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/Statistic.h" 19 #include "llvm/ADT/PostOrderIterator.h" 20 #include "llvm/Analysis/InstructionSimplify.h" 21 #include "llvm/Analysis/TargetFolder.h" 22 #include "llvm/Analysis/ValueTracking.h" 23 #include "llvm/IR/IRBuilder.h" 24 #include "llvm/IR/InstVisitor.h" 25 #include "llvm/IR/PatternMatch.h" 26 #include "llvm/IR/Value.h" 27 #include "llvm/Support/Debug.h" 28 #include "llvm/Support/KnownBits.h" 29 #include "llvm/Transforms/InstCombine/InstCombiner.h" 30 #include "llvm/Transforms/Utils/Local.h" 31 #include <cassert> 32 33 #define DEBUG_TYPE "instcombine" 34 #include "llvm/Transforms/Utils/InstructionWorklist.h" 35 36 using namespace llvm::PatternMatch; 37 38 // As a default, let's assume that we want to be aggressive, 39 // and attempt to traverse with no limits in attempt to sink negation. 40 static constexpr unsigned NegatorDefaultMaxDepth = ~0U; 41 42 // Let's guesstimate that most often we will end up visiting/producing 43 // fairly small number of new instructions. 44 static constexpr unsigned NegatorMaxNodesSSO = 16; 45 46 namespace llvm { 47 48 class AAResults; 49 class APInt; 50 class AssumptionCache; 51 class BlockFrequencyInfo; 52 class DataLayout; 53 class DominatorTree; 54 class GEPOperator; 55 class GlobalVariable; 56 class LoopInfo; 57 class OptimizationRemarkEmitter; 58 class ProfileSummaryInfo; 59 class TargetLibraryInfo; 60 class User; 61 62 class LLVM_LIBRARY_VISIBILITY InstCombinerImpl final 63 : public InstCombiner, 64 public InstVisitor<InstCombinerImpl, Instruction *> { 65 public: 66 InstCombinerImpl(InstructionWorklist &Worklist, BuilderTy &Builder, 67 bool MinimizeSize, AAResults *AA, AssumptionCache &AC, 68 TargetLibraryInfo &TLI, TargetTransformInfo &TTI, 69 DominatorTree &DT, OptimizationRemarkEmitter &ORE, 70 BlockFrequencyInfo *BFI, ProfileSummaryInfo *PSI, 71 const DataLayout &DL, LoopInfo *LI) 72 : InstCombiner(Worklist, Builder, MinimizeSize, AA, AC, TLI, TTI, DT, ORE, 73 BFI, PSI, DL, LI) {} 74 75 virtual ~InstCombinerImpl() = default; 76 77 /// Perform early cleanup and prepare the InstCombine worklist. 78 bool prepareWorklist(Function &F, 79 ReversePostOrderTraversal<BasicBlock *> &RPOT); 80 81 /// Run the combiner over the entire worklist until it is empty. 82 /// 83 /// \returns true if the IR is changed. 84 bool run(); 85 86 // Visitation implementation - Implement instruction combining for different 87 // instruction types. The semantics are as follows: 88 // Return Value: 89 // null - No change was made 90 // I - Change was made, I is still valid, I may be dead though 91 // otherwise - Change was made, replace I with returned instruction 92 // 93 Instruction *visitFNeg(UnaryOperator &I); 94 Instruction *visitAdd(BinaryOperator &I); 95 Instruction *visitFAdd(BinaryOperator &I); 96 Value *OptimizePointerDifference( 97 Value *LHS, Value *RHS, Type *Ty, bool isNUW); 98 Instruction *visitSub(BinaryOperator &I); 99 Instruction *visitFSub(BinaryOperator &I); 100 Instruction *visitMul(BinaryOperator &I); 101 Instruction *foldFMulReassoc(BinaryOperator &I); 102 Instruction *visitFMul(BinaryOperator &I); 103 Instruction *visitURem(BinaryOperator &I); 104 Instruction *visitSRem(BinaryOperator &I); 105 Instruction *visitFRem(BinaryOperator &I); 106 bool simplifyDivRemOfSelectWithZeroOp(BinaryOperator &I); 107 Instruction *commonIRemTransforms(BinaryOperator &I); 108 Instruction *commonIDivTransforms(BinaryOperator &I); 109 Instruction *visitUDiv(BinaryOperator &I); 110 Instruction *visitSDiv(BinaryOperator &I); 111 Instruction *visitFDiv(BinaryOperator &I); 112 Value *simplifyRangeCheck(ICmpInst *Cmp0, ICmpInst *Cmp1, bool Inverted); 113 Instruction *visitAnd(BinaryOperator &I); 114 Instruction *visitOr(BinaryOperator &I); 115 bool sinkNotIntoLogicalOp(Instruction &I); 116 bool sinkNotIntoOtherHandOfLogicalOp(Instruction &I); 117 Instruction *visitXor(BinaryOperator &I); 118 Instruction *visitShl(BinaryOperator &I); 119 Value *reassociateShiftAmtsOfTwoSameDirectionShifts( 120 BinaryOperator *Sh0, const SimplifyQuery &SQ, 121 bool AnalyzeForSignBitExtraction = false); 122 Instruction *canonicalizeCondSignextOfHighBitExtractToSignextHighBitExtract( 123 BinaryOperator &I); 124 Instruction *foldVariableSignZeroExtensionOfVariableHighBitExtract( 125 BinaryOperator &OldAShr); 126 Instruction *visitAShr(BinaryOperator &I); 127 Instruction *visitLShr(BinaryOperator &I); 128 Instruction *commonShiftTransforms(BinaryOperator &I); 129 Instruction *visitFCmpInst(FCmpInst &I); 130 CmpInst *canonicalizeICmpPredicate(CmpInst &I); 131 Instruction *visitICmpInst(ICmpInst &I); 132 Instruction *FoldShiftByConstant(Value *Op0, Constant *Op1, 133 BinaryOperator &I); 134 Instruction *commonCastTransforms(CastInst &CI); 135 Instruction *visitTrunc(TruncInst &CI); 136 Instruction *visitZExt(ZExtInst &Zext); 137 Instruction *visitSExt(SExtInst &Sext); 138 Instruction *visitFPTrunc(FPTruncInst &CI); 139 Instruction *visitFPExt(CastInst &CI); 140 Instruction *visitFPToUI(FPToUIInst &FI); 141 Instruction *visitFPToSI(FPToSIInst &FI); 142 Instruction *visitUIToFP(CastInst &CI); 143 Instruction *visitSIToFP(CastInst &CI); 144 Instruction *visitPtrToInt(PtrToIntInst &CI); 145 Instruction *visitIntToPtr(IntToPtrInst &CI); 146 Instruction *visitBitCast(BitCastInst &CI); 147 Instruction *visitAddrSpaceCast(AddrSpaceCastInst &CI); 148 Instruction *foldItoFPtoI(CastInst &FI); 149 Instruction *visitSelectInst(SelectInst &SI); 150 Instruction *visitCallInst(CallInst &CI); 151 Instruction *visitInvokeInst(InvokeInst &II); 152 Instruction *visitCallBrInst(CallBrInst &CBI); 153 154 Instruction *SliceUpIllegalIntegerPHI(PHINode &PN); 155 Instruction *visitPHINode(PHINode &PN); 156 Instruction *visitGetElementPtrInst(GetElementPtrInst &GEP); 157 Instruction *visitGEPOfGEP(GetElementPtrInst &GEP, GEPOperator *Src); 158 Instruction *visitAllocaInst(AllocaInst &AI); 159 Instruction *visitAllocSite(Instruction &FI); 160 Instruction *visitFree(CallInst &FI, Value *FreedOp); 161 Instruction *visitLoadInst(LoadInst &LI); 162 Instruction *visitStoreInst(StoreInst &SI); 163 Instruction *visitAtomicRMWInst(AtomicRMWInst &SI); 164 Instruction *visitUnconditionalBranchInst(BranchInst &BI); 165 Instruction *visitBranchInst(BranchInst &BI); 166 Instruction *visitFenceInst(FenceInst &FI); 167 Instruction *visitSwitchInst(SwitchInst &SI); 168 Instruction *visitReturnInst(ReturnInst &RI); 169 Instruction *visitUnreachableInst(UnreachableInst &I); 170 Instruction * 171 foldAggregateConstructionIntoAggregateReuse(InsertValueInst &OrigIVI); 172 Instruction *visitInsertValueInst(InsertValueInst &IV); 173 Instruction *visitInsertElementInst(InsertElementInst &IE); 174 Instruction *visitExtractElementInst(ExtractElementInst &EI); 175 Instruction *simplifyBinOpSplats(ShuffleVectorInst &SVI); 176 Instruction *visitShuffleVectorInst(ShuffleVectorInst &SVI); 177 Instruction *visitExtractValueInst(ExtractValueInst &EV); 178 Instruction *visitLandingPadInst(LandingPadInst &LI); 179 Instruction *visitVAEndInst(VAEndInst &I); 180 Value *pushFreezeToPreventPoisonFromPropagating(FreezeInst &FI); 181 bool freezeOtherUses(FreezeInst &FI); 182 Instruction *foldFreezeIntoRecurrence(FreezeInst &I, PHINode *PN); 183 Instruction *visitFreeze(FreezeInst &I); 184 185 /// Specify what to return for unhandled instructions. 186 Instruction *visitInstruction(Instruction &I) { return nullptr; } 187 188 /// True when DB dominates all uses of DI except UI. 189 /// UI must be in the same block as DI. 190 /// The routine checks that the DI parent and DB are different. 191 bool dominatesAllUses(const Instruction *DI, const Instruction *UI, 192 const BasicBlock *DB) const; 193 194 /// Try to replace select with select operand SIOpd in SI-ICmp sequence. 195 bool replacedSelectWithOperand(SelectInst *SI, const ICmpInst *Icmp, 196 const unsigned SIOpd); 197 198 LoadInst *combineLoadToNewType(LoadInst &LI, Type *NewTy, 199 const Twine &Suffix = ""); 200 201 KnownFPClass computeKnownFPClass(Value *Val, FastMathFlags FMF, 202 FPClassTest Interested = fcAllFlags, 203 const Instruction *CtxI = nullptr, 204 unsigned Depth = 0) const { 205 return llvm::computeKnownFPClass(Val, FMF, DL, Interested, Depth, &TLI, &AC, 206 CtxI, &DT); 207 } 208 209 KnownFPClass computeKnownFPClass(Value *Val, 210 FPClassTest Interested = fcAllFlags, 211 const Instruction *CtxI = nullptr, 212 unsigned Depth = 0) const { 213 return llvm::computeKnownFPClass(Val, DL, Interested, Depth, &TLI, &AC, 214 CtxI, &DT); 215 } 216 217 /// Check if fmul \p MulVal, +0.0 will yield +0.0 (or signed zero is 218 /// ignorable). 219 bool fmulByZeroIsZero(Value *MulVal, FastMathFlags FMF, 220 const Instruction *CtxI) const; 221 222 Constant *getLosslessTrunc(Constant *C, Type *TruncTy, unsigned ExtOp) { 223 Constant *TruncC = ConstantExpr::getTrunc(C, TruncTy); 224 Constant *ExtTruncC = 225 ConstantFoldCastOperand(ExtOp, TruncC, C->getType(), DL); 226 if (ExtTruncC && ExtTruncC == C) 227 return TruncC; 228 return nullptr; 229 } 230 231 Constant *getLosslessUnsignedTrunc(Constant *C, Type *TruncTy) { 232 return getLosslessTrunc(C, TruncTy, Instruction::ZExt); 233 } 234 235 Constant *getLosslessSignedTrunc(Constant *C, Type *TruncTy) { 236 return getLosslessTrunc(C, TruncTy, Instruction::SExt); 237 } 238 239 private: 240 bool annotateAnyAllocSite(CallBase &Call, const TargetLibraryInfo *TLI); 241 bool isDesirableIntType(unsigned BitWidth) const; 242 bool shouldChangeType(unsigned FromBitWidth, unsigned ToBitWidth) const; 243 bool shouldChangeType(Type *From, Type *To) const; 244 Value *dyn_castNegVal(Value *V) const; 245 246 /// Classify whether a cast is worth optimizing. 247 /// 248 /// This is a helper to decide whether the simplification of 249 /// logic(cast(A), cast(B)) to cast(logic(A, B)) should be performed. 250 /// 251 /// \param CI The cast we are interested in. 252 /// 253 /// \return true if this cast actually results in any code being generated and 254 /// if it cannot already be eliminated by some other transformation. 255 bool shouldOptimizeCast(CastInst *CI); 256 257 /// Try to optimize a sequence of instructions checking if an operation 258 /// on LHS and RHS overflows. 259 /// 260 /// If this overflow check is done via one of the overflow check intrinsics, 261 /// then CtxI has to be the call instruction calling that intrinsic. If this 262 /// overflow check is done by arithmetic followed by a compare, then CtxI has 263 /// to be the arithmetic instruction. 264 /// 265 /// If a simplification is possible, stores the simplified result of the 266 /// operation in OperationResult and result of the overflow check in 267 /// OverflowResult, and return true. If no simplification is possible, 268 /// returns false. 269 bool OptimizeOverflowCheck(Instruction::BinaryOps BinaryOp, bool IsSigned, 270 Value *LHS, Value *RHS, 271 Instruction &CtxI, Value *&OperationResult, 272 Constant *&OverflowResult); 273 274 Instruction *visitCallBase(CallBase &Call); 275 Instruction *tryOptimizeCall(CallInst *CI); 276 bool transformConstExprCastCall(CallBase &Call); 277 Instruction *transformCallThroughTrampoline(CallBase &Call, 278 IntrinsicInst &Tramp); 279 280 // Return (a, b) if (LHS, RHS) is known to be (a, b) or (b, a). 281 // Otherwise, return std::nullopt 282 // Currently it matches: 283 // - LHS = (select c, a, b), RHS = (select c, b, a) 284 // - LHS = (phi [a, BB0], [b, BB1]), RHS = (phi [b, BB0], [a, BB1]) 285 // - LHS = min(a, b), RHS = max(a, b) 286 std::optional<std::pair<Value *, Value *>> matchSymmetricPair(Value *LHS, 287 Value *RHS); 288 289 Value *simplifyMaskedLoad(IntrinsicInst &II); 290 Instruction *simplifyMaskedStore(IntrinsicInst &II); 291 Instruction *simplifyMaskedGather(IntrinsicInst &II); 292 Instruction *simplifyMaskedScatter(IntrinsicInst &II); 293 294 /// Transform (zext icmp) to bitwise / integer operations in order to 295 /// eliminate it. 296 /// 297 /// \param ICI The icmp of the (zext icmp) pair we are interested in. 298 /// \parem CI The zext of the (zext icmp) pair we are interested in. 299 /// 300 /// \return null if the transformation cannot be performed. If the 301 /// transformation can be performed the new instruction that replaces the 302 /// (zext icmp) pair will be returned. 303 Instruction *transformZExtICmp(ICmpInst *Cmp, ZExtInst &Zext); 304 305 Instruction *transformSExtICmp(ICmpInst *Cmp, SExtInst &Sext); 306 307 bool willNotOverflowSignedAdd(const WithCache<const Value *> &LHS, 308 const WithCache<const Value *> &RHS, 309 const Instruction &CxtI) const { 310 return computeOverflowForSignedAdd(LHS, RHS, &CxtI) == 311 OverflowResult::NeverOverflows; 312 } 313 314 bool willNotOverflowUnsignedAdd(const WithCache<const Value *> &LHS, 315 const WithCache<const Value *> &RHS, 316 const Instruction &CxtI) const { 317 return computeOverflowForUnsignedAdd(LHS, RHS, &CxtI) == 318 OverflowResult::NeverOverflows; 319 } 320 321 bool willNotOverflowAdd(const Value *LHS, const Value *RHS, 322 const Instruction &CxtI, bool IsSigned) const { 323 return IsSigned ? willNotOverflowSignedAdd(LHS, RHS, CxtI) 324 : willNotOverflowUnsignedAdd(LHS, RHS, CxtI); 325 } 326 327 bool willNotOverflowSignedSub(const Value *LHS, const Value *RHS, 328 const Instruction &CxtI) const { 329 return computeOverflowForSignedSub(LHS, RHS, &CxtI) == 330 OverflowResult::NeverOverflows; 331 } 332 333 bool willNotOverflowUnsignedSub(const Value *LHS, const Value *RHS, 334 const Instruction &CxtI) const { 335 return computeOverflowForUnsignedSub(LHS, RHS, &CxtI) == 336 OverflowResult::NeverOverflows; 337 } 338 339 bool willNotOverflowSub(const Value *LHS, const Value *RHS, 340 const Instruction &CxtI, bool IsSigned) const { 341 return IsSigned ? willNotOverflowSignedSub(LHS, RHS, CxtI) 342 : willNotOverflowUnsignedSub(LHS, RHS, CxtI); 343 } 344 345 bool willNotOverflowSignedMul(const Value *LHS, const Value *RHS, 346 const Instruction &CxtI) const { 347 return computeOverflowForSignedMul(LHS, RHS, &CxtI) == 348 OverflowResult::NeverOverflows; 349 } 350 351 bool willNotOverflowUnsignedMul(const Value *LHS, const Value *RHS, 352 const Instruction &CxtI) const { 353 return computeOverflowForUnsignedMul(LHS, RHS, &CxtI) == 354 OverflowResult::NeverOverflows; 355 } 356 357 bool willNotOverflowMul(const Value *LHS, const Value *RHS, 358 const Instruction &CxtI, bool IsSigned) const { 359 return IsSigned ? willNotOverflowSignedMul(LHS, RHS, CxtI) 360 : willNotOverflowUnsignedMul(LHS, RHS, CxtI); 361 } 362 363 bool willNotOverflow(BinaryOperator::BinaryOps Opcode, const Value *LHS, 364 const Value *RHS, const Instruction &CxtI, 365 bool IsSigned) const { 366 switch (Opcode) { 367 case Instruction::Add: return willNotOverflowAdd(LHS, RHS, CxtI, IsSigned); 368 case Instruction::Sub: return willNotOverflowSub(LHS, RHS, CxtI, IsSigned); 369 case Instruction::Mul: return willNotOverflowMul(LHS, RHS, CxtI, IsSigned); 370 default: llvm_unreachable("Unexpected opcode for overflow query"); 371 } 372 } 373 374 Value *EmitGEPOffset(User *GEP); 375 Instruction *scalarizePHI(ExtractElementInst &EI, PHINode *PN); 376 Instruction *foldBitcastExtElt(ExtractElementInst &ExtElt); 377 Instruction *foldCastedBitwiseLogic(BinaryOperator &I); 378 Instruction *foldBinopOfSextBoolToSelect(BinaryOperator &I); 379 Instruction *narrowBinOp(TruncInst &Trunc); 380 Instruction *narrowMaskedBinOp(BinaryOperator &And); 381 Instruction *narrowMathIfNoOverflow(BinaryOperator &I); 382 Instruction *narrowFunnelShift(TruncInst &Trunc); 383 Instruction *optimizeBitCastFromPhi(CastInst &CI, PHINode *PN); 384 Instruction *matchSAddSubSat(IntrinsicInst &MinMax1); 385 Instruction *foldNot(BinaryOperator &I); 386 Instruction *foldBinOpOfDisplacedShifts(BinaryOperator &I); 387 388 /// Determine if a pair of casts can be replaced by a single cast. 389 /// 390 /// \param CI1 The first of a pair of casts. 391 /// \param CI2 The second of a pair of casts. 392 /// 393 /// \return 0 if the cast pair cannot be eliminated, otherwise returns an 394 /// Instruction::CastOps value for a cast that can replace the pair, casting 395 /// CI1->getSrcTy() to CI2->getDstTy(). 396 /// 397 /// \see CastInst::isEliminableCastPair 398 Instruction::CastOps isEliminableCastPair(const CastInst *CI1, 399 const CastInst *CI2); 400 Value *simplifyIntToPtrRoundTripCast(Value *Val); 401 402 Value *foldAndOrOfICmps(ICmpInst *LHS, ICmpInst *RHS, Instruction &I, 403 bool IsAnd, bool IsLogical = false); 404 Value *foldXorOfICmps(ICmpInst *LHS, ICmpInst *RHS, BinaryOperator &Xor); 405 406 Value *foldEqOfParts(ICmpInst *Cmp0, ICmpInst *Cmp1, bool IsAnd); 407 408 Value *foldAndOrOfICmpsUsingRanges(ICmpInst *ICmp1, ICmpInst *ICmp2, 409 bool IsAnd); 410 411 /// Optimize (fcmp)&(fcmp) or (fcmp)|(fcmp). 412 /// NOTE: Unlike most of instcombine, this returns a Value which should 413 /// already be inserted into the function. 414 Value *foldLogicOfFCmps(FCmpInst *LHS, FCmpInst *RHS, bool IsAnd, 415 bool IsLogicalSelect = false); 416 417 Instruction *foldLogicOfIsFPClass(BinaryOperator &Operator, Value *LHS, 418 Value *RHS); 419 420 Instruction * 421 canonicalizeConditionalNegationViaMathToSelect(BinaryOperator &i); 422 423 Value *foldAndOrOfICmpsOfAndWithPow2(ICmpInst *LHS, ICmpInst *RHS, 424 Instruction *CxtI, bool IsAnd, 425 bool IsLogical = false); 426 Value *matchSelectFromAndOr(Value *A, Value *B, Value *C, Value *D, 427 bool InvertFalseVal = false); 428 Value *getSelectCondition(Value *A, Value *B, bool ABIsTheSame); 429 430 Instruction *foldLShrOverflowBit(BinaryOperator &I); 431 Instruction *foldExtractOfOverflowIntrinsic(ExtractValueInst &EV); 432 Instruction *foldIntrinsicWithOverflowCommon(IntrinsicInst *II); 433 Instruction *foldIntrinsicIsFPClass(IntrinsicInst &II); 434 Instruction *foldFPSignBitOps(BinaryOperator &I); 435 Instruction *foldFDivConstantDivisor(BinaryOperator &I); 436 437 // Optimize one of these forms: 438 // and i1 Op, SI / select i1 Op, i1 SI, i1 false (if IsAnd = true) 439 // or i1 Op, SI / select i1 Op, i1 true, i1 SI (if IsAnd = false) 440 // into simplier select instruction using isImpliedCondition. 441 Instruction *foldAndOrOfSelectUsingImpliedCond(Value *Op, SelectInst &SI, 442 bool IsAnd); 443 444 Instruction *hoistFNegAboveFMulFDiv(Value *FNegOp, Instruction &FMFSource); 445 446 public: 447 /// Create and insert the idiom we use to indicate a block is unreachable 448 /// without having to rewrite the CFG from within InstCombine. 449 void CreateNonTerminatorUnreachable(Instruction *InsertAt) { 450 auto &Ctx = InsertAt->getContext(); 451 auto *SI = new StoreInst(ConstantInt::getTrue(Ctx), 452 PoisonValue::get(PointerType::getUnqual(Ctx)), 453 /*isVolatile*/ false, Align(1)); 454 InsertNewInstBefore(SI, InsertAt->getIterator()); 455 } 456 457 /// Combiner aware instruction erasure. 458 /// 459 /// When dealing with an instruction that has side effects or produces a void 460 /// value, we can't rely on DCE to delete the instruction. Instead, visit 461 /// methods should return the value returned by this function. 462 Instruction *eraseInstFromFunction(Instruction &I) override { 463 LLVM_DEBUG(dbgs() << "IC: ERASE " << I << '\n'); 464 assert(I.use_empty() && "Cannot erase instruction that is used!"); 465 salvageDebugInfo(I); 466 467 // Make sure that we reprocess all operands now that we reduced their 468 // use counts. 469 SmallVector<Value *> Ops(I.operands()); 470 Worklist.remove(&I); 471 DC.removeValue(&I); 472 I.eraseFromParent(); 473 for (Value *Op : Ops) 474 Worklist.handleUseCountDecrement(Op); 475 MadeIRChange = true; 476 return nullptr; // Don't do anything with FI 477 } 478 479 OverflowResult computeOverflow( 480 Instruction::BinaryOps BinaryOp, bool IsSigned, 481 Value *LHS, Value *RHS, Instruction *CxtI) const; 482 483 /// Performs a few simplifications for operators which are associative 484 /// or commutative. 485 bool SimplifyAssociativeOrCommutative(BinaryOperator &I); 486 487 /// Tries to simplify binary operations which some other binary 488 /// operation distributes over. 489 /// 490 /// It does this by either by factorizing out common terms (eg "(A*B)+(A*C)" 491 /// -> "A*(B+C)") or expanding out if this results in simplifications (eg: "A 492 /// & (B | C) -> (A&B) | (A&C)" if this is a win). Returns the simplified 493 /// value, or null if it didn't simplify. 494 Value *foldUsingDistributiveLaws(BinaryOperator &I); 495 496 /// Tries to simplify add operations using the definition of remainder. 497 /// 498 /// The definition of remainder is X % C = X - (X / C ) * C. The add 499 /// expression X % C0 + (( X / C0 ) % C1) * C0 can be simplified to 500 /// X % (C0 * C1) 501 Value *SimplifyAddWithRemainder(BinaryOperator &I); 502 503 // Binary Op helper for select operations where the expression can be 504 // efficiently reorganized. 505 Value *SimplifySelectsFeedingBinaryOp(BinaryOperator &I, Value *LHS, 506 Value *RHS); 507 508 // If `I` has operand `(ctpop (not x))`, fold `I` with `(sub nuw nsw 509 // BitWidth(x), (ctpop x))`. 510 Instruction *tryFoldInstWithCtpopWithNot(Instruction *I); 511 512 // (Binop1 (Binop2 (logic_shift X, C), C1), (logic_shift Y, C)) 513 // -> (logic_shift (Binop1 (Binop2 X, inv_logic_shift(C1, C)), Y), C) 514 // (Binop1 (Binop2 (logic_shift X, Amt), Mask), (logic_shift Y, Amt)) 515 // -> (BinOp (logic_shift (BinOp X, Y)), Mask) 516 Instruction *foldBinOpShiftWithShift(BinaryOperator &I); 517 518 /// Tries to simplify binops of select and cast of the select condition. 519 /// 520 /// (Binop (cast C), (select C, T, F)) 521 /// -> (select C, C0, C1) 522 Instruction *foldBinOpOfSelectAndCastOfSelectCondition(BinaryOperator &I); 523 524 /// This tries to simplify binary operations by factorizing out common terms 525 /// (e. g. "(A*B)+(A*C)" -> "A*(B+C)"). 526 Value *tryFactorizationFolds(BinaryOperator &I); 527 528 /// Match a select chain which produces one of three values based on whether 529 /// the LHS is less than, equal to, or greater than RHS respectively. 530 /// Return true if we matched a three way compare idiom. The LHS, RHS, Less, 531 /// Equal and Greater values are saved in the matching process and returned to 532 /// the caller. 533 bool matchThreeWayIntCompare(SelectInst *SI, Value *&LHS, Value *&RHS, 534 ConstantInt *&Less, ConstantInt *&Equal, 535 ConstantInt *&Greater); 536 537 /// Attempts to replace V with a simpler value based on the demanded 538 /// bits. 539 Value *SimplifyDemandedUseBits(Value *V, APInt DemandedMask, KnownBits &Known, 540 unsigned Depth, Instruction *CxtI); 541 bool SimplifyDemandedBits(Instruction *I, unsigned Op, 542 const APInt &DemandedMask, KnownBits &Known, 543 unsigned Depth = 0) override; 544 545 /// Helper routine of SimplifyDemandedUseBits. It computes KnownZero/KnownOne 546 /// bits. It also tries to handle simplifications that can be done based on 547 /// DemandedMask, but without modifying the Instruction. 548 Value *SimplifyMultipleUseDemandedBits(Instruction *I, 549 const APInt &DemandedMask, 550 KnownBits &Known, 551 unsigned Depth, Instruction *CxtI); 552 553 /// Helper routine of SimplifyDemandedUseBits. It tries to simplify demanded 554 /// bit for "r1 = shr x, c1; r2 = shl r1, c2" instruction sequence. 555 Value *simplifyShrShlDemandedBits( 556 Instruction *Shr, const APInt &ShrOp1, Instruction *Shl, 557 const APInt &ShlOp1, const APInt &DemandedMask, KnownBits &Known); 558 559 /// Tries to simplify operands to an integer instruction based on its 560 /// demanded bits. 561 bool SimplifyDemandedInstructionBits(Instruction &Inst); 562 bool SimplifyDemandedInstructionBits(Instruction &Inst, KnownBits &Known); 563 564 Value *SimplifyDemandedVectorElts(Value *V, APInt DemandedElts, 565 APInt &PoisonElts, unsigned Depth = 0, 566 bool AllowMultipleUsers = false) override; 567 568 /// Canonicalize the position of binops relative to shufflevector. 569 Instruction *foldVectorBinop(BinaryOperator &Inst); 570 Instruction *foldVectorSelect(SelectInst &Sel); 571 Instruction *foldSelectShuffle(ShuffleVectorInst &Shuf); 572 573 /// Given a binary operator, cast instruction, or select which has a PHI node 574 /// as operand #0, see if we can fold the instruction into the PHI (which is 575 /// only possible if all operands to the PHI are constants). 576 Instruction *foldOpIntoPhi(Instruction &I, PHINode *PN); 577 578 /// For a binary operator with 2 phi operands, try to hoist the binary 579 /// operation before the phi. This can result in fewer instructions in 580 /// patterns where at least one set of phi operands simplifies. 581 /// Example: 582 /// BB3: binop (phi [X, BB1], [C1, BB2]), (phi [Y, BB1], [C2, BB2]) 583 /// --> 584 /// BB1: BO = binop X, Y 585 /// BB3: phi [BO, BB1], [(binop C1, C2), BB2] 586 Instruction *foldBinopWithPhiOperands(BinaryOperator &BO); 587 588 /// Given an instruction with a select as one operand and a constant as the 589 /// other operand, try to fold the binary operator into the select arguments. 590 /// This also works for Cast instructions, which obviously do not have a 591 /// second operand. 592 Instruction *FoldOpIntoSelect(Instruction &Op, SelectInst *SI, 593 bool FoldWithMultiUse = false); 594 595 /// This is a convenience wrapper function for the above two functions. 596 Instruction *foldBinOpIntoSelectOrPhi(BinaryOperator &I); 597 598 Instruction *foldAddWithConstant(BinaryOperator &Add); 599 600 Instruction *foldSquareSumInt(BinaryOperator &I); 601 Instruction *foldSquareSumFP(BinaryOperator &I); 602 603 /// Try to rotate an operation below a PHI node, using PHI nodes for 604 /// its operands. 605 Instruction *foldPHIArgOpIntoPHI(PHINode &PN); 606 Instruction *foldPHIArgBinOpIntoPHI(PHINode &PN); 607 Instruction *foldPHIArgInsertValueInstructionIntoPHI(PHINode &PN); 608 Instruction *foldPHIArgExtractValueInstructionIntoPHI(PHINode &PN); 609 Instruction *foldPHIArgGEPIntoPHI(PHINode &PN); 610 Instruction *foldPHIArgLoadIntoPHI(PHINode &PN); 611 Instruction *foldPHIArgZextsIntoPHI(PHINode &PN); 612 Instruction *foldPHIArgIntToPtrToPHI(PHINode &PN); 613 614 /// If an integer typed PHI has only one use which is an IntToPtr operation, 615 /// replace the PHI with an existing pointer typed PHI if it exists. Otherwise 616 /// insert a new pointer typed PHI and replace the original one. 617 bool foldIntegerTypedPHI(PHINode &PN); 618 619 /// Helper function for FoldPHIArgXIntoPHI() to set debug location for the 620 /// folded operation. 621 void PHIArgMergedDebugLoc(Instruction *Inst, PHINode &PN); 622 623 Instruction *foldGEPICmp(GEPOperator *GEPLHS, Value *RHS, 624 ICmpInst::Predicate Cond, Instruction &I); 625 Instruction *foldSelectICmp(ICmpInst::Predicate Pred, SelectInst *SI, 626 Value *RHS, const ICmpInst &I); 627 bool foldAllocaCmp(AllocaInst *Alloca); 628 Instruction *foldCmpLoadFromIndexedGlobal(LoadInst *LI, 629 GetElementPtrInst *GEP, 630 GlobalVariable *GV, CmpInst &ICI, 631 ConstantInt *AndCst = nullptr); 632 Instruction *foldFCmpIntToFPConst(FCmpInst &I, Instruction *LHSI, 633 Constant *RHSC); 634 Instruction *foldICmpAddOpConst(Value *X, const APInt &C, 635 ICmpInst::Predicate Pred); 636 Instruction *foldICmpWithCastOp(ICmpInst &ICmp); 637 Instruction *foldICmpWithZextOrSext(ICmpInst &ICmp); 638 639 Instruction *foldICmpUsingKnownBits(ICmpInst &Cmp); 640 Instruction *foldICmpWithDominatingICmp(ICmpInst &Cmp); 641 Instruction *foldICmpWithConstant(ICmpInst &Cmp); 642 Instruction *foldICmpUsingBoolRange(ICmpInst &I); 643 Instruction *foldICmpInstWithConstant(ICmpInst &Cmp); 644 Instruction *foldICmpInstWithConstantNotInt(ICmpInst &Cmp); 645 Instruction *foldICmpInstWithConstantAllowUndef(ICmpInst &Cmp, 646 const APInt &C); 647 Instruction *foldICmpBinOp(ICmpInst &Cmp, const SimplifyQuery &SQ); 648 Instruction *foldICmpWithMinMax(Instruction &I, MinMaxIntrinsic *MinMax, 649 Value *Z, ICmpInst::Predicate Pred); 650 Instruction *foldICmpEquality(ICmpInst &Cmp); 651 Instruction *foldIRemByPowerOfTwoToBitTest(ICmpInst &I); 652 Instruction *foldSignBitTest(ICmpInst &I); 653 Instruction *foldICmpWithZero(ICmpInst &Cmp); 654 655 Value *foldMultiplicationOverflowCheck(ICmpInst &Cmp); 656 657 Instruction *foldICmpBinOpWithConstant(ICmpInst &Cmp, BinaryOperator *BO, 658 const APInt &C); 659 Instruction *foldICmpSelectConstant(ICmpInst &Cmp, SelectInst *Select, 660 ConstantInt *C); 661 Instruction *foldICmpTruncConstant(ICmpInst &Cmp, TruncInst *Trunc, 662 const APInt &C); 663 Instruction *foldICmpTruncWithTruncOrExt(ICmpInst &Cmp, 664 const SimplifyQuery &Q); 665 Instruction *foldICmpAndConstant(ICmpInst &Cmp, BinaryOperator *And, 666 const APInt &C); 667 Instruction *foldICmpXorConstant(ICmpInst &Cmp, BinaryOperator *Xor, 668 const APInt &C); 669 Instruction *foldICmpOrConstant(ICmpInst &Cmp, BinaryOperator *Or, 670 const APInt &C); 671 Instruction *foldICmpMulConstant(ICmpInst &Cmp, BinaryOperator *Mul, 672 const APInt &C); 673 Instruction *foldICmpShlConstant(ICmpInst &Cmp, BinaryOperator *Shl, 674 const APInt &C); 675 Instruction *foldICmpShrConstant(ICmpInst &Cmp, BinaryOperator *Shr, 676 const APInt &C); 677 Instruction *foldICmpSRemConstant(ICmpInst &Cmp, BinaryOperator *UDiv, 678 const APInt &C); 679 Instruction *foldICmpUDivConstant(ICmpInst &Cmp, BinaryOperator *UDiv, 680 const APInt &C); 681 Instruction *foldICmpDivConstant(ICmpInst &Cmp, BinaryOperator *Div, 682 const APInt &C); 683 Instruction *foldICmpSubConstant(ICmpInst &Cmp, BinaryOperator *Sub, 684 const APInt &C); 685 Instruction *foldICmpAddConstant(ICmpInst &Cmp, BinaryOperator *Add, 686 const APInt &C); 687 Instruction *foldICmpAndConstConst(ICmpInst &Cmp, BinaryOperator *And, 688 const APInt &C1); 689 Instruction *foldICmpAndShift(ICmpInst &Cmp, BinaryOperator *And, 690 const APInt &C1, const APInt &C2); 691 Instruction *foldICmpXorShiftConst(ICmpInst &Cmp, BinaryOperator *Xor, 692 const APInt &C); 693 Instruction *foldICmpShrConstConst(ICmpInst &I, Value *ShAmt, const APInt &C1, 694 const APInt &C2); 695 Instruction *foldICmpShlConstConst(ICmpInst &I, Value *ShAmt, const APInt &C1, 696 const APInt &C2); 697 698 Instruction *foldICmpBinOpEqualityWithConstant(ICmpInst &Cmp, 699 BinaryOperator *BO, 700 const APInt &C); 701 Instruction *foldICmpIntrinsicWithConstant(ICmpInst &ICI, IntrinsicInst *II, 702 const APInt &C); 703 Instruction *foldICmpEqIntrinsicWithConstant(ICmpInst &ICI, IntrinsicInst *II, 704 const APInt &C); 705 Instruction *foldICmpBitCast(ICmpInst &Cmp); 706 Instruction *foldICmpWithTrunc(ICmpInst &Cmp); 707 Instruction *foldICmpCommutative(ICmpInst::Predicate Pred, Value *Op0, 708 Value *Op1, ICmpInst &CxtI); 709 710 // Helpers of visitSelectInst(). 711 Instruction *foldSelectOfBools(SelectInst &SI); 712 Instruction *foldSelectExtConst(SelectInst &Sel); 713 Instruction *foldSelectOpOp(SelectInst &SI, Instruction *TI, Instruction *FI); 714 Instruction *foldSelectIntoOp(SelectInst &SI, Value *, Value *); 715 Instruction *foldSPFofSPF(Instruction *Inner, SelectPatternFlavor SPF1, 716 Value *A, Value *B, Instruction &Outer, 717 SelectPatternFlavor SPF2, Value *C); 718 Instruction *foldSelectInstWithICmp(SelectInst &SI, ICmpInst *ICI); 719 Instruction *foldSelectValueEquivalence(SelectInst &SI, ICmpInst &ICI); 720 bool replaceInInstruction(Value *V, Value *Old, Value *New, 721 unsigned Depth = 0); 722 723 Value *insertRangeTest(Value *V, const APInt &Lo, const APInt &Hi, 724 bool isSigned, bool Inside); 725 bool mergeStoreIntoSuccessor(StoreInst &SI); 726 727 /// Given an initial instruction, check to see if it is the root of a 728 /// bswap/bitreverse idiom. If so, return the equivalent bswap/bitreverse 729 /// intrinsic. 730 Instruction *matchBSwapOrBitReverse(Instruction &I, bool MatchBSwaps, 731 bool MatchBitReversals); 732 733 Instruction *SimplifyAnyMemTransfer(AnyMemTransferInst *MI); 734 Instruction *SimplifyAnyMemSet(AnyMemSetInst *MI); 735 736 Value *EvaluateInDifferentType(Value *V, Type *Ty, bool isSigned); 737 738 bool tryToSinkInstruction(Instruction *I, BasicBlock *DestBlock); 739 740 bool removeInstructionsBeforeUnreachable(Instruction &I); 741 void addDeadEdge(BasicBlock *From, BasicBlock *To, 742 SmallVectorImpl<BasicBlock *> &Worklist); 743 void handleUnreachableFrom(Instruction *I, 744 SmallVectorImpl<BasicBlock *> &Worklist); 745 void handlePotentiallyDeadBlocks(SmallVectorImpl<BasicBlock *> &Worklist); 746 void handlePotentiallyDeadSuccessors(BasicBlock *BB, BasicBlock *LiveSucc); 747 void freelyInvertAllUsersOf(Value *V, Value *IgnoredUser = nullptr); 748 }; 749 750 class Negator final { 751 /// Top-to-bottom, def-to-use negated instruction tree we produced. 752 SmallVector<Instruction *, NegatorMaxNodesSSO> NewInstructions; 753 754 using BuilderTy = IRBuilder<TargetFolder, IRBuilderCallbackInserter>; 755 BuilderTy Builder; 756 757 const bool IsTrulyNegation; 758 759 SmallDenseMap<Value *, Value *> NegationsCache; 760 761 Negator(LLVMContext &C, const DataLayout &DL, bool IsTrulyNegation); 762 763 #if LLVM_ENABLE_STATS 764 unsigned NumValuesVisitedInThisNegator = 0; 765 ~Negator(); 766 #endif 767 768 using Result = std::pair<ArrayRef<Instruction *> /*NewInstructions*/, 769 Value * /*NegatedRoot*/>; 770 771 std::array<Value *, 2> getSortedOperandsOfBinOp(Instruction *I); 772 773 [[nodiscard]] Value *visitImpl(Value *V, bool IsNSW, unsigned Depth); 774 775 [[nodiscard]] Value *negate(Value *V, bool IsNSW, unsigned Depth); 776 777 /// Recurse depth-first and attempt to sink the negation. 778 /// FIXME: use worklist? 779 [[nodiscard]] std::optional<Result> run(Value *Root, bool IsNSW); 780 781 Negator(const Negator &) = delete; 782 Negator(Negator &&) = delete; 783 Negator &operator=(const Negator &) = delete; 784 Negator &operator=(Negator &&) = delete; 785 786 public: 787 /// Attempt to negate \p Root. Retuns nullptr if negation can't be performed, 788 /// otherwise returns negated value. 789 [[nodiscard]] static Value *Negate(bool LHSIsZero, bool IsNSW, Value *Root, 790 InstCombinerImpl &IC); 791 }; 792 793 } // end namespace llvm 794 795 #undef DEBUG_TYPE 796 797 #endif // LLVM_LIB_TRANSFORMS_INSTCOMBINE_INSTCOMBINEINTERNAL_H 798