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 34 using 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. 38 static 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. 42 static constexpr unsigned NegatorMaxNodesSSO = 16; 43 44 namespace llvm { 45 46 class AAResults; 47 class APInt; 48 class AssumptionCache; 49 class BlockFrequencyInfo; 50 class DataLayout; 51 class DominatorTree; 52 class GEPOperator; 53 class GlobalVariable; 54 class LoopInfo; 55 class OptimizationRemarkEmitter; 56 class ProfileSummaryInfo; 57 class TargetLibraryInfo; 58 class User; 59 60 class LLVM_LIBRARY_VISIBILITY InstCombinerImpl final 61 : public InstCombiner, 62 public InstVisitor<InstCombinerImpl, Instruction *> { 63 public: InstCombinerImpl(InstCombineWorklist & Worklist,BuilderTy & Builder,bool MinimizeSize,AAResults * AA,AssumptionCache & AC,TargetLibraryInfo & TLI,TargetTransformInfo & TTI,DominatorTree & DT,OptimizationRemarkEmitter & ORE,BlockFrequencyInfo * BFI,ProfileSummaryInfo * PSI,const DataLayout & DL,LoopInfo * LI)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 ~InstCombinerImpl()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 Instruction *visitXor(BinaryOperator &I); 109 Instruction *visitShl(BinaryOperator &I); 110 Value *reassociateShiftAmtsOfTwoSameDirectionShifts( 111 BinaryOperator *Sh0, const SimplifyQuery &SQ, 112 bool AnalyzeForSignBitExtraction = false); 113 Instruction *canonicalizeCondSignextOfHighBitExtractToSignextHighBitExtract( 114 BinaryOperator &I); 115 Instruction *foldVariableSignZeroExtensionOfVariableHighBitExtract( 116 BinaryOperator &OldAShr); 117 Instruction *visitAShr(BinaryOperator &I); 118 Instruction *visitLShr(BinaryOperator &I); 119 Instruction *commonShiftTransforms(BinaryOperator &I); 120 Instruction *visitFCmpInst(FCmpInst &I); 121 CmpInst *canonicalizeICmpPredicate(CmpInst &I); 122 Instruction *visitICmpInst(ICmpInst &I); 123 Instruction *FoldShiftByConstant(Value *Op0, Constant *Op1, 124 BinaryOperator &I); 125 Instruction *commonCastTransforms(CastInst &CI); 126 Instruction *commonPointerCastTransforms(CastInst &CI); 127 Instruction *visitTrunc(TruncInst &CI); 128 Instruction *visitZExt(ZExtInst &CI); 129 Instruction *visitSExt(SExtInst &CI); 130 Instruction *visitFPTrunc(FPTruncInst &CI); 131 Instruction *visitFPExt(CastInst &CI); 132 Instruction *visitFPToUI(FPToUIInst &FI); 133 Instruction *visitFPToSI(FPToSIInst &FI); 134 Instruction *visitUIToFP(CastInst &CI); 135 Instruction *visitSIToFP(CastInst &CI); 136 Instruction *visitPtrToInt(PtrToIntInst &CI); 137 Instruction *visitIntToPtr(IntToPtrInst &CI); 138 Instruction *visitBitCast(BitCastInst &CI); 139 Instruction *visitAddrSpaceCast(AddrSpaceCastInst &CI); 140 Instruction *foldItoFPtoI(CastInst &FI); 141 Instruction *visitSelectInst(SelectInst &SI); 142 Instruction *visitCallInst(CallInst &CI); 143 Instruction *visitInvokeInst(InvokeInst &II); 144 Instruction *visitCallBrInst(CallBrInst &CBI); 145 146 Instruction *SliceUpIllegalIntegerPHI(PHINode &PN); 147 Instruction *visitPHINode(PHINode &PN); 148 Instruction *visitGetElementPtrInst(GetElementPtrInst &GEP); 149 Instruction *visitAllocaInst(AllocaInst &AI); 150 Instruction *visitAllocSite(Instruction &FI); 151 Instruction *visitFree(CallInst &FI); 152 Instruction *visitLoadInst(LoadInst &LI); 153 Instruction *visitStoreInst(StoreInst &SI); 154 Instruction *visitAtomicRMWInst(AtomicRMWInst &SI); 155 Instruction *visitUnconditionalBranchInst(BranchInst &BI); 156 Instruction *visitBranchInst(BranchInst &BI); 157 Instruction *visitFenceInst(FenceInst &FI); 158 Instruction *visitSwitchInst(SwitchInst &SI); 159 Instruction *visitReturnInst(ReturnInst &RI); 160 Instruction *visitUnreachableInst(UnreachableInst &I); 161 Instruction * 162 foldAggregateConstructionIntoAggregateReuse(InsertValueInst &OrigIVI); 163 Instruction *visitInsertValueInst(InsertValueInst &IV); 164 Instruction *visitInsertElementInst(InsertElementInst &IE); 165 Instruction *visitExtractElementInst(ExtractElementInst &EI); 166 Instruction *visitShuffleVectorInst(ShuffleVectorInst &SVI); 167 Instruction *visitExtractValueInst(ExtractValueInst &EV); 168 Instruction *visitLandingPadInst(LandingPadInst &LI); 169 Instruction *visitVAEndInst(VAEndInst &I); 170 Instruction *visitFreeze(FreezeInst &I); 171 172 /// Specify what to return for unhandled instructions. visitInstruction(Instruction & I)173 Instruction *visitInstruction(Instruction &I) { return nullptr; } 174 175 /// True when DB dominates all uses of DI except UI. 176 /// UI must be in the same block as DI. 177 /// The routine checks that the DI parent and DB are different. 178 bool dominatesAllUses(const Instruction *DI, const Instruction *UI, 179 const BasicBlock *DB) const; 180 181 /// Try to replace select with select operand SIOpd in SI-ICmp sequence. 182 bool replacedSelectWithOperand(SelectInst *SI, const ICmpInst *Icmp, 183 const unsigned SIOpd); 184 185 /// Try to replace instruction \p I with value \p V which are pointers 186 /// in different address space. 187 /// \return true if successful. 188 bool replacePointer(Instruction &I, Value *V); 189 190 LoadInst *combineLoadToNewType(LoadInst &LI, Type *NewTy, 191 const Twine &Suffix = ""); 192 193 private: 194 bool shouldChangeType(unsigned FromBitWidth, unsigned ToBitWidth) const; 195 bool shouldChangeType(Type *From, Type *To) const; 196 Value *dyn_castNegVal(Value *V) const; 197 Type *FindElementAtOffset(PointerType *PtrTy, int64_t Offset, 198 SmallVectorImpl<Value *> &NewIndices); 199 200 /// Classify whether a cast is worth optimizing. 201 /// 202 /// This is a helper to decide whether the simplification of 203 /// logic(cast(A), cast(B)) to cast(logic(A, B)) should be performed. 204 /// 205 /// \param CI The cast we are interested in. 206 /// 207 /// \return true if this cast actually results in any code being generated and 208 /// if it cannot already be eliminated by some other transformation. 209 bool shouldOptimizeCast(CastInst *CI); 210 211 /// Try to optimize a sequence of instructions checking if an operation 212 /// on LHS and RHS overflows. 213 /// 214 /// If this overflow check is done via one of the overflow check intrinsics, 215 /// then CtxI has to be the call instruction calling that intrinsic. If this 216 /// overflow check is done by arithmetic followed by a compare, then CtxI has 217 /// to be the arithmetic instruction. 218 /// 219 /// If a simplification is possible, stores the simplified result of the 220 /// operation in OperationResult and result of the overflow check in 221 /// OverflowResult, and return true. If no simplification is possible, 222 /// returns false. 223 bool OptimizeOverflowCheck(Instruction::BinaryOps BinaryOp, bool IsSigned, 224 Value *LHS, Value *RHS, 225 Instruction &CtxI, Value *&OperationResult, 226 Constant *&OverflowResult); 227 228 Instruction *visitCallBase(CallBase &Call); 229 Instruction *tryOptimizeCall(CallInst *CI); 230 bool transformConstExprCastCall(CallBase &Call); 231 Instruction *transformCallThroughTrampoline(CallBase &Call, 232 IntrinsicInst &Tramp); 233 234 Value *simplifyMaskedLoad(IntrinsicInst &II); 235 Instruction *simplifyMaskedStore(IntrinsicInst &II); 236 Instruction *simplifyMaskedGather(IntrinsicInst &II); 237 Instruction *simplifyMaskedScatter(IntrinsicInst &II); 238 239 /// Transform (zext icmp) to bitwise / integer operations in order to 240 /// eliminate it. 241 /// 242 /// \param ICI The icmp of the (zext icmp) pair we are interested in. 243 /// \parem CI The zext of the (zext icmp) pair we are interested in. 244 /// \param DoTransform Pass false to just test whether the given (zext icmp) 245 /// would be transformed. Pass true to actually perform the transformation. 246 /// 247 /// \return null if the transformation cannot be performed. If the 248 /// transformation can be performed the new instruction that replaces the 249 /// (zext icmp) pair will be returned (if \p DoTransform is false the 250 /// unmodified \p ICI will be returned in this case). 251 Instruction *transformZExtICmp(ICmpInst *ICI, ZExtInst &CI, 252 bool DoTransform = true); 253 254 Instruction *transformSExtICmp(ICmpInst *ICI, Instruction &CI); 255 willNotOverflowSignedAdd(const Value * LHS,const Value * RHS,const Instruction & CxtI)256 bool willNotOverflowSignedAdd(const Value *LHS, const Value *RHS, 257 const Instruction &CxtI) const { 258 return computeOverflowForSignedAdd(LHS, RHS, &CxtI) == 259 OverflowResult::NeverOverflows; 260 } 261 willNotOverflowUnsignedAdd(const Value * LHS,const Value * RHS,const Instruction & CxtI)262 bool willNotOverflowUnsignedAdd(const Value *LHS, const Value *RHS, 263 const Instruction &CxtI) const { 264 return computeOverflowForUnsignedAdd(LHS, RHS, &CxtI) == 265 OverflowResult::NeverOverflows; 266 } 267 willNotOverflowAdd(const Value * LHS,const Value * RHS,const Instruction & CxtI,bool IsSigned)268 bool willNotOverflowAdd(const Value *LHS, const Value *RHS, 269 const Instruction &CxtI, bool IsSigned) const { 270 return IsSigned ? willNotOverflowSignedAdd(LHS, RHS, CxtI) 271 : willNotOverflowUnsignedAdd(LHS, RHS, CxtI); 272 } 273 willNotOverflowSignedSub(const Value * LHS,const Value * RHS,const Instruction & CxtI)274 bool willNotOverflowSignedSub(const Value *LHS, const Value *RHS, 275 const Instruction &CxtI) const { 276 return computeOverflowForSignedSub(LHS, RHS, &CxtI) == 277 OverflowResult::NeverOverflows; 278 } 279 willNotOverflowUnsignedSub(const Value * LHS,const Value * RHS,const Instruction & CxtI)280 bool willNotOverflowUnsignedSub(const Value *LHS, const Value *RHS, 281 const Instruction &CxtI) const { 282 return computeOverflowForUnsignedSub(LHS, RHS, &CxtI) == 283 OverflowResult::NeverOverflows; 284 } 285 willNotOverflowSub(const Value * LHS,const Value * RHS,const Instruction & CxtI,bool IsSigned)286 bool willNotOverflowSub(const Value *LHS, const Value *RHS, 287 const Instruction &CxtI, bool IsSigned) const { 288 return IsSigned ? willNotOverflowSignedSub(LHS, RHS, CxtI) 289 : willNotOverflowUnsignedSub(LHS, RHS, CxtI); 290 } 291 willNotOverflowSignedMul(const Value * LHS,const Value * RHS,const Instruction & CxtI)292 bool willNotOverflowSignedMul(const Value *LHS, const Value *RHS, 293 const Instruction &CxtI) const { 294 return computeOverflowForSignedMul(LHS, RHS, &CxtI) == 295 OverflowResult::NeverOverflows; 296 } 297 willNotOverflowUnsignedMul(const Value * LHS,const Value * RHS,const Instruction & CxtI)298 bool willNotOverflowUnsignedMul(const Value *LHS, const Value *RHS, 299 const Instruction &CxtI) const { 300 return computeOverflowForUnsignedMul(LHS, RHS, &CxtI) == 301 OverflowResult::NeverOverflows; 302 } 303 willNotOverflowMul(const Value * LHS,const Value * RHS,const Instruction & CxtI,bool IsSigned)304 bool willNotOverflowMul(const Value *LHS, const Value *RHS, 305 const Instruction &CxtI, bool IsSigned) const { 306 return IsSigned ? willNotOverflowSignedMul(LHS, RHS, CxtI) 307 : willNotOverflowUnsignedMul(LHS, RHS, CxtI); 308 } 309 willNotOverflow(BinaryOperator::BinaryOps Opcode,const Value * LHS,const Value * RHS,const Instruction & CxtI,bool IsSigned)310 bool willNotOverflow(BinaryOperator::BinaryOps Opcode, const Value *LHS, 311 const Value *RHS, const Instruction &CxtI, 312 bool IsSigned) const { 313 switch (Opcode) { 314 case Instruction::Add: return willNotOverflowAdd(LHS, RHS, CxtI, IsSigned); 315 case Instruction::Sub: return willNotOverflowSub(LHS, RHS, CxtI, IsSigned); 316 case Instruction::Mul: return willNotOverflowMul(LHS, RHS, CxtI, IsSigned); 317 default: llvm_unreachable("Unexpected opcode for overflow query"); 318 } 319 } 320 321 Value *EmitGEPOffset(User *GEP); 322 Instruction *scalarizePHI(ExtractElementInst &EI, PHINode *PN); 323 Instruction *foldCastedBitwiseLogic(BinaryOperator &I); 324 Instruction *narrowBinOp(TruncInst &Trunc); 325 Instruction *narrowMaskedBinOp(BinaryOperator &And); 326 Instruction *narrowMathIfNoOverflow(BinaryOperator &I); 327 Instruction *narrowFunnelShift(TruncInst &Trunc); 328 Instruction *optimizeBitCastFromPhi(CastInst &CI, PHINode *PN); 329 Instruction *matchSAddSubSat(SelectInst &MinMax1); 330 331 /// Determine if a pair of casts can be replaced by a single cast. 332 /// 333 /// \param CI1 The first of a pair of casts. 334 /// \param CI2 The second of a pair of casts. 335 /// 336 /// \return 0 if the cast pair cannot be eliminated, otherwise returns an 337 /// Instruction::CastOps value for a cast that can replace the pair, casting 338 /// CI1->getSrcTy() to CI2->getDstTy(). 339 /// 340 /// \see CastInst::isEliminableCastPair 341 Instruction::CastOps isEliminableCastPair(const CastInst *CI1, 342 const CastInst *CI2); 343 344 Value *foldAndOfICmps(ICmpInst *LHS, ICmpInst *RHS, BinaryOperator &And); 345 Value *foldOrOfICmps(ICmpInst *LHS, ICmpInst *RHS, BinaryOperator &Or); 346 Value *foldXorOfICmps(ICmpInst *LHS, ICmpInst *RHS, BinaryOperator &Xor); 347 348 /// Optimize (fcmp)&(fcmp) or (fcmp)|(fcmp). 349 /// NOTE: Unlike most of instcombine, this returns a Value which should 350 /// already be inserted into the function. 351 Value *foldLogicOfFCmps(FCmpInst *LHS, FCmpInst *RHS, bool IsAnd); 352 353 Value *foldAndOrOfICmpsOfAndWithPow2(ICmpInst *LHS, ICmpInst *RHS, 354 BinaryOperator &Logic); 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 public: 362 /// Inserts an instruction \p New before instruction \p Old 363 /// 364 /// Also adds the new instruction to the worklist and returns \p New so that 365 /// it is suitable for use as the return from the visitation patterns. InsertNewInstBefore(Instruction * New,Instruction & Old)366 Instruction *InsertNewInstBefore(Instruction *New, Instruction &Old) { 367 assert(New && !New->getParent() && 368 "New instruction already inserted into a basic block!"); 369 BasicBlock *BB = Old.getParent(); 370 BB->getInstList().insert(Old.getIterator(), New); // Insert inst 371 Worklist.add(New); 372 return New; 373 } 374 375 /// Same as InsertNewInstBefore, but also sets the debug loc. InsertNewInstWith(Instruction * New,Instruction & Old)376 Instruction *InsertNewInstWith(Instruction *New, Instruction &Old) { 377 New->setDebugLoc(Old.getDebugLoc()); 378 return InsertNewInstBefore(New, Old); 379 } 380 381 /// A combiner-aware RAUW-like routine. 382 /// 383 /// This method is to be used when an instruction is found to be dead, 384 /// replaceable with another preexisting expression. Here we add all uses of 385 /// I to the worklist, replace all uses of I with the new value, then return 386 /// I, so that the inst combiner will know that I was modified. replaceInstUsesWith(Instruction & I,Value * V)387 Instruction *replaceInstUsesWith(Instruction &I, Value *V) { 388 // If there are no uses to replace, then we return nullptr to indicate that 389 // no changes were made to the program. 390 if (I.use_empty()) return nullptr; 391 392 Worklist.pushUsersToWorkList(I); // Add all modified instrs to worklist. 393 394 // If we are replacing the instruction with itself, this must be in a 395 // segment of unreachable code, so just clobber the instruction. 396 if (&I == V) 397 V = UndefValue::get(I.getType()); 398 399 LLVM_DEBUG(dbgs() << "IC: Replacing " << I << "\n" 400 << " with " << *V << '\n'); 401 402 I.replaceAllUsesWith(V); 403 return &I; 404 } 405 406 /// Replace operand of instruction and add old operand to the worklist. replaceOperand(Instruction & I,unsigned OpNum,Value * V)407 Instruction *replaceOperand(Instruction &I, unsigned OpNum, Value *V) { 408 Worklist.addValue(I.getOperand(OpNum)); 409 I.setOperand(OpNum, V); 410 return &I; 411 } 412 413 /// Replace use and add the previously used value to the worklist. replaceUse(Use & U,Value * NewValue)414 void replaceUse(Use &U, Value *NewValue) { 415 Worklist.addValue(U); 416 U = NewValue; 417 } 418 419 /// Creates a result tuple for an overflow intrinsic \p II with a given 420 /// \p Result and a constant \p Overflow value. CreateOverflowTuple(IntrinsicInst * II,Value * Result,Constant * Overflow)421 Instruction *CreateOverflowTuple(IntrinsicInst *II, Value *Result, 422 Constant *Overflow) { 423 Constant *V[] = {UndefValue::get(Result->getType()), Overflow}; 424 StructType *ST = cast<StructType>(II->getType()); 425 Constant *Struct = ConstantStruct::get(ST, V); 426 return InsertValueInst::Create(Struct, Result, 0); 427 } 428 429 /// Create and insert the idiom we use to indicate a block is unreachable 430 /// without having to rewrite the CFG from within InstCombine. CreateNonTerminatorUnreachable(Instruction * InsertAt)431 void CreateNonTerminatorUnreachable(Instruction *InsertAt) { 432 auto &Ctx = InsertAt->getContext(); 433 new StoreInst(ConstantInt::getTrue(Ctx), 434 UndefValue::get(Type::getInt1PtrTy(Ctx)), 435 InsertAt); 436 } 437 438 439 /// Combiner aware instruction erasure. 440 /// 441 /// When dealing with an instruction that has side effects or produces a void 442 /// value, we can't rely on DCE to delete the instruction. Instead, visit 443 /// methods should return the value returned by this function. eraseInstFromFunction(Instruction & I)444 Instruction *eraseInstFromFunction(Instruction &I) override { 445 LLVM_DEBUG(dbgs() << "IC: ERASE " << I << '\n'); 446 assert(I.use_empty() && "Cannot erase instruction that is used!"); 447 salvageDebugInfo(I); 448 449 // Make sure that we reprocess all operands now that we reduced their 450 // use counts. 451 for (Use &Operand : I.operands()) 452 if (auto *Inst = dyn_cast<Instruction>(Operand)) 453 Worklist.add(Inst); 454 455 Worklist.remove(&I); 456 I.eraseFromParent(); 457 MadeIRChange = true; 458 return nullptr; // Don't do anything with FI 459 } 460 computeKnownBits(const Value * V,KnownBits & Known,unsigned Depth,const Instruction * CxtI)461 void computeKnownBits(const Value *V, KnownBits &Known, 462 unsigned Depth, const Instruction *CxtI) const { 463 llvm::computeKnownBits(V, Known, DL, Depth, &AC, CxtI, &DT); 464 } 465 computeKnownBits(const Value * V,unsigned Depth,const Instruction * CxtI)466 KnownBits computeKnownBits(const Value *V, unsigned Depth, 467 const Instruction *CxtI) const { 468 return llvm::computeKnownBits(V, DL, Depth, &AC, CxtI, &DT); 469 } 470 471 bool isKnownToBeAPowerOfTwo(const Value *V, bool OrZero = false, 472 unsigned Depth = 0, 473 const Instruction *CxtI = nullptr) { 474 return llvm::isKnownToBeAPowerOfTwo(V, DL, OrZero, Depth, &AC, CxtI, &DT); 475 } 476 477 bool MaskedValueIsZero(const Value *V, const APInt &Mask, unsigned Depth = 0, 478 const Instruction *CxtI = nullptr) const { 479 return llvm::MaskedValueIsZero(V, Mask, DL, Depth, &AC, CxtI, &DT); 480 } 481 482 unsigned ComputeNumSignBits(const Value *Op, unsigned Depth = 0, 483 const Instruction *CxtI = nullptr) const { 484 return llvm::ComputeNumSignBits(Op, DL, Depth, &AC, CxtI, &DT); 485 } 486 computeOverflowForUnsignedMul(const Value * LHS,const Value * RHS,const Instruction * CxtI)487 OverflowResult computeOverflowForUnsignedMul(const Value *LHS, 488 const Value *RHS, 489 const Instruction *CxtI) const { 490 return llvm::computeOverflowForUnsignedMul(LHS, RHS, DL, &AC, CxtI, &DT); 491 } 492 computeOverflowForSignedMul(const Value * LHS,const Value * RHS,const Instruction * CxtI)493 OverflowResult computeOverflowForSignedMul(const Value *LHS, 494 const Value *RHS, 495 const Instruction *CxtI) const { 496 return llvm::computeOverflowForSignedMul(LHS, RHS, DL, &AC, CxtI, &DT); 497 } 498 computeOverflowForUnsignedAdd(const Value * LHS,const Value * RHS,const Instruction * CxtI)499 OverflowResult computeOverflowForUnsignedAdd(const Value *LHS, 500 const Value *RHS, 501 const Instruction *CxtI) const { 502 return llvm::computeOverflowForUnsignedAdd(LHS, RHS, DL, &AC, CxtI, &DT); 503 } 504 computeOverflowForSignedAdd(const Value * LHS,const Value * RHS,const Instruction * CxtI)505 OverflowResult computeOverflowForSignedAdd(const Value *LHS, 506 const Value *RHS, 507 const Instruction *CxtI) const { 508 return llvm::computeOverflowForSignedAdd(LHS, RHS, DL, &AC, CxtI, &DT); 509 } 510 computeOverflowForUnsignedSub(const Value * LHS,const Value * RHS,const Instruction * CxtI)511 OverflowResult computeOverflowForUnsignedSub(const Value *LHS, 512 const Value *RHS, 513 const Instruction *CxtI) const { 514 return llvm::computeOverflowForUnsignedSub(LHS, RHS, DL, &AC, CxtI, &DT); 515 } 516 computeOverflowForSignedSub(const Value * LHS,const Value * RHS,const Instruction * CxtI)517 OverflowResult computeOverflowForSignedSub(const Value *LHS, const Value *RHS, 518 const Instruction *CxtI) const { 519 return llvm::computeOverflowForSignedSub(LHS, RHS, DL, &AC, CxtI, &DT); 520 } 521 522 OverflowResult computeOverflow( 523 Instruction::BinaryOps BinaryOp, bool IsSigned, 524 Value *LHS, Value *RHS, Instruction *CxtI) const; 525 526 /// Performs a few simplifications for operators which are associative 527 /// or commutative. 528 bool SimplifyAssociativeOrCommutative(BinaryOperator &I); 529 530 /// Tries to simplify binary operations which some other binary 531 /// operation distributes over. 532 /// 533 /// It does this by either by factorizing out common terms (eg "(A*B)+(A*C)" 534 /// -> "A*(B+C)") or expanding out if this results in simplifications (eg: "A 535 /// & (B | C) -> (A&B) | (A&C)" if this is a win). Returns the simplified 536 /// value, or null if it didn't simplify. 537 Value *SimplifyUsingDistributiveLaws(BinaryOperator &I); 538 539 /// Tries to simplify add operations using the definition of remainder. 540 /// 541 /// The definition of remainder is X % C = X - (X / C ) * C. The add 542 /// expression X % C0 + (( X / C0 ) % C1) * C0 can be simplified to 543 /// X % (C0 * C1) 544 Value *SimplifyAddWithRemainder(BinaryOperator &I); 545 546 // Binary Op helper for select operations where the expression can be 547 // efficiently reorganized. 548 Value *SimplifySelectsFeedingBinaryOp(BinaryOperator &I, Value *LHS, 549 Value *RHS); 550 551 /// This tries to simplify binary operations by factorizing out common terms 552 /// (e. g. "(A*B)+(A*C)" -> "A*(B+C)"). 553 Value *tryFactorization(BinaryOperator &, Instruction::BinaryOps, Value *, 554 Value *, Value *, Value *); 555 556 /// Match a select chain which produces one of three values based on whether 557 /// the LHS is less than, equal to, or greater than RHS respectively. 558 /// Return true if we matched a three way compare idiom. The LHS, RHS, Less, 559 /// Equal and Greater values are saved in the matching process and returned to 560 /// the caller. 561 bool matchThreeWayIntCompare(SelectInst *SI, Value *&LHS, Value *&RHS, 562 ConstantInt *&Less, ConstantInt *&Equal, 563 ConstantInt *&Greater); 564 565 /// Attempts to replace V with a simpler value based on the demanded 566 /// bits. 567 Value *SimplifyDemandedUseBits(Value *V, APInt DemandedMask, KnownBits &Known, 568 unsigned Depth, Instruction *CxtI); 569 bool SimplifyDemandedBits(Instruction *I, unsigned Op, 570 const APInt &DemandedMask, KnownBits &Known, 571 unsigned Depth = 0) override; 572 573 /// Helper routine of SimplifyDemandedUseBits. It computes KnownZero/KnownOne 574 /// bits. It also tries to handle simplifications that can be done based on 575 /// DemandedMask, but without modifying the Instruction. 576 Value *SimplifyMultipleUseDemandedBits(Instruction *I, 577 const APInt &DemandedMask, 578 KnownBits &Known, 579 unsigned Depth, Instruction *CxtI); 580 581 /// Helper routine of SimplifyDemandedUseBits. It tries to simplify demanded 582 /// bit for "r1 = shr x, c1; r2 = shl r1, c2" instruction sequence. 583 Value *simplifyShrShlDemandedBits( 584 Instruction *Shr, const APInt &ShrOp1, Instruction *Shl, 585 const APInt &ShlOp1, const APInt &DemandedMask, KnownBits &Known); 586 587 /// Tries to simplify operands to an integer instruction based on its 588 /// demanded bits. 589 bool SimplifyDemandedInstructionBits(Instruction &Inst); 590 591 virtual Value * 592 SimplifyDemandedVectorElts(Value *V, APInt DemandedElts, APInt &UndefElts, 593 unsigned Depth = 0, 594 bool AllowMultipleUsers = false) override; 595 596 /// Canonicalize the position of binops relative to shufflevector. 597 Instruction *foldVectorBinop(BinaryOperator &Inst); 598 Instruction *foldVectorSelect(SelectInst &Sel); 599 600 /// Given a binary operator, cast instruction, or select which has a PHI node 601 /// as operand #0, see if we can fold the instruction into the PHI (which is 602 /// only possible if all operands to the PHI are constants). 603 Instruction *foldOpIntoPhi(Instruction &I, PHINode *PN); 604 605 /// Given an instruction with a select as one operand and a constant as the 606 /// other operand, try to fold the binary operator into the select arguments. 607 /// This also works for Cast instructions, which obviously do not have a 608 /// second operand. 609 Instruction *FoldOpIntoSelect(Instruction &Op, SelectInst *SI); 610 611 /// This is a convenience wrapper function for the above two functions. 612 Instruction *foldBinOpIntoSelectOrPhi(BinaryOperator &I); 613 614 Instruction *foldAddWithConstant(BinaryOperator &Add); 615 616 /// Try to rotate an operation below a PHI node, using PHI nodes for 617 /// its operands. 618 Instruction *foldPHIArgOpIntoPHI(PHINode &PN); 619 Instruction *foldPHIArgBinOpIntoPHI(PHINode &PN); 620 Instruction *foldPHIArgInsertValueInstructionIntoPHI(PHINode &PN); 621 Instruction *foldPHIArgExtractValueInstructionIntoPHI(PHINode &PN); 622 Instruction *foldPHIArgGEPIntoPHI(PHINode &PN); 623 Instruction *foldPHIArgLoadIntoPHI(PHINode &PN); 624 Instruction *foldPHIArgZextsIntoPHI(PHINode &PN); 625 626 /// If an integer typed PHI has only one use which is an IntToPtr operation, 627 /// replace the PHI with an existing pointer typed PHI if it exists. Otherwise 628 /// insert a new pointer typed PHI and replace the original one. 629 Instruction *foldIntegerTypedPHI(PHINode &PN); 630 631 /// Helper function for FoldPHIArgXIntoPHI() to set debug location for the 632 /// folded operation. 633 void PHIArgMergedDebugLoc(Instruction *Inst, PHINode &PN); 634 635 Instruction *foldGEPICmp(GEPOperator *GEPLHS, Value *RHS, 636 ICmpInst::Predicate Cond, Instruction &I); 637 Instruction *foldAllocaCmp(ICmpInst &ICI, const AllocaInst *Alloca, 638 const Value *Other); 639 Instruction *foldCmpLoadFromIndexedGlobal(GetElementPtrInst *GEP, 640 GlobalVariable *GV, CmpInst &ICI, 641 ConstantInt *AndCst = nullptr); 642 Instruction *foldFCmpIntToFPConst(FCmpInst &I, Instruction *LHSI, 643 Constant *RHSC); 644 Instruction *foldICmpAddOpConst(Value *X, const APInt &C, 645 ICmpInst::Predicate Pred); 646 Instruction *foldICmpWithCastOp(ICmpInst &ICI); 647 648 Instruction *foldICmpUsingKnownBits(ICmpInst &Cmp); 649 Instruction *foldICmpWithDominatingICmp(ICmpInst &Cmp); 650 Instruction *foldICmpWithConstant(ICmpInst &Cmp); 651 Instruction *foldICmpInstWithConstant(ICmpInst &Cmp); 652 Instruction *foldICmpInstWithConstantNotInt(ICmpInst &Cmp); 653 Instruction *foldICmpBinOp(ICmpInst &Cmp, const SimplifyQuery &SQ); 654 Instruction *foldICmpEquality(ICmpInst &Cmp); 655 Instruction *foldIRemByPowerOfTwoToBitTest(ICmpInst &I); 656 Instruction *foldSignBitTest(ICmpInst &I); 657 Instruction *foldICmpWithZero(ICmpInst &Cmp); 658 659 Value *foldUnsignedMultiplicationOverflowCheck(ICmpInst &Cmp); 660 661 Instruction *foldICmpSelectConstant(ICmpInst &Cmp, SelectInst *Select, 662 ConstantInt *C); 663 Instruction *foldICmpTruncConstant(ICmpInst &Cmp, TruncInst *Trunc, 664 const APInt &C); 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 *foldICmpShrConstConst(ICmpInst &I, Value *ShAmt, const APInt &C1, 692 const APInt &C2); 693 Instruction *foldICmpShlConstConst(ICmpInst &I, Value *ShAmt, const APInt &C1, 694 const APInt &C2); 695 696 Instruction *foldICmpBinOpEqualityWithConstant(ICmpInst &Cmp, 697 BinaryOperator *BO, 698 const APInt &C); 699 Instruction *foldICmpIntrinsicWithConstant(ICmpInst &ICI, IntrinsicInst *II, 700 const APInt &C); 701 Instruction *foldICmpEqIntrinsicWithConstant(ICmpInst &ICI, IntrinsicInst *II, 702 const APInt &C); 703 704 // Helpers of visitSelectInst(). 705 Instruction *foldSelectExtConst(SelectInst &Sel); 706 Instruction *foldSelectOpOp(SelectInst &SI, Instruction *TI, Instruction *FI); 707 Instruction *foldSelectIntoOp(SelectInst &SI, Value *, Value *); 708 Instruction *foldSPFofSPF(Instruction *Inner, SelectPatternFlavor SPF1, 709 Value *A, Value *B, Instruction &Outer, 710 SelectPatternFlavor SPF2, Value *C); 711 Instruction *foldSelectInstWithICmp(SelectInst &SI, ICmpInst *ICI); 712 Instruction *foldSelectValueEquivalence(SelectInst &SI, ICmpInst &ICI); 713 714 Instruction *OptAndOp(BinaryOperator *Op, ConstantInt *OpRHS, 715 ConstantInt *AndRHS, BinaryOperator &TheAnd); 716 717 Value *insertRangeTest(Value *V, const APInt &Lo, const APInt &Hi, 718 bool isSigned, bool Inside); 719 Instruction *PromoteCastOfAllocation(BitCastInst &CI, AllocaInst &AI); 720 bool mergeStoreIntoSuccessor(StoreInst &SI); 721 722 /// Given an 'or' instruction, check to see if it is part of a 723 /// bswap/bitreverse idiom. If so, return the equivalent bswap/bitreverse 724 /// intrinsic. 725 Instruction *matchBSwapOrBitReverse(BinaryOperator &Or, bool MatchBSwaps, 726 bool MatchBitReversals); 727 728 Instruction *SimplifyAnyMemTransfer(AnyMemTransferInst *MI); 729 Instruction *SimplifyAnyMemSet(AnyMemSetInst *MI); 730 731 Value *EvaluateInDifferentType(Value *V, Type *Ty, bool isSigned); 732 733 /// Returns a value X such that Val = X * Scale, or null if none. 734 /// 735 /// If the multiplication is known not to overflow then NoSignedWrap is set. 736 Value *Descale(Value *Val, APInt Scale, bool &NoSignedWrap); 737 }; 738 739 class Negator final { 740 /// Top-to-bottom, def-to-use negated instruction tree we produced. 741 SmallVector<Instruction *, NegatorMaxNodesSSO> NewInstructions; 742 743 using BuilderTy = IRBuilder<TargetFolder, IRBuilderCallbackInserter>; 744 BuilderTy Builder; 745 746 const DataLayout &DL; 747 AssumptionCache &AC; 748 const DominatorTree &DT; 749 750 const bool IsTrulyNegation; 751 752 SmallDenseMap<Value *, Value *> NegationsCache; 753 754 Negator(LLVMContext &C, const DataLayout &DL, AssumptionCache &AC, 755 const DominatorTree &DT, bool IsTrulyNegation); 756 757 #if LLVM_ENABLE_STATS 758 unsigned NumValuesVisitedInThisNegator = 0; 759 ~Negator(); 760 #endif 761 762 using Result = std::pair<ArrayRef<Instruction *> /*NewInstructions*/, 763 Value * /*NegatedRoot*/>; 764 765 std::array<Value *, 2> getSortedOperandsOfBinOp(Instruction *I); 766 767 LLVM_NODISCARD Value *visitImpl(Value *V, unsigned Depth); 768 769 LLVM_NODISCARD Value *negate(Value *V, unsigned Depth); 770 771 /// Recurse depth-first and attempt to sink the negation. 772 /// FIXME: use worklist? 773 LLVM_NODISCARD Optional<Result> run(Value *Root); 774 775 Negator(const Negator &) = delete; 776 Negator(Negator &&) = delete; 777 Negator &operator=(const Negator &) = delete; 778 Negator &operator=(Negator &&) = delete; 779 780 public: 781 /// Attempt to negate \p Root. Retuns nullptr if negation can't be performed, 782 /// otherwise returns negated value. 783 LLVM_NODISCARD static Value *Negate(bool LHSIsZero, Value *Root, 784 InstCombinerImpl &IC); 785 }; 786 787 } // end namespace llvm 788 789 #undef DEBUG_TYPE 790 791 #endif // LLVM_LIB_TRANSFORMS_INSTCOMBINE_INSTCOMBINEINTERNAL_H 792