1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py 2; 3; Test numbering remains continuous across: 4; - InstCombine/fast-basictest.ll 5; - PhaseOrdering/fast-basictest.ll 6; - PhaseOrdering/fast-reassociate-gvn.ll 7; - Reassociate/fast-basictest.ll 8; 9; RUN: opt < %s -passes=reassociate -S | FileCheck %s 10 11; ((a + (-3)) + b) + 3 -> a + b 12; That only works with both instcombine and reassociate passes enabled. 13; Check that reassociate is not enough. 14 15; TODO: check if we can eliminate zero add. 16define float @test2(float %reg109, float %reg1111) { 17; CHECK-LABEL: @test2( 18; CHECK-NEXT: [[REG116:%.*]] = fadd fast float [[REG109:%.*]], 0.000000e+00 19; CHECK-NEXT: [[REG117:%.*]] = fadd fast float [[REG116]], [[REG1111:%.*]] 20; CHECK-NEXT: ret float [[REG117]] 21; 22 %reg115 = fadd fast float %reg109, -3.000000e+01 23 %reg116 = fadd fast float %reg115, %reg1111 24 %reg117 = fadd fast float %reg116, 3.000000e+01 25 ret float %reg117 26} 27 28define float @test2_no_FMF(float %reg109, float %reg1111) { 29; CHECK-LABEL: @test2_no_FMF( 30; CHECK-NEXT: [[REG115:%.*]] = fadd float [[REG109:%.*]], -3.000000e+01 31; CHECK-NEXT: [[REG116:%.*]] = fadd float [[REG115]], [[REG1111:%.*]] 32; CHECK-NEXT: [[REG117:%.*]] = fadd float [[REG116]], 3.000000e+01 33; CHECK-NEXT: ret float [[REG117]] 34; 35 %reg115 = fadd float %reg109, -3.000000e+01 36 %reg116 = fadd float %reg115, %reg1111 37 %reg117 = fadd float %reg116, 3.000000e+01 38 ret float %reg117 39} 40 41define float @test2_reassoc(float %reg109, float %reg1111) { 42; CHECK-LABEL: @test2_reassoc( 43; CHECK-NEXT: [[REG115:%.*]] = fadd reassoc float [[REG109:%.*]], -3.000000e+01 44; CHECK-NEXT: [[REG116:%.*]] = fadd reassoc float [[REG115]], [[REG1111:%.*]] 45; CHECK-NEXT: [[REG117:%.*]] = fadd reassoc float [[REG116]], 3.000000e+01 46; CHECK-NEXT: ret float [[REG117]] 47; 48 %reg115 = fadd reassoc float %reg109, -3.000000e+01 49 %reg116 = fadd reassoc float %reg115, %reg1111 50 %reg117 = fadd reassoc float %reg116, 3.000000e+01 51 ret float %reg117 52} 53 54@fe = external global float 55@fa = external global float 56@fb = external global float 57@fc = external global float 58@ff = external global float 59 60; If two sums of the same operands in different order are counted with 'fast' 61; flag and then stored to global variables, we can reuse the same value twice. 62; Sums: 63; - test3: (a+b)+c and (a+c)+b 64; - test4: c+(a+b) and (c+a)+b 65; - test5: c+(b+a) and (c+a)+b 66; 67; That only works with both gvn and reassociate passes enabled. 68; Check that reassociate is not enough. 69 70define void @test3() { 71; CHECK-LABEL: @test3( 72; CHECK-NEXT: [[A:%.*]] = load float, ptr @fa, align 4 73; CHECK-NEXT: [[B:%.*]] = load float, ptr @fb, align 4 74; CHECK-NEXT: [[C:%.*]] = load float, ptr @fc, align 4 75; CHECK-NEXT: [[T1:%.*]] = fadd fast float [[B]], [[A]] 76; CHECK-NEXT: [[T2:%.*]] = fadd fast float [[T1]], [[C]] 77; CHECK-NEXT: [[T3:%.*]] = fadd fast float [[B]], [[A]] 78; CHECK-NEXT: [[T4:%.*]] = fadd fast float [[T3]], [[C]] 79; CHECK-NEXT: store float [[T2]], ptr @fe, align 4 80; CHECK-NEXT: store float [[T4]], ptr @ff, align 4 81; CHECK-NEXT: ret void 82; 83 %A = load float, ptr @fa 84 %B = load float, ptr @fb 85 %C = load float, ptr @fc 86 %t1 = fadd fast float %A, %B 87 %t2 = fadd fast float %t1, %C 88 %t3 = fadd fast float %A, %C 89 %t4 = fadd fast float %t3, %B 90 ; e = (a+b)+c; 91 store float %t2, ptr @fe 92 ; f = (a+c)+b 93 store float %t4, ptr @ff 94 ret void 95} 96 97define void @test4() { 98; CHECK-LABEL: @test4( 99; CHECK-NEXT: [[A:%.*]] = load float, ptr @fa, align 4 100; CHECK-NEXT: [[B:%.*]] = load float, ptr @fb, align 4 101; CHECK-NEXT: [[C:%.*]] = load float, ptr @fc, align 4 102; CHECK-NEXT: [[T1:%.*]] = fadd fast float [[B]], [[A]] 103; CHECK-NEXT: [[T2:%.*]] = fadd fast float [[T1]], [[C]] 104; CHECK-NEXT: [[T3:%.*]] = fadd fast float [[B]], [[A]] 105; CHECK-NEXT: [[T4:%.*]] = fadd fast float [[T3]], [[C]] 106; CHECK-NEXT: store float [[T2]], ptr @fe, align 4 107; CHECK-NEXT: store float [[T4]], ptr @ff, align 4 108; CHECK-NEXT: ret void 109; 110 %A = load float, ptr @fa 111 %B = load float, ptr @fb 112 %C = load float, ptr @fc 113 %t1 = fadd fast float %A, %B 114 %t2 = fadd fast float %C, %t1 115 %t3 = fadd fast float %C, %A 116 %t4 = fadd fast float %t3, %B 117 ; e = c+(a+b) 118 store float %t2, ptr @fe 119 ; f = (c+a)+b 120 store float %t4, ptr @ff 121 ret void 122} 123 124define void @test5() { 125; CHECK-LABEL: @test5( 126; CHECK-NEXT: [[A:%.*]] = load float, ptr @fa, align 4 127; CHECK-NEXT: [[B:%.*]] = load float, ptr @fb, align 4 128; CHECK-NEXT: [[C:%.*]] = load float, ptr @fc, align 4 129; CHECK-NEXT: [[T1:%.*]] = fadd fast float [[B]], [[A]] 130; CHECK-NEXT: [[T2:%.*]] = fadd fast float [[T1]], [[C]] 131; CHECK-NEXT: [[T3:%.*]] = fadd fast float [[B]], [[A]] 132; CHECK-NEXT: [[T4:%.*]] = fadd fast float [[T3]], [[C]] 133; CHECK-NEXT: store float [[T2]], ptr @fe, align 4 134; CHECK-NEXT: store float [[T4]], ptr @ff, align 4 135; CHECK-NEXT: ret void 136; 137 %A = load float, ptr @fa 138 %B = load float, ptr @fb 139 %C = load float, ptr @fc 140 %t1 = fadd fast float %B, %A 141 %t2 = fadd fast float %C, %t1 142 %t3 = fadd fast float %C, %A 143 %t4 = fadd fast float %t3, %B 144 ; e = c+(b+a) 145 store float %t2, ptr @fe 146 ; f = (c+a)+b 147 store float %t4, ptr @ff 148 ret void 149} 150 151define float @test6(float %A, float %B, float %C) { 152; CHECK-LABEL: @test6( 153; CHECK-NEXT: [[REASS_ADD1:%.*]] = fadd fast float [[C:%.*]], [[B:%.*]] 154; CHECK-NEXT: [[REASS_MUL2:%.*]] = fmul fast float [[A:%.*]], [[A]] 155; CHECK-NEXT: [[REASS_MUL:%.*]] = fmul fast float [[REASS_MUL2]], [[REASS_ADD1]] 156; CHECK-NEXT: ret float [[REASS_MUL]] 157; 158 %aa = fmul fast float %A, %A 159 %aab = fmul fast float %aa, %B 160 %ac = fmul fast float %A, %C 161 %aac = fmul fast float %ac, %A 162 %r = fadd fast float %aab, %aac 163 ret float %r 164} 165 166define float @test6_reassoc(float %A, float %B, float %C) { 167; CHECK-LABEL: @test6_reassoc( 168; CHECK-NEXT: [[AA:%.*]] = fmul reassoc float [[A:%.*]], [[A]] 169; CHECK-NEXT: [[AAB:%.*]] = fmul reassoc float [[AA]], [[B:%.*]] 170; CHECK-NEXT: [[AC:%.*]] = fmul reassoc float [[A]], [[C:%.*]] 171; CHECK-NEXT: [[AAC:%.*]] = fmul reassoc float [[A]], [[AC]] 172; CHECK-NEXT: [[R:%.*]] = fadd reassoc float [[AAB]], [[AAC]] 173; CHECK-NEXT: ret float [[R]] 174; 175 %aa = fmul reassoc float %A, %A 176 %aab = fmul reassoc float %aa, %B 177 %ac = fmul reassoc float %A, %C 178 %aac = fmul reassoc float %ac, %A 179 %r = fadd reassoc float %aab, %aac 180 ret float %r 181} 182 183; (-X)*Y + Z -> Z-X*Y 184define float @test7(float %X, float %Y, float %Z) { 185; CHECK-LABEL: @test7( 186; CHECK-NEXT: [[B:%.*]] = fmul fast float [[Y:%.*]], [[X:%.*]] 187; CHECK-NEXT: [[TMP1:%.*]] = fsub fast float [[Z:%.*]], [[B]] 188; CHECK-NEXT: ret float [[TMP1]] 189; 190 %A = fsub fast float 0.0, %X 191 %B = fmul fast float %A, %Y 192 %C = fadd fast float %B, %Z 193 ret float %C 194} 195 196define float @test7_unary_fneg(float %X, float %Y, float %Z) { 197; CHECK-LABEL: @test7_unary_fneg( 198; CHECK-NEXT: [[B:%.*]] = fmul fast float [[Y:%.*]], [[X:%.*]] 199; CHECK-NEXT: [[TMP1:%.*]] = fsub fast float [[Z:%.*]], [[B]] 200; CHECK-NEXT: ret float [[TMP1]] 201; 202 %A = fneg fast float %X 203 %B = fmul fast float %A, %Y 204 %C = fadd fast float %B, %Z 205 ret float %C 206} 207 208define float @test7_reassoc_nsz(float %X, float %Y, float %Z) { 209; CHECK-LABEL: @test7_reassoc_nsz( 210; CHECK-NEXT: [[B:%.*]] = fmul reassoc nsz float [[Y:%.*]], [[X:%.*]] 211; CHECK-NEXT: [[TMP1:%.*]] = fsub reassoc nsz float [[Z:%.*]], [[B]] 212; CHECK-NEXT: ret float [[TMP1]] 213; 214 %A = fsub reassoc nsz float 0.0, %X 215 %B = fmul reassoc nsz float %A, %Y 216 %C = fadd reassoc nsz float %B, %Z 217 ret float %C 218} 219 220; Verify that fold is not done only with 'reassoc' ('nsz' is required) 221define float @test7_reassoc(float %X, float %Y, float %Z) { 222; CHECK-LABEL: @test7_reassoc( 223; CHECK-NEXT: [[A:%.*]] = fsub reassoc float 0.000000e+00, [[X:%.*]] 224; CHECK-NEXT: [[B:%.*]] = fmul reassoc float [[A]], [[Y:%.*]] 225; CHECK-NEXT: [[C:%.*]] = fadd reassoc float [[B]], [[Z:%.*]] 226; CHECK-NEXT: ret float [[C]] 227; 228 %A = fsub reassoc float 0.0, %X 229 %B = fmul reassoc float %A, %Y 230 %C = fadd reassoc float %B, %Z 231 ret float %C 232} 233 234; Integer version of: 235; (-X)*Y + Z -> Z-X*Y 236; TODO: check if we can change the mul of -1 and the add to a sub. 237define i32 @test7_int(i32 %X, i32 %Y, i32 %Z) { 238; CHECK-LABEL: @test7_int( 239; CHECK-NEXT: [[A:%.*]] = mul i32 [[Y:%.*]], [[X:%.*]] 240; CHECK-NEXT: [[B:%.*]] = mul i32 [[A]], -1 241; CHECK-NEXT: [[C:%.*]] = add i32 [[B]], [[Z:%.*]] 242; CHECK-NEXT: ret i32 [[C]] 243; 244 %A = sub i32 0, %X 245 %B = mul i32 %A, %Y 246 %C = add i32 %B, %Z 247 ret i32 %C 248} 249 250define float @test8(float %X) { 251; CHECK-LABEL: @test8( 252; CHECK-NEXT: [[FACTOR:%.*]] = fmul fast float [[X:%.*]], 9.400000e+01 253; CHECK-NEXT: ret float [[FACTOR]] 254; 255 %Y = fmul fast float %X, 4.700000e+01 256 %Z = fadd fast float %Y, %Y 257 ret float %Z 258} 259 260; Side note: (x + x + x) and (3*x) each have only a single rounding. So 261; transforming x+x+x to 3*x is always safe, even without any FMF. 262; To avoid that special-case, we have the addition of 'x' four times, here. 263define float @test9(float %X) { 264; CHECK-LABEL: @test9( 265; CHECK-NEXT: [[FACTOR:%.*]] = fmul fast float [[X:%.*]], 4.000000e+00 266; CHECK-NEXT: ret float [[FACTOR]] 267; 268 %Y = fadd fast float %X ,%X 269 %Z = fadd fast float %Y, %X 270 %W = fadd fast float %Z, %X 271 ret float %W 272} 273 274define float @test10(float %W) { 275; CHECK-LABEL: @test10( 276; CHECK-NEXT: [[FACTOR:%.*]] = fmul fast float [[W:%.*]], 3.810000e+02 277; CHECK-NEXT: ret float [[FACTOR]] 278; 279 %X = fmul fast float %W, 127.0 280 %Y = fadd fast float %X ,%X 281 %Z = fadd fast float %Y, %X 282 ret float %Z 283} 284 285define float @test11(float %X) { 286; CHECK-LABEL: @test11( 287; CHECK-NEXT: [[FACTOR:%.*]] = fmul fast float [[X:%.*]], -3.000000e+00 288; CHECK-NEXT: [[Z:%.*]] = fadd fast float [[FACTOR]], 6.000000e+00 289; CHECK-NEXT: ret float [[Z]] 290; 291 %A = fsub fast float 1.000000e+00, %X 292 %B = fsub fast float 2.000000e+00, %X 293 %C = fsub fast float 3.000000e+00, %X 294 %Y = fadd fast float %A ,%B 295 %Z = fadd fast float %Y, %C 296 ret float %Z 297} 298 299define float @test12(float %X1, float %X2, float %X3) { 300; CHECK-LABEL: @test12( 301; CHECK-NEXT: [[B:%.*]] = fmul fast float [[X2:%.*]], [[X1:%.*]] 302; CHECK-NEXT: [[C:%.*]] = fmul fast float [[X3:%.*]], [[X1]] 303; CHECK-NEXT: [[TMP1:%.*]] = fsub fast float [[C]], [[B]] 304; CHECK-NEXT: ret float [[TMP1]] 305; 306 %A = fsub fast float 0.000000e+00, %X1 307 %B = fmul fast float %A, %X2 ; -X1*X2 308 %C = fmul fast float %X1, %X3 ; X1*X3 309 %D = fadd fast float %B, %C ; -X1*X2 + X1*X3 -> X1*(X3-X2) 310 ret float %D 311} 312 313define float @test12_unary_fneg(float %X1, float %X2, float %X3) { 314; CHECK-LABEL: @test12_unary_fneg( 315; CHECK-NEXT: [[B:%.*]] = fmul fast float [[X2:%.*]], [[X1:%.*]] 316; CHECK-NEXT: [[C:%.*]] = fmul fast float [[X3:%.*]], [[X1]] 317; CHECK-NEXT: [[TMP1:%.*]] = fsub fast float [[C]], [[B]] 318; CHECK-NEXT: ret float [[TMP1]] 319; 320 %A = fneg fast float %X1 321 %B = fmul fast float %A, %X2 ; -X1*X2 322 %C = fmul fast float %X1, %X3 ; X1*X3 323 %D = fadd fast float %B, %C ; -X1*X2 + X1*X3 -> X1*(X3-X2) 324 ret float %D 325} 326 327define float @test12_reassoc_nsz(float %X1, float %X2, float %X3) { 328; CHECK-LABEL: @test12_reassoc_nsz( 329; CHECK-NEXT: [[B:%.*]] = fmul reassoc nsz float [[X2:%.*]], [[X1:%.*]] 330; CHECK-NEXT: [[C:%.*]] = fmul reassoc nsz float [[X3:%.*]], [[X1]] 331; CHECK-NEXT: [[TMP1:%.*]] = fsub reassoc nsz float [[C]], [[B]] 332; CHECK-NEXT: ret float [[TMP1]] 333; 334 %A = fsub reassoc nsz float 0.000000e+00, %X1 335 %B = fmul reassoc nsz float %A, %X2 ; -X1*X2 336 %C = fmul reassoc nsz float %X1, %X3 ; X1*X3 337 %D = fadd reassoc nsz float %B, %C ; -X1*X2 + X1*X3 -> X1*(X3-X2) 338 ret float %D 339} 340 341; TODO: check if 'nsz' is technically required. Currently the optimization 342; is not done with only 'reassoc' without 'nsz'. 343define float @test12_reassoc(float %X1, float %X2, float %X3) { 344; CHECK-LABEL: @test12_reassoc( 345; CHECK-NEXT: [[A:%.*]] = fsub reassoc float 0.000000e+00, [[X1:%.*]] 346; CHECK-NEXT: [[B:%.*]] = fmul reassoc float [[A]], [[X2:%.*]] 347; CHECK-NEXT: [[C:%.*]] = fmul reassoc float [[X1]], [[X3:%.*]] 348; CHECK-NEXT: [[D:%.*]] = fadd reassoc float [[B]], [[C]] 349; CHECK-NEXT: ret float [[D]] 350; 351 %A = fsub reassoc float 0.000000e+00, %X1 352 %B = fmul reassoc float %A, %X2 ; -X1*X2 353 %C = fmul reassoc float %X1, %X3 ; X1*X3 354 %D = fadd reassoc float %B, %C ; -X1*X2 + X1*X3 -> X1*(X3-X2) 355 ret float %D 356} 357 358; (x1 * 47) + (x2 * -47) => (x1 - x2) * 47 359; That only works with both instcombine and reassociate passes enabled. 360; Check that reassociate is not enough. 361 362define float @test13(float %X1, float %X2) { 363; CHECK-LABEL: @test13( 364; CHECK-NEXT: [[B:%.*]] = fmul fast float [[X1:%.*]], 4.700000e+01 365; CHECK-NEXT: [[C:%.*]] = fmul fast float [[X2:%.*]], 4.700000e+01 366; CHECK-NEXT: [[TMP1:%.*]] = fsub fast float [[B]], [[C]] 367; CHECK-NEXT: ret float [[TMP1]] 368; 369 %B = fmul fast float %X1, 47. ; X1*47 370 %C = fmul fast float %X2, -47. ; X2*-47 371 %D = fadd fast float %B, %C ; X1*47 + X2*-47 -> 47*(X1-X2) 372 ret float %D 373} 374 375define float @test13_reassoc_nsz(float %X1, float %X2) { 376; CHECK-LABEL: @test13_reassoc_nsz( 377; CHECK-NEXT: [[B:%.*]] = fmul reassoc nsz float [[X1:%.*]], 4.700000e+01 378; CHECK-NEXT: [[C:%.*]] = fmul reassoc nsz float [[X2:%.*]], 4.700000e+01 379; CHECK-NEXT: [[TMP1:%.*]] = fsub reassoc nsz float [[B]], [[C]] 380; CHECK-NEXT: ret float [[TMP1]] 381; 382 %B = fmul reassoc nsz float %X1, 47. ; X1*47 383 %C = fmul reassoc nsz float %X2, -47. ; X2*-47 384 %D = fadd reassoc nsz float %B, %C ; X1*47 + X2*-47 -> 47*(X1-X2) 385 ret float %D 386} 387 388define float @test13_reassoc(float %X1, float %X2) { 389; CHECK-LABEL: @test13_reassoc( 390; CHECK-NEXT: [[B:%.*]] = fmul reassoc float [[X1:%.*]], 4.700000e+01 391; CHECK-NEXT: [[C:%.*]] = fmul reassoc float [[X2:%.*]], 4.700000e+01 392; CHECK-NEXT: [[TMP1:%.*]] = fsub reassoc float [[B]], [[C]] 393; CHECK-NEXT: ret float [[TMP1]] 394; 395 %B = fmul reassoc float %X1, 47. ; X1*47 396 %C = fmul reassoc float %X2, -47. ; X2*-47 397 %D = fadd reassoc float %B, %C ; X1*47 + X2*-47 -> 47*(X1-X2) 398 ret float %D 399} 400 401define float @test14(float %arg) { 402; CHECK-LABEL: @test14( 403; CHECK-NEXT: [[T2:%.*]] = fmul fast float [[ARG:%.*]], 1.440000e+02 404; CHECK-NEXT: ret float [[T2]] 405; 406 %t1 = fmul fast float 1.200000e+01, %arg 407 %t2 = fmul fast float %t1, 1.200000e+01 408 ret float %t2 409} 410 411; TODO: check if we can transform the code with 'reassoc' only. 412; The same IR is transformed to one fmul in instcombine pass. 413define float @test14_reassoc(float %arg) { 414; CHECK-LABEL: @test14_reassoc( 415; CHECK-NEXT: [[T1:%.*]] = fmul reassoc float [[ARG:%.*]], 1.200000e+01 416; CHECK-NEXT: [[T2:%.*]] = fmul reassoc float [[T1]], 1.200000e+01 417; CHECK-NEXT: ret float [[T2]] 418; 419 %t1 = fmul reassoc float 1.200000e+01, %arg 420 %t2 = fmul reassoc float %t1, 1.200000e+01 421 ret float %t2 422} 423 424; (b+(a+1234))+-a -> b+1234 425; That only works with both instcombine and reassociate passes enabled. 426; Check that reassociate is not enough. 427 428; TODO: check if we can remove dead fsub. 429define float @test15(float %b, float %a) { 430; CHECK-LABEL: @test15( 431; CHECK-NEXT: [[TMP1:%.*]] = fsub fast float 0.000000e+00, [[A:%.*]] 432; CHECK-NEXT: [[TMP2:%.*]] = fadd fast float [[B:%.*]], 1.234000e+03 433; CHECK-NEXT: ret float [[TMP2]] 434; 435 %1 = fadd fast float %a, 1234.0 436 %2 = fadd fast float %b, %1 437 %3 = fsub fast float 0.0, %a 438 %4 = fadd fast float %2, %3 439 ret float %4 440} 441 442; TODO: check if we can remove dead fneg. 443define float @test15_unary_fneg(float %b, float %a) { 444; CHECK-LABEL: @test15_unary_fneg( 445; CHECK-NEXT: [[TMP1:%.*]] = fneg fast float [[A:%.*]] 446; CHECK-NEXT: [[TMP2:%.*]] = fadd fast float [[B:%.*]], 1.234000e+03 447; CHECK-NEXT: ret float [[TMP2]] 448; 449 %1 = fadd fast float %a, 1234.0 450 %2 = fadd fast float %b, %1 451 %3 = fneg fast float %a 452 %4 = fadd fast float %2, %3 453 ret float %4 454} 455 456; TODO: check if we can remove dead fsub. 457define float @test15_reassoc_nsz(float %b, float %a) { 458; CHECK-LABEL: @test15_reassoc_nsz( 459; CHECK-NEXT: [[TMP1:%.*]] = fsub reassoc nsz float 0.000000e+00, [[A:%.*]] 460; CHECK-NEXT: [[TMP2:%.*]] = fadd reassoc nsz float [[B:%.*]], 1.234000e+03 461; CHECK-NEXT: ret float [[TMP2]] 462; 463 %1 = fadd reassoc nsz float %a, 1234.0 464 %2 = fadd reassoc nsz float %b, %1 465 %3 = fsub reassoc nsz float 0.0, %a 466 %4 = fadd reassoc nsz float %2, %3 467 ret float %4 468} 469 470define float @test15_reassoc(float %b, float %a) { 471; CHECK-LABEL: @test15_reassoc( 472; CHECK-NEXT: [[TMP1:%.*]] = fadd reassoc float [[A:%.*]], 1.234000e+03 473; CHECK-NEXT: [[TMP2:%.*]] = fadd reassoc float [[B:%.*]], [[TMP1]] 474; CHECK-NEXT: [[TMP3:%.*]] = fsub reassoc float 0.000000e+00, [[A]] 475; CHECK-NEXT: [[TMP4:%.*]] = fadd reassoc float [[TMP3]], [[TMP2]] 476; CHECK-NEXT: ret float [[TMP4]] 477; 478 %1 = fadd reassoc float %a, 1234.0 479 %2 = fadd reassoc float %b, %1 480 %3 = fsub reassoc float 0.0, %a 481 %4 = fadd reassoc float %2, %3 482 ret float %4 483} 484 485; Test that we can turn things like X*-(Y*Z) -> X*-1*Y*Z. 486; That only works with both instcombine and reassociate passes enabled. 487; Check that reassociate is not enough. 488 489define float @test16(float %a, float %b, float %z) { 490; CHECK-LABEL: @test16( 491; CHECK-NEXT: [[C:%.*]] = fmul fast float [[A:%.*]], 1.234500e+04 492; CHECK-NEXT: [[E:%.*]] = fmul fast float [[C]], [[B:%.*]] 493; CHECK-NEXT: [[F:%.*]] = fmul fast float [[E]], [[Z:%.*]] 494; CHECK-NEXT: [[TMP1:%.*]] = fadd fast float [[F]], 0.000000e+00 495; CHECK-NEXT: ret float [[TMP1]] 496; 497 %c = fsub fast float 0.000000e+00, %z 498 %d = fmul fast float %a, %b 499 %e = fmul fast float %c, %d 500 %f = fmul fast float %e, 1.234500e+04 501 %g = fsub fast float 0.000000e+00, %f 502 ret float %g 503} 504 505define float @test16_unary_fneg(float %a, float %b, float %z) { 506; CHECK-LABEL: @test16_unary_fneg( 507; CHECK-NEXT: [[E:%.*]] = fmul fast float [[A:%.*]], 1.234500e+04 508; CHECK-NEXT: [[F:%.*]] = fmul fast float [[E]], [[B:%.*]] 509; CHECK-NEXT: [[G:%.*]] = fmul fast float [[F]], [[Z:%.*]] 510; CHECK-NEXT: ret float [[G]] 511; 512 %c = fneg fast float %z 513 %d = fmul fast float %a, %b 514 %e = fmul fast float %c, %d 515 %f = fmul fast float %e, 1.234500e+04 516 %g = fneg fast float %f 517 ret float %g 518} 519 520define float @test16_reassoc(float %a, float %b, float %z) { 521; CHECK-LABEL: @test16_reassoc( 522; CHECK-NEXT: [[C:%.*]] = fsub reassoc float 0.000000e+00, [[Z:%.*]] 523; CHECK-NEXT: [[D:%.*]] = fmul reassoc float [[A:%.*]], [[B:%.*]] 524; CHECK-NEXT: [[E:%.*]] = fmul reassoc float [[D]], [[C]] 525; CHECK-NEXT: [[F:%.*]] = fmul reassoc float [[E]], 1.234500e+04 526; CHECK-NEXT: [[G:%.*]] = fsub reassoc float 0.000000e+00, [[F]] 527; CHECK-NEXT: ret float [[G]] 528; 529 %c = fsub reassoc float 0.000000e+00, %z 530 %d = fmul reassoc float %a, %b 531 %e = fmul reassoc float %c, %d 532 %f = fmul reassoc float %e, 1.234500e+04 533 %g = fsub reassoc float 0.000000e+00, %f 534 ret float %g 535} 536 537; TODO: check if we can remove: 538; - fadd fast x, 0 539; ... as 'fast' implies 'nsz' 540define float @test17(float %a, float %b, float %z) { 541; CHECK-LABEL: @test17( 542; CHECK-NEXT: [[C:%.*]] = fmul fast float [[A:%.*]], 4.000000e+01 543; CHECK-NEXT: [[E:%.*]] = fmul fast float [[C]], [[Z:%.*]] 544; CHECK-NEXT: [[TMP1:%.*]] = fadd fast float [[E]], 0.000000e+00 545; CHECK-NEXT: ret float [[TMP1]] 546; 547 %d = fmul fast float %z, 4.000000e+01 548 %c = fsub fast float 0.000000e+00, %d 549 %e = fmul fast float %a, %c 550 %f = fsub fast float 0.000000e+00, %e 551 ret float %f 552} 553 554define float @test17_unary_fneg(float %a, float %b, float %z) { 555; CHECK-LABEL: @test17_unary_fneg( 556; CHECK-NEXT: [[E:%.*]] = fmul fast float [[A:%.*]], 4.000000e+01 557; CHECK-NEXT: [[F:%.*]] = fmul fast float [[E]], [[Z:%.*]] 558; CHECK-NEXT: ret float [[F]] 559; 560 %d = fmul fast float %z, 4.000000e+01 561 %c = fneg fast float %d 562 %e = fmul fast float %a, %c 563 %f = fneg fast float %e 564 ret float %f 565} 566 567; With sub reassociation, constant folding can eliminate the 12 and -12 constants. 568; TODO: check if we can remove fadd fast x, 0 as 'fast' implies 'nsz' 569define float @test18(float %A, float %B) { 570; CHECK-LABEL: @test18( 571; CHECK-NEXT: [[B_NEG:%.*]] = fneg fast float [[B:%.*]] 572; CHECK-NEXT: [[Y:%.*]] = fadd fast float [[A:%.*]], 0.000000e+00 573; CHECK-NEXT: [[Z:%.*]] = fadd fast float [[Y]], [[B_NEG]] 574; CHECK-NEXT: ret float [[Z]] 575; 576 %X = fadd fast float -1.200000e+01, %A 577 %Y = fsub fast float %X, %B 578 %Z = fadd fast float %Y, 1.200000e+01 579 ret float %Z 580} 581