1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py 2; RUN: opt -mtriple=x86_64-unknown-linux-gnu < %s -instcombine -S | FileCheck %s 3 4; Make sure libcalls are replaced with intrinsic calls. 5 6declare float @llvm.fabs.f32(float) 7declare double @llvm.fabs.f64(double) 8declare fp128 @llvm.fabs.f128(fp128) 9 10declare float @fabsf(float) 11declare double @fabs(double) 12declare fp128 @fabsl(fp128) 13declare float @llvm.fma.f32(float, float, float) 14declare float @llvm.fmuladd.f32(float, float, float) 15 16define float @replace_fabs_call_f32(float %x) { 17; CHECK-LABEL: @replace_fabs_call_f32( 18; CHECK-NEXT: [[FABSF:%.*]] = call float @llvm.fabs.f32(float [[X:%.*]]) 19; CHECK-NEXT: ret float [[FABSF]] 20; 21 %fabsf = tail call float @fabsf(float %x) 22 ret float %fabsf 23} 24 25define double @replace_fabs_call_f64(double %x) { 26; CHECK-LABEL: @replace_fabs_call_f64( 27; CHECK-NEXT: [[FABS:%.*]] = call double @llvm.fabs.f64(double [[X:%.*]]) 28; CHECK-NEXT: ret double [[FABS]] 29; 30 %fabs = tail call double @fabs(double %x) 31 ret double %fabs 32} 33 34define fp128 @replace_fabs_call_f128(fp128 %x) { 35; CHECK-LABEL: @replace_fabs_call_f128( 36; CHECK-NEXT: [[FABSL:%.*]] = call fp128 @llvm.fabs.f128(fp128 [[X:%.*]]) 37; CHECK-NEXT: ret fp128 [[FABSL]] 38; 39 %fabsl = tail call fp128 @fabsl(fp128 %x) 40 ret fp128 %fabsl 41} 42 43; Make sure fast math flags are preserved when replacing the libcall. 44define float @fmf_replace_fabs_call_f32(float %x) { 45; CHECK-LABEL: @fmf_replace_fabs_call_f32( 46; CHECK-NEXT: [[FABSF:%.*]] = call nnan float @llvm.fabs.f32(float [[X:%.*]]) 47; CHECK-NEXT: ret float [[FABSF]] 48; 49 %fabsf = tail call nnan float @fabsf(float %x) 50 ret float %fabsf 51} 52 53; Make sure all intrinsic calls are eliminated when the input is known 54; positive. 55 56; The fabs cannot be eliminated because %x may be a NaN 57 58define float @square_fabs_intrinsic_f32(float %x) { 59; CHECK-LABEL: @square_fabs_intrinsic_f32( 60; CHECK-NEXT: [[MUL:%.*]] = fmul float [[X:%.*]], [[X]] 61; CHECK-NEXT: [[FABSF:%.*]] = tail call float @llvm.fabs.f32(float [[MUL]]) 62; CHECK-NEXT: ret float [[FABSF]] 63; 64 %mul = fmul float %x, %x 65 %fabsf = tail call float @llvm.fabs.f32(float %mul) 66 ret float %fabsf 67} 68 69define double @square_fabs_intrinsic_f64(double %x) { 70; CHECK-LABEL: @square_fabs_intrinsic_f64( 71; CHECK-NEXT: [[MUL:%.*]] = fmul double [[X:%.*]], [[X]] 72; CHECK-NEXT: [[FABS:%.*]] = tail call double @llvm.fabs.f64(double [[MUL]]) 73; CHECK-NEXT: ret double [[FABS]] 74; 75 %mul = fmul double %x, %x 76 %fabs = tail call double @llvm.fabs.f64(double %mul) 77 ret double %fabs 78} 79 80define fp128 @square_fabs_intrinsic_f128(fp128 %x) { 81; CHECK-LABEL: @square_fabs_intrinsic_f128( 82; CHECK-NEXT: [[MUL:%.*]] = fmul fp128 [[X:%.*]], [[X]] 83; CHECK-NEXT: [[FABSL:%.*]] = tail call fp128 @llvm.fabs.f128(fp128 [[MUL]]) 84; CHECK-NEXT: ret fp128 [[FABSL]] 85; 86 %mul = fmul fp128 %x, %x 87 %fabsl = tail call fp128 @llvm.fabs.f128(fp128 %mul) 88 ret fp128 %fabsl 89} 90 91define float @square_nnan_fabs_intrinsic_f32(float %x) { 92; CHECK-LABEL: @square_nnan_fabs_intrinsic_f32( 93; CHECK-NEXT: [[MUL:%.*]] = fmul nnan float [[X:%.*]], [[X]] 94; CHECK-NEXT: ret float [[MUL]] 95; 96 %mul = fmul nnan float %x, %x 97 %fabsf = call float @llvm.fabs.f32(float %mul) 98 ret float %fabsf 99} 100 101; Shrinking a library call to a smaller type should not be inhibited by nor inhibit the square optimization. 102 103define float @square_fabs_shrink_call1(float %x) { 104; CHECK-LABEL: @square_fabs_shrink_call1( 105; CHECK-NEXT: [[TMP1:%.*]] = fmul float [[X:%.*]], [[X]] 106; CHECK-NEXT: [[TRUNC:%.*]] = call float @llvm.fabs.f32(float [[TMP1]]) 107; CHECK-NEXT: ret float [[TRUNC]] 108; 109 %ext = fpext float %x to double 110 %sq = fmul double %ext, %ext 111 %fabs = call double @fabs(double %sq) 112 %trunc = fptrunc double %fabs to float 113 ret float %trunc 114} 115 116define float @square_fabs_shrink_call2(float %x) { 117; CHECK-LABEL: @square_fabs_shrink_call2( 118; CHECK-NEXT: [[SQ:%.*]] = fmul float [[X:%.*]], [[X]] 119; CHECK-NEXT: [[TRUNC:%.*]] = call float @llvm.fabs.f32(float [[SQ]]) 120; CHECK-NEXT: ret float [[TRUNC]] 121; 122 %sq = fmul float %x, %x 123 %ext = fpext float %sq to double 124 %fabs = call double @fabs(double %ext) 125 %trunc = fptrunc double %fabs to float 126 ret float %trunc 127} 128 129define float @fabs_select_constant_negative_positive(i32 %c) { 130; CHECK-LABEL: @fabs_select_constant_negative_positive( 131; CHECK-NEXT: [[CMP:%.*]] = icmp eq i32 [[C:%.*]], 0 132; CHECK-NEXT: [[FABS:%.*]] = select i1 [[CMP]], float 1.000000e+00, float 2.000000e+00 133; CHECK-NEXT: ret float [[FABS]] 134; 135 %cmp = icmp eq i32 %c, 0 136 %select = select i1 %cmp, float -1.0, float 2.0 137 %fabs = call float @llvm.fabs.f32(float %select) 138 ret float %fabs 139} 140 141define float @fabs_select_constant_positive_negative(i32 %c) { 142; CHECK-LABEL: @fabs_select_constant_positive_negative( 143; CHECK-NEXT: [[CMP:%.*]] = icmp eq i32 [[C:%.*]], 0 144; CHECK-NEXT: [[FABS:%.*]] = select i1 [[CMP]], float 1.000000e+00, float 2.000000e+00 145; CHECK-NEXT: ret float [[FABS]] 146; 147 %cmp = icmp eq i32 %c, 0 148 %select = select i1 %cmp, float 1.0, float -2.0 149 %fabs = call float @llvm.fabs.f32(float %select) 150 ret float %fabs 151} 152 153define float @fabs_select_constant_negative_negative(i32 %c) { 154; CHECK-LABEL: @fabs_select_constant_negative_negative( 155; CHECK-NEXT: [[CMP:%.*]] = icmp eq i32 [[C:%.*]], 0 156; CHECK-NEXT: [[FABS:%.*]] = select i1 [[CMP]], float 1.000000e+00, float 2.000000e+00 157; CHECK-NEXT: ret float [[FABS]] 158; 159 %cmp = icmp eq i32 %c, 0 160 %select = select i1 %cmp, float -1.0, float -2.0 161 %fabs = call float @llvm.fabs.f32(float %select) 162 ret float %fabs 163} 164 165define float @fabs_select_constant_neg0(i32 %c) { 166; CHECK-LABEL: @fabs_select_constant_neg0( 167; CHECK-NEXT: ret float 0.000000e+00 168; 169 %cmp = icmp eq i32 %c, 0 170 %select = select i1 %cmp, float -0.0, float 0.0 171 %fabs = call float @llvm.fabs.f32(float %select) 172 ret float %fabs 173} 174 175define float @fabs_select_var_constant_negative(i32 %c, float %x) { 176; CHECK-LABEL: @fabs_select_var_constant_negative( 177; CHECK-NEXT: [[CMP:%.*]] = icmp eq i32 [[C:%.*]], 0 178; CHECK-NEXT: [[SELECT:%.*]] = select i1 [[CMP]], float [[X:%.*]], float -1.000000e+00 179; CHECK-NEXT: [[FABS:%.*]] = call float @llvm.fabs.f32(float [[SELECT]]) 180; CHECK-NEXT: ret float [[FABS]] 181; 182 %cmp = icmp eq i32 %c, 0 183 %select = select i1 %cmp, float %x, float -1.0 184 %fabs = call float @llvm.fabs.f32(float %select) 185 ret float %fabs 186} 187 188; The fabs cannot be eliminated because %x may be a NaN 189 190define float @square_fma_fabs_intrinsic_f32(float %x) { 191; CHECK-LABEL: @square_fma_fabs_intrinsic_f32( 192; CHECK-NEXT: [[FMA:%.*]] = call float @llvm.fma.f32(float [[X:%.*]], float [[X]], float 1.000000e+00) 193; CHECK-NEXT: [[FABSF:%.*]] = call float @llvm.fabs.f32(float [[FMA]]) 194; CHECK-NEXT: ret float [[FABSF]] 195; 196 %fma = call float @llvm.fma.f32(float %x, float %x, float 1.0) 197 %fabsf = call float @llvm.fabs.f32(float %fma) 198 ret float %fabsf 199} 200 201; The fabs cannot be eliminated because %x may be a NaN 202 203define float @square_nnan_fma_fabs_intrinsic_f32(float %x) { 204; CHECK-LABEL: @square_nnan_fma_fabs_intrinsic_f32( 205; CHECK-NEXT: [[FMA:%.*]] = call nnan float @llvm.fma.f32(float [[X:%.*]], float [[X]], float 1.000000e+00) 206; CHECK-NEXT: ret float [[FMA]] 207; 208 %fma = call nnan float @llvm.fma.f32(float %x, float %x, float 1.0) 209 %fabsf = call float @llvm.fabs.f32(float %fma) 210 ret float %fabsf 211} 212 213define float @square_fmuladd_fabs_intrinsic_f32(float %x) { 214; CHECK-LABEL: @square_fmuladd_fabs_intrinsic_f32( 215; CHECK-NEXT: [[FMULADD:%.*]] = call float @llvm.fmuladd.f32(float [[X:%.*]], float [[X]], float 1.000000e+00) 216; CHECK-NEXT: [[FABSF:%.*]] = call float @llvm.fabs.f32(float [[FMULADD]]) 217; CHECK-NEXT: ret float [[FABSF]] 218; 219 %fmuladd = call float @llvm.fmuladd.f32(float %x, float %x, float 1.0) 220 %fabsf = call float @llvm.fabs.f32(float %fmuladd) 221 ret float %fabsf 222} 223 224define float @square_nnan_fmuladd_fabs_intrinsic_f32(float %x) { 225; CHECK-LABEL: @square_nnan_fmuladd_fabs_intrinsic_f32( 226; CHECK-NEXT: [[FMULADD:%.*]] = call nnan float @llvm.fmuladd.f32(float [[X:%.*]], float [[X]], float 1.000000e+00) 227; CHECK-NEXT: ret float [[FMULADD]] 228; 229 %fmuladd = call nnan float @llvm.fmuladd.f32(float %x, float %x, float 1.0) 230 %fabsf = call float @llvm.fabs.f32(float %fmuladd) 231 ret float %fabsf 232} 233 234; Don't introduce a second fpext 235 236define double @multi_use_fabs_fpext(float %x) { 237; CHECK-LABEL: @multi_use_fabs_fpext( 238; CHECK-NEXT: [[FPEXT:%.*]] = fpext float [[X:%.*]] to double 239; CHECK-NEXT: [[FABS:%.*]] = call double @llvm.fabs.f64(double [[FPEXT]]) 240; CHECK-NEXT: store volatile double [[FPEXT]], double* undef, align 8 241; CHECK-NEXT: ret double [[FABS]] 242; 243 %fpext = fpext float %x to double 244 %fabs = call double @llvm.fabs.f64(double %fpext) 245 store volatile double %fpext, double* undef 246 ret double %fabs 247} 248 249; Negative test for the fabs folds below: we require nnan, so 250; we won't always clear the sign bit of a NaN value. 251 252define double @select_fcmp_ole_zero(double %x) { 253; CHECK-LABEL: @select_fcmp_ole_zero( 254; CHECK-NEXT: [[LEZERO:%.*]] = fcmp ole double [[X:%.*]], 0.000000e+00 255; CHECK-NEXT: [[NEGX:%.*]] = fsub double 0.000000e+00, [[X]] 256; CHECK-NEXT: [[FABS:%.*]] = select i1 [[LEZERO]], double [[NEGX]], double [[X]] 257; CHECK-NEXT: ret double [[FABS]] 258; 259 %lezero = fcmp ole double %x, 0.0 260 %negx = fsub double 0.0, %x 261 %fabs = select i1 %lezero, double %negx, double %x 262 ret double %fabs 263} 264 265; X <= 0.0 ? (0.0 - X) : X --> fabs(X) 266 267define double @select_fcmp_nnan_ole_zero(double %x) { 268; CHECK-LABEL: @select_fcmp_nnan_ole_zero( 269; CHECK-NEXT: [[TMP1:%.*]] = call nnan double @llvm.fabs.f64(double [[X:%.*]]) 270; CHECK-NEXT: ret double [[TMP1]] 271; 272 %lezero = fcmp nnan ole double %x, 0.0 273 %negx = fsub double 0.0, %x 274 %fabs = select i1 %lezero, double %negx, double %x 275 ret double %fabs 276} 277 278; X <= -0.0 ? (0.0 - X) : X --> fabs(X) 279 280define <2 x float> @select_fcmp_nnan_ole_negzero(<2 x float> %x) { 281; CHECK-LABEL: @select_fcmp_nnan_ole_negzero( 282; CHECK-NEXT: [[TMP1:%.*]] = call nnan <2 x float> @llvm.fabs.v2f32(<2 x float> [[X:%.*]]) 283; CHECK-NEXT: ret <2 x float> [[TMP1]] 284; 285 %lezero = fcmp nnan ole <2 x float> %x, <float -0.0, float -0.0> 286 %negx = fsub <2 x float> <float 0.0, float undef>, %x 287 %fabs = select <2 x i1> %lezero, <2 x float> %negx, <2 x float> %x 288 ret <2 x float> %fabs 289} 290 291; X > 0.0 ? X : (0.0 - X) --> fabs(X) 292 293define fp128 @select_fcmp_nnan_ogt_zero(fp128 %x) { 294; CHECK-LABEL: @select_fcmp_nnan_ogt_zero( 295; CHECK-NEXT: [[TMP1:%.*]] = call nnan fp128 @llvm.fabs.f128(fp128 [[X:%.*]]) 296; CHECK-NEXT: ret fp128 [[TMP1]] 297; 298 %gtzero = fcmp nnan ogt fp128 %x, zeroinitializer 299 %negx = fsub fp128 zeroinitializer, %x 300 %fabs = select i1 %gtzero, fp128 %x, fp128 %negx 301 ret fp128 %fabs 302} 303 304; X > -0.0 ? X : (0.0 - X) --> fabs(X) 305 306define half @select_fcmp_nnan_ogt_negzero(half %x) { 307; CHECK-LABEL: @select_fcmp_nnan_ogt_negzero( 308; CHECK-NEXT: [[TMP1:%.*]] = call nnan half @llvm.fabs.f16(half [[X:%.*]]) 309; CHECK-NEXT: ret half [[TMP1]] 310; 311 %gtzero = fcmp nnan ogt half %x, -0.0 312 %negx = fsub half 0.0, %x 313 %fabs = select i1 %gtzero, half %x, half %negx 314 ret half %fabs 315} 316 317; X < 0.0 ? -X : X --> fabs(X) 318 319define double @select_fcmp_nnan_nsz_olt_zero(double %x) { 320; CHECK-LABEL: @select_fcmp_nnan_nsz_olt_zero( 321; CHECK-NEXT: [[TMP1:%.*]] = call nnan nsz double @llvm.fabs.f64(double [[X:%.*]]) 322; CHECK-NEXT: ret double [[TMP1]] 323; 324 %ltzero = fcmp nnan nsz olt double %x, 0.0 325 %negx = fsub double -0.0, %x 326 %fabs = select i1 %ltzero, double %negx, double %x 327 ret double %fabs 328} 329 330; X < -0.0 ? -X : X --> fabs(X) 331 332define float @select_fcmp_nnan_nsz_olt_negzero(float %x) { 333; CHECK-LABEL: @select_fcmp_nnan_nsz_olt_negzero( 334; CHECK-NEXT: [[TMP1:%.*]] = call nnan ninf nsz float @llvm.fabs.f32(float [[X:%.*]]) 335; CHECK-NEXT: ret float [[TMP1]] 336; 337 %ltzero = fcmp nnan nsz ninf olt float %x, -0.0 338 %negx = fsub float -0.0, %x 339 %fabs = select i1 %ltzero, float %negx, float %x 340 ret float %fabs 341} 342 343; X <= 0.0 ? -X : X --> fabs(X) 344 345define double @select_fcmp_nnan_nsz_ole_zero(double %x) { 346; CHECK-LABEL: @select_fcmp_nnan_nsz_ole_zero( 347; CHECK-NEXT: [[TMP1:%.*]] = call fast double @llvm.fabs.f64(double [[X:%.*]]) 348; CHECK-NEXT: ret double [[TMP1]] 349; 350 %lezero = fcmp fast ole double %x, 0.0 351 %negx = fsub double -0.0, %x 352 %fabs = select i1 %lezero, double %negx, double %x 353 ret double %fabs 354} 355 356; X <= -0.0 ? -X : X --> fabs(X) 357 358define float @select_fcmp_nnan_nsz_ole_negzero(float %x) { 359; CHECK-LABEL: @select_fcmp_nnan_nsz_ole_negzero( 360; CHECK-NEXT: [[TMP1:%.*]] = call nnan nsz float @llvm.fabs.f32(float [[X:%.*]]) 361; CHECK-NEXT: ret float [[TMP1]] 362; 363 %lezero = fcmp nnan nsz ole float %x, -0.0 364 %negx = fsub float -0.0, %x 365 %fabs = select i1 %lezero, float %negx, float %x 366 ret float %fabs 367} 368 369; X > 0.0 ? X : (0.0 - X) --> fabs(X) 370 371define <2 x float> @select_fcmp_nnan_nsz_ogt_zero(<2 x float> %x) { 372; CHECK-LABEL: @select_fcmp_nnan_nsz_ogt_zero( 373; CHECK-NEXT: [[TMP1:%.*]] = call nnan nsz arcp <2 x float> @llvm.fabs.v2f32(<2 x float> [[X:%.*]]) 374; CHECK-NEXT: ret <2 x float> [[TMP1]] 375; 376 %gtzero = fcmp nnan nsz arcp ogt <2 x float> %x, zeroinitializer 377 %negx = fsub <2 x float> <float -0.0, float -0.0>, %x 378 %fabs = select <2 x i1> %gtzero, <2 x float> %x, <2 x float> %negx 379 ret <2 x float> %fabs 380} 381 382; X > -0.0 ? X : (0.0 - X) --> fabs(X) 383 384define half @select_fcmp_nnan_nsz_ogt_negzero(half %x) { 385; CHECK-LABEL: @select_fcmp_nnan_nsz_ogt_negzero( 386; CHECK-NEXT: [[TMP1:%.*]] = call fast half @llvm.fabs.f16(half [[X:%.*]]) 387; CHECK-NEXT: ret half [[TMP1]] 388; 389 %gtzero = fcmp fast ogt half %x, -0.0 390 %negx = fsub half 0.0, %x 391 %fabs = select i1 %gtzero, half %x, half %negx 392 ret half %fabs 393} 394 395; X > 0.0 ? X : (0.0 - X) --> fabs(X) 396 397define <2 x double> @select_fcmp_nnan_nsz_oge_zero(<2 x double> %x) { 398; CHECK-LABEL: @select_fcmp_nnan_nsz_oge_zero( 399; CHECK-NEXT: [[TMP1:%.*]] = call reassoc nnan nsz <2 x double> @llvm.fabs.v2f64(<2 x double> [[X:%.*]]) 400; CHECK-NEXT: ret <2 x double> [[TMP1]] 401; 402 %gezero = fcmp nnan nsz reassoc oge <2 x double> %x, zeroinitializer 403 %negx = fsub <2 x double> <double -0.0, double -0.0>, %x 404 %fabs = select <2 x i1> %gezero, <2 x double> %x, <2 x double> %negx 405 ret <2 x double> %fabs 406} 407 408; X > -0.0 ? X : (0.0 - X) --> fabs(X) 409 410define half @select_fcmp_nnan_nsz_oge_negzero(half %x) { 411; CHECK-LABEL: @select_fcmp_nnan_nsz_oge_negzero( 412; CHECK-NEXT: [[TMP1:%.*]] = call nnan nsz half @llvm.fabs.f16(half [[X:%.*]]) 413; CHECK-NEXT: ret half [[TMP1]] 414; 415 %gezero = fcmp nnan nsz oge half %x, -0.0 416 %negx = fsub half -0.0, %x 417 %fabs = select i1 %gezero, half %x, half %negx 418 ret half %fabs 419} 420 421