1; RUN: llc -march=amdgcn -mcpu=tonga -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=GCN,VI,GFX89 %s 2; RUN: llc -march=amdgcn -mcpu=gfx900 -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=GCN,GFX9,GFX89 %s 3; RUN: llc -march=amdgcn -mcpu=kaveri -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=GCN,CI %s 4 5declare half @llvm.fabs.f16(half) #0 6declare half @llvm.canonicalize.f16(half) #0 7declare <2 x half> @llvm.fabs.v2f16(<2 x half>) #0 8declare <2 x half> @llvm.canonicalize.v2f16(<2 x half>) #0 9declare <3 x half> @llvm.canonicalize.v3f16(<3 x half>) #0 10declare <4 x half> @llvm.canonicalize.v4f16(<4 x half>) #0 11declare i32 @llvm.amdgcn.workitem.id.x() #0 12 13; GCN-LABEL: {{^}}test_fold_canonicalize_undef_value_f16: 14; GFX89: v_mov_b32_e32 [[REG:v[0-9]+]], 0x7e00{{$}} 15; GFX89: {{flat|global}}_store_short v{{\[[0-9]+:[0-9]+\]}}, [[REG]] 16define amdgpu_kernel void @test_fold_canonicalize_undef_value_f16(half addrspace(1)* %out) #1 { 17 %canonicalized = call half @llvm.canonicalize.f16(half undef) 18 store half %canonicalized, half addrspace(1)* %out 19 ret void 20} 21 22; GCN-LABEL: {{^}}v_test_canonicalize_var_f16: 23; GFX89: v_max_f16_e32 [[REG:v[0-9]+]], {{v[0-9]+}}, {{v[0-9]+}} 24; GFX89: {{flat|global}}_store_short v{{\[[0-9]+:[0-9]+\]}}, [[REG]] 25 26; CI: v_cvt_f32_f16_e32 27; CI: v_mul_f32_e32 {{v[0-9]+}}, 1.0, {{v[0-9]+}} 28define amdgpu_kernel void @v_test_canonicalize_var_f16(half addrspace(1)* %out) #1 { 29 %val = load half, half addrspace(1)* %out 30 %canonicalized = call half @llvm.canonicalize.f16(half %val) 31 store half %canonicalized, half addrspace(1)* undef 32 ret void 33} 34 35; GCN-LABEL: {{^}}s_test_canonicalize_var_f16: 36; GFX89: v_max_f16_e64 [[REG:v[0-9]+]], {{s[0-9]+}}, {{s[0-9]+}} 37; GFX89: {{flat|global}}_store_short v{{\[[0-9]+:[0-9]+\]}}, [[REG]] 38define amdgpu_kernel void @s_test_canonicalize_var_f16(half addrspace(1)* %out, i16 zeroext %val.arg) #1 { 39 %val = bitcast i16 %val.arg to half 40 %canonicalized = call half @llvm.canonicalize.f16(half %val) 41 store half %canonicalized, half addrspace(1)* %out 42 ret void 43} 44 45; GCN-LABEL: {{^}}v_test_canonicalize_build_vector_v2f16: 46; GFX9: v_and_b32_e32 v0, 0xffff, v0 47; GFX9-NEXT: v_lshl_or_b32 v0, v1, 16, v0 48; GFX9-NEXT: v_pk_max_f16 v0, v0, v0 49 50; VI: v_max_f16_sdwa v1, v1, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD 51; VI: v_max_f16_e32 v0, v0, v0 52; VI: v_or_b32_e32 v0, v0, v1 53define <2 x half> @v_test_canonicalize_build_vector_v2f16(half %lo, half %hi) #1 { 54 %ins0 = insertelement <2 x half> undef, half %lo, i32 0 55 %ins1 = insertelement <2 x half> %ins0, half %hi, i32 1 56 %canonicalized = call <2 x half> @llvm.canonicalize.v2f16(<2 x half> %ins1) 57 ret <2 x half> %canonicalized 58} 59 60; GCN-LABEL: {{^}}v_test_canonicalize_fabs_var_f16: 61; GFX89: v_max_f16_e64 [[REG:v[0-9]+]], |{{v[0-9]+}}|, |{{v[0-9]+}}| 62; GFX89: {{flat|global}}_store_short v{{\[[0-9]+:[0-9]+\]}}, [[REG]] 63define amdgpu_kernel void @v_test_canonicalize_fabs_var_f16(half addrspace(1)* %out) #1 { 64 %val = load half, half addrspace(1)* %out 65 %val.fabs = call half @llvm.fabs.f16(half %val) 66 %canonicalized = call half @llvm.canonicalize.f16(half %val.fabs) 67 store half %canonicalized, half addrspace(1)* %out 68 ret void 69} 70 71; GCN-LABEL: {{^}}v_test_canonicalize_fneg_fabs_var_f16: 72; GFX89: v_max_f16_e64 [[REG:v[0-9]+]], -|{{v[0-9]+}}|, -|{{v[0-9]+}}| 73; GFX89: {{flat|global}}_store_short v{{\[[0-9]+:[0-9]+\]}}, [[REG]] 74 75; CI: v_cvt_f32_f16_e64 v{{[0-9]+}}, -|v{{[0-9]+}}| 76; CI: v_mul_f32_e32 {{v[0-9]+}}, 1.0, {{v[0-9]+}} 77define amdgpu_kernel void @v_test_canonicalize_fneg_fabs_var_f16(half addrspace(1)* %out) #1 { 78 %val = load half, half addrspace(1)* %out 79 %val.fabs = call half @llvm.fabs.f16(half %val) 80 %val.fabs.fneg = fsub half -0.0, %val.fabs 81 %canonicalized = call half @llvm.canonicalize.f16(half %val.fabs.fneg) 82 store half %canonicalized, half addrspace(1)* %out 83 ret void 84} 85 86; GCN-LABEL: {{^}}v_test_canonicalize_fneg_var_f16: 87; GFX89: v_max_f16_e64 [[REG:v[0-9]+]], -{{v[0-9]+}}, -{{v[0-9]+}} 88; GFX89: {{flat|global}}_store_short v{{\[[0-9]+:[0-9]+\]}}, [[REG]] 89 90; CI: v_cvt_f32_f16_e64 {{v[0-9]+}}, -{{v[0-9]+}} 91; CI: v_mul_f32_e32 {{v[0-9]+}}, 1.0, {{v[0-9]+}} 92define amdgpu_kernel void @v_test_canonicalize_fneg_var_f16(half addrspace(1)* %out) #1 { 93 %val = load half, half addrspace(1)* %out 94 %val.fneg = fsub half -0.0, %val 95 %canonicalized = call half @llvm.canonicalize.f16(half %val.fneg) 96 store half %canonicalized, half addrspace(1)* %out 97 ret void 98} 99 100; GCN-LABEL: {{^}}v_test_no_denormals_canonicalize_fneg_var_f16: 101; GFX89: v_mul_f16_e32 [[REG:v[0-9]+]], -1.0, v{{[0-9]+}} 102; GFX89: {{flat|global}}_store_short v{{\[[0-9]+:[0-9]+\]}}, [[REG]] 103define amdgpu_kernel void @v_test_no_denormals_canonicalize_fneg_var_f16(half addrspace(1)* %out) #2 { 104 %val = load half, half addrspace(1)* %out 105 %val.fneg = fsub half -0.0, %val 106 %canonicalized = call half @llvm.canonicalize.f16(half %val.fneg) 107 store half %canonicalized, half addrspace(1)* %out 108 ret void 109} 110 111; GCN-LABEL: {{^}}v_test_no_denormals_canonicalize_fneg_fabs_var_f16: 112; GFX89: v_mul_f16_e64 [[REG:v[0-9]+]], -1.0, |v{{[0-9]+}}| 113; GFX89: {{flat|global}}_store_short v{{\[[0-9]+:[0-9]+\]}}, [[REG]] 114 115; CI: v_cvt_f32_f16_e64 {{v[0-9]+}}, -|{{v[0-9]+}}| 116; CI: v_mul_f32_e32 {{v[0-9]+}}, 1.0, {{v[0-9]+}} 117define amdgpu_kernel void @v_test_no_denormals_canonicalize_fneg_fabs_var_f16(half addrspace(1)* %out) #2 { 118 %val = load half, half addrspace(1)* %out 119 %val.fabs = call half @llvm.fabs.f16(half %val) 120 %val.fabs.fneg = fsub half -0.0, %val.fabs 121 %canonicalized = call half @llvm.canonicalize.f16(half %val.fabs.fneg) 122 store half %canonicalized, half addrspace(1)* %out 123 ret void 124} 125 126; GCN-LABEL: {{^}}test_fold_canonicalize_p0_f16: 127; GFX89: v_mov_b32_e32 [[REG:v[0-9]+]], 0{{$}} 128; GFX89: {{flat|global}}_store_short v{{\[[0-9]+:[0-9]+\]}}, [[REG]] 129define amdgpu_kernel void @test_fold_canonicalize_p0_f16(half addrspace(1)* %out) #1 { 130 %canonicalized = call half @llvm.canonicalize.f16(half 0.0) 131 store half %canonicalized, half addrspace(1)* %out 132 ret void 133} 134 135; GCN-LABEL: {{^}}test_fold_canonicalize_n0_f16: 136; GFX89: v_mov_b32_e32 [[REG:v[0-9]+]], 0xffff8000{{$}} 137; GFX89: {{flat|global}}_store_short v{{\[[0-9]+:[0-9]+\]}}, [[REG]] 138define amdgpu_kernel void @test_fold_canonicalize_n0_f16(half addrspace(1)* %out) #1 { 139 %canonicalized = call half @llvm.canonicalize.f16(half -0.0) 140 store half %canonicalized, half addrspace(1)* %out 141 ret void 142} 143 144; GCN-LABEL: {{^}}test_fold_canonicalize_p1_f16: 145; GFX89: v_mov_b32_e32 [[REG:v[0-9]+]], 0x3c00{{$}} 146; GFX89: {{flat|global}}_store_short v{{\[[0-9]+:[0-9]+\]}}, [[REG]] 147define amdgpu_kernel void @test_fold_canonicalize_p1_f16(half addrspace(1)* %out) #1 { 148 %canonicalized = call half @llvm.canonicalize.f16(half 1.0) 149 store half %canonicalized, half addrspace(1)* %out 150 ret void 151} 152 153; GCN-LABEL: {{^}}test_fold_canonicalize_n1_f16: 154; GFX89: v_mov_b32_e32 [[REG:v[0-9]+]], 0xffffbc00{{$}} 155; GFX89: {{flat|global}}_store_short v{{\[[0-9]+:[0-9]+\]}}, [[REG]] 156define amdgpu_kernel void @test_fold_canonicalize_n1_f16(half addrspace(1)* %out) #1 { 157 %canonicalized = call half @llvm.canonicalize.f16(half -1.0) 158 store half %canonicalized, half addrspace(1)* %out 159 ret void 160} 161 162; GCN-LABEL: {{^}}test_fold_canonicalize_literal_f16: 163; GFX89: v_mov_b32_e32 [[REG:v[0-9]+]], 0x4c00{{$}} 164; GFX89: {{flat|global}}_store_short v{{\[[0-9]+:[0-9]+\]}}, [[REG]] 165define amdgpu_kernel void @test_fold_canonicalize_literal_f16(half addrspace(1)* %out) #1 { 166 %canonicalized = call half @llvm.canonicalize.f16(half 16.0) 167 store half %canonicalized, half addrspace(1)* %out 168 ret void 169} 170 171; GCN-LABEL: {{^}}test_default_denormals_fold_canonicalize_denormal0_f16: 172; GFX89: v_mov_b32_e32 [[REG:v[0-9]+]], 0x3ff{{$}} 173; GFX89: {{flat|global}}_store_short v{{\[[0-9]+:[0-9]+\]}}, [[REG]] 174define amdgpu_kernel void @test_default_denormals_fold_canonicalize_denormal0_f16(half addrspace(1)* %out) #1 { 175 %canonicalized = call half @llvm.canonicalize.f16(half 0xH03FF) 176 store half %canonicalized, half addrspace(1)* %out 177 ret void 178} 179 180; GCN-LABEL: {{^}}test_denormals_fold_canonicalize_denormal0_f16: 181; GFX89: v_mov_b32_e32 [[REG:v[0-9]+]], 0x3ff{{$}} 182; GFX89: {{flat|global}}_store_short v{{\[[0-9]+:[0-9]+\]}}, [[REG]] 183define amdgpu_kernel void @test_denormals_fold_canonicalize_denormal0_f16(half addrspace(1)* %out) #3 { 184 %canonicalized = call half @llvm.canonicalize.f16(half 0xH03FF) 185 store half %canonicalized, half addrspace(1)* %out 186 ret void 187} 188 189; GCN-LABEL: {{^}}test_default_denormals_fold_canonicalize_denormal1_f16: 190; GFX89: v_mov_b32_e32 [[REG:v[0-9]+]], 0xffff83ff{{$}} 191; GFX89: {{flat|global}}_store_short v{{\[[0-9]+:[0-9]+\]}}, [[REG]] 192define amdgpu_kernel void @test_default_denormals_fold_canonicalize_denormal1_f16(half addrspace(1)* %out) #1 { 193 %canonicalized = call half @llvm.canonicalize.f16(half 0xH83FF) 194 store half %canonicalized, half addrspace(1)* %out 195 ret void 196} 197 198; GCN-LABEL: {{^}}test_denormals_fold_canonicalize_denormal1_f16: 199; GFX89: v_mov_b32_e32 [[REG:v[0-9]+]], 0xffff83ff{{$}} 200; GFX89: {{flat|global}}_store_short v{{\[[0-9]+:[0-9]+\]}}, [[REG]] 201define amdgpu_kernel void @test_denormals_fold_canonicalize_denormal1_f16(half addrspace(1)* %out) #3 { 202 %canonicalized = call half @llvm.canonicalize.f16(half 0xH83FF) 203 store half %canonicalized, half addrspace(1)* %out 204 ret void 205} 206 207; GCN-LABEL: {{^}}test_fold_canonicalize_qnan_f16: 208; GFX89: v_mov_b32_e32 [[REG:v[0-9]+]], 0x7c00{{$}} 209; GFX89: {{flat|global}}_store_short v{{\[[0-9]+:[0-9]+\]}}, [[REG]] 210define amdgpu_kernel void @test_fold_canonicalize_qnan_f16(half addrspace(1)* %out) #1 { 211 %canonicalized = call half @llvm.canonicalize.f16(half 0xH7C00) 212 store half %canonicalized, half addrspace(1)* %out 213 ret void 214} 215 216; GCN-LABEL: {{^}}test_fold_canonicalize_qnan_value_neg1_f16: 217; GFX89: v_mov_b32_e32 [[REG:v[0-9]+]], 0x7e00{{$}} 218; GFX89: {{flat|global}}_store_short v{{\[[0-9]+:[0-9]+\]}}, [[REG]] 219define amdgpu_kernel void @test_fold_canonicalize_qnan_value_neg1_f16(half addrspace(1)* %out) #1 { 220 %canonicalized = call half @llvm.canonicalize.f16(half bitcast (i16 -1 to half)) 221 store half %canonicalized, half addrspace(1)* %out 222 ret void 223} 224 225; GCN-LABEL: {{^}}test_fold_canonicalize_qnan_value_neg2_f16: 226; GFX89: v_mov_b32_e32 [[REG:v[0-9]+]], 0x7e00{{$}} 227; GFX89: {{flat|global}}_store_short v{{\[[0-9]+:[0-9]+\]}}, [[REG]] 228define amdgpu_kernel void @test_fold_canonicalize_qnan_value_neg2_f16(half addrspace(1)* %out) #1 { 229 %canonicalized = call half @llvm.canonicalize.f16(half bitcast (i16 -2 to half)) 230 store half %canonicalized, half addrspace(1)* %out 231 ret void 232} 233 234; GCN-LABEL: {{^}}test_fold_canonicalize_snan0_value_f16: 235; GFX89: v_mov_b32_e32 [[REG:v[0-9]+]], 0x7e00{{$}} 236; GFX89: {{flat|global}}_store_short v{{\[[0-9]+:[0-9]+\]}}, [[REG]] 237define amdgpu_kernel void @test_fold_canonicalize_snan0_value_f16(half addrspace(1)* %out) #1 { 238 %canonicalized = call half @llvm.canonicalize.f16(half 0xH7C01) 239 store half %canonicalized, half addrspace(1)* %out 240 ret void 241} 242 243; GCN-LABEL: {{^}}test_fold_canonicalize_snan1_value_f16: 244; GFX89: v_mov_b32_e32 [[REG:v[0-9]+]], 0x7e00{{$}} 245; GFX89: {{flat|global}}_store_short v{{\[[0-9]+:[0-9]+\]}}, [[REG]] 246define amdgpu_kernel void @test_fold_canonicalize_snan1_value_f16(half addrspace(1)* %out) #1 { 247 %canonicalized = call half @llvm.canonicalize.f16(half 0xH7DFF) 248 store half %canonicalized, half addrspace(1)* %out 249 ret void 250} 251 252; GCN-LABEL: {{^}}test_fold_canonicalize_snan2_value_f16: 253; GFX89: v_mov_b32_e32 [[REG:v[0-9]+]], 0x7e00{{$}} 254; GFX89: {{flat|global}}_store_short v{{\[[0-9]+:[0-9]+\]}}, [[REG]] 255define amdgpu_kernel void @test_fold_canonicalize_snan2_value_f16(half addrspace(1)* %out) #1 { 256 %canonicalized = call half @llvm.canonicalize.f16(half 0xHFDFF) 257 store half %canonicalized, half addrspace(1)* %out 258 ret void 259} 260 261; GCN-LABEL: {{^}}test_fold_canonicalize_snan3_value_f16: 262; GFX89: v_mov_b32_e32 [[REG:v[0-9]+]], 0x7e00{{$}} 263; GFX89: {{flat|global}}_store_short v{{\[[0-9]+:[0-9]+\]}}, [[REG]] 264define amdgpu_kernel void @test_fold_canonicalize_snan3_value_f16(half addrspace(1)* %out) #1 { 265 %canonicalized = call half @llvm.canonicalize.f16(half 0xHFC01) 266 store half %canonicalized, half addrspace(1)* %out 267 ret void 268} 269 270; GCN-LABEL: {{^}}v_test_canonicalize_var_v2f16: 271; VI-DAG: v_max_f16_sdwa [[REG0:v[0-9]+]], {{v[0-9]+}}, {{v[0-9]+}} dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1 272; VI-DAG: v_max_f16_e32 [[REG1:v[0-9]+]], {{v[0-9]+}}, {{v[0-9]+}} 273; VI-NOT: v_and_b32 274 275; GFX9: v_pk_max_f16 [[REG:v[0-9]+]], {{v[0-9]+}}, {{v[0-9]+$}} 276; GFX9: {{flat|global}}_store_dword v{{\[[0-9]+:[0-9]+\]}}, [[REG]] 277define amdgpu_kernel void @v_test_canonicalize_var_v2f16(<2 x half> addrspace(1)* %out) #1 { 278 %tid = call i32 @llvm.amdgcn.workitem.id.x() 279 %gep = getelementptr <2 x half>, <2 x half> addrspace(1)* %out, i32 %tid 280 %val = load <2 x half>, <2 x half> addrspace(1)* %gep 281 %canonicalized = call <2 x half> @llvm.canonicalize.v2f16(<2 x half> %val) 282 store <2 x half> %canonicalized, <2 x half> addrspace(1)* %out 283 ret void 284} 285 286; GCN-LABEL: {{^}}v_test_canonicalize_fabs_var_v2f16: 287; VI: v_max_f16_sdwa [[REG0:v[0-9]+]], |v{{[0-9]+}}|, |v{{[0-9]+}}| dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1 288; VI: v_max_f16_e64 [[REG1:v[0-9]+]], |v{{[0-9]+}}|, |v{{[0-9]+}}| 289; VI-NOT: 0xffff 290; VI: v_or_b32 291 292; GFX9: v_and_b32_e32 [[ABS:v[0-9]+]], 0x7fff7fff, v{{[0-9]+}} 293; GFX9: v_pk_max_f16 [[REG:v[0-9]+]], [[ABS]], [[ABS]]{{$}} 294; GFX89: {{flat|global}}_store_dword 295define amdgpu_kernel void @v_test_canonicalize_fabs_var_v2f16(<2 x half> addrspace(1)* %out) #1 { 296 %tid = call i32 @llvm.amdgcn.workitem.id.x() 297 %gep = getelementptr <2 x half>, <2 x half> addrspace(1)* %out, i32 %tid 298 %val = load <2 x half>, <2 x half> addrspace(1)* %gep 299 %val.fabs = call <2 x half> @llvm.fabs.v2f16(<2 x half> %val) 300 %canonicalized = call <2 x half> @llvm.canonicalize.v2f16(<2 x half> %val.fabs) 301 store <2 x half> %canonicalized, <2 x half> addrspace(1)* %out 302 ret void 303} 304 305; GCN-LABEL: {{^}}v_test_canonicalize_fneg_fabs_var_v2f16: 306; VI-DAG: v_max_f16_sdwa [[REG0:v[0-9]+]], -|v{{[0-9]+}}|, -|v{{[0-9]+}}| dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1 307; VI-DAG: v_max_f16_e64 [[REG1:v[0-9]+]], -|v{{[0-9]+}}|, -|v{{[0-9]+}}| 308; VI: v_or_b32 309 310; GFX9: v_and_b32_e32 [[ABS:v[0-9]+]], 0x7fff7fff, v{{[0-9]+}} 311; GFX9: v_pk_max_f16 [[REG:v[0-9]+]], [[ABS]], [[ABS]] neg_lo:[1,1] neg_hi:[1,1]{{$}} 312; GFX89: {{flat|global}}_store_dword 313 314; CI: v_cvt_f32_f16 315; CI: v_cvt_f32_f16 316; CI: v_mul_f32_e32 v{{[0-9]+}}, 1.0 317; CI: v_mul_f32_e32 v{{[0-9]+}}, 1.0 318define amdgpu_kernel void @v_test_canonicalize_fneg_fabs_var_v2f16(<2 x half> addrspace(1)* %out) #1 { 319 %tid = call i32 @llvm.amdgcn.workitem.id.x() 320 %gep = getelementptr <2 x half>, <2 x half> addrspace(1)* %out, i32 %tid 321 %val = load <2 x half>, <2 x half> addrspace(1)* %gep 322 %val.fabs = call <2 x half> @llvm.fabs.v2f16(<2 x half> %val) 323 %val.fabs.fneg = fsub <2 x half> <half -0.0, half -0.0>, %val.fabs 324 %canonicalized = call <2 x half> @llvm.canonicalize.v2f16(<2 x half> %val.fabs.fneg) 325 store <2 x half> %canonicalized, <2 x half> addrspace(1)* %out 326 ret void 327} 328 329; GCN-LABEL: {{^}}v_test_canonicalize_fneg_var_v2f16: 330; VI-DAG: v_max_f16_sdwa [[REG1:v[0-9]+]], -v{{[0-9]+}}, -v{{[0-9]+}} dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1 331; VI-DAG: v_max_f16_e64 [[REG0:v[0-9]+]], -v{{[0-9]+}}, -v{{[0-9]+}} 332; VI-NOT: 0xffff 333 334; GFX9: v_pk_max_f16 [[REG:v[0-9]+]], {{v[0-9]+}}, {{v[0-9]+}} neg_lo:[1,1] neg_hi:[1,1]{{$}} 335; GFX9: {{flat|global}}_store_dword v{{\[[0-9]+:[0-9]+\]}}, [[REG]] 336define amdgpu_kernel void @v_test_canonicalize_fneg_var_v2f16(<2 x half> addrspace(1)* %out) #1 { 337 %tid = call i32 @llvm.amdgcn.workitem.id.x() 338 %gep = getelementptr <2 x half>, <2 x half> addrspace(1)* %out, i32 %tid 339 %val = load <2 x half>, <2 x half> addrspace(1)* %gep 340 %fneg.val = fsub <2 x half> <half -0.0, half -0.0>, %val 341 %canonicalized = call <2 x half> @llvm.canonicalize.v2f16(<2 x half> %fneg.val) 342 store <2 x half> %canonicalized, <2 x half> addrspace(1)* %out 343 ret void 344} 345 346; GCN-LABEL: {{^}}s_test_canonicalize_var_v2f16: 347; VI-DAG: v_max_f16_sdwa [[REG0:v[0-9]+]], {{v[0-9]+}}, {{v[0-9]+}} dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD 348; VI-DAG: v_max_f16_e64 [[REG1:v[0-9]+]], {{s[0-9]+}}, {{s[0-9]+}} 349; VI-NOT: v_and_b32 350 351; GFX9: v_pk_max_f16 [[REG:v[0-9]+]], {{s[0-9]+}}, {{s[0-9]+$}} 352; GFX9: {{flat|global}}_store_dword v{{\[[0-9]+:[0-9]+\]}}, [[REG]] 353define amdgpu_kernel void @s_test_canonicalize_var_v2f16(<2 x half> addrspace(1)* %out, i32 zeroext %val.arg) #1 { 354 %val = bitcast i32 %val.arg to <2 x half> 355 %canonicalized = call <2 x half> @llvm.canonicalize.v2f16(<2 x half> %val) 356 store <2 x half> %canonicalized, <2 x half> addrspace(1)* %out 357 ret void 358} 359 360; GCN-LABEL: {{^}}test_fold_canonicalize_p0_v2f16: 361; GFX89: v_mov_b32_e32 [[REG:v[0-9]+]], 0{{$}} 362; GFX89: {{flat|global}}_store_dword v{{\[[0-9]+:[0-9]+\]}}, [[REG]] 363define amdgpu_kernel void @test_fold_canonicalize_p0_v2f16(<2 x half> addrspace(1)* %out) #1 { 364 %canonicalized = call <2 x half> @llvm.canonicalize.v2f16(<2 x half> zeroinitializer) 365 store <2 x half> %canonicalized, <2 x half> addrspace(1)* %out 366 ret void 367} 368 369; GCN-LABEL: {{^}}test_fold_canonicalize_n0_v2f16: 370; GFX89: v_mov_b32_e32 [[REG:v[0-9]+]], 0x80008000{{$}} 371; GFX89: {{flat|global}}_store_dword v{{\[[0-9]+:[0-9]+\]}}, [[REG]] 372define amdgpu_kernel void @test_fold_canonicalize_n0_v2f16(<2 x half> addrspace(1)* %out) #1 { 373 %canonicalized = call <2 x half> @llvm.canonicalize.v2f16(<2 x half> <half -0.0, half -0.0>) 374 store <2 x half> %canonicalized, <2 x half> addrspace(1)* %out 375 ret void 376} 377 378; GCN-LABEL: {{^}}test_fold_canonicalize_p1_v2f16: 379; GFX89: v_mov_b32_e32 [[REG:v[0-9]+]], 0x3c003c00{{$}} 380; GFX89: {{flat|global}}_store_dword v{{\[[0-9]+:[0-9]+\]}}, [[REG]] 381define amdgpu_kernel void @test_fold_canonicalize_p1_v2f16(<2 x half> addrspace(1)* %out) #1 { 382 %canonicalized = call <2 x half> @llvm.canonicalize.v2f16(<2 x half> <half 1.0, half 1.0>) 383 store <2 x half> %canonicalized, <2 x half> addrspace(1)* %out 384 ret void 385} 386 387; GCN-LABEL: {{^}}test_fold_canonicalize_n1_v2f16: 388; GFX89: v_mov_b32_e32 [[REG:v[0-9]+]], 0xbc00bc00{{$}} 389; GFX89: {{flat|global}}_store_dword v{{\[[0-9]+:[0-9]+\]}}, [[REG]] 390define amdgpu_kernel void @test_fold_canonicalize_n1_v2f16(<2 x half> addrspace(1)* %out) #1 { 391 %canonicalized = call <2 x half> @llvm.canonicalize.v2f16(<2 x half> <half -1.0, half -1.0>) 392 store <2 x half> %canonicalized, <2 x half> addrspace(1)* %out 393 ret void 394} 395 396; GCN-LABEL: {{^}}test_fold_canonicalize_literal_v2f16: 397; GFX89: v_mov_b32_e32 [[REG:v[0-9]+]], 0x4c004c00{{$}} 398; GFX89: {{flat|global}}_store_dword v{{\[[0-9]+:[0-9]+\]}}, [[REG]] 399define amdgpu_kernel void @test_fold_canonicalize_literal_v2f16(<2 x half> addrspace(1)* %out) #1 { 400 %canonicalized = call <2 x half> @llvm.canonicalize.v2f16(<2 x half> <half 16.0, half 16.0>) 401 store <2 x half> %canonicalized, <2 x half> addrspace(1)* %out 402 ret void 403} 404 405; GCN-LABEL: {{^}}test_no_denormals_fold_canonicalize_denormal0_v2f16: 406; GFX89: v_mov_b32_e32 [[REG:v[0-9]+]], 0x3ff03ff{{$}} 407; GFX89: {{flat|global}}_store_dword v{{\[[0-9]+:[0-9]+\]}}, [[REG]] 408define amdgpu_kernel void @test_no_denormals_fold_canonicalize_denormal0_v2f16(<2 x half> addrspace(1)* %out) #1 { 409 %canonicalized = call <2 x half> @llvm.canonicalize.v2f16(<2 x half> <half 0xH03FF, half 0xH03FF>) 410 store <2 x half> %canonicalized, <2 x half> addrspace(1)* %out 411 ret void 412} 413 414; GCN-LABEL: {{^}}test_denormals_fold_canonicalize_denormal0_v2f16: 415; GFX89: v_mov_b32_e32 [[REG:v[0-9]+]], 0x3ff03ff{{$}} 416; GFX89: {{flat|global}}_store_dword v{{\[[0-9]+:[0-9]+\]}}, [[REG]] 417define amdgpu_kernel void @test_denormals_fold_canonicalize_denormal0_v2f16(<2 x half> addrspace(1)* %out) #3 { 418 %canonicalized = call <2 x half> @llvm.canonicalize.v2f16(<2 x half> <half 0xH03FF, half 0xH03FF>) 419 store <2 x half> %canonicalized, <2 x half> addrspace(1)* %out 420 ret void 421} 422 423; GCN-LABEL: {{^}}test_no_denormals_fold_canonicalize_denormal1_v2f16: 424; GFX89: v_mov_b32_e32 [[REG:v[0-9]+]], 0x83ff83ff{{$}} 425; GFX89: {{flat|global}}_store_dword v{{\[[0-9]+:[0-9]+\]}}, [[REG]] 426define amdgpu_kernel void @test_no_denormals_fold_canonicalize_denormal1_v2f16(<2 x half> addrspace(1)* %out) #1 { 427 %canonicalized = call <2 x half> @llvm.canonicalize.v2f16(<2 x half> <half 0xH83FF, half 0xH83FF>) 428 store <2 x half> %canonicalized, <2 x half> addrspace(1)* %out 429 ret void 430} 431 432; GCN-LABEL: {{^}}test_denormals_fold_canonicalize_denormal1_v2f16: 433; GFX89: v_mov_b32_e32 [[REG:v[0-9]+]], 0x83ff83ff{{$}} 434; GFX89: {{flat|global}}_store_dword v{{\[[0-9]+:[0-9]+\]}}, [[REG]] 435define amdgpu_kernel void @test_denormals_fold_canonicalize_denormal1_v2f16(<2 x half> addrspace(1)* %out) #3 { 436 %canonicalized = call <2 x half> @llvm.canonicalize.v2f16(<2 x half> <half 0xH83FF, half 0xH83FF>) 437 store <2 x half> %canonicalized, <2 x half> addrspace(1)* %out 438 ret void 439} 440 441; GCN-LABEL: {{^}}test_fold_canonicalize_qnan_v2f16: 442; GFX89: v_mov_b32_e32 [[REG:v[0-9]+]], 0x7c007c00{{$}} 443; GFX89: {{flat|global}}_store_dword v{{\[[0-9]+:[0-9]+\]}}, [[REG]] 444define amdgpu_kernel void @test_fold_canonicalize_qnan_v2f16(<2 x half> addrspace(1)* %out) #1 { 445 %canonicalized = call <2 x half> @llvm.canonicalize.v2f16(<2 x half> <half 0xH7C00, half 0xH7C00>) 446 store <2 x half> %canonicalized, <2 x half> addrspace(1)* %out 447 ret void 448} 449 450; GCN-LABEL: {{^}}test_fold_canonicalize_qnan_value_neg1_v2f16: 451; GFX89: v_mov_b32_e32 [[REG:v[0-9]+]], 0x7e007e00{{$}} 452; GFX89: {{flat|global}}_store_dword v{{\[[0-9]+:[0-9]+\]}}, [[REG]] 453define amdgpu_kernel void @test_fold_canonicalize_qnan_value_neg1_v2f16(<2 x half> addrspace(1)* %out) #1 { 454 %canonicalized = call <2 x half> @llvm.canonicalize.v2f16(<2 x half> bitcast (i32 -1 to <2 x half>)) 455 store <2 x half> %canonicalized, <2 x half> addrspace(1)* %out 456 ret void 457} 458 459; GCN-LABEL: {{^}}test_fold_canonicalize_qnan_value_neg2_v2f16: 460; GFX89: v_mov_b32_e32 [[REG:v[0-9]+]], 0x7e007e00{{$}} 461; GFX89: {{flat|global}}_store_dword v{{\[[0-9]+:[0-9]+\]}}, [[REG]] 462define amdgpu_kernel void @test_fold_canonicalize_qnan_value_neg2_v2f16(<2 x half> addrspace(1)* %out) #1 { 463 %canonicalized = call <2 x half> @llvm.canonicalize.v2f16(<2 x half> <half bitcast (i16 -2 to half), half bitcast (i16 -2 to half)>) 464 store <2 x half> %canonicalized, <2 x half> addrspace(1)* %out 465 ret void 466} 467 468; GCN-LABEL: {{^}}test_fold_canonicalize_snan0_value_v2f16: 469; GFX89: v_mov_b32_e32 [[REG:v[0-9]+]], 0x7e007e00{{$}} 470; GFX89: {{flat|global}}_store_dword v{{\[[0-9]+:[0-9]+\]}}, [[REG]] 471define amdgpu_kernel void @test_fold_canonicalize_snan0_value_v2f16(<2 x half> addrspace(1)* %out) #1 { 472 %canonicalized = call <2 x half> @llvm.canonicalize.v2f16(<2 x half> <half 0xH7C01, half 0xH7C01>) 473 store <2 x half> %canonicalized, <2 x half> addrspace(1)* %out 474 ret void 475} 476 477; GCN-LABEL: {{^}}test_fold_canonicalize_snan1_value_v2f16: 478; GFX89: v_mov_b32_e32 [[REG:v[0-9]+]], 0x7e007e00{{$}} 479; GFX89: {{flat|global}}_store_dword v{{\[[0-9]+:[0-9]+\]}}, [[REG]] 480define amdgpu_kernel void @test_fold_canonicalize_snan1_value_v2f16(<2 x half> addrspace(1)* %out) #1 { 481 %canonicalized = call <2 x half> @llvm.canonicalize.v2f16(<2 x half> <half 0xH7DFF, half 0xH7DFF>) 482 store <2 x half> %canonicalized, <2 x half> addrspace(1)* %out 483 ret void 484} 485 486; GCN-LABEL: {{^}}test_fold_canonicalize_snan2_value_v2f16: 487; GFX89: v_mov_b32_e32 [[REG:v[0-9]+]], 0x7e007e00{{$}} 488; GFX89: {{flat|global}}_store_dword v{{\[[0-9]+:[0-9]+\]}}, [[REG]] 489define amdgpu_kernel void @test_fold_canonicalize_snan2_value_v2f16(<2 x half> addrspace(1)* %out) #1 { 490 %canonicalized = call <2 x half> @llvm.canonicalize.v2f16(<2 x half> <half 0xHFDFF, half 0xHFDFF>) 491 store <2 x half> %canonicalized, <2 x half> addrspace(1)* %out 492 ret void 493} 494 495; GCN-LABEL: {{^}}test_fold_canonicalize_snan3_value_v2f16: 496; GFX89: v_mov_b32_e32 [[REG:v[0-9]+]], 0x7e007e00{{$}} 497; GFX89: {{flat|global}}_store_dword v{{\[[0-9]+:[0-9]+\]}}, [[REG]] 498define amdgpu_kernel void @test_fold_canonicalize_snan3_value_v2f16(<2 x half> addrspace(1)* %out) #1 { 499 %canonicalized = call <2 x half> @llvm.canonicalize.v2f16(<2 x half> <half 0xHFC01, half 0xHFC01>) 500 store <2 x half> %canonicalized, <2 x half> addrspace(1)* %out 501 ret void 502} 503 504; FIXME: Extra 4th component handled 505; GCN-LABEL: {{^}}v_test_canonicalize_var_v3f16: 506; GFX9: s_waitcnt 507; GFX9-NEXT: v_pk_max_f16 v0, v0, v0 508; GFX9-NEXT: v_pk_max_f16 v1, v1, v1 509; GFX9-NEXT: s_setpc_b64 510 511; VI-DAG: v_max_f16_sdwa [[CANON_ELT1:v[0-9]+]], v0, v0 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1 512; VI-DAG: v_max_f16_e32 [[CANON_ELT0:v[0-9]+]], v0, v0 513; VI-DAG: v_max_f16_e32 v1, v1, v1 514; VI-DAG: v_or_b32_e32 v0, [[CANON_ELT0]], [[CANON_ELT1]] 515 516; VI: s_setpc_b64 517define <3 x half> @v_test_canonicalize_var_v3f16(<3 x half> %val) #1 { 518 %canonicalized = call <3 x half> @llvm.canonicalize.v3f16(<3 x half> %val) 519 ret <3 x half> %canonicalized 520} 521 522; GCN-LABEL: {{^}}v_test_canonicalize_var_v4f16: 523; GFX9: s_waitcnt 524; GFX9-NEXT: v_pk_max_f16 v0, v0, v0 525; GFX9-NEXT: v_pk_max_f16 v1, v1, v1 526; GFX9-NEXT: s_setpc_b64 527 528; VI-DAG: v_max_f16_sdwa [[CANON_ELT3:v[0-9]+]], v1, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1 529; VI-DAG: v_max_f16_e32 [[CANON_ELT2:v[0-9]+]], v1, v1 530; VI-DAG: v_max_f16_sdwa [[CANON_ELT1:v[0-9]+]], v0, v0 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1 531; VI-DAG: v_max_f16_e32 [[CANON_ELT0:v[0-9]+]], v0, v0 532; VI-DAG: v_or_b32_e32 v0, [[CANON_ELT0]], [[CANON_ELT1]] 533; VI-DAG: v_or_b32_e32 v1, [[CANON_ELT2]], [[CANON_ELT3]] 534; VI: s_setpc_b64 535define <4 x half> @v_test_canonicalize_var_v4f16(<4 x half> %val) #1 { 536 %canonicalized = call <4 x half> @llvm.canonicalize.v4f16(<4 x half> %val) 537 ret <4 x half> %canonicalized 538} 539 540; GCN-LABEL: {{^}}s_test_canonicalize_undef_v2f16: 541; GCN: v_mov_b32_e32 [[REG:v[0-9]+]], 0x7e007e00 542; GFX89: {{flat|global}}_store_dword v{{\[[0-9]+:[0-9]+\]}}, [[REG]] 543define amdgpu_kernel void @s_test_canonicalize_undef_v2f16(<2 x half> addrspace(1)* %out) #1 { 544 %canonicalized = call <2 x half> @llvm.canonicalize.v2f16(<2 x half> undef) 545 store <2 x half> %canonicalized, <2 x half> addrspace(1)* %out 546 ret void 547} 548 549; GCN-LABEL: {{^}}v_test_canonicalize_reg_undef_v2f16: 550; GFX9: s_waitcnt 551; GFX9-NEXT: v_max_f16_e32 v0, v0, v0 552; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v0 553; GFX9-NEXT: s_setpc_b64 554 555; High bits known zero 556; FIXME: Should also be true on gfx9 by default? 557; VI: s_waitcnt 558; VI-NEXT: v_max_f16_e32 v0, v0, v0 559; VI-NEXT: s_setpc_b64 560define <2 x half> @v_test_canonicalize_reg_undef_v2f16(half %val) #1 { 561 %vec = insertelement <2 x half> undef, half %val, i32 0 562 %canonicalized = call <2 x half> @llvm.canonicalize.v2f16(<2 x half> %vec) 563 ret <2 x half> %canonicalized 564} 565 566; GCN-LABEL: {{^}}v_test_canonicalize_undef_reg_v2f16: 567; GFX89: s_waitcnt 568; GFX89-NEXT: v_max_f16_sdwa v0, v0, v0 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD 569; GFX89-NEXT: s_setpc_b64 570define <2 x half> @v_test_canonicalize_undef_reg_v2f16(half %val) #1 { 571 %vec = insertelement <2 x half> undef, half %val, i32 1 572 %canonicalized = call <2 x half> @llvm.canonicalize.v2f16(<2 x half> %vec) 573 ret <2 x half> %canonicalized 574} 575 576; GCN-LABEL: {{^}}v_test_canonicalize_undef_lo_imm_hi_v2f16: 577; GCN: s_waitcnt 578; GFX89-NEXT: v_mov_b32_e32 v0, 0x3c003c00 579; GFX89-NEXT: s_setpc_b64 580 581; CI-NEXT: v_mov_b32_e32 v0, 0x7fc00000 582; CI-NEXT: v_mov_b32_e32 v1, 1.0 583; CI-NEXT: s_setpc_b64 584define <2 x half> @v_test_canonicalize_undef_lo_imm_hi_v2f16() #1 { 585 %vec = insertelement <2 x half> undef, half 1.0, i32 1 586 %canonicalized = call <2 x half> @llvm.canonicalize.v2f16(<2 x half> %vec) 587 ret <2 x half> %canonicalized 588} 589 590; GCN-LABEL: {{^}}v_test_canonicalize_imm_lo_undef_hi_v2f16: 591; GCN: s_waitcnt 592; GFX89-NEXT: v_mov_b32_e32 v0, 0x3c003c00 593; GFX89-NEXT: s_setpc_b64 594 595; CI-NEXT: v_mov_b32_e32 v0, 1.0 596; CI-NEXT: v_mov_b32_e32 v1, 0x7fc00000 597; CI-NEXT: s_setpc_b64 598define <2 x half> @v_test_canonicalize_imm_lo_undef_hi_v2f16() #1 { 599 %vec = insertelement <2 x half> undef, half 1.0, i32 0 600 %canonicalized = call <2 x half> @llvm.canonicalize.v2f16(<2 x half> %vec) 601 ret <2 x half> %canonicalized 602} 603 604; GCN-LABEL: {{^}}v_test_canonicalize_undef_lo_k_hi_v2f16: 605; GCN: s_waitcnt 606; GFX89-NEXT: v_mov_b32_e32 v0, 0x4c004c00 607; GFX89-NEXT: s_setpc_b64 608 609; CI-NEXT: v_mov_b32_e32 v0, 0x7fc00000 610; CI-NEXT: v_mov_b32_e32 v1, 0x41800000 611; CI-NEXT: s_setpc_b64 612define <2 x half> @v_test_canonicalize_undef_lo_k_hi_v2f16() #1 { 613 %vec = insertelement <2 x half> undef, half 16.0, i32 1 614 %canonicalized = call <2 x half> @llvm.canonicalize.v2f16(<2 x half> %vec) 615 ret <2 x half> %canonicalized 616} 617 618; GCN-LABEL: {{^}}v_test_canonicalize_k_lo_undef_hi_v2f16: 619; GCN: s_waitcnt 620; GFX89-NEXT: v_mov_b32_e32 v0, 0x4c004c00 621; GFX89-NEXT: s_setpc_b64 622 623; CI-NEXT: v_mov_b32_e32 v0, 0x41800000 624; CI-NEXT: v_mov_b32_e32 v1, 0x7fc00000 625; CI-NEXT: s_setpc_b64 626define <2 x half> @v_test_canonicalize_k_lo_undef_hi_v2f16() #1 { 627 %vec = insertelement <2 x half> undef, half 16.0, i32 0 628 %canonicalized = call <2 x half> @llvm.canonicalize.v2f16(<2 x half> %vec) 629 ret <2 x half> %canonicalized 630} 631 632; GCN-LABEL: {{^}}v_test_canonicalize_reg_k_v2f16: 633; GFX9: s_waitcnt 634; GFX9-DAG: v_max_f16_e32 v0, v0, v0 635; GFX9-DAG: s_movk_i32 [[K:s[0-9]+]], 0x4000 636; GFX9: v_and_b32_e32 v0, 0xffff, v0 637; GFX9: v_lshl_or_b32 v0, [[K]], 16, v0 638; GFX9: s_setpc_b64 639 640; VI: s_waitcnt 641; VI-NEXT: v_max_f16_e32 v0, v0, v0 642; VI-NEXT: v_or_b32_e32 v0, 2.0, v0 643; VI-NEXT: s_setpc_b64 644define <2 x half> @v_test_canonicalize_reg_k_v2f16(half %val) #1 { 645 %vec0 = insertelement <2 x half> undef, half %val, i32 0 646 %vec1 = insertelement <2 x half> %vec0, half 2.0, i32 1 647 %canonicalized = call <2 x half> @llvm.canonicalize.v2f16(<2 x half> %vec1) 648 ret <2 x half> %canonicalized 649} 650 651; GCN-LABEL: {{^}}v_test_canonicalize_k_reg_v2f16: 652; GFX9: v_max_f16_e32 v0, v0, v0 653; GFX9: v_mov_b32_e32 [[K:v[0-9]+]], 0x4000 654; GFX9: v_lshl_or_b32 v0, v0, 16, [[K]] 655; GFX9: s_setpc_b64 656 657; VI: s_waitcnt 658; VI-NEXT: v_max_f16_sdwa v0, v0, v0 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD 659; VI-NEXT: v_or_b32_e32 v0, 0x4000, v0 660; VI-NEXT: s_setpc_b64 661define <2 x half> @v_test_canonicalize_k_reg_v2f16(half %val) #1 { 662 %vec0 = insertelement <2 x half> undef, half 2.0, i32 0 663 %vec1 = insertelement <2 x half> %vec0, half %val, i32 1 664 %canonicalized = call <2 x half> @llvm.canonicalize.v2f16(<2 x half> %vec1) 665 ret <2 x half> %canonicalized 666} 667 668; GCN-LABEL: {{^}}s_test_canonicalize_undef_v4f16: 669; GCN: v_mov_b32_e32 v0, 0x7e007e00 670; GCN: v_mov_b32_e32 v1, v0 671define amdgpu_kernel void @s_test_canonicalize_undef_v4f16(<4 x half> addrspace(1)* %out) #1 { 672 %canonicalized = call <4 x half> @llvm.canonicalize.v4f16(<4 x half> undef) 673 store <4 x half> %canonicalized, <4 x half> addrspace(1)* %out 674 ret void 675} 676 677; GCN-LABEL: {{^}}v_test_canonicalize_reg_undef_undef_undef_v4f16: 678; GFX9: s_waitcnt 679; GFX9-NEXT: v_max_f16_e32 v0, v0, v0 680; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v0 681; GFX9-NEXT: v_mov_b32_e32 v1, 0 682; GFX9-NEXT: s_setpc_b64 683 684; VI: s_waitcnt 685; VI-NEXT: v_max_f16_e32 v0, v0, v0 686; VI-NEXT: v_or_b32_e32 v0, 0x7e000000, v0 687; VI-NEXT: v_mov_b32_e32 v1, 0x7e007e00 688; VI-NEXT: s_setpc_b64 689define <4 x half> @v_test_canonicalize_reg_undef_undef_undef_v4f16(half %val) #1 { 690 %vec = insertelement <4 x half> undef, half %val, i32 0 691 %canonicalized = call <4 x half> @llvm.canonicalize.v4f16(<4 x half> %vec) 692 ret <4 x half> %canonicalized 693} 694 695; GCN-LABEL: {{^}}v_test_canonicalize_reg_reg_undef_undef_v4f16: 696; GFX9: s_waitcnt 697; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v0 698; GFX9-NEXT: v_lshl_or_b32 v0, v1, 16, v0 699; GFX9-NEXT: v_pk_max_f16 v0, v0, v0 700; GFX9-NEXT: v_mov_b32_e32 v1, 0 701; GFX9-NEXT: s_setpc_b64 702 703; VI: s_waitcnt 704; VI-DAG: v_max_f16_e32 v0, v0, v0 705; VI-DAG: v_max_f16_sdwa v1, v1, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD 706; VI: v_or_b32_e32 v0, v0, v1 707; VI-NEXT: v_mov_b32_e32 v1, 0x7e007e00 708; VI-NEXT: s_setpc_b64 709define <4 x half> @v_test_canonicalize_reg_reg_undef_undef_v4f16(half %val0, half %val1) #1 { 710 %vec0 = insertelement <4 x half> undef, half %val0, i32 0 711 %vec1 = insertelement <4 x half> %vec0, half %val1, i32 1 712 %canonicalized = call <4 x half> @llvm.canonicalize.v4f16(<4 x half> %vec1) 713 ret <4 x half> %canonicalized 714} 715 716; GCN-LABEL: {{^}}v_test_canonicalize_reg_undef_reg_reg_v4f16: 717; GFX9: s_waitcnt 718; GFX9-NEXT: v_mov_b32_e32 [[MASK:v[0-9]+]], 0xffff 719; GFX9-NEXT: v_and_b32_e32 v1, [[MASK]], v1 720; GFX9-NEXT: v_max_f16_e32 v0, v0, v0 721; GFX9-NEXT: v_lshl_or_b32 v1, v2, 16, v1 722; GFX9-NEXT: v_and_b32_e32 v0, [[MASK]], v0 723; GFX9-NEXT: v_pk_max_f16 v1, v1, v1 724; GFX9-NEXT: s_setpc_b64 725 726; VI: s_waitcnt 727; VI-NEXT: v_max_f16_e32 v0, v0, v0 728; VI-NEXT: v_max_f16_e32 v1, v1, v1 729; VI-NEXT: v_max_f16_sdwa v2, v2, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD 730; VI-NEXT: v_or_b32_e32 v0, 0x7e000000, v0 731; VI-NEXT: v_or_b32_e32 v1, v1, v2 732; VI-NEXT: s_setpc_b64 733define <4 x half> @v_test_canonicalize_reg_undef_reg_reg_v4f16(half %val0, half %val1, half %val2) #1 { 734 %vec0 = insertelement <4 x half> undef, half %val0, i32 0 735 %vec1 = insertelement <4 x half> %vec0, half %val1, i32 2 736 %vec2 = insertelement <4 x half> %vec1, half %val2, i32 3 737 %canonicalized = call <4 x half> @llvm.canonicalize.v4f16(<4 x half> %vec2) 738 ret <4 x half> %canonicalized 739} 740 741attributes #0 = { nounwind readnone } 742attributes #1 = { nounwind "denormal-fp-math-f32"="preserve-sign,preserve-sign" } 743attributes #2 = { nounwind "denormal-fp-math"="preserve-sign,preserve-sign" } 744attributes #3 = { nounwind "denormal-fp-math-f32"="preserve-sign,preserve-sign" } 745