1 //===- InstCombineInternal.h - InstCombine pass internals -------*- C++ -*-===// 2 // 3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4 // See https://llvm.org/LICENSE.txt for license information. 5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6 // 7 //===----------------------------------------------------------------------===// 8 // 9 /// \file 10 /// 11 /// This file provides internal interfaces used to implement the InstCombine. 12 // 13 //===----------------------------------------------------------------------===// 14 15 #ifndef LLVM_LIB_TRANSFORMS_INSTCOMBINE_INSTCOMBINEINTERNAL_H 16 #define LLVM_LIB_TRANSFORMS_INSTCOMBINE_INSTCOMBINEINTERNAL_H 17 18 #include "llvm/ADT/ArrayRef.h" 19 #include "llvm/Analysis/AliasAnalysis.h" 20 #include "llvm/Analysis/InstructionSimplify.h" 21 #include "llvm/Analysis/TargetFolder.h" 22 #include "llvm/Analysis/ValueTracking.h" 23 #include "llvm/IR/Argument.h" 24 #include "llvm/IR/BasicBlock.h" 25 #include "llvm/IR/Constant.h" 26 #include "llvm/IR/Constants.h" 27 #include "llvm/IR/DerivedTypes.h" 28 #include "llvm/IR/IRBuilder.h" 29 #include "llvm/IR/InstVisitor.h" 30 #include "llvm/IR/InstrTypes.h" 31 #include "llvm/IR/Instruction.h" 32 #include "llvm/IR/IntrinsicInst.h" 33 #include "llvm/IR/Intrinsics.h" 34 #include "llvm/IR/PatternMatch.h" 35 #include "llvm/IR/Use.h" 36 #include "llvm/IR/Value.h" 37 #include "llvm/Support/Casting.h" 38 #include "llvm/Support/Compiler.h" 39 #include "llvm/Support/Debug.h" 40 #include "llvm/Support/KnownBits.h" 41 #include "llvm/Support/raw_ostream.h" 42 #include "llvm/Transforms/InstCombine/InstCombineWorklist.h" 43 #include "llvm/Transforms/Utils/Local.h" 44 #include <cassert> 45 #include <cstdint> 46 47 #define DEBUG_TYPE "instcombine" 48 49 using namespace llvm::PatternMatch; 50 51 namespace llvm { 52 53 class APInt; 54 class AssumptionCache; 55 class BlockFrequencyInfo; 56 class DataLayout; 57 class DominatorTree; 58 class GEPOperator; 59 class GlobalVariable; 60 class LoopInfo; 61 class OptimizationRemarkEmitter; 62 class ProfileSummaryInfo; 63 class TargetLibraryInfo; 64 class User; 65 66 /// Assign a complexity or rank value to LLVM Values. This is used to reduce 67 /// the amount of pattern matching needed for compares and commutative 68 /// instructions. For example, if we have: 69 /// icmp ugt X, Constant 70 /// or 71 /// xor (add X, Constant), cast Z 72 /// 73 /// We do not have to consider the commuted variants of these patterns because 74 /// canonicalization based on complexity guarantees the above ordering. 75 /// 76 /// This routine maps IR values to various complexity ranks: 77 /// 0 -> undef 78 /// 1 -> Constants 79 /// 2 -> Other non-instructions 80 /// 3 -> Arguments 81 /// 4 -> Cast and (f)neg/not instructions 82 /// 5 -> Other instructions 83 static inline unsigned getComplexity(Value *V) { 84 if (isa<Instruction>(V)) { 85 if (isa<CastInst>(V) || match(V, m_Neg(m_Value())) || 86 match(V, m_Not(m_Value())) || match(V, m_FNeg(m_Value()))) 87 return 4; 88 return 5; 89 } 90 if (isa<Argument>(V)) 91 return 3; 92 return isa<Constant>(V) ? (isa<UndefValue>(V) ? 0 : 1) : 2; 93 } 94 95 /// Predicate canonicalization reduces the number of patterns that need to be 96 /// matched by other transforms. For example, we may swap the operands of a 97 /// conditional branch or select to create a compare with a canonical (inverted) 98 /// predicate which is then more likely to be matched with other values. 99 static inline bool isCanonicalPredicate(CmpInst::Predicate Pred) { 100 switch (Pred) { 101 case CmpInst::ICMP_NE: 102 case CmpInst::ICMP_ULE: 103 case CmpInst::ICMP_SLE: 104 case CmpInst::ICMP_UGE: 105 case CmpInst::ICMP_SGE: 106 // TODO: There are 16 FCMP predicates. Should others be (not) canonical? 107 case CmpInst::FCMP_ONE: 108 case CmpInst::FCMP_OLE: 109 case CmpInst::FCMP_OGE: 110 return false; 111 default: 112 return true; 113 } 114 } 115 116 /// Return the source operand of a potentially bitcasted value while optionally 117 /// checking if it has one use. If there is no bitcast or the one use check is 118 /// not met, return the input value itself. 119 static inline Value *peekThroughBitcast(Value *V, bool OneUseOnly = false) { 120 if (auto *BitCast = dyn_cast<BitCastInst>(V)) 121 if (!OneUseOnly || BitCast->hasOneUse()) 122 return BitCast->getOperand(0); 123 124 // V is not a bitcast or V has more than one use and OneUseOnly is true. 125 return V; 126 } 127 128 /// Add one to a Constant 129 static inline Constant *AddOne(Constant *C) { 130 return ConstantExpr::getAdd(C, ConstantInt::get(C->getType(), 1)); 131 } 132 133 /// Subtract one from a Constant 134 static inline Constant *SubOne(Constant *C) { 135 return ConstantExpr::getSub(C, ConstantInt::get(C->getType(), 1)); 136 } 137 138 /// Return true if the specified value is free to invert (apply ~ to). 139 /// This happens in cases where the ~ can be eliminated. If WillInvertAllUses 140 /// is true, work under the assumption that the caller intends to remove all 141 /// uses of V and only keep uses of ~V. 142 static inline bool IsFreeToInvert(Value *V, bool WillInvertAllUses) { 143 // ~(~(X)) -> X. 144 if (match(V, m_Not(m_Value()))) 145 return true; 146 147 // Constants can be considered to be not'ed values. 148 if (isa<ConstantInt>(V)) 149 return true; 150 151 // A vector of constant integers can be inverted easily. 152 if (V->getType()->isVectorTy() && isa<Constant>(V)) { 153 unsigned NumElts = V->getType()->getVectorNumElements(); 154 for (unsigned i = 0; i != NumElts; ++i) { 155 Constant *Elt = cast<Constant>(V)->getAggregateElement(i); 156 if (!Elt) 157 return false; 158 159 if (isa<UndefValue>(Elt)) 160 continue; 161 162 if (!isa<ConstantInt>(Elt)) 163 return false; 164 } 165 return true; 166 } 167 168 // Compares can be inverted if all of their uses are being modified to use the 169 // ~V. 170 if (isa<CmpInst>(V)) 171 return WillInvertAllUses; 172 173 // If `V` is of the form `A + Constant` then `-1 - V` can be folded into `(-1 174 // - Constant) - A` if we are willing to invert all of the uses. 175 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(V)) 176 if (BO->getOpcode() == Instruction::Add || 177 BO->getOpcode() == Instruction::Sub) 178 if (isa<Constant>(BO->getOperand(0)) || isa<Constant>(BO->getOperand(1))) 179 return WillInvertAllUses; 180 181 // Selects with invertible operands are freely invertible 182 if (match(V, m_Select(m_Value(), m_Not(m_Value()), m_Not(m_Value())))) 183 return WillInvertAllUses; 184 185 return false; 186 } 187 188 /// Some binary operators require special handling to avoid poison and undefined 189 /// behavior. If a constant vector has undef elements, replace those undefs with 190 /// identity constants if possible because those are always safe to execute. 191 /// If no identity constant exists, replace undef with some other safe constant. 192 static inline Constant *getSafeVectorConstantForBinop( 193 BinaryOperator::BinaryOps Opcode, Constant *In, bool IsRHSConstant) { 194 assert(In->getType()->isVectorTy() && "Not expecting scalars here"); 195 196 Type *EltTy = In->getType()->getVectorElementType(); 197 auto *SafeC = ConstantExpr::getBinOpIdentity(Opcode, EltTy, IsRHSConstant); 198 if (!SafeC) { 199 // TODO: Should this be available as a constant utility function? It is 200 // similar to getBinOpAbsorber(). 201 if (IsRHSConstant) { 202 switch (Opcode) { 203 case Instruction::SRem: // X % 1 = 0 204 case Instruction::URem: // X %u 1 = 0 205 SafeC = ConstantInt::get(EltTy, 1); 206 break; 207 case Instruction::FRem: // X % 1.0 (doesn't simplify, but it is safe) 208 SafeC = ConstantFP::get(EltTy, 1.0); 209 break; 210 default: 211 llvm_unreachable("Only rem opcodes have no identity constant for RHS"); 212 } 213 } else { 214 switch (Opcode) { 215 case Instruction::Shl: // 0 << X = 0 216 case Instruction::LShr: // 0 >>u X = 0 217 case Instruction::AShr: // 0 >> X = 0 218 case Instruction::SDiv: // 0 / X = 0 219 case Instruction::UDiv: // 0 /u X = 0 220 case Instruction::SRem: // 0 % X = 0 221 case Instruction::URem: // 0 %u X = 0 222 case Instruction::Sub: // 0 - X (doesn't simplify, but it is safe) 223 case Instruction::FSub: // 0.0 - X (doesn't simplify, but it is safe) 224 case Instruction::FDiv: // 0.0 / X (doesn't simplify, but it is safe) 225 case Instruction::FRem: // 0.0 % X = 0 226 SafeC = Constant::getNullValue(EltTy); 227 break; 228 default: 229 llvm_unreachable("Expected to find identity constant for opcode"); 230 } 231 } 232 } 233 assert(SafeC && "Must have safe constant for binop"); 234 unsigned NumElts = In->getType()->getVectorNumElements(); 235 SmallVector<Constant *, 16> Out(NumElts); 236 for (unsigned i = 0; i != NumElts; ++i) { 237 Constant *C = In->getAggregateElement(i); 238 Out[i] = isa<UndefValue>(C) ? SafeC : C; 239 } 240 return ConstantVector::get(Out); 241 } 242 243 /// The core instruction combiner logic. 244 /// 245 /// This class provides both the logic to recursively visit instructions and 246 /// combine them. 247 class LLVM_LIBRARY_VISIBILITY InstCombiner 248 : public InstVisitor<InstCombiner, Instruction *> { 249 // FIXME: These members shouldn't be public. 250 public: 251 /// A worklist of the instructions that need to be simplified. 252 InstCombineWorklist &Worklist; 253 254 /// An IRBuilder that automatically inserts new instructions into the 255 /// worklist. 256 using BuilderTy = IRBuilder<TargetFolder, IRBuilderCallbackInserter>; 257 BuilderTy &Builder; 258 259 private: 260 // Mode in which we are running the combiner. 261 const bool MinimizeSize; 262 263 /// Enable combines that trigger rarely but are costly in compiletime. 264 const bool ExpensiveCombines; 265 266 AliasAnalysis *AA; 267 268 // Required analyses. 269 AssumptionCache &AC; 270 TargetLibraryInfo &TLI; 271 DominatorTree &DT; 272 const DataLayout &DL; 273 const SimplifyQuery SQ; 274 OptimizationRemarkEmitter &ORE; 275 BlockFrequencyInfo *BFI; 276 ProfileSummaryInfo *PSI; 277 278 // Optional analyses. When non-null, these can both be used to do better 279 // combining and will be updated to reflect any changes. 280 LoopInfo *LI; 281 282 bool MadeIRChange = false; 283 284 public: 285 InstCombiner(InstCombineWorklist &Worklist, BuilderTy &Builder, 286 bool MinimizeSize, bool ExpensiveCombines, AliasAnalysis *AA, 287 AssumptionCache &AC, TargetLibraryInfo &TLI, DominatorTree &DT, 288 OptimizationRemarkEmitter &ORE, BlockFrequencyInfo *BFI, 289 ProfileSummaryInfo *PSI, const DataLayout &DL, LoopInfo *LI) 290 : Worklist(Worklist), Builder(Builder), MinimizeSize(MinimizeSize), 291 ExpensiveCombines(ExpensiveCombines), AA(AA), AC(AC), TLI(TLI), DT(DT), 292 DL(DL), SQ(DL, &TLI, &DT, &AC), ORE(ORE), BFI(BFI), PSI(PSI), LI(LI) {} 293 294 /// Run the combiner over the entire worklist until it is empty. 295 /// 296 /// \returns true if the IR is changed. 297 bool run(); 298 299 AssumptionCache &getAssumptionCache() const { return AC; } 300 301 const DataLayout &getDataLayout() const { return DL; } 302 303 DominatorTree &getDominatorTree() const { return DT; } 304 305 LoopInfo *getLoopInfo() const { return LI; } 306 307 TargetLibraryInfo &getTargetLibraryInfo() const { return TLI; } 308 309 // Visitation implementation - Implement instruction combining for different 310 // instruction types. The semantics are as follows: 311 // Return Value: 312 // null - No change was made 313 // I - Change was made, I is still valid, I may be dead though 314 // otherwise - Change was made, replace I with returned instruction 315 // 316 Instruction *visitFNeg(UnaryOperator &I); 317 Instruction *visitAdd(BinaryOperator &I); 318 Instruction *visitFAdd(BinaryOperator &I); 319 Value *OptimizePointerDifference(Value *LHS, Value *RHS, Type *Ty); 320 Instruction *visitSub(BinaryOperator &I); 321 Instruction *visitFSub(BinaryOperator &I); 322 Instruction *visitMul(BinaryOperator &I); 323 Instruction *visitFMul(BinaryOperator &I); 324 Instruction *visitURem(BinaryOperator &I); 325 Instruction *visitSRem(BinaryOperator &I); 326 Instruction *visitFRem(BinaryOperator &I); 327 bool simplifyDivRemOfSelectWithZeroOp(BinaryOperator &I); 328 Instruction *commonRemTransforms(BinaryOperator &I); 329 Instruction *commonIRemTransforms(BinaryOperator &I); 330 Instruction *commonDivTransforms(BinaryOperator &I); 331 Instruction *commonIDivTransforms(BinaryOperator &I); 332 Instruction *visitUDiv(BinaryOperator &I); 333 Instruction *visitSDiv(BinaryOperator &I); 334 Instruction *visitFDiv(BinaryOperator &I); 335 Value *simplifyRangeCheck(ICmpInst *Cmp0, ICmpInst *Cmp1, bool Inverted); 336 Instruction *visitAnd(BinaryOperator &I); 337 Instruction *visitOr(BinaryOperator &I); 338 Instruction *visitXor(BinaryOperator &I); 339 Instruction *visitShl(BinaryOperator &I); 340 Instruction *visitAShr(BinaryOperator &I); 341 Instruction *visitLShr(BinaryOperator &I); 342 Instruction *commonShiftTransforms(BinaryOperator &I); 343 Instruction *visitFCmpInst(FCmpInst &I); 344 Instruction *visitICmpInst(ICmpInst &I); 345 Instruction *FoldShiftByConstant(Value *Op0, Constant *Op1, 346 BinaryOperator &I); 347 Instruction *commonCastTransforms(CastInst &CI); 348 Instruction *commonPointerCastTransforms(CastInst &CI); 349 Instruction *visitTrunc(TruncInst &CI); 350 Instruction *visitZExt(ZExtInst &CI); 351 Instruction *visitSExt(SExtInst &CI); 352 Instruction *visitFPTrunc(FPTruncInst &CI); 353 Instruction *visitFPExt(CastInst &CI); 354 Instruction *visitFPToUI(FPToUIInst &FI); 355 Instruction *visitFPToSI(FPToSIInst &FI); 356 Instruction *visitUIToFP(CastInst &CI); 357 Instruction *visitSIToFP(CastInst &CI); 358 Instruction *visitPtrToInt(PtrToIntInst &CI); 359 Instruction *visitIntToPtr(IntToPtrInst &CI); 360 Instruction *visitBitCast(BitCastInst &CI); 361 Instruction *visitAddrSpaceCast(AddrSpaceCastInst &CI); 362 Instruction *FoldItoFPtoI(Instruction &FI); 363 Instruction *visitSelectInst(SelectInst &SI); 364 Instruction *visitCallInst(CallInst &CI); 365 Instruction *visitInvokeInst(InvokeInst &II); 366 Instruction *visitCallBrInst(CallBrInst &CBI); 367 368 Instruction *SliceUpIllegalIntegerPHI(PHINode &PN); 369 Instruction *visitPHINode(PHINode &PN); 370 Instruction *visitGetElementPtrInst(GetElementPtrInst &GEP); 371 Instruction *visitAllocaInst(AllocaInst &AI); 372 Instruction *visitAllocSite(Instruction &FI); 373 Instruction *visitFree(CallInst &FI); 374 Instruction *visitLoadInst(LoadInst &LI); 375 Instruction *visitStoreInst(StoreInst &SI); 376 Instruction *visitAtomicRMWInst(AtomicRMWInst &SI); 377 Instruction *visitBranchInst(BranchInst &BI); 378 Instruction *visitFenceInst(FenceInst &FI); 379 Instruction *visitSwitchInst(SwitchInst &SI); 380 Instruction *visitReturnInst(ReturnInst &RI); 381 Instruction *visitInsertValueInst(InsertValueInst &IV); 382 Instruction *visitInsertElementInst(InsertElementInst &IE); 383 Instruction *visitExtractElementInst(ExtractElementInst &EI); 384 Instruction *visitShuffleVectorInst(ShuffleVectorInst &SVI); 385 Instruction *visitExtractValueInst(ExtractValueInst &EV); 386 Instruction *visitLandingPadInst(LandingPadInst &LI); 387 Instruction *visitVAStartInst(VAStartInst &I); 388 Instruction *visitVACopyInst(VACopyInst &I); 389 390 /// Specify what to return for unhandled instructions. 391 Instruction *visitInstruction(Instruction &I) { return nullptr; } 392 393 /// True when DB dominates all uses of DI except UI. 394 /// UI must be in the same block as DI. 395 /// The routine checks that the DI parent and DB are different. 396 bool dominatesAllUses(const Instruction *DI, const Instruction *UI, 397 const BasicBlock *DB) const; 398 399 /// Try to replace select with select operand SIOpd in SI-ICmp sequence. 400 bool replacedSelectWithOperand(SelectInst *SI, const ICmpInst *Icmp, 401 const unsigned SIOpd); 402 403 /// Try to replace instruction \p I with value \p V which are pointers 404 /// in different address space. 405 /// \return true if successful. 406 bool replacePointer(Instruction &I, Value *V); 407 408 private: 409 bool shouldChangeType(unsigned FromBitWidth, unsigned ToBitWidth) const; 410 bool shouldChangeType(Type *From, Type *To) const; 411 Value *dyn_castNegVal(Value *V) const; 412 Type *FindElementAtOffset(PointerType *PtrTy, int64_t Offset, 413 SmallVectorImpl<Value *> &NewIndices); 414 415 /// Classify whether a cast is worth optimizing. 416 /// 417 /// This is a helper to decide whether the simplification of 418 /// logic(cast(A), cast(B)) to cast(logic(A, B)) should be performed. 419 /// 420 /// \param CI The cast we are interested in. 421 /// 422 /// \return true if this cast actually results in any code being generated and 423 /// if it cannot already be eliminated by some other transformation. 424 bool shouldOptimizeCast(CastInst *CI); 425 426 /// Try to optimize a sequence of instructions checking if an operation 427 /// on LHS and RHS overflows. 428 /// 429 /// If this overflow check is done via one of the overflow check intrinsics, 430 /// then CtxI has to be the call instruction calling that intrinsic. If this 431 /// overflow check is done by arithmetic followed by a compare, then CtxI has 432 /// to be the arithmetic instruction. 433 /// 434 /// If a simplification is possible, stores the simplified result of the 435 /// operation in OperationResult and result of the overflow check in 436 /// OverflowResult, and return true. If no simplification is possible, 437 /// returns false. 438 bool OptimizeOverflowCheck(Instruction::BinaryOps BinaryOp, bool IsSigned, 439 Value *LHS, Value *RHS, 440 Instruction &CtxI, Value *&OperationResult, 441 Constant *&OverflowResult); 442 443 Instruction *visitCallBase(CallBase &Call); 444 Instruction *tryOptimizeCall(CallInst *CI); 445 bool transformConstExprCastCall(CallBase &Call); 446 Instruction *transformCallThroughTrampoline(CallBase &Call, 447 IntrinsicInst &Tramp); 448 449 Value *simplifyMaskedLoad(IntrinsicInst &II); 450 Instruction *simplifyMaskedStore(IntrinsicInst &II); 451 Instruction *simplifyMaskedGather(IntrinsicInst &II); 452 Instruction *simplifyMaskedScatter(IntrinsicInst &II); 453 454 /// Transform (zext icmp) to bitwise / integer operations in order to 455 /// eliminate it. 456 /// 457 /// \param ICI The icmp of the (zext icmp) pair we are interested in. 458 /// \parem CI The zext of the (zext icmp) pair we are interested in. 459 /// \param DoTransform Pass false to just test whether the given (zext icmp) 460 /// would be transformed. Pass true to actually perform the transformation. 461 /// 462 /// \return null if the transformation cannot be performed. If the 463 /// transformation can be performed the new instruction that replaces the 464 /// (zext icmp) pair will be returned (if \p DoTransform is false the 465 /// unmodified \p ICI will be returned in this case). 466 Instruction *transformZExtICmp(ICmpInst *ICI, ZExtInst &CI, 467 bool DoTransform = true); 468 469 Instruction *transformSExtICmp(ICmpInst *ICI, Instruction &CI); 470 471 bool willNotOverflowSignedAdd(const Value *LHS, const Value *RHS, 472 const Instruction &CxtI) const { 473 return computeOverflowForSignedAdd(LHS, RHS, &CxtI) == 474 OverflowResult::NeverOverflows; 475 } 476 477 bool willNotOverflowUnsignedAdd(const Value *LHS, const Value *RHS, 478 const Instruction &CxtI) const { 479 return computeOverflowForUnsignedAdd(LHS, RHS, &CxtI) == 480 OverflowResult::NeverOverflows; 481 } 482 483 bool willNotOverflowAdd(const Value *LHS, const Value *RHS, 484 const Instruction &CxtI, bool IsSigned) const { 485 return IsSigned ? willNotOverflowSignedAdd(LHS, RHS, CxtI) 486 : willNotOverflowUnsignedAdd(LHS, RHS, CxtI); 487 } 488 489 bool willNotOverflowSignedSub(const Value *LHS, const Value *RHS, 490 const Instruction &CxtI) const { 491 return computeOverflowForSignedSub(LHS, RHS, &CxtI) == 492 OverflowResult::NeverOverflows; 493 } 494 495 bool willNotOverflowUnsignedSub(const Value *LHS, const Value *RHS, 496 const Instruction &CxtI) const { 497 return computeOverflowForUnsignedSub(LHS, RHS, &CxtI) == 498 OverflowResult::NeverOverflows; 499 } 500 501 bool willNotOverflowSub(const Value *LHS, const Value *RHS, 502 const Instruction &CxtI, bool IsSigned) const { 503 return IsSigned ? willNotOverflowSignedSub(LHS, RHS, CxtI) 504 : willNotOverflowUnsignedSub(LHS, RHS, CxtI); 505 } 506 507 bool willNotOverflowSignedMul(const Value *LHS, const Value *RHS, 508 const Instruction &CxtI) const { 509 return computeOverflowForSignedMul(LHS, RHS, &CxtI) == 510 OverflowResult::NeverOverflows; 511 } 512 513 bool willNotOverflowUnsignedMul(const Value *LHS, const Value *RHS, 514 const Instruction &CxtI) const { 515 return computeOverflowForUnsignedMul(LHS, RHS, &CxtI) == 516 OverflowResult::NeverOverflows; 517 } 518 519 bool willNotOverflowMul(const Value *LHS, const Value *RHS, 520 const Instruction &CxtI, bool IsSigned) const { 521 return IsSigned ? willNotOverflowSignedMul(LHS, RHS, CxtI) 522 : willNotOverflowUnsignedMul(LHS, RHS, CxtI); 523 } 524 525 bool willNotOverflow(BinaryOperator::BinaryOps Opcode, const Value *LHS, 526 const Value *RHS, const Instruction &CxtI, 527 bool IsSigned) const { 528 switch (Opcode) { 529 case Instruction::Add: return willNotOverflowAdd(LHS, RHS, CxtI, IsSigned); 530 case Instruction::Sub: return willNotOverflowSub(LHS, RHS, CxtI, IsSigned); 531 case Instruction::Mul: return willNotOverflowMul(LHS, RHS, CxtI, IsSigned); 532 default: llvm_unreachable("Unexpected opcode for overflow query"); 533 } 534 } 535 536 Value *EmitGEPOffset(User *GEP); 537 Instruction *scalarizePHI(ExtractElementInst &EI, PHINode *PN); 538 Instruction *foldCastedBitwiseLogic(BinaryOperator &I); 539 Instruction *narrowBinOp(TruncInst &Trunc); 540 Instruction *narrowMaskedBinOp(BinaryOperator &And); 541 Instruction *narrowMathIfNoOverflow(BinaryOperator &I); 542 Instruction *narrowRotate(TruncInst &Trunc); 543 Instruction *optimizeBitCastFromPhi(CastInst &CI, PHINode *PN); 544 545 /// Determine if a pair of casts can be replaced by a single cast. 546 /// 547 /// \param CI1 The first of a pair of casts. 548 /// \param CI2 The second of a pair of casts. 549 /// 550 /// \return 0 if the cast pair cannot be eliminated, otherwise returns an 551 /// Instruction::CastOps value for a cast that can replace the pair, casting 552 /// CI1->getSrcTy() to CI2->getDstTy(). 553 /// 554 /// \see CastInst::isEliminableCastPair 555 Instruction::CastOps isEliminableCastPair(const CastInst *CI1, 556 const CastInst *CI2); 557 558 Value *foldAndOfICmps(ICmpInst *LHS, ICmpInst *RHS, Instruction &CxtI); 559 Value *foldOrOfICmps(ICmpInst *LHS, ICmpInst *RHS, Instruction &CxtI); 560 Value *foldXorOfICmps(ICmpInst *LHS, ICmpInst *RHS); 561 562 /// Optimize (fcmp)&(fcmp) or (fcmp)|(fcmp). 563 /// NOTE: Unlike most of instcombine, this returns a Value which should 564 /// already be inserted into the function. 565 Value *foldLogicOfFCmps(FCmpInst *LHS, FCmpInst *RHS, bool IsAnd); 566 567 Value *foldAndOrOfICmpsOfAndWithPow2(ICmpInst *LHS, ICmpInst *RHS, 568 bool JoinedByAnd, Instruction &CxtI); 569 Value *matchSelectFromAndOr(Value *A, Value *B, Value *C, Value *D); 570 Value *getSelectCondition(Value *A, Value *B); 571 572 Instruction *foldIntrinsicWithOverflowCommon(IntrinsicInst *II); 573 574 public: 575 /// Inserts an instruction \p New before instruction \p Old 576 /// 577 /// Also adds the new instruction to the worklist and returns \p New so that 578 /// it is suitable for use as the return from the visitation patterns. 579 Instruction *InsertNewInstBefore(Instruction *New, Instruction &Old) { 580 assert(New && !New->getParent() && 581 "New instruction already inserted into a basic block!"); 582 BasicBlock *BB = Old.getParent(); 583 BB->getInstList().insert(Old.getIterator(), New); // Insert inst 584 Worklist.Add(New); 585 return New; 586 } 587 588 /// Same as InsertNewInstBefore, but also sets the debug loc. 589 Instruction *InsertNewInstWith(Instruction *New, Instruction &Old) { 590 New->setDebugLoc(Old.getDebugLoc()); 591 return InsertNewInstBefore(New, Old); 592 } 593 594 /// A combiner-aware RAUW-like routine. 595 /// 596 /// This method is to be used when an instruction is found to be dead, 597 /// replaceable with another preexisting expression. Here we add all uses of 598 /// I to the worklist, replace all uses of I with the new value, then return 599 /// I, so that the inst combiner will know that I was modified. 600 Instruction *replaceInstUsesWith(Instruction &I, Value *V) { 601 // If there are no uses to replace, then we return nullptr to indicate that 602 // no changes were made to the program. 603 if (I.use_empty()) return nullptr; 604 605 Worklist.AddUsersToWorkList(I); // Add all modified instrs to worklist. 606 607 // If we are replacing the instruction with itself, this must be in a 608 // segment of unreachable code, so just clobber the instruction. 609 if (&I == V) 610 V = UndefValue::get(I.getType()); 611 612 LLVM_DEBUG(dbgs() << "IC: Replacing " << I << "\n" 613 << " with " << *V << '\n'); 614 615 I.replaceAllUsesWith(V); 616 return &I; 617 } 618 619 /// Creates a result tuple for an overflow intrinsic \p II with a given 620 /// \p Result and a constant \p Overflow value. 621 Instruction *CreateOverflowTuple(IntrinsicInst *II, Value *Result, 622 Constant *Overflow) { 623 Constant *V[] = {UndefValue::get(Result->getType()), Overflow}; 624 StructType *ST = cast<StructType>(II->getType()); 625 Constant *Struct = ConstantStruct::get(ST, V); 626 return InsertValueInst::Create(Struct, Result, 0); 627 } 628 629 /// Create and insert the idiom we use to indicate a block is unreachable 630 /// without having to rewrite the CFG from within InstCombine. 631 void CreateNonTerminatorUnreachable(Instruction *InsertAt) { 632 auto &Ctx = InsertAt->getContext(); 633 new StoreInst(ConstantInt::getTrue(Ctx), 634 UndefValue::get(Type::getInt1PtrTy(Ctx)), 635 InsertAt); 636 } 637 638 639 /// Combiner aware instruction erasure. 640 /// 641 /// When dealing with an instruction that has side effects or produces a void 642 /// value, we can't rely on DCE to delete the instruction. Instead, visit 643 /// methods should return the value returned by this function. 644 Instruction *eraseInstFromFunction(Instruction &I) { 645 LLVM_DEBUG(dbgs() << "IC: ERASE " << I << '\n'); 646 assert(I.use_empty() && "Cannot erase instruction that is used!"); 647 salvageDebugInfo(I); 648 649 // Make sure that we reprocess all operands now that we reduced their 650 // use counts. 651 if (I.getNumOperands() < 8) { 652 for (Use &Operand : I.operands()) 653 if (auto *Inst = dyn_cast<Instruction>(Operand)) 654 Worklist.Add(Inst); 655 } 656 Worklist.Remove(&I); 657 I.eraseFromParent(); 658 MadeIRChange = true; 659 return nullptr; // Don't do anything with FI 660 } 661 662 void computeKnownBits(const Value *V, KnownBits &Known, 663 unsigned Depth, const Instruction *CxtI) const { 664 llvm::computeKnownBits(V, Known, DL, Depth, &AC, CxtI, &DT); 665 } 666 667 KnownBits computeKnownBits(const Value *V, unsigned Depth, 668 const Instruction *CxtI) const { 669 return llvm::computeKnownBits(V, DL, Depth, &AC, CxtI, &DT); 670 } 671 672 bool isKnownToBeAPowerOfTwo(const Value *V, bool OrZero = false, 673 unsigned Depth = 0, 674 const Instruction *CxtI = nullptr) { 675 return llvm::isKnownToBeAPowerOfTwo(V, DL, OrZero, Depth, &AC, CxtI, &DT); 676 } 677 678 bool MaskedValueIsZero(const Value *V, const APInt &Mask, unsigned Depth = 0, 679 const Instruction *CxtI = nullptr) const { 680 return llvm::MaskedValueIsZero(V, Mask, DL, Depth, &AC, CxtI, &DT); 681 } 682 683 unsigned ComputeNumSignBits(const Value *Op, unsigned Depth = 0, 684 const Instruction *CxtI = nullptr) const { 685 return llvm::ComputeNumSignBits(Op, DL, Depth, &AC, CxtI, &DT); 686 } 687 688 OverflowResult computeOverflowForUnsignedMul(const Value *LHS, 689 const Value *RHS, 690 const Instruction *CxtI) const { 691 return llvm::computeOverflowForUnsignedMul(LHS, RHS, DL, &AC, CxtI, &DT); 692 } 693 694 OverflowResult computeOverflowForSignedMul(const Value *LHS, 695 const Value *RHS, 696 const Instruction *CxtI) const { 697 return llvm::computeOverflowForSignedMul(LHS, RHS, DL, &AC, CxtI, &DT); 698 } 699 700 OverflowResult computeOverflowForUnsignedAdd(const Value *LHS, 701 const Value *RHS, 702 const Instruction *CxtI) const { 703 return llvm::computeOverflowForUnsignedAdd(LHS, RHS, DL, &AC, CxtI, &DT); 704 } 705 706 OverflowResult computeOverflowForSignedAdd(const Value *LHS, 707 const Value *RHS, 708 const Instruction *CxtI) const { 709 return llvm::computeOverflowForSignedAdd(LHS, RHS, DL, &AC, CxtI, &DT); 710 } 711 712 OverflowResult computeOverflowForUnsignedSub(const Value *LHS, 713 const Value *RHS, 714 const Instruction *CxtI) const { 715 return llvm::computeOverflowForUnsignedSub(LHS, RHS, DL, &AC, CxtI, &DT); 716 } 717 718 OverflowResult computeOverflowForSignedSub(const Value *LHS, const Value *RHS, 719 const Instruction *CxtI) const { 720 return llvm::computeOverflowForSignedSub(LHS, RHS, DL, &AC, CxtI, &DT); 721 } 722 723 OverflowResult computeOverflow( 724 Instruction::BinaryOps BinaryOp, bool IsSigned, 725 Value *LHS, Value *RHS, Instruction *CxtI) const; 726 727 /// Maximum size of array considered when transforming. 728 uint64_t MaxArraySizeForCombine; 729 730 private: 731 /// Performs a few simplifications for operators which are associative 732 /// or commutative. 733 bool SimplifyAssociativeOrCommutative(BinaryOperator &I); 734 735 /// Tries to simplify binary operations which some other binary 736 /// operation distributes over. 737 /// 738 /// It does this by either by factorizing out common terms (eg "(A*B)+(A*C)" 739 /// -> "A*(B+C)") or expanding out if this results in simplifications (eg: "A 740 /// & (B | C) -> (A&B) | (A&C)" if this is a win). Returns the simplified 741 /// value, or null if it didn't simplify. 742 Value *SimplifyUsingDistributiveLaws(BinaryOperator &I); 743 744 /// Tries to simplify add operations using the definition of remainder. 745 /// 746 /// The definition of remainder is X % C = X - (X / C ) * C. The add 747 /// expression X % C0 + (( X / C0 ) % C1) * C0 can be simplified to 748 /// X % (C0 * C1) 749 Value *SimplifyAddWithRemainder(BinaryOperator &I); 750 751 // Binary Op helper for select operations where the expression can be 752 // efficiently reorganized. 753 Value *SimplifySelectsFeedingBinaryOp(BinaryOperator &I, Value *LHS, 754 Value *RHS); 755 756 /// This tries to simplify binary operations by factorizing out common terms 757 /// (e. g. "(A*B)+(A*C)" -> "A*(B+C)"). 758 Value *tryFactorization(BinaryOperator &, Instruction::BinaryOps, Value *, 759 Value *, Value *, Value *); 760 761 /// Match a select chain which produces one of three values based on whether 762 /// the LHS is less than, equal to, or greater than RHS respectively. 763 /// Return true if we matched a three way compare idiom. The LHS, RHS, Less, 764 /// Equal and Greater values are saved in the matching process and returned to 765 /// the caller. 766 bool matchThreeWayIntCompare(SelectInst *SI, Value *&LHS, Value *&RHS, 767 ConstantInt *&Less, ConstantInt *&Equal, 768 ConstantInt *&Greater); 769 770 /// Attempts to replace V with a simpler value based on the demanded 771 /// bits. 772 Value *SimplifyDemandedUseBits(Value *V, APInt DemandedMask, KnownBits &Known, 773 unsigned Depth, Instruction *CxtI); 774 bool SimplifyDemandedBits(Instruction *I, unsigned Op, 775 const APInt &DemandedMask, KnownBits &Known, 776 unsigned Depth = 0); 777 778 /// Helper routine of SimplifyDemandedUseBits. It computes KnownZero/KnownOne 779 /// bits. It also tries to handle simplifications that can be done based on 780 /// DemandedMask, but without modifying the Instruction. 781 Value *SimplifyMultipleUseDemandedBits(Instruction *I, 782 const APInt &DemandedMask, 783 KnownBits &Known, 784 unsigned Depth, Instruction *CxtI); 785 786 /// Helper routine of SimplifyDemandedUseBits. It tries to simplify demanded 787 /// bit for "r1 = shr x, c1; r2 = shl r1, c2" instruction sequence. 788 Value *simplifyShrShlDemandedBits( 789 Instruction *Shr, const APInt &ShrOp1, Instruction *Shl, 790 const APInt &ShlOp1, const APInt &DemandedMask, KnownBits &Known); 791 792 /// Tries to simplify operands to an integer instruction based on its 793 /// demanded bits. 794 bool SimplifyDemandedInstructionBits(Instruction &Inst); 795 796 Value *simplifyAMDGCNMemoryIntrinsicDemanded(IntrinsicInst *II, 797 APInt DemandedElts, 798 int DmaskIdx = -1); 799 800 Value *SimplifyDemandedVectorElts(Value *V, APInt DemandedElts, 801 APInt &UndefElts, unsigned Depth = 0); 802 803 /// Canonicalize the position of binops relative to shufflevector. 804 Instruction *foldVectorBinop(BinaryOperator &Inst); 805 806 /// Given a binary operator, cast instruction, or select which has a PHI node 807 /// as operand #0, see if we can fold the instruction into the PHI (which is 808 /// only possible if all operands to the PHI are constants). 809 Instruction *foldOpIntoPhi(Instruction &I, PHINode *PN); 810 811 /// Given an instruction with a select as one operand and a constant as the 812 /// other operand, try to fold the binary operator into the select arguments. 813 /// This also works for Cast instructions, which obviously do not have a 814 /// second operand. 815 Instruction *FoldOpIntoSelect(Instruction &Op, SelectInst *SI); 816 817 /// This is a convenience wrapper function for the above two functions. 818 Instruction *foldBinOpIntoSelectOrPhi(BinaryOperator &I); 819 820 Instruction *foldAddWithConstant(BinaryOperator &Add); 821 822 /// Try to rotate an operation below a PHI node, using PHI nodes for 823 /// its operands. 824 Instruction *FoldPHIArgOpIntoPHI(PHINode &PN); 825 Instruction *FoldPHIArgBinOpIntoPHI(PHINode &PN); 826 Instruction *FoldPHIArgGEPIntoPHI(PHINode &PN); 827 Instruction *FoldPHIArgLoadIntoPHI(PHINode &PN); 828 Instruction *FoldPHIArgZextsIntoPHI(PHINode &PN); 829 830 /// If an integer typed PHI has only one use which is an IntToPtr operation, 831 /// replace the PHI with an existing pointer typed PHI if it exists. Otherwise 832 /// insert a new pointer typed PHI and replace the original one. 833 Instruction *FoldIntegerTypedPHI(PHINode &PN); 834 835 /// Helper function for FoldPHIArgXIntoPHI() to set debug location for the 836 /// folded operation. 837 void PHIArgMergedDebugLoc(Instruction *Inst, PHINode &PN); 838 839 Instruction *foldGEPICmp(GEPOperator *GEPLHS, Value *RHS, 840 ICmpInst::Predicate Cond, Instruction &I); 841 Instruction *foldAllocaCmp(ICmpInst &ICI, const AllocaInst *Alloca, 842 const Value *Other); 843 Instruction *foldCmpLoadFromIndexedGlobal(GetElementPtrInst *GEP, 844 GlobalVariable *GV, CmpInst &ICI, 845 ConstantInt *AndCst = nullptr); 846 Instruction *foldFCmpIntToFPConst(FCmpInst &I, Instruction *LHSI, 847 Constant *RHSC); 848 Instruction *foldICmpAddOpConst(Value *X, const APInt &C, 849 ICmpInst::Predicate Pred); 850 Instruction *foldICmpWithCastAndCast(ICmpInst &ICI); 851 852 Instruction *foldICmpUsingKnownBits(ICmpInst &Cmp); 853 Instruction *foldICmpWithDominatingICmp(ICmpInst &Cmp); 854 Instruction *foldICmpWithConstant(ICmpInst &Cmp); 855 Instruction *foldICmpInstWithConstant(ICmpInst &Cmp); 856 Instruction *foldICmpInstWithConstantNotInt(ICmpInst &Cmp); 857 Instruction *foldICmpBinOp(ICmpInst &Cmp); 858 Instruction *foldICmpEquality(ICmpInst &Cmp); 859 Instruction *foldICmpWithZero(ICmpInst &Cmp); 860 861 Instruction *foldICmpSelectConstant(ICmpInst &Cmp, SelectInst *Select, 862 ConstantInt *C); 863 Instruction *foldICmpTruncConstant(ICmpInst &Cmp, TruncInst *Trunc, 864 const APInt &C); 865 Instruction *foldICmpAndConstant(ICmpInst &Cmp, BinaryOperator *And, 866 const APInt &C); 867 Instruction *foldICmpXorConstant(ICmpInst &Cmp, BinaryOperator *Xor, 868 const APInt &C); 869 Instruction *foldICmpOrConstant(ICmpInst &Cmp, BinaryOperator *Or, 870 const APInt &C); 871 Instruction *foldICmpMulConstant(ICmpInst &Cmp, BinaryOperator *Mul, 872 const APInt &C); 873 Instruction *foldICmpShlConstant(ICmpInst &Cmp, BinaryOperator *Shl, 874 const APInt &C); 875 Instruction *foldICmpShrConstant(ICmpInst &Cmp, BinaryOperator *Shr, 876 const APInt &C); 877 Instruction *foldICmpUDivConstant(ICmpInst &Cmp, BinaryOperator *UDiv, 878 const APInt &C); 879 Instruction *foldICmpDivConstant(ICmpInst &Cmp, BinaryOperator *Div, 880 const APInt &C); 881 Instruction *foldICmpSubConstant(ICmpInst &Cmp, BinaryOperator *Sub, 882 const APInt &C); 883 Instruction *foldICmpAddConstant(ICmpInst &Cmp, BinaryOperator *Add, 884 const APInt &C); 885 Instruction *foldICmpAndConstConst(ICmpInst &Cmp, BinaryOperator *And, 886 const APInt &C1); 887 Instruction *foldICmpAndShift(ICmpInst &Cmp, BinaryOperator *And, 888 const APInt &C1, const APInt &C2); 889 Instruction *foldICmpShrConstConst(ICmpInst &I, Value *ShAmt, const APInt &C1, 890 const APInt &C2); 891 Instruction *foldICmpShlConstConst(ICmpInst &I, Value *ShAmt, const APInt &C1, 892 const APInt &C2); 893 894 Instruction *foldICmpBinOpEqualityWithConstant(ICmpInst &Cmp, 895 BinaryOperator *BO, 896 const APInt &C); 897 Instruction *foldICmpIntrinsicWithConstant(ICmpInst &ICI, IntrinsicInst *II, 898 const APInt &C); 899 Instruction *foldICmpEqIntrinsicWithConstant(ICmpInst &ICI, IntrinsicInst *II, 900 const APInt &C); 901 902 // Helpers of visitSelectInst(). 903 Instruction *foldSelectExtConst(SelectInst &Sel); 904 Instruction *foldSelectOpOp(SelectInst &SI, Instruction *TI, Instruction *FI); 905 Instruction *foldSelectIntoOp(SelectInst &SI, Value *, Value *); 906 Instruction *foldSPFofSPF(Instruction *Inner, SelectPatternFlavor SPF1, 907 Value *A, Value *B, Instruction &Outer, 908 SelectPatternFlavor SPF2, Value *C); 909 Instruction *foldSelectInstWithICmp(SelectInst &SI, ICmpInst *ICI); 910 911 Instruction *OptAndOp(BinaryOperator *Op, ConstantInt *OpRHS, 912 ConstantInt *AndRHS, BinaryOperator &TheAnd); 913 914 Value *insertRangeTest(Value *V, const APInt &Lo, const APInt &Hi, 915 bool isSigned, bool Inside); 916 Instruction *PromoteCastOfAllocation(BitCastInst &CI, AllocaInst &AI); 917 bool mergeStoreIntoSuccessor(StoreInst &SI); 918 919 /// Given an 'or' instruction, check to see if it is part of a bswap idiom. 920 /// If so, return the equivalent bswap intrinsic. 921 Instruction *matchBSwap(BinaryOperator &Or); 922 923 Instruction *SimplifyAnyMemTransfer(AnyMemTransferInst *MI); 924 Instruction *SimplifyAnyMemSet(AnyMemSetInst *MI); 925 926 Value *EvaluateInDifferentType(Value *V, Type *Ty, bool isSigned); 927 928 /// Returns a value X such that Val = X * Scale, or null if none. 929 /// 930 /// If the multiplication is known not to overflow then NoSignedWrap is set. 931 Value *Descale(Value *Val, APInt Scale, bool &NoSignedWrap); 932 }; 933 934 } // end namespace llvm 935 936 #undef DEBUG_TYPE 937 938 #endif // LLVM_LIB_TRANSFORMS_INSTCOMBINE_INSTCOMBINEINTERNAL_H 939