1 //===- InstCombineSelect.cpp ----------------------------------------------===// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 // 10 // This file implements the visitSelect function. 11 // 12 //===----------------------------------------------------------------------===// 13 14 #include "InstCombineInternal.h" 15 #include "llvm/Analysis/ConstantFolding.h" 16 #include "llvm/Analysis/InstructionSimplify.h" 17 #include "llvm/Analysis/ValueTracking.h" 18 #include "llvm/IR/PatternMatch.h" 19 using namespace llvm; 20 using namespace PatternMatch; 21 22 #define DEBUG_TYPE "instcombine" 23 24 static SelectPatternFlavor 25 getInverseMinMaxSelectPattern(SelectPatternFlavor SPF) { 26 switch (SPF) { 27 default: 28 llvm_unreachable("unhandled!"); 29 30 case SPF_SMIN: 31 return SPF_SMAX; 32 case SPF_UMIN: 33 return SPF_UMAX; 34 case SPF_SMAX: 35 return SPF_SMIN; 36 case SPF_UMAX: 37 return SPF_UMIN; 38 } 39 } 40 41 static CmpInst::Predicate getCmpPredicateForMinMax(SelectPatternFlavor SPF, 42 bool Ordered=false) { 43 switch (SPF) { 44 default: 45 llvm_unreachable("unhandled!"); 46 47 case SPF_SMIN: 48 return ICmpInst::ICMP_SLT; 49 case SPF_UMIN: 50 return ICmpInst::ICMP_ULT; 51 case SPF_SMAX: 52 return ICmpInst::ICMP_SGT; 53 case SPF_UMAX: 54 return ICmpInst::ICMP_UGT; 55 case SPF_FMINNUM: 56 return Ordered ? FCmpInst::FCMP_OLT : FCmpInst::FCMP_ULT; 57 case SPF_FMAXNUM: 58 return Ordered ? FCmpInst::FCMP_OGT : FCmpInst::FCMP_UGT; 59 } 60 } 61 62 static Value *generateMinMaxSelectPattern(InstCombiner::BuilderTy *Builder, 63 SelectPatternFlavor SPF, Value *A, 64 Value *B) { 65 CmpInst::Predicate Pred = getCmpPredicateForMinMax(SPF); 66 assert(CmpInst::isIntPredicate(Pred)); 67 return Builder->CreateSelect(Builder->CreateICmp(Pred, A, B), A, B); 68 } 69 70 /// We want to turn code that looks like this: 71 /// %C = or %A, %B 72 /// %D = select %cond, %C, %A 73 /// into: 74 /// %C = select %cond, %B, 0 75 /// %D = or %A, %C 76 /// 77 /// Assuming that the specified instruction is an operand to the select, return 78 /// a bitmask indicating which operands of this instruction are foldable if they 79 /// equal the other incoming value of the select. 80 /// 81 static unsigned GetSelectFoldableOperands(Instruction *I) { 82 switch (I->getOpcode()) { 83 case Instruction::Add: 84 case Instruction::Mul: 85 case Instruction::And: 86 case Instruction::Or: 87 case Instruction::Xor: 88 return 3; // Can fold through either operand. 89 case Instruction::Sub: // Can only fold on the amount subtracted. 90 case Instruction::Shl: // Can only fold on the shift amount. 91 case Instruction::LShr: 92 case Instruction::AShr: 93 return 1; 94 default: 95 return 0; // Cannot fold 96 } 97 } 98 99 /// For the same transformation as the previous function, return the identity 100 /// constant that goes into the select. 101 static Constant *GetSelectFoldableConstant(Instruction *I) { 102 switch (I->getOpcode()) { 103 default: llvm_unreachable("This cannot happen!"); 104 case Instruction::Add: 105 case Instruction::Sub: 106 case Instruction::Or: 107 case Instruction::Xor: 108 case Instruction::Shl: 109 case Instruction::LShr: 110 case Instruction::AShr: 111 return Constant::getNullValue(I->getType()); 112 case Instruction::And: 113 return Constant::getAllOnesValue(I->getType()); 114 case Instruction::Mul: 115 return ConstantInt::get(I->getType(), 1); 116 } 117 } 118 119 /// Here we have (select c, TI, FI), and we know that TI and FI 120 /// have the same opcode and only one use each. Try to simplify this. 121 Instruction *InstCombiner::FoldSelectOpOp(SelectInst &SI, Instruction *TI, 122 Instruction *FI) { 123 if (TI->getNumOperands() == 1) { 124 // If this is a non-volatile load or a cast from the same type, 125 // merge. 126 if (TI->isCast()) { 127 Type *FIOpndTy = FI->getOperand(0)->getType(); 128 if (TI->getOperand(0)->getType() != FIOpndTy) 129 return nullptr; 130 // The select condition may be a vector. We may only change the operand 131 // type if the vector width remains the same (and matches the condition). 132 Type *CondTy = SI.getCondition()->getType(); 133 if (CondTy->isVectorTy() && (!FIOpndTy->isVectorTy() || 134 CondTy->getVectorNumElements() != FIOpndTy->getVectorNumElements())) 135 return nullptr; 136 } else { 137 return nullptr; // unknown unary op. 138 } 139 140 // Fold this by inserting a select from the input values. 141 Value *NewSI = Builder->CreateSelect(SI.getCondition(), TI->getOperand(0), 142 FI->getOperand(0), SI.getName()+".v"); 143 return CastInst::Create(Instruction::CastOps(TI->getOpcode()), NewSI, 144 TI->getType()); 145 } 146 147 // Only handle binary operators here. 148 if (!isa<BinaryOperator>(TI)) 149 return nullptr; 150 151 // Figure out if the operations have any operands in common. 152 Value *MatchOp, *OtherOpT, *OtherOpF; 153 bool MatchIsOpZero; 154 if (TI->getOperand(0) == FI->getOperand(0)) { 155 MatchOp = TI->getOperand(0); 156 OtherOpT = TI->getOperand(1); 157 OtherOpF = FI->getOperand(1); 158 MatchIsOpZero = true; 159 } else if (TI->getOperand(1) == FI->getOperand(1)) { 160 MatchOp = TI->getOperand(1); 161 OtherOpT = TI->getOperand(0); 162 OtherOpF = FI->getOperand(0); 163 MatchIsOpZero = false; 164 } else if (!TI->isCommutative()) { 165 return nullptr; 166 } else if (TI->getOperand(0) == FI->getOperand(1)) { 167 MatchOp = TI->getOperand(0); 168 OtherOpT = TI->getOperand(1); 169 OtherOpF = FI->getOperand(0); 170 MatchIsOpZero = true; 171 } else if (TI->getOperand(1) == FI->getOperand(0)) { 172 MatchOp = TI->getOperand(1); 173 OtherOpT = TI->getOperand(0); 174 OtherOpF = FI->getOperand(1); 175 MatchIsOpZero = true; 176 } else { 177 return nullptr; 178 } 179 180 // If we reach here, they do have operations in common. 181 Value *NewSI = Builder->CreateSelect(SI.getCondition(), OtherOpT, 182 OtherOpF, SI.getName()+".v"); 183 184 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TI)) { 185 if (MatchIsOpZero) 186 return BinaryOperator::Create(BO->getOpcode(), MatchOp, NewSI); 187 else 188 return BinaryOperator::Create(BO->getOpcode(), NewSI, MatchOp); 189 } 190 llvm_unreachable("Shouldn't get here"); 191 } 192 193 static bool isSelect01(Constant *C1, Constant *C2) { 194 ConstantInt *C1I = dyn_cast<ConstantInt>(C1); 195 if (!C1I) 196 return false; 197 ConstantInt *C2I = dyn_cast<ConstantInt>(C2); 198 if (!C2I) 199 return false; 200 if (!C1I->isZero() && !C2I->isZero()) // One side must be zero. 201 return false; 202 return C1I->isOne() || C1I->isAllOnesValue() || 203 C2I->isOne() || C2I->isAllOnesValue(); 204 } 205 206 /// Try to fold the select into one of the operands to allow further 207 /// optimization. 208 Instruction *InstCombiner::FoldSelectIntoOp(SelectInst &SI, Value *TrueVal, 209 Value *FalseVal) { 210 // See the comment above GetSelectFoldableOperands for a description of the 211 // transformation we are doing here. 212 if (Instruction *TVI = dyn_cast<Instruction>(TrueVal)) { 213 if (TVI->hasOneUse() && TVI->getNumOperands() == 2 && 214 !isa<Constant>(FalseVal)) { 215 if (unsigned SFO = GetSelectFoldableOperands(TVI)) { 216 unsigned OpToFold = 0; 217 if ((SFO & 1) && FalseVal == TVI->getOperand(0)) { 218 OpToFold = 1; 219 } else if ((SFO & 2) && FalseVal == TVI->getOperand(1)) { 220 OpToFold = 2; 221 } 222 223 if (OpToFold) { 224 Constant *C = GetSelectFoldableConstant(TVI); 225 Value *OOp = TVI->getOperand(2-OpToFold); 226 // Avoid creating select between 2 constants unless it's selecting 227 // between 0, 1 and -1. 228 if (!isa<Constant>(OOp) || isSelect01(C, cast<Constant>(OOp))) { 229 Value *NewSel = Builder->CreateSelect(SI.getCondition(), OOp, C); 230 NewSel->takeName(TVI); 231 BinaryOperator *TVI_BO = cast<BinaryOperator>(TVI); 232 BinaryOperator *BO = BinaryOperator::Create(TVI_BO->getOpcode(), 233 FalseVal, NewSel); 234 if (isa<PossiblyExactOperator>(BO)) 235 BO->setIsExact(TVI_BO->isExact()); 236 if (isa<OverflowingBinaryOperator>(BO)) { 237 BO->setHasNoUnsignedWrap(TVI_BO->hasNoUnsignedWrap()); 238 BO->setHasNoSignedWrap(TVI_BO->hasNoSignedWrap()); 239 } 240 return BO; 241 } 242 } 243 } 244 } 245 } 246 247 if (Instruction *FVI = dyn_cast<Instruction>(FalseVal)) { 248 if (FVI->hasOneUse() && FVI->getNumOperands() == 2 && 249 !isa<Constant>(TrueVal)) { 250 if (unsigned SFO = GetSelectFoldableOperands(FVI)) { 251 unsigned OpToFold = 0; 252 if ((SFO & 1) && TrueVal == FVI->getOperand(0)) { 253 OpToFold = 1; 254 } else if ((SFO & 2) && TrueVal == FVI->getOperand(1)) { 255 OpToFold = 2; 256 } 257 258 if (OpToFold) { 259 Constant *C = GetSelectFoldableConstant(FVI); 260 Value *OOp = FVI->getOperand(2-OpToFold); 261 // Avoid creating select between 2 constants unless it's selecting 262 // between 0, 1 and -1. 263 if (!isa<Constant>(OOp) || isSelect01(C, cast<Constant>(OOp))) { 264 Value *NewSel = Builder->CreateSelect(SI.getCondition(), C, OOp); 265 NewSel->takeName(FVI); 266 BinaryOperator *FVI_BO = cast<BinaryOperator>(FVI); 267 BinaryOperator *BO = BinaryOperator::Create(FVI_BO->getOpcode(), 268 TrueVal, NewSel); 269 if (isa<PossiblyExactOperator>(BO)) 270 BO->setIsExact(FVI_BO->isExact()); 271 if (isa<OverflowingBinaryOperator>(BO)) { 272 BO->setHasNoUnsignedWrap(FVI_BO->hasNoUnsignedWrap()); 273 BO->setHasNoSignedWrap(FVI_BO->hasNoSignedWrap()); 274 } 275 return BO; 276 } 277 } 278 } 279 } 280 } 281 282 return nullptr; 283 } 284 285 /// We want to turn: 286 /// (select (icmp eq (and X, C1), 0), Y, (or Y, C2)) 287 /// into: 288 /// (or (shl (and X, C1), C3), y) 289 /// iff: 290 /// C1 and C2 are both powers of 2 291 /// where: 292 /// C3 = Log(C2) - Log(C1) 293 /// 294 /// This transform handles cases where: 295 /// 1. The icmp predicate is inverted 296 /// 2. The select operands are reversed 297 /// 3. The magnitude of C2 and C1 are flipped 298 static Value *foldSelectICmpAndOr(const SelectInst &SI, Value *TrueVal, 299 Value *FalseVal, 300 InstCombiner::BuilderTy *Builder) { 301 const ICmpInst *IC = dyn_cast<ICmpInst>(SI.getCondition()); 302 if (!IC || !IC->isEquality() || !SI.getType()->isIntegerTy()) 303 return nullptr; 304 305 Value *CmpLHS = IC->getOperand(0); 306 Value *CmpRHS = IC->getOperand(1); 307 308 if (!match(CmpRHS, m_Zero())) 309 return nullptr; 310 311 Value *X; 312 const APInt *C1; 313 if (!match(CmpLHS, m_And(m_Value(X), m_Power2(C1)))) 314 return nullptr; 315 316 const APInt *C2; 317 bool OrOnTrueVal = false; 318 bool OrOnFalseVal = match(FalseVal, m_Or(m_Specific(TrueVal), m_Power2(C2))); 319 if (!OrOnFalseVal) 320 OrOnTrueVal = match(TrueVal, m_Or(m_Specific(FalseVal), m_Power2(C2))); 321 322 if (!OrOnFalseVal && !OrOnTrueVal) 323 return nullptr; 324 325 Value *V = CmpLHS; 326 Value *Y = OrOnFalseVal ? TrueVal : FalseVal; 327 328 unsigned C1Log = C1->logBase2(); 329 unsigned C2Log = C2->logBase2(); 330 if (C2Log > C1Log) { 331 V = Builder->CreateZExtOrTrunc(V, Y->getType()); 332 V = Builder->CreateShl(V, C2Log - C1Log); 333 } else if (C1Log > C2Log) { 334 V = Builder->CreateLShr(V, C1Log - C2Log); 335 V = Builder->CreateZExtOrTrunc(V, Y->getType()); 336 } else 337 V = Builder->CreateZExtOrTrunc(V, Y->getType()); 338 339 ICmpInst::Predicate Pred = IC->getPredicate(); 340 if ((Pred == ICmpInst::ICMP_NE && OrOnFalseVal) || 341 (Pred == ICmpInst::ICMP_EQ && OrOnTrueVal)) 342 V = Builder->CreateXor(V, *C2); 343 344 return Builder->CreateOr(V, Y); 345 } 346 347 /// Attempt to fold a cttz/ctlz followed by a icmp plus select into a single 348 /// call to cttz/ctlz with flag 'is_zero_undef' cleared. 349 /// 350 /// For example, we can fold the following code sequence: 351 /// \code 352 /// %0 = tail call i32 @llvm.cttz.i32(i32 %x, i1 true) 353 /// %1 = icmp ne i32 %x, 0 354 /// %2 = select i1 %1, i32 %0, i32 32 355 /// \code 356 /// 357 /// into: 358 /// %0 = tail call i32 @llvm.cttz.i32(i32 %x, i1 false) 359 static Value *foldSelectCttzCtlz(ICmpInst *ICI, Value *TrueVal, Value *FalseVal, 360 InstCombiner::BuilderTy *Builder) { 361 ICmpInst::Predicate Pred = ICI->getPredicate(); 362 Value *CmpLHS = ICI->getOperand(0); 363 Value *CmpRHS = ICI->getOperand(1); 364 365 // Check if the condition value compares a value for equality against zero. 366 if (!ICI->isEquality() || !match(CmpRHS, m_Zero())) 367 return nullptr; 368 369 Value *Count = FalseVal; 370 Value *ValueOnZero = TrueVal; 371 if (Pred == ICmpInst::ICMP_NE) 372 std::swap(Count, ValueOnZero); 373 374 // Skip zero extend/truncate. 375 Value *V = nullptr; 376 if (match(Count, m_ZExt(m_Value(V))) || 377 match(Count, m_Trunc(m_Value(V)))) 378 Count = V; 379 380 // Check if the value propagated on zero is a constant number equal to the 381 // sizeof in bits of 'Count'. 382 unsigned SizeOfInBits = Count->getType()->getScalarSizeInBits(); 383 if (!match(ValueOnZero, m_SpecificInt(SizeOfInBits))) 384 return nullptr; 385 386 // Check that 'Count' is a call to intrinsic cttz/ctlz. Also check that the 387 // input to the cttz/ctlz is used as LHS for the compare instruction. 388 if (match(Count, m_Intrinsic<Intrinsic::cttz>(m_Specific(CmpLHS))) || 389 match(Count, m_Intrinsic<Intrinsic::ctlz>(m_Specific(CmpLHS)))) { 390 IntrinsicInst *II = cast<IntrinsicInst>(Count); 391 IRBuilder<> Builder(II); 392 // Explicitly clear the 'undef_on_zero' flag. 393 IntrinsicInst *NewI = cast<IntrinsicInst>(II->clone()); 394 Type *Ty = NewI->getArgOperand(1)->getType(); 395 NewI->setArgOperand(1, Constant::getNullValue(Ty)); 396 Builder.Insert(NewI); 397 return Builder.CreateZExtOrTrunc(NewI, ValueOnZero->getType()); 398 } 399 400 return nullptr; 401 } 402 403 /// Visit a SelectInst that has an ICmpInst as its first operand. 404 Instruction *InstCombiner::visitSelectInstWithICmp(SelectInst &SI, 405 ICmpInst *ICI) { 406 bool Changed = false; 407 ICmpInst::Predicate Pred = ICI->getPredicate(); 408 Value *CmpLHS = ICI->getOperand(0); 409 Value *CmpRHS = ICI->getOperand(1); 410 Value *TrueVal = SI.getTrueValue(); 411 Value *FalseVal = SI.getFalseValue(); 412 413 // Check cases where the comparison is with a constant that 414 // can be adjusted to fit the min/max idiom. We may move or edit ICI 415 // here, so make sure the select is the only user. 416 if (ICI->hasOneUse()) 417 if (ConstantInt *CI = dyn_cast<ConstantInt>(CmpRHS)) { 418 switch (Pred) { 419 default: break; 420 case ICmpInst::ICMP_ULT: 421 case ICmpInst::ICMP_SLT: 422 case ICmpInst::ICMP_UGT: 423 case ICmpInst::ICMP_SGT: { 424 // These transformations only work for selects over integers. 425 IntegerType *SelectTy = dyn_cast<IntegerType>(SI.getType()); 426 if (!SelectTy) 427 break; 428 429 Constant *AdjustedRHS; 430 if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_SGT) 431 AdjustedRHS = ConstantInt::get(CI->getContext(), CI->getValue() + 1); 432 else // (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_SLT) 433 AdjustedRHS = ConstantInt::get(CI->getContext(), CI->getValue() - 1); 434 435 // X > C ? X : C+1 --> X < C+1 ? C+1 : X 436 // X < C ? X : C-1 --> X > C-1 ? C-1 : X 437 if ((CmpLHS == TrueVal && AdjustedRHS == FalseVal) || 438 (CmpLHS == FalseVal && AdjustedRHS == TrueVal)) 439 ; // Nothing to do here. Values match without any sign/zero extension. 440 441 // Types do not match. Instead of calculating this with mixed types 442 // promote all to the larger type. This enables scalar evolution to 443 // analyze this expression. 444 else if (CmpRHS->getType()->getScalarSizeInBits() 445 < SelectTy->getBitWidth()) { 446 Constant *sextRHS = ConstantExpr::getSExt(AdjustedRHS, SelectTy); 447 448 // X = sext x; x >s c ? X : C+1 --> X = sext x; X <s C+1 ? C+1 : X 449 // X = sext x; x <s c ? X : C-1 --> X = sext x; X >s C-1 ? C-1 : X 450 // X = sext x; x >u c ? X : C+1 --> X = sext x; X <u C+1 ? C+1 : X 451 // X = sext x; x <u c ? X : C-1 --> X = sext x; X >u C-1 ? C-1 : X 452 if (match(TrueVal, m_SExt(m_Specific(CmpLHS))) && 453 sextRHS == FalseVal) { 454 CmpLHS = TrueVal; 455 AdjustedRHS = sextRHS; 456 } else if (match(FalseVal, m_SExt(m_Specific(CmpLHS))) && 457 sextRHS == TrueVal) { 458 CmpLHS = FalseVal; 459 AdjustedRHS = sextRHS; 460 } else if (ICI->isUnsigned()) { 461 Constant *zextRHS = ConstantExpr::getZExt(AdjustedRHS, SelectTy); 462 // X = zext x; x >u c ? X : C+1 --> X = zext x; X <u C+1 ? C+1 : X 463 // X = zext x; x <u c ? X : C-1 --> X = zext x; X >u C-1 ? C-1 : X 464 // zext + signed compare cannot be changed: 465 // 0xff <s 0x00, but 0x00ff >s 0x0000 466 if (match(TrueVal, m_ZExt(m_Specific(CmpLHS))) && 467 zextRHS == FalseVal) { 468 CmpLHS = TrueVal; 469 AdjustedRHS = zextRHS; 470 } else if (match(FalseVal, m_ZExt(m_Specific(CmpLHS))) && 471 zextRHS == TrueVal) { 472 CmpLHS = FalseVal; 473 AdjustedRHS = zextRHS; 474 } else 475 break; 476 } else 477 break; 478 } else 479 break; 480 481 Pred = ICmpInst::getSwappedPredicate(Pred); 482 CmpRHS = AdjustedRHS; 483 std::swap(FalseVal, TrueVal); 484 ICI->setPredicate(Pred); 485 ICI->setOperand(0, CmpLHS); 486 ICI->setOperand(1, CmpRHS); 487 SI.setOperand(1, TrueVal); 488 SI.setOperand(2, FalseVal); 489 490 // Move ICI instruction right before the select instruction. Otherwise 491 // the sext/zext value may be defined after the ICI instruction uses it. 492 ICI->moveBefore(&SI); 493 494 Changed = true; 495 break; 496 } 497 } 498 } 499 500 // Transform (X >s -1) ? C1 : C2 --> ((X >>s 31) & (C2 - C1)) + C1 501 // and (X <s 0) ? C2 : C1 --> ((X >>s 31) & (C2 - C1)) + C1 502 // FIXME: Type and constness constraints could be lifted, but we have to 503 // watch code size carefully. We should consider xor instead of 504 // sub/add when we decide to do that. 505 if (IntegerType *Ty = dyn_cast<IntegerType>(CmpLHS->getType())) { 506 if (TrueVal->getType() == Ty) { 507 if (ConstantInt *Cmp = dyn_cast<ConstantInt>(CmpRHS)) { 508 ConstantInt *C1 = nullptr, *C2 = nullptr; 509 if (Pred == ICmpInst::ICMP_SGT && Cmp->isAllOnesValue()) { 510 C1 = dyn_cast<ConstantInt>(TrueVal); 511 C2 = dyn_cast<ConstantInt>(FalseVal); 512 } else if (Pred == ICmpInst::ICMP_SLT && Cmp->isNullValue()) { 513 C1 = dyn_cast<ConstantInt>(FalseVal); 514 C2 = dyn_cast<ConstantInt>(TrueVal); 515 } 516 if (C1 && C2) { 517 // This shift results in either -1 or 0. 518 Value *AShr = Builder->CreateAShr(CmpLHS, Ty->getBitWidth()-1); 519 520 // Check if we can express the operation with a single or. 521 if (C2->isAllOnesValue()) 522 return ReplaceInstUsesWith(SI, Builder->CreateOr(AShr, C1)); 523 524 Value *And = Builder->CreateAnd(AShr, C2->getValue()-C1->getValue()); 525 return ReplaceInstUsesWith(SI, Builder->CreateAdd(And, C1)); 526 } 527 } 528 } 529 } 530 531 // NOTE: if we wanted to, this is where to detect integer MIN/MAX 532 533 if (CmpRHS != CmpLHS && isa<Constant>(CmpRHS)) { 534 if (CmpLHS == TrueVal && Pred == ICmpInst::ICMP_EQ) { 535 // Transform (X == C) ? X : Y -> (X == C) ? C : Y 536 SI.setOperand(1, CmpRHS); 537 Changed = true; 538 } else if (CmpLHS == FalseVal && Pred == ICmpInst::ICMP_NE) { 539 // Transform (X != C) ? Y : X -> (X != C) ? Y : C 540 SI.setOperand(2, CmpRHS); 541 Changed = true; 542 } 543 } 544 545 { 546 unsigned BitWidth = DL.getTypeSizeInBits(TrueVal->getType()); 547 APInt MinSignedValue = APInt::getSignBit(BitWidth); 548 Value *X; 549 const APInt *Y, *C; 550 bool TrueWhenUnset; 551 bool IsBitTest = false; 552 if (ICmpInst::isEquality(Pred) && 553 match(CmpLHS, m_And(m_Value(X), m_Power2(Y))) && 554 match(CmpRHS, m_Zero())) { 555 IsBitTest = true; 556 TrueWhenUnset = Pred == ICmpInst::ICMP_EQ; 557 } else if (Pred == ICmpInst::ICMP_SLT && match(CmpRHS, m_Zero())) { 558 X = CmpLHS; 559 Y = &MinSignedValue; 560 IsBitTest = true; 561 TrueWhenUnset = false; 562 } else if (Pred == ICmpInst::ICMP_SGT && match(CmpRHS, m_AllOnes())) { 563 X = CmpLHS; 564 Y = &MinSignedValue; 565 IsBitTest = true; 566 TrueWhenUnset = true; 567 } 568 if (IsBitTest) { 569 Value *V = nullptr; 570 // (X & Y) == 0 ? X : X ^ Y --> X & ~Y 571 if (TrueWhenUnset && TrueVal == X && 572 match(FalseVal, m_Xor(m_Specific(X), m_APInt(C))) && *Y == *C) 573 V = Builder->CreateAnd(X, ~(*Y)); 574 // (X & Y) != 0 ? X ^ Y : X --> X & ~Y 575 else if (!TrueWhenUnset && FalseVal == X && 576 match(TrueVal, m_Xor(m_Specific(X), m_APInt(C))) && *Y == *C) 577 V = Builder->CreateAnd(X, ~(*Y)); 578 // (X & Y) == 0 ? X ^ Y : X --> X | Y 579 else if (TrueWhenUnset && FalseVal == X && 580 match(TrueVal, m_Xor(m_Specific(X), m_APInt(C))) && *Y == *C) 581 V = Builder->CreateOr(X, *Y); 582 // (X & Y) != 0 ? X : X ^ Y --> X | Y 583 else if (!TrueWhenUnset && TrueVal == X && 584 match(FalseVal, m_Xor(m_Specific(X), m_APInt(C))) && *Y == *C) 585 V = Builder->CreateOr(X, *Y); 586 587 if (V) 588 return ReplaceInstUsesWith(SI, V); 589 } 590 } 591 592 if (Value *V = foldSelectICmpAndOr(SI, TrueVal, FalseVal, Builder)) 593 return ReplaceInstUsesWith(SI, V); 594 595 if (Value *V = foldSelectCttzCtlz(ICI, TrueVal, FalseVal, Builder)) 596 return ReplaceInstUsesWith(SI, V); 597 598 return Changed ? &SI : nullptr; 599 } 600 601 602 /// SI is a select whose condition is a PHI node (but the two may be in 603 /// different blocks). See if the true/false values (V) are live in all of the 604 /// predecessor blocks of the PHI. For example, cases like this can't be mapped: 605 /// 606 /// X = phi [ C1, BB1], [C2, BB2] 607 /// Y = add 608 /// Z = select X, Y, 0 609 /// 610 /// because Y is not live in BB1/BB2. 611 /// 612 static bool CanSelectOperandBeMappingIntoPredBlock(const Value *V, 613 const SelectInst &SI) { 614 // If the value is a non-instruction value like a constant or argument, it 615 // can always be mapped. 616 const Instruction *I = dyn_cast<Instruction>(V); 617 if (!I) return true; 618 619 // If V is a PHI node defined in the same block as the condition PHI, we can 620 // map the arguments. 621 const PHINode *CondPHI = cast<PHINode>(SI.getCondition()); 622 623 if (const PHINode *VP = dyn_cast<PHINode>(I)) 624 if (VP->getParent() == CondPHI->getParent()) 625 return true; 626 627 // Otherwise, if the PHI and select are defined in the same block and if V is 628 // defined in a different block, then we can transform it. 629 if (SI.getParent() == CondPHI->getParent() && 630 I->getParent() != CondPHI->getParent()) 631 return true; 632 633 // Otherwise we have a 'hard' case and we can't tell without doing more 634 // detailed dominator based analysis, punt. 635 return false; 636 } 637 638 /// We have an SPF (e.g. a min or max) of an SPF of the form: 639 /// SPF2(SPF1(A, B), C) 640 Instruction *InstCombiner::FoldSPFofSPF(Instruction *Inner, 641 SelectPatternFlavor SPF1, 642 Value *A, Value *B, 643 Instruction &Outer, 644 SelectPatternFlavor SPF2, Value *C) { 645 if (C == A || C == B) { 646 // MAX(MAX(A, B), B) -> MAX(A, B) 647 // MIN(MIN(a, b), a) -> MIN(a, b) 648 if (SPF1 == SPF2) 649 return ReplaceInstUsesWith(Outer, Inner); 650 651 // MAX(MIN(a, b), a) -> a 652 // MIN(MAX(a, b), a) -> a 653 if ((SPF1 == SPF_SMIN && SPF2 == SPF_SMAX) || 654 (SPF1 == SPF_SMAX && SPF2 == SPF_SMIN) || 655 (SPF1 == SPF_UMIN && SPF2 == SPF_UMAX) || 656 (SPF1 == SPF_UMAX && SPF2 == SPF_UMIN)) 657 return ReplaceInstUsesWith(Outer, C); 658 } 659 660 if (SPF1 == SPF2) { 661 if (ConstantInt *CB = dyn_cast<ConstantInt>(B)) { 662 if (ConstantInt *CC = dyn_cast<ConstantInt>(C)) { 663 APInt ACB = CB->getValue(); 664 APInt ACC = CC->getValue(); 665 666 // MIN(MIN(A, 23), 97) -> MIN(A, 23) 667 // MAX(MAX(A, 97), 23) -> MAX(A, 97) 668 if ((SPF1 == SPF_UMIN && ACB.ule(ACC)) || 669 (SPF1 == SPF_SMIN && ACB.sle(ACC)) || 670 (SPF1 == SPF_UMAX && ACB.uge(ACC)) || 671 (SPF1 == SPF_SMAX && ACB.sge(ACC))) 672 return ReplaceInstUsesWith(Outer, Inner); 673 674 // MIN(MIN(A, 97), 23) -> MIN(A, 23) 675 // MAX(MAX(A, 23), 97) -> MAX(A, 97) 676 if ((SPF1 == SPF_UMIN && ACB.ugt(ACC)) || 677 (SPF1 == SPF_SMIN && ACB.sgt(ACC)) || 678 (SPF1 == SPF_UMAX && ACB.ult(ACC)) || 679 (SPF1 == SPF_SMAX && ACB.slt(ACC))) { 680 Outer.replaceUsesOfWith(Inner, A); 681 return &Outer; 682 } 683 } 684 } 685 } 686 687 // ABS(ABS(X)) -> ABS(X) 688 // NABS(NABS(X)) -> NABS(X) 689 if (SPF1 == SPF2 && (SPF1 == SPF_ABS || SPF1 == SPF_NABS)) { 690 return ReplaceInstUsesWith(Outer, Inner); 691 } 692 693 // ABS(NABS(X)) -> ABS(X) 694 // NABS(ABS(X)) -> NABS(X) 695 if ((SPF1 == SPF_ABS && SPF2 == SPF_NABS) || 696 (SPF1 == SPF_NABS && SPF2 == SPF_ABS)) { 697 SelectInst *SI = cast<SelectInst>(Inner); 698 Value *NewSI = Builder->CreateSelect( 699 SI->getCondition(), SI->getFalseValue(), SI->getTrueValue()); 700 return ReplaceInstUsesWith(Outer, NewSI); 701 } 702 703 auto IsFreeOrProfitableToInvert = 704 [&](Value *V, Value *&NotV, bool &ElidesXor) { 705 if (match(V, m_Not(m_Value(NotV)))) { 706 // If V has at most 2 uses then we can get rid of the xor operation 707 // entirely. 708 ElidesXor |= !V->hasNUsesOrMore(3); 709 return true; 710 } 711 712 if (IsFreeToInvert(V, !V->hasNUsesOrMore(3))) { 713 NotV = nullptr; 714 return true; 715 } 716 717 return false; 718 }; 719 720 Value *NotA, *NotB, *NotC; 721 bool ElidesXor = false; 722 723 // MIN(MIN(~A, ~B), ~C) == ~MAX(MAX(A, B), C) 724 // MIN(MAX(~A, ~B), ~C) == ~MAX(MIN(A, B), C) 725 // MAX(MIN(~A, ~B), ~C) == ~MIN(MAX(A, B), C) 726 // MAX(MAX(~A, ~B), ~C) == ~MIN(MIN(A, B), C) 727 // 728 // This transform is performance neutral if we can elide at least one xor from 729 // the set of three operands, since we'll be tacking on an xor at the very 730 // end. 731 if (IsFreeOrProfitableToInvert(A, NotA, ElidesXor) && 732 IsFreeOrProfitableToInvert(B, NotB, ElidesXor) && 733 IsFreeOrProfitableToInvert(C, NotC, ElidesXor) && ElidesXor) { 734 if (!NotA) 735 NotA = Builder->CreateNot(A); 736 if (!NotB) 737 NotB = Builder->CreateNot(B); 738 if (!NotC) 739 NotC = Builder->CreateNot(C); 740 741 Value *NewInner = generateMinMaxSelectPattern( 742 Builder, getInverseMinMaxSelectPattern(SPF1), NotA, NotB); 743 Value *NewOuter = Builder->CreateNot(generateMinMaxSelectPattern( 744 Builder, getInverseMinMaxSelectPattern(SPF2), NewInner, NotC)); 745 return ReplaceInstUsesWith(Outer, NewOuter); 746 } 747 748 return nullptr; 749 } 750 751 /// If one of the constants is zero (we know they can't both be) and we have an 752 /// icmp instruction with zero, and we have an 'and' with the non-constant value 753 /// and a power of two we can turn the select into a shift on the result of the 754 /// 'and'. 755 static Value *foldSelectICmpAnd(const SelectInst &SI, ConstantInt *TrueVal, 756 ConstantInt *FalseVal, 757 InstCombiner::BuilderTy *Builder) { 758 const ICmpInst *IC = dyn_cast<ICmpInst>(SI.getCondition()); 759 if (!IC || !IC->isEquality() || !SI.getType()->isIntegerTy()) 760 return nullptr; 761 762 if (!match(IC->getOperand(1), m_Zero())) 763 return nullptr; 764 765 ConstantInt *AndRHS; 766 Value *LHS = IC->getOperand(0); 767 if (!match(LHS, m_And(m_Value(), m_ConstantInt(AndRHS)))) 768 return nullptr; 769 770 // If both select arms are non-zero see if we have a select of the form 771 // 'x ? 2^n + C : C'. Then we can offset both arms by C, use the logic 772 // for 'x ? 2^n : 0' and fix the thing up at the end. 773 ConstantInt *Offset = nullptr; 774 if (!TrueVal->isZero() && !FalseVal->isZero()) { 775 if ((TrueVal->getValue() - FalseVal->getValue()).isPowerOf2()) 776 Offset = FalseVal; 777 else if ((FalseVal->getValue() - TrueVal->getValue()).isPowerOf2()) 778 Offset = TrueVal; 779 else 780 return nullptr; 781 782 // Adjust TrueVal and FalseVal to the offset. 783 TrueVal = ConstantInt::get(Builder->getContext(), 784 TrueVal->getValue() - Offset->getValue()); 785 FalseVal = ConstantInt::get(Builder->getContext(), 786 FalseVal->getValue() - Offset->getValue()); 787 } 788 789 // Make sure the mask in the 'and' and one of the select arms is a power of 2. 790 if (!AndRHS->getValue().isPowerOf2() || 791 (!TrueVal->getValue().isPowerOf2() && 792 !FalseVal->getValue().isPowerOf2())) 793 return nullptr; 794 795 // Determine which shift is needed to transform result of the 'and' into the 796 // desired result. 797 ConstantInt *ValC = !TrueVal->isZero() ? TrueVal : FalseVal; 798 unsigned ValZeros = ValC->getValue().logBase2(); 799 unsigned AndZeros = AndRHS->getValue().logBase2(); 800 801 // If types don't match we can still convert the select by introducing a zext 802 // or a trunc of the 'and'. The trunc case requires that all of the truncated 803 // bits are zero, we can figure that out by looking at the 'and' mask. 804 if (AndZeros >= ValC->getBitWidth()) 805 return nullptr; 806 807 Value *V = Builder->CreateZExtOrTrunc(LHS, SI.getType()); 808 if (ValZeros > AndZeros) 809 V = Builder->CreateShl(V, ValZeros - AndZeros); 810 else if (ValZeros < AndZeros) 811 V = Builder->CreateLShr(V, AndZeros - ValZeros); 812 813 // Okay, now we know that everything is set up, we just don't know whether we 814 // have a icmp_ne or icmp_eq and whether the true or false val is the zero. 815 bool ShouldNotVal = !TrueVal->isZero(); 816 ShouldNotVal ^= IC->getPredicate() == ICmpInst::ICMP_NE; 817 if (ShouldNotVal) 818 V = Builder->CreateXor(V, ValC); 819 820 // Apply an offset if needed. 821 if (Offset) 822 V = Builder->CreateAdd(V, Offset); 823 return V; 824 } 825 826 Instruction *InstCombiner::visitSelectInst(SelectInst &SI) { 827 Value *CondVal = SI.getCondition(); 828 Value *TrueVal = SI.getTrueValue(); 829 Value *FalseVal = SI.getFalseValue(); 830 831 if (Value *V = 832 SimplifySelectInst(CondVal, TrueVal, FalseVal, DL, TLI, DT, AC)) 833 return ReplaceInstUsesWith(SI, V); 834 835 if (SI.getType()->isIntegerTy(1)) { 836 if (ConstantInt *C = dyn_cast<ConstantInt>(TrueVal)) { 837 if (C->getZExtValue()) { 838 // Change: A = select B, true, C --> A = or B, C 839 return BinaryOperator::CreateOr(CondVal, FalseVal); 840 } 841 // Change: A = select B, false, C --> A = and !B, C 842 Value *NotCond = Builder->CreateNot(CondVal, "not."+CondVal->getName()); 843 return BinaryOperator::CreateAnd(NotCond, FalseVal); 844 } 845 if (ConstantInt *C = dyn_cast<ConstantInt>(FalseVal)) { 846 if (!C->getZExtValue()) { 847 // Change: A = select B, C, false --> A = and B, C 848 return BinaryOperator::CreateAnd(CondVal, TrueVal); 849 } 850 // Change: A = select B, C, true --> A = or !B, C 851 Value *NotCond = Builder->CreateNot(CondVal, "not."+CondVal->getName()); 852 return BinaryOperator::CreateOr(NotCond, TrueVal); 853 } 854 855 // select a, b, a -> a&b 856 // select a, a, b -> a|b 857 if (CondVal == TrueVal) 858 return BinaryOperator::CreateOr(CondVal, FalseVal); 859 if (CondVal == FalseVal) 860 return BinaryOperator::CreateAnd(CondVal, TrueVal); 861 862 // select a, ~a, b -> (~a)&b 863 // select a, b, ~a -> (~a)|b 864 if (match(TrueVal, m_Not(m_Specific(CondVal)))) 865 return BinaryOperator::CreateAnd(TrueVal, FalseVal); 866 if (match(FalseVal, m_Not(m_Specific(CondVal)))) 867 return BinaryOperator::CreateOr(TrueVal, FalseVal); 868 } 869 870 // Selecting between two integer constants? 871 if (ConstantInt *TrueValC = dyn_cast<ConstantInt>(TrueVal)) 872 if (ConstantInt *FalseValC = dyn_cast<ConstantInt>(FalseVal)) { 873 // select C, 1, 0 -> zext C to int 874 if (FalseValC->isZero() && TrueValC->getValue() == 1) 875 return new ZExtInst(CondVal, SI.getType()); 876 877 // select C, -1, 0 -> sext C to int 878 if (FalseValC->isZero() && TrueValC->isAllOnesValue()) 879 return new SExtInst(CondVal, SI.getType()); 880 881 // select C, 0, 1 -> zext !C to int 882 if (TrueValC->isZero() && FalseValC->getValue() == 1) { 883 Value *NotCond = Builder->CreateNot(CondVal, "not."+CondVal->getName()); 884 return new ZExtInst(NotCond, SI.getType()); 885 } 886 887 // select C, 0, -1 -> sext !C to int 888 if (TrueValC->isZero() && FalseValC->isAllOnesValue()) { 889 Value *NotCond = Builder->CreateNot(CondVal, "not."+CondVal->getName()); 890 return new SExtInst(NotCond, SI.getType()); 891 } 892 893 if (Value *V = foldSelectICmpAnd(SI, TrueValC, FalseValC, Builder)) 894 return ReplaceInstUsesWith(SI, V); 895 } 896 897 // See if we are selecting two values based on a comparison of the two values. 898 if (FCmpInst *FCI = dyn_cast<FCmpInst>(CondVal)) { 899 if (FCI->getOperand(0) == TrueVal && FCI->getOperand(1) == FalseVal) { 900 // Transform (X == Y) ? X : Y -> Y 901 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ) { 902 // This is not safe in general for floating point: 903 // consider X== -0, Y== +0. 904 // It becomes safe if either operand is a nonzero constant. 905 ConstantFP *CFPt, *CFPf; 906 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) && 907 !CFPt->getValueAPF().isZero()) || 908 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) && 909 !CFPf->getValueAPF().isZero())) 910 return ReplaceInstUsesWith(SI, FalseVal); 911 } 912 // Transform (X une Y) ? X : Y -> X 913 if (FCI->getPredicate() == FCmpInst::FCMP_UNE) { 914 // This is not safe in general for floating point: 915 // consider X== -0, Y== +0. 916 // It becomes safe if either operand is a nonzero constant. 917 ConstantFP *CFPt, *CFPf; 918 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) && 919 !CFPt->getValueAPF().isZero()) || 920 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) && 921 !CFPf->getValueAPF().isZero())) 922 return ReplaceInstUsesWith(SI, TrueVal); 923 } 924 925 // Canonicalize to use ordered comparisons by swapping the select 926 // operands. 927 // 928 // e.g. 929 // (X ugt Y) ? X : Y -> (X ole Y) ? Y : X 930 if (FCI->hasOneUse() && FCmpInst::isUnordered(FCI->getPredicate())) { 931 FCmpInst::Predicate InvPred = FCI->getInversePredicate(); 932 IRBuilder<>::FastMathFlagGuard FMFG(*Builder); 933 Builder->SetFastMathFlags(FCI->getFastMathFlags()); 934 Value *NewCond = Builder->CreateFCmp(InvPred, TrueVal, FalseVal, 935 FCI->getName() + ".inv"); 936 937 return SelectInst::Create(NewCond, FalseVal, TrueVal, 938 SI.getName() + ".p"); 939 } 940 941 // NOTE: if we wanted to, this is where to detect MIN/MAX 942 } else if (FCI->getOperand(0) == FalseVal && FCI->getOperand(1) == TrueVal){ 943 // Transform (X == Y) ? Y : X -> X 944 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ) { 945 // This is not safe in general for floating point: 946 // consider X== -0, Y== +0. 947 // It becomes safe if either operand is a nonzero constant. 948 ConstantFP *CFPt, *CFPf; 949 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) && 950 !CFPt->getValueAPF().isZero()) || 951 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) && 952 !CFPf->getValueAPF().isZero())) 953 return ReplaceInstUsesWith(SI, FalseVal); 954 } 955 // Transform (X une Y) ? Y : X -> Y 956 if (FCI->getPredicate() == FCmpInst::FCMP_UNE) { 957 // This is not safe in general for floating point: 958 // consider X== -0, Y== +0. 959 // It becomes safe if either operand is a nonzero constant. 960 ConstantFP *CFPt, *CFPf; 961 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) && 962 !CFPt->getValueAPF().isZero()) || 963 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) && 964 !CFPf->getValueAPF().isZero())) 965 return ReplaceInstUsesWith(SI, TrueVal); 966 } 967 968 // Canonicalize to use ordered comparisons by swapping the select 969 // operands. 970 // 971 // e.g. 972 // (X ugt Y) ? X : Y -> (X ole Y) ? X : Y 973 if (FCI->hasOneUse() && FCmpInst::isUnordered(FCI->getPredicate())) { 974 FCmpInst::Predicate InvPred = FCI->getInversePredicate(); 975 IRBuilder<>::FastMathFlagGuard FMFG(*Builder); 976 Builder->SetFastMathFlags(FCI->getFastMathFlags()); 977 Value *NewCond = Builder->CreateFCmp(InvPred, FalseVal, TrueVal, 978 FCI->getName() + ".inv"); 979 980 return SelectInst::Create(NewCond, FalseVal, TrueVal, 981 SI.getName() + ".p"); 982 } 983 984 // NOTE: if we wanted to, this is where to detect MIN/MAX 985 } 986 // NOTE: if we wanted to, this is where to detect ABS 987 } 988 989 // See if we are selecting two values based on a comparison of the two values. 990 if (ICmpInst *ICI = dyn_cast<ICmpInst>(CondVal)) 991 if (Instruction *Result = visitSelectInstWithICmp(SI, ICI)) 992 return Result; 993 994 if (Instruction *TI = dyn_cast<Instruction>(TrueVal)) 995 if (Instruction *FI = dyn_cast<Instruction>(FalseVal)) 996 if (TI->hasOneUse() && FI->hasOneUse()) { 997 Instruction *AddOp = nullptr, *SubOp = nullptr; 998 999 // Turn (select C, (op X, Y), (op X, Z)) -> (op X, (select C, Y, Z)) 1000 if (TI->getOpcode() == FI->getOpcode()) 1001 if (Instruction *IV = FoldSelectOpOp(SI, TI, FI)) 1002 return IV; 1003 1004 // Turn select C, (X+Y), (X-Y) --> (X+(select C, Y, (-Y))). This is 1005 // even legal for FP. 1006 if ((TI->getOpcode() == Instruction::Sub && 1007 FI->getOpcode() == Instruction::Add) || 1008 (TI->getOpcode() == Instruction::FSub && 1009 FI->getOpcode() == Instruction::FAdd)) { 1010 AddOp = FI; SubOp = TI; 1011 } else if ((FI->getOpcode() == Instruction::Sub && 1012 TI->getOpcode() == Instruction::Add) || 1013 (FI->getOpcode() == Instruction::FSub && 1014 TI->getOpcode() == Instruction::FAdd)) { 1015 AddOp = TI; SubOp = FI; 1016 } 1017 1018 if (AddOp) { 1019 Value *OtherAddOp = nullptr; 1020 if (SubOp->getOperand(0) == AddOp->getOperand(0)) { 1021 OtherAddOp = AddOp->getOperand(1); 1022 } else if (SubOp->getOperand(0) == AddOp->getOperand(1)) { 1023 OtherAddOp = AddOp->getOperand(0); 1024 } 1025 1026 if (OtherAddOp) { 1027 // So at this point we know we have (Y -> OtherAddOp): 1028 // select C, (add X, Y), (sub X, Z) 1029 Value *NegVal; // Compute -Z 1030 if (SI.getType()->isFPOrFPVectorTy()) { 1031 NegVal = Builder->CreateFNeg(SubOp->getOperand(1)); 1032 if (Instruction *NegInst = dyn_cast<Instruction>(NegVal)) { 1033 FastMathFlags Flags = AddOp->getFastMathFlags(); 1034 Flags &= SubOp->getFastMathFlags(); 1035 NegInst->setFastMathFlags(Flags); 1036 } 1037 } else { 1038 NegVal = Builder->CreateNeg(SubOp->getOperand(1)); 1039 } 1040 1041 Value *NewTrueOp = OtherAddOp; 1042 Value *NewFalseOp = NegVal; 1043 if (AddOp != TI) 1044 std::swap(NewTrueOp, NewFalseOp); 1045 Value *NewSel = 1046 Builder->CreateSelect(CondVal, NewTrueOp, 1047 NewFalseOp, SI.getName() + ".p"); 1048 1049 if (SI.getType()->isFPOrFPVectorTy()) { 1050 Instruction *RI = 1051 BinaryOperator::CreateFAdd(SubOp->getOperand(0), NewSel); 1052 1053 FastMathFlags Flags = AddOp->getFastMathFlags(); 1054 Flags &= SubOp->getFastMathFlags(); 1055 RI->setFastMathFlags(Flags); 1056 return RI; 1057 } else 1058 return BinaryOperator::CreateAdd(SubOp->getOperand(0), NewSel); 1059 } 1060 } 1061 } 1062 1063 // See if we can fold the select into one of our operands. 1064 if (SI.getType()->isIntOrIntVectorTy() || SI.getType()->isFPOrFPVectorTy()) { 1065 if (Instruction *FoldI = FoldSelectIntoOp(SI, TrueVal, FalseVal)) 1066 return FoldI; 1067 1068 Value *LHS, *RHS, *LHS2, *RHS2; 1069 Instruction::CastOps CastOp; 1070 SelectPatternResult SPR = matchSelectPattern(&SI, LHS, RHS, &CastOp); 1071 auto SPF = SPR.Flavor; 1072 1073 if (SelectPatternResult::isMinOrMax(SPF)) { 1074 // Canonicalize so that type casts are outside select patterns. 1075 if (LHS->getType()->getPrimitiveSizeInBits() != 1076 SI.getType()->getPrimitiveSizeInBits()) { 1077 CmpInst::Predicate Pred = getCmpPredicateForMinMax(SPF, SPR.Ordered); 1078 1079 Value *Cmp; 1080 if (CmpInst::isIntPredicate(Pred)) { 1081 Cmp = Builder->CreateICmp(Pred, LHS, RHS); 1082 } else { 1083 IRBuilder<>::FastMathFlagGuard FMFG(*Builder); 1084 auto FMF = cast<FPMathOperator>(SI.getCondition())->getFastMathFlags(); 1085 Builder->SetFastMathFlags(FMF); 1086 Cmp = Builder->CreateFCmp(Pred, LHS, RHS); 1087 } 1088 1089 Value *NewSI = Builder->CreateCast(CastOp, 1090 Builder->CreateSelect(Cmp, LHS, RHS), 1091 SI.getType()); 1092 return ReplaceInstUsesWith(SI, NewSI); 1093 } 1094 } 1095 1096 if (SPF) { 1097 // MAX(MAX(a, b), a) -> MAX(a, b) 1098 // MIN(MIN(a, b), a) -> MIN(a, b) 1099 // MAX(MIN(a, b), a) -> a 1100 // MIN(MAX(a, b), a) -> a 1101 // ABS(ABS(a)) -> ABS(a) 1102 // NABS(NABS(a)) -> NABS(a) 1103 if (SelectPatternFlavor SPF2 = matchSelectPattern(LHS, LHS2, RHS2).Flavor) 1104 if (Instruction *R = FoldSPFofSPF(cast<Instruction>(LHS),SPF2,LHS2,RHS2, 1105 SI, SPF, RHS)) 1106 return R; 1107 if (SelectPatternFlavor SPF2 = matchSelectPattern(RHS, LHS2, RHS2).Flavor) 1108 if (Instruction *R = FoldSPFofSPF(cast<Instruction>(RHS),SPF2,LHS2,RHS2, 1109 SI, SPF, LHS)) 1110 return R; 1111 } 1112 1113 // MAX(~a, ~b) -> ~MIN(a, b) 1114 if (SPF == SPF_SMAX || SPF == SPF_UMAX) { 1115 if (IsFreeToInvert(LHS, LHS->hasNUses(2)) && 1116 IsFreeToInvert(RHS, RHS->hasNUses(2))) { 1117 1118 // This transform adds a xor operation and that extra cost needs to be 1119 // justified. We look for simplifications that will result from 1120 // applying this rule: 1121 1122 bool Profitable = 1123 (LHS->hasNUses(2) && match(LHS, m_Not(m_Value()))) || 1124 (RHS->hasNUses(2) && match(RHS, m_Not(m_Value()))) || 1125 (SI.hasOneUse() && match(*SI.user_begin(), m_Not(m_Value()))); 1126 1127 if (Profitable) { 1128 Value *NewLHS = Builder->CreateNot(LHS); 1129 Value *NewRHS = Builder->CreateNot(RHS); 1130 Value *NewCmp = SPF == SPF_SMAX 1131 ? Builder->CreateICmpSLT(NewLHS, NewRHS) 1132 : Builder->CreateICmpULT(NewLHS, NewRHS); 1133 Value *NewSI = 1134 Builder->CreateNot(Builder->CreateSelect(NewCmp, NewLHS, NewRHS)); 1135 return ReplaceInstUsesWith(SI, NewSI); 1136 } 1137 } 1138 } 1139 1140 // TODO. 1141 // ABS(-X) -> ABS(X) 1142 } 1143 1144 // See if we can fold the select into a phi node if the condition is a select. 1145 if (isa<PHINode>(SI.getCondition())) 1146 // The true/false values have to be live in the PHI predecessor's blocks. 1147 if (CanSelectOperandBeMappingIntoPredBlock(TrueVal, SI) && 1148 CanSelectOperandBeMappingIntoPredBlock(FalseVal, SI)) 1149 if (Instruction *NV = FoldOpIntoPhi(SI)) 1150 return NV; 1151 1152 if (SelectInst *TrueSI = dyn_cast<SelectInst>(TrueVal)) { 1153 if (TrueSI->getCondition()->getType() == CondVal->getType()) { 1154 // select(C, select(C, a, b), c) -> select(C, a, c) 1155 if (TrueSI->getCondition() == CondVal) { 1156 if (SI.getTrueValue() == TrueSI->getTrueValue()) 1157 return nullptr; 1158 SI.setOperand(1, TrueSI->getTrueValue()); 1159 return &SI; 1160 } 1161 // select(C0, select(C1, a, b), b) -> select(C0&C1, a, b) 1162 // We choose this as normal form to enable folding on the And and shortening 1163 // paths for the values (this helps GetUnderlyingObjects() for example). 1164 if (TrueSI->getFalseValue() == FalseVal && TrueSI->hasOneUse()) { 1165 Value *And = Builder->CreateAnd(CondVal, TrueSI->getCondition()); 1166 SI.setOperand(0, And); 1167 SI.setOperand(1, TrueSI->getTrueValue()); 1168 return &SI; 1169 } 1170 } 1171 } 1172 if (SelectInst *FalseSI = dyn_cast<SelectInst>(FalseVal)) { 1173 if (FalseSI->getCondition()->getType() == CondVal->getType()) { 1174 // select(C, a, select(C, b, c)) -> select(C, a, c) 1175 if (FalseSI->getCondition() == CondVal) { 1176 if (SI.getFalseValue() == FalseSI->getFalseValue()) 1177 return nullptr; 1178 SI.setOperand(2, FalseSI->getFalseValue()); 1179 return &SI; 1180 } 1181 // select(C0, a, select(C1, a, b)) -> select(C0|C1, a, b) 1182 if (FalseSI->getTrueValue() == TrueVal && FalseSI->hasOneUse()) { 1183 Value *Or = Builder->CreateOr(CondVal, FalseSI->getCondition()); 1184 SI.setOperand(0, Or); 1185 SI.setOperand(2, FalseSI->getFalseValue()); 1186 return &SI; 1187 } 1188 } 1189 } 1190 1191 if (BinaryOperator::isNot(CondVal)) { 1192 SI.setOperand(0, BinaryOperator::getNotArgument(CondVal)); 1193 SI.setOperand(1, FalseVal); 1194 SI.setOperand(2, TrueVal); 1195 return &SI; 1196 } 1197 1198 if (VectorType* VecTy = dyn_cast<VectorType>(SI.getType())) { 1199 unsigned VWidth = VecTy->getNumElements(); 1200 APInt UndefElts(VWidth, 0); 1201 APInt AllOnesEltMask(APInt::getAllOnesValue(VWidth)); 1202 if (Value *V = SimplifyDemandedVectorElts(&SI, AllOnesEltMask, UndefElts)) { 1203 if (V != &SI) 1204 return ReplaceInstUsesWith(SI, V); 1205 return &SI; 1206 } 1207 1208 if (isa<ConstantAggregateZero>(CondVal)) { 1209 return ReplaceInstUsesWith(SI, FalseVal); 1210 } 1211 } 1212 1213 return nullptr; 1214 } 1215