT:    v_and_b32_e32 v3, 0xffff0000, v5
; GFX8-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v3
; GFX8-NEXT:    v_lshlrev_b32_e32 v3, 16, v0
; GFX8-NEXT:    v_cndmask_b32_e32 v1, v6, v1, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; GFX8-NEXT:    v_lshlrev_b32_e32 v3, 16, v2
; GFX8-NEXT:    v_cndmask_b32_e32 v0, v0, v2, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; GFX8-NEXT:    v_cndmask_b32_e32 v2, v2, v0, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v3, 16, v2
; GFX8-NEXT:    v_lshlrev_b32_e32 v5, 16, v0
; GFX8-NEXT:    v_cmp_gt_f32_e32 vcc, v5, v3
; GFX8-NEXT:    v_cndmask_b32_e32 v3, v2, v0, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; GFX8-NEXT:    v_mul_f32_e32 v3, 1.0, v3
; GFX8-NEXT:    v_bfe_u32 v5, v3, 16, 1
; GFX8-NEXT:    v_add_u32_e32 v5, vcc, v5, v3
; GFX8-NEXT:    v_add_u32_e32 v5, vcc, s4, v5
; GFX8-NEXT:    v_or_b32_e32 v6, 0x400000, v3
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; GFX8-NEXT:    v_cndmask_b32_e32 v3, v5, v6, vcc
; GFX8-NEXT:    v_lshrrev_b32_e32 v5, 16, v3
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v0
; GFX8-NEXT:    v_cndmask_b32_e32 v0, v5, v0, vcc
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v2
; GFX8-NEXT:    v_cndmask_b32_e32 v0, v0, v2, vcc
; GFX8-NEXT:    v_and_b32_e32 v2, 0xffff0000, v3
; GFX8-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v2
; GFX8-NEXT:    v_cndmask_b32_e32 v0, v5, v0, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v2, 16, v4
; GFX8-NEXT:    v_or_b32_sdwa v0, v0, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX8-NEXT:    s_setpc_b64 s[30:31]
;
; GFX900-LABEL: v_maximumnum_v3bf16_no_ieee:
; GFX900:       ; %bb.0:
; GFX900-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX900-NEXT:    v_and_b32_e32 v4, 0xffff0000, v0
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v4, v4
; GFX900-NEXT:    v_lshrrev_b32_e32 v4, 16, v2
; GFX900-NEXT:    v_and_b32_e32 v7, 0xffff0000, v2
; GFX900-NEXT:    v_cndmask_b32_sdwa v5, v0, v4, vcc dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v7, v7
; GFX900-NEXT:    v_cndmask_b32_e32 v4, v4, v5, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v6, 16, v5
; GFX900-NEXT:    v_lshlrev_b32_e32 v7, 16, v4
; GFX900-NEXT:    v_cmp_gt_f32_e32 vcc, v6, v7
; GFX900-NEXT:    v_cndmask_b32_e32 v6, v4, v5, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; GFX900-NEXT:    v_max_f32_e32 v6, v6, v6
; GFX900-NEXT:    v_bfe_u32 v7, v6, 16, 1
; GFX900-NEXT:    s_movk_i32 s4, 0x7fff
; GFX900-NEXT:    v_add3_u32 v7, v7, v6, s4
; GFX900-NEXT:    v_or_b32_e32 v8, 0x400000, v6
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v6, v6
; GFX900-NEXT:    v_cndmask_b32_e32 v6, v7, v8, vcc
; GFX900-NEXT:    v_lshrrev_b32_e32 v7, 16, v6
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v5
; GFX900-NEXT:    v_cndmask_b32_e32 v5, v7, v5, vcc
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v4
; GFX900-NEXT:    v_cndmask_b32_e32 v4, v5, v4, vcc
; GFX900-NEXT:    v_and_b32_e32 v5, 0xffff0000, v6
; GFX900-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v5
; GFX900-NEXT:    v_lshlrev_b32_e32 v5, 16, v1
; GFX900-NEXT:    v_cndmask_b32_e32 v4, v7, v4, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; GFX900-NEXT:    v_lshlrev_b32_e32 v5, 16, v3
; GFX900-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; GFX900-NEXT:    v_cndmask_b32_e32 v3, v3, v1, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v5, 16, v3
; GFX900-NEXT:    v_lshlrev_b32_e32 v6, 16, v1
; GFX900-NEXT:    v_cmp_gt_f32_e32 vcc, v6, v5
; GFX900-NEXT:    v_cndmask_b32_e32 v5, v3, v1, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; GFX900-NEXT:    v_max_f32_e32 v5, v5, v5
; GFX900-NEXT:    v_bfe_u32 v6, v5, 16, 1
; GFX900-NEXT:    v_add3_u32 v6, v6, v5, s4
; GFX900-NEXT:    v_or_b32_e32 v7, 0x400000, v5
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; GFX900-NEXT:    v_cndmask_b32_e32 v5, v6, v7, vcc
; GFX900-NEXT:    v_lshrrev_b32_e32 v6, 16, v5
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v1
; GFX900-NEXT:    v_cndmask_b32_e32 v1, v6, v1, vcc
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v3
; GFX900-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc
; GFX900-NEXT:    v_and_b32_e32 v3, 0xffff0000, v5
; GFX900-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v3
; GFX900-NEXT:    v_lshlrev_b32_e32 v3, 16, v0
; GFX900-NEXT:    v_cndmask_b32_e32 v1, v6, v1, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; GFX900-NEXT:    v_lshlrev_b32_e32 v3, 16, v2
; GFX900-NEXT:    v_cndmask_b32_e32 v0, v0, v2, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; GFX900-NEXT:    v_cndmask_b32_e32 v2, v2, v0, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v3, 16, v2
; GFX900-NEXT:    v_lshlrev_b32_e32 v5, 16, v0
; GFX900-NEXT:    v_cmp_gt_f32_e32 vcc, v5, v3
; GFX900-NEXT:    v_cndmask_b32_e32 v3, v2, v0, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; GFX900-NEXT:    v_max_f32_e32 v3, v3, v3
; GFX900-NEXT:    v_bfe_u32 v5, v3, 16, 1
; GFX900-NEXT:    v_add3_u32 v5, v5, v3, s4
; GFX900-NEXT:    v_or_b32_e32 v6, 0x400000, v3
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; GFX900-NEXT:    v_cndmask_b32_e32 v3, v5, v6, vcc
; GFX900-NEXT:    v_lshrrev_b32_e32 v5, 16, v3
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v0
; GFX900-NEXT:    v_cndmask_b32_e32 v0, v5, v0, vcc
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v2
; GFX900-NEXT:    v_cndmask_b32_e32 v0, v0, v2, vcc
; GFX900-NEXT:    v_and_b32_e32 v2, 0xffff0000, v3
; GFX900-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v2
; GFX900-NEXT:    v_cndmask_b32_e32 v0, v5, v0, vcc
; GFX900-NEXT:    s_mov_b32 s4, 0x5040100
; GFX900-NEXT:    v_perm_b32 v0, v4, v0, s4
; GFX900-NEXT:    s_setpc_b64 s[30:31]
;
; GFX950-LABEL: v_maximumnum_v3bf16_no_ieee:
; GFX950:       ; %bb.0:
; GFX950-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX950-NEXT:    v_and_b32_e32 v4, 0xffff0000, v0
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v4, v4
; GFX950-NEXT:    v_lshrrev_b32_e32 v4, 16, v2
; GFX950-NEXT:    v_and_b32_e32 v7, 0xffff0000, v2
; GFX950-NEXT:    v_cndmask_b32_sdwa v5, v0, v4, vcc dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v7, v7
; GFX950-NEXT:    v_lshlrev_b32_e32 v6, 16, v5
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v4, v4, v5, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v7, 16, v4
; GFX950-NEXT:    v_cmp_gt_f32_e32 vcc, v6, v7
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v6, v4, v5, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; GFX950-NEXT:    v_max_f32_e32 v6, v6, v6
; GFX950-NEXT:    v_cvt_pk_bf16_f32 v6, v6, s0
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v5
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v5, v6, v5, vcc
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v4
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v4, v5, v4, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v5, 16, v6
; GFX950-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v5
; GFX950-NEXT:    v_lshlrev_b32_e32 v5, 16, v1
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v4, v6, v4, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; GFX950-NEXT:    v_lshlrev_b32_e32 v5, 16, v3
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; GFX950-NEXT:    v_lshlrev_b32_e32 v6, 16, v1
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v3, v3, v1, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v5, 16, v3
; GFX950-NEXT:    v_cmp_gt_f32_e32 vcc, v6, v5
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v5, v3, v1, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; GFX950-NEXT:    v_max_f32_e32 v5, v5, v5
; GFX950-NEXT:    v_cvt_pk_bf16_f32 v5, v5, s0
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v1
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v1, v5, v1, vcc
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v3
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v3, 16, v5
; GFX950-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v3
; GFX950-NEXT:    v_lshlrev_b32_e32 v3, 16, v0
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v1, v5, v1, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; GFX950-NEXT:    v_lshlrev_b32_e32 v3, 16, v2
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v0, v0, v2, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; GFX950-NEXT:    v_lshlrev_b32_e32 v5, 16, v0
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v2, v2, v0, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v3, 16, v2
; GFX950-NEXT:    v_cmp_gt_f32_e32 vcc, v5, v3
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v3, v2, v0, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; GFX950-NEXT:    v_max_f32_e32 v3, v3, v3
; GFX950-NEXT:    v_cvt_pk_bf16_f32 v3, v3, s0
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v0
; GFX950-NEXT:    s_mov_b32 s0, 0x5040100
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v0, v3, v0, vcc
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v2
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v0, v0, v2, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v2, 16, v3
; GFX950-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v2
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v0, v3, v0, vcc
; GFX950-NEXT:    v_perm_b32 v0, v4, v0, s0
; GFX950-NEXT:    s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_maximumnum_v3bf16_no_ieee:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT:    v_lshlrev_b32_e32 v5, 16, v1
; GFX10-NEXT:    v_and_b32_e32 v4, 0xffff0000, v0
; GFX10-NEXT:    v_lshlrev_b32_e32 v8, 16, v0
; GFX10-NEXT:    v_lshrrev_b32_e32 v6, 16, v2
; GFX10-NEXT:    v_and_b32_e32 v7, 0xffff0000, v2
; GFX10-NEXT:    v_cmp_u_f32_e64 s4, v5, v5
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX10-NEXT:    v_lshlrev_b32_e32 v4, 16, v3
; GFX10-NEXT:    v_cndmask_b32_e64 v1, v1, v3, s4
; GFX10-NEXT:    v_cmp_u_f32_e64 s4, v8, v8
; GFX10-NEXT:    v_lshlrev_b32_e32 v8, 16, v2
; GFX10-NEXT:    v_cndmask_b32_e64 v5, v0, v2, s4
; GFX10-NEXT:    v_cndmask_b32_sdwa v0, v0, v6, vcc_lo dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX10-NEXT:    v_lshlrev_b32_e32 v9, 16, v5
; GFX10-NEXT:    v_lshlrev_b32_e32 v10, 16, v0
; GFX10-NEXT:    v_cndmask_b32_e32 v3, v3, v1, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
; GFX10-NEXT:    v_cndmask_b32_e32 v2, v2, v5, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
; GFX10-NEXT:    v_lshlrev_b32_e32 v7, 16, v1
; GFX10-NEXT:    v_lshlrev_b32_e32 v8, 16, v2
; GFX10-NEXT:    v_cndmask_b32_e32 v4, v6, v0, vcc_lo
; GFX10-NEXT:    v_lshlrev_b32_e32 v6, 16, v3
; GFX10-NEXT:    v_lshlrev_b32_e32 v11, 16, v4
; GFX10-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v7, v6
; GFX10-NEXT:    v_cndmask_b32_e32 v6, v3, v1, vcc_lo
; GFX10-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v9, v8
; GFX10-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; GFX10-NEXT:    v_cndmask_b32_e32 v7, v2, v5, vcc_lo
; GFX10-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v10, v11
; GFX10-NEXT:    v_max_f32_e32 v6, v6, v6
; GFX10-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; GFX10-NEXT:    v_cndmask_b32_e32 v8, v4, v0, vcc_lo
; GFX10-NEXT:    v_bfe_u32 v9, v6, 16, 1
; GFX10-NEXT:    v_max_f32_e32 v7, v7, v7
; GFX10-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; GFX10-NEXT:    v_or_b32_e32 v11, 0x400000, v6
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX10-NEXT:    v_add3_u32 v9, v9, v6, 0x7fff
; GFX10-NEXT:    v_bfe_u32 v10, v7, 16, 1
; GFX10-NEXT:    v_max_f32_e32 v8, v8, v8
; GFX10-NEXT:    v_or_b32_e32 v13, 0x400000, v7
; GFX10-NEXT:    v_cndmask_b32_e32 v6, v9, v11, vcc_lo
; GFX10-NEXT:    v_add3_u32 v10, v10, v7, 0x7fff
; GFX10-NEXT:    v_bfe_u32 v12, v8, 16, 1
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
; GFX10-NEXT:    v_or_b32_e32 v14, 0x400000, v8
; GFX10-NEXT:    v_lshrrev_b32_e32 v9, 16, v6
; GFX10-NEXT:    v_and_b32_e32 v6, 0xffff0000, v6
; GFX10-NEXT:    v_add3_u32 v12, v12, v8, 0x7fff
; GFX10-NEXT:    v_cndmask_b32_e32 v7, v10, v13, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
; GFX10-NEXT:    v_lshrrev_b32_e32 v10, 16, v7
; GFX10-NEXT:    v_cndmask_b32_e32 v8, v12, v14, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v5
; GFX10-NEXT:    v_and_b32_e32 v7, 0xffff0000, v7
; GFX10-NEXT:    v_lshrrev_b32_e32 v11, 16, v8
; GFX10-NEXT:    v_cndmask_b32_e32 v5, v10, v5, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v0
; GFX10-NEXT:    v_and_b32_e32 v8, 0xffff0000, v8
; GFX10-NEXT:    v_cndmask_b32_e32 v0, v11, v0, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v2
; GFX10-NEXT:    v_cndmask_b32_e32 v2, v5, v2, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v4
; GFX10-NEXT:    v_cndmask_b32_e32 v0, v0, v4, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v1
; GFX10-NEXT:    v_cndmask_b32_e32 v1, v9, v1, vcc_lo
; GFX10-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v7
; GFX10-NEXT:    v_cndmask_b32_e32 v2, v10, v2, vcc_lo
; GFX10-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v8
; GFX10-NEXT:    v_cndmask_b32_e32 v0, v11, v0, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v3
; GFX10-NEXT:    v_perm_b32 v0, v0, v2, 0x5040100
; GFX10-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc_lo
; GFX10-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v6
; GFX10-NEXT:    v_cndmask_b32_e32 v1, v9, v1, vcc_lo
; GFX10-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-TRUE16-LABEL: v_maximumnum_v3bf16_no_ieee:
; GFX11-TRUE16:       ; %bb.0:
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xffff0000, v0
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xffff0000, v2
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 16, v1
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 16, v2
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 16, v3
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s1, v5, v5
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s0, v6, v6
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 16, v0
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v4.l, v0.h, v2.h, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v1.l, v1.l, v3.l, s0
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s0, v9, v9
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s2, v8, v8
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v5.l, v2.h, v4.l, s1
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, v4.l
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v3.l, v3.l, v1.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.l, v1.l
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v0.l, v0.l, v2.l, s2
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v9.l, v5.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v10.l, v3.l
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s2, 0, v5.l
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v2.l, v2.l, v0.l, s0
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, v0.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s3, 0, v3.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v11.l, v2.l
; GFX11-TRUE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v6, v9
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s4, 0, v2.l
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v6.l, v5.l, v4.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v4.l
; GFX11-TRUE16-NEXT:    v_cmp_gt_f32_e64 s1, v8, v11
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v8.l, v2.l, v0.l, s1
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_dual_max_f32 v6, v6, v6 :: v_dual_lshlrev_b32 v7, 16, v7
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s1, 0, v0.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v9, v6, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_gt_f32_e64 s0, v7, v10
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, 0x400000, v6
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s5, v6, v6
; GFX11-TRUE16-NEXT:    v_max_f32_e32 v8, v8, v8
; GFX11-TRUE16-NEXT:    v_add3_u32 v9, v9, v6, 0x7fff
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v7.l, v3.l, v1.l, s0
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s0, 0, v1.l
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v12, v8, 16, 1
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e64 v6, v9, v10, s5
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, 0x400000, v8
; GFX11-TRUE16-NEXT:    v_add3_u32 v12, v12, v8, 0x7fff
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v0.h, v6.h, v4.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_dual_max_f32 v7, v7, v7 :: v_dual_and_b32 v4, 0xffff0000, v6
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v0.h, v0.h, v5.l, s2
; GFX11-TRUE16-NEXT:    v_bfe_u32 v11, v7, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v13, 0x400000, v7
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s5, v7, v7
; GFX11-TRUE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v4
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT:    v_add3_u32 v11, v11, v7, 0x7fff
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v0.h, v6.h, v0.h, vcc_lo
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e64 v7, v11, v13, s5
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s5, v8, v8
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v1.l, v7.h, v1.l, s0
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e64 v8, v12, v14, s5
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xffff0000, v7
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v1.l, v1.l, v3.l, s3
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v0.l, v8.h, v0.l, s1
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xffff0000, v8
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_cmp_eq_f32_e64 s1, 0, v9
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v0.l, v0.l, v2.l, s4
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_cmp_eq_f32_e64 s0, 0, v10
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v1.l, v7.h, v1.l, s1
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v0.l, v8.h, v0.l, s0
; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-FAKE16-LABEL: v_maximumnum_v3bf16_no_ieee:
; GFX11-FAKE16:       ; %bb.0:
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xffff0000, v0
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v5, 16, v2
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v6, 16, v0
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v9, 16, v0
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v4, v6, v5, vcc_lo
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v6, 0xffff0000, v2
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v10, 16, v4
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v5, v5, v4, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v0, v0, v2 :: v_dual_lshlrev_b32 v7, 16, v1
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v6, 16, v2
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc_lo
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v8, 16, v3
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v2, v2, v0 :: v_dual_lshlrev_b32 v7, 16, v5
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v3, v3, v1 :: v_dual_lshlrev_b32 v8, 16, v0
; GFX11-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v10, v7
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v7, 16, v2
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v10, 16, v1
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v6, v5, v4 :: v_dual_lshlrev_b32 v9, 16, v3
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v8, v7
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v7, v2, v0 :: v_dual_lshlrev_b32 v6, 16, v6
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v10, v9
; GFX11-FAKE16-NEXT:    v_dual_max_f32 v6, v6, v6 :: v_dual_lshlrev_b32 v7, 16, v7
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v8, v3, v1, vcc_lo
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v9, v6, 16, 1
; GFX11-FAKE16-NEXT:    v_max_f32_e32 v7, v7, v7
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v10, 0x400000, v6
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; GFX11-FAKE16-NEXT:    v_add3_u32 v9, v9, v6, 0x7fff
; GFX11-FAKE16-NEXT:    v_bfe_u32 v11, v7, 16, 1
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v6, v9, v10, vcc_lo
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, 0x400000, v7
; GFX11-FAKE16-NEXT:    v_add3_u32 v10, v11, v7, 0x7fff
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v13, 16, v6
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v7, v10, v9 :: v_dual_and_b32 v6, 0xffff0000, v6
; GFX11-FAKE16-NEXT:    v_max_f32_e32 v8, v8, v8
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v9, 16, v7
; GFX11-FAKE16-NEXT:    v_bfe_u32 v12, v8, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v11, 0x400000, v8
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_add3_u32 v12, v12, v8, 0x7fff
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v8, v12, v11, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v4
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v10, 16, v8
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v4, v13, v4, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v5
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v4, v4, v5 :: v_dual_and_b32 v5, 0xffff0000, v7
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v0
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v9, v0, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v6
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v4, v13, v4, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v2
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v0, v2, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v1
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v1, v10, v1 :: v_dual_and_b32 v2, 0xffff0000, v8
; GFX11-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v5
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v9, v0, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v3
; GFX11-FAKE16-NEXT:    v_perm_b32 v0, v4, v0, 0x5040100
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v2
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v10, v1, vcc_lo
; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
;
; GFX12-TRUE16-LABEL: v_maximumnum_v3bf16_no_ieee:
; GFX12-TRUE16:       ; %bb.0:
; GFX12-TRUE16-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX12-TRUE16-NEXT:    s_wait_expcnt 0x0
; GFX12-TRUE16-NEXT:    s_wait_samplecnt 0x0
; GFX12-TRUE16-NEXT:    s_wait_bvhcnt 0x0
; GFX12-TRUE16-NEXT:    s_wait_kmcnt 0x0
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v4, 0xffff0000, v0
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v5, 0xffff0000, v2
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 16, v1
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 16, v2
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 16, v3
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s1, v5, v5
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s0, v6, v6
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 16, v0
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v4.l, v0.h, v2.h, vcc_lo
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v1.l, v1.l, v3.l, s0
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s0, v9, v9
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s2, v8, v8
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v5.l, v2.h, v4.l, s1
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v6.l, v4.l
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v3.l, v3.l, v1.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v7.l, v1.l
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v0.l, v0.l, v2.l, s2
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v9.l, v5.l
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v10.l, v3.l
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s2, 0, v5.l
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v2.l, v2.l, v0.l, s0
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v8.l, v0.l
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s3, 0, v3.l
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v11.l, v2.l
; GFX12-TRUE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v6, v9
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s4, 0, v2.l
; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v6.l, v5.l, v4.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v4.l
; GFX12-TRUE16-NEXT:    v_cmp_gt_f32_e64 s1, v8, v11
; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_2)
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v8.l, v2.l, v0.l, s1
; GFX12-TRUE16-NEXT:    v_dual_max_num_f32 v6, v6, v6 :: v_dual_lshlrev_b32 v7, 16, v7
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s1, 0, v0.l
; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; GFX12-TRUE16-NEXT:    v_bfe_u32 v9, v6, 16, 1
; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX12-TRUE16-NEXT:    v_cmp_gt_f32_e64 s0, v7, v10
; GFX12-TRUE16-NEXT:    v_or_b32_e32 v10, 0x400000, v6
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s5, v6, v6
; GFX12-TRUE16-NEXT:    v_max_num_f32_e32 v8, v8, v8
; GFX12-TRUE16-NEXT:    v_add3_u32 v9, v9, v6, 0x7fff
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v7.l, v3.l, v1.l, s0
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s0, 0, v1.l
; GFX12-TRUE16-NEXT:    v_bfe_u32 v12, v8, 16, 1
; GFX12-TRUE16-NEXT:    v_cndmask_b32_e64 v6, v9, v10, s5
; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; GFX12-TRUE16-NEXT:    v_or_b32_e32 v14, 0x400000, v8
; GFX12-TRUE16-NEXT:    v_add3_u32 v12, v12, v8, 0x7fff
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v0.h, v6.h, v4.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_dual_max_num_f32 v7, v7, v7 :: v_dual_and_b32 v4, 0xffff0000, v6
; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v0.h, v0.h, v5.l, s2
; GFX12-TRUE16-NEXT:    v_bfe_u32 v11, v7, 16, 1
; GFX12-TRUE16-NEXT:    v_or_b32_e32 v13, 0x400000, v7
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s5, v7, v7
; GFX12-TRUE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v4
; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_2)
; GFX12-TRUE16-NEXT:    v_add3_u32 v11, v11, v7, 0x7fff
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v0.h, v6.h, v0.h, vcc_lo
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b32_e64 v7, v11, v13, s5
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s5, v8, v8
; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v1.l, v7.h, v1.l, s0
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b32_e64 v8, v12, v14, s5
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v9, 0xffff0000, v7
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v1.l, v1.l, v3.l, s3
; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v0.l, v8.h, v0.l, s1
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v10, 0xffff0000, v8
; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX12-TRUE16-NEXT:    v_cmp_eq_f32_e64 s1, 0, v9
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v0.l, v0.l, v2.l, s4
; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX12-TRUE16-NEXT:    v_cmp_eq_f32_e64 s0, 0, v10
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v1.l, v7.h, v1.l, s1
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v0.l, v8.h, v0.l, s0
; GFX12-TRUE16-NEXT:    s_setpc_b64 s[30:31]
;
; GFX12-FAKE16-LABEL: v_maximumnum_v3bf16_no_ieee:
; GFX12-FAKE16:       ; %bb.0:
; GFX12-FAKE16-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX12-FAKE16-NEXT:    s_wait_expcnt 0x0
; GFX12-FAKE16-NEXT:    s_wait_samplecnt 0x0
; GFX12-FAKE16-NEXT:    s_wait_bvhcnt 0x0
; GFX12-FAKE16-NEXT:    s_wait_kmcnt 0x0
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v4, 0xffff0000, v0
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v5, 16, v2
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v6, 16, v0
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v9, 16, v0
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v4, v6, v5, vcc_lo
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v6, 0xffff0000, v2
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v10, 16, v4
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v5, v5, v4, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v0, v0, v2 :: v_dual_lshlrev_b32 v7, 16, v1
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v6, 16, v2
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc_lo
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v8, 16, v3
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v2, v2, v0 :: v_dual_lshlrev_b32 v7, 16, v5
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v3, v3, v1 :: v_dual_lshlrev_b32 v8, 16, v0
; GFX12-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v10, v7
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v7, 16, v2
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v10, 16, v1
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v6, v5, v4 :: v_dual_lshlrev_b32 v9, 16, v3
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX12-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v8, v7
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v7, v2, v0 :: v_dual_lshlrev_b32 v6, 16, v6
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX12-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v10, v9
; GFX12-FAKE16-NEXT:    v_dual_max_num_f32 v6, v6, v6 :: v_dual_lshlrev_b32 v7, 16, v7
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v8, v3, v1, vcc_lo
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX12-FAKE16-NEXT:    v_bfe_u32 v9, v6, 16, 1
; GFX12-FAKE16-NEXT:    v_max_num_f32_e32 v7, v7, v7
; GFX12-FAKE16-NEXT:    v_or_b32_e32 v10, 0x400000, v6
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; GFX12-FAKE16-NEXT:    v_add3_u32 v9, v9, v6, 0x7fff
; GFX12-FAKE16-NEXT:    v_bfe_u32 v11, v7, 16, 1
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v6, v9, v10, vcc_lo
; GFX12-FAKE16-NEXT:    v_or_b32_e32 v9, 0x400000, v7
; GFX12-FAKE16-NEXT:    v_add3_u32 v10, v11, v7, 0x7fff
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v13, 16, v6
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v7, v10, v9 :: v_dual_and_b32 v6, 0xffff0000, v6
; GFX12-FAKE16-NEXT:    v_max_num_f32_e32 v8, v8, v8
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v9, 16, v7
; GFX12-FAKE16-NEXT:    v_bfe_u32 v12, v8, 16, 1
; GFX12-FAKE16-NEXT:    v_or_b32_e32 v11, 0x400000, v8
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX12-FAKE16-NEXT:    v_add3_u32 v12, v12, v8, 0x7fff
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v8, v12, v11, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v4
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_4) | instid1(VALU_DEP_2)
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v10, 16, v8
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v4, v13, v4, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v5
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v4, v4, v5 :: v_dual_and_b32 v5, 0xffff0000, v7
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v0
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v9, v0, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v6
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v4, v13, v4, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v2
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v0, v2, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v1
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v1, v10, v1 :: v_dual_and_b32 v2, 0xffff0000, v8
; GFX12-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v5
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v9, v0, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v3
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_4) | instid1(VALU_DEP_2)
; GFX12-FAKE16-NEXT:    v_perm_b32 v0, v4, v0, 0x5040100
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v2
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v10, v1, vcc_lo
; GFX12-FAKE16-NEXT:    s_setpc_b64 s[30:31]
  %result = call <3 x bfloat> @llvm.maximumnum.v3bf16(<3 x bfloat> %x, <3 x bfloat> %y)
  ret <3 x bfloat> %result
}

define <4 x bfloat> @v_maximumnum_v4bf16_no_ieee(<4 x bfloat> %x, <4 x bfloat> %y) #0 {
; GFX7-LABEL: v_maximumnum_v4bf16_no_ieee:
; GFX7:       ; %bb.0:
; GFX7-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT:    v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT:    v_mul_f32_e32 v4, 1.0, v4
; GFX7-NEXT:    v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT:    v_mul_f32_e32 v5, 1.0, v5
; GFX7-NEXT:    v_mul_f32_e32 v2, 1.0, v2
; GFX7-NEXT:    v_mul_f32_e32 v6, 1.0, v6
; GFX7-NEXT:    v_mul_f32_e32 v3, 1.0, v3
; GFX7-NEXT:    v_mul_f32_e32 v7, 1.0, v7
; GFX7-NEXT:    v_and_b32_e32 v7, 0xffff0000, v7
; GFX7-NEXT:    v_and_b32_e32 v3, 0xffff0000, v3
; GFX7-NEXT:    v_and_b32_e32 v6, 0xffff0000, v6
; GFX7-NEXT:    v_and_b32_e32 v2, 0xffff0000, v2
; GFX7-NEXT:    v_and_b32_e32 v5, 0xffff0000, v5
; GFX7-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT:    v_and_b32_e32 v4, 0xffff0000, v4
; GFX7-NEXT:    v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT:    v_max_f32_e32 v3, v3, v7
; GFX7-NEXT:    v_max_f32_e32 v2, v2, v6
; GFX7-NEXT:    v_max_f32_e32 v1, v1, v5
; GFX7-NEXT:    v_max_f32_e32 v0, v0, v4
; GFX7-NEXT:    v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT:    v_and_b32_e32 v2, 0xffff0000, v2
; GFX7-NEXT:    v_and_b32_e32 v3, 0xffff0000, v3
; GFX7-NEXT:    s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_maximumnum_v4bf16_no_ieee:
; GFX8:       ; %bb.0:
; GFX8-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT:    v_and_b32_e32 v4, 0xffff0000, v1
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v4, v4
; GFX8-NEXT:    v_lshrrev_b32_e32 v4, 16, v3
; GFX8-NEXT:    v_and_b32_e32 v7, 0xffff0000, v3
; GFX8-NEXT:    v_cndmask_b32_sdwa v5, v1, v4, vcc dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v7, v7
; GFX8-NEXT:    v_cndmask_b32_e32 v4, v4, v5, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v6, 16, v5
; GFX8-NEXT:    v_lshlrev_b32_e32 v7, 16, v4
; GFX8-NEXT:    v_cmp_gt_f32_e32 vcc, v6, v7
; GFX8-NEXT:    v_cndmask_b32_e32 v6, v4, v5, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; GFX8-NEXT:    v_mul_f32_e32 v6, 1.0, v6
; GFX8-NEXT:    v_bfe_u32 v7, v6, 16, 1
; GFX8-NEXT:    v_add_u32_e32 v7, vcc, v7, v6
; GFX8-NEXT:    s_movk_i32 s4, 0x7fff
; GFX8-NEXT:    v_add_u32_e32 v7, vcc, s4, v7
; GFX8-NEXT:    v_or_b32_e32 v8, 0x400000, v6
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v6, v6
; GFX8-NEXT:    v_cndmask_b32_e32 v6, v7, v8, vcc
; GFX8-NEXT:    v_lshrrev_b32_e32 v7, 16, v6
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v5
; GFX8-NEXT:    v_cndmask_b32_e32 v5, v7, v5, vcc
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v4
; GFX8-NEXT:    v_cndmask_b32_e32 v4, v5, v4, vcc
; GFX8-NEXT:    v_and_b32_e32 v5, 0xffff0000, v6
; GFX8-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v5
; GFX8-NEXT:    v_and_b32_e32 v5, 0xffff0000, v0
; GFX8-NEXT:    v_cndmask_b32_e32 v4, v7, v4, vcc
; GFX8-NEXT:    v_lshrrev_b32_e32 v6, 16, v2
; GFX8-NEXT:    v_lshrrev_b32_e32 v7, 16, v0
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; GFX8-NEXT:    v_and_b32_e32 v8, 0xffff0000, v2
; GFX8-NEXT:    v_cndmask_b32_e32 v5, v7, v6, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v8, v8
; GFX8-NEXT:    v_cndmask_b32_e32 v6, v6, v5, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v7, 16, v5
; GFX8-NEXT:    v_lshlrev_b32_e32 v8, 16, v6
; GFX8-NEXT:    v_cmp_gt_f32_e32 vcc, v7, v8
; GFX8-NEXT:    v_cndmask_b32_e32 v7, v6, v5, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; GFX8-NEXT:    v_mul_f32_e32 v7, 1.0, v7
; GFX8-NEXT:    v_bfe_u32 v8, v7, 16, 1
; GFX8-NEXT:    v_add_u32_e32 v8, vcc, v8, v7
; GFX8-NEXT:    v_add_u32_e32 v8, vcc, s4, v8
; GFX8-NEXT:    v_or_b32_e32 v9, 0x400000, v7
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v7, v7
; GFX8-NEXT:    v_cndmask_b32_e32 v7, v8, v9, vcc
; GFX8-NEXT:    v_lshrrev_b32_e32 v8, 16, v7
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v5
; GFX8-NEXT:    v_cndmask_b32_e32 v5, v8, v5, vcc
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v6
; GFX8-NEXT:    v_cndmask_b32_e32 v5, v5, v6, vcc
; GFX8-NEXT:    v_and_b32_e32 v6, 0xffff0000, v7
; GFX8-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v6
; GFX8-NEXT:    v_lshlrev_b32_e32 v6, 16, v1
; GFX8-NEXT:    v_cndmask_b32_e32 v5, v8, v5, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v6, v6
; GFX8-NEXT:    v_lshlrev_b32_e32 v6, 16, v3
; GFX8-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v6, v6
; GFX8-NEXT:    v_cndmask_b32_e32 v3, v3, v1, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v6, 16, v3
; GFX8-NEXT:    v_lshlrev_b32_e32 v7, 16, v1
; GFX8-NEXT:    v_cmp_gt_f32_e32 vcc, v7, v6
; GFX8-NEXT:    v_cndmask_b32_e32 v6, v3, v1, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; GFX8-NEXT:    v_mul_f32_e32 v6, 1.0, v6
; GFX8-NEXT:    v_bfe_u32 v7, v6, 16, 1
; GFX8-NEXT:    v_add_u32_e32 v7, vcc, v7, v6
; GFX8-NEXT:    v_add_u32_e32 v7, vcc, s4, v7
; GFX8-NEXT:    v_or_b32_e32 v8, 0x400000, v6
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v6, v6
; GFX8-NEXT:    v_cndmask_b32_e32 v6, v7, v8, vcc
; GFX8-NEXT:    v_lshrrev_b32_e32 v7, 16, v6
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v1
; GFX8-NEXT:    v_cndmask_b32_e32 v1, v7, v1, vcc
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v3
; GFX8-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc
; GFX8-NEXT:    v_and_b32_e32 v3, 0xffff0000, v6
; GFX8-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v3
; GFX8-NEXT:    v_lshlrev_b32_e32 v3, 16, v0
; GFX8-NEXT:    v_cndmask_b32_e32 v1, v7, v1, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; GFX8-NEXT:    v_lshlrev_b32_e32 v3, 16, v2
; GFX8-NEXT:    v_cndmask_b32_e32 v0, v0, v2, vcc
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; GFX8-NEXT:    v_cndmask_b32_e32 v2, v2, v0, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v3, 16, v2
; GFX8-NEXT:    v_lshlrev_b32_e32 v6, 16, v0
; GFX8-NEXT:    v_cmp_gt_f32_e32 vcc, v6, v3
; GFX8-NEXT:    v_cndmask_b32_e32 v3, v2, v0, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; GFX8-NEXT:    v_mul_f32_e32 v3, 1.0, v3
; GFX8-NEXT:    v_bfe_u32 v6, v3, 16, 1
; GFX8-NEXT:    v_add_u32_e32 v6, vcc, v6, v3
; GFX8-NEXT:    v_add_u32_e32 v6, vcc, s4, v6
; GFX8-NEXT:    v_or_b32_e32 v7, 0x400000, v3
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; GFX8-NEXT:    v_cndmask_b32_e32 v3, v6, v7, vcc
; GFX8-NEXT:    v_lshrrev_b32_e32 v6, 16, v3
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v0
; GFX8-NEXT:    v_cndmask_b32_e32 v0, v6, v0, vcc
; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v2
; GFX8-NEXT:    v_cndmask_b32_e32 v0, v0, v2, vcc
; GFX8-NEXT:    v_and_b32_e32 v2, 0xffff0000, v3
; GFX8-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v2
; GFX8-NEXT:    v_cndmask_b32_e32 v0, v6, v0, vcc
; GFX8-NEXT:    v_lshlrev_b32_e32 v2, 16, v5
; GFX8-NEXT:    v_or_b32_sdwa v0, v0, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX8-NEXT:    v_lshlrev_b32_e32 v2, 16, v4
; GFX8-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX8-NEXT:    s_setpc_b64 s[30:31]
;
; GFX900-LABEL: v_maximumnum_v4bf16_no_ieee:
; GFX900:       ; %bb.0:
; GFX900-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX900-NEXT:    v_and_b32_e32 v4, 0xffff0000, v1
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v4, v4
; GFX900-NEXT:    v_lshrrev_b32_e32 v4, 16, v3
; GFX900-NEXT:    v_and_b32_e32 v7, 0xffff0000, v3
; GFX900-NEXT:    v_cndmask_b32_sdwa v5, v1, v4, vcc dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v7, v7
; GFX900-NEXT:    v_cndmask_b32_e32 v4, v4, v5, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v6, 16, v5
; GFX900-NEXT:    v_lshlrev_b32_e32 v7, 16, v4
; GFX900-NEXT:    v_cmp_gt_f32_e32 vcc, v6, v7
; GFX900-NEXT:    v_cndmask_b32_e32 v6, v4, v5, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; GFX900-NEXT:    v_max_f32_e32 v6, v6, v6
; GFX900-NEXT:    v_bfe_u32 v7, v6, 16, 1
; GFX900-NEXT:    s_movk_i32 s4, 0x7fff
; GFX900-NEXT:    v_add3_u32 v7, v7, v6, s4
; GFX900-NEXT:    v_or_b32_e32 v8, 0x400000, v6
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v6, v6
; GFX900-NEXT:    v_cndmask_b32_e32 v6, v7, v8, vcc
; GFX900-NEXT:    v_lshrrev_b32_e32 v7, 16, v6
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v5
; GFX900-NEXT:    v_cndmask_b32_e32 v5, v7, v5, vcc
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v4
; GFX900-NEXT:    v_cndmask_b32_e32 v4, v5, v4, vcc
; GFX900-NEXT:    v_and_b32_e32 v5, 0xffff0000, v6
; GFX900-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v5
; GFX900-NEXT:    v_and_b32_e32 v5, 0xffff0000, v0
; GFX900-NEXT:    v_cndmask_b32_e32 v4, v7, v4, vcc
; GFX900-NEXT:    v_lshrrev_b32_e32 v6, 16, v2
; GFX900-NEXT:    v_lshrrev_b32_e32 v7, 16, v0
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; GFX900-NEXT:    v_and_b32_e32 v8, 0xffff0000, v2
; GFX900-NEXT:    v_cndmask_b32_e32 v5, v7, v6, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v8, v8
; GFX900-NEXT:    v_cndmask_b32_e32 v6, v6, v5, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v7, 16, v5
; GFX900-NEXT:    v_lshlrev_b32_e32 v8, 16, v6
; GFX900-NEXT:    v_cmp_gt_f32_e32 vcc, v7, v8
; GFX900-NEXT:    v_cndmask_b32_e32 v7, v6, v5, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; GFX900-NEXT:    v_max_f32_e32 v7, v7, v7
; GFX900-NEXT:    v_bfe_u32 v8, v7, 16, 1
; GFX900-NEXT:    v_add3_u32 v8, v8, v7, s4
; GFX900-NEXT:    v_or_b32_e32 v9, 0x400000, v7
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v7, v7
; GFX900-NEXT:    v_cndmask_b32_e32 v7, v8, v9, vcc
; GFX900-NEXT:    v_lshrrev_b32_e32 v8, 16, v7
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v5
; GFX900-NEXT:    v_cndmask_b32_e32 v5, v8, v5, vcc
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v6
; GFX900-NEXT:    v_cndmask_b32_e32 v5, v5, v6, vcc
; GFX900-NEXT:    v_and_b32_e32 v6, 0xffff0000, v7
; GFX900-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v6
; GFX900-NEXT:    v_lshlrev_b32_e32 v6, 16, v1
; GFX900-NEXT:    v_cndmask_b32_e32 v5, v8, v5, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v6, v6
; GFX900-NEXT:    v_lshlrev_b32_e32 v6, 16, v3
; GFX900-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v6, v6
; GFX900-NEXT:    v_cndmask_b32_e32 v3, v3, v1, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v6, 16, v3
; GFX900-NEXT:    v_lshlrev_b32_e32 v7, 16, v1
; GFX900-NEXT:    v_cmp_gt_f32_e32 vcc, v7, v6
; GFX900-NEXT:    v_cndmask_b32_e32 v6, v3, v1, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; GFX900-NEXT:    v_max_f32_e32 v6, v6, v6
; GFX900-NEXT:    v_bfe_u32 v7, v6, 16, 1
; GFX900-NEXT:    v_add3_u32 v7, v7, v6, s4
; GFX900-NEXT:    v_or_b32_e32 v8, 0x400000, v6
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v6, v6
; GFX900-NEXT:    v_cndmask_b32_e32 v6, v7, v8, vcc
; GFX900-NEXT:    v_lshrrev_b32_e32 v7, 16, v6
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v1
; GFX900-NEXT:    v_cndmask_b32_e32 v1, v7, v1, vcc
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v3
; GFX900-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc
; GFX900-NEXT:    v_and_b32_e32 v3, 0xffff0000, v6
; GFX900-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v3
; GFX900-NEXT:    v_lshlrev_b32_e32 v3, 16, v0
; GFX900-NEXT:    v_cndmask_b32_e32 v1, v7, v1, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; GFX900-NEXT:    v_lshlrev_b32_e32 v3, 16, v2
; GFX900-NEXT:    v_cndmask_b32_e32 v0, v0, v2, vcc
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; GFX900-NEXT:    v_cndmask_b32_e32 v2, v2, v0, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v3, 16, v2
; GFX900-NEXT:    v_lshlrev_b32_e32 v6, 16, v0
; GFX900-NEXT:    v_cmp_gt_f32_e32 vcc, v6, v3
; GFX900-NEXT:    v_cndmask_b32_e32 v3, v2, v0, vcc
; GFX900-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; GFX900-NEXT:    v_max_f32_e32 v3, v3, v3
; GFX900-NEXT:    v_bfe_u32 v6, v3, 16, 1
; GFX900-NEXT:    v_add3_u32 v6, v6, v3, s4
; GFX900-NEXT:    v_or_b32_e32 v7, 0x400000, v3
; GFX900-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; GFX900-NEXT:    v_cndmask_b32_e32 v3, v6, v7, vcc
; GFX900-NEXT:    v_lshrrev_b32_e32 v6, 16, v3
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v0
; GFX900-NEXT:    v_cndmask_b32_e32 v0, v6, v0, vcc
; GFX900-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v2
; GFX900-NEXT:    v_cndmask_b32_e32 v0, v0, v2, vcc
; GFX900-NEXT:    v_and_b32_e32 v2, 0xffff0000, v3
; GFX900-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v2
; GFX900-NEXT:    v_cndmask_b32_e32 v0, v6, v0, vcc
; GFX900-NEXT:    s_mov_b32 s4, 0x5040100
; GFX900-NEXT:    v_perm_b32 v0, v5, v0, s4
; GFX900-NEXT:    v_perm_b32 v1, v4, v1, s4
; GFX900-NEXT:    s_setpc_b64 s[30:31]
;
; GFX950-LABEL: v_maximumnum_v4bf16_no_ieee:
; GFX950:       ; %bb.0:
; GFX950-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX950-NEXT:    v_and_b32_e32 v4, 0xffff0000, v1
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v4, v4
; GFX950-NEXT:    v_lshrrev_b32_e32 v4, 16, v3
; GFX950-NEXT:    v_and_b32_e32 v7, 0xffff0000, v3
; GFX950-NEXT:    v_cndmask_b32_sdwa v5, v1, v4, vcc dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v7, v7
; GFX950-NEXT:    v_lshlrev_b32_e32 v6, 16, v5
; GFX950-NEXT:    v_and_b32_e32 v8, 0xffff0000, v2
; GFX950-NEXT:    v_cndmask_b32_e32 v4, v4, v5, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v7, 16, v4
; GFX950-NEXT:    v_cmp_gt_f32_e32 vcc, v6, v7
; GFX950-NEXT:    v_lshrrev_b32_e32 v7, 16, v0
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v6, v4, v5, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; GFX950-NEXT:    v_max_f32_e32 v6, v6, v6
; GFX950-NEXT:    v_cvt_pk_bf16_f32 v6, v6, s0
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v5
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v5, v6, v5, vcc
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v4
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v4, v5, v4, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v5, 16, v6
; GFX950-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v5
; GFX950-NEXT:    v_and_b32_e32 v5, 0xffff0000, v0
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v4, v6, v4, vcc
; GFX950-NEXT:    v_lshrrev_b32_e32 v6, 16, v2
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v5, v7, v6, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v8, v8
; GFX950-NEXT:    v_lshlrev_b32_e32 v7, 16, v5
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v6, v6, v5, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v8, 16, v6
; GFX950-NEXT:    v_cmp_gt_f32_e32 vcc, v7, v8
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v7, v6, v5, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; GFX950-NEXT:    v_max_f32_e32 v7, v7, v7
; GFX950-NEXT:    v_cvt_pk_bf16_f32 v7, v7, s0
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v5
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v5, v7, v5, vcc
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v6
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v5, v5, v6, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v6, 16, v7
; GFX950-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v6
; GFX950-NEXT:    v_lshlrev_b32_e32 v6, 16, v1
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v5, v7, v5, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v6, v6
; GFX950-NEXT:    v_lshlrev_b32_e32 v6, 16, v3
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v6, v6
; GFX950-NEXT:    v_lshlrev_b32_e32 v7, 16, v1
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v3, v3, v1, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v6, 16, v3
; GFX950-NEXT:    v_cmp_gt_f32_e32 vcc, v7, v6
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v6, v3, v1, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; GFX950-NEXT:    v_max_f32_e32 v6, v6, v6
; GFX950-NEXT:    v_cvt_pk_bf16_f32 v6, v6, s0
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v1
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v1, v6, v1, vcc
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v3
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v3, 16, v6
; GFX950-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v3
; GFX950-NEXT:    v_lshlrev_b32_e32 v3, 16, v0
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v1, v6, v1, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; GFX950-NEXT:    v_lshlrev_b32_e32 v3, 16, v2
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v0, v0, v2, vcc
; GFX950-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; GFX950-NEXT:    v_lshlrev_b32_e32 v6, 16, v0
; GFX950-NEXT:    s_nop 0
; GFX950-NEXT:    v_cndmask_b32_e32 v2, v2, v0, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v3, 16, v2
; GFX950-NEXT:    v_cmp_gt_f32_e32 vcc, v6, v3
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v3, v2, v0, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; GFX950-NEXT:    v_max_f32_e32 v3, v3, v3
; GFX950-NEXT:    v_cvt_pk_bf16_f32 v3, v3, s0
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v0
; GFX950-NEXT:    s_mov_b32 s0, 0x5040100
; GFX950-NEXT:    v_perm_b32 v1, v4, v1, s0
; GFX950-NEXT:    v_cndmask_b32_e32 v0, v3, v0, vcc
; GFX950-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v2
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v0, v0, v2, vcc
; GFX950-NEXT:    v_lshlrev_b32_e32 v2, 16, v3
; GFX950-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v2
; GFX950-NEXT:    s_nop 1
; GFX950-NEXT:    v_cndmask_b32_e32 v0, v3, v0, vcc
; GFX950-NEXT:    v_perm_b32 v0, v5, v0, s0
; GFX950-NEXT:    s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_maximumnum_v4bf16_no_ieee:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT:    v_and_b32_e32 v4, 0xffff0000, v1
; GFX10-NEXT:    v_lshrrev_b32_e32 v5, 16, v3
; GFX10-NEXT:    v_and_b32_e32 v6, 0xffff0000, v3
; GFX10-NEXT:    v_lshrrev_b32_e32 v7, 16, v2
; GFX10-NEXT:    v_lshrrev_b32_e32 v8, 16, v0
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX10-NEXT:    v_and_b32_e32 v4, 0xffff0000, v0
; GFX10-NEXT:    v_lshlrev_b32_e32 v11, 16, v1
; GFX10-NEXT:    v_lshlrev_b32_e32 v13, 16, v0
; GFX10-NEXT:    v_and_b32_e32 v9, 0xffff0000, v2
; GFX10-NEXT:    v_cndmask_b32_sdwa v10, v1, v5, vcc_lo dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX10-NEXT:    v_lshlrev_b32_e32 v12, 16, v3
; GFX10-NEXT:    v_lshlrev_b32_e32 v6, 16, v2
; GFX10-NEXT:    v_cndmask_b32_e32 v5, v5, v10, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX10-NEXT:    v_lshlrev_b32_e32 v14, 16, v5
; GFX10-NEXT:    v_cndmask_b32_e32 v4, v8, v7, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v11, v11
; GFX10-NEXT:    v_lshlrev_b32_e32 v8, 16, v10
; GFX10-NEXT:    v_lshlrev_b32_e32 v11, 16, v4
; GFX10-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v13, v13
; GFX10-NEXT:    v_cndmask_b32_e32 v0, v0, v2, vcc_lo
; GFX10-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v8, v14
; GFX10-NEXT:    v_cndmask_b32_e32 v8, v5, v10, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
; GFX10-NEXT:    v_lshlrev_b32_e32 v9, 16, v1
; GFX10-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; GFX10-NEXT:    v_cndmask_b32_e32 v7, v7, v4, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v12, v12
; GFX10-NEXT:    v_max_f32_e32 v8, v8, v8
; GFX10-NEXT:    v_lshlrev_b32_e32 v12, 16, v7
; GFX10-NEXT:    v_cndmask_b32_e32 v3, v3, v1, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX10-NEXT:    v_lshlrev_b32_e32 v6, 16, v0
; GFX10-NEXT:    v_bfe_u32 v14, v8, 16, 1
; GFX10-NEXT:    v_lshlrev_b32_e32 v13, 16, v3
; GFX10-NEXT:    v_cndmask_b32_e32 v2, v2, v0, vcc_lo
; GFX10-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v11, v12
; GFX10-NEXT:    v_lshlrev_b32_e32 v12, 16, v2
; GFX10-NEXT:    v_cndmask_b32_e32 v11, v7, v4, vcc_lo
; GFX10-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v9, v13
; GFX10-NEXT:    v_or_b32_e32 v13, 0x400000, v8
; GFX10-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
; GFX10-NEXT:    v_cndmask_b32_e32 v9, v3, v1, vcc_lo
; GFX10-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v6, v12
; GFX10-NEXT:    v_add3_u32 v12, v14, v8, 0x7fff
; GFX10-NEXT:    v_max_f32_e32 v11, v11, v11
; GFX10-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; GFX10-NEXT:    v_cndmask_b32_e32 v6, v2, v0, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
; GFX10-NEXT:    v_or_b32_e32 v14, 0x400000, v11
; GFX10-NEXT:    v_max_f32_e32 v9, v9, v9
; GFX10-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; GFX10-NEXT:    v_cndmask_b32_e32 v8, v12, v13, vcc_lo
; GFX10-NEXT:    v_bfe_u32 v12, v11, 16, 1
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v10
; GFX10-NEXT:    v_bfe_u32 v15, v9, 16, 1
; GFX10-NEXT:    v_max_f32_e32 v6, v6, v6
; GFX10-NEXT:    v_lshrrev_b32_e32 v13, 16, v8
; GFX10-NEXT:    v_add3_u32 v12, v12, v11, 0x7fff
; GFX10-NEXT:    v_and_b32_e32 v8, 0xffff0000, v8
; GFX10-NEXT:    v_or_b32_e32 v16, 0x400000, v6
; GFX10-NEXT:    v_cndmask_b32_e32 v10, v13, v10, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v11, v11
; GFX10-NEXT:    v_cndmask_b32_e32 v11, v12, v14, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v5
; GFX10-NEXT:    v_bfe_u32 v12, v6, 16, 1
; GFX10-NEXT:    v_add3_u32 v14, v15, v9, 0x7fff
; GFX10-NEXT:    v_or_b32_e32 v15, 0x400000, v9
; GFX10-NEXT:    v_cndmask_b32_e32 v5, v10, v5, vcc_lo
; GFX10-NEXT:    v_lshrrev_b32_e32 v10, 16, v11
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v4
; GFX10-NEXT:    v_add3_u32 v12, v12, v6, 0x7fff
; GFX10-NEXT:    v_and_b32_e32 v11, 0xffff0000, v11
; GFX10-NEXT:    v_cndmask_b32_e32 v4, v10, v4, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
; GFX10-NEXT:    v_cndmask_b32_e32 v9, v14, v15, vcc_lo
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX10-NEXT:    v_cndmask_b32_e32 v6, v12, v16, vcc_lo
; GFX10-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v8
; GFX10-NEXT:    v_lshrrev_b32_e32 v8, 16, v9
; GFX10-NEXT:    v_cndmask_b32_e32 v5, v13, v5, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v7
; GFX10-NEXT:    v_cndmask_b32_e32 v4, v4, v7, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v1
; GFX10-NEXT:    v_lshrrev_b32_e32 v7, 16, v6
; GFX10-NEXT:    v_cndmask_b32_e32 v1, v8, v1, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v0
; GFX10-NEXT:    v_cndmask_b32_e32 v0, v7, v0, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v3
; GFX10-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc_lo
; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v2
; GFX10-NEXT:    v_and_b32_e32 v3, 0xffff0000, v6
; GFX10-NEXT:    v_cndmask_b32_e32 v0, v0, v2, vcc_lo
; GFX10-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v11
; GFX10-NEXT:    v_and_b32_e32 v2, 0xffff0000, v9
; GFX10-NEXT:    v_cndmask_b32_e32 v4, v10, v4, vcc_lo
; GFX10-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v3
; GFX10-NEXT:    v_cndmask_b32_e32 v0, v7, v0, vcc_lo
; GFX10-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v2
; GFX10-NEXT:    v_perm_b32 v0, v4, v0, 0x5040100
; GFX10-NEXT:    v_cndmask_b32_e32 v1, v8, v1, vcc_lo
; GFX10-NEXT:    v_perm_b32 v1, v5, v1, 0x5040100
; GFX10-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-TRUE16-LABEL: v_maximumnum_v4bf16_no_ieee:
; GFX11-TRUE16:       ; %bb.0:
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xffff0000, v1
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xffff0000, v3
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xffff0000, v0
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff0000, v2
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 16, v1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s0, v5, v5
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 16, v3
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s1, v6, v6
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s2, v8, v8
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v4.l, v1.h, v3.h, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s3, v9, v9
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v6.l, v0.h, v2.h, s1
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v10, 16, v0
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v5.l, v3.h, v4.l, s0
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, v4.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v11, 16, v2
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v7.l, v2.h, v6.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s4, v10, v10
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v9.l, v5.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v1.l, v1.l, v3.l, s2
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v10.l, v6.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v12.l, v7.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s0, v11, v11
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v0.l, v0.l, v2.l, s4
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v3.l, v3.l, v1.l, s3
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
; GFX11-TRUE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v8, v9
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v11.l, v1.l
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v2.l, v2.l, v0.l, s0
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v13.l, v3.l
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v8.l, v5.l, v4.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v10, v12
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v9.l, v0.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v14.l, v2.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v10.l, v8.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 16, v13
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v8.l, v7.l, v6.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v13, 16, v14
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
; GFX11-TRUE16-NEXT:    v_cmp_gt_f32_e64 s0, v11, v12
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v11.l, v8.l
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v4.l
; GFX11-TRUE16-NEXT:    v_cmp_gt_f32_e64 s1, v9, v13
; GFX11-TRUE16-NEXT:    v_max_f32_e32 v10, v10, v10
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v8.l, v3.l, v1.l, s0
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s2, 0, v6.l
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v9.l, v2.l, v0.l, s1
; GFX11-TRUE16-NEXT:    v_bfe_u32 v12, v10, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v13, 0x400000, v10
; GFX11-TRUE16-NEXT:    v_max_f32_e32 v11, v11, v11
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s1, v10, v10
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; GFX11-TRUE16-NEXT:    v_add3_u32 v12, v12, v10, 0x7fff
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; GFX11-TRUE16-NEXT:    v_bfe_u32 v14, v11, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s0, 0, v5.l
; GFX11-TRUE16-NEXT:    v_max_f32_e32 v8, v8, v8
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e64 v10, v12, v13, s1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, 0x400000, v11
; GFX11-TRUE16-NEXT:    v_add3_u32 v13, v14, v11, 0x7fff
; GFX11-TRUE16-NEXT:    v_max_f32_e32 v9, v9, v9
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s1, v8, v8
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v0.h, v10.h, v4.l, vcc_lo
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xffff0000, v10
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v11, v11
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v0.h, v0.h, v5.l, s0
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s0, 0, v7.l
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v11, v13, v12, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v12, v8, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v4
; GFX11-TRUE16-NEXT:    v_bfe_u32 v4, v9, 16, 1
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v1.h, v11.h, v6.l, s2
; GFX11-TRUE16-NEXT:    v_add3_u32 v5, v12, v8, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, 0x400000, v8
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_add3_u32 v4, v4, v9, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, 0x400000, v9
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e64 s2, v9, v9
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v2.h, v1.h, v7.l, s0
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e64 v5, v5, v6, s1
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s1, 0, v1.l
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xffff0000, v11
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e64 v4, v4, v12, s2
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s2, 0, v0.l
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s0, 0, v3.l
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v1.l, v5.h, v1.l, s1
; GFX11-TRUE16-NEXT:    v_cmp_eq_u16_e64 s1, 0, v2.l
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff0000, v4
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v0.l, v4.h, v0.l, s2
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xffff0000, v5
; GFX11-TRUE16-NEXT:    v_cmp_eq_f32_e64 s2, 0, v6
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v1.l, v1.l, v3.l, s0
; GFX11-TRUE16-NEXT:    v_cmp_eq_f32_e64 s0, 0, v7
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v0.l, v0.l, v2.l, s1
; GFX11-TRUE16-NEXT:    v_cmp_eq_f32_e64 s1, 0, v8
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v1.h, v10.h, v0.h, vcc_lo
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v0.h, v11.h, v2.h, s2
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v0.l, v4.h, v0.l, s0
; GFX11-TRUE16-NEXT:    v_cndmask_b16 v1.l, v5.h, v1.l, s1
; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-FAKE16-LABEL: v_maximumnum_v4bf16_no_ieee:
; GFX11-FAKE16:       ; %bb.0:
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v8, 16, v2
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xffff0000, v1
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v5, 16, v3
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v6, 16, v1
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v9, 16, v0
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v12, 16, v3
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v4, v6, v5 :: v_dual_and_b32 v7, 0xffff0000, v3
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v6, 0xffff0000, v0
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v5, v5, v4, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v7, 0xffff0000, v2
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v10, 16, v5
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v6, v9, v8, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v7, v8, v6, vcc_lo
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v9, 16, v4
; GFX11-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v9, v10
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v11, 16, v1
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v9, 16, v0
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v8, v5, v4, vcc_lo
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v11, v11
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v13, 16, v7
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v1, v1, v3 :: v_dual_lshlrev_b32 v10, 16, v6
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_max_f32_e32 v8, v8, v8
; GFX11-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v10, v13
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v13, 16, v1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v11, 0x400000, v8
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v10, v7, v6, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
; GFX11-FAKE16-NEXT:    v_bfe_u32 v9, v8, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v0, v2, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v12, v12
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_add3_u32 v9, v9, v8, 0x7fff
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v3, v3, v1 :: v_dual_lshlrev_b32 v14, 16, v2
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v12, 16, v3
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v14, v14
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v2, v0, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v15, 16, v0
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v14, 16, v2
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v8, v9, v11, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v13, v12
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v13, 16, v8
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v9, v3, v1, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v15, v14
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v8, 0xffff0000, v8
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v12, v2, v0 :: v_dual_lshlrev_b32 v9, 16, v9
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v4
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_dual_max_f32 v9, v9, v9 :: v_dual_lshlrev_b32 v12, 16, v12
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v4, v13, v4, vcc_lo
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v15, 0x400000, v9
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_max_f32_e32 v12, v12, v12
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v16, 0x400000, v12
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_max_f32_e32 v10, v10, v10
; GFX11-FAKE16-NEXT:    v_bfe_u32 v11, v10, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v14, 0x400000, v10
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v10, v10
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_add3_u32 v11, v11, v10, 0x7fff
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v10, v11, v14, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v5
; GFX11-FAKE16-NEXT:    v_bfe_u32 v11, v9, 16, 1
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v14, 16, v10
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v4, v4, v5, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v6
; GFX11-FAKE16-NEXT:    v_bfe_u32 v5, v12, 16, 1
; GFX11-FAKE16-NEXT:    v_add3_u32 v11, v11, v9, 0x7fff
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xffff0000, v10
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v6, v14, v6, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
; GFX11-FAKE16-NEXT:    v_add3_u32 v5, v5, v12, 0x7fff
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v9, v11, v15, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v12, v12
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v5, v5, v16, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v8
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v8, 16, v9
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v4, v13, v4, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v7
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v6, v6, v7, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v1
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v7, 16, v5
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v8, v1, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v0
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v7, v0, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v3
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v2
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v0, v0, v2 :: v_dual_and_b32 v3, 0xffff0000, v5
; GFX11-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v10
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v5, v14, v6 :: v_dual_and_b32 v2, 0xffff0000, v9
; GFX11-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v3
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v7, v0, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v2
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_perm_b32 v0, v5, v0, 0x5040100
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v8, v1, vcc_lo
; GFX11-FAKE16-NEXT:    v_perm_b32 v1, v4, v1, 0x5040100
; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
;
; GFX12-TRUE16-LABEL: v_maximumnum_v4bf16_no_ieee:
; GFX12-TRUE16:       ; %bb.0:
; GFX12-TRUE16-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX12-TRUE16-NEXT:    s_wait_expcnt 0x0
; GFX12-TRUE16-NEXT:    s_wait_samplecnt 0x0
; GFX12-TRUE16-NEXT:    s_wait_bvhcnt 0x0
; GFX12-TRUE16-NEXT:    s_wait_kmcnt 0x0
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v4, 0xffff0000, v1
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v5, 0xffff0000, v3
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v6, 0xffff0000, v0
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff0000, v2
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 16, v1
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s0, v5, v5
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 16, v3
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s1, v6, v6
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s2, v8, v8
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v4.l, v1.h, v3.h, vcc_lo
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s3, v9, v9
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v6.l, v0.h, v2.h, s1
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v10, 16, v0
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v5.l, v3.h, v4.l, s0
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v8.l, v4.l
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v11, 16, v2
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v7.l, v2.h, v6.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s4, v10, v10
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v9.l, v5.l
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v1.l, v1.l, v3.l, s2
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v10.l, v6.l
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v12.l, v7.l
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s0, v11, v11
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v0.l, v0.l, v2.l, s4
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v3.l, v3.l, v1.l, s3
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
; GFX12-TRUE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v8, v9
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v11.l, v1.l
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v2.l, v2.l, v0.l, s0
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v13.l, v3.l
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v8.l, v5.l, v4.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v10, v12
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v9.l, v0.l
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v14.l, v2.l
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v10.l, v8.l
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 16, v13
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v8.l, v7.l, v6.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v13, 16, v14
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
; GFX12-TRUE16-NEXT:    v_cmp_gt_f32_e64 s0, v11, v12
; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v11.l, v8.l
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v4.l
; GFX12-TRUE16-NEXT:    v_cmp_gt_f32_e64 s1, v9, v13
; GFX12-TRUE16-NEXT:    v_max_num_f32_e32 v10, v10, v10
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v8.l, v3.l, v1.l, s0
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s2, 0, v6.l
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v9.l, v2.l, v0.l, s1
; GFX12-TRUE16-NEXT:    v_bfe_u32 v12, v10, 16, 1
; GFX12-TRUE16-NEXT:    v_or_b32_e32 v13, 0x400000, v10
; GFX12-TRUE16-NEXT:    v_max_num_f32_e32 v11, v11, v11
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s1, v10, v10
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; GFX12-TRUE16-NEXT:    v_add3_u32 v12, v12, v10, 0x7fff
; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; GFX12-TRUE16-NEXT:    v_bfe_u32 v14, v11, 16, 1
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s0, 0, v5.l
; GFX12-TRUE16-NEXT:    v_max_num_f32_e32 v8, v8, v8
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b32_e64 v10, v12, v13, s1
; GFX12-TRUE16-NEXT:    v_or_b32_e32 v12, 0x400000, v11
; GFX12-TRUE16-NEXT:    v_add3_u32 v13, v14, v11, 0x7fff
; GFX12-TRUE16-NEXT:    v_max_num_f32_e32 v9, v9, v9
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s1, v8, v8
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v0.h, v10.h, v4.l, vcc_lo
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v4, 0xffff0000, v10
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v11, v11
; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v0.h, v0.h, v5.l, s0
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s0, 0, v7.l
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b32_e32 v11, v13, v12, vcc_lo
; GFX12-TRUE16-NEXT:    v_bfe_u32 v12, v8, 16, 1
; GFX12-TRUE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v4
; GFX12-TRUE16-NEXT:    v_bfe_u32 v4, v9, 16, 1
; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v1.h, v11.h, v6.l, s2
; GFX12-TRUE16-NEXT:    v_add3_u32 v5, v12, v8, 0x7fff
; GFX12-TRUE16-NEXT:    v_or_b32_e32 v6, 0x400000, v8
; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX12-TRUE16-NEXT:    v_add3_u32 v4, v4, v9, 0x7fff
; GFX12-TRUE16-NEXT:    v_or_b32_e32 v12, 0x400000, v9
; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e64 s2, v9, v9
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v2.h, v1.h, v7.l, s0
; GFX12-TRUE16-NEXT:    v_cndmask_b32_e64 v5, v5, v6, s1
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s1, 0, v1.l
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v6, 0xffff0000, v11
; GFX12-TRUE16-NEXT:    v_cndmask_b32_e64 v4, v4, v12, s2
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s2, 0, v0.l
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s0, 0, v3.l
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v1.l, v5.h, v1.l, s1
; GFX12-TRUE16-NEXT:    v_cmp_eq_u16_e64 s1, 0, v2.l
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff0000, v4
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v0.l, v4.h, v0.l, s2
; GFX12-TRUE16-NEXT:    v_and_b32_e32 v8, 0xffff0000, v5
; GFX12-TRUE16-NEXT:    v_cmp_eq_f32_e64 s2, 0, v6
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v1.l, v1.l, v3.l, s0
; GFX12-TRUE16-NEXT:    v_cmp_eq_f32_e64 s0, 0, v7
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v0.l, v0.l, v2.l, s1
; GFX12-TRUE16-NEXT:    v_cmp_eq_f32_e64 s1, 0, v8
; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v1.h, v10.h, v0.h, vcc_lo
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v0.h, v11.h, v2.h, s2
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v0.l, v4.h, v0.l, s0
; GFX12-TRUE16-NEXT:    s_wait_alu 0xf1ff
; GFX12-TRUE16-NEXT:    v_cndmask_b16 v1.l, v5.h, v1.l, s1
; GFX12-TRUE16-NEXT:    s_setpc_b64 s[30:31]
;
; GFX12-FAKE16-LABEL: v_maximumnum_v4bf16_no_ieee:
; GFX12-FAKE16:       ; %bb.0:
; GFX12-FAKE16-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX12-FAKE16-NEXT:    s_wait_expcnt 0x0
; GFX12-FAKE16-NEXT:    s_wait_samplecnt 0x0
; GFX12-FAKE16-NEXT:    s_wait_bvhcnt 0x0
; GFX12-FAKE16-NEXT:    s_wait_kmcnt 0x0
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v8, 16, v2
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v4, 0xffff0000, v1
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v5, 16, v3
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v6, 16, v1
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v9, 16, v0
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v12, 16, v3
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v4, v6, v5 :: v_dual_and_b32 v7, 0xffff0000, v3
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v6, 0xffff0000, v0
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v5, v5, v4, vcc_lo
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v7, 0xffff0000, v2
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v10, 16, v5
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v6, v9, v8, vcc_lo
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v7, v8, v6, vcc_lo
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v9, 16, v4
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_4) | instid1(VALU_DEP_3)
; GFX12-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v9, v10
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v11, 16, v1
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v9, 16, v0
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v8, v5, v4, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v11, v11
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v13, 16, v7
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v1, v1, v3 :: v_dual_lshlrev_b32 v10, 16, v6
; GFX12-FAKE16-NEXT:    v_max_num_f32_e32 v8, v8, v8
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX12-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v10, v13
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v13, 16, v1
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX12-FAKE16-NEXT:    v_or_b32_e32 v11, 0x400000, v8
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v10, v7, v6, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
; GFX12-FAKE16-NEXT:    v_bfe_u32 v9, v8, 16, 1
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v0, v2, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v12, v12
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_1)
; GFX12-FAKE16-NEXT:    v_add3_u32 v9, v9, v8, 0x7fff
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v3, v3, v1 :: v_dual_lshlrev_b32 v14, 16, v2
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v12, 16, v3
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_4) | instid1(VALU_DEP_3)
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v14, v14
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v2, v0, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v15, 16, v0
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v14, 16, v2
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v8, v9, v11, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v13, v12
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v13, 16, v8
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v9, v3, v1, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v15, v14
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v8, 0xffff0000, v8
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v12, v2, v0 :: v_dual_lshlrev_b32 v9, 16, v9
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v4
; GFX12-FAKE16-NEXT:    v_dual_max_num_f32 v9, v9, v9 :: v_dual_lshlrev_b32 v12, 16, v12
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v4, v13, v4, vcc_lo
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX12-FAKE16-NEXT:    v_or_b32_e32 v15, 0x400000, v9
; GFX12-FAKE16-NEXT:    v_max_num_f32_e32 v12, v12, v12
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX12-FAKE16-NEXT:    v_or_b32_e32 v16, 0x400000, v12
; GFX12-FAKE16-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
; GFX12-FAKE16-NEXT:    v_max_num_f32_e32 v10, v10, v10
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX12-FAKE16-NEXT:    v_bfe_u32 v11, v10, 16, 1
; GFX12-FAKE16-NEXT:    v_or_b32_e32 v14, 0x400000, v10
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v10, v10
; GFX12-FAKE16-NEXT:    v_add3_u32 v11, v11, v10, 0x7fff
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v10, v11, v14, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v5
; GFX12-FAKE16-NEXT:    v_bfe_u32 v11, v9, 16, 1
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v14, 16, v10
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v4, v4, v5, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v6
; GFX12-FAKE16-NEXT:    v_bfe_u32 v5, v12, 16, 1
; GFX12-FAKE16-NEXT:    v_add3_u32 v11, v11, v9, 0x7fff
; GFX12-FAKE16-NEXT:    v_and_b32_e32 v10, 0xffff0000, v10
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v6, v14, v6, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
; GFX12-FAKE16-NEXT:    v_add3_u32 v5, v5, v12, 0x7fff
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v9, v11, v15, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v12, v12
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v5, v5, v16, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v8
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v8, 16, v9
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v4, v13, v4, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v7
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v6, v6, v7, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v1
; GFX12-FAKE16-NEXT:    v_lshrrev_b32_e32 v7, 16, v5
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v8, v1, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v0
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v7, v0, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v3
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v2
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v0, v0, v2 :: v_dual_and_b32 v3, 0xffff0000, v5
; GFX12-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v10
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_dual_cndmask_b32 v5, v14, v6 :: v_dual_and_b32 v2, 0xffff0000, v9
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX12-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v3
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v7, v0, vcc_lo
; GFX12-FAKE16-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v2
; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1)
; GFX12-FAKE16-NEXT:    v_perm_b32 v0, v5, v0, 0x5040100
; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffd
; GFX12-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v8, v1, vcc_lo
; GFX12-FAKE16-NEXT:    v_perm_b32 v1, v4, v1, 0x5040100
; GFX12-FAKE16-NEXT:    s_setpc_b64 s[30:31]
  %result = call <4 x bfloat> @llvm.maximumnum.v4bf16(<4 x bfloat> %x, <4 x bfloat> %y)
  ret <4 x bfloat> %result
}

attributes #0 = { "amdgpu-ieee"="false" }
;; NOTE: These prefixes are unused and the list is autogenerated. Do not add tests below this line:
; GFX11: {{.*}}
; GFX12: {{.*}}
; GFX9: {{.*}}