1; RUN: llc -amdgpu-scalarize-global-loads=false -march=amdgcn -verify-machineinstrs < %s | FileCheck -check-prefix=GCN %s 2 3declare float @llvm.fabs.f32(float) #0 4declare float @llvm.canonicalize.f32(float) #0 5declare double @llvm.fabs.f64(double) #0 6declare double @llvm.canonicalize.f64(double) #0 7declare half @llvm.canonicalize.f16(half) #0 8declare <2 x half> @llvm.canonicalize.v2f16(<2 x half>) #0 9declare i32 @llvm.amdgcn.workitem.id.x() #0 10 11; GCN-LABEL: {{^}}v_test_canonicalize_var_f32: 12; GCN: v_mul_f32_e32 [[REG:v[0-9]+]], 1.0, {{v[0-9]+}} 13; GCN: buffer_store_dword [[REG]] 14define amdgpu_kernel void @v_test_canonicalize_var_f32(float addrspace(1)* %out) #1 { 15 %val = load float, float addrspace(1)* %out 16 %canonicalized = call float @llvm.canonicalize.f32(float %val) 17 store float %canonicalized, float addrspace(1)* %out 18 ret void 19} 20 21; GCN-LABEL: {{^}}s_test_canonicalize_var_f32: 22; GCN: v_mul_f32_e64 [[REG:v[0-9]+]], 1.0, {{s[0-9]+}} 23; GCN: buffer_store_dword [[REG]] 24define amdgpu_kernel void @s_test_canonicalize_var_f32(float addrspace(1)* %out, float %val) #1 { 25 %canonicalized = call float @llvm.canonicalize.f32(float %val) 26 store float %canonicalized, float addrspace(1)* %out 27 ret void 28} 29 30; GCN-LABEL: {{^}}v_test_canonicalize_fabs_var_f32: 31; GCN: v_mul_f32_e64 [[REG:v[0-9]+]], 1.0, |{{v[0-9]+}}| 32; GCN: buffer_store_dword [[REG]] 33define amdgpu_kernel void @v_test_canonicalize_fabs_var_f32(float addrspace(1)* %out) #1 { 34 %val = load float, float addrspace(1)* %out 35 %val.fabs = call float @llvm.fabs.f32(float %val) 36 %canonicalized = call float @llvm.canonicalize.f32(float %val.fabs) 37 store float %canonicalized, float addrspace(1)* %out 38 ret void 39} 40 41; GCN-LABEL: {{^}}v_test_canonicalize_fneg_fabs_var_f32: 42; GCN: v_mul_f32_e64 [[REG:v[0-9]+]], 1.0, -|{{v[0-9]+}}| 43; GCN: buffer_store_dword [[REG]] 44define amdgpu_kernel void @v_test_canonicalize_fneg_fabs_var_f32(float addrspace(1)* %out) #1 { 45 %val = load float, float addrspace(1)* %out 46 %val.fabs = call float @llvm.fabs.f32(float %val) 47 %val.fabs.fneg = fsub float -0.0, %val.fabs 48 %canonicalized = call float @llvm.canonicalize.f32(float %val.fabs.fneg) 49 store float %canonicalized, float addrspace(1)* %out 50 ret void 51} 52 53; GCN-LABEL: {{^}}v_test_canonicalize_fneg_var_f32: 54; GCN: v_mul_f32_e64 [[REG:v[0-9]+]], 1.0, -{{v[0-9]+}} 55; GCN: buffer_store_dword [[REG]] 56define amdgpu_kernel void @v_test_canonicalize_fneg_var_f32(float addrspace(1)* %out) #1 { 57 %val = load float, float addrspace(1)* %out 58 %val.fneg = fsub float -0.0, %val 59 %canonicalized = call float @llvm.canonicalize.f32(float %val.fneg) 60 store float %canonicalized, float addrspace(1)* %out 61 ret void 62} 63 64; GCN-LABEL: {{^}}test_fold_canonicalize_p0_f32: 65; GCN: v_mov_b32_e32 [[REG:v[0-9]+]], 0{{$}} 66; GCN: buffer_store_dword [[REG]] 67define amdgpu_kernel void @test_fold_canonicalize_p0_f32(float addrspace(1)* %out) #1 { 68 %canonicalized = call float @llvm.canonicalize.f32(float 0.0) 69 store float %canonicalized, float addrspace(1)* %out 70 ret void 71} 72 73; GCN-LABEL: {{^}}test_fold_canonicalize_n0_f32: 74; GCN: v_bfrev_b32_e32 [[REG:v[0-9]+]], 1{{$}} 75; GCN: buffer_store_dword [[REG]] 76define amdgpu_kernel void @test_fold_canonicalize_n0_f32(float addrspace(1)* %out) #1 { 77 %canonicalized = call float @llvm.canonicalize.f32(float -0.0) 78 store float %canonicalized, float addrspace(1)* %out 79 ret void 80} 81 82; GCN-LABEL: {{^}}test_fold_canonicalize_p1_f32: 83; GCN: v_mov_b32_e32 [[REG:v[0-9]+]], 1.0{{$}} 84; GCN: buffer_store_dword [[REG]] 85define amdgpu_kernel void @test_fold_canonicalize_p1_f32(float addrspace(1)* %out) #1 { 86 %canonicalized = call float @llvm.canonicalize.f32(float 1.0) 87 store float %canonicalized, float addrspace(1)* %out 88 ret void 89} 90 91; GCN-LABEL: {{^}}test_fold_canonicalize_n1_f32: 92; GCN: v_mov_b32_e32 [[REG:v[0-9]+]], -1.0{{$}} 93; GCN: buffer_store_dword [[REG]] 94define amdgpu_kernel void @test_fold_canonicalize_n1_f32(float addrspace(1)* %out) #1 { 95 %canonicalized = call float @llvm.canonicalize.f32(float -1.0) 96 store float %canonicalized, float addrspace(1)* %out 97 ret void 98} 99 100; GCN-LABEL: {{^}}test_fold_canonicalize_literal_f32: 101; GCN: v_mov_b32_e32 [[REG:v[0-9]+]], 0x41800000{{$}} 102; GCN: buffer_store_dword [[REG]] 103define amdgpu_kernel void @test_fold_canonicalize_literal_f32(float addrspace(1)* %out) #1 { 104 %canonicalized = call float @llvm.canonicalize.f32(float 16.0) 105 store float %canonicalized, float addrspace(1)* %out 106 ret void 107} 108 109; GCN-LABEL: {{^}}test_no_denormals_fold_canonicalize_denormal0_f32: 110; GCN: v_mov_b32_e32 [[REG:v[0-9]+]], 0{{$}} 111; GCN: buffer_store_dword [[REG]] 112define amdgpu_kernel void @test_no_denormals_fold_canonicalize_denormal0_f32(float addrspace(1)* %out) #1 { 113 %canonicalized = call float @llvm.canonicalize.f32(float bitcast (i32 8388607 to float)) 114 store float %canonicalized, float addrspace(1)* %out 115 ret void 116} 117 118; GCN-LABEL: {{^}}test_denormals_fold_canonicalize_denormal0_f32: 119; GCN: v_mov_b32_e32 [[REG:v[0-9]+]], 0x7fffff{{$}} 120; GCN: buffer_store_dword [[REG]] 121define amdgpu_kernel void @test_denormals_fold_canonicalize_denormal0_f32(float addrspace(1)* %out) #3 { 122 %canonicalized = call float @llvm.canonicalize.f32(float bitcast (i32 8388607 to float)) 123 store float %canonicalized, float addrspace(1)* %out 124 ret void 125} 126 127; GCN-LABEL: {{^}}test_no_denormals_fold_canonicalize_denormal1_f32: 128; GCN: v_mov_b32_e32 [[REG:v[0-9]+]], 0{{$}} 129; GCN: buffer_store_dword [[REG]] 130define amdgpu_kernel void @test_no_denormals_fold_canonicalize_denormal1_f32(float addrspace(1)* %out) #1 { 131 %canonicalized = call float @llvm.canonicalize.f32(float bitcast (i32 2155872255 to float)) 132 store float %canonicalized, float addrspace(1)* %out 133 ret void 134} 135 136; GCN-LABEL: {{^}}test_denormals_fold_canonicalize_denormal1_f32: 137; GCN: v_mov_b32_e32 [[REG:v[0-9]+]], 0x807fffff{{$}} 138; GCN: buffer_store_dword [[REG]] 139define amdgpu_kernel void @test_denormals_fold_canonicalize_denormal1_f32(float addrspace(1)* %out) #3 { 140 %canonicalized = call float @llvm.canonicalize.f32(float bitcast (i32 2155872255 to float)) 141 store float %canonicalized, float addrspace(1)* %out 142 ret void 143} 144 145; GCN-LABEL: {{^}}test_fold_canonicalize_qnan_f32: 146; GCN: v_mov_b32_e32 [[REG:v[0-9]+]], 0x7fc00000{{$}} 147; GCN: buffer_store_dword [[REG]] 148define amdgpu_kernel void @test_fold_canonicalize_qnan_f32(float addrspace(1)* %out) #1 { 149 %canonicalized = call float @llvm.canonicalize.f32(float 0x7FF8000000000000) 150 store float %canonicalized, float addrspace(1)* %out 151 ret void 152} 153 154; GCN-LABEL: {{^}}test_fold_canonicalize_qnan_value_neg1_f32: 155; GCN: v_mov_b32_e32 [[REG:v[0-9]+]], 0x7fc00000{{$}} 156; GCN: buffer_store_dword [[REG]] 157define amdgpu_kernel void @test_fold_canonicalize_qnan_value_neg1_f32(float addrspace(1)* %out) #1 { 158 %canonicalized = call float @llvm.canonicalize.f32(float bitcast (i32 -1 to float)) 159 store float %canonicalized, float addrspace(1)* %out 160 ret void 161} 162 163; GCN-LABEL: {{^}}test_fold_canonicalize_qnan_value_neg2_f32: 164; GCN: v_mov_b32_e32 [[REG:v[0-9]+]], 0x7fc00000{{$}} 165; GCN: buffer_store_dword [[REG]] 166define amdgpu_kernel void @test_fold_canonicalize_qnan_value_neg2_f32(float addrspace(1)* %out) #1 { 167 %canonicalized = call float @llvm.canonicalize.f32(float bitcast (i32 -2 to float)) 168 store float %canonicalized, float addrspace(1)* %out 169 ret void 170} 171 172; GCN-LABEL: {{^}}test_fold_canonicalize_snan0_value_f32: 173; GCN: v_mov_b32_e32 [[REG:v[0-9]+]], 0x7fc00000{{$}} 174; GCN: buffer_store_dword [[REG]] 175define amdgpu_kernel void @test_fold_canonicalize_snan0_value_f32(float addrspace(1)* %out) #1 { 176 %canonicalized = call float @llvm.canonicalize.f32(float bitcast (i32 2139095041 to float)) 177 store float %canonicalized, float addrspace(1)* %out 178 ret void 179} 180 181; GCN-LABEL: {{^}}test_fold_canonicalize_snan1_value_f32: 182; GCN: v_mov_b32_e32 [[REG:v[0-9]+]], 0x7fc00000{{$}} 183; GCN: buffer_store_dword [[REG]] 184define amdgpu_kernel void @test_fold_canonicalize_snan1_value_f32(float addrspace(1)* %out) #1 { 185 %canonicalized = call float @llvm.canonicalize.f32(float bitcast (i32 2143289343 to float)) 186 store float %canonicalized, float addrspace(1)* %out 187 ret void 188} 189 190; GCN-LABEL: {{^}}test_fold_canonicalize_snan2_value_f32: 191; GCN: v_mov_b32_e32 [[REG:v[0-9]+]], 0x7fc00000{{$}} 192; GCN: buffer_store_dword [[REG]] 193define amdgpu_kernel void @test_fold_canonicalize_snan2_value_f32(float addrspace(1)* %out) #1 { 194 %canonicalized = call float @llvm.canonicalize.f32(float bitcast (i32 4286578689 to float)) 195 store float %canonicalized, float addrspace(1)* %out 196 ret void 197} 198 199; GCN-LABEL: {{^}}test_fold_canonicalize_snan3_value_f32: 200; GCN: v_mov_b32_e32 [[REG:v[0-9]+]], 0x7fc00000{{$}} 201; GCN: buffer_store_dword [[REG]] 202define amdgpu_kernel void @test_fold_canonicalize_snan3_value_f32(float addrspace(1)* %out) #1 { 203 %canonicalized = call float @llvm.canonicalize.f32(float bitcast (i32 4290772991 to float)) 204 store float %canonicalized, float addrspace(1)* %out 205 ret void 206} 207 208; GCN-LABEL: {{^}}v_test_canonicalize_var_f64: 209; GCN: v_max_f64 [[REG:v\[[0-9]+:[0-9]+\]]], {{v\[[0-9]+:[0-9]+\]}}, {{v\[[0-9]+:[0-9]+\]}} 210; GCN: buffer_store_dwordx2 [[REG]] 211define amdgpu_kernel void @v_test_canonicalize_var_f64(double addrspace(1)* %out) #1 { 212 %val = load double, double addrspace(1)* %out 213 %canonicalized = call double @llvm.canonicalize.f64(double %val) 214 store double %canonicalized, double addrspace(1)* %out 215 ret void 216} 217 218; GCN-LABEL: {{^}}s_test_canonicalize_var_f64: 219; GCN: v_max_f64 [[REG:v\[[0-9]+:[0-9]+\]]], {{s\[[0-9]+:[0-9]+\]}}, {{s\[[0-9]+:[0-9]+\]}} 220; GCN: buffer_store_dwordx2 [[REG]] 221define amdgpu_kernel void @s_test_canonicalize_var_f64(double addrspace(1)* %out, double %val) #1 { 222 %canonicalized = call double @llvm.canonicalize.f64(double %val) 223 store double %canonicalized, double addrspace(1)* %out 224 ret void 225} 226 227; GCN-LABEL: {{^}}v_test_canonicalize_fabs_var_f64: 228; GCN: v_max_f64 [[REG:v\[[0-9]+:[0-9]+\]]], |{{v\[[0-9]+:[0-9]+\]}}|, |{{v\[[0-9]+:[0-9]+\]}}| 229; GCN: buffer_store_dwordx2 [[REG]] 230define amdgpu_kernel void @v_test_canonicalize_fabs_var_f64(double addrspace(1)* %out) #1 { 231 %val = load double, double addrspace(1)* %out 232 %val.fabs = call double @llvm.fabs.f64(double %val) 233 %canonicalized = call double @llvm.canonicalize.f64(double %val.fabs) 234 store double %canonicalized, double addrspace(1)* %out 235 ret void 236} 237 238; GCN-LABEL: {{^}}v_test_canonicalize_fneg_fabs_var_f64: 239; GCN: v_max_f64 [[REG:v\[[0-9]+:[0-9]\]]], -|{{v\[[0-9]+:[0-9]+\]}}|, -|{{v\[[0-9]+:[0-9]+\]}}| 240; GCN: buffer_store_dwordx2 [[REG]] 241define amdgpu_kernel void @v_test_canonicalize_fneg_fabs_var_f64(double addrspace(1)* %out) #1 { 242 %val = load double, double addrspace(1)* %out 243 %val.fabs = call double @llvm.fabs.f64(double %val) 244 %val.fabs.fneg = fsub double -0.0, %val.fabs 245 %canonicalized = call double @llvm.canonicalize.f64(double %val.fabs.fneg) 246 store double %canonicalized, double addrspace(1)* %out 247 ret void 248} 249 250; GCN-LABEL: {{^}}v_test_canonicalize_fneg_var_f64: 251; GCN: v_max_f64 [[REG:v\[[0-9]+:[0-9]+\]]], -{{v\[[0-9]+:[0-9]+\]}}, -{{v\[[0-9]+:[0-9]+\]}} 252; GCN: buffer_store_dwordx2 [[REG]] 253define amdgpu_kernel void @v_test_canonicalize_fneg_var_f64(double addrspace(1)* %out) #1 { 254 %val = load double, double addrspace(1)* %out 255 %val.fneg = fsub double -0.0, %val 256 %canonicalized = call double @llvm.canonicalize.f64(double %val.fneg) 257 store double %canonicalized, double addrspace(1)* %out 258 ret void 259} 260 261; GCN-LABEL: {{^}}test_fold_canonicalize_p0_f64: 262; GCN: v_mov_b32_e32 v[[LO:[0-9]+]], 0{{$}} 263; GCN: v_mov_b32_e32 v[[HI:[0-9]+]], v[[LO]]{{$}} 264; GCN: buffer_store_dwordx2 v{{\[}}[[LO]]:[[HI]]{{\]}} 265define amdgpu_kernel void @test_fold_canonicalize_p0_f64(double addrspace(1)* %out) #1 { 266 %canonicalized = call double @llvm.canonicalize.f64(double 0.0) 267 store double %canonicalized, double addrspace(1)* %out 268 ret void 269} 270 271; GCN-LABEL: {{^}}test_fold_canonicalize_n0_f64: 272; GCN-DAG: v_mov_b32_e32 v[[LO:[0-9]+]], 0{{$}} 273; GCN-DAG: v_bfrev_b32_e32 v[[HI:[0-9]+]], 1{{$}} 274; GCN: buffer_store_dwordx2 v{{\[}}[[LO]]:[[HI]]{{\]}} 275define amdgpu_kernel void @test_fold_canonicalize_n0_f64(double addrspace(1)* %out) #1 { 276 %canonicalized = call double @llvm.canonicalize.f64(double -0.0) 277 store double %canonicalized, double addrspace(1)* %out 278 ret void 279} 280 281; GCN-LABEL: {{^}}test_fold_canonicalize_p1_f64: 282; GCN-DAG: v_mov_b32_e32 v[[LO:[0-9]+]], 0{{$}} 283; GCN-DAG: v_mov_b32_e32 v[[HI:[0-9]+]], 0x3ff00000{{$}} 284; GCN: buffer_store_dwordx2 v{{\[}}[[LO]]:[[HI]]{{\]}} 285define amdgpu_kernel void @test_fold_canonicalize_p1_f64(double addrspace(1)* %out) #1 { 286 %canonicalized = call double @llvm.canonicalize.f64(double 1.0) 287 store double %canonicalized, double addrspace(1)* %out 288 ret void 289} 290 291; GCN-LABEL: {{^}}test_fold_canonicalize_n1_f64: 292; GCN-DAG: v_mov_b32_e32 v[[LO:[0-9]+]], 0{{$}} 293; GCN-DAG: v_mov_b32_e32 v[[HI:[0-9]+]], 0xbff00000{{$}} 294; GCN: buffer_store_dwordx2 v{{\[}}[[LO]]:[[HI]]{{\]}} 295define amdgpu_kernel void @test_fold_canonicalize_n1_f64(double addrspace(1)* %out) #1 { 296 %canonicalized = call double @llvm.canonicalize.f64(double -1.0) 297 store double %canonicalized, double addrspace(1)* %out 298 ret void 299} 300 301; GCN-LABEL: {{^}}test_fold_canonicalize_literal_f64: 302; GCN-DAG: v_mov_b32_e32 v[[LO:[0-9]+]], 0{{$}} 303; GCN-DAG: v_mov_b32_e32 v[[HI:[0-9]+]], 0x40300000{{$}} 304; GCN: buffer_store_dwordx2 v{{\[}}[[LO]]:[[HI]]{{\]}} 305define amdgpu_kernel void @test_fold_canonicalize_literal_f64(double addrspace(1)* %out) #1 { 306 %canonicalized = call double @llvm.canonicalize.f64(double 16.0) 307 store double %canonicalized, double addrspace(1)* %out 308 ret void 309} 310 311; GCN-LABEL: {{^}}test_no_denormals_fold_canonicalize_denormal0_f64: 312; GCN: v_mov_b32_e32 v[[LO:[0-9]+]], 0{{$}} 313; GCN: v_mov_b32_e32 v[[HI:[0-9]+]], v[[LO]]{{$}} 314; GCN: buffer_store_dwordx2 v{{\[}}[[LO]]:[[HI]]{{\]}} 315define amdgpu_kernel void @test_no_denormals_fold_canonicalize_denormal0_f64(double addrspace(1)* %out) #2 { 316 %canonicalized = call double @llvm.canonicalize.f64(double bitcast (i64 4503599627370495 to double)) 317 store double %canonicalized, double addrspace(1)* %out 318 ret void 319} 320 321; GCN-LABEL: {{^}}test_denormals_fold_canonicalize_denormal0_f64: 322; GCN-DAG: v_mov_b32_e32 v[[LO:[0-9]+]], -1{{$}} 323; GCN-DAG: v_mov_b32_e32 v[[HI:[0-9]+]], 0xfffff{{$}} 324; GCN: buffer_store_dwordx2 v{{\[}}[[LO]]:[[HI]]{{\]}} 325define amdgpu_kernel void @test_denormals_fold_canonicalize_denormal0_f64(double addrspace(1)* %out) #3 { 326 %canonicalized = call double @llvm.canonicalize.f64(double bitcast (i64 4503599627370495 to double)) 327 store double %canonicalized, double addrspace(1)* %out 328 ret void 329} 330 331; GCN-LABEL: {{^}}test_no_denormals_fold_canonicalize_denormal1_f64: 332; GCN: v_mov_b32_e32 v[[LO:[0-9]+]], 0{{$}} 333; GCN: v_mov_b32_e32 v[[HI:[0-9]+]], v[[LO]]{{$}} 334; GCN: buffer_store_dwordx2 v{{\[}}[[LO]]:[[HI]]{{\]}} 335define amdgpu_kernel void @test_no_denormals_fold_canonicalize_denormal1_f64(double addrspace(1)* %out) #2 { 336 %canonicalized = call double @llvm.canonicalize.f64(double bitcast (i64 9227875636482146303 to double)) 337 store double %canonicalized, double addrspace(1)* %out 338 ret void 339} 340 341; GCN-LABEL: {{^}}test_denormals_fold_canonicalize_denormal1_f64: 342; GCN-DAG: v_mov_b32_e32 v[[LO:[0-9]+]], -1{{$}} 343; GCN-DAG: v_mov_b32_e32 v[[HI:[0-9]+]], 0x800fffff{{$}} 344; GCN: buffer_store_dwordx2 v{{\[}}[[LO]]:[[HI]]{{\]}} 345define amdgpu_kernel void @test_denormals_fold_canonicalize_denormal1_f64(double addrspace(1)* %out) #3 { 346 %canonicalized = call double @llvm.canonicalize.f64(double bitcast (i64 9227875636482146303 to double)) 347 store double %canonicalized, double addrspace(1)* %out 348 ret void 349} 350 351; GCN-LABEL: {{^}}test_fold_canonicalize_qnan_f64: 352; GCN-DAG: v_mov_b32_e32 v[[HI:[0-9]+]], 0x7ff80000{{$}} 353; GCN-DAG: v_mov_b32_e32 v[[LO:[0-9]+]], 0{{$}} 354; GCN: buffer_store_dwordx2 v{{\[}}[[LO]]:[[HI]]{{\]}} 355define amdgpu_kernel void @test_fold_canonicalize_qnan_f64(double addrspace(1)* %out) #1 { 356 %canonicalized = call double @llvm.canonicalize.f64(double 0x7FF8000000000000) 357 store double %canonicalized, double addrspace(1)* %out 358 ret void 359} 360 361; GCN-LABEL: {{^}}test_fold_canonicalize_qnan_value_neg1_f64: 362; GCN-DAG: v_mov_b32_e32 v[[HI:[0-9]+]], 0x7ff80000{{$}} 363; GCN-DAG: v_mov_b32_e32 v[[LO:[0-9]+]], 0{{$}} 364; GCN: buffer_store_dwordx2 v{{\[}}[[LO]]:[[HI]]{{\]}} 365define amdgpu_kernel void @test_fold_canonicalize_qnan_value_neg1_f64(double addrspace(1)* %out) #1 { 366 %canonicalized = call double @llvm.canonicalize.f64(double bitcast (i64 -1 to double)) 367 store double %canonicalized, double addrspace(1)* %out 368 ret void 369} 370 371; GCN-LABEL: {{^}}test_fold_canonicalize_qnan_value_neg2_f64: 372; GCN-DAG: v_mov_b32_e32 v[[HI:[0-9]+]], 0x7ff80000{{$}} 373; GCN-DAG: v_mov_b32_e32 v[[LO:[0-9]+]], 0{{$}} 374; GCN: buffer_store_dwordx2 v{{\[}}[[LO]]:[[HI]]{{\]}} 375define amdgpu_kernel void @test_fold_canonicalize_qnan_value_neg2_f64(double addrspace(1)* %out) #1 { 376 %canonicalized = call double @llvm.canonicalize.f64(double bitcast (i64 -2 to double)) 377 store double %canonicalized, double addrspace(1)* %out 378 ret void 379} 380 381; GCN-LABEL: {{^}}test_fold_canonicalize_snan0_value_f64: 382; GCN-DAG: v_mov_b32_e32 v[[HI:[0-9]+]], 0x7ff80000{{$}} 383; GCN-DAG: v_mov_b32_e32 v[[LO:[0-9]+]], 0{{$}} 384; GCN: buffer_store_dwordx2 v{{\[}}[[LO]]:[[HI]]{{\]}} 385define amdgpu_kernel void @test_fold_canonicalize_snan0_value_f64(double addrspace(1)* %out) #1 { 386 %canonicalized = call double @llvm.canonicalize.f64(double bitcast (i64 9218868437227405313 to double)) 387 store double %canonicalized, double addrspace(1)* %out 388 ret void 389} 390 391; GCN-LABEL: {{^}}test_fold_canonicalize_snan1_value_f64: 392; GCN-DAG: v_mov_b32_e32 v[[HI:[0-9]+]], 0x7ff80000{{$}} 393; GCN-DAG: v_mov_b32_e32 v[[LO:[0-9]+]], 0{{$}} 394; GCN: buffer_store_dwordx2 v{{\[}}[[LO]]:[[HI]]{{\]}} 395define amdgpu_kernel void @test_fold_canonicalize_snan1_value_f64(double addrspace(1)* %out) #1 { 396 %canonicalized = call double @llvm.canonicalize.f64(double bitcast (i64 9223372036854775807 to double)) 397 store double %canonicalized, double addrspace(1)* %out 398 ret void 399} 400 401; GCN-LABEL: {{^}}test_fold_canonicalize_snan2_value_f64: 402; GCN-DAG: v_mov_b32_e32 v[[HI:[0-9]+]], 0x7ff80000{{$}} 403; GCN-DAG: v_mov_b32_e32 v[[LO:[0-9]+]], 0{{$}} 404; GCN: buffer_store_dwordx2 v{{\[}}[[LO]]:[[HI]]{{\]}} 405define amdgpu_kernel void @test_fold_canonicalize_snan2_value_f64(double addrspace(1)* %out) #1 { 406 %canonicalized = call double @llvm.canonicalize.f64(double bitcast (i64 18442240474082181121 to double)) 407 store double %canonicalized, double addrspace(1)* %out 408 ret void 409} 410 411; GCN-LABEL: {{^}}test_fold_canonicalize_snan3_value_f64: 412; GCN-DAG: v_mov_b32_e32 v[[HI:[0-9]+]], 0x7ff80000{{$}} 413; GCN-DAG: v_mov_b32_e32 v[[LO:[0-9]+]], 0{{$}} 414; GCN: buffer_store_dwordx2 v{{\[}}[[LO]]:[[HI]]{{\]}} 415define amdgpu_kernel void @test_fold_canonicalize_snan3_value_f64(double addrspace(1)* %out) #1 { 416 %canonicalized = call double @llvm.canonicalize.f64(double bitcast (i64 18446744073709551615 to double)) 417 store double %canonicalized, double addrspace(1)* %out 418 ret void 419} 420 421; GCN-LABEL: {{^}}test_canonicalize_value_f64_flush: 422; GCN: v_mul_f64 v[{{[0-9:]+}}], 1.0, v[{{[0-9:]+}}] 423define amdgpu_kernel void @test_canonicalize_value_f64_flush(double addrspace(1)* %arg, double addrspace(1)* %out) #4 { 424 %id = tail call i32 @llvm.amdgcn.workitem.id.x() 425 %gep = getelementptr inbounds double, double addrspace(1)* %arg, i32 %id 426 %v = load double, double addrspace(1)* %gep, align 8 427 %canonicalized = tail call double @llvm.canonicalize.f64(double %v) 428 %gep2 = getelementptr inbounds double, double addrspace(1)* %out, i32 %id 429 store double %canonicalized, double addrspace(1)* %gep2, align 8 430 ret void 431} 432 433; GCN-LABEL: {{^}}test_canonicalize_value_f32_flush: 434; GCN: v_mul_f32_e32 {{v[0-9]+}}, 1.0, {{v[0-9]+}} 435define amdgpu_kernel void @test_canonicalize_value_f32_flush(float addrspace(1)* %arg, float addrspace(1)* %out) #4 { 436 %id = tail call i32 @llvm.amdgcn.workitem.id.x() 437 %gep = getelementptr inbounds float, float addrspace(1)* %arg, i32 %id 438 %v = load float, float addrspace(1)* %gep, align 4 439 %canonicalized = tail call float @llvm.canonicalize.f32(float %v) 440 %gep2 = getelementptr inbounds float, float addrspace(1)* %out, i32 %id 441 store float %canonicalized, float addrspace(1)* %gep2, align 4 442 ret void 443} 444 445; GCN-LABEL: {{^}}test_canonicalize_value_f16_flush: 446; GCN: v_mul_f16_e32 {{v[0-9]+}}, 1.0, {{v[0-9]+}} 447define amdgpu_kernel void @test_canonicalize_value_f16_flush(half addrspace(1)* %arg, half addrspace(1)* %out) #4 { 448 %id = tail call i32 @llvm.amdgcn.workitem.id.x() 449 %gep = getelementptr inbounds half, half addrspace(1)* %arg, i32 %id 450 %v = load half, half addrspace(1)* %gep, align 2 451 %canonicalized = tail call half @llvm.canonicalize.f16(half %v) 452 %gep2 = getelementptr inbounds half, half addrspace(1)* %out, i32 %id 453 store half %canonicalized, half addrspace(1)* %gep2, align 2 454 ret void 455} 456 457; GCN-LABEL: {{^}}test_canonicalize_value_v2f16_flush_gfx8: 458; GCN: v_mov_b32_e32 [[ONE:v[0-9]+]], 0x3c00 459; GCN-DAG: v_mul_f16_sdwa v{{[0-9]+}}, [[ONE]], v{{[0-9]+}} dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 460; GCN-DAG: v_mul_f16_e32 v{{[0-9]+}}, 1.0, v{{[0-9]+}} 461define amdgpu_kernel void @test_canonicalize_value_v2f16_flush_gfx8(<2 x half> addrspace(1)* %arg, <2 x half> addrspace(1)* %out) #4 { 462 %id = tail call i32 @llvm.amdgcn.workitem.id.x() 463 %gep = getelementptr inbounds <2 x half>, <2 x half> addrspace(1)* %arg, i32 %id 464 %v = load <2 x half>, <2 x half> addrspace(1)* %gep, align 4 465 %canonicalized = tail call <2 x half> @llvm.canonicalize.v2f16(<2 x half> %v) 466 %gep2 = getelementptr inbounds <2 x half>, <2 x half> addrspace(1)* %out, i32 %id 467 store <2 x half> %canonicalized, <2 x half> addrspace(1)* %gep2, align 2 468 ret void 469} 470 471; GCN-LABEL: {{^}}test_canonicalize_value_v2f16_flush_gfx9: 472; GCN-DAG: v_pk_mul_f16 v{{[0-9]+}}, 1.0, v{{[0-9]+}} 473define amdgpu_kernel void @test_canonicalize_value_v2f16_flush_gfx9(<2 x half> addrspace(1)* %arg, <2 x half> addrspace(1)* %out) #6 { 474 %id = tail call i32 @llvm.amdgcn.workitem.id.x() 475 %gep = getelementptr inbounds <2 x half>, <2 x half> addrspace(1)* %arg, i32 %id 476 %v = load <2 x half>, <2 x half> addrspace(1)* %gep, align 4 477 %canonicalized = tail call <2 x half> @llvm.canonicalize.v2f16(<2 x half> %v) 478 %gep2 = getelementptr inbounds <2 x half>, <2 x half> addrspace(1)* %out, i32 %id 479 store <2 x half> %canonicalized, <2 x half> addrspace(1)* %gep2, align 2 480 ret void 481} 482 483; GCN-LABEL: {{^}}test_canonicalize_value_f64_denorm: 484; GCN: v_max_f64 v[{{[0-9:]+}}], v[{{[0-9:]+}}], v[{{[0-9:]+}}] 485define amdgpu_kernel void @test_canonicalize_value_f64_denorm(double addrspace(1)* %arg, double addrspace(1)* %out) #5 { 486 %id = tail call i32 @llvm.amdgcn.workitem.id.x() 487 %gep = getelementptr inbounds double, double addrspace(1)* %arg, i32 %id 488 %v = load double, double addrspace(1)* %gep, align 8 489 %canonicalized = tail call double @llvm.canonicalize.f64(double %v) 490 %gep2 = getelementptr inbounds double, double addrspace(1)* %out, i32 %id 491 store double %canonicalized, double addrspace(1)* %gep2, align 8 492 ret void 493} 494 495; GCN-LABEL: {{^}}test_canonicalize_value_f32_denorm: 496; GCN: v_max_f32_e32 {{v[0-9]+}}, {{v[0-9]+}}, {{v[0-9]+}} 497define amdgpu_kernel void @test_canonicalize_value_f32_denorm(float addrspace(1)* %arg, float addrspace(1)* %out) #5 { 498 %id = tail call i32 @llvm.amdgcn.workitem.id.x() 499 %gep = getelementptr inbounds float, float addrspace(1)* %arg, i32 %id 500 %v = load float, float addrspace(1)* %gep, align 4 501 %canonicalized = tail call float @llvm.canonicalize.f32(float %v) 502 %gep2 = getelementptr inbounds float, float addrspace(1)* %out, i32 %id 503 store float %canonicalized, float addrspace(1)* %gep2, align 4 504 ret void 505} 506 507; GCN-LABEL: {{^}}test_canonicalize_value_f16_denorm: 508; GCN: v_max_f16_e32 {{v[0-9]+}}, {{v[0-9]+}}, {{v[0-9]+}} 509define amdgpu_kernel void @test_canonicalize_value_f16_denorm(half addrspace(1)* %arg, half addrspace(1)* %out) #5 { 510 %id = tail call i32 @llvm.amdgcn.workitem.id.x() 511 %gep = getelementptr inbounds half, half addrspace(1)* %arg, i32 %id 512 %v = load half, half addrspace(1)* %gep, align 2 513 %canonicalized = tail call half @llvm.canonicalize.f16(half %v) 514 %gep2 = getelementptr inbounds half, half addrspace(1)* %out, i32 %id 515 store half %canonicalized, half addrspace(1)* %gep2, align 2 516 ret void 517} 518 519; GCN-LABEL: {{^}}test_canonicalize_value_v2f16_denorm: 520; GCN: v_pk_max_f16 {{v[0-9]+}}, {{v[0-9]+}}, {{v[0-9]+}} 521define amdgpu_kernel void @test_canonicalize_value_v2f16_denorm(<2 x half> addrspace(1)* %arg, <2 x half> addrspace(1)* %out) #5 { 522 %id = tail call i32 @llvm.amdgcn.workitem.id.x() 523 %gep = getelementptr inbounds <2 x half>, <2 x half> addrspace(1)* %arg, i32 %id 524 %v = load <2 x half>, <2 x half> addrspace(1)* %gep, align 4 525 %canonicalized = tail call <2 x half> @llvm.canonicalize.v2f16(<2 x half> %v) 526 %gep2 = getelementptr inbounds <2 x half>, <2 x half> addrspace(1)* %out, i32 %id 527 store <2 x half> %canonicalized, <2 x half> addrspace(1)* %gep2, align 2 528 ret void 529} 530 531attributes #0 = { nounwind readnone } 532attributes #1 = { nounwind } 533attributes #2 = { nounwind "target-features"="-fp32-denormals,-fp64-fp16-denormals" } 534attributes #3 = { nounwind "target-features"="+fp32-denormals,+fp64-fp16-denormals" } 535attributes #4 = { nounwind "target-features"="-fp32-denormals,-fp64-fp16-denormals" "target-cpu"="tonga" } 536attributes #5 = { nounwind "target-features"="+fp32-denormals,+fp64-fp16-denormals" "target-cpu"="gfx900" } 537attributes #6 = { nounwind "target-features"="-fp32-denormals,-fp64-fp16-denormals" "target-cpu"="gfx900" } 538