1 //===- InstCombineInternal.h - InstCombine pass internals -------*- C++ -*-===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
11 /// This file provides internal interfaces used to implement the InstCombine.
13 //===----------------------------------------------------------------------===//
15 #ifndef LLVM_LIB_TRANSFORMS_INSTCOMBINE_INSTCOMBINEINTERNAL_H
16 #define LLVM_LIB_TRANSFORMS_INSTCOMBINE_INSTCOMBINEINTERNAL_H
18 #include "llvm/Analysis/AliasAnalysis.h"
19 #include "llvm/Analysis/AssumptionCache.h"
20 #include "llvm/Analysis/InstructionSimplify.h"
21 #include "llvm/Analysis/LoopInfo.h"
22 #include "llvm/Analysis/TargetFolder.h"
23 #include "llvm/Analysis/ValueTracking.h"
24 #include "llvm/IR/DIBuilder.h"
25 #include "llvm/IR/Dominators.h"
26 #include "llvm/IR/IRBuilder.h"
27 #include "llvm/IR/InstVisitor.h"
28 #include "llvm/IR/IntrinsicInst.h"
29 #include "llvm/IR/Operator.h"
30 #include "llvm/IR/PatternMatch.h"
31 #include "llvm/Pass.h"
32 #include "llvm/Support/Dwarf.h"
33 #include "llvm/Support/KnownBits.h"
34 #include "llvm/Transforms/InstCombine/InstCombineWorklist.h"
35 #include "llvm/Transforms/Utils/Local.h"
37 #define DEBUG_TYPE "instcombine"
43 class TargetLibraryInfo;
48 /// Assign a complexity or rank value to LLVM Values. This is used to reduce
49 /// the amount of pattern matching needed for compares and commutative
50 /// instructions. For example, if we have:
51 /// icmp ugt X, Constant
53 /// xor (add X, Constant), cast Z
55 /// We do not have to consider the commuted variants of these patterns because
56 /// canonicalization based on complexity guarantees the above ordering.
58 /// This routine maps IR values to various complexity ranks:
61 /// 2 -> Other non-instructions
63 /// 4 -> Cast and (f)neg/not instructions
64 /// 5 -> Other instructions
65 static inline unsigned getComplexity(Value *V) {
66 if (isa<Instruction>(V)) {
67 if (isa<CastInst>(V) || BinaryOperator::isNeg(V) ||
68 BinaryOperator::isFNeg(V) || BinaryOperator::isNot(V))
74 return isa<Constant>(V) ? (isa<UndefValue>(V) ? 0 : 1) : 2;
77 /// Predicate canonicalization reduces the number of patterns that need to be
78 /// matched by other transforms. For example, we may swap the operands of a
79 /// conditional branch or select to create a compare with a canonical (inverted)
80 /// predicate which is then more likely to be matched with other values.
81 static inline bool isCanonicalPredicate(CmpInst::Predicate Pred) {
83 case CmpInst::ICMP_NE:
84 case CmpInst::ICMP_ULE:
85 case CmpInst::ICMP_SLE:
86 case CmpInst::ICMP_UGE:
87 case CmpInst::ICMP_SGE:
88 // TODO: There are 16 FCMP predicates. Should others be (not) canonical?
89 case CmpInst::FCMP_ONE:
90 case CmpInst::FCMP_OLE:
91 case CmpInst::FCMP_OGE:
98 /// \brief Add one to a Constant
99 static inline Constant *AddOne(Constant *C) {
100 return ConstantExpr::getAdd(C, ConstantInt::get(C->getType(), 1));
102 /// \brief Subtract one from a Constant
103 static inline Constant *SubOne(Constant *C) {
104 return ConstantExpr::getSub(C, ConstantInt::get(C->getType(), 1));
107 /// \brief Return true if the specified value is free to invert (apply ~ to).
108 /// This happens in cases where the ~ can be eliminated. If WillInvertAllUses
109 /// is true, work under the assumption that the caller intends to remove all
110 /// uses of V and only keep uses of ~V.
112 static inline bool IsFreeToInvert(Value *V, bool WillInvertAllUses) {
114 if (BinaryOperator::isNot(V))
117 // Constants can be considered to be not'ed values.
118 if (isa<ConstantInt>(V))
121 // A vector of constant integers can be inverted easily.
123 if (V->getType()->isVectorTy() && match(V, PatternMatch::m_Constant(CV))) {
124 unsigned NumElts = V->getType()->getVectorNumElements();
125 for (unsigned i = 0; i != NumElts; ++i) {
126 Constant *Elt = CV->getAggregateElement(i);
130 if (isa<UndefValue>(Elt))
133 if (!isa<ConstantInt>(Elt))
139 // Compares can be inverted if all of their uses are being modified to use the
142 return WillInvertAllUses;
144 // If `V` is of the form `A + Constant` then `-1 - V` can be folded into `(-1
145 // - Constant) - A` if we are willing to invert all of the uses.
146 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(V))
147 if (BO->getOpcode() == Instruction::Add ||
148 BO->getOpcode() == Instruction::Sub)
149 if (isa<Constant>(BO->getOperand(0)) || isa<Constant>(BO->getOperand(1)))
150 return WillInvertAllUses;
156 /// \brief Specific patterns of overflow check idioms that we match.
157 enum OverflowCheckFlavor {
168 /// \brief Returns the OverflowCheckFlavor corresponding to a overflow_with_op
170 static inline OverflowCheckFlavor
171 IntrinsicIDToOverflowCheckFlavor(unsigned ID) {
175 case Intrinsic::uadd_with_overflow:
176 return OCF_UNSIGNED_ADD;
177 case Intrinsic::sadd_with_overflow:
178 return OCF_SIGNED_ADD;
179 case Intrinsic::usub_with_overflow:
180 return OCF_UNSIGNED_SUB;
181 case Intrinsic::ssub_with_overflow:
182 return OCF_SIGNED_SUB;
183 case Intrinsic::umul_with_overflow:
184 return OCF_UNSIGNED_MUL;
185 case Intrinsic::smul_with_overflow:
186 return OCF_SIGNED_MUL;
190 /// \brief The core instruction combiner logic.
192 /// This class provides both the logic to recursively visit instructions and
194 class LLVM_LIBRARY_VISIBILITY InstCombiner
195 : public InstVisitor<InstCombiner, Instruction *> {
196 // FIXME: These members shouldn't be public.
198 /// \brief A worklist of the instructions that need to be simplified.
199 InstCombineWorklist &Worklist;
201 /// \brief An IRBuilder that automatically inserts new instructions into the
203 typedef IRBuilder<TargetFolder, IRBuilderCallbackInserter> BuilderTy;
207 // Mode in which we are running the combiner.
208 const bool MinimizeSize;
209 /// Enable combines that trigger rarely but are costly in compiletime.
210 const bool ExpensiveCombines;
214 // Required analyses.
216 TargetLibraryInfo &TLI;
218 const DataLayout &DL;
219 const SimplifyQuery SQ;
220 // Optional analyses. When non-null, these can both be used to do better
221 // combining and will be updated to reflect any changes.
227 InstCombiner(InstCombineWorklist &Worklist, BuilderTy *Builder,
228 bool MinimizeSize, bool ExpensiveCombines, AliasAnalysis *AA,
229 AssumptionCache &AC, TargetLibraryInfo &TLI, DominatorTree &DT,
230 const DataLayout &DL, LoopInfo *LI)
231 : Worklist(Worklist), Builder(Builder), MinimizeSize(MinimizeSize),
232 ExpensiveCombines(ExpensiveCombines), AA(AA), AC(AC), TLI(TLI), DT(DT),
233 DL(DL), SQ(DL, &TLI, &DT, &AC), LI(LI), MadeIRChange(false) {}
235 /// \brief Run the combiner over the entire worklist until it is empty.
237 /// \returns true if the IR is changed.
240 AssumptionCache &getAssumptionCache() const { return AC; }
242 const DataLayout &getDataLayout() const { return DL; }
244 DominatorTree &getDominatorTree() const { return DT; }
246 LoopInfo *getLoopInfo() const { return LI; }
248 TargetLibraryInfo &getTargetLibraryInfo() const { return TLI; }
250 // Visitation implementation - Implement instruction combining for different
251 // instruction types. The semantics are as follows:
253 // null - No change was made
254 // I - Change was made, I is still valid, I may be dead though
255 // otherwise - Change was made, replace I with returned instruction
257 Instruction *visitAdd(BinaryOperator &I);
258 Instruction *visitFAdd(BinaryOperator &I);
259 Value *OptimizePointerDifference(Value *LHS, Value *RHS, Type *Ty);
260 Instruction *visitSub(BinaryOperator &I);
261 Instruction *visitFSub(BinaryOperator &I);
262 Instruction *visitMul(BinaryOperator &I);
263 Value *foldFMulConst(Instruction *FMulOrDiv, Constant *C,
264 Instruction *InsertBefore);
265 Instruction *visitFMul(BinaryOperator &I);
266 Instruction *visitURem(BinaryOperator &I);
267 Instruction *visitSRem(BinaryOperator &I);
268 Instruction *visitFRem(BinaryOperator &I);
269 bool SimplifyDivRemOfSelect(BinaryOperator &I);
270 Instruction *commonRemTransforms(BinaryOperator &I);
271 Instruction *commonIRemTransforms(BinaryOperator &I);
272 Instruction *commonDivTransforms(BinaryOperator &I);
273 Instruction *commonIDivTransforms(BinaryOperator &I);
274 Instruction *visitUDiv(BinaryOperator &I);
275 Instruction *visitSDiv(BinaryOperator &I);
276 Instruction *visitFDiv(BinaryOperator &I);
277 Value *simplifyRangeCheck(ICmpInst *Cmp0, ICmpInst *Cmp1, bool Inverted);
278 Instruction *visitAnd(BinaryOperator &I);
279 Instruction *FoldOrWithConstants(BinaryOperator &I, Value *Op, Value *A,
281 Instruction *FoldXorWithConstants(BinaryOperator &I, Value *Op, Value *A,
283 Instruction *visitOr(BinaryOperator &I);
284 Instruction *visitXor(BinaryOperator &I);
285 Instruction *visitShl(BinaryOperator &I);
286 Instruction *visitAShr(BinaryOperator &I);
287 Instruction *visitLShr(BinaryOperator &I);
288 Instruction *commonShiftTransforms(BinaryOperator &I);
289 Instruction *visitFCmpInst(FCmpInst &I);
290 Instruction *visitICmpInst(ICmpInst &I);
291 Instruction *FoldShiftByConstant(Value *Op0, Constant *Op1,
293 Instruction *commonCastTransforms(CastInst &CI);
294 Instruction *commonPointerCastTransforms(CastInst &CI);
295 Instruction *visitTrunc(TruncInst &CI);
296 Instruction *visitZExt(ZExtInst &CI);
297 Instruction *visitSExt(SExtInst &CI);
298 Instruction *visitFPTrunc(FPTruncInst &CI);
299 Instruction *visitFPExt(CastInst &CI);
300 Instruction *visitFPToUI(FPToUIInst &FI);
301 Instruction *visitFPToSI(FPToSIInst &FI);
302 Instruction *visitUIToFP(CastInst &CI);
303 Instruction *visitSIToFP(CastInst &CI);
304 Instruction *visitPtrToInt(PtrToIntInst &CI);
305 Instruction *visitIntToPtr(IntToPtrInst &CI);
306 Instruction *visitBitCast(BitCastInst &CI);
307 Instruction *visitAddrSpaceCast(AddrSpaceCastInst &CI);
308 Instruction *FoldItoFPtoI(Instruction &FI);
309 Instruction *visitSelectInst(SelectInst &SI);
310 Instruction *visitCallInst(CallInst &CI);
311 Instruction *visitInvokeInst(InvokeInst &II);
313 Instruction *SliceUpIllegalIntegerPHI(PHINode &PN);
314 Instruction *visitPHINode(PHINode &PN);
315 Instruction *visitGetElementPtrInst(GetElementPtrInst &GEP);
316 Instruction *visitAllocaInst(AllocaInst &AI);
317 Instruction *visitAllocSite(Instruction &FI);
318 Instruction *visitFree(CallInst &FI);
319 Instruction *visitLoadInst(LoadInst &LI);
320 Instruction *visitStoreInst(StoreInst &SI);
321 Instruction *visitBranchInst(BranchInst &BI);
322 Instruction *visitFenceInst(FenceInst &FI);
323 Instruction *visitSwitchInst(SwitchInst &SI);
324 Instruction *visitReturnInst(ReturnInst &RI);
325 Instruction *visitInsertValueInst(InsertValueInst &IV);
326 Instruction *visitInsertElementInst(InsertElementInst &IE);
327 Instruction *visitExtractElementInst(ExtractElementInst &EI);
328 Instruction *visitShuffleVectorInst(ShuffleVectorInst &SVI);
329 Instruction *visitExtractValueInst(ExtractValueInst &EV);
330 Instruction *visitLandingPadInst(LandingPadInst &LI);
331 Instruction *visitVAStartInst(VAStartInst &I);
332 Instruction *visitVACopyInst(VACopyInst &I);
334 /// Specify what to return for unhandled instructions.
335 Instruction *visitInstruction(Instruction &I) { return nullptr; }
337 /// True when DB dominates all uses of DI except UI.
338 /// UI must be in the same block as DI.
339 /// The routine checks that the DI parent and DB are different.
340 bool dominatesAllUses(const Instruction *DI, const Instruction *UI,
341 const BasicBlock *DB) const;
343 /// Try to replace select with select operand SIOpd in SI-ICmp sequence.
344 bool replacedSelectWithOperand(SelectInst *SI, const ICmpInst *Icmp,
345 const unsigned SIOpd);
347 /// Try to replace instruction \p I with value \p V which are pointers
348 /// in different address space.
349 /// \return true if successful.
350 bool replacePointer(Instruction &I, Value *V);
353 bool shouldChangeType(unsigned FromBitWidth, unsigned ToBitWidth) const;
354 bool shouldChangeType(Type *From, Type *To) const;
355 Value *dyn_castNegVal(Value *V) const;
356 Value *dyn_castFNegVal(Value *V, bool NoSignedZero = false) const;
357 Type *FindElementAtOffset(PointerType *PtrTy, int64_t Offset,
358 SmallVectorImpl<Value *> &NewIndices);
360 /// Classify whether a cast is worth optimizing.
362 /// This is a helper to decide whether the simplification of
363 /// logic(cast(A), cast(B)) to cast(logic(A, B)) should be performed.
365 /// \param CI The cast we are interested in.
367 /// \return true if this cast actually results in any code being generated and
368 /// if it cannot already be eliminated by some other transformation.
369 bool shouldOptimizeCast(CastInst *CI);
371 /// \brief Try to optimize a sequence of instructions checking if an operation
372 /// on LHS and RHS overflows.
374 /// If this overflow check is done via one of the overflow check intrinsics,
375 /// then CtxI has to be the call instruction calling that intrinsic. If this
376 /// overflow check is done by arithmetic followed by a compare, then CtxI has
377 /// to be the arithmetic instruction.
379 /// If a simplification is possible, stores the simplified result of the
380 /// operation in OperationResult and result of the overflow check in
381 /// OverflowResult, and return true. If no simplification is possible,
383 bool OptimizeOverflowCheck(OverflowCheckFlavor OCF, Value *LHS, Value *RHS,
384 Instruction &CtxI, Value *&OperationResult,
385 Constant *&OverflowResult);
387 Instruction *visitCallSite(CallSite CS);
388 Instruction *tryOptimizeCall(CallInst *CI);
389 bool transformConstExprCastCall(CallSite CS);
390 Instruction *transformCallThroughTrampoline(CallSite CS,
391 IntrinsicInst *Tramp);
393 /// Transform (zext icmp) to bitwise / integer operations in order to
396 /// \param ICI The icmp of the (zext icmp) pair we are interested in.
397 /// \parem CI The zext of the (zext icmp) pair we are interested in.
398 /// \param DoTransform Pass false to just test whether the given (zext icmp)
399 /// would be transformed. Pass true to actually perform the transformation.
401 /// \return null if the transformation cannot be performed. If the
402 /// transformation can be performed the new instruction that replaces the
403 /// (zext icmp) pair will be returned (if \p DoTransform is false the
404 /// unmodified \p ICI will be returned in this case).
405 Instruction *transformZExtICmp(ICmpInst *ICI, ZExtInst &CI,
406 bool DoTransform = true);
408 Instruction *transformSExtICmp(ICmpInst *ICI, Instruction &CI);
409 bool willNotOverflowSignedAdd(const Value *LHS, const Value *RHS,
410 const Instruction &CxtI) const {
411 return computeOverflowForSignedAdd(LHS, RHS, &CxtI) ==
412 OverflowResult::NeverOverflows;
414 bool willNotOverflowUnsignedAdd(const Value *LHS, const Value *RHS,
415 const Instruction &CxtI) const {
416 return computeOverflowForUnsignedAdd(LHS, RHS, &CxtI) ==
417 OverflowResult::NeverOverflows;
419 bool willNotOverflowSignedSub(const Value *LHS, const Value *RHS,
420 const Instruction &CxtI) const;
421 bool willNotOverflowUnsignedSub(const Value *LHS, const Value *RHS,
422 const Instruction &CxtI) const;
423 bool willNotOverflowSignedMul(const Value *LHS, const Value *RHS,
424 const Instruction &CxtI) const;
425 bool willNotOverflowUnsignedMul(const Value *LHS, const Value *RHS,
426 const Instruction &CxtI) const {
427 return computeOverflowForUnsignedMul(LHS, RHS, &CxtI) ==
428 OverflowResult::NeverOverflows;
430 Value *EmitGEPOffset(User *GEP);
431 Instruction *scalarizePHI(ExtractElementInst &EI, PHINode *PN);
432 Value *EvaluateInDifferentElementOrder(Value *V, ArrayRef<int> Mask);
433 Instruction *foldCastedBitwiseLogic(BinaryOperator &I);
434 Instruction *shrinkBitwiseLogic(TruncInst &Trunc);
435 Instruction *optimizeBitCastFromPhi(CastInst &CI, PHINode *PN);
437 /// Determine if a pair of casts can be replaced by a single cast.
439 /// \param CI1 The first of a pair of casts.
440 /// \param CI2 The second of a pair of casts.
442 /// \return 0 if the cast pair cannot be eliminated, otherwise returns an
443 /// Instruction::CastOps value for a cast that can replace the pair, casting
444 /// CI1->getSrcTy() to CI2->getDstTy().
446 /// \see CastInst::isEliminableCastPair
447 Instruction::CastOps isEliminableCastPair(const CastInst *CI1,
448 const CastInst *CI2);
450 Value *foldAndOfICmps(ICmpInst *LHS, ICmpInst *RHS);
451 Value *foldAndOfFCmps(FCmpInst *LHS, FCmpInst *RHS);
452 Value *foldOrOfICmps(ICmpInst *LHS, ICmpInst *RHS, Instruction *CxtI);
453 Value *foldOrOfFCmps(FCmpInst *LHS, FCmpInst *RHS);
454 Value *foldXorOfICmps(ICmpInst *LHS, ICmpInst *RHS);
457 /// \brief Inserts an instruction \p New before instruction \p Old
459 /// Also adds the new instruction to the worklist and returns \p New so that
460 /// it is suitable for use as the return from the visitation patterns.
461 Instruction *InsertNewInstBefore(Instruction *New, Instruction &Old) {
462 assert(New && !New->getParent() &&
463 "New instruction already inserted into a basic block!");
464 BasicBlock *BB = Old.getParent();
465 BB->getInstList().insert(Old.getIterator(), New); // Insert inst
470 /// \brief Same as InsertNewInstBefore, but also sets the debug loc.
471 Instruction *InsertNewInstWith(Instruction *New, Instruction &Old) {
472 New->setDebugLoc(Old.getDebugLoc());
473 return InsertNewInstBefore(New, Old);
476 /// \brief A combiner-aware RAUW-like routine.
478 /// This method is to be used when an instruction is found to be dead,
479 /// replaceable with another preexisting expression. Here we add all uses of
480 /// I to the worklist, replace all uses of I with the new value, then return
481 /// I, so that the inst combiner will know that I was modified.
482 Instruction *replaceInstUsesWith(Instruction &I, Value *V) {
483 // If there are no uses to replace, then we return nullptr to indicate that
484 // no changes were made to the program.
485 if (I.use_empty()) return nullptr;
487 Worklist.AddUsersToWorkList(I); // Add all modified instrs to worklist.
489 // If we are replacing the instruction with itself, this must be in a
490 // segment of unreachable code, so just clobber the instruction.
492 V = UndefValue::get(I.getType());
494 DEBUG(dbgs() << "IC: Replacing " << I << "\n"
495 << " with " << *V << '\n');
497 I.replaceAllUsesWith(V);
501 /// Creates a result tuple for an overflow intrinsic \p II with a given
502 /// \p Result and a constant \p Overflow value.
503 Instruction *CreateOverflowTuple(IntrinsicInst *II, Value *Result,
504 Constant *Overflow) {
505 Constant *V[] = {UndefValue::get(Result->getType()), Overflow};
506 StructType *ST = cast<StructType>(II->getType());
507 Constant *Struct = ConstantStruct::get(ST, V);
508 return InsertValueInst::Create(Struct, Result, 0);
511 /// \brief Combiner aware instruction erasure.
513 /// When dealing with an instruction that has side effects or produces a void
514 /// value, we can't rely on DCE to delete the instruction. Instead, visit
515 /// methods should return the value returned by this function.
516 Instruction *eraseInstFromFunction(Instruction &I) {
517 DEBUG(dbgs() << "IC: ERASE " << I << '\n');
518 assert(I.use_empty() && "Cannot erase instruction that is used!");
521 // Make sure that we reprocess all operands now that we reduced their
523 if (I.getNumOperands() < 8) {
524 for (Use &Operand : I.operands())
525 if (auto *Inst = dyn_cast<Instruction>(Operand))
531 return nullptr; // Don't do anything with FI
534 void computeKnownBits(const Value *V, KnownBits &Known,
535 unsigned Depth, const Instruction *CxtI) const {
536 llvm::computeKnownBits(V, Known, DL, Depth, &AC, CxtI, &DT);
538 KnownBits computeKnownBits(const Value *V, unsigned Depth,
539 const Instruction *CxtI) const {
540 return llvm::computeKnownBits(V, DL, Depth, &AC, CxtI, &DT);
543 bool isKnownToBeAPowerOfTwo(const Value *V, bool OrZero = false,
545 const Instruction *CxtI = nullptr) {
546 return llvm::isKnownToBeAPowerOfTwo(V, DL, OrZero, Depth, &AC, CxtI, &DT);
549 bool MaskedValueIsZero(const Value *V, const APInt &Mask, unsigned Depth = 0,
550 const Instruction *CxtI = nullptr) const {
551 return llvm::MaskedValueIsZero(V, Mask, DL, Depth, &AC, CxtI, &DT);
553 unsigned ComputeNumSignBits(const Value *Op, unsigned Depth = 0,
554 const Instruction *CxtI = nullptr) const {
555 return llvm::ComputeNumSignBits(Op, DL, Depth, &AC, CxtI, &DT);
557 OverflowResult computeOverflowForUnsignedMul(const Value *LHS,
559 const Instruction *CxtI) const {
560 return llvm::computeOverflowForUnsignedMul(LHS, RHS, DL, &AC, CxtI, &DT);
562 OverflowResult computeOverflowForUnsignedAdd(const Value *LHS,
564 const Instruction *CxtI) const {
565 return llvm::computeOverflowForUnsignedAdd(LHS, RHS, DL, &AC, CxtI, &DT);
567 OverflowResult computeOverflowForSignedAdd(const Value *LHS,
569 const Instruction *CxtI) const {
570 return llvm::computeOverflowForSignedAdd(LHS, RHS, DL, &AC, CxtI, &DT);
573 /// Maximum size of array considered when transforming.
574 uint64_t MaxArraySizeForCombine;
577 /// \brief Performs a few simplifications for operators which are associative
579 bool SimplifyAssociativeOrCommutative(BinaryOperator &I);
581 /// \brief Tries to simplify binary operations which some other binary
582 /// operation distributes over.
584 /// It does this by either by factorizing out common terms (eg "(A*B)+(A*C)"
585 /// -> "A*(B+C)") or expanding out if this results in simplifications (eg: "A
586 /// & (B | C) -> (A&B) | (A&C)" if this is a win). Returns the simplified
587 /// value, or null if it didn't simplify.
588 Value *SimplifyUsingDistributiveLaws(BinaryOperator &I);
590 /// This tries to simplify binary operations by factorizing out common terms
591 /// (e. g. "(A*B)+(A*C)" -> "A*(B+C)").
592 Value *tryFactorization(InstCombiner::BuilderTy *, BinaryOperator &,
593 Instruction::BinaryOps, Value *, Value *, Value *,
596 /// \brief Attempts to replace V with a simpler value based on the demanded
598 Value *SimplifyDemandedUseBits(Value *V, APInt DemandedMask, KnownBits &Known,
599 unsigned Depth, Instruction *CxtI);
600 bool SimplifyDemandedBits(Instruction *I, unsigned Op,
601 const APInt &DemandedMask, KnownBits &Known,
603 /// Helper routine of SimplifyDemandedUseBits. It computes KnownZero/KnownOne
604 /// bits. It also tries to handle simplifications that can be done based on
605 /// DemandedMask, but without modifying the Instruction.
606 Value *SimplifyMultipleUseDemandedBits(Instruction *I,
607 const APInt &DemandedMask,
609 unsigned Depth, Instruction *CxtI);
610 /// Helper routine of SimplifyDemandedUseBits. It tries to simplify demanded
611 /// bit for "r1 = shr x, c1; r2 = shl r1, c2" instruction sequence.
612 Value *simplifyShrShlDemandedBits(
613 Instruction *Shr, const APInt &ShrOp1, Instruction *Shl,
614 const APInt &ShlOp1, const APInt &DemandedMask, KnownBits &Known);
616 /// \brief Tries to simplify operands to an integer instruction based on its
618 bool SimplifyDemandedInstructionBits(Instruction &Inst);
620 Value *SimplifyDemandedVectorElts(Value *V, APInt DemandedElts,
621 APInt &UndefElts, unsigned Depth = 0);
623 Value *SimplifyVectorOp(BinaryOperator &Inst);
624 Value *SimplifyBSwap(BinaryOperator &Inst);
627 /// Given a binary operator, cast instruction, or select which has a PHI node
628 /// as operand #0, see if we can fold the instruction into the PHI (which is
629 /// only possible if all operands to the PHI are constants).
630 Instruction *foldOpIntoPhi(Instruction &I, PHINode *PN);
632 /// Given an instruction with a select as one operand and a constant as the
633 /// other operand, try to fold the binary operator into the select arguments.
634 /// This also works for Cast instructions, which obviously do not have a
636 Instruction *FoldOpIntoSelect(Instruction &Op, SelectInst *SI);
638 /// This is a convenience wrapper function for the above two functions.
639 Instruction *foldOpWithConstantIntoOperand(BinaryOperator &I);
641 /// \brief Try to rotate an operation below a PHI node, using PHI nodes for
643 Instruction *FoldPHIArgOpIntoPHI(PHINode &PN);
644 Instruction *FoldPHIArgBinOpIntoPHI(PHINode &PN);
645 Instruction *FoldPHIArgGEPIntoPHI(PHINode &PN);
646 Instruction *FoldPHIArgLoadIntoPHI(PHINode &PN);
647 Instruction *FoldPHIArgZextsIntoPHI(PHINode &PN);
649 /// Helper function for FoldPHIArgXIntoPHI() to get debug location for the
650 /// folded operation.
651 DebugLoc PHIArgMergedDebugLoc(PHINode &PN);
653 Instruction *foldGEPICmp(GEPOperator *GEPLHS, Value *RHS,
654 ICmpInst::Predicate Cond, Instruction &I);
655 Instruction *foldAllocaCmp(ICmpInst &ICI, const AllocaInst *Alloca,
657 Instruction *foldCmpLoadFromIndexedGlobal(GetElementPtrInst *GEP,
658 GlobalVariable *GV, CmpInst &ICI,
659 ConstantInt *AndCst = nullptr);
660 Instruction *foldFCmpIntToFPConst(FCmpInst &I, Instruction *LHSI,
662 Instruction *foldICmpAddOpConst(Instruction &ICI, Value *X, ConstantInt *CI,
663 ICmpInst::Predicate Pred);
664 Instruction *foldICmpWithCastAndCast(ICmpInst &ICI);
666 Instruction *foldICmpUsingKnownBits(ICmpInst &Cmp);
667 Instruction *foldICmpWithConstant(ICmpInst &Cmp);
668 Instruction *foldICmpInstWithConstant(ICmpInst &Cmp);
669 Instruction *foldICmpInstWithConstantNotInt(ICmpInst &Cmp);
670 Instruction *foldICmpBinOp(ICmpInst &Cmp);
671 Instruction *foldICmpEquality(ICmpInst &Cmp);
673 Instruction *foldICmpTruncConstant(ICmpInst &Cmp, Instruction *Trunc,
675 Instruction *foldICmpAndConstant(ICmpInst &Cmp, BinaryOperator *And,
677 Instruction *foldICmpXorConstant(ICmpInst &Cmp, BinaryOperator *Xor,
679 Instruction *foldICmpOrConstant(ICmpInst &Cmp, BinaryOperator *Or,
681 Instruction *foldICmpMulConstant(ICmpInst &Cmp, BinaryOperator *Mul,
683 Instruction *foldICmpShlConstant(ICmpInst &Cmp, BinaryOperator *Shl,
685 Instruction *foldICmpShrConstant(ICmpInst &Cmp, BinaryOperator *Shr,
687 Instruction *foldICmpUDivConstant(ICmpInst &Cmp, BinaryOperator *UDiv,
689 Instruction *foldICmpDivConstant(ICmpInst &Cmp, BinaryOperator *Div,
691 Instruction *foldICmpSubConstant(ICmpInst &Cmp, BinaryOperator *Sub,
693 Instruction *foldICmpAddConstant(ICmpInst &Cmp, BinaryOperator *Add,
695 Instruction *foldICmpAndConstConst(ICmpInst &Cmp, BinaryOperator *And,
697 Instruction *foldICmpAndShift(ICmpInst &Cmp, BinaryOperator *And,
698 const APInt *C1, const APInt *C2);
699 Instruction *foldICmpShrConstConst(ICmpInst &I, Value *ShAmt, const APInt &C1,
701 Instruction *foldICmpShlConstConst(ICmpInst &I, Value *ShAmt, const APInt &C1,
704 Instruction *foldICmpBinOpEqualityWithConstant(ICmpInst &Cmp,
707 Instruction *foldICmpIntrinsicWithConstant(ICmpInst &ICI, const APInt *C);
709 // Helpers of visitSelectInst().
710 Instruction *foldSelectExtConst(SelectInst &Sel);
711 Instruction *foldSelectOpOp(SelectInst &SI, Instruction *TI, Instruction *FI);
712 Instruction *foldSelectIntoOp(SelectInst &SI, Value *, Value *);
713 Instruction *foldSPFofSPF(Instruction *Inner, SelectPatternFlavor SPF1,
714 Value *A, Value *B, Instruction &Outer,
715 SelectPatternFlavor SPF2, Value *C);
716 Instruction *foldSelectInstWithICmp(SelectInst &SI, ICmpInst *ICI);
718 Instruction *OptAndOp(BinaryOperator *Op, ConstantInt *OpRHS,
719 ConstantInt *AndRHS, BinaryOperator &TheAnd);
721 Value *insertRangeTest(Value *V, const APInt &Lo, const APInt &Hi,
722 bool isSigned, bool Inside);
723 Instruction *PromoteCastOfAllocation(BitCastInst &CI, AllocaInst &AI);
724 Instruction *MatchBSwap(BinaryOperator &I);
725 bool SimplifyStoreAtEndOfBlock(StoreInst &SI);
727 Instruction *SimplifyElementAtomicMemCpy(ElementAtomicMemCpyInst *AMI);
728 Instruction *SimplifyMemTransfer(MemIntrinsic *MI);
729 Instruction *SimplifyMemSet(MemSetInst *MI);
731 Value *EvaluateInDifferentType(Value *V, Type *Ty, bool isSigned);
733 /// \brief Returns a value X such that Val = X * Scale, or null if none.
735 /// If the multiplication is known not to overflow then NoSignedWrap is set.
736 Value *Descale(Value *Val, APInt Scale, bool &NoSignedWrap);
739 } // end namespace llvm.