1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py 2; RUN: opt < %s -reassociate -gvn -instcombine -S | FileCheck %s 3 4; With reassociation, constant folding can eliminate the 12 and -12 constants. 5define float @test1(float %arg) { 6; CHECK-LABEL: @test1( 7; CHECK-NEXT: [[ARG_NEG:%.*]] = fneg fast float [[ARG:%.*]] 8; CHECK-NEXT: ret float [[ARG_NEG]] 9; 10 %t1 = fsub fast float -1.200000e+01, %arg 11 %t2 = fadd fast float %t1, 1.200000e+01 12 ret float %t2 13} 14 15; Check again using the minimal subset of FMF. 16; Both 'reassoc' and 'nsz' are required. 17define float @test1_minimal(float %arg) { 18; CHECK-LABEL: @test1_minimal( 19; CHECK-NEXT: [[TMP1:%.*]] = fneg reassoc nsz float [[ARG:%.*]] 20; CHECK-NEXT: ret float [[TMP1]] 21; 22 %t1 = fsub reassoc nsz float -1.200000e+01, %arg 23 %t2 = fadd reassoc nsz float %t1, 1.200000e+01 24 ret float %t2 25} 26 27; Verify the fold is not done with only 'reassoc' ('nsz' is required). 28define float @test1_reassoc(float %arg) { 29; CHECK-LABEL: @test1_reassoc( 30; CHECK-NEXT: [[T1:%.*]] = fsub reassoc float -1.200000e+01, [[ARG:%.*]] 31; CHECK-NEXT: [[T2:%.*]] = fadd reassoc float [[T1]], 1.200000e+01 32; CHECK-NEXT: ret float [[T2]] 33; 34 %t1 = fsub reassoc float -1.200000e+01, %arg 35 %t2 = fadd reassoc float %t1, 1.200000e+01 36 ret float %t2 37} 38 39define float @test2(float %reg109, float %reg1111) { 40; CHECK-LABEL: @test2( 41; CHECK-NEXT: [[REG115:%.*]] = fadd float [[REG109:%.*]], -3.000000e+01 42; CHECK-NEXT: [[REG116:%.*]] = fadd float [[REG115]], [[REG1111:%.*]] 43; CHECK-NEXT: [[REG117:%.*]] = fadd float [[REG116]], 3.000000e+01 44; CHECK-NEXT: ret float [[REG117]] 45; 46 %reg115 = fadd float %reg109, -3.000000e+01 47 %reg116 = fadd float %reg115, %reg1111 48 %reg117 = fadd float %reg116, 3.000000e+01 49 ret float %reg117 50} 51 52define float @test3(float %reg109, float %reg1111) { 53; CHECK-LABEL: @test3( 54; CHECK-NEXT: [[REG117:%.*]] = fadd fast float [[REG109:%.*]], [[REG1111:%.*]] 55; CHECK-NEXT: ret float [[REG117]] 56; 57 %reg115 = fadd fast float %reg109, -3.000000e+01 58 %reg116 = fadd fast float %reg115, %reg1111 59 %reg117 = fadd fast float %reg116, 3.000000e+01 60 ret float %reg117 61} 62 63define float @test3_reassoc(float %reg109, float %reg1111) { 64; CHECK-LABEL: @test3_reassoc( 65; CHECK-NEXT: [[REG115:%.*]] = fadd reassoc float [[REG109:%.*]], -3.000000e+01 66; CHECK-NEXT: [[REG116:%.*]] = fadd reassoc float [[REG115]], [[REG1111:%.*]] 67; CHECK-NEXT: [[REG117:%.*]] = fadd reassoc float [[REG116]], 3.000000e+01 68; CHECK-NEXT: ret float [[REG117]] 69; 70 %reg115 = fadd reassoc float %reg109, -3.000000e+01 71 %reg116 = fadd reassoc float %reg115, %reg1111 72 %reg117 = fadd reassoc float %reg116, 3.000000e+01 73 ret float %reg117 74} 75 76@fe = external global float 77@fa = external global float 78@fb = external global float 79@fc = external global float 80@ff = external global float 81 82define void @test4() { 83; CHECK-LABEL: @test4( 84; CHECK-NEXT: [[A:%.*]] = load float, float* @fa, align 4 85; CHECK-NEXT: [[B:%.*]] = load float, float* @fb, align 4 86; CHECK-NEXT: [[C:%.*]] = load float, float* @fc, align 4 87; CHECK-NEXT: [[T1:%.*]] = fadd fast float [[B]], [[A]] 88; CHECK-NEXT: [[T2:%.*]] = fadd fast float [[T1]], [[C]] 89; CHECK-NEXT: store float [[T2]], float* @fe, align 4 90; CHECK-NEXT: store float [[T2]], float* @ff, align 4 91; CHECK-NEXT: ret void 92; 93 %A = load float, float* @fa 94 %B = load float, float* @fb 95 %C = load float, float* @fc 96 %t1 = fadd fast float %A, %B 97 %t2 = fadd fast float %t1, %C 98 %t3 = fadd fast float %C, %A 99 %t4 = fadd fast float %t3, %B 100 ; e = (a+b)+c; 101 store float %t2, float* @fe 102 ; f = (a+c)+b 103 store float %t4, float* @ff 104 ret void 105} 106 107define void @test5() { 108; CHECK-LABEL: @test5( 109; CHECK-NEXT: [[A:%.*]] = load float, float* @fa, align 4 110; CHECK-NEXT: [[B:%.*]] = load float, float* @fb, align 4 111; CHECK-NEXT: [[C:%.*]] = load float, float* @fc, align 4 112; CHECK-NEXT: [[T1:%.*]] = fadd fast float [[B]], [[A]] 113; CHECK-NEXT: [[T2:%.*]] = fadd fast float [[T1]], [[C]] 114; CHECK-NEXT: store float [[T2]], float* @fe, align 4 115; CHECK-NEXT: store float [[T2]], float* @ff, align 4 116; CHECK-NEXT: ret void 117; 118 %A = load float, float* @fa 119 %B = load float, float* @fb 120 %C = load float, float* @fc 121 %t1 = fadd fast float %A, %B 122 %t2 = fadd fast float %t1, %C 123 %t3 = fadd fast float %C, %A 124 %t4 = fadd fast float %t3, %B 125 ; e = c+(a+b) 126 store float %t2, float* @fe 127 ; f = (c+a)+b 128 store float %t4, float* @ff 129 ret void 130} 131 132define void @test6() { 133; CHECK-LABEL: @test6( 134; CHECK-NEXT: [[A:%.*]] = load float, float* @fa, align 4 135; CHECK-NEXT: [[B:%.*]] = load float, float* @fb, align 4 136; CHECK-NEXT: [[C:%.*]] = load float, float* @fc, align 4 137; CHECK-NEXT: [[T1:%.*]] = fadd fast float [[B]], [[A]] 138; CHECK-NEXT: [[T2:%.*]] = fadd fast float [[T1]], [[C]] 139; CHECK-NEXT: store float [[T2]], float* @fe, align 4 140; CHECK-NEXT: store float [[T2]], float* @ff, align 4 141; CHECK-NEXT: ret void 142; 143 %A = load float, float* @fa 144 %B = load float, float* @fb 145 %C = load float, float* @fc 146 %t1 = fadd fast float %B, %A 147 %t2 = fadd fast float %t1, %C 148 %t3 = fadd fast float %C, %A 149 %t4 = fadd fast float %t3, %B 150 ; e = c+(b+a) 151 store float %t2, float* @fe 152 ; f = (c+a)+b 153 store float %t4, float* @ff 154 ret void 155} 156 157define float @test7(float %A, float %B, float %C) { 158; CHECK-LABEL: @test7( 159; CHECK-NEXT: [[REASS_ADD1:%.*]] = fadd fast float [[C:%.*]], [[B:%.*]] 160; CHECK-NEXT: [[REASS_MUL2:%.*]] = fmul fast float [[A:%.*]], [[A]] 161; CHECK-NEXT: [[REASS_MUL:%.*]] = fmul fast float [[REASS_MUL2]], [[REASS_ADD1]] 162; CHECK-NEXT: ret float [[REASS_MUL]] 163; 164 %aa = fmul fast float %A, %A 165 %aab = fmul fast float %aa, %B 166 %ac = fmul fast float %A, %C 167 %aac = fmul fast float %ac, %A 168 %r = fadd fast float %aab, %aac 169 ret float %r 170} 171 172define float @test7_reassoc(float %A, float %B, float %C) { 173; CHECK-LABEL: @test7_reassoc( 174; CHECK-NEXT: [[AA:%.*]] = fmul reassoc float [[A:%.*]], [[A]] 175; CHECK-NEXT: [[AAB:%.*]] = fmul reassoc float [[AA]], [[B:%.*]] 176; CHECK-NEXT: [[TMP1:%.*]] = fmul reassoc float [[A]], [[A]] 177; CHECK-NEXT: [[AAC:%.*]] = fmul reassoc float [[TMP1]], [[C:%.*]] 178; CHECK-NEXT: [[R:%.*]] = fadd reassoc float [[AAB]], [[AAC]] 179; CHECK-NEXT: ret float [[R]] 180; 181 %aa = fmul reassoc float %A, %A 182 %aab = fmul reassoc float %aa, %B 183 %ac = fmul reassoc float %A, %C 184 %aac = fmul reassoc float %ac, %A 185 %r = fadd reassoc float %aab, %aac 186 ret float %r 187} 188 189; (-X)*Y + Z -> Z-X*Y 190 191define float @test8(float %X, float %Y, float %Z) { 192; CHECK-LABEL: @test8( 193; CHECK-NEXT: [[A:%.*]] = fmul fast float [[Y:%.*]], [[X:%.*]] 194; CHECK-NEXT: [[C:%.*]] = fsub fast float [[Z:%.*]], [[A]] 195; CHECK-NEXT: ret float [[C]] 196; 197 %A = fsub fast float 0.0, %X 198 %B = fmul fast float %A, %Y 199 %C = fadd fast float %B, %Z 200 ret float %C 201} 202 203define float @test8_unary_fneg(float %X, float %Y, float %Z) { 204; CHECK-LABEL: @test8_unary_fneg( 205; CHECK-NEXT: [[TMP1:%.*]] = fmul fast float [[X:%.*]], [[Y:%.*]] 206; CHECK-NEXT: [[C:%.*]] = fsub fast float [[Z:%.*]], [[TMP1]] 207; CHECK-NEXT: ret float [[C]] 208; 209 %A = fneg fast float %X 210 %B = fmul fast float %A, %Y 211 %C = fadd fast float %B, %Z 212 ret float %C 213} 214 215define float @test8_reassoc(float %X, float %Y, float %Z) { 216; CHECK-LABEL: @test8_reassoc( 217; CHECK-NEXT: [[A:%.*]] = fsub reassoc float 0.000000e+00, [[X:%.*]] 218; CHECK-NEXT: [[B:%.*]] = fmul reassoc float [[A]], [[Y:%.*]] 219; CHECK-NEXT: [[C:%.*]] = fadd reassoc float [[B]], [[Z:%.*]] 220; CHECK-NEXT: ret float [[C]] 221; 222 %A = fsub reassoc float 0.0, %X 223 %B = fmul reassoc float %A, %Y 224 %C = fadd reassoc float %B, %Z 225 ret float %C 226} 227 228define float @test9(float %X) { 229; CHECK-LABEL: @test9( 230; CHECK-NEXT: [[FACTOR:%.*]] = fmul fast float [[X:%.*]], 9.400000e+01 231; CHECK-NEXT: ret float [[FACTOR]] 232; 233 %Y = fmul fast float %X, 4.700000e+01 234 %Z = fadd fast float %Y, %Y 235 ret float %Z 236} 237 238; Check again with 'reassoc' and 'nsz' ('nsz' not technically required). 239define float @test9_reassoc_nsz(float %X) { 240; CHECK-LABEL: @test9_reassoc_nsz( 241; CHECK-NEXT: [[TMP1:%.*]] = fmul reassoc nsz float [[X:%.*]], 9.400000e+01 242; CHECK-NEXT: ret float [[TMP1]] 243; 244 %Y = fmul reassoc nsz float %X, 4.700000e+01 245 %Z = fadd reassoc nsz float %Y, %Y 246 ret float %Z 247} 248 249; TODO: This doesn't require 'nsz'. It should fold to X * 94.0 250define float @test9_reassoc(float %X) { 251; CHECK-LABEL: @test9_reassoc( 252; CHECK-NEXT: [[Y:%.*]] = fmul reassoc float [[X:%.*]], 4.700000e+01 253; CHECK-NEXT: [[Z:%.*]] = fadd reassoc float [[Y]], [[Y]] 254; CHECK-NEXT: ret float [[Z]] 255; 256 %Y = fmul reassoc float %X, 4.700000e+01 257 %Z = fadd reassoc float %Y, %Y 258 ret float %Z 259} 260 261; Side note: (x + x + x) and (3*x) each have only a single rounding. So 262; transforming x+x+x to 3*x is always safe, even without any FMF. 263; To avoid that special-case, we have the addition of 'x' four times, here. 264define float @test10(float %X) { 265; CHECK-LABEL: @test10( 266; CHECK-NEXT: [[FACTOR:%.*]] = fmul fast float [[X:%.*]], 4.000000e+00 267; CHECK-NEXT: ret float [[FACTOR]] 268; 269 %Y = fadd fast float %X ,%X 270 %Z = fadd fast float %Y, %X 271 %W = fadd fast float %Z, %X 272 ret float %W 273} 274 275; Check again with 'reassoc' and 'nsz' ('nsz' not technically required). 276define float @test10_reassoc_nsz(float %X) { 277; CHECK-LABEL: @test10_reassoc_nsz( 278; CHECK-NEXT: [[TMP1:%.*]] = fmul reassoc nsz float [[X:%.*]], 4.000000e+00 279; CHECK-NEXT: ret float [[TMP1]] 280; 281 %Y = fadd reassoc nsz float %X ,%X 282 %Z = fadd reassoc nsz float %Y, %X 283 %W = fadd reassoc nsz float %Z, %X 284 ret float %W 285} 286 287; TODO: This doesn't require 'nsz'. It should fold to 4 * x 288define float @test10_reassoc(float %X) { 289; CHECK-LABEL: @test10_reassoc( 290; CHECK-NEXT: [[Y:%.*]] = fadd reassoc float [[X:%.*]], [[X]] 291; CHECK-NEXT: [[Z:%.*]] = fadd reassoc float [[Y]], [[X]] 292; CHECK-NEXT: [[W:%.*]] = fadd reassoc float [[Z]], [[X]] 293; CHECK-NEXT: ret float [[W]] 294; 295 %Y = fadd reassoc float %X ,%X 296 %Z = fadd reassoc float %Y, %X 297 %W = fadd reassoc float %Z, %X 298 ret float %W 299} 300 301define float @test11(float %W) { 302; CHECK-LABEL: @test11( 303; CHECK-NEXT: [[FACTOR:%.*]] = fmul fast float [[W:%.*]], 3.810000e+02 304; CHECK-NEXT: ret float [[FACTOR]] 305; 306 %X = fmul fast float %W, 127.0 307 %Y = fadd fast float %X ,%X 308 %Z = fadd fast float %Y, %X 309 ret float %Z 310} 311 312; Check again using the minimal subset of FMF. 313; Check again with 'reassoc' and 'nsz' ('nsz' not technically required). 314define float @test11_reassoc_nsz(float %W) { 315; CHECK-LABEL: @test11_reassoc_nsz( 316; CHECK-NEXT: [[Z:%.*]] = fmul reassoc nsz float [[W:%.*]], 3.810000e+02 317; CHECK-NEXT: ret float [[Z]] 318; 319 %X = fmul reassoc nsz float %W, 127.0 320 %Y = fadd reassoc nsz float %X ,%X 321 %Z = fadd reassoc nsz float %Y, %X 322 ret float %Z 323} 324 325; TODO: This doesn't require 'nsz'. It should fold to W*381.0. 326define float @test11_reassoc(float %W) { 327; CHECK-LABEL: @test11_reassoc( 328; CHECK-NEXT: [[X:%.*]] = fmul reassoc float [[W:%.*]], 1.270000e+02 329; CHECK-NEXT: [[Y:%.*]] = fadd reassoc float [[X]], [[X]] 330; CHECK-NEXT: [[Z:%.*]] = fadd reassoc float [[X]], [[Y]] 331; CHECK-NEXT: ret float [[Z]] 332; 333 %X = fmul reassoc float %W, 127.0 334 %Y = fadd reassoc float %X ,%X 335 %Z = fadd reassoc float %Y, %X 336 ret float %Z 337} 338 339define float @test12(float %X) { 340; CHECK-LABEL: @test12( 341; CHECK-NEXT: [[FACTOR:%.*]] = fmul fast float [[X:%.*]], -3.000000e+00 342; CHECK-NEXT: [[Z:%.*]] = fadd fast float [[FACTOR]], 6.000000e+00 343; CHECK-NEXT: ret float [[Z]] 344; 345 %A = fsub fast float 1.000000e+00, %X 346 %B = fsub fast float 2.000000e+00, %X 347 %C = fsub fast float 3.000000e+00, %X 348 %Y = fadd fast float %A ,%B 349 %Z = fadd fast float %Y, %C 350 ret float %Z 351} 352 353; Check again with 'reassoc' and 'nsz' ('nsz' not technically required). 354define float @test12_reassoc_nsz(float %X) { 355; CHECK-LABEL: @test12_reassoc_nsz( 356; CHECK-NEXT: [[TMP1:%.*]] = fmul reassoc nsz float [[X:%.*]], -3.000000e+00 357; CHECK-NEXT: [[TMP2:%.*]] = fadd reassoc nsz float [[TMP1]], 6.000000e+00 358; CHECK-NEXT: ret float [[TMP2]] 359; 360 %A = fsub reassoc nsz float 1.000000e+00, %X 361 %B = fsub reassoc nsz float 2.000000e+00, %X 362 %C = fsub reassoc nsz float 3.000000e+00, %X 363 %Y = fadd reassoc nsz float %A ,%B 364 %Z = fadd reassoc nsz float %Y, %C 365 ret float %Z 366} 367 368; TODO: This doesn't require 'nsz'. It should fold to (6.0 - 3.0*x) 369define float @test12_reassoc(float %X) { 370; CHECK-LABEL: @test12_reassoc( 371; CHECK-NEXT: [[A:%.*]] = fsub reassoc float 1.000000e+00, [[X:%.*]] 372; CHECK-NEXT: [[B:%.*]] = fsub reassoc float 2.000000e+00, [[X]] 373; CHECK-NEXT: [[C:%.*]] = fsub reassoc float 3.000000e+00, [[X]] 374; CHECK-NEXT: [[Y:%.*]] = fadd reassoc float [[A]], [[B]] 375; CHECK-NEXT: [[Z:%.*]] = fadd reassoc float [[C]], [[Y]] 376; CHECK-NEXT: ret float [[Z]] 377; 378 %A = fsub reassoc float 1.000000e+00, %X 379 %B = fsub reassoc float 2.000000e+00, %X 380 %C = fsub reassoc float 3.000000e+00, %X 381 %Y = fadd reassoc float %A ,%B 382 %Z = fadd reassoc float %Y, %C 383 ret float %Z 384} 385 386define float @test13(float %X1, float %X2, float %X3) { 387; CHECK-LABEL: @test13( 388; CHECK-NEXT: [[REASS_ADD:%.*]] = fsub fast float [[X3:%.*]], [[X2:%.*]] 389; CHECK-NEXT: [[REASS_MUL:%.*]] = fmul fast float [[REASS_ADD]], [[X1:%.*]] 390; CHECK-NEXT: ret float [[REASS_MUL]] 391; 392 %A = fsub fast float 0.000000e+00, %X1 393 %B = fmul fast float %A, %X2 ; -X1*X2 394 %C = fmul fast float %X1, %X3 ; X1*X3 395 %D = fadd fast float %B, %C ; -X1*X2 + X1*X3 -> X1*(X3-X2) 396 ret float %D 397} 398 399define float @test13_unary_fneg(float %X1, float %X2, float %X3) { 400; CHECK-LABEL: @test13_unary_fneg( 401; CHECK-NEXT: [[TMP1:%.*]] = fsub fast float [[X3:%.*]], [[X2:%.*]] 402; CHECK-NEXT: [[D:%.*]] = fmul fast float [[TMP1]], [[X1:%.*]] 403; CHECK-NEXT: ret float [[D]] 404; 405 %A = fneg fast float %X1 406 %B = fmul fast float %A, %X2 ; -X1*X2 407 %C = fmul fast float %X1, %X3 ; X1*X3 408 %D = fadd fast float %B, %C ; -X1*X2 + X1*X3 -> X1*(X3-X2) 409 ret float %D 410} 411 412define float @test13_reassoc(float %X1, float %X2, float %X3) { 413; CHECK-LABEL: @test13_reassoc( 414; CHECK-NEXT: [[A:%.*]] = fsub reassoc float 0.000000e+00, [[X1:%.*]] 415; CHECK-NEXT: [[B:%.*]] = fmul reassoc float [[A]], [[X2:%.*]] 416; CHECK-NEXT: [[C:%.*]] = fmul reassoc float [[X1]], [[X3:%.*]] 417; CHECK-NEXT: [[D:%.*]] = fadd reassoc float [[B]], [[C]] 418; CHECK-NEXT: ret float [[D]] 419; 420 %A = fsub reassoc float 0.000000e+00, %X1 421 %B = fmul reassoc float %A, %X2 ; -X1*X2 422 %C = fmul reassoc float %X1, %X3 ; X1*X3 423 %D = fadd reassoc float %B, %C ; -X1*X2 + X1*X3 -> X1*(X3-X2) 424 ret float %D 425} 426 427define float @test14(float %X1, float %X2) { 428; CHECK-LABEL: @test14( 429; CHECK-NEXT: [[TMP1:%.*]] = fsub fast float [[X1:%.*]], [[X2:%.*]] 430; CHECK-NEXT: [[D1:%.*]] = fmul fast float [[TMP1]], 4.700000e+01 431; CHECK-NEXT: ret float [[D1]] 432; 433 %B = fmul fast float %X1, 47. ; X1*47 434 %C = fmul fast float %X2, -47. ; X2*-47 435 %D = fadd fast float %B, %C ; X1*47 + X2*-47 -> 47*(X1-X2) 436 ret float %D 437} 438 439; (x1 * 47) + (x2 * -47) => (x1 - x2) * 47 440; Check again with 'reassoc' and 'nsz' ('nsz' not technically required). 441define float @test14_reassoc_nsz(float %X1, float %X2) { 442; CHECK-LABEL: @test14_reassoc_nsz( 443; CHECK-NEXT: [[TMP1:%.*]] = fsub reassoc nsz float [[X1:%.*]], [[X2:%.*]] 444; CHECK-NEXT: [[D1:%.*]] = fmul reassoc nsz float [[TMP1]], 4.700000e+01 445; CHECK-NEXT: ret float [[D1]] 446; 447 %B = fmul reassoc nsz float %X1, 47. ; X1*47 448 %C = fmul reassoc nsz float %X2, -47. ; X2*-47 449 %D = fadd reassoc nsz float %B, %C ; X1*47 + X2*-47 -> 47*(X1-X2) 450 ret float %D 451} 452 453; TODO: This doesn't require 'nsz'. It should fold to ((x1 - x2) * 47.0) 454define float @test14_reassoc(float %X1, float %X2) { 455; CHECK-LABEL: @test14_reassoc( 456; CHECK-NEXT: [[B:%.*]] = fmul reassoc float [[X1:%.*]], 4.700000e+01 457; CHECK-NEXT: [[C:%.*]] = fmul reassoc float [[X2:%.*]], 4.700000e+01 458; CHECK-NEXT: [[D1:%.*]] = fsub reassoc float [[B]], [[C]] 459; CHECK-NEXT: ret float [[D1]] 460; 461 %B = fmul reassoc float %X1, 47. ; X1*47 462 %C = fmul reassoc float %X2, -47. ; X2*-47 463 %D = fadd reassoc float %B, %C ; X1*47 + X2*-47 -> 47*(X1-X2) 464 ret float %D 465} 466 467define float @test15(float %arg) { 468; CHECK-LABEL: @test15( 469; CHECK-NEXT: [[T2:%.*]] = fmul fast float [[ARG:%.*]], 1.440000e+02 470; CHECK-NEXT: ret float [[T2]] 471; 472 %t1 = fmul fast float 1.200000e+01, %arg 473 %t2 = fmul fast float %t1, 1.200000e+01 474 ret float %t2 475} 476 477define float @test15_reassoc(float %arg) { 478; CHECK-LABEL: @test15_reassoc( 479; CHECK-NEXT: [[T2:%.*]] = fmul reassoc float [[ARG:%.*]], 1.440000e+02 480; CHECK-NEXT: ret float [[T2]] 481; 482 %t1 = fmul reassoc float 1.200000e+01, %arg 483 %t2 = fmul reassoc float %t1, 1.200000e+01 484 ret float %t2 485} 486 487; (b+(a+1234))+-a -> b+1234 488define float @test16(float %b, float %a) { 489; CHECK-LABEL: @test16( 490; CHECK-NEXT: [[TMP1:%.*]] = fadd fast float [[B:%.*]], 1.234000e+03 491; CHECK-NEXT: ret float [[TMP1]] 492; 493 %1 = fadd fast float %a, 1234.0 494 %2 = fadd fast float %b, %1 495 %3 = fsub fast float 0.0, %a 496 %4 = fadd fast float %2, %3 497 ret float %4 498} 499 500define float @test16_unary_fneg(float %b, float %a) { 501; CHECK-LABEL: @test16_unary_fneg( 502; CHECK-NEXT: [[TMP1:%.*]] = fadd fast float [[B:%.*]], 1.234000e+03 503; CHECK-NEXT: ret float [[TMP1]] 504; 505 %1 = fadd fast float %a, 1234.0 506 %2 = fadd fast float %b, %1 507 %3 = fneg fast float %a 508 %4 = fadd fast float %2, %3 509 ret float %4 510} 511 512define float @test16_reassoc(float %b, float %a) { 513; CHECK-LABEL: @test16_reassoc( 514; CHECK-NEXT: [[TMP1:%.*]] = fadd reassoc float [[A:%.*]], 1.234000e+03 515; CHECK-NEXT: [[TMP2:%.*]] = fadd reassoc float [[TMP1]], [[B:%.*]] 516; CHECK-NEXT: [[TMP3:%.*]] = fsub reassoc float 0.000000e+00, [[A]] 517; CHECK-NEXT: [[TMP4:%.*]] = fadd reassoc float [[TMP3]], [[TMP2]] 518; CHECK-NEXT: ret float [[TMP4]] 519; 520 %1 = fadd reassoc float %a, 1234.0 521 %2 = fadd reassoc float %b, %1 522 %3 = fsub reassoc float 0.0, %a 523 %4 = fadd reassoc float %2, %3 524 ret float %4 525} 526 527; Test that we can turn things like X*-(Y*Z) -> X*-1*Y*Z. 528 529define float @test17(float %a, float %b, float %z) { 530; CHECK-LABEL: @test17( 531; CHECK-NEXT: [[E:%.*]] = fmul fast float [[A:%.*]], 1.234500e+04 532; CHECK-NEXT: [[F:%.*]] = fmul fast float [[E]], [[B:%.*]] 533; CHECK-NEXT: [[G:%.*]] = fmul fast float [[F]], [[Z:%.*]] 534; CHECK-NEXT: ret float [[G]] 535; 536 %c = fsub fast float 0.000000e+00, %z 537 %d = fmul fast float %a, %b 538 %e = fmul fast float %c, %d 539 %f = fmul fast float %e, 1.234500e+04 540 %g = fsub fast float 0.000000e+00, %f 541 ret float %g 542} 543 544define float @test17_unary_fneg(float %a, float %b, float %z) { 545; CHECK-LABEL: @test17_unary_fneg( 546; CHECK-NEXT: [[D:%.*]] = fmul fast float [[A:%.*]], 1.234500e+04 547; CHECK-NEXT: [[E:%.*]] = fmul fast float [[D]], [[B:%.*]] 548; CHECK-NEXT: [[TMP1:%.*]] = fmul fast float [[E]], [[Z:%.*]] 549; CHECK-NEXT: ret float [[TMP1]] 550; 551 %c = fneg fast float %z 552 %d = fmul fast float %a, %b 553 %e = fmul fast float %c, %d 554 %f = fmul fast float %e, 1.234500e+04 555 %g = fneg fast float %f 556 ret float %g 557} 558 559define float @test17_reassoc(float %a, float %b, float %z) { 560; CHECK-LABEL: @test17_reassoc( 561; CHECK-NEXT: [[C:%.*]] = fsub reassoc float 0.000000e+00, [[Z:%.*]] 562; CHECK-NEXT: [[D:%.*]] = fmul reassoc float [[A:%.*]], [[B:%.*]] 563; CHECK-NEXT: [[E:%.*]] = fmul reassoc float [[D]], [[C]] 564; CHECK-NEXT: [[F:%.*]] = fmul reassoc float [[E]], 1.234500e+04 565; CHECK-NEXT: [[G:%.*]] = fsub reassoc float 0.000000e+00, [[F]] 566; CHECK-NEXT: ret float [[G]] 567; 568 %c = fsub reassoc float 0.000000e+00, %z 569 %d = fmul reassoc float %a, %b 570 %e = fmul reassoc float %c, %d 571 %f = fmul reassoc float %e, 1.234500e+04 572 %g = fsub reassoc float 0.000000e+00, %f 573 ret float %g 574} 575 576define float @test18(float %a, float %b, float %z) { 577; CHECK-LABEL: @test18( 578; CHECK-NEXT: [[E:%.*]] = fmul fast float [[A:%.*]], 4.000000e+01 579; CHECK-NEXT: [[F:%.*]] = fmul fast float [[E]], [[Z:%.*]] 580; CHECK-NEXT: ret float [[F]] 581; 582 %d = fmul fast float %z, 4.000000e+01 583 %c = fsub fast float 0.000000e+00, %d 584 %e = fmul fast float %a, %c 585 %f = fsub fast float 0.000000e+00, %e 586 ret float %f 587} 588 589define float @test18_unary_fneg(float %a, float %b, float %z) { 590; CHECK-LABEL: @test18_unary_fneg( 591; CHECK-NEXT: [[E:%.*]] = fmul fast float [[A:%.*]], 4.000000e+01 592; CHECK-NEXT: [[F:%.*]] = fmul fast float [[E]], [[Z:%.*]] 593; CHECK-NEXT: ret float [[F]] 594; 595 %d = fmul fast float %z, 4.000000e+01 596 %c = fneg fast float %d 597 %e = fmul fast float %a, %c 598 %f = fneg fast float %e 599 ret float %f 600} 601 602define float @test18_reassoc(float %a, float %b, float %z) { 603; CHECK-LABEL: @test18_reassoc( 604; CHECK-NEXT: [[D:%.*]] = fmul reassoc float [[Z:%.*]], 4.000000e+01 605; CHECK-NEXT: [[C:%.*]] = fsub reassoc float 0.000000e+00, [[D]] 606; CHECK-NEXT: [[E:%.*]] = fmul reassoc float [[C]], [[A:%.*]] 607; CHECK-NEXT: [[F:%.*]] = fsub reassoc float 0.000000e+00, [[E]] 608; CHECK-NEXT: ret float [[F]] 609; 610 %d = fmul reassoc float %z, 4.000000e+01 611 %c = fsub reassoc float 0.000000e+00, %d 612 %e = fmul reassoc float %a, %c 613 %f = fsub reassoc float 0.000000e+00, %e 614 ret float %f 615} 616 617; fneg of fneg is an identity operation, so no FMF are needed to remove those instructions. 618 619define float @test18_unary_fneg_no_FMF(float %a, float %b, float %z) { 620; CHECK-LABEL: @test18_unary_fneg_no_FMF( 621; CHECK-NEXT: [[TMP1:%.*]] = fmul float [[Z:%.*]], 4.000000e+01 622; CHECK-NEXT: [[F:%.*]] = fmul float [[TMP1]], [[A:%.*]] 623; CHECK-NEXT: ret float [[F]] 624; 625 %d = fmul float %z, 4.000000e+01 626 %c = fneg float %d 627 %e = fmul float %a, %c 628 %f = fneg float %e 629 ret float %f 630} 631 632define float @test18_reassoc_unary_fneg(float %a, float %b, float %z) { 633; CHECK-LABEL: @test18_reassoc_unary_fneg( 634; CHECK-NEXT: [[TMP1:%.*]] = fmul reassoc float [[Z:%.*]], 4.000000e+01 635; CHECK-NEXT: [[F:%.*]] = fmul reassoc float [[TMP1]], [[A:%.*]] 636; CHECK-NEXT: ret float [[F]] 637; 638 %d = fmul reassoc float %z, 4.000000e+01 639 %c = fneg reassoc float %d 640 %e = fmul reassoc float %a, %c 641 %f = fneg reassoc float %e 642 ret float %f 643} 644 645; With sub reassociation, constant folding can eliminate the 12 and -12 constants. 646define float @test19(float %A, float %B) { 647; CHECK-LABEL: @test19( 648; CHECK-NEXT: [[Z:%.*]] = fsub fast float [[A:%.*]], [[B:%.*]] 649; CHECK-NEXT: ret float [[Z]] 650; 651 %X = fadd fast float -1.200000e+01, %A 652 %Y = fsub fast float %X, %B 653 %Z = fadd fast float %Y, 1.200000e+01 654 ret float %Z 655} 656 657define float @test19_reassoc(float %A, float %B) { 658; CHECK-LABEL: @test19_reassoc( 659; CHECK-NEXT: [[X:%.*]] = fadd reassoc float [[A:%.*]], -1.200000e+01 660; CHECK-NEXT: [[Y:%.*]] = fsub reassoc float [[X]], [[B:%.*]] 661; CHECK-NEXT: [[Z:%.*]] = fadd reassoc float [[Y]], 1.200000e+01 662; CHECK-NEXT: ret float [[Z]] 663; 664 %X = fadd reassoc float -1.200000e+01, %A 665 %Y = fsub reassoc float %X, %B 666 %Z = fadd reassoc float %Y, 1.200000e+01 667 ret float %Z 668} 669 670; With sub reassociation, constant folding can eliminate the uses of %a. 671define float @test20(float %a, float %b, float %c) nounwind { 672; CHECK-LABEL: @test20( 673; CHECK-NEXT: [[TMP1:%.*]] = fadd fast float [[B:%.*]], [[C:%.*]] 674; CHECK-NEXT: [[T7:%.*]] = fneg fast float [[TMP1]] 675; CHECK-NEXT: ret float [[T7]] 676; 677 %t3 = fsub fast float %a, %b 678 %t5 = fsub fast float %t3, %c 679 %t7 = fsub fast float %t5, %a 680 ret float %t7 681} 682 683define float @test20_reassoc(float %a, float %b, float %c) nounwind { 684; CHECK-LABEL: @test20_reassoc( 685; CHECK-NEXT: [[T3:%.*]] = fsub reassoc float [[A:%.*]], [[B:%.*]] 686; CHECK-NEXT: [[T5:%.*]] = fsub reassoc float [[T3]], [[C:%.*]] 687; CHECK-NEXT: [[T7:%.*]] = fsub reassoc float [[T5]], [[A]] 688; CHECK-NEXT: ret float [[T7]] 689; 690 %t3 = fsub reassoc float %a, %b 691 %t5 = fsub reassoc float %t3, %c 692 %t7 = fsub reassoc float %t5, %a 693 ret float %t7 694} 695 696