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/Analysis/InstructionSimplify.h" 20#include "llvm/Analysis/TargetFolder.h" 21#include "llvm/Analysis/ValueTracking.h" 22#include "llvm/IR/IRBuilder.h" 23#include "llvm/IR/InstVisitor.h" 24#include "llvm/IR/PatternMatch.h" 25#include "llvm/Support/Debug.h" 26#include "llvm/Support/KnownBits.h" 27#include "llvm/Transforms/InstCombine/InstCombineWorklist.h" 28#include "llvm/Transforms/InstCombine/InstCombiner.h" 29#include "llvm/Transforms/Utils/Local.h" 30#include <cassert> 31 32#define DEBUG_TYPE "instcombine" 33 34using namespace llvm::PatternMatch; 35 36// As a default, let's assume that we want to be aggressive, 37// and attempt to traverse with no limits in attempt to sink negation. 38static constexpr unsigned NegatorDefaultMaxDepth = ~0U; 39 40// Let's guesstimate that most often we will end up visiting/producing 41// fairly small number of new instructions. 42static constexpr unsigned NegatorMaxNodesSSO = 16; 43 44namespace llvm { 45 46class AAResults; 47class APInt; 48class AssumptionCache; 49class BlockFrequencyInfo; 50class DataLayout; 51class DominatorTree; 52class GEPOperator; 53class GlobalVariable; 54class LoopInfo; 55class OptimizationRemarkEmitter; 56class ProfileSummaryInfo; 57class TargetLibraryInfo; 58class User; 59 60class LLVM_LIBRARY_VISIBILITY InstCombinerImpl final 61 : public InstCombiner, 62 public InstVisitor<InstCombinerImpl, Instruction *> { 63public: 64 InstCombinerImpl(InstCombineWorklist &Worklist, BuilderTy &Builder, 65 bool MinimizeSize, AAResults *AA, AssumptionCache &AC, 66 TargetLibraryInfo &TLI, TargetTransformInfo &TTI, 67 DominatorTree &DT, OptimizationRemarkEmitter &ORE, 68 BlockFrequencyInfo *BFI, ProfileSummaryInfo *PSI, 69 const DataLayout &DL, LoopInfo *LI) 70 : InstCombiner(Worklist, Builder, MinimizeSize, AA, AC, TLI, TTI, DT, ORE, 71 BFI, PSI, DL, LI) {} 72 73 virtual ~InstCombinerImpl() {} 74 75 /// Run the combiner over the entire worklist until it is empty. 76 /// 77 /// \returns true if the IR is changed. 78 bool run(); 79 80 // Visitation implementation - Implement instruction combining for different 81 // instruction types. The semantics are as follows: 82 // Return Value: 83 // null - No change was made 84 // I - Change was made, I is still valid, I may be dead though 85 // otherwise - Change was made, replace I with returned instruction 86 // 87 Instruction *visitFNeg(UnaryOperator &I); 88 Instruction *visitAdd(BinaryOperator &I); 89 Instruction *visitFAdd(BinaryOperator &I); 90 Value *OptimizePointerDifference( 91 Value *LHS, Value *RHS, Type *Ty, bool isNUW); 92 Instruction *visitSub(BinaryOperator &I); 93 Instruction *visitFSub(BinaryOperator &I); 94 Instruction *visitMul(BinaryOperator &I); 95 Instruction *visitFMul(BinaryOperator &I); 96 Instruction *visitURem(BinaryOperator &I); 97 Instruction *visitSRem(BinaryOperator &I); 98 Instruction *visitFRem(BinaryOperator &I); 99 bool simplifyDivRemOfSelectWithZeroOp(BinaryOperator &I); 100 Instruction *commonIRemTransforms(BinaryOperator &I); 101 Instruction *commonIDivTransforms(BinaryOperator &I); 102 Instruction *visitUDiv(BinaryOperator &I); 103 Instruction *visitSDiv(BinaryOperator &I); 104 Instruction *visitFDiv(BinaryOperator &I); 105 Value *simplifyRangeCheck(ICmpInst *Cmp0, ICmpInst *Cmp1, bool Inverted); 106 Instruction *visitAnd(BinaryOperator &I); 107 Instruction *visitOr(BinaryOperator &I); 108 bool sinkNotIntoOtherHandOfAndOrOr(BinaryOperator &I); 109 Instruction *visitXor(BinaryOperator &I); 110 Instruction *visitShl(BinaryOperator &I); 111 Value *reassociateShiftAmtsOfTwoSameDirectionShifts( 112 BinaryOperator *Sh0, const SimplifyQuery &SQ, 113 bool AnalyzeForSignBitExtraction = false); 114 Instruction *canonicalizeCondSignextOfHighBitExtractToSignextHighBitExtract( 115 BinaryOperator &I); 116 Instruction *foldVariableSignZeroExtensionOfVariableHighBitExtract( 117 BinaryOperator &OldAShr); 118 Instruction *visitAShr(BinaryOperator &I); 119 Instruction *visitLShr(BinaryOperator &I); 120 Instruction *commonShiftTransforms(BinaryOperator &I); 121 Instruction *visitFCmpInst(FCmpInst &I); 122 CmpInst *canonicalizeICmpPredicate(CmpInst &I); 123 Instruction *visitICmpInst(ICmpInst &I); 124 Instruction *FoldShiftByConstant(Value *Op0, Constant *Op1, 125 BinaryOperator &I); 126 Instruction *commonCastTransforms(CastInst &CI); 127 Instruction *commonPointerCastTransforms(CastInst &CI); 128 Instruction *visitTrunc(TruncInst &CI); 129 Instruction *visitZExt(ZExtInst &CI); 130 Instruction *visitSExt(SExtInst &CI); 131 Instruction *visitFPTrunc(FPTruncInst &CI); 132 Instruction *visitFPExt(CastInst &CI); 133 Instruction *visitFPToUI(FPToUIInst &FI); 134 Instruction *visitFPToSI(FPToSIInst &FI); 135 Instruction *visitUIToFP(CastInst &CI); 136 Instruction *visitSIToFP(CastInst &CI); 137 Instruction *visitPtrToInt(PtrToIntInst &CI); 138 Instruction *visitIntToPtr(IntToPtrInst &CI); 139 Instruction *visitBitCast(BitCastInst &CI); 140 Instruction *visitAddrSpaceCast(AddrSpaceCastInst &CI); 141 Instruction *foldItoFPtoI(CastInst &FI); 142 Instruction *visitSelectInst(SelectInst &SI); 143 Instruction *visitCallInst(CallInst &CI); 144 Instruction *visitInvokeInst(InvokeInst &II); 145 Instruction *visitCallBrInst(CallBrInst &CBI); 146 147 Instruction *SliceUpIllegalIntegerPHI(PHINode &PN); 148 Instruction *visitPHINode(PHINode &PN); 149 Instruction *visitGetElementPtrInst(GetElementPtrInst &GEP); 150 Instruction *visitAllocaInst(AllocaInst &AI); 151 Instruction *visitAllocSite(Instruction &FI); 152 Instruction *visitFree(CallInst &FI); 153 Instruction *visitLoadInst(LoadInst &LI); 154 Instruction *visitStoreInst(StoreInst &SI); 155 Instruction *visitAtomicRMWInst(AtomicRMWInst &SI); 156 Instruction *visitUnconditionalBranchInst(BranchInst &BI); 157 Instruction *visitBranchInst(BranchInst &BI); 158 Instruction *visitFenceInst(FenceInst &FI); 159 Instruction *visitSwitchInst(SwitchInst &SI); 160 Instruction *visitReturnInst(ReturnInst &RI); 161 Instruction *visitUnreachableInst(UnreachableInst &I); 162 Instruction * 163 foldAggregateConstructionIntoAggregateReuse(InsertValueInst &OrigIVI); 164 Instruction *visitInsertValueInst(InsertValueInst &IV); 165 Instruction *visitInsertElementInst(InsertElementInst &IE); 166 Instruction *visitExtractElementInst(ExtractElementInst &EI); 167 Instruction *visitShuffleVectorInst(ShuffleVectorInst &SVI); 168 Instruction *visitExtractValueInst(ExtractValueInst &EV); 169 Instruction *visitLandingPadInst(LandingPadInst &LI); 170 Instruction *visitVAEndInst(VAEndInst &I); 171 Instruction *visitFreeze(FreezeInst &I); 172 173 /// Specify what to return for unhandled instructions. 174 Instruction *visitInstruction(Instruction &I) { return nullptr; } 175 176 /// True when DB dominates all uses of DI except UI. 177 /// UI must be in the same block as DI. 178 /// The routine checks that the DI parent and DB are different. 179 bool dominatesAllUses(const Instruction *DI, const Instruction *UI, 180 const BasicBlock *DB) const; 181 182 /// Try to replace select with select operand SIOpd in SI-ICmp sequence. 183 bool replacedSelectWithOperand(SelectInst *SI, const ICmpInst *Icmp, 184 const unsigned SIOpd); 185 186 LoadInst *combineLoadToNewType(LoadInst &LI, Type *NewTy, 187 const Twine &Suffix = ""); 188 189private: 190 void annotateAnyAllocSite(CallBase &Call, const TargetLibraryInfo *TLI); 191 bool shouldChangeType(unsigned FromBitWidth, unsigned ToBitWidth) const; 192 bool shouldChangeType(Type *From, Type *To) const; 193 Value *dyn_castNegVal(Value *V) const; 194 Type *FindElementAtOffset(PointerType *PtrTy, int64_t Offset, 195 SmallVectorImpl<Value *> &NewIndices); 196 197 /// Classify whether a cast is worth optimizing. 198 /// 199 /// This is a helper to decide whether the simplification of 200 /// logic(cast(A), cast(B)) to cast(logic(A, B)) should be performed. 201 /// 202 /// \param CI The cast we are interested in. 203 /// 204 /// \return true if this cast actually results in any code being generated and 205 /// if it cannot already be eliminated by some other transformation. 206 bool shouldOptimizeCast(CastInst *CI); 207 208 /// Try to optimize a sequence of instructions checking if an operation 209 /// on LHS and RHS overflows. 210 /// 211 /// If this overflow check is done via one of the overflow check intrinsics, 212 /// then CtxI has to be the call instruction calling that intrinsic. If this 213 /// overflow check is done by arithmetic followed by a compare, then CtxI has 214 /// to be the arithmetic instruction. 215 /// 216 /// If a simplification is possible, stores the simplified result of the 217 /// operation in OperationResult and result of the overflow check in 218 /// OverflowResult, and return true. If no simplification is possible, 219 /// returns false. 220 bool OptimizeOverflowCheck(Instruction::BinaryOps BinaryOp, bool IsSigned, 221 Value *LHS, Value *RHS, 222 Instruction &CtxI, Value *&OperationResult, 223 Constant *&OverflowResult); 224 225 Instruction *visitCallBase(CallBase &Call); 226 Instruction *tryOptimizeCall(CallInst *CI); 227 bool transformConstExprCastCall(CallBase &Call); 228 Instruction *transformCallThroughTrampoline(CallBase &Call, 229 IntrinsicInst &Tramp); 230 231 Value *simplifyMaskedLoad(IntrinsicInst &II); 232 Instruction *simplifyMaskedStore(IntrinsicInst &II); 233 Instruction *simplifyMaskedGather(IntrinsicInst &II); 234 Instruction *simplifyMaskedScatter(IntrinsicInst &II); 235 236 /// Transform (zext icmp) to bitwise / integer operations in order to 237 /// eliminate it. 238 /// 239 /// \param ICI The icmp of the (zext icmp) pair we are interested in. 240 /// \parem CI The zext of the (zext icmp) pair we are interested in. 241 /// \param DoTransform Pass false to just test whether the given (zext icmp) 242 /// would be transformed. Pass true to actually perform the transformation. 243 /// 244 /// \return null if the transformation cannot be performed. If the 245 /// transformation can be performed the new instruction that replaces the 246 /// (zext icmp) pair will be returned (if \p DoTransform is false the 247 /// unmodified \p ICI will be returned in this case). 248 Instruction *transformZExtICmp(ICmpInst *ICI, ZExtInst &CI, 249 bool DoTransform = true); 250 251 Instruction *transformSExtICmp(ICmpInst *ICI, Instruction &CI); 252 253 bool willNotOverflowSignedAdd(const Value *LHS, const Value *RHS, 254 const Instruction &CxtI) const { 255 return computeOverflowForSignedAdd(LHS, RHS, &CxtI) == 256 OverflowResult::NeverOverflows; 257 } 258 259 bool willNotOverflowUnsignedAdd(const Value *LHS, const Value *RHS, 260 const Instruction &CxtI) const { 261 return computeOverflowForUnsignedAdd(LHS, RHS, &CxtI) == 262 OverflowResult::NeverOverflows; 263 } 264 265 bool willNotOverflowAdd(const Value *LHS, const Value *RHS, 266 const Instruction &CxtI, bool IsSigned) const { 267 return IsSigned ? willNotOverflowSignedAdd(LHS, RHS, CxtI) 268 : willNotOverflowUnsignedAdd(LHS, RHS, CxtI); 269 } 270 271 bool willNotOverflowSignedSub(const Value *LHS, const Value *RHS, 272 const Instruction &CxtI) const { 273 return computeOverflowForSignedSub(LHS, RHS, &CxtI) == 274 OverflowResult::NeverOverflows; 275 } 276 277 bool willNotOverflowUnsignedSub(const Value *LHS, const Value *RHS, 278 const Instruction &CxtI) const { 279 return computeOverflowForUnsignedSub(LHS, RHS, &CxtI) == 280 OverflowResult::NeverOverflows; 281 } 282 283 bool willNotOverflowSub(const Value *LHS, const Value *RHS, 284 const Instruction &CxtI, bool IsSigned) const { 285 return IsSigned ? willNotOverflowSignedSub(LHS, RHS, CxtI) 286 : willNotOverflowUnsignedSub(LHS, RHS, CxtI); 287 } 288 289 bool willNotOverflowSignedMul(const Value *LHS, const Value *RHS, 290 const Instruction &CxtI) const { 291 return computeOverflowForSignedMul(LHS, RHS, &CxtI) == 292 OverflowResult::NeverOverflows; 293 } 294 295 bool willNotOverflowUnsignedMul(const Value *LHS, const Value *RHS, 296 const Instruction &CxtI) const { 297 return computeOverflowForUnsignedMul(LHS, RHS, &CxtI) == 298 OverflowResult::NeverOverflows; 299 } 300 301 bool willNotOverflowMul(const Value *LHS, const Value *RHS, 302 const Instruction &CxtI, bool IsSigned) const { 303 return IsSigned ? willNotOverflowSignedMul(LHS, RHS, CxtI) 304 : willNotOverflowUnsignedMul(LHS, RHS, CxtI); 305 } 306 307 bool willNotOverflow(BinaryOperator::BinaryOps Opcode, const Value *LHS, 308 const Value *RHS, const Instruction &CxtI, 309 bool IsSigned) const { 310 switch (Opcode) { 311 case Instruction::Add: return willNotOverflowAdd(LHS, RHS, CxtI, IsSigned); 312 case Instruction::Sub: return willNotOverflowSub(LHS, RHS, CxtI, IsSigned); 313 case Instruction::Mul: return willNotOverflowMul(LHS, RHS, CxtI, IsSigned); 314 default: llvm_unreachable("Unexpected opcode for overflow query"); 315 } 316 } 317 318 Value *EmitGEPOffset(User *GEP); 319 Instruction *scalarizePHI(ExtractElementInst &EI, PHINode *PN); 320 Instruction *foldCastedBitwiseLogic(BinaryOperator &I); 321 Instruction *narrowBinOp(TruncInst &Trunc); 322 Instruction *narrowMaskedBinOp(BinaryOperator &And); 323 Instruction *narrowMathIfNoOverflow(BinaryOperator &I); 324 Instruction *narrowFunnelShift(TruncInst &Trunc); 325 Instruction *optimizeBitCastFromPhi(CastInst &CI, PHINode *PN); 326 Instruction *matchSAddSubSat(SelectInst &MinMax1); 327 328 void freelyInvertAllUsersOf(Value *V); 329 330 /// Determine if a pair of casts can be replaced by a single cast. 331 /// 332 /// \param CI1 The first of a pair of casts. 333 /// \param CI2 The second of a pair of casts. 334 /// 335 /// \return 0 if the cast pair cannot be eliminated, otherwise returns an 336 /// Instruction::CastOps value for a cast that can replace the pair, casting 337 /// CI1->getSrcTy() to CI2->getDstTy(). 338 /// 339 /// \see CastInst::isEliminableCastPair 340 Instruction::CastOps isEliminableCastPair(const CastInst *CI1, 341 const CastInst *CI2); 342 343 Value *foldAndOfICmps(ICmpInst *LHS, ICmpInst *RHS, BinaryOperator &And); 344 Value *foldOrOfICmps(ICmpInst *LHS, ICmpInst *RHS, BinaryOperator &Or); 345 Value *foldXorOfICmps(ICmpInst *LHS, ICmpInst *RHS, BinaryOperator &Xor); 346 347 /// Optimize (fcmp)&(fcmp) or (fcmp)|(fcmp). 348 /// NOTE: Unlike most of instcombine, this returns a Value which should 349 /// already be inserted into the function. 350 Value *foldLogicOfFCmps(FCmpInst *LHS, FCmpInst *RHS, bool IsAnd); 351 352 Value *foldAndOrOfICmpsOfAndWithPow2(ICmpInst *LHS, ICmpInst *RHS, 353 Instruction *CxtI, bool IsAnd, 354 bool IsLogical = false); 355 Value *matchSelectFromAndOr(Value *A, Value *B, Value *C, Value *D); 356 Value *getSelectCondition(Value *A, Value *B); 357 358 Instruction *foldIntrinsicWithOverflowCommon(IntrinsicInst *II); 359 Instruction *foldFPSignBitOps(BinaryOperator &I); 360 361 // Optimize one of these forms: 362 // and i1 Op, SI / select i1 Op, i1 SI, i1 false (if IsAnd = true) 363 // or i1 Op, SI / select i1 Op, i1 true, i1 SI (if IsAnd = false) 364 // into simplier select instruction using isImpliedCondition. 365 Instruction *foldAndOrOfSelectUsingImpliedCond(Value *Op, SelectInst &SI, 366 bool IsAnd); 367 368public: 369 /// Inserts an instruction \p New before instruction \p Old 370 /// 371 /// Also adds the new instruction to the worklist and returns \p New so that 372 /// it is suitable for use as the return from the visitation patterns. 373 Instruction *InsertNewInstBefore(Instruction *New, Instruction &Old) { 374 assert(New && !New->getParent() && 375 "New instruction already inserted into a basic block!"); 376 BasicBlock *BB = Old.getParent(); 377 BB->getInstList().insert(Old.getIterator(), New); // Insert inst 378 Worklist.add(New); 379 return New; 380 } 381 382 /// Same as InsertNewInstBefore, but also sets the debug loc. 383 Instruction *InsertNewInstWith(Instruction *New, Instruction &Old) { 384 New->setDebugLoc(Old.getDebugLoc()); 385 return InsertNewInstBefore(New, Old); 386 } 387 388 /// A combiner-aware RAUW-like routine. 389 /// 390 /// This method is to be used when an instruction is found to be dead, 391 /// replaceable with another preexisting expression. Here we add all uses of 392 /// I to the worklist, replace all uses of I with the new value, then return 393 /// I, so that the inst combiner will know that I was modified. 394 Instruction *replaceInstUsesWith(Instruction &I, Value *V) { 395 // If there are no uses to replace, then we return nullptr to indicate that 396 // no changes were made to the program. 397 if (I.use_empty()) return nullptr; 398 399 Worklist.pushUsersToWorkList(I); // Add all modified instrs to worklist. 400 401 // If we are replacing the instruction with itself, this must be in a 402 // segment of unreachable code, so just clobber the instruction. 403 if (&I == V) 404 V = UndefValue::get(I.getType()); 405 406 LLVM_DEBUG(dbgs() << "IC: Replacing " << I << "\n" 407 << " with " << *V << '\n'); 408 409 I.replaceAllUsesWith(V); 410 MadeIRChange = true; 411 return &I; 412 } 413 414 /// Replace operand of instruction and add old operand to the worklist. 415 Instruction *replaceOperand(Instruction &I, unsigned OpNum, Value *V) { 416 Worklist.addValue(I.getOperand(OpNum)); 417 I.setOperand(OpNum, V); 418 return &I; 419 } 420 421 /// Replace use and add the previously used value to the worklist. 422 void replaceUse(Use &U, Value *NewValue) { 423 Worklist.addValue(U); 424 U = NewValue; 425 } 426 427 /// Creates a result tuple for an overflow intrinsic \p II with a given 428 /// \p Result and a constant \p Overflow value. 429 Instruction *CreateOverflowTuple(IntrinsicInst *II, Value *Result, 430 Constant *Overflow) { 431 Constant *V[] = {UndefValue::get(Result->getType()), Overflow}; 432 StructType *ST = cast<StructType>(II->getType()); 433 Constant *Struct = ConstantStruct::get(ST, V); 434 return InsertValueInst::Create(Struct, Result, 0); 435 } 436 437 /// Create and insert the idiom we use to indicate a block is unreachable 438 /// without having to rewrite the CFG from within InstCombine. 439 void CreateNonTerminatorUnreachable(Instruction *InsertAt) { 440 auto &Ctx = InsertAt->getContext(); 441 new StoreInst(ConstantInt::getTrue(Ctx), 442 UndefValue::get(Type::getInt1PtrTy(Ctx)), 443 InsertAt); 444 } 445 446 447 /// Combiner aware instruction erasure. 448 /// 449 /// When dealing with an instruction that has side effects or produces a void 450 /// value, we can't rely on DCE to delete the instruction. Instead, visit 451 /// methods should return the value returned by this function. 452 Instruction *eraseInstFromFunction(Instruction &I) override { 453 LLVM_DEBUG(dbgs() << "IC: ERASE " << I << '\n'); 454 assert(I.use_empty() && "Cannot erase instruction that is used!"); 455 salvageDebugInfo(I); 456 457 // Make sure that we reprocess all operands now that we reduced their 458 // use counts. 459 for (Use &Operand : I.operands()) 460 if (auto *Inst = dyn_cast<Instruction>(Operand)) 461 Worklist.add(Inst); 462 463 Worklist.remove(&I); 464 I.eraseFromParent(); 465 MadeIRChange = true; 466 return nullptr; // Don't do anything with FI 467 } 468 469 void computeKnownBits(const Value *V, KnownBits &Known, 470 unsigned Depth, const Instruction *CxtI) const { 471 llvm::computeKnownBits(V, Known, DL, Depth, &AC, CxtI, &DT); 472 } 473 474 KnownBits computeKnownBits(const Value *V, unsigned Depth, 475 const Instruction *CxtI) const { 476 return llvm::computeKnownBits(V, DL, Depth, &AC, CxtI, &DT); 477 } 478 479 bool isKnownToBeAPowerOfTwo(const Value *V, bool OrZero = false, 480 unsigned Depth = 0, 481 const Instruction *CxtI = nullptr) { 482 return llvm::isKnownToBeAPowerOfTwo(V, DL, OrZero, Depth, &AC, CxtI, &DT); 483 } 484 485 bool MaskedValueIsZero(const Value *V, const APInt &Mask, unsigned Depth = 0, 486 const Instruction *CxtI = nullptr) const { 487 return llvm::MaskedValueIsZero(V, Mask, DL, Depth, &AC, CxtI, &DT); 488 } 489 490 unsigned ComputeNumSignBits(const Value *Op, unsigned Depth = 0, 491 const Instruction *CxtI = nullptr) const { 492 return llvm::ComputeNumSignBits(Op, DL, Depth, &AC, CxtI, &DT); 493 } 494 495 OverflowResult computeOverflowForUnsignedMul(const Value *LHS, 496 const Value *RHS, 497 const Instruction *CxtI) const { 498 return llvm::computeOverflowForUnsignedMul(LHS, RHS, DL, &AC, CxtI, &DT); 499 } 500 501 OverflowResult computeOverflowForSignedMul(const Value *LHS, 502 const Value *RHS, 503 const Instruction *CxtI) const { 504 return llvm::computeOverflowForSignedMul(LHS, RHS, DL, &AC, CxtI, &DT); 505 } 506 507 OverflowResult computeOverflowForUnsignedAdd(const Value *LHS, 508 const Value *RHS, 509 const Instruction *CxtI) const { 510 return llvm::computeOverflowForUnsignedAdd(LHS, RHS, DL, &AC, CxtI, &DT); 511 } 512 513 OverflowResult computeOverflowForSignedAdd(const Value *LHS, 514 const Value *RHS, 515 const Instruction *CxtI) const { 516 return llvm::computeOverflowForSignedAdd(LHS, RHS, DL, &AC, CxtI, &DT); 517 } 518 519 OverflowResult computeOverflowForUnsignedSub(const Value *LHS, 520 const Value *RHS, 521 const Instruction *CxtI) const { 522 return llvm::computeOverflowForUnsignedSub(LHS, RHS, DL, &AC, CxtI, &DT); 523 } 524 525 OverflowResult computeOverflowForSignedSub(const Value *LHS, const Value *RHS, 526 const Instruction *CxtI) const { 527 return llvm::computeOverflowForSignedSub(LHS, RHS, DL, &AC, CxtI, &DT); 528 } 529 530 OverflowResult computeOverflow( 531 Instruction::BinaryOps BinaryOp, bool IsSigned, 532 Value *LHS, Value *RHS, Instruction *CxtI) const; 533 534 /// Performs a few simplifications for operators which are associative 535 /// or commutative. 536 bool SimplifyAssociativeOrCommutative(BinaryOperator &I); 537 538 /// Tries to simplify binary operations which some other binary 539 /// operation distributes over. 540 /// 541 /// It does this by either by factorizing out common terms (eg "(A*B)+(A*C)" 542 /// -> "A*(B+C)") or expanding out if this results in simplifications (eg: "A 543 /// & (B | C) -> (A&B) | (A&C)" if this is a win). Returns the simplified 544 /// value, or null if it didn't simplify. 545 Value *SimplifyUsingDistributiveLaws(BinaryOperator &I); 546 547 /// Tries to simplify add operations using the definition of remainder. 548 /// 549 /// The definition of remainder is X % C = X - (X / C ) * C. The add 550 /// expression X % C0 + (( X / C0 ) % C1) * C0 can be simplified to 551 /// X % (C0 * C1) 552 Value *SimplifyAddWithRemainder(BinaryOperator &I); 553 554 // Binary Op helper for select operations where the expression can be 555 // efficiently reorganized. 556 Value *SimplifySelectsFeedingBinaryOp(BinaryOperator &I, Value *LHS, 557 Value *RHS); 558 559 /// This tries to simplify binary operations by factorizing out common terms 560 /// (e. g. "(A*B)+(A*C)" -> "A*(B+C)"). 561 Value *tryFactorization(BinaryOperator &, Instruction::BinaryOps, Value *, 562 Value *, Value *, Value *); 563 564 /// Match a select chain which produces one of three values based on whether 565 /// the LHS is less than, equal to, or greater than RHS respectively. 566 /// Return true if we matched a three way compare idiom. The LHS, RHS, Less, 567 /// Equal and Greater values are saved in the matching process and returned to 568 /// the caller. 569 bool matchThreeWayIntCompare(SelectInst *SI, Value *&LHS, Value *&RHS, 570 ConstantInt *&Less, ConstantInt *&Equal, 571 ConstantInt *&Greater); 572 573 /// Attempts to replace V with a simpler value based on the demanded 574 /// bits. 575 Value *SimplifyDemandedUseBits(Value *V, APInt DemandedMask, KnownBits &Known, 576 unsigned Depth, Instruction *CxtI); 577 bool SimplifyDemandedBits(Instruction *I, unsigned Op, 578 const APInt &DemandedMask, KnownBits &Known, 579 unsigned Depth = 0) override; 580 581 /// Helper routine of SimplifyDemandedUseBits. It computes KnownZero/KnownOne 582 /// bits. It also tries to handle simplifications that can be done based on 583 /// DemandedMask, but without modifying the Instruction. 584 Value *SimplifyMultipleUseDemandedBits(Instruction *I, 585 const APInt &DemandedMask, 586 KnownBits &Known, 587 unsigned Depth, Instruction *CxtI); 588 589 /// Helper routine of SimplifyDemandedUseBits. It tries to simplify demanded 590 /// bit for "r1 = shr x, c1; r2 = shl r1, c2" instruction sequence. 591 Value *simplifyShrShlDemandedBits( 592 Instruction *Shr, const APInt &ShrOp1, Instruction *Shl, 593 const APInt &ShlOp1, const APInt &DemandedMask, KnownBits &Known); 594 595 /// Tries to simplify operands to an integer instruction based on its 596 /// demanded bits. 597 bool SimplifyDemandedInstructionBits(Instruction &Inst); 598 599 virtual Value * 600 SimplifyDemandedVectorElts(Value *V, APInt DemandedElts, APInt &UndefElts, 601 unsigned Depth = 0, 602 bool AllowMultipleUsers = false) override; 603 604 /// Canonicalize the position of binops relative to shufflevector. 605 Instruction *foldVectorBinop(BinaryOperator &Inst); 606 Instruction *foldVectorSelect(SelectInst &Sel); 607 608 /// Given a binary operator, cast instruction, or select which has a PHI node 609 /// as operand #0, see if we can fold the instruction into the PHI (which is 610 /// only possible if all operands to the PHI are constants). 611 Instruction *foldOpIntoPhi(Instruction &I, PHINode *PN); 612 613 /// Given an instruction with a select as one operand and a constant as the 614 /// other operand, try to fold the binary operator into the select arguments. 615 /// This also works for Cast instructions, which obviously do not have a 616 /// second operand. 617 Instruction *FoldOpIntoSelect(Instruction &Op, SelectInst *SI); 618 619 /// This is a convenience wrapper function for the above two functions. 620 Instruction *foldBinOpIntoSelectOrPhi(BinaryOperator &I); 621 622 Instruction *foldAddWithConstant(BinaryOperator &Add); 623 624 /// Try to rotate an operation below a PHI node, using PHI nodes for 625 /// its operands. 626 Instruction *foldPHIArgOpIntoPHI(PHINode &PN); 627 Instruction *foldPHIArgBinOpIntoPHI(PHINode &PN); 628 Instruction *foldPHIArgInsertValueInstructionIntoPHI(PHINode &PN); 629 Instruction *foldPHIArgExtractValueInstructionIntoPHI(PHINode &PN); 630 Instruction *foldPHIArgGEPIntoPHI(PHINode &PN); 631 Instruction *foldPHIArgLoadIntoPHI(PHINode &PN); 632 Instruction *foldPHIArgZextsIntoPHI(PHINode &PN); 633 634 /// If an integer typed PHI has only one use which is an IntToPtr operation, 635 /// replace the PHI with an existing pointer typed PHI if it exists. Otherwise 636 /// insert a new pointer typed PHI and replace the original one. 637 Instruction *foldIntegerTypedPHI(PHINode &PN); 638 639 /// Helper function for FoldPHIArgXIntoPHI() to set debug location for the 640 /// folded operation. 641 void PHIArgMergedDebugLoc(Instruction *Inst, PHINode &PN); 642 643 Instruction *foldGEPICmp(GEPOperator *GEPLHS, Value *RHS, 644 ICmpInst::Predicate Cond, Instruction &I); 645 Instruction *foldAllocaCmp(ICmpInst &ICI, const AllocaInst *Alloca, 646 const Value *Other); 647 Instruction *foldCmpLoadFromIndexedGlobal(GetElementPtrInst *GEP, 648 GlobalVariable *GV, CmpInst &ICI, 649 ConstantInt *AndCst = nullptr); 650 Instruction *foldFCmpIntToFPConst(FCmpInst &I, Instruction *LHSI, 651 Constant *RHSC); 652 Instruction *foldICmpAddOpConst(Value *X, const APInt &C, 653 ICmpInst::Predicate Pred); 654 Instruction *foldICmpWithCastOp(ICmpInst &ICI); 655 656 Instruction *foldICmpUsingKnownBits(ICmpInst &Cmp); 657 Instruction *foldICmpWithDominatingICmp(ICmpInst &Cmp); 658 Instruction *foldICmpWithConstant(ICmpInst &Cmp); 659 Instruction *foldICmpInstWithConstant(ICmpInst &Cmp); 660 Instruction *foldICmpInstWithConstantNotInt(ICmpInst &Cmp); 661 Instruction *foldICmpBinOp(ICmpInst &Cmp, const SimplifyQuery &SQ); 662 Instruction *foldICmpEquality(ICmpInst &Cmp); 663 Instruction *foldIRemByPowerOfTwoToBitTest(ICmpInst &I); 664 Instruction *foldSignBitTest(ICmpInst &I); 665 Instruction *foldICmpWithZero(ICmpInst &Cmp); 666 667 Value *foldUnsignedMultiplicationOverflowCheck(ICmpInst &Cmp); 668 669 Instruction *foldICmpSelectConstant(ICmpInst &Cmp, SelectInst *Select, 670 ConstantInt *C); 671 Instruction *foldICmpTruncConstant(ICmpInst &Cmp, TruncInst *Trunc, 672 const APInt &C); 673 Instruction *foldICmpAndConstant(ICmpInst &Cmp, BinaryOperator *And, 674 const APInt &C); 675 Instruction *foldICmpXorConstant(ICmpInst &Cmp, BinaryOperator *Xor, 676 const APInt &C); 677 Instruction *foldICmpOrConstant(ICmpInst &Cmp, BinaryOperator *Or, 678 const APInt &C); 679 Instruction *foldICmpMulConstant(ICmpInst &Cmp, BinaryOperator *Mul, 680 const APInt &C); 681 Instruction *foldICmpShlConstant(ICmpInst &Cmp, BinaryOperator *Shl, 682 const APInt &C); 683 Instruction *foldICmpShrConstant(ICmpInst &Cmp, BinaryOperator *Shr, 684 const APInt &C); 685 Instruction *foldICmpSRemConstant(ICmpInst &Cmp, BinaryOperator *UDiv, 686 const APInt &C); 687 Instruction *foldICmpUDivConstant(ICmpInst &Cmp, BinaryOperator *UDiv, 688 const APInt &C); 689 Instruction *foldICmpDivConstant(ICmpInst &Cmp, BinaryOperator *Div, 690 const APInt &C); 691 Instruction *foldICmpSubConstant(ICmpInst &Cmp, BinaryOperator *Sub, 692 const APInt &C); 693 Instruction *foldICmpAddConstant(ICmpInst &Cmp, BinaryOperator *Add, 694 const APInt &C); 695 Instruction *foldICmpAndConstConst(ICmpInst &Cmp, BinaryOperator *And, 696 const APInt &C1); 697 Instruction *foldICmpAndShift(ICmpInst &Cmp, BinaryOperator *And, 698 const APInt &C1, const APInt &C2); 699 Instruction *foldICmpShrConstConst(ICmpInst &I, Value *ShAmt, const APInt &C1, 700 const APInt &C2); 701 Instruction *foldICmpShlConstConst(ICmpInst &I, Value *ShAmt, const APInt &C1, 702 const APInt &C2); 703 704 Instruction *foldICmpBinOpEqualityWithConstant(ICmpInst &Cmp, 705 BinaryOperator *BO, 706 const APInt &C); 707 Instruction *foldICmpIntrinsicWithConstant(ICmpInst &ICI, IntrinsicInst *II, 708 const APInt &C); 709 Instruction *foldICmpEqIntrinsicWithConstant(ICmpInst &ICI, IntrinsicInst *II, 710 const APInt &C); 711 712 // Helpers of visitSelectInst(). 713 Instruction *foldSelectExtConst(SelectInst &Sel); 714 Instruction *foldSelectOpOp(SelectInst &SI, Instruction *TI, Instruction *FI); 715 Instruction *foldSelectIntoOp(SelectInst &SI, Value *, Value *); 716 Instruction *foldSPFofSPF(Instruction *Inner, SelectPatternFlavor SPF1, 717 Value *A, Value *B, Instruction &Outer, 718 SelectPatternFlavor SPF2, Value *C); 719 Instruction *foldSelectInstWithICmp(SelectInst &SI, ICmpInst *ICI); 720 Instruction *foldSelectValueEquivalence(SelectInst &SI, ICmpInst &ICI); 721 722 Value *insertRangeTest(Value *V, const APInt &Lo, const APInt &Hi, 723 bool isSigned, bool Inside); 724 Instruction *PromoteCastOfAllocation(BitCastInst &CI, AllocaInst &AI); 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 /// Returns a value X such that Val = X * Scale, or null if none. 739 /// 740 /// If the multiplication is known not to overflow then NoSignedWrap is set. 741 Value *Descale(Value *Val, APInt Scale, bool &NoSignedWrap); 742}; 743 744class Negator final { 745 /// Top-to-bottom, def-to-use negated instruction tree we produced. 746 SmallVector<Instruction *, NegatorMaxNodesSSO> NewInstructions; 747 748 using BuilderTy = IRBuilder<TargetFolder, IRBuilderCallbackInserter>; 749 BuilderTy Builder; 750 751 const DataLayout &DL; 752 AssumptionCache &AC; 753 const DominatorTree &DT; 754 755 const bool IsTrulyNegation; 756 757 SmallDenseMap<Value *, Value *> NegationsCache; 758 759 Negator(LLVMContext &C, const DataLayout &DL, AssumptionCache &AC, 760 const DominatorTree &DT, bool IsTrulyNegation); 761 762#if LLVM_ENABLE_STATS 763 unsigned NumValuesVisitedInThisNegator = 0; 764 ~Negator(); 765#endif 766 767 using Result = std::pair<ArrayRef<Instruction *> /*NewInstructions*/, 768 Value * /*NegatedRoot*/>; 769 770 std::array<Value *, 2> getSortedOperandsOfBinOp(Instruction *I); 771 772 LLVM_NODISCARD Value *visitImpl(Value *V, unsigned Depth); 773 774 LLVM_NODISCARD Value *negate(Value *V, unsigned Depth); 775 776 /// Recurse depth-first and attempt to sink the negation. 777 /// FIXME: use worklist? 778 LLVM_NODISCARD Optional<Result> run(Value *Root); 779 780 Negator(const Negator &) = delete; 781 Negator(Negator &&) = delete; 782 Negator &operator=(const Negator &) = delete; 783 Negator &operator=(Negator &&) = delete; 784 785public: 786 /// Attempt to negate \p Root. Retuns nullptr if negation can't be performed, 787 /// otherwise returns negated value. 788 LLVM_NODISCARD static Value *Negate(bool LHSIsZero, Value *Root, 789 InstCombinerImpl &IC); 790}; 791 792} // end namespace llvm 793 794#undef DEBUG_TYPE 795 796#endif // LLVM_LIB_TRANSFORMS_INSTCOMBINE_INSTCOMBINEINTERNAL_H 797