llvm/test/CodeGen/AMDGPU/fcanonicalize.bf16.ll

; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
; RUN: llc -mtriple=amdgcn -mcpu=gfx1250  < %s | FileCheck -enable-var-scope -check-prefixes=GFX1250 %s

declare bfloat @llvm.fabs.bf16(bfloat) #0
declare bfloat @llvm.canonicalize.bf16(bfloat) #0
declare <2 x bfloat> @llvm.fabs.v2bf16(<2 x bfloat>) #0
declare <2 x bfloat> @llvm.canonicalize.v2bf16(<2 x bfloat>) #0
declare <3 x bfloat> @llvm.canonicalize.v3bf16(<3 x bfloat>) #0
declare <4 x bfloat> @llvm.canonicalize.v4bf16(<4 x bfloat>) #0
declare <6 x bfloat> @llvm.canonicalize.v6bf16(<6 x bfloat>) #0
declare <8 x bfloat> @llvm.canonicalize.v8bf16(<8 x bfloat>) #0
declare <12 x bfloat> @llvm.canonicalize.v12bf16(<12 x bfloat>) #0
declare <16 x bfloat> @llvm.canonicalize.v16bf16(<16 x bfloat>) #0
declare <32 x bfloat> @llvm.canonicalize.v32bf16(<32 x bfloat>) #0
declare <64 x bfloat> @llvm.canonicalize.v64bf16(<64 x bfloat>) #0
declare i32 @llvm.amdgcn.workitem.id.x() #0

define amdgpu_kernel void @test_fold_canonicalize_undef_value_bf16(ptr addrspace(1) %out) #1 {
; GFX1250-LABEL: test_fold_canonicalize_undef_value_bf16:
; GFX1250:       ; %bb.0:
; GFX1250-NEXT:    s_load_b64 s[0:1], s[4:5], 0x24
; GFX1250-NEXT:    v_mov_b32_e32 v0, 0
; GFX1250-NEXT:    s_wait_kmcnt 0x0
; GFX1250-NEXT:    global_store_b16 v0, v0, s[0:1]
; GFX1250-NEXT:    s_endpgm
  %canonicalized = call bfloat @llvm.canonicalize.bf16(bfloat undef)
  store bfloat %canonicalized, ptr addrspace(1) %out
  ret void
}

define amdgpu_kernel void @v_test_canonicalize_var_bf16(ptr addrspace(1) %out) #1 {
; GFX1250-LABEL: v_test_canonicalize_var_bf16:
; GFX1250:       ; %bb.0:
; GFX1250-NEXT:    s_load_b64 s[0:1], s[4:5], 0x24
; GFX1250-NEXT:    v_mov_b32_e32 v0, 0
; GFX1250-NEXT:    s_wait_kmcnt 0x0
; GFX1250-NEXT:    global_load_u16 v0, v0, s[0:1]
; GFX1250-NEXT:    s_wait_loadcnt 0x0
; GFX1250-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
; GFX1250-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX1250-NEXT:    v_max_num_f32_e32 v0, v0, v0
; GFX1250-NEXT:    v_cvt_pk_bf16_f32 v0, v0, s0
; GFX1250-NEXT:    global_store_b16 v[0:1], v0, off
; GFX1250-NEXT:    s_endpgm
  %val = load bfloat, ptr addrspace(1) %out
  %canonicalized = call bfloat @llvm.canonicalize.bf16(bfloat %val)
  store bfloat %canonicalized, ptr addrspace(1) poison
  ret void
}

define amdgpu_kernel void @s_test_canonicalize_var_bf16(ptr addrspace(1) %out, i16 zeroext %val.arg) #1 {
; GFX1250-LABEL: s_test_canonicalize_var_bf16:
; GFX1250:       ; %bb.0:
; GFX1250-NEXT:    s_load_b96 s[0:2], s[4:5], 0x24
; GFX1250-NEXT:    v_mov_b32_e32 v1, 0
; GFX1250-NEXT:    s_wait_kmcnt 0x0
; GFX1250-NEXT:    s_lshl_b32 s2, s2, 16
; GFX1250-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX1250-NEXT:    v_max_num_f32_e64 v0, s2, s2
; GFX1250-NEXT:    v_cvt_pk_bf16_f32 v0, v0, s0
; GFX1250-NEXT:    global_store_b16 v1, v0, s[0:1]
; GFX1250-NEXT:    s_endpgm
  %val = bitcast i16 %val.arg to bfloat
  %canonicalized = call bfloat @llvm.canonicalize.bf16(bfloat %val)
  store bfloat %canonicalized, ptr addrspace(1) %out
  ret void
}

define <2 x bfloat> @v_test_canonicalize_build_vector_v2bf16(bfloat %lo, bfloat %hi) #1 {
; GFX1250-LABEL: v_test_canonicalize_build_vector_v2bf16:
; GFX1250:       ; %bb.0:
; GFX1250-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX1250-NEXT:    s_wait_kmcnt 0x0
; GFX1250-NEXT:    v_dual_lshlrev_b32 v1, 16, v1 :: v_dual_lshlrev_b32 v0, 16, v0
; GFX1250-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX1250-NEXT:    v_dual_max_num_f32 v1, v1, v1 :: v_dual_max_num_f32 v0, v0, v0
; GFX1250-NEXT:    v_cvt_pk_bf16_f32 v0, v0, v1
; GFX1250-NEXT:    s_set_pc_i64 s[30:31]
  %ins0 = insertelement <2 x bfloat> poison, bfloat %lo, i32 0
  %ins1 = insertelement <2 x bfloat> %ins0, bfloat %hi, i32 1
  %canonicalized = call <2 x bfloat> @llvm.canonicalize.v2bf16(<2 x bfloat> %ins1)
  ret <2 x bfloat> %canonicalized
}


define amdgpu_kernel void @v_test_canonicalize_fabs_var_bf16(ptr addrspace(1) %out) #1 {
; GFX1250-LABEL: v_test_canonicalize_fabs_var_bf16:
; GFX1250:       ; %bb.0:
; GFX1250-NEXT:    s_load_b64 s[0:1], s[4:5], 0x24
; GFX1250-NEXT:    v_mov_b32_e32 v0, 0
; GFX1250-NEXT:    s_wait_kmcnt 0x0
; GFX1250-NEXT:    global_load_u16 v1, v0, s[0:1]
; GFX1250-NEXT:    s_wait_loadcnt 0x0
; GFX1250-NEXT:    v_and_b32_e32 v1, 0x7fff, v1
; GFX1250-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX1250-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; GFX1250-NEXT:    v_max_num_f32_e32 v1, v1, v1
; GFX1250-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX1250-NEXT:    v_cvt_pk_bf16_f32 v1, v1, s0
; GFX1250-NEXT:    global_store_b16 v0, v1, s[0:1]
; GFX1250-NEXT:    s_endpgm
  %val = load bfloat, ptr addrspace(1) %out
  %val.fabs = call bfloat @llvm.fabs.bf16(bfloat %val)
  %canonicalized = call bfloat @llvm.canonicalize.bf16(bfloat %val.fabs)
  store bfloat %canonicalized, ptr addrspace(1) %out
  ret void
}


define amdgpu_kernel void @v_test_canonicalize_fneg_fabs_var_bf16(ptr addrspace(1) %out) #1 {
; GFX1250-LABEL: v_test_canonicalize_fneg_fabs_var_bf16:
; GFX1250:       ; %bb.0:
; GFX1250-NEXT:    s_load_b64 s[0:1], s[4:5], 0x24
; GFX1250-NEXT:    v_mov_b32_e32 v0, 0
; GFX1250-NEXT:    s_wait_kmcnt 0x0
; GFX1250-NEXT:    global_load_u16 v1, v0, s[0:1]
; GFX1250-NEXT:    s_wait_loadcnt 0x0
; GFX1250-NEXT:    v_or_b32_e32 v1, 0x8000, v1
; GFX1250-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX1250-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; GFX1250-NEXT:    v_max_num_f32_e32 v1, v1, v1
; GFX1250-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX1250-NEXT:    v_cvt_pk_bf16_f32 v1, v1, s0
; GFX1250-NEXT:    global_store_b16 v0, v1, s[0:1]
; GFX1250-NEXT:    s_endpgm
  %val = load bfloat, ptr addrspace(1) %out
  %val.fabs = call bfloat @llvm.fabs.bf16(bfloat %val)
  %val.fabs.fneg = fneg bfloat %val.fabs
  %canonicalized = call bfloat @llvm.canonicalize.bf16(bfloat %val.fabs.fneg)
  store bfloat %canonicalized, ptr addrspace(1) %out
  ret void
}

define amdgpu_kernel void @v_test_canonicalize_fneg_var_bf16(ptr addrspace(1) %out) #1 {
; GFX1250-LABEL: v_test_canonicalize_fneg_var_bf16:
; GFX1250:       ; %bb.0:
; GFX1250-NEXT:    s_load_b64 s[0:1], s[4:5], 0x24
; GFX1250-NEXT:    v_mov_b32_e32 v0, 0
; GFX1250-NEXT:    s_wait_kmcnt 0x0
; GFX1250-NEXT:    global_load_u16 v1, v0, s[0:1]
; GFX1250-NEXT:    s_wait_loadcnt 0x0
; GFX1250-NEXT:    v_xor_b32_e32 v1, 0x8000, v1
; GFX1250-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX1250-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; GFX1250-NEXT:    v_max_num_f32_e32 v1, v1, v1
; GFX1250-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX1250-NEXT:    v_cvt_pk_bf16_f32 v1, v1, s0
; GFX1250-NEXT:    global_store_b16 v0, v1, s[0:1]
; GFX1250-NEXT:    s_endpgm
  %val = load bfloat, ptr addrspace(1) %out
  %val.fneg = fneg bfloat %val
  %canonicalized = call bfloat @llvm.canonicalize.bf16(bfloat %val.fneg)
  store bfloat %canonicalized, ptr addrspace(1) %out
  ret void
}

define amdgpu_kernel void @v_test_no_denormals_canonicalize_fneg_var_bf16(ptr addrspace(1) %out) #2 {
; GFX1250-LABEL: v_test_no_denormals_canonicalize_fneg_var_bf16:
; GFX1250:       ; %bb.0:
; GFX1250-NEXT:    s_load_b64 s[0:1], s[4:5], 0x24
; GFX1250-NEXT:    v_mov_b32_e32 v0, 0
; GFX1250-NEXT:    s_wait_kmcnt 0x0
; GFX1250-NEXT:    global_load_u16 v1, v0, s[0:1]
; GFX1250-NEXT:    s_wait_loadcnt 0x0
; GFX1250-NEXT:    v_xor_b32_e32 v1, 0x8000, v1
; GFX1250-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX1250-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; GFX1250-NEXT:    v_max_num_f32_e32 v1, v1, v1
; GFX1250-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX1250-NEXT:    v_cvt_pk_bf16_f32 v1, v1, s0
; GFX1250-NEXT:    global_store_b16 v0, v1, s[0:1]
; GFX1250-NEXT:    s_endpgm
  %val = load bfloat, ptr addrspace(1) %out
  %val.fneg = fneg bfloat %val
  %canonicalized = call bfloat @llvm.canonicalize.bf16(bfloat %val.fneg)
  store bfloat %canonicalized, ptr addrspace(1) %out
  ret void
}

define amdgpu_kernel void @v_test_no_denormals_canonicalize_fneg_fabs_var_bf16(ptr addrspace(1) %out) #2 {
; GFX1250-LABEL: v_test_no_denormals_canonicalize_fneg_fabs_var_bf16:
; GFX1250:       ; %bb.0:
; GFX1250-NEXT:    s_load_b64 s[0:1], s[4:5], 0x24
; GFX1250-NEXT:    v_mov_b32_e32 v0, 0
; GFX1250-NEXT:    s_wait_kmcnt 0x0
; GFX1250-NEXT:    global_load_u16 v1, v0, s[0:1]
; GFX1250-NEXT:    s_wait_loadcnt 0x0
; GFX1250-NEXT:    v_or_b32_e32 v1, 0x8000, v1
; GFX1250-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX1250-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; GFX1250-NEXT:    v_max_num_f32_e32 v1, v1, v1
; GFX1250-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX1250-NEXT:    v_cvt_pk_bf16_f32 v1, v1, s0
; GFX1250-NEXT:    global_store_b16 v0, v1, s[0:1]
; GFX1250-NEXT:    s_endpgm
  %val = load bfloat, ptr addrspace(1) %out
  %val.fabs = call bfloat @llvm.fabs.bf16(bfloat %val)
  %val.fabs.fneg = fneg bfloat %val.fabs
  %canonicalized = call bfloat @llvm.canonicalize.bf16(bfloat %val.fabs.fneg)
  store bfloat %canonicalized, ptr addrspace(1) %out
  ret void
}

define amdgpu_kernel void @test_fold_canonicalize_p0_bf16(ptr addrspace(1) %out) #1 {
; GFX1250-LABEL: test_fold_canonicalize_p0_bf16:
; GFX1250:       ; %bb.0:
; GFX1250-NEXT:    s_load_b64 s[0:1], s[4:5], 0x24
; GFX1250-NEXT:    v_mov_b32_e32 v0, 0
; GFX1250-NEXT:    s_wait_kmcnt 0x0
; GFX1250-NEXT:    global_store_b16 v0, v0, s[0:1]
; GFX1250-NEXT:    s_endpgm
   %canonicalized = call bfloat @llvm.canonicalize.bf16(bfloat 0.0)
   store bfloat %canonicalized, ptr addrspace(1) %out
   ret void
}

define amdgpu_kernel void @test_fold_canonicalize_n0_bf16(ptr addrspace(1) %out) #1 {
; GFX1250-LABEL: test_fold_canonicalize_n0_bf16:
; GFX1250:       ; %bb.0:
; GFX1250-NEXT:    s_load_b64 s[0:1], s[4:5], 0x24
; GFX1250-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, 0xffff8000
; GFX1250-NEXT:    s_wait_kmcnt 0x0
; GFX1250-NEXT:    global_store_b16 v0, v1, s[0:1]
; GFX1250-NEXT:    s_endpgm
  %canonicalized = call bfloat @llvm.canonicalize.bf16(bfloat -0.0)
  store bfloat %canonicalized, ptr addrspace(1) %out
  ret void
}

define amdgpu_kernel void @test_fold_canonicalize_p1_bf16(ptr addrspace(1) %out) #1 {
; GFX1250-LABEL: test_fold_canonicalize_p1_bf16:
; GFX1250:       ; %bb.0:
; GFX1250-NEXT:    s_load_b64 s[0:1], s[4:5], 0x24
; GFX1250-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, 0x3f80
; GFX1250-NEXT:    s_wait_kmcnt 0x0
; GFX1250-NEXT:    global_store_b16 v0, v1, s[0:1]
; GFX1250-NEXT:    s_endpgm
  %canonicalized = call bfloat @llvm.canonicalize.bf16(bfloat 1.0)
  store bfloat %canonicalized, ptr addrspace(1) %out
  ret void
}

define amdgpu_kernel void @test_fold_canonicalize_n1_bf16(ptr addrspace(1) %out) #1 {
; GFX1250-LABEL: test_fold_canonicalize_n1_bf16:
; GFX1250:       ; %bb.0:
; GFX1250-NEXT:    s_load_b64 s[0:1], s[4:5], 0x24
; GFX1250-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, 0xffffbf80
; GFX1250-NEXT:    s_wait_kmcnt 0x0
; GFX1250-NEXT:    global_store_b16 v0, v1, s[0:1]
; GFX1250-NEXT:    s_endpgm
  %canonicalized = call bfloat @llvm.canonicalize.bf16(bfloat -1.0)
  store bfloat %canonicalized, ptr addrspace(1) %out
  ret void
}

define amdgpu_kernel void @test_fold_canonicalize_literal_bf16(ptr addrspace(1) %out) #1 {
; GFX1250-LABEL: test_fold_canonicalize_literal_bf16:
; GFX1250:       ; %bb.0:
; GFX1250-NEXT:    s_load_b64 s[0:1], s[4:5], 0x24
; GFX1250-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, 0x4180
; GFX1250-NEXT:    s_wait_kmcnt 0x0
; GFX1250-NEXT:    global_store_b16 v0, v1, s[0:1]
; GFX1250-NEXT:    s_endpgm
  %canonicalized = call bfloat @llvm.canonicalize.bf16(bfloat 16.0)
  store bfloat %canonicalized, ptr addrspace(1) %out
  ret void
}

define amdgpu_kernel void @test_default_denormals_fold_canonicalize_denormal0_bf16(ptr addrspace(1) %out) #1 {
; GFX1250-LABEL: test_default_denormals_fold_canonicalize_denormal0_bf16:
; GFX1250:       ; %bb.0:
; GFX1250-NEXT:    s_load_b64 s[0:1], s[4:5], 0x24
; GFX1250-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, 0x3ff
; GFX1250-NEXT:    s_wait_kmcnt 0x0
; GFX1250-NEXT:    global_store_b16 v0, v1, s[0:1]
; GFX1250-NEXT:    s_endpgm
  %canonicalized = call bfloat @llvm.canonicalize.bf16(bfloat 0xR03FF)
  store bfloat %canonicalized, ptr addrspace(1) %out
  ret void
}

define amdgpu_kernel void @test_denormals_fold_canonicalize_denormal0_bf16(ptr addrspace(1) %out) #3 {
; GFX1250-LABEL: test_denormals_fold_canonicalize_denormal0_bf16:
; GFX1250:       ; %bb.0:
; GFX1250-NEXT:    s_load_b64 s[0:1], s[4:5], 0x24
; GFX1250-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, 0x3ff
; GFX1250-NEXT:    s_wait_kmcnt 0x0
; GFX1250-NEXT:    global_store_b16 v0, v1, s[0:1]
; GFX1250-NEXT:    s_endpgm
  %canonicalized = call bfloat @llvm.canonicalize.bf16(bfloat 0xR03FF)
  store bfloat %canonicalized, ptr addrspace(1) %out
  ret void
}

define amdgpu_kernel void @test_default_denormals_fold_canonicalize_denormal1_bf16(ptr addrspace(1) %out) #1 {
; GFX1250-LABEL: test_default_denormals_fold_canonicalize_denormal1_bf16:
; GFX1250:       ; %bb.0:
; GFX1250-NEXT:    s_load_b64 s[0:1], s[4:5], 0x24
; GFX1250-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, 0xffff83ff
; GFX1250-NEXT:    s_wait_kmcnt 0x0
; GFX1250-NEXT:    global_store_b16 v0, v1, s[0:1]
; GFX1250-NEXT:    s_endpgm
  %canonicalized = call bfloat @llvm.canonicalize.bf16(bfloat 0xR83FF)
  store bfloat %canonicalized, ptr addrspace(1) %out
  ret void
}

define amdgpu_kernel void @test_denormals_fold_canonicalize_denormal1_bf16(ptr addrspace(1) %out) #3 {
; GFX1250-LABEL: test_denormals_fold_canonicalize_denormal1_bf16:
; GFX1250:       ; %bb.0:
; GFX1250-NEXT:    s_load_b64 s[0:1], s[4:5], 0x24
; GFX1250-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, 0xffff83ff
; GFX1250-NEXT:    s_wait_kmcnt 0x0
; GFX1250-NEXT:    global_store_b16 v0, v1, s[0:1]
; GFX1250-NEXT:    s_endpgm
  %canonicalized = call bfloat @llvm.canonicalize.bf16(bfloat 0xR83FF)
  store bfloat %canonicalized, ptr addrspace(1) %out
  ret void
}

define amdgpu_kernel void @test_fold_canonicalize_qnan_bf16(ptr addrspace(1) %out) #1 {
; GFX1250-LABEL: test_fold_canonicalize_qnan_bf16:
; GFX1250:       ; %bb.0:
; GFX1250-NEXT:    s_load_b64 s[0:1], s[4:5], 0x24
; GFX1250-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, 0x7c00
; GFX1250-NEXT:    s_wait_kmcnt 0x0
; GFX1250-NEXT:    global_store_b16 v0, v1, s[0:1]
; GFX1250-NEXT:    s_endpgm
  %canonicalized = call bfloat @llvm.canonicalize.bf16(bfloat 0xR7C00)
  store bfloat %canonicalized, ptr addrspace(1) %out
  ret void
}

define amdgpu_kernel void @test_fold_canonicalize_qnan_value_neg1_bf16(ptr addrspace(1) %out) #1 {
; GFX1250-LABEL: test_fold_canonicalize_qnan_value_neg1_bf16:
; GFX1250:       ; %bb.0:
; GFX1250-NEXT:    s_load_b64 s[0:1], s[4:5], 0x24
; GFX1250-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, 0x7fc0
; GFX1250-NEXT:    s_wait_kmcnt 0x0
; GFX1250-NEXT:    global_store_b16 v0, v1, s[0:1]
; GFX1250-NEXT:    s_endpgm
  %canonicalized = call bfloat @llvm.canonicalize.bf16(bfloat bitcast (i16 -1 to bfloat))
  store bfloat %canonicalized, ptr addrspace(1) %out
  ret void
}

define amdgpu_kernel void @test_fold_canonicalize_qnan_value_neg2_bf16(ptr addrspace(1) %out) #1 {
; GFX1250-LABEL: test_fold_canonicalize_qnan_value_neg2_bf16:
; GFX1250:       ; %bb.0:
; GFX1250-NEXT:    s_load_b64 s[0:1], s[4:5], 0x24
; GFX1250-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, 0x7fc0
; GFX1250-NEXT:    s_wait_kmcnt 0x0
; GFX1250-NEXT:    global_store_b16 v0, v1, s[0:1]
; GFX1250-NEXT:    s_endpgm
  %canonicalized = call bfloat @llvm.canonicalize.bf16(bfloat bitcast (i16 -2 to bfloat))
  store bfloat %canonicalized, ptr addrspace(1) %out
  ret void
}

define amdgpu_kernel void @test_fold_canonicalize_snan0_value_bf16(ptr addrspace(1) %out) #1 {
; GFX1250-LABEL: test_fold_canonicalize_snan0_value_bf16:
; GFX1250:       ; %bb.0:
; GFX1250-NEXT:    s_load_b64 s[0:1], s[4:5], 0x24
; GFX1250-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, 0x7c01
; GFX1250-NEXT:    s_wait_kmcnt 0x0
; GFX1250-NEXT:    global_store_b16 v0, v1, s[0:1]
; GFX1250-NEXT:    s_endpgm
  %canonicalized = call bfloat @llvm.canonicalize.bf16(bfloat 0xR7C01)
  store bfloat %canonicalized, ptr addrspace(1) %out
  ret void
}

define amdgpu_kernel void @test_fold_canonicalize_snan1_value_bf16(ptr addrspace(1) %out) #1 {
; GFX1250-LABEL: test_fold_canonicalize_snan1_value_bf16:
; GFX1250:       ; %bb.0:
; GFX1250-NEXT:    s_load_b64 s[0:1], s[4:5], 0x24
; GFX1250-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, 0x7dff
; GFX1250-NEXT:    s_wait_kmcnt 0x0
; GFX1250-NEXT:    global_store_b16 v0, v1, s[0:1]
; GFX1250-NEXT:    s_endpgm
  %canonicalized = call bfloat @llvm.canonicalize.bf16(bfloat 0xR7DFF)
  store bfloat %canonicalized, ptr addrspace(1) %out
  ret void
}

define amdgpu_kernel void @test_fold_canonicalize_snan2_value_bf16(ptr addrspace(1) %out) #1 {
; GFX1250-LABEL: test_fold_canonicalize_snan2_value_bf16:
; GFX1250:       ; %bb.0:
; GFX1250-NEXT:    s_load_b64 s[0:1], s[4:5], 0x24
; GFX1250-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, 0xfffffdff
; GFX1250-NEXT:    s_wait_kmcnt 0x0
; GFX1250-NEXT:    global_store_b16 v0, v1, s[0:1]
; GFX1250-NEXT:    s_endpgm
  %canonicalized = call bfloat @llvm.canonicalize.bf16(bfloat 0xRFDFF)
  store bfloat %canonicalized, ptr addrspace(1) %out
  ret void
}

define amdgpu_kernel void @test_fold_canonicalize_snan3_value_bf16(ptr addrspace(1) %out) #1 {
; GFX1250-LABEL: test_fold_canonicalize_snan3_value_bf16:
; GFX1250:       ; %bb.0:
; GFX1250-NEXT:    s_load_b64 s[0:1], s[4:5], 0x24
; GFX1250-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, 0xfffffc01
; GFX1250-NEXT:    s_wait_kmcnt 0x0
; GFX1250-NEXT:    global_store_b16 v0, v1, s[0:1]
; GFX1250-NEXT:    s_endpgm
  %canonicalized = call bfloat @llvm.canonicalize.bf16(bfloat 0xRFC01)
  store bfloat %canonicalized, ptr addrspace(1) %out
  ret void
}

define amdgpu_kernel void @v_test_canonicalize_var_v2bf16(ptr addrspace(1) %out) #1 {
; GFX1250-LABEL: v_test_canonicalize_var_v2bf16:
; GFX1250:       ; %bb.0:
; GFX1250-NEXT:    s_load_b64 s[0:1], s[4:5], 0x24
; GFX1250-NEXT:    v_and_b32_e32 v0, 0x3ff, v0
; GFX1250-NEXT:    v_mov_b32_e32 v2, 0
; GFX1250-NEXT:    s_wait_kmcnt 0x0
; GFX1250-NEXT:    global_load_b32 v0, v0, s[0:1] scale_offset
; GFX1250-NEXT:    s_wait_loadcnt 0x0
; GFX1250-NEXT:    v_and_b32_e32 v1, 0xffff0000, v0
; GFX1250-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX1250-NEXT:    v_dual_max_num_f32 v1, v1, v1 :: v_dual_lshlrev_b32 v0, 16, v0
; GFX1250-NEXT:    v_max_num_f32_e32 v0, v0, v0
; GFX1250-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX1250-NEXT:    v_cvt_pk_bf16_f32 v0, v0, v1
; GFX1250-NEXT:    global_store_b32 v2, v0, s[0:1]
; GFX1250-NEXT:    s_endpgm
  %tid = call i32 @llvm.amdgcn.workitem.id.x()
  %gep = getelementptr <2 x bfloat>, ptr addrspace(1) %out, i32 %tid
  %val = load <2 x bfloat>, ptr addrspace(1) %gep
  %canonicalized = call <2 x bfloat> @llvm.canonicalize.v2bf16(<2 x bfloat> %val)
  store <2 x bfloat> %canonicalized, ptr addrspace(1) %out
  ret void
}

define amdgpu_kernel void @v_test_canonicalize_fabs_var_v2bf16(ptr addrspace(1) %out) #1 {
; GFX1250-LABEL: v_test_canonicalize_fabs_var_v2bf16:
; GFX1250:       ; %bb.0:
; GFX1250-NEXT:    s_load_b64 s[0:1], s[4:5], 0x24
; GFX1250-NEXT:    v_and_b32_e32 v0, 0x3ff, v0
; GFX1250-NEXT:    v_mov_b32_e32 v2, 0
; GFX1250-NEXT:    s_wait_kmcnt 0x0
; GFX1250-NEXT:    global_load_b32 v0, v0, s[0:1] scale_offset
; GFX1250-NEXT:    s_wait_loadcnt 0x0
; GFX1250-NEXT:    v_lshrrev_b32_e32 v1, 16, v0
; GFX1250-NEXT:    v_and_b32_e32 v0, 0x7fff, v0
; GFX1250-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX1250-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
; GFX1250-NEXT:    v_and_b32_e32 v1, 0x7fff, v1
; GFX1250-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX1250-NEXT:    v_dual_max_num_f32 v0, v0, v0 :: v_dual_lshlrev_b32 v1, 16, v1
; GFX1250-NEXT:    v_max_num_f32_e32 v1, v1, v1
; GFX1250-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX1250-NEXT:    v_cvt_pk_bf16_f32 v0, v0, v1
; GFX1250-NEXT:    global_store_b32 v2, v0, s[0:1]
; GFX1250-NEXT:    s_endpgm
  %tid = call i32 @llvm.amdgcn.workitem.id.x()
  %gep = getelementptr <2 x bfloat>, ptr addrspace(1) %out, i32 %tid
  %val = load <2 x bfloat>, ptr addrspace(1) %gep
  %val.fabs = call <2 x bfloat> @llvm.fabs.v2bf16(<2 x bfloat> %val)
  %canonicalized = call <2 x bfloat> @llvm.canonicalize.v2bf16(<2 x bfloat> %val.fabs)
  store <2 x bfloat> %canonicalized, ptr addrspace(1) %out
  ret void
}

define amdgpu_kernel void @v_test_canonicalize_fneg_fabs_var_v2bf16(ptr addrspace(1) %out) #1 {
; GFX1250-LABEL: v_test_canonicalize_fneg_fabs_var_v2bf16:
; GFX1250:       ; %bb.0:
; GFX1250-NEXT:    s_load_b64 s[0:1], s[4:5], 0x24
; GFX1250-NEXT:    v_and_b32_e32 v0, 0x3ff, v0
; GFX1250-NEXT:    v_mov_b32_e32 v2, 0
; GFX1250-NEXT:    s_wait_kmcnt 0x0
; GFX1250-NEXT:    global_load_b32 v0, v0, s[0:1] scale_offset
; GFX1250-NEXT:    s_wait_loadcnt 0x0
; GFX1250-NEXT:    v_lshrrev_b32_e32 v1, 16, v0
; GFX1250-NEXT:    v_or_b32_e32 v0, 0x8000, v0
; GFX1250-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX1250-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
; GFX1250-NEXT:    v_or_b32_e32 v1, 0x8000, v1
; GFX1250-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX1250-NEXT:    v_dual_max_num_f32 v0, v0, v0 :: v_dual_lshlrev_b32 v1, 16, v1
; GFX1250-NEXT:    v_max_num_f32_e32 v1, v1, v1
; GFX1250-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX1250-NEXT:    v_cvt_pk_bf16_f32 v0, v0, v1
; GFX1250-NEXT:    global_store_b32 v2, v0, s[0:1]
; GFX1250-NEXT:    s_endpgm
  %tid = call i32 @llvm.amdgcn.workitem.id.x()
  %gep = getelementptr <2 x bfloat>, ptr addrspace(1) %out, i32 %tid
  %val = load <2 x bfloat>, ptr addrspace(1) %gep
  %val.fabs = call <2 x bfloat> @llvm.fabs.v2bf16(<2 x bfloat> %val)
  %val.fabs.fneg = fneg <2 x bfloat> %val.fabs
  %canonicalized = call <2 x bfloat> @llvm.canonicalize.v2bf16(<2 x bfloat> %val.fabs.fneg)
  store <2 x bfloat> %canonicalized, ptr addrspace(1) %out
  ret void
}

define amdgpu_kernel void @v_test_canonicalize_fneg_var_v2bf16(ptr addrspace(1) %out) #1 {
; GFX1250-LABEL: v_test_canonicalize_fneg_var_v2bf16:
; GFX1250:       ; %bb.0:
; GFX1250-NEXT:    s_load_b64 s[0:1], s[4:5], 0x24
; GFX1250-NEXT:    v_and_b32_e32 v0, 0x3ff, v0
; GFX1250-NEXT:    v_mov_b32_e32 v2, 0
; GFX1250-NEXT:    s_wait_kmcnt 0x0
; GFX1250-NEXT:    global_load_b32 v0, v0, s[0:1] scale_offset
; GFX1250-NEXT:    s_wait_loadcnt 0x0
; GFX1250-NEXT:    v_lshrrev_b32_e32 v1, 16, v0
; GFX1250-NEXT:    v_xor_b32_e32 v0, 0x8000, v0
; GFX1250-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX1250-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
; GFX1250-NEXT:    v_xor_b32_e32 v1, 0x8000, v1
; GFX1250-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX1250-NEXT:    v_dual_max_num_f32 v0, v0, v0 :: v_dual_lshlrev_b32 v1, 16, v1
; GFX1250-NEXT:    v_max_num_f32_e32 v1, v1, v1
; GFX1250-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX1250-NEXT:    v_cvt_pk_bf16_f32 v0, v0, v1
; GFX1250-NEXT:    global_store_b32 v2, v0, s[0:1]
; GFX1250-NEXT:    s_endpgm
  %tid = call i32 @llvm.amdgcn.workitem.id.x()
  %gep = getelementptr <2 x bfloat>, ptr addrspace(1) %out, i32 %tid
  %val = load <2 x bfloat>, ptr addrspace(1) %gep
  %fneg.val = fneg <2 x bfloat> %val
  %canonicalized = call <2 x bfloat> @llvm.canonicalize.v2bf16(<2 x bfloat> %fneg.val)
  store <2 x bfloat> %canonicalized, ptr addrspace(1) %out
  ret void
}

define amdgpu_kernel void @s_test_canonicalize_var_v2bf16(ptr addrspace(1) %out, i32 zeroext %val.arg) #1 {
; GFX1250-LABEL: s_test_canonicalize_var_v2bf16:
; GFX1250:       ; %bb.0:
; GFX1250-NEXT:    s_load_b96 s[0:2], s[4:5], 0x24
; GFX1250-NEXT:    v_mov_b32_e32 v2, 0
; GFX1250-NEXT:    s_wait_kmcnt 0x0
; GFX1250-NEXT:    s_and_b32 s3, s2, 0xffff0000
; GFX1250-NEXT:    s_lshl_b32 s2, s2, 16
; GFX1250-NEXT:    v_max_num_f32_e64 v0, s3, s3
; GFX1250-NEXT:    v_max_num_f32_e64 v1, s2, s2
; GFX1250-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX1250-NEXT:    v_cvt_pk_bf16_f32 v0, v1, v0
; GFX1250-NEXT:    global_store_b32 v2, v0, s[0:1]
; GFX1250-NEXT:    s_endpgm
  %val = bitcast i32 %val.arg to <2 x bfloat>
  %canonicalized = call <2 x bfloat> @llvm.canonicalize.v2bf16(<2 x bfloat> %val)
  store <2 x bfloat> %canonicalized, ptr addrspace(1) %out
  ret void
}

define amdgpu_kernel void @test_fold_canonicalize_p0_v2bf16(ptr addrspace(1) %out) #1 {
; GFX1250-LABEL: test_fold_canonicalize_p0_v2bf16:
; GFX1250:       ; %bb.0:
; GFX1250-NEXT:    s_load_b64 s[0:1], s[4:5], 0x24
; GFX1250-NEXT:    v_mov_b32_e32 v0, 0
; GFX1250-NEXT:    s_wait_kmcnt 0x0
; GFX1250-NEXT:    global_store_b32 v0, v0, s[0:1]
; GFX1250-NEXT:    s_endpgm
  %canonicalized = call <2 x bfloat> @llvm.canonicalize.v2bf16(<2 x bfloat> zeroinitializer)
  store <2 x bfloat> %canonicalized, ptr addrspace(1) %out
  ret void
}

define amdgpu_kernel void @test_fold_canonicalize_n0_v2bf16(ptr addrspace(1) %out) #1 {
; GFX1250-LABEL: test_fold_canonicalize_n0_v2bf16:
; GFX1250:       ; %bb.0:
; GFX1250-NEXT:    s_load_b64 s[0:1], s[4:5], 0x24
; GFX1250-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, 0x80008000
; GFX1250-NEXT:    s_wait_kmcnt 0x0
; GFX1250-NEXT:    global_store_b32 v0, v1, s[0:1]
; GFX1250-NEXT:    s_endpgm
  %canonicalized = call <2 x bfloat> @llvm.canonicalize.v2bf16(<2 x bfloat> <bfloat -0.0, bfloat -0.0>)
  store <2 x bfloat> %canonicalized, ptr addrspace(1) %out
  ret void
}

define amdgpu_kernel void @test_fold_canonicalize_p1_v2bf16(ptr addrspace(1) %out) #1 {
; GFX1250-LABEL: test_fold_canonicalize_p1_v2bf16:
; GFX1250:       ; %bb.0:
; GFX1250-NEXT:    s_load_b64 s[0:1], s[4:5], 0x24
; GFX1250-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, 0x3f803f80
; GFX1250-NEXT:    s_wait_kmcnt 0x0
; GFX1250-NEXT:    global_store_b32 v0, v1, s[0:1]
; GFX1250-NEXT:    s_endpgm
  %canonicalized = call <2 x bfloat> @llvm.canonicalize.v2bf16(<2 x bfloat> <bfloat 1.0, bfloat 1.0>)
  store <2 x bfloat> %canonicalized, ptr addrspace(1) %out
  ret void
}

define amdgpu_kernel void @test_fold_canonicalize_n1_v2bf16(ptr addrspace(1) %out) #1 {
; GFX1250-LABEL: test_fold_canonicalize_n1_v2bf16:
; GFX1250:       ; %bb.0:
; GFX1250-NEXT:    s_load_b64 s[0:1], s[4:5], 0x24
; GFX1250-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, 0xbf80bf80
; GFX1250-NEXT:    s_wait_kmcnt 0x0
; GFX1250-NEXT:    global_store_b32 v0, v1, s[0:1]
; GFX1250-NEXT:    s_endpgm
  %canonicalized = call <2 x bfloat> @llvm.canonicalize.v2bf16(<2 x bfloat> <bfloat -1.0, bfloat -1.0>)
  store <2 x bfloat> %canonicalized, ptr addrspace(1) %out
  ret void
}

define amdgpu_kernel void @test_fold_canonicalize_literal_v2bf16(ptr addrspace(1) %out) #1 {
; GFX1250-LABEL: test_fold_canonicalize_literal_v2bf16:
; GFX1250:       ; %bb.0:
; GFX1250-NEXT:    s_load_b64 s[0:1], s[4:5], 0x24
; GFX1250-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, 0x41804180
; GFX1250-NEXT:    s_wait_kmcnt 0x0
; GFX1250-NEXT:    global_store_b32 v0, v1, s[0:1]
; GFX1250-NEXT:    s_endpgm
  %canonicalized = call <2 x bfloat> @llvm.canonicalize.v2bf16(<2 x bfloat> <bfloat 16.0, bfloat 16.0>)
  store <2 x bfloat> %canonicalized, ptr addrspace(1) %out
  ret void
}

define amdgpu_kernel void @test_no_denormals_fold_canonicalize_denormal0_v2bf16(ptr addrspace(1) %out) #1 {
; GFX1250-LABEL: test_no_denormals_fold_canonicalize_denormal0_v2bf16:
; GFX1250:       ; %bb.0:
; GFX1250-NEXT:    s_load_b64 s[0:1], s[4:5], 0x24
; GFX1250-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, 0x3ff03ff
; GFX1250-NEXT:    s_wait_kmcnt 0x0
; GFX1250-NEXT:    global_store_b32 v0, v1, s[0:1]
; GFX1250-NEXT:    s_endpgm
  %canonicalized = call <2 x bfloat> @llvm.canonicalize.v2bf16(<2 x bfloat> <bfloat 0xR03FF, bfloat 0xR03FF>)
  store <2 x bfloat> %canonicalized, ptr addrspace(1) %out
  ret void
}

define amdgpu_kernel void @test_denormals_fold_canonicalize_denormal0_v2bf16(ptr addrspace(1) %out) #3 {
; GFX1250-LABEL: test_denormals_fold_canonicalize_denormal0_v2bf16:
; GFX1250:       ; %bb.0:
; GFX1250-NEXT:    s_load_b64 s[0:1], s[4:5], 0x24
; GFX1250-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, 0x3ff03ff
; GFX1250-NEXT:    s_wait_kmcnt 0x0
; GFX1250-NEXT:    global_store_b32 v0, v1, s[0:1]
; GFX1250-NEXT:    s_endpgm
  %canonicalized = call <2 x bfloat> @llvm.canonicalize.v2bf16(<2 x bfloat> <bfloat 0xR03FF, bfloat 0xR03FF>)
  store <2 x bfloat> %canonicalized, ptr addrspace(1) %out
  ret void
}

define amdgpu_kernel void @test_no_denormals_fold_canonicalize_denormal1_v2bf16(ptr addrspace(1) %out) #1 {
; GFX1250-LABEL: test_no_denormals_fold_canonicalize_denormal1_v2bf16:
; GFX1250:       ; %bb.0:
; GFX1250-NEXT:    s_load_b64 s[0:1], s[4:5], 0x24
; GFX1250-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, 0x83ff83ff
; GFX1250-NEXT:    s_wait_kmcnt 0x0
; GFX1250-NEXT:    global_store_b32 v0, v1, s[0:1]
; GFX1250-NEXT:    s_endpgm
  %canonicalized = call <2 x bfloat> @llvm.canonicalize.v2bf16(<2 x bfloat> <bfloat 0xR83FF, bfloat 0xR83FF>)
  store <2 x bfloat> %canonicalized, ptr addrspace(1) %out
  ret void
}

define amdgpu_kernel void @test_denormals_fold_canonicalize_denormal1_v2bf16(ptr addrspace(1) %out) #3 {
; GFX1250-LABEL: test_denormals_fold_canonicalize_denormal1_v2bf16:
; GFX1250:       ; %bb.0:
; GFX1250-NEXT:    s_load_b64 s[0:1], s[4:5], 0x24
; GFX1250-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, 0x83ff83ff
; GFX1250-NEXT:    s_wait_kmcnt 0x0
; GFX1250-NEXT:    global_store_b32 v0, v1, s[0:1]
; GFX1250-NEXT:    s_endpgm
  %canonicalized = call <2 x bfloat> @llvm.canonicalize.v2bf16(<2 x bfloat> <bfloat 0xR83FF, bfloat 0xR83FF>)
  store <2 x bfloat> %canonicalized, ptr addrspace(1) %out
  ret void
}

define amdgpu_kernel void @test_fold_canonicalize_qnan_v2bf16(ptr addrspace(1) %out) #1 {
; GFX1250-LABEL: test_fold_canonicalize_qnan_v2bf16:
; GFX1250:       ; %bb.0:
; GFX1250-NEXT:    s_load_b64 s[0:1], s[4:5], 0x24
; GFX1250-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, 0x7c007c00
; GFX1250-NEXT:    s_wait_kmcnt 0x0
; GFX1250-NEXT:    global_store_b32 v0, v1, s[0:1]
; GFX1250-NEXT:    s_endpgm
  %canonicalized = call <2 x bfloat> @llvm.canonicalize.v2bf16(<2 x bfloat> <bfloat 0xR7C00, bfloat 0xR7C00>)
  store <2 x bfloat> %canonicalized, ptr addrspace(1) %out
  ret void
}

define amdgpu_kernel void @test_fold_canonicalize_qnan_value_neg1_v2bf16(ptr addrspace(1) %out) #1 {
; GFX1250-LABEL: test_fold_canonicalize_qnan_value_neg1_v2bf16:
; GFX1250:       ; %bb.0:
; GFX1250-NEXT:    s_load_b64 s[0:1], s[4:5], 0x24
; GFX1250-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, 0x7fc07fc0
; GFX1250-NEXT:    s_wait_kmcnt 0x0
; GFX1250-NEXT:    global_store_b32 v0, v1, s[0:1]
; GFX1250-NEXT:    s_endpgm
  %canonicalized = call <2 x bfloat> @llvm.canonicalize.v2bf16(<2 x bfloat> bitcast (i32 -1 to <2 x bfloat>))
  store <2 x bfloat> %canonicalized, ptr addrspace(1) %out
  ret void
}

define amdgpu_kernel void @test_fold_canonicalize_qnan_value_neg2_v2bf16(ptr addrspace(1) %out) #1 {
; GFX1250-LABEL: test_fold_canonicalize_qnan_value_neg2_v2bf16:
; GFX1250:       ; %bb.0:
; GFX1250-NEXT:    s_load_b64 s[0:1], s[4:5], 0x24
; GFX1250-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, 0x7fc07fc0
; GFX1250-NEXT:    s_wait_kmcnt 0x0
; GFX1250-NEXT:    global_store_b32 v0, v1, s[0:1]
; GFX1250-NEXT:    s_endpgm
  %canonicalized = call <2 x bfloat> @llvm.canonicalize.v2bf16(<2 x bfloat> <bfloat bitcast (i16 -2 to bfloat), bfloat bitcast (i16 -2 to bfloat)>)
  store <2 x bfloat> %canonicalized, ptr addrspace(1) %out
  ret void
}

define amdgpu_kernel void @test_fold_canonicalize_snan0_value_v2bf16(ptr addrspace(1) %out) #1 {
; GFX1250-LABEL: test_fold_canonicalize_snan0_value_v2bf16:
; GFX1250:       ; %bb.0:
; GFX1250-NEXT:    s_load_b64 s[0:1], s[4:5], 0x24
; GFX1250-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, 0x7c017c01
; GFX1250-NEXT:    s_wait_kmcnt 0x0
; GFX1250-NEXT:    global_store_b32 v0, v1, s[0:1]
; GFX1250-NEXT:    s_endpgm
  %canonicalized = call <2 x bfloat> @llvm.canonicalize.v2bf16(<2 x bfloat> <bfloat 0xR7C01, bfloat 0xR7C01>)
  store <2 x bfloat> %canonicalized, ptr addrspace(1) %out
  ret void
}

define amdgpu_kernel void @test_fold_canonicalize_snan1_value_v2bf16(ptr addrspace(1) %out) #1 {
; GFX1250-LABEL: test_fold_canonicalize_snan1_value_v2bf16:
; GFX1250:       ; %bb.0:
; GFX1250-NEXT:    s_load_b64 s[0:1], s[4:5], 0x24
; GFX1250-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, 0x7dff7dff
; GFX1250-NEXT:    s_wait_kmcnt 0x0
; GFX1250-NEXT:    global_store_b32 v0, v1, s[0:1]
; GFX1250-NEXT:    s_endpgm
  %canonicalized = call <2 x bfloat> @llvm.canonicalize.v2bf16(<2 x bfloat> <bfloat 0xR7DFF, bfloat 0xR7DFF>)
  store <2 x bfloat> %canonicalized, ptr addrspace(1) %out
  ret void
}

define amdgpu_kernel void @test_fold_canonicalize_snan2_value_v2bf16(ptr addrspace(1) %out) #1 {
; GFX1250-LABEL: test_fold_canonicalize_snan2_value_v2bf16:
; GFX1250:       ; %bb.0:
; GFX1250-NEXT:    s_load_b64 s[0:1], s[4:5], 0x24
; GFX1250-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, 0xfdfffdff
; GFX1250-NEXT:    s_wait_kmcnt 0x0
; GFX1250-NEXT:    global_store_b32 v0, v1, s[0:1]
; GFX1250-NEXT:    s_endpgm
  %canonicalized = call <2 x bfloat> @llvm.canonicalize.v2bf16(<2 x bfloat> <bfloat 0xRFDFF, bfloat 0xRFDFF>)
  store <2 x bfloat> %canonicalized, ptr addrspace(1) %out
  ret void
}

define amdgpu_kernel void @test_fold_canonicalize_snan3_value_v2bf16(ptr addrspace(1) %out) #1 {
; GFX1250-LABEL: test_fold_canonicalize_snan3_value_v2bf16:
; GFX1250:       ; %bb.0:
; GFX1250-NEXT:    s_load_b64 s[0:1], s[4:5], 0x24
; GFX1250-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, 0xfc01fc01
; GFX1250-NEXT:    s_wait_kmcnt 0x0
; GFX1250-NEXT:    global_store_b32 v0, v1, s[0:1]
; GFX1250-NEXT:    s_endpgm
  %canonicalized = call <2 x bfloat> @llvm.canonicalize.v2bf16(<2 x bfloat> <bfloat 0xRFC01, bfloat 0xRFC01>)
  store <2 x bfloat> %canonicalized, ptr addrspace(1) %out
  ret void
}

define <3 x bfloat> @v_test_canonicalize_var_v3bf16(<3 x bfloat> %val) #1 {
; GFX1250-LABEL: v_test_canonicalize_var_v3bf16:
; GFX1250:       ; %bb.0:
; GFX1250-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX1250-NEXT:    s_wait_kmcnt 0x0
; GFX1250-NEXT:    v_and_b32_e32 v2, 0xffff0000, v0
; GFX1250-NEXT:    v_dual_lshlrev_b32 v0, 16, v0 :: v_dual_lshlrev_b32 v1, 16, v1
; GFX1250-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX1250-NEXT:    v_dual_max_num_f32 v2, v2, v2 :: v_dual_max_num_f32 v0, v0, v0
; GFX1250-NEXT:    v_max_num_f32_e32 v1, v1, v1
; GFX1250-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX1250-NEXT:    v_cvt_pk_bf16_f32 v0, v0, v2
; GFX1250-NEXT:    v_cvt_pk_bf16_f32 v1, v1, s0
; GFX1250-NEXT:    s_set_pc_i64 s[30:31]
  %canonicalized = call <3 x bfloat> @llvm.canonicalize.v3bf16(<3 x bfloat> %val)
  ret <3 x bfloat> %canonicalized
}

define <4 x bfloat> @v_test_canonicalize_var_v4bf16(<4 x bfloat> %val) #1 {
; GFX1250-LABEL: v_test_canonicalize_var_v4bf16:
; GFX1250:       ; %bb.0:
; GFX1250-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX1250-NEXT:    s_wait_kmcnt 0x0
; GFX1250-NEXT:    v_and_b32_e32 v2, 0xffff0000, v1
; GFX1250-NEXT:    v_and_b32_e32 v3, 0xffff0000, v0
; GFX1250-NEXT:    v_dual_lshlrev_b32 v0, 16, v0 :: v_dual_lshlrev_b32 v1, 16, v1
; GFX1250-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX1250-NEXT:    v_dual_max_num_f32 v2, v2, v2 :: v_dual_max_num_f32 v3, v3, v3
; GFX1250-NEXT:    v_dual_max_num_f32 v0, v0, v0 :: v_dual_max_num_f32 v1, v1, v1
; GFX1250-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX1250-NEXT:    v_cvt_pk_bf16_f32 v0, v0, v3
; GFX1250-NEXT:    v_cvt_pk_bf16_f32 v1, v1, v2
; GFX1250-NEXT:    s_set_pc_i64 s[30:31]
  %canonicalized = call <4 x bfloat> @llvm.canonicalize.v4bf16(<4 x bfloat> %val)
  ret <4 x bfloat> %canonicalized
}

define amdgpu_kernel void @s_test_canonicalize_undef_v2bf16(ptr addrspace(1) %out) #1 {
; GFX1250-LABEL: s_test_canonicalize_undef_v2bf16:
; GFX1250:       ; %bb.0:
; GFX1250-NEXT:    s_load_b64 s[0:1], s[4:5], 0x24
; GFX1250-NEXT:    v_mov_b32_e32 v0, 0
; GFX1250-NEXT:    s_wait_kmcnt 0x0
; GFX1250-NEXT:    global_store_b32 v0, v0, s[0:1]
; GFX1250-NEXT:    s_endpgm
  %canonicalized = call <2 x bfloat> @llvm.canonicalize.v2bf16(<2 x bfloat> undef)
  store <2 x bfloat> %canonicalized, ptr addrspace(1) %out
  ret void
}

define <2 x bfloat> @v_test_canonicalize_reg_undef_v2bf16(bfloat %val) #1 {
; GFX1250-LABEL: v_test_canonicalize_reg_undef_v2bf16:
; GFX1250:       ; %bb.0:
; GFX1250-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX1250-NEXT:    s_wait_kmcnt 0x0
; GFX1250-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
; GFX1250-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX1250-NEXT:    v_max_num_f32_e32 v0, v0, v0
; GFX1250-NEXT:    v_cvt_pk_bf16_f32 v0, v0, s0
; GFX1250-NEXT:    s_movk_i32 s0, 0x7fc0
; GFX1250-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instid1(SALU_CYCLE_1)
; GFX1250-NEXT:    v_perm_b32 v0, s0, v0, 0x5040100
; GFX1250-NEXT:    s_set_pc_i64 s[30:31]
  %vec = insertelement <2 x bfloat> poison, bfloat %val, i32 0
  %canonicalized = call <2 x bfloat> @llvm.canonicalize.v2bf16(<2 x bfloat> %vec)
  ret <2 x bfloat> %canonicalized
}

define <2 x bfloat> @v_test_canonicalize_undef_reg_v2bf16(bfloat %val) #1 {
; GFX1250-LABEL: v_test_canonicalize_undef_reg_v2bf16:
; GFX1250:       ; %bb.0:
; GFX1250-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX1250-NEXT:    s_wait_kmcnt 0x0
; GFX1250-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
; GFX1250-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX1250-NEXT:    v_max_num_f32_e32 v0, v0, v0
; GFX1250-NEXT:    v_cvt_pk_bf16_f32 v0, v0, s0
; GFX1250-NEXT:    s_movk_i32 s0, 0x7fc0
; GFX1250-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instid1(SALU_CYCLE_1)
; GFX1250-NEXT:    v_perm_b32 v0, v0, s0, 0x5040100
; GFX1250-NEXT:    s_set_pc_i64 s[30:31]
  %vec = insertelement <2 x bfloat> poison, bfloat %val, i32 1
  %canonicalized = call <2 x bfloat> @llvm.canonicalize.v2bf16(<2 x bfloat> %vec)
  ret <2 x bfloat> %canonicalized
}

define <2 x bfloat> @v_test_canonicalize_undef_lo_imm_hi_v2bf16() #1 {
; GFX1250-LABEL: v_test_canonicalize_undef_lo_imm_hi_v2bf16:
; GFX1250:       ; %bb.0:
; GFX1250-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX1250-NEXT:    s_wait_kmcnt 0x0
; GFX1250-NEXT:    v_mov_b32_e32 v0, 1.0
; GFX1250-NEXT:    s_set_pc_i64 s[30:31]
  %vec = insertelement <2 x bfloat> undef, bfloat 1.0, i32 1
  %canonicalized = call <2 x bfloat> @llvm.canonicalize.v2bf16(<2 x bfloat> %vec)
  ret <2 x bfloat> %canonicalized
}

define <2 x bfloat> @v_test_canonicalize_imm_lo_undef_hi_v2bf16() #1 {
; GFX1250-LABEL: v_test_canonicalize_imm_lo_undef_hi_v2bf16:
; GFX1250:       ; %bb.0:
; GFX1250-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX1250-NEXT:    s_wait_kmcnt 0x0
; GFX1250-NEXT:    v_mov_b32_e32 v0, 0x3f80
; GFX1250-NEXT:    s_set_pc_i64 s[30:31]
  %vec = insertelement <2 x bfloat> undef, bfloat 1.0, i32 0
  %canonicalized = call <2 x bfloat> @llvm.canonicalize.v2bf16(<2 x bfloat> %vec)
  ret <2 x bfloat> %canonicalized
}

define <2 x bfloat> @v_test_canonicalize_undef_lo_k_hi_v2bf16() #1 {
; GFX1250-LABEL: v_test_canonicalize_undef_lo_k_hi_v2bf16:
; GFX1250:       ; %bb.0:
; GFX1250-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX1250-NEXT:    s_wait_kmcnt 0x0
; GFX1250-NEXT:    v_mov_b32_e32 v0, 0x41800000
; GFX1250-NEXT:    s_set_pc_i64 s[30:31]
  %vec = insertelement <2 x bfloat> undef, bfloat 16.0, i32 1
  %canonicalized = call <2 x bfloat> @llvm.canonicalize.v2bf16(<2 x bfloat> %vec)
  ret <2 x bfloat> %canonicalized
}

define <2 x bfloat> @v_test_canonicalize_k_lo_undef_hi_v2bf16() #1 {
; GFX1250-LABEL: v_test_canonicalize_k_lo_undef_hi_v2bf16:
; GFX1250:       ; %bb.0:
; GFX1250-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX1250-NEXT:    s_wait_kmcnt 0x0
; GFX1250-NEXT:    v_mov_b32_e32 v0, 0x4180
; GFX1250-NEXT:    s_set_pc_i64 s[30:31]
  %vec = insertelement <2 x bfloat> undef, bfloat 16.0, i32 0
  %canonicalized = call <2 x bfloat> @llvm.canonicalize.v2bf16(<2 x bfloat> %vec)
  ret <2 x bfloat> %canonicalized
}

define <2 x bfloat> @v_test_canonicalize_reg_k_v2bf16(bfloat %val) #1 {
; GFX1250-LABEL: v_test_canonicalize_reg_k_v2bf16:
; GFX1250:       ; %bb.0:
; GFX1250-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX1250-NEXT:    s_wait_kmcnt 0x0
; GFX1250-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
; GFX1250-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX1250-NEXT:    v_max_num_f32_e32 v0, v0, v0
; GFX1250-NEXT:    v_cvt_pk_bf16_f32 v0, v0, s0
; GFX1250-NEXT:    s_movk_i32 s0, 0x4000
; GFX1250-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instid1(SALU_CYCLE_1)
; GFX1250-NEXT:    v_perm_b32 v0, s0, v0, 0x5040100
; GFX1250-NEXT:    s_set_pc_i64 s[30:31]
  %vec0 = insertelement <2 x bfloat> poison, bfloat %val, i32 0
  %vec1 = insertelement <2 x bfloat> %vec0, bfloat 2.0, i32 1
  %canonicalized = call <2 x bfloat> @llvm.canonicalize.v2bf16(<2 x bfloat> %vec1)
  ret <2 x bfloat> %canonicalized
}

define <2 x bfloat> @v_test_canonicalize_k_reg_v2bf16(bfloat %val) #1 {
; GFX1250-LABEL: v_test_canonicalize_k_reg_v2bf16:
; GFX1250:       ; %bb.0:
; GFX1250-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX1250-NEXT:    s_wait_kmcnt 0x0
; GFX1250-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
; GFX1250-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX1250-NEXT:    v_max_num_f32_e32 v0, v0, v0
; GFX1250-NEXT:    v_cvt_pk_bf16_f32 v0, v0, s0
; GFX1250-NEXT:    s_movk_i32 s0, 0x4000
; GFX1250-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instid1(SALU_CYCLE_1)
; GFX1250-NEXT:    v_perm_b32 v0, v0, s0, 0x5040100
; GFX1250-NEXT:    s_set_pc_i64 s[30:31]
  %vec0 = insertelement <2 x bfloat> poison, bfloat 2.0, i32 0
  %vec1 = insertelement <2 x bfloat> %vec0, bfloat %val, i32 1
  %canonicalized = call <2 x bfloat> @llvm.canonicalize.v2bf16(<2 x bfloat> %vec1)
  ret <2 x bfloat> %canonicalized
}

define amdgpu_kernel void @s_test_canonicalize_undef_v4bf16(ptr addrspace(1) %out) #1 {
; GFX1250-LABEL: s_test_canonicalize_undef_v4bf16:
; GFX1250:       ; %bb.0:
; GFX1250-NEXT:    s_load_b64 s[0:1], s[4:5], 0x24
; GFX1250-NEXT:    v_mov_b32_e32 v0, 0
; GFX1250-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX1250-NEXT:    v_mov_b32_e32 v1, v0
; GFX1250-NEXT:    s_wait_kmcnt 0x0
; GFX1250-NEXT:    global_store_b64 v0, v[0:1], s[0:1]
; GFX1250-NEXT:    s_endpgm
  %canonicalized = call <4 x bfloat> @llvm.canonicalize.v4bf16(<4 x bfloat> undef)
  store <4 x bfloat> %canonicalized, ptr addrspace(1) %out
  ret void
}

define <4 x bfloat> @v_test_canonicalize_reg_undef_undef_undef_v4bf16(bfloat %val) #1 {
; GFX1250-LABEL: v_test_canonicalize_reg_undef_undef_undef_v4bf16:
; GFX1250:       ; %bb.0:
; GFX1250-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX1250-NEXT:    s_wait_kmcnt 0x0
; GFX1250-NEXT:    v_dual_mov_b32 v1, 0x7fc07fc0 :: v_dual_lshlrev_b32 v0, 16, v0
; GFX1250-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX1250-NEXT:    v_max_num_f32_e32 v0, v0, v0
; GFX1250-NEXT:    v_cvt_pk_bf16_f32 v0, v0, s0
; GFX1250-NEXT:    s_movk_i32 s0, 0x7fc0
; GFX1250-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instid1(SALU_CYCLE_1)
; GFX1250-NEXT:    v_perm_b32 v0, s0, v0, 0x5040100
; GFX1250-NEXT:    s_set_pc_i64 s[30:31]
  %vec = insertelement <4 x bfloat> poison, bfloat %val, i32 0
  %canonicalized = call <4 x bfloat> @llvm.canonicalize.v4bf16(<4 x bfloat> %vec)
  ret <4 x bfloat> %canonicalized
}

define <4 x bfloat> @v_test_canonicalize_reg_reg_undef_undef_v4bf16(bfloat %val0, bfloat %val1) #1 {
; GFX1250-LABEL: v_test_canonicalize_reg_reg_undef_undef_v4bf16:
; GFX1250:       ; %bb.0:
; GFX1250-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX1250-NEXT:    s_wait_kmcnt 0x0
; GFX1250-NEXT:    v_dual_lshlrev_b32 v1, 16, v1 :: v_dual_lshlrev_b32 v0, 16, v0
; GFX1250-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX1250-NEXT:    v_dual_max_num_f32 v1, v1, v1 :: v_dual_max_num_f32 v0, v0, v0
; GFX1250-NEXT:    v_cvt_pk_bf16_f32 v0, v0, v1
; GFX1250-NEXT:    v_mov_b32_e32 v1, 0x7fc07fc0
; GFX1250-NEXT:    s_set_pc_i64 s[30:31]
  %vec0 = insertelement <4 x bfloat> poison, bfloat %val0, i32 0
  %vec1 = insertelement <4 x bfloat> %vec0, bfloat %val1, i32 1
  %canonicalized = call <4 x bfloat> @llvm.canonicalize.v4bf16(<4 x bfloat> %vec1)
  ret <4 x bfloat> %canonicalized
}

define <4 x bfloat> @v_test_canonicalize_reg_undef_reg_reg_v4bf16(bfloat %val0, bfloat %val1, bfloat %val2) #1 {
; GFX1250-LABEL: v_test_canonicalize_reg_undef_reg_reg_v4bf16:
; GFX1250:       ; %bb.0:
; GFX1250-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX1250-NEXT:    s_wait_kmcnt 0x0
; GFX1250-NEXT:    v_dual_lshlrev_b32 v0, 16, v0 :: v_dual_lshlrev_b32 v2, 16, v2
; GFX1250-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX1250-NEXT:    v_dual_max_num_f32 v0, v0, v0 :: v_dual_lshlrev_b32 v1, 16, v1
; GFX1250-NEXT:    v_dual_max_num_f32 v2, v2, v2 :: v_dual_max_num_f32 v1, v1, v1
; GFX1250-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX1250-NEXT:    v_cvt_pk_bf16_f32 v0, v0, s0
; GFX1250-NEXT:    s_movk_i32 s0, 0x7fc0
; GFX1250-NEXT:    v_cvt_pk_bf16_f32 v1, v1, v2
; GFX1250-NEXT:    s_delay_alu instid0(VALU_DEP_2)
; GFX1250-NEXT:    v_perm_b32 v0, s0, v0, 0x5040100
; GFX1250-NEXT:    s_set_pc_i64 s[30:31]
  %vec0 = insertelement <4 x bfloat> poison, bfloat %val0, i32 0
  %vec1 = insertelement <4 x bfloat> %vec0, bfloat %val1, i32 2
  %vec2 = insertelement <4 x bfloat> %vec1, bfloat %val2, i32 3
  %canonicalized = call <4 x bfloat> @llvm.canonicalize.v4bf16(<4 x bfloat> %vec2)
  ret <4 x bfloat> %canonicalized
}

define <6 x bfloat> @v_test_canonicalize_var_v6bf16(<6 x bfloat> %val) #1 {
; GFX1250-LABEL: v_test_canonicalize_var_v6bf16:
; GFX1250:       ; %bb.0:
; GFX1250-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX1250-NEXT:    s_wait_kmcnt 0x0
; GFX1250-NEXT:    v_and_b32_e32 v3, 0xffff0000, v2
; GFX1250-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; GFX1250-NEXT:    v_and_b32_e32 v4, 0xffff0000, v1
; GFX1250-NEXT:    v_and_b32_e32 v5, 0xffff0000, v0
; GFX1250-NEXT:    v_dual_lshlrev_b32 v0, 16, v0 :: v_dual_lshlrev_b32 v1, 16, v1
; GFX1250-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX1250-NEXT:    v_dual_max_num_f32 v3, v3, v3 :: v_dual_max_num_f32 v4, v4, v4
; GFX1250-NEXT:    v_dual_max_num_f32 v5, v5, v5 :: v_dual_max_num_f32 v0, v0, v0
; GFX1250-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX1250-NEXT:    v_dual_max_num_f32 v1, v1, v1 :: v_dual_max_num_f32 v2, v2, v2
; GFX1250-NEXT:    v_cvt_pk_bf16_f32 v0, v0, v5
; GFX1250-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX1250-NEXT:    v_cvt_pk_bf16_f32 v1, v1, v4
; GFX1250-NEXT:    v_cvt_pk_bf16_f32 v2, v2, v3
; GFX1250-NEXT:    s_set_pc_i64 s[30:31]
  %canonicalized = call <6 x bfloat> @llvm.canonicalize.v6bf16(<6 x bfloat> %val)
  ret <6 x bfloat> %canonicalized
}

define <8 x bfloat> @v_test_canonicalize_var_v8bf16(<8 x bfloat> %val) #1 {
; GFX1250-LABEL: v_test_canonicalize_var_v8bf16:
; GFX1250:       ; %bb.0:
; GFX1250-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX1250-NEXT:    s_wait_kmcnt 0x0
; GFX1250-NEXT:    v_and_b32_e32 v5, 0xffff0000, v2
; GFX1250-NEXT:    v_and_b32_e32 v4, 0xffff0000, v3
; GFX1250-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; GFX1250-NEXT:    v_and_b32_e32 v6, 0xffff0000, v1
; GFX1250-NEXT:    v_and_b32_e32 v7, 0xffff0000, v0
; GFX1250-NEXT:    v_dual_max_num_f32 v5, v5, v5 :: v_dual_lshlrev_b32 v0, 16, v0
; GFX1250-NEXT:    v_dual_lshlrev_b32 v1, 16, v1 :: v_dual_lshlrev_b32 v2, 16, v2
; GFX1250-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX1250-NEXT:    v_dual_max_num_f32 v4, v4, v4 :: v_dual_max_num_f32 v6, v6, v6
; GFX1250-NEXT:    v_dual_max_num_f32 v7, v7, v7 :: v_dual_max_num_f32 v0, v0, v0
; GFX1250-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX1250-NEXT:    v_dual_max_num_f32 v1, v1, v1 :: v_dual_max_num_f32 v2, v2, v2
; GFX1250-NEXT:    v_max_num_f32_e32 v3, v3, v3
; GFX1250-NEXT:    v_cvt_pk_bf16_f32 v0, v0, v7
; GFX1250-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX1250-NEXT:    v_cvt_pk_bf16_f32 v1, v1, v6
; GFX1250-NEXT:    v_cvt_pk_bf16_f32 v2, v2, v5
; GFX1250-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX1250-NEXT:    v_cvt_pk_bf16_f32 v3, v3, v4
; GFX1250-NEXT:    s_set_pc_i64 s[30:31]
  %canonicalized = call <8 x bfloat> @llvm.canonicalize.v8bf16(<8 x bfloat> %val)
  ret <8 x bfloat> %canonicalized
}

define <12 x bfloat> @v_test_canonicalize_var_v12bf16(<12 x bfloat> %val) #1 {
; GFX1250-LABEL: v_test_canonicalize_var_v12bf16:
; GFX1250:       ; %bb.0:
; GFX1250-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX1250-NEXT:    s_wait_kmcnt 0x0
; GFX1250-NEXT:    v_and_b32_e32 v6, 0xffff0000, v5
; GFX1250-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; GFX1250-NEXT:    v_and_b32_e32 v7, 0xffff0000, v4
; GFX1250-NEXT:    v_and_b32_e32 v8, 0xffff0000, v3
; GFX1250-NEXT:    v_and_b32_e32 v9, 0xffff0000, v2
; GFX1250-NEXT:    v_dual_lshlrev_b32 v4, 16, v4 :: v_dual_lshlrev_b32 v3, 16, v3
; GFX1250-NEXT:    v_and_b32_e32 v10, 0xffff0000, v1
; GFX1250-NEXT:    v_and_b32_e32 v11, 0xffff0000, v0
; GFX1250-NEXT:    v_dual_lshlrev_b32 v0, 16, v0 :: v_dual_lshlrev_b32 v1, 16, v1
; GFX1250-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; GFX1250-NEXT:    v_dual_max_num_f32 v6, v6, v6 :: v_dual_max_num_f32 v5, v5, v5
; GFX1250-NEXT:    v_dual_max_num_f32 v7, v7, v7 :: v_dual_max_num_f32 v8, v8, v8
; GFX1250-NEXT:    v_dual_max_num_f32 v9, v9, v9 :: v_dual_max_num_f32 v10, v10, v10
; GFX1250-NEXT:    v_dual_max_num_f32 v0, v0, v0 :: v_dual_max_num_f32 v11, v11, v11
; GFX1250-NEXT:    v_dual_max_num_f32 v1, v1, v1 :: v_dual_max_num_f32 v2, v2, v2
; GFX1250-NEXT:    v_dual_max_num_f32 v3, v3, v3 :: v_dual_max_num_f32 v4, v4, v4
; GFX1250-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX1250-NEXT:    v_cvt_pk_bf16_f32 v0, v0, v11
; GFX1250-NEXT:    v_cvt_pk_bf16_f32 v1, v1, v10
; GFX1250-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX1250-NEXT:    v_cvt_pk_bf16_f32 v2, v2, v9
; GFX1250-NEXT:    v_cvt_pk_bf16_f32 v3, v3, v8
; GFX1250-NEXT:    v_cvt_pk_bf16_f32 v4, v4, v7
; GFX1250-NEXT:    v_cvt_pk_bf16_f32 v5, v5, v6
; GFX1250-NEXT:    s_set_pc_i64 s[30:31]
  %canonicalized = call <12 x bfloat> @llvm.canonicalize.v12bf16(<12 x bfloat> %val)
  ret <12 x bfloat> %canonicalized
}

define <16 x bfloat> @v_test_canonicalize_var_v16bf16(<16 x bfloat> %val) #1 {
; GFX1250-LABEL: v_test_canonicalize_var_v16bf16:
; GFX1250:       ; %bb.0:
; GFX1250-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX1250-NEXT:    s_wait_kmcnt 0x0
; GFX1250-NEXT:    v_and_b32_e32 v8, 0xffff0000, v7
; GFX1250-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; GFX1250-NEXT:    v_and_b32_e32 v9, 0xffff0000, v6
; GFX1250-NEXT:    v_and_b32_e32 v10, 0xffff0000, v5
; GFX1250-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX1250-NEXT:    v_dual_max_num_f32 v8, v8, v8 :: v_dual_lshlrev_b32 v5, 16, v5
; GFX1250-NEXT:    v_dual_max_num_f32 v7, v7, v7 :: v_dual_lshlrev_b32 v6, 16, v6
; GFX1250-NEXT:    v_and_b32_e32 v11, 0xffff0000, v4
; GFX1250-NEXT:    v_and_b32_e32 v12, 0xffff0000, v3
; GFX1250-NEXT:    v_and_b32_e32 v13, 0xffff0000, v2
; GFX1250-NEXT:    v_dual_lshlrev_b32 v4, 16, v4 :: v_dual_lshlrev_b32 v3, 16, v3
; GFX1250-NEXT:    v_and_b32_e32 v14, 0xffff0000, v1
; GFX1250-NEXT:    v_and_b32_e32 v15, 0xffff0000, v0
; GFX1250-NEXT:    v_dual_lshlrev_b32 v0, 16, v0 :: v_dual_lshlrev_b32 v1, 16, v1
; GFX1250-NEXT:    v_dual_max_num_f32 v9, v9, v9 :: v_dual_lshlrev_b32 v2, 16, v2
; GFX1250-NEXT:    v_max_num_f32_e32 v6, v6, v6
; GFX1250-NEXT:    v_dual_max_num_f32 v10, v10, v10 :: v_dual_max_num_f32 v5, v5, v5
; GFX1250-NEXT:    v_dual_max_num_f32 v11, v11, v11 :: v_dual_max_num_f32 v12, v12, v12
; GFX1250-NEXT:    v_dual_max_num_f32 v13, v13, v13 :: v_dual_max_num_f32 v14, v14, v14
; GFX1250-NEXT:    v_dual_max_num_f32 v0, v0, v0 :: v_dual_max_num_f32 v15, v15, v15
; GFX1250-NEXT:    v_dual_max_num_f32 v1, v1, v1 :: v_dual_max_num_f32 v2, v2, v2
; GFX1250-NEXT:    v_dual_max_num_f32 v3, v3, v3 :: v_dual_max_num_f32 v4, v4, v4
; GFX1250-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX1250-NEXT:    v_cvt_pk_bf16_f32 v0, v0, v15
; GFX1250-NEXT:    v_cvt_pk_bf16_f32 v1, v1, v14
; GFX1250-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX1250-NEXT:    v_cvt_pk_bf16_f32 v2, v2, v13
; GFX1250-NEXT:    v_cvt_pk_bf16_f32 v3, v3, v12
; GFX1250-NEXT:    v_cvt_pk_bf16_f32 v4, v4, v11
; GFX1250-NEXT:    v_cvt_pk_bf16_f32 v5, v5, v10
; GFX1250-NEXT:    v_cvt_pk_bf16_f32 v6, v6, v9
; GFX1250-NEXT:    v_cvt_pk_bf16_f32 v7, v7, v8
; GFX1250-NEXT:    s_set_pc_i64 s[30:31]
  %canonicalized = call <16 x bfloat> @llvm.canonicalize.v16bf16(<16 x bfloat> %val)
  ret <16 x bfloat> %canonicalized
}

define <32 x bfloat> @v_test_canonicalize_var_v32bf16(<32 x bfloat> %val) #1 {
; GFX1250-LABEL: v_test_canonicalize_var_v32bf16:
; GFX1250:       ; %bb.0:
; GFX1250-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX1250-NEXT:    s_wait_kmcnt 0x0
; GFX1250-NEXT:    v_and_b32_e32 v16, 0xffff0000, v15
; GFX1250-NEXT:    v_and_b32_e32 v18, 0xffff0000, v13
; GFX1250-NEXT:    v_and_b32_e32 v20, 0xffff0000, v11
; GFX1250-NEXT:    v_and_b32_e32 v22, 0xffff0000, v9
; GFX1250-NEXT:    v_and_b32_e32 v24, 0xffff0000, v7
; GFX1250-NEXT:    v_dual_max_num_f32 v16, v16, v16 :: v_dual_lshlrev_b32 v15, 16, v15
; GFX1250-NEXT:    v_and_b32_e32 v17, 0xffff0000, v14
; GFX1250-NEXT:    v_dual_lshlrev_b32 v14, 16, v14 :: v_dual_lshlrev_b32 v13, 16, v13
; GFX1250-NEXT:    v_max_num_f32_e32 v18, v18, v18
; GFX1250-NEXT:    v_and_b32_e32 v19, 0xffff0000, v12
; GFX1250-NEXT:    v_dual_lshlrev_b32 v12, 16, v12 :: v_dual_lshlrev_b32 v11, 16, v11
; GFX1250-NEXT:    v_max_num_f32_e32 v20, v20, v20
; GFX1250-NEXT:    v_and_b32_e32 v21, 0xffff0000, v10
; GFX1250-NEXT:    v_dual_lshlrev_b32 v10, 16, v10 :: v_dual_lshlrev_b32 v9, 16, v9
; GFX1250-NEXT:    v_max_num_f32_e32 v22, v22, v22
; GFX1250-NEXT:    v_and_b32_e32 v23, 0xffff0000, v8
; GFX1250-NEXT:    v_dual_lshlrev_b32 v8, 16, v8 :: v_dual_lshlrev_b32 v7, 16, v7
; GFX1250-NEXT:    v_max_num_f32_e32 v24, v24, v24
; GFX1250-NEXT:    v_and_b32_e32 v25, 0xffff0000, v6
; GFX1250-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; GFX1250-NEXT:    v_and_b32_e32 v26, 0xffff0000, v5
; GFX1250-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; GFX1250-NEXT:    v_and_b32_e32 v27, 0xffff0000, v4
; GFX1250-NEXT:    v_and_b32_e32 v28, 0xffff0000, v3
; GFX1250-NEXT:    v_and_b32_e32 v29, 0xffff0000, v2
; GFX1250-NEXT:    v_dual_lshlrev_b32 v4, 16, v4 :: v_dual_lshlrev_b32 v3, 16, v3
; GFX1250-NEXT:    v_and_b32_e32 v30, 0xffff0000, v1
; GFX1250-NEXT:    v_and_b32_e32 v31, 0xffff0000, v0
; GFX1250-NEXT:    v_dual_lshlrev_b32 v0, 16, v0 :: v_dual_lshlrev_b32 v1, 16, v1
; GFX1250-NEXT:    v_dual_max_num_f32 v15, v15, v15 :: v_dual_lshlrev_b32 v2, 16, v2
; GFX1250-NEXT:    v_dual_max_num_f32 v17, v17, v17 :: v_dual_max_num_f32 v14, v14, v14
; GFX1250-NEXT:    v_dual_max_num_f32 v13, v13, v13 :: v_dual_max_num_f32 v19, v19, v19
; GFX1250-NEXT:    v_dual_max_num_f32 v12, v12, v12 :: v_dual_max_num_f32 v11, v11, v11
; GFX1250-NEXT:    v_dual_max_num_f32 v21, v21, v21 :: v_dual_max_num_f32 v10, v10, v10
; GFX1250-NEXT:    v_dual_max_num_f32 v9, v9, v9 :: v_dual_max_num_f32 v23, v23, v23
; GFX1250-NEXT:    v_dual_max_num_f32 v8, v8, v8 :: v_dual_max_num_f32 v7, v7, v7
; GFX1250-NEXT:    v_dual_max_num_f32 v25, v25, v25 :: v_dual_max_num_f32 v6, v6, v6
; GFX1250-NEXT:    v_dual_max_num_f32 v26, v26, v26 :: v_dual_max_num_f32 v5, v5, v5
; GFX1250-NEXT:    v_dual_max_num_f32 v27, v27, v27 :: v_dual_max_num_f32 v28, v28, v28
; GFX1250-NEXT:    v_dual_max_num_f32 v29, v29, v29 :: v_dual_max_num_f32 v30, v30, v30
; GFX1250-NEXT:    v_dual_max_num_f32 v0, v0, v0 :: v_dual_max_num_f32 v31, v31, v31
; GFX1250-NEXT:    v_dual_max_num_f32 v1, v1, v1 :: v_dual_max_num_f32 v2, v2, v2
; GFX1250-NEXT:    v_dual_max_num_f32 v3, v3, v3 :: v_dual_max_num_f32 v4, v4, v4
; GFX1250-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX1250-NEXT:    v_cvt_pk_bf16_f32 v0, v0, v31
; GFX1250-NEXT:    v_cvt_pk_bf16_f32 v1, v1, v30
; GFX1250-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX1250-NEXT:    v_cvt_pk_bf16_f32 v2, v2, v29
; GFX1250-NEXT:    v_cvt_pk_bf16_f32 v3, v3, v28
; GFX1250-NEXT:    v_cvt_pk_bf16_f32 v4, v4, v27
; GFX1250-NEXT:    v_cvt_pk_bf16_f32 v5, v5, v26
; GFX1250-NEXT:    v_cvt_pk_bf16_f32 v6, v6, v25
; GFX1250-NEXT:    v_cvt_pk_bf16_f32 v7, v7, v24
; GFX1250-NEXT:    v_cvt_pk_bf16_f32 v8, v8, v23
; GFX1250-NEXT:    v_cvt_pk_bf16_f32 v9, v9, v22
; GFX1250-NEXT:    v_cvt_pk_bf16_f32 v10, v10, v21
; GFX1250-NEXT:    v_cvt_pk_bf16_f32 v11, v11, v20
; GFX1250-NEXT:    v_cvt_pk_bf16_f32 v12, v12, v19
; GFX1250-NEXT:    v_cvt_pk_bf16_f32 v13, v13, v18
; GFX1250-NEXT:    v_cvt_pk_bf16_f32 v14, v14, v17
; GFX1250-NEXT:    v_cvt_pk_bf16_f32 v15, v15, v16
; GFX1250-NEXT:    s_set_pc_i64 s[30:31]
  %canonicalized = call <32 x bfloat> @llvm.canonicalize.v32bf16(<32 x bfloat> %val)
  ret <32 x bfloat> %canonicalized
}

define <64 x bfloat> @v_test_canonicalize_var_v64bf16(<64 x bfloat> %val) #1 {
; GFX1250-LABEL: v_test_canonicalize_var_v64bf16:
; GFX1250:       ; %bb.0:
; GFX1250-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX1250-NEXT:    s_wait_kmcnt 0x0
; GFX1250-NEXT:    scratch_load_b32 v31, off, s32
; GFX1250-NEXT:    v_and_b32_e32 v81, 0xffff0000, v0
; GFX1250-NEXT:    v_and_b32_e32 v38, 0xffff0000, v24
; GFX1250-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
; GFX1250-NEXT:    v_and_b32_e32 v39, 0xffff0000, v23
; GFX1250-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; GFX1250-NEXT:    v_and_b32_e32 v80, 0xffff0000, v6
; GFX1250-NEXT:    v_dual_lshlrev_b32 v0, 16, v0 :: v_dual_lshlrev_b32 v6, 16, v6
; GFX1250-NEXT:    v_and_b32_e32 v82, 0xffff0000, v1
; GFX1250-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; GFX1250-NEXT:    v_max_num_f32_e32 v81, v81, v81
; GFX1250-NEXT:    v_and_b32_e32 v83, 0xffff0000, v2
; GFX1250-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; GFX1250-NEXT:    v_and_b32_e32 v34, 0xffff0000, v28
; GFX1250-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
; GFX1250-NEXT:    v_and_b32_e32 v35, 0xffff0000, v27
; GFX1250-NEXT:    v_lshlrev_b32_e32 v27, 16, v27
; GFX1250-NEXT:    v_and_b32_e32 v36, 0xffff0000, v26
; GFX1250-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
; GFX1250-NEXT:    v_and_b32_e32 v48, 0xffff0000, v22
; GFX1250-NEXT:    v_dual_max_num_f32 v0, v0, v0 :: v_dual_max_num_f32 v82, v82, v82
; GFX1250-NEXT:    v_dual_max_num_f32 v1, v1, v1 :: v_dual_max_num_f32 v83, v83, v83
; GFX1250-NEXT:    v_dual_max_num_f32 v2, v2, v2 :: v_dual_max_num_f32 v24, v24, v24
; GFX1250-NEXT:    v_max_num_f32_e32 v39, v39, v39
; GFX1250-NEXT:    v_dual_max_num_f32 v23, v23, v23 :: v_dual_max_num_f32 v48, v48, v48
; GFX1250-NEXT:    v_and_b32_e32 v32, 0xffff0000, v30
; GFX1250-NEXT:    v_lshlrev_b32_e32 v30, 16, v30
; GFX1250-NEXT:    v_and_b32_e32 v33, 0xffff0000, v29
; GFX1250-NEXT:    v_lshlrev_b32_e32 v29, 16, v29
; GFX1250-NEXT:    v_and_b32_e32 v37, 0xffff0000, v25
; GFX1250-NEXT:    v_dual_lshlrev_b32 v25, 16, v25 :: v_dual_lshlrev_b32 v22, 16, v22
; GFX1250-NEXT:    v_and_b32_e32 v49, 0xffff0000, v21
; GFX1250-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
; GFX1250-NEXT:    v_and_b32_e32 v50, 0xffff0000, v20
; GFX1250-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
; GFX1250-NEXT:    v_and_b32_e32 v51, 0xffff0000, v19
; GFX1250-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
; GFX1250-NEXT:    v_and_b32_e32 v52, 0xffff0000, v18
; GFX1250-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
; GFX1250-NEXT:    v_and_b32_e32 v53, 0xffff0000, v17
; GFX1250-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
; GFX1250-NEXT:    v_and_b32_e32 v54, 0xffff0000, v16
; GFX1250-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
; GFX1250-NEXT:    v_and_b32_e32 v55, 0xffff0000, v15
; GFX1250-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; GFX1250-NEXT:    v_and_b32_e32 v64, 0xffff0000, v14
; GFX1250-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
; GFX1250-NEXT:    v_and_b32_e32 v65, 0xffff0000, v13
; GFX1250-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
; GFX1250-NEXT:    v_and_b32_e32 v66, 0xffff0000, v12
; GFX1250-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; GFX1250-NEXT:    v_and_b32_e32 v67, 0xffff0000, v11
; GFX1250-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
; GFX1250-NEXT:    v_and_b32_e32 v68, 0xffff0000, v10
; GFX1250-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
; GFX1250-NEXT:    v_and_b32_e32 v69, 0xffff0000, v9
; GFX1250-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; GFX1250-NEXT:    v_and_b32_e32 v70, 0xffff0000, v8
; GFX1250-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; GFX1250-NEXT:    v_and_b32_e32 v71, 0xffff0000, v7
; GFX1250-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; GFX1250-NEXT:    v_cvt_pk_bf16_f32 v0, v0, v81
; GFX1250-NEXT:    v_and_b32_e32 v81, 0xffff0000, v5
; GFX1250-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; GFX1250-NEXT:    v_cvt_pk_bf16_f32 v1, v1, v82
; GFX1250-NEXT:    v_and_b32_e32 v82, 0xffff0000, v4
; GFX1250-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; GFX1250-NEXT:    v_cvt_pk_bf16_f32 v2, v2, v83
; GFX1250-NEXT:    v_and_b32_e32 v83, 0xffff0000, v3
; GFX1250-NEXT:    v_dual_max_num_f32 v32, v32, v32 :: v_dual_lshlrev_b32 v3, 16, v3
; GFX1250-NEXT:    v_dual_max_num_f32 v27, v27, v27 :: v_dual_max_num_f32 v36, v36, v36
; GFX1250-NEXT:    v_dual_max_num_f32 v26, v26, v26 :: v_dual_max_num_f32 v37, v37, v37
; GFX1250-NEXT:    v_cvt_pk_bf16_f32 v23, v23, v39
; GFX1250-NEXT:    v_dual_max_num_f32 v30, v30, v30 :: v_dual_max_num_f32 v33, v33, v33
; GFX1250-NEXT:    v_dual_max_num_f32 v29, v29, v29 :: v_dual_max_num_f32 v34, v34, v34
; GFX1250-NEXT:    v_dual_max_num_f32 v28, v28, v28 :: v_dual_max_num_f32 v35, v35, v35
; GFX1250-NEXT:    v_dual_max_num_f32 v25, v25, v25 :: v_dual_max_num_f32 v38, v38, v38
; GFX1250-NEXT:    v_dual_max_num_f32 v22, v22, v22 :: v_dual_max_num_f32 v49, v49, v49
; GFX1250-NEXT:    v_dual_max_num_f32 v21, v21, v21 :: v_dual_max_num_f32 v50, v50, v50
; GFX1250-NEXT:    v_dual_max_num_f32 v20, v20, v20 :: v_dual_max_num_f32 v51, v51, v51
; GFX1250-NEXT:    v_dual_max_num_f32 v19, v19, v19 :: v_dual_max_num_f32 v52, v52, v52
; GFX1250-NEXT:    v_dual_max_num_f32 v18, v18, v18 :: v_dual_max_num_f32 v53, v53, v53
; GFX1250-NEXT:    v_dual_max_num_f32 v17, v17, v17 :: v_dual_max_num_f32 v54, v54, v54
; GFX1250-NEXT:    v_dual_max_num_f32 v16, v16, v16 :: v_dual_max_num_f32 v55, v55, v55
; GFX1250-NEXT:    v_dual_max_num_f32 v15, v15, v15 :: v_dual_max_num_f32 v64, v64, v64
; GFX1250-NEXT:    v_dual_max_num_f32 v14, v14, v14 :: v_dual_max_num_f32 v65, v65, v65
; GFX1250-NEXT:    v_dual_max_num_f32 v13, v13, v13 :: v_dual_max_num_f32 v66, v66, v66
; GFX1250-NEXT:    v_dual_max_num_f32 v12, v12, v12 :: v_dual_max_num_f32 v67, v67, v67
; GFX1250-NEXT:    v_dual_max_num_f32 v11, v11, v11 :: v_dual_max_num_f32 v68, v68, v68
; GFX1250-NEXT:    v_dual_max_num_f32 v10, v10, v10 :: v_dual_max_num_f32 v69, v69, v69
; GFX1250-NEXT:    v_dual_max_num_f32 v9, v9, v9 :: v_dual_max_num_f32 v70, v70, v70
; GFX1250-NEXT:    v_dual_max_num_f32 v8, v8, v8 :: v_dual_max_num_f32 v71, v71, v71
; GFX1250-NEXT:    v_dual_max_num_f32 v80, v80, v80 :: v_dual_max_num_f32 v81, v81, v81
; GFX1250-NEXT:    v_dual_max_num_f32 v82, v82, v82 :: v_dual_max_num_f32 v83, v83, v83
; GFX1250-NEXT:    v_dual_max_num_f32 v3, v3, v3 :: v_dual_max_num_f32 v4, v4, v4
; GFX1250-NEXT:    v_dual_max_num_f32 v5, v5, v5 :: v_dual_max_num_f32 v6, v6, v6
; GFX1250-NEXT:    v_max_num_f32_e32 v7, v7, v7
; GFX1250-NEXT:    v_cvt_pk_bf16_f32 v26, v26, v36
; GFX1250-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX1250-NEXT:    v_cvt_pk_bf16_f32 v3, v3, v83
; GFX1250-NEXT:    v_cvt_pk_bf16_f32 v4, v4, v82
; GFX1250-NEXT:    v_cvt_pk_bf16_f32 v5, v5, v81
; GFX1250-NEXT:    v_cvt_pk_bf16_f32 v6, v6, v80
; GFX1250-NEXT:    v_cvt_pk_bf16_f32 v7, v7, v71
; GFX1250-NEXT:    v_cvt_pk_bf16_f32 v8, v8, v70
; GFX1250-NEXT:    v_cvt_pk_bf16_f32 v9, v9, v69
; GFX1250-NEXT:    v_cvt_pk_bf16_f32 v10, v10, v68
; GFX1250-NEXT:    v_cvt_pk_bf16_f32 v11, v11, v67
; GFX1250-NEXT:    v_cvt_pk_bf16_f32 v12, v12, v66
; GFX1250-NEXT:    v_cvt_pk_bf16_f32 v13, v13, v65
; GFX1250-NEXT:    v_cvt_pk_bf16_f32 v14, v14, v64
; GFX1250-NEXT:    v_cvt_pk_bf16_f32 v15, v15, v55
; GFX1250-NEXT:    v_cvt_pk_bf16_f32 v16, v16, v54
; GFX1250-NEXT:    v_cvt_pk_bf16_f32 v17, v17, v53
; GFX1250-NEXT:    v_cvt_pk_bf16_f32 v18, v18, v52
; GFX1250-NEXT:    v_cvt_pk_bf16_f32 v19, v19, v51
; GFX1250-NEXT:    v_cvt_pk_bf16_f32 v20, v20, v50
; GFX1250-NEXT:    v_cvt_pk_bf16_f32 v21, v21, v49
; GFX1250-NEXT:    v_cvt_pk_bf16_f32 v22, v22, v48
; GFX1250-NEXT:    v_cvt_pk_bf16_f32 v24, v24, v38
; GFX1250-NEXT:    v_cvt_pk_bf16_f32 v25, v25, v37
; GFX1250-NEXT:    v_cvt_pk_bf16_f32 v27, v27, v35
; GFX1250-NEXT:    v_cvt_pk_bf16_f32 v28, v28, v34
; GFX1250-NEXT:    v_cvt_pk_bf16_f32 v29, v29, v33
; GFX1250-NEXT:    v_cvt_pk_bf16_f32 v30, v30, v32
; GFX1250-NEXT:    s_wait_loadcnt 0x0
; GFX1250-NEXT:    v_and_b32_e32 v39, 0xffff0000, v31
; GFX1250-NEXT:    v_lshlrev_b32_e32 v31, 16, v31
; GFX1250-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX1250-NEXT:    v_max_num_f32_e32 v36, v39, v39
; GFX1250-NEXT:    v_max_num_f32_e32 v31, v31, v31
; GFX1250-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX1250-NEXT:    v_cvt_pk_bf16_f32 v31, v31, v36
; GFX1250-NEXT:    s_set_pc_i64 s[30:31]
  %canonicalized = call <64 x bfloat> @llvm.canonicalize.v64bf16(<64 x bfloat> %val)
  ret <64 x bfloat> %canonicalized
}

attributes #0 = { nounwind readnone }
attributes #1 = { nounwind "denormal-fp-math-f32"="preserve-sign,preserve-sign" }
attributes #2 = { nounwind "denormal-fp-math"="preserve-sign,preserve-sign" }
attributes #3 = { nounwind "denormal-fp-math-f32"="preserve-sign,preserve-sign" }