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/PostOrderIterator.h"
19#include "llvm/ADT/Statistic.h"
20#include "llvm/Analysis/InstructionSimplify.h"
21#include "llvm/Analysis/TargetFolder.h"
22#include "llvm/Analysis/ValueTracking.h"
23#include "llvm/IR/IRBuilder.h"
24#include "llvm/IR/InstVisitor.h"
25#include "llvm/IR/PatternMatch.h"
26#include "llvm/IR/ProfDataUtils.h"
27#include "llvm/IR/Value.h"
28#include "llvm/Support/Debug.h"
29#include "llvm/Support/KnownBits.h"
30#include "llvm/Support/KnownFPClass.h"
31#include "llvm/Transforms/InstCombine/InstCombiner.h"
32#include "llvm/Transforms/Utils/Local.h"
33#include <cassert>
34
35#define DEBUG_TYPE "instcombine"
36#include "llvm/Transforms/Utils/InstructionWorklist.h"
37
38// As a default, let's assume that we want to be aggressive,
39// and attempt to traverse with no limits in attempt to sink negation.
40static constexpr unsigned NegatorDefaultMaxDepth = ~0U;
41
42// Let's guesstimate that most often we will end up visiting/producing
43// fairly small number of new instructions.
44static constexpr unsigned NegatorMaxNodesSSO = 16;
45
46namespace llvm {
47
48class AAResults;
49class APInt;
50class AssumptionCache;
51class BlockFrequencyInfo;
52class DataLayout;
53class DominatorTree;
54class GEPOperator;
55class GlobalVariable;
56class OptimizationRemarkEmitter;
57class ProfileSummaryInfo;
58class TargetLibraryInfo;
59class User;
60
61class LLVM_LIBRARY_VISIBILITY InstCombinerImpl final
62 : public InstCombiner,
63 public InstVisitor<InstCombinerImpl, Instruction *> {
64public:
65 InstCombinerImpl(InstructionWorklist &Worklist, BuilderTy &Builder,
66 Function &F, AAResults *AA, AssumptionCache &AC,
67 TargetLibraryInfo &TLI, TargetTransformInfo &TTI,
68 DominatorTree &DT, OptimizationRemarkEmitter &ORE,
69 BlockFrequencyInfo *BFI, BranchProbabilityInfo *BPI,
70 ProfileSummaryInfo *PSI, const DataLayout &DL,
71 ReversePostOrderTraversal<BasicBlock *> &RPOT)
72 : InstCombiner(Worklist, Builder, F, AA, AC, TLI, TTI, DT, ORE, BFI, BPI,
73 PSI, DL, RPOT) {}
74
75 ~InstCombinerImpl() override = default;
76
77 /// Perform early cleanup and prepare the InstCombine worklist.
78 bool prepareWorklist(Function &F);
79
80 /// Run the combiner over the entire worklist until it is empty.
81 ///
82 /// \returns true if the IR is changed.
83 bool run();
84
85 // Visitation implementation - Implement instruction combining for different
86 // instruction types. The semantics are as follows:
87 // Return Value:
88 // null - No change was made
89 // I - Change was made, I is still valid, I may be dead though
90 // otherwise - Change was made, replace I with returned instruction
91 //
92 Instruction *visitFNeg(UnaryOperator &I);
93 Instruction *visitAdd(BinaryOperator &I);
94 Instruction *visitFAdd(BinaryOperator &I);
95 Value *OptimizePointerDifference(
96 Value *LHS, Value *RHS, Type *Ty, bool isNUW);
97 Instruction *visitSub(BinaryOperator &I);
98 Instruction *visitFSub(BinaryOperator &I);
99 Instruction *visitMul(BinaryOperator &I);
100 Instruction *foldPowiReassoc(BinaryOperator &I);
101 Instruction *foldFMulReassoc(BinaryOperator &I);
102 Instruction *visitFMul(BinaryOperator &I);
103 Instruction *visitURem(BinaryOperator &I);
104 Instruction *visitSRem(BinaryOperator &I);
105 Instruction *visitFRem(BinaryOperator &I);
106 bool simplifyDivRemOfSelectWithZeroOp(BinaryOperator &I);
107 Instruction *commonIDivRemTransforms(BinaryOperator &I);
108 Instruction *commonIRemTransforms(BinaryOperator &I);
109 Instruction *commonIDivTransforms(BinaryOperator &I);
110 Instruction *visitUDiv(BinaryOperator &I);
111 Instruction *visitSDiv(BinaryOperator &I);
112 Instruction *visitFDiv(BinaryOperator &I);
113 Value *simplifyRangeCheck(ICmpInst *Cmp0, ICmpInst *Cmp1, bool Inverted);
114 Instruction *FoldOrOfLogicalAnds(Value *Op0, Value *Op1);
115 Instruction *visitAnd(BinaryOperator &I);
116 Instruction *visitOr(BinaryOperator &I);
117 bool sinkNotIntoLogicalOp(Instruction &I);
118 bool sinkNotIntoOtherHandOfLogicalOp(Instruction &I);
119 Instruction *visitXor(BinaryOperator &I);
120 Instruction *visitShl(BinaryOperator &I);
121 Value *reassociateShiftAmtsOfTwoSameDirectionShifts(
122 BinaryOperator *Sh0, const SimplifyQuery &SQ,
123 bool AnalyzeForSignBitExtraction = false);
124 Instruction *canonicalizeCondSignextOfHighBitExtractToSignextHighBitExtract(
125 BinaryOperator &I);
126 Instruction *foldVariableSignZeroExtensionOfVariableHighBitExtract(
127 BinaryOperator &OldAShr);
128 Instruction *visitAShr(BinaryOperator &I);
129 Instruction *visitLShr(BinaryOperator &I);
130 Instruction *commonShiftTransforms(BinaryOperator &I);
131 Instruction *visitFCmpInst(FCmpInst &I);
132 CmpInst *canonicalizeICmpPredicate(CmpInst &I);
133 Instruction *visitICmpInst(ICmpInst &I);
134 Instruction *FoldShiftByConstant(Value *Op0, Constant *Op1,
135 BinaryOperator &I);
136 Instruction *commonCastTransforms(CastInst &CI);
137 Instruction *visitTrunc(TruncInst &CI);
138 Instruction *visitZExt(ZExtInst &Zext);
139 Instruction *visitSExt(SExtInst &Sext);
140 Instruction *visitFPTrunc(FPTruncInst &CI);
141 Instruction *visitFPExt(CastInst &CI);
142 Instruction *visitFPToUI(FPToUIInst &FI);
143 Instruction *visitFPToSI(FPToSIInst &FI);
144 Instruction *visitUIToFP(CastInst &CI);
145 Instruction *visitSIToFP(CastInst &CI);
146 Instruction *visitPtrToInt(PtrToIntInst &CI);
147 Instruction *visitPtrToAddr(PtrToAddrInst &CI);
148 Instruction *visitIntToPtr(IntToPtrInst &CI);
149 Instruction *visitBitCast(BitCastInst &CI);
150 Instruction *visitAddrSpaceCast(AddrSpaceCastInst &CI);
151 Instruction *foldItoFPtoI(CastInst &FI);
152 Instruction *visitSelectInst(SelectInst &SI);
153 Instruction *foldShuffledIntrinsicOperands(IntrinsicInst *II);
154 Value *foldReversedIntrinsicOperands(IntrinsicInst *II);
155 Instruction *visitCallInst(CallInst &CI);
156 Instruction *visitInvokeInst(InvokeInst &II);
157 Instruction *visitCallBrInst(CallBrInst &CBI);
158
159 Instruction *SliceUpIllegalIntegerPHI(PHINode &PN);
160 Instruction *visitPHINode(PHINode &PN);
161 Instruction *visitGetElementPtrInst(GetElementPtrInst &GEP);
162 Instruction *visitGEPOfGEP(GetElementPtrInst &GEP, GEPOperator *Src);
163 Instruction *visitAllocaInst(AllocaInst &AI);
164 Instruction *visitAllocSite(Instruction &FI);
165 Instruction *visitFree(CallInst &FI, Value *FreedOp);
166 Instruction *visitLoadInst(LoadInst &LI);
167 Instruction *visitStoreInst(StoreInst &SI);
168 Instruction *visitAtomicRMWInst(AtomicRMWInst &SI);
169 Instruction *visitUncondBrInst(UncondBrInst &BI);
170 Instruction *visitCondBrInst(CondBrInst &BI);
171 Instruction *visitFenceInst(FenceInst &FI);
172 Instruction *visitSwitchInst(SwitchInst &SI);
173 Instruction *visitReturnInst(ReturnInst &RI);
174 Instruction *visitUnreachableInst(UnreachableInst &I);
175 Instruction *
176 foldAggregateConstructionIntoAggregateReuse(InsertValueInst &OrigIVI);
177 Instruction *visitInsertValueInst(InsertValueInst &IV);
178 Instruction *visitInsertElementInst(InsertElementInst &IE);
179 Instruction *visitExtractElementInst(ExtractElementInst &EI);
180 Instruction *simplifyBinOpSplats(ShuffleVectorInst &SVI);
181 Instruction *visitShuffleVectorInst(ShuffleVectorInst &SVI);
182 Instruction *visitExtractValueInst(ExtractValueInst &EV);
183 Instruction *visitLandingPadInst(LandingPadInst &LI);
184 Instruction *visitVAEndInst(VAEndInst &I);
185 Value *pushFreezeToPreventPoisonFromPropagating(FreezeInst &FI);
186 bool freezeOtherUses(FreezeInst &FI);
187 Instruction *foldFreezeIntoRecurrence(FreezeInst &I, PHINode *PN);
188 Instruction *visitFreeze(FreezeInst &I);
189
190 /// Specify what to return for unhandled instructions.
191 Instruction *visitInstruction(Instruction &I) { return nullptr; }
192
193 /// True when DB dominates all uses of DI except UI.
194 /// UI must be in the same block as DI.
195 /// The routine checks that the DI parent and DB are different.
196 bool dominatesAllUses(const Instruction *DI, const Instruction *UI,
197 const BasicBlock *DB) const;
198
199 /// Try to replace select with select operand SIOpd in SI-ICmp sequence.
200 bool replacedSelectWithOperand(SelectInst *SI, const ICmpInst *Icmp,
201 const unsigned SIOpd);
202
203 LoadInst *combineLoadToNewType(LoadInst &LI, Type *NewTy,
204 const Twine &Suffix = "");
205
206 /// Check if fmul \p MulVal, +0.0 will yield +0.0 (or signed zero is
207 /// ignorable).
208 bool fmulByZeroIsZero(Value *MulVal, FastMathFlags FMF,
209 const Instruction *CtxI) const;
210
211 std::optional<std::pair<Intrinsic::ID, SmallVector<Value *, 3>>>
212 convertOrOfShiftsToFunnelShift(Instruction &Or);
213
214private:
215 bool annotateAnyAllocSite(CallBase &Call, const TargetLibraryInfo *TLI);
216 bool isDesirableIntType(unsigned BitWidth) const;
217 bool shouldChangeType(unsigned FromBitWidth, unsigned ToBitWidth) const;
218 bool shouldChangeType(Type *From, Type *To) const;
219 Value *dyn_castNegVal(Value *V) const;
220
221 /// Classify whether a cast is worth optimizing.
222 ///
223 /// This is a helper to decide whether the simplification of
224 /// logic(cast(A), cast(B)) to cast(logic(A, B)) should be performed.
225 ///
226 /// \param CI The cast we are interested in.
227 ///
228 /// \return true if this cast actually results in any code being generated and
229 /// if it cannot already be eliminated by some other transformation.
230 bool shouldOptimizeCast(CastInst *CI);
231
232 /// Try to optimize a sequence of instructions checking if an operation
233 /// on LHS and RHS overflows.
234 ///
235 /// If this overflow check is done via one of the overflow check intrinsics,
236 /// then CtxI has to be the call instruction calling that intrinsic. If this
237 /// overflow check is done by arithmetic followed by a compare, then CtxI has
238 /// to be the arithmetic instruction.
239 ///
240 /// If a simplification is possible, stores the simplified result of the
241 /// operation in OperationResult and result of the overflow check in
242 /// OverflowResult, and return true. If no simplification is possible,
243 /// returns false.
244 bool OptimizeOverflowCheck(Instruction::BinaryOps BinaryOp, bool IsSigned,
245 Value *LHS, Value *RHS,
246 Instruction &CtxI, Value *&OperationResult,
247 Constant *&OverflowResult);
248
249 Instruction *visitCallBase(CallBase &Call);
250 Instruction *tryOptimizeCall(CallInst *CI);
251 bool transformConstExprCastCall(CallBase &Call);
252 Instruction *transformCallThroughTrampoline(CallBase &Call,
253 IntrinsicInst &Tramp);
254
255 /// Try to optimize a call to the result of a ptrauth intrinsic, potentially
256 /// into the ptrauth call bundle:
257 /// - call(ptrauth.resign(p)), ["ptrauth"()] -> call p, ["ptrauth"()]
258 /// - call(ptrauth.sign(p)), ["ptrauth"()] -> call p
259 /// as long as the key/discriminator are the same in sign and auth-bundle,
260 /// and we don't change the key in the bundle (to a potentially-invalid key.)
261 Instruction *foldPtrAuthIntrinsicCallee(CallBase &Call);
262
263 /// Try to optimize a call to a ptrauth constant, into its ptrauth bundle:
264 /// call(ptrauth(f)), ["ptrauth"()] -> call f
265 /// as long as the key/discriminator are the same in constant and bundle.
266 Instruction *foldPtrAuthConstantCallee(CallBase &Call);
267
268 // Return (a, b) if (LHS, RHS) is known to be (a, b) or (b, a).
269 // Otherwise, return std::nullopt
270 // Currently it matches:
271 // - LHS = (select c, a, b), RHS = (select c, b, a)
272 // - LHS = (phi [a, BB0], [b, BB1]), RHS = (phi [b, BB0], [a, BB1])
273 // - LHS = min(a, b), RHS = max(a, b)
274 std::optional<std::pair<Value *, Value *>> matchSymmetricPair(Value *LHS,
275 Value *RHS);
276
277 Value *simplifyMaskedLoad(IntrinsicInst &II);
278 Instruction *simplifyMaskedStore(IntrinsicInst &II);
279 Instruction *simplifyMaskedGather(IntrinsicInst &II);
280 Instruction *simplifyMaskedScatter(IntrinsicInst &II);
281
282 /// Transform (zext icmp) to bitwise / integer operations in order to
283 /// eliminate it.
284 ///
285 /// \param ICI The icmp of the (zext icmp) pair we are interested in.
286 /// \parem CI The zext of the (zext icmp) pair we are interested in.
287 ///
288 /// \return null if the transformation cannot be performed. If the
289 /// transformation can be performed the new instruction that replaces the
290 /// (zext icmp) pair will be returned.
291 Instruction *transformZExtICmp(ICmpInst *Cmp, ZExtInst &Zext);
292
293 Instruction *transformSExtICmp(ICmpInst *Cmp, SExtInst &Sext);
294
295 bool willNotOverflowSignedAdd(const WithCache<const Value *> &LHS,
296 const WithCache<const Value *> &RHS,
297 const Instruction &CxtI) const {
298 return computeOverflowForSignedAdd(LHS, RHS, CxtI: &CxtI) ==
299 OverflowResult::NeverOverflows;
300 }
301
302 bool willNotOverflowUnsignedAdd(const WithCache<const Value *> &LHS,
303 const WithCache<const Value *> &RHS,
304 const Instruction &CxtI) const {
305 return computeOverflowForUnsignedAdd(LHS, RHS, CxtI: &CxtI) ==
306 OverflowResult::NeverOverflows;
307 }
308
309 bool willNotOverflowAdd(const Value *LHS, const Value *RHS,
310 const Instruction &CxtI, bool IsSigned) const {
311 return IsSigned ? willNotOverflowSignedAdd(LHS, RHS, CxtI)
312 : willNotOverflowUnsignedAdd(LHS, RHS, CxtI);
313 }
314
315 bool willNotOverflowSignedSub(const Value *LHS, const Value *RHS,
316 const Instruction &CxtI) const {
317 return computeOverflowForSignedSub(LHS, RHS, CxtI: &CxtI) ==
318 OverflowResult::NeverOverflows;
319 }
320
321 bool willNotOverflowUnsignedSub(const Value *LHS, const Value *RHS,
322 const Instruction &CxtI) const {
323 return computeOverflowForUnsignedSub(LHS, RHS, CxtI: &CxtI) ==
324 OverflowResult::NeverOverflows;
325 }
326
327 bool willNotOverflowSub(const Value *LHS, const Value *RHS,
328 const Instruction &CxtI, bool IsSigned) const {
329 return IsSigned ? willNotOverflowSignedSub(LHS, RHS, CxtI)
330 : willNotOverflowUnsignedSub(LHS, RHS, CxtI);
331 }
332
333 bool willNotOverflowSignedMul(const Value *LHS, const Value *RHS,
334 const Instruction &CxtI) const {
335 return computeOverflowForSignedMul(LHS, RHS, CxtI: &CxtI) ==
336 OverflowResult::NeverOverflows;
337 }
338
339 bool willNotOverflowUnsignedMul(const Value *LHS, const Value *RHS,
340 const Instruction &CxtI,
341 bool IsNSW = false) const {
342 return computeOverflowForUnsignedMul(LHS, RHS, CxtI: &CxtI, IsNSW) ==
343 OverflowResult::NeverOverflows;
344 }
345
346 bool willNotOverflowMul(const Value *LHS, const Value *RHS,
347 const Instruction &CxtI, bool IsSigned) const {
348 return IsSigned ? willNotOverflowSignedMul(LHS, RHS, CxtI)
349 : willNotOverflowUnsignedMul(LHS, RHS, CxtI);
350 }
351
352 bool willNotOverflow(BinaryOperator::BinaryOps Opcode, const Value *LHS,
353 const Value *RHS, const Instruction &CxtI,
354 bool IsSigned) const {
355 switch (Opcode) {
356 case Instruction::Add: return willNotOverflowAdd(LHS, RHS, CxtI, IsSigned);
357 case Instruction::Sub: return willNotOverflowSub(LHS, RHS, CxtI, IsSigned);
358 case Instruction::Mul: return willNotOverflowMul(LHS, RHS, CxtI, IsSigned);
359 default: llvm_unreachable("Unexpected opcode for overflow query");
360 }
361 }
362
363 Value *EmitGEPOffset(GEPOperator *GEP, bool RewriteGEP = false);
364 /// Emit sum of multiple GEP offsets. The GEPs are processed in reverse
365 /// order.
366 Value *EmitGEPOffsets(ArrayRef<GEPOperator *> GEPs, GEPNoWrapFlags NW,
367 Type *IdxTy, bool RewriteGEPs);
368 Instruction *scalarizePHI(ExtractElementInst &EI, PHINode *PN);
369 Instruction *foldBitcastExtElt(ExtractElementInst &ExtElt);
370 Instruction *foldCastedBitwiseLogic(BinaryOperator &I);
371 Instruction *foldFBinOpOfIntCasts(BinaryOperator &I);
372 // Should only be called by `foldFBinOpOfIntCasts`.
373 Instruction *foldFBinOpOfIntCastsFromSign(
374 BinaryOperator &BO, bool OpsFromSigned, std::array<Value *, 2> IntOps,
375 Constant *Op1FpC, SmallVectorImpl<WithCache<const Value *>> &OpsKnown);
376 Instruction *foldBinopOfSextBoolToSelect(BinaryOperator &I);
377 Instruction *narrowBinOp(TruncInst &Trunc);
378 Instruction *narrowMaskedBinOp(BinaryOperator &And);
379 Instruction *narrowMathIfNoOverflow(BinaryOperator &I);
380 Instruction *narrowFunnelShift(TruncInst &Trunc);
381 Instruction *optimizeBitCastFromPhi(CastInst &CI, PHINode *PN);
382 Instruction *matchSAddSubSat(IntrinsicInst &MinMax1);
383 Instruction *foldNot(BinaryOperator &I);
384 Instruction *foldBinOpOfDisplacedShifts(BinaryOperator &I);
385
386 /// Determine if a pair of casts can be replaced by a single cast.
387 ///
388 /// \param CI1 The first of a pair of casts.
389 /// \param CI2 The second of a pair of casts.
390 ///
391 /// \return 0 if the cast pair cannot be eliminated, otherwise returns an
392 /// Instruction::CastOps value for a cast that can replace the pair, casting
393 /// CI1->getSrcTy() to CI2->getDstTy().
394 ///
395 /// \see CastInst::isEliminableCastPair
396 Instruction::CastOps isEliminableCastPair(const CastInst *CI1,
397 const CastInst *CI2);
398 Value *simplifyIntToPtrRoundTripCast(Value *Val);
399
400 Value *foldAndOrOfICmps(ICmpInst *LHS, ICmpInst *RHS, Instruction &I,
401 bool IsAnd, bool IsLogical = false);
402 Value *foldXorOfICmps(ICmpInst *LHS, ICmpInst *RHS, BinaryOperator &Xor);
403
404 Value *foldEqOfParts(Value *Cmp0, Value *Cmp1, bool IsAnd);
405
406 Value *foldAndOrOfICmpsUsingRanges(ICmpInst *ICmp1, ICmpInst *ICmp2,
407 bool IsAnd);
408
409 /// Optimize (fcmp)&(fcmp) or (fcmp)|(fcmp).
410 /// NOTE: Unlike most of instcombine, this returns a Value which should
411 /// already be inserted into the function.
412 Value *foldLogicOfFCmps(FCmpInst *LHS, FCmpInst *RHS, bool IsAnd,
413 bool IsLogicalSelect = false);
414
415 Instruction *foldLogicOfIsFPClass(BinaryOperator &Operator, Value *LHS,
416 Value *RHS);
417
418 Value *foldBooleanAndOr(Value *LHS, Value *RHS, Instruction &I, bool IsAnd,
419 bool IsLogical);
420
421 Value *reassociateBooleanAndOr(Value *LHS, Value *X, Value *Y, Instruction &I,
422 bool IsAnd, bool RHSIsLogical);
423
424 Value *foldDisjointOr(Value *LHS, Value *RHS);
425
426 Value *reassociateDisjointOr(Value *LHS, Value *RHS);
427
428 Instruction *
429 canonicalizeConditionalNegationViaMathToSelect(BinaryOperator &i);
430
431 Value *matchSelectFromAndOr(Value *A, Value *B, Value *C, Value *D,
432 bool InvertFalseVal = false);
433 Value *getSelectCondition(Value *A, Value *B, bool ABIsTheSame);
434
435 Instruction *foldLShrOverflowBit(BinaryOperator &I);
436 Instruction *foldExtractOfOverflowIntrinsic(ExtractValueInst &EV);
437 Instruction *foldIntrinsicWithOverflowCommon(IntrinsicInst *II);
438 Instruction *foldIntrinsicIsFPClass(IntrinsicInst &II);
439 Instruction *foldFPSignBitOps(BinaryOperator &I);
440 Instruction *foldFDivConstantDivisor(BinaryOperator &I);
441
442 // Optimize one of these forms:
443 // and i1 Op, SI / select i1 Op, i1 SI, i1 false (if IsAnd = true)
444 // or i1 Op, SI / select i1 Op, i1 true, i1 SI (if IsAnd = false)
445 // into simplier select instruction using isImpliedCondition.
446 Instruction *foldAndOrOfSelectUsingImpliedCond(Value *Op, SelectInst &SI,
447 bool IsAnd);
448
449 Instruction *hoistFNegAboveFMulFDiv(Value *FNegOp, Instruction &FMFSource);
450
451 /// Simplify \p V given that it is known to be non-null.
452 /// Returns the simplified value if possible, otherwise returns nullptr.
453 /// If \p HasDereferenceable is true, the simplification will not perform
454 /// same object checks.
455 Value *simplifyNonNullOperand(Value *V, bool HasDereferenceable,
456 unsigned Depth = 0);
457
458 /// Create `select C, S1, S2`. Use only when the profile cannot be calculated
459 /// from existing profile metadata: if the Function has profiles, this will
460 /// set the profile of this select to "unknown".
461 SelectInst *
462 createSelectInstWithUnknownProfile(Value *C, Value *S1, Value *S2,
463 const Twine &NameStr = "",
464 InsertPosition InsertBefore = nullptr) {
465 auto *Sel = SelectInst::Create(C, S1, S2, NameStr, InsertBefore, MDFrom: nullptr);
466 setExplicitlyUnknownBranchWeightsIfProfiled(I&: *Sel, DEBUG_TYPE, F: &F);
467 return Sel;
468 }
469
470public:
471 /// Create and insert the idiom we use to indicate a block is unreachable
472 /// without having to rewrite the CFG from within InstCombine.
473 void CreateNonTerminatorUnreachable(Instruction *InsertAt) {
474 auto &Ctx = InsertAt->getContext();
475 auto *SI = new StoreInst(ConstantInt::getTrue(Context&: Ctx),
476 PoisonValue::get(T: PointerType::getUnqual(C&: Ctx)),
477 /*isVolatile*/ false, Align(1));
478 InsertNewInstWith(New: SI, Old: InsertAt->getIterator());
479 }
480
481 /// Combiner aware instruction erasure.
482 ///
483 /// When dealing with an instruction that has side effects or produces a void
484 /// value, we can't rely on DCE to delete the instruction. Instead, visit
485 /// methods should return the value returned by this function.
486 Instruction *eraseInstFromFunction(Instruction &I) override {
487 LLVM_DEBUG(dbgs() << "IC: ERASE " << I << '\n');
488 assert(I.use_empty() && "Cannot erase instruction that is used!");
489 salvageDebugInfo(I);
490
491 // Make sure that we reprocess all operands now that we reduced their
492 // use counts.
493 SmallVector<Value *> Ops(I.operands());
494 Worklist.remove(I: &I);
495 DC.removeValue(V: &I);
496 I.eraseFromParent();
497 for (Value *Op : Ops)
498 Worklist.handleUseCountDecrement(V: Op);
499 MadeIRChange = true;
500 return nullptr; // Don't do anything with FI
501 }
502
503 OverflowResult computeOverflow(
504 Instruction::BinaryOps BinaryOp, bool IsSigned,
505 Value *LHS, Value *RHS, Instruction *CxtI) const;
506
507 /// Performs a few simplifications for operators which are associative
508 /// or commutative.
509 bool SimplifyAssociativeOrCommutative(BinaryOperator &I);
510
511 /// Tries to simplify binary operations which some other binary
512 /// operation distributes over.
513 ///
514 /// It does this by either by factorizing out common terms (eg "(A*B)+(A*C)"
515 /// -> "A*(B+C)") or expanding out if this results in simplifications (eg: "A
516 /// & (B | C) -> (A&B) | (A&C)" if this is a win). Returns the simplified
517 /// value, or null if it didn't simplify.
518 Value *foldUsingDistributiveLaws(BinaryOperator &I);
519
520 /// Tries to simplify add operations using the definition of remainder.
521 ///
522 /// The definition of remainder is X % C = X - (X / C ) * C. The add
523 /// expression X % C0 + (( X / C0 ) % C1) * C0 can be simplified to
524 /// X % (C0 * C1)
525 Value *SimplifyAddWithRemainder(BinaryOperator &I);
526
527 // Binary Op helper for select operations where the expression can be
528 // efficiently reorganized.
529 Value *SimplifySelectsFeedingBinaryOp(BinaryOperator &I, Value *LHS,
530 Value *RHS);
531
532 // If `I` has operand `(ctpop (not x))`, fold `I` with `(sub nuw nsw
533 // BitWidth(x), (ctpop x))`.
534 Instruction *tryFoldInstWithCtpopWithNot(Instruction *I);
535
536 // (Binop1 (Binop2 (logic_shift X, C), C1), (logic_shift Y, C))
537 // -> (logic_shift (Binop1 (Binop2 X, inv_logic_shift(C1, C)), Y), C)
538 // (Binop1 (Binop2 (logic_shift X, Amt), Mask), (logic_shift Y, Amt))
539 // -> (BinOp (logic_shift (BinOp X, Y)), Mask)
540 Instruction *foldBinOpShiftWithShift(BinaryOperator &I);
541
542 /// Tries to simplify binops of select and cast of the select condition.
543 ///
544 /// (Binop (cast C), (select C, T, F))
545 /// -> (select C, C0, C1)
546 Instruction *foldBinOpOfSelectAndCastOfSelectCondition(BinaryOperator &I);
547
548 /// This tries to simplify binary operations by factorizing out common terms
549 /// (e. g. "(A*B)+(A*C)" -> "A*(B+C)").
550 Value *tryFactorizationFolds(BinaryOperator &I);
551
552 /// Match a select chain which produces one of three values based on whether
553 /// the LHS is less than, equal to, or greater than RHS respectively.
554 /// Return true if we matched a three way compare idiom. The LHS, RHS, Less,
555 /// Equal and Greater values are saved in the matching process and returned to
556 /// the caller.
557 bool matchThreeWayIntCompare(SelectInst *SI, Value *&LHS, Value *&RHS,
558 ConstantInt *&Less, ConstantInt *&Equal,
559 ConstantInt *&Greater);
560
561 /// Attempts to replace I with a simpler value based on the demanded
562 /// bits.
563 Value *SimplifyDemandedUseBits(Instruction *I, const APInt &DemandedMask,
564 KnownBits &Known, const SimplifyQuery &Q,
565 unsigned Depth = 0);
566 using InstCombiner::SimplifyDemandedBits;
567 bool SimplifyDemandedBits(Instruction *I, unsigned Op,
568 const APInt &DemandedMask, KnownBits &Known,
569 const SimplifyQuery &Q,
570 unsigned Depth = 0) override;
571
572 /// Helper routine of SimplifyDemandedUseBits. It computes KnownZero/KnownOne
573 /// bits. It also tries to handle simplifications that can be done based on
574 /// DemandedMask, but without modifying the Instruction.
575 Value *SimplifyMultipleUseDemandedBits(Instruction *I,
576 const APInt &DemandedMask,
577 KnownBits &Known,
578 const SimplifyQuery &Q,
579 unsigned Depth = 0);
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 bool SimplifyDemandedInstructionBits(Instruction &Inst, KnownBits &Known);
591
592 Value *SimplifyDemandedVectorElts(Value *V, APInt DemandedElts,
593 APInt &PoisonElts, unsigned Depth = 0,
594 bool AllowMultipleUsers = false) override;
595
596 /// Attempts to replace V with a simpler value based on the demanded
597 /// floating-point classes
598 Value *SimplifyDemandedUseFPClass(Instruction *I, FPClassTest DemandedMask,
599 KnownFPClass &Known, const SimplifyQuery &Q,
600 unsigned Depth = 0);
601 Value *SimplifyMultipleUseDemandedFPClass(Instruction *I,
602 FPClassTest DemandedMask,
603 KnownFPClass &Known,
604 const SimplifyQuery &Q,
605 unsigned Depth);
606
607 bool SimplifyDemandedFPClass(Instruction *I, unsigned Op,
608 FPClassTest DemandedMask, KnownFPClass &Known,
609 const SimplifyQuery &Q, unsigned Depth = 0);
610
611 bool SimplifyDemandedInstructionFPClass(Instruction &Inst);
612
613 /// Common transforms for add / disjoint or
614 Instruction *foldAddLikeCommutative(Value *LHS, Value *RHS, bool NSW,
615 bool NUW);
616
617 /// Canonicalize the position of binops relative to shufflevector.
618 Instruction *foldVectorBinop(BinaryOperator &Inst);
619 Instruction *foldVectorSelect(SelectInst &Sel);
620 Instruction *foldSelectShuffle(ShuffleVectorInst &Shuf);
621 Constant *unshuffleConstant(ArrayRef<int> ShMask, Constant *C,
622 VectorType *NewCTy);
623
624 /// Given a binary operator, cast instruction, or select which has a PHI node
625 /// as operand #0, see if we can fold the instruction into the PHI (which is
626 /// only possible if all operands to the PHI are constants).
627 Instruction *foldOpIntoPhi(Instruction &I, PHINode *PN,
628 bool AllowMultipleUses = false);
629
630 /// Try to fold binary operators whose operands are simple interleaved
631 /// recurrences to a single recurrence. This is a common pattern in reduction
632 /// operations.
633 /// Example:
634 /// %phi1 = phi [init1, %BB1], [%op1, %BB2]
635 /// %phi2 = phi [init2, %BB1], [%op2, %BB2]
636 /// %op1 = binop %phi1, constant1
637 /// %op2 = binop %phi2, constant2
638 /// %rdx = binop %op1, %op2
639 /// -->
640 /// %phi_combined = phi [init_combined, %BB1], [%op_combined, %BB2]
641 /// %rdx_combined = binop %phi_combined, constant_combined
642 Instruction *foldBinopWithRecurrence(BinaryOperator &BO);
643
644 /// For a binary operator with 2 phi operands, try to hoist the binary
645 /// operation before the phi. This can result in fewer instructions in
646 /// patterns where at least one set of phi operands simplifies.
647 /// Example:
648 /// BB3: binop (phi [X, BB1], [C1, BB2]), (phi [Y, BB1], [C2, BB2])
649 /// -->
650 /// BB1: BO = binop X, Y
651 /// BB3: phi [BO, BB1], [(binop C1, C2), BB2]
652 Instruction *foldBinopWithPhiOperands(BinaryOperator &BO);
653
654 /// Given an instruction with a select as one operand and a constant as the
655 /// other operand, try to fold the binary operator into the select arguments.
656 /// This also works for Cast instructions, which obviously do not have a
657 /// second operand.
658 Instruction *FoldOpIntoSelect(Instruction &Op, SelectInst *SI,
659 bool FoldWithMultiUse = false,
660 bool SimplifyBothArms = false);
661
662 Instruction *foldBinOpSelectBinOp(BinaryOperator &Op);
663
664 /// This is a convenience wrapper function for the above two functions.
665 Instruction *foldBinOpIntoSelectOrPhi(BinaryOperator &I);
666
667 Instruction *foldAddWithConstant(BinaryOperator &Add);
668
669 Instruction *foldSquareSumInt(BinaryOperator &I);
670 Instruction *foldSquareSumFP(BinaryOperator &I);
671
672 /// Try to rotate an operation below a PHI node, using PHI nodes for
673 /// its operands.
674 Instruction *foldPHIArgOpIntoPHI(PHINode &PN);
675 Instruction *foldPHIArgBinOpIntoPHI(PHINode &PN);
676 Instruction *foldPHIArgInsertValueInstructionIntoPHI(PHINode &PN);
677 Instruction *foldPHIArgExtractValueInstructionIntoPHI(PHINode &PN);
678 Instruction *foldPHIArgGEPIntoPHI(PHINode &PN);
679 Instruction *foldPHIArgLoadIntoPHI(PHINode &PN);
680 Instruction *foldPHIArgZextsIntoPHI(PHINode &PN);
681 Instruction *foldPHIArgIntToPtrToPHI(PHINode &PN);
682
683 /// If the phi is within a phi web, which is formed by the def-use chain
684 /// of phis and all the phis in the web are only used in the other phis.
685 /// In this case, these phis are dead and we will remove all of them.
686 bool foldDeadPhiWeb(PHINode &PN);
687
688 /// If an integer typed PHI has only one use which is an IntToPtr operation,
689 /// replace the PHI with an existing pointer typed PHI if it exists. Otherwise
690 /// insert a new pointer typed PHI and replace the original one.
691 bool foldIntegerTypedPHI(PHINode &PN);
692
693 /// Helper function for FoldPHIArgXIntoPHI() to set debug location for the
694 /// folded operation.
695 void PHIArgMergedDebugLoc(Instruction *Inst, PHINode &PN);
696
697 Value *foldPtrToIntOrAddrOfGEP(Type *IntTy, Value *Ptr);
698 Instruction *foldGEPICmp(GEPOperator *GEPLHS, Value *RHS, CmpPredicate Cond,
699 Instruction &I);
700 Instruction *foldSelectICmp(CmpPredicate Pred, SelectInst *SI, Value *RHS,
701 const ICmpInst &I);
702 bool foldAllocaCmp(AllocaInst *Alloca);
703 Instruction *foldCmpLoadFromIndexedGlobal(LoadInst *LI,
704 GetElementPtrInst *GEP,
705 CmpInst &ICI,
706 ConstantInt *AndCst = nullptr);
707 Instruction *foldFCmpIntToFPConst(FCmpInst &I, Instruction *LHSI,
708 Constant *RHSC);
709 Instruction *foldICmpAddOpConst(Value *X, const APInt &C, CmpPredicate Pred);
710 Instruction *foldICmpWithCastOp(ICmpInst &ICmp);
711 Instruction *foldICmpWithZextOrSext(ICmpInst &ICmp);
712
713 Instruction *foldICmpUsingKnownBits(ICmpInst &Cmp);
714 Instruction *foldICmpWithDominatingICmp(ICmpInst &Cmp);
715 Instruction *foldICmpWithConstant(ICmpInst &Cmp);
716 Instruction *foldIsMultipleOfAPowerOfTwo(ICmpInst &Cmp);
717 Instruction *foldICmpUsingBoolRange(ICmpInst &I);
718 Instruction *foldICmpInstWithConstant(ICmpInst &Cmp);
719 Instruction *foldICmpInstWithConstantNotInt(ICmpInst &Cmp);
720 Instruction *foldICmpInstWithConstantAllowPoison(ICmpInst &Cmp,
721 const APInt &C);
722 Instruction *foldICmpBinOp(ICmpInst &Cmp, const SimplifyQuery &SQ);
723 Instruction *foldICmpWithMinMax(Instruction &I, MinMaxIntrinsic *MinMax,
724 Value *Z, CmpPredicate Pred);
725 Instruction *foldICmpWithClamp(ICmpInst &Cmp, Value *X, MinMaxIntrinsic *Min);
726 Instruction *foldICmpEquality(ICmpInst &Cmp);
727 Instruction *foldIRemByPowerOfTwoToBitTest(ICmpInst &I);
728 Instruction *foldSignBitTest(ICmpInst &I);
729 Instruction *foldICmpWithZero(ICmpInst &Cmp);
730
731 Value *foldMultiplicationOverflowCheck(ICmpInst &Cmp);
732
733 Instruction *foldICmpBinOpWithConstant(ICmpInst &Cmp, BinaryOperator *BO,
734 const APInt &C);
735 Instruction *foldICmpSelectConstant(ICmpInst &Cmp, SelectInst *Select,
736 ConstantInt *C);
737 Instruction *foldICmpTruncConstant(ICmpInst &Cmp, TruncInst *Trunc,
738 const APInt &C);
739 Instruction *foldICmpTruncWithTruncOrExt(ICmpInst &Cmp,
740 const SimplifyQuery &Q);
741 Instruction *foldICmpAndConstant(ICmpInst &Cmp, BinaryOperator *And,
742 const APInt &C);
743 Instruction *foldICmpXorConstant(ICmpInst &Cmp, BinaryOperator *Xor,
744 const APInt &C);
745 Instruction *foldICmpOrConstant(ICmpInst &Cmp, BinaryOperator *Or,
746 const APInt &C);
747 Instruction *foldICmpMulConstant(ICmpInst &Cmp, BinaryOperator *Mul,
748 const APInt &C);
749 Instruction *foldICmpShlConstant(ICmpInst &Cmp, BinaryOperator *Shl,
750 const APInt &C);
751 Instruction *foldICmpShrConstant(ICmpInst &Cmp, BinaryOperator *Shr,
752 const APInt &C);
753 Instruction *foldICmpSRemConstant(ICmpInst &Cmp, BinaryOperator *UDiv,
754 const APInt &C);
755 Instruction *foldICmpUDivConstant(ICmpInst &Cmp, BinaryOperator *UDiv,
756 const APInt &C);
757 Instruction *foldICmpDivConstant(ICmpInst &Cmp, BinaryOperator *Div,
758 const APInt &C);
759 Instruction *foldICmpSubConstant(ICmpInst &Cmp, BinaryOperator *Sub,
760 const APInt &C);
761 Instruction *foldICmpAddConstant(ICmpInst &Cmp, BinaryOperator *Add,
762 const APInt &C);
763 Instruction *foldICmpAndConstConst(ICmpInst &Cmp, BinaryOperator *And,
764 const APInt &C1);
765 Instruction *foldICmpAndShift(ICmpInst &Cmp, BinaryOperator *And,
766 const APInt &C1, const APInt &C2);
767 Instruction *foldICmpXorShiftConst(ICmpInst &Cmp, BinaryOperator *Xor,
768 const APInt &C);
769 Instruction *foldICmpShrConstConst(ICmpInst &I, Value *ShAmt, const APInt &C1,
770 const APInt &C2);
771 Instruction *foldICmpShlConstConst(ICmpInst &I, Value *ShAmt, const APInt &C1,
772 const APInt &C2);
773
774 Instruction *foldICmpBinOpWithConstantViaTruthTable(ICmpInst &Cmp,
775 BinaryOperator *BO,
776 const APInt &C);
777 Instruction *foldICmpBinOpEqualityWithConstant(ICmpInst &Cmp,
778 BinaryOperator *BO,
779 const APInt &C);
780 Instruction *foldICmpIntrinsicWithConstant(ICmpInst &ICI, IntrinsicInst *II,
781 const APInt &C);
782 Instruction *foldICmpEqIntrinsicWithConstant(ICmpInst &ICI, IntrinsicInst *II,
783 const APInt &C);
784 Instruction *foldICmpBitCast(ICmpInst &Cmp);
785 Instruction *foldICmpWithTrunc(ICmpInst &Cmp);
786 Instruction *foldICmpCommutative(CmpPredicate Pred, Value *Op0, Value *Op1,
787 ICmpInst &CxtI);
788
789 // Helpers of visitSelectInst().
790 Instruction *foldSelectOfBools(SelectInst &SI);
791 Instruction *foldSelectToCmp(SelectInst &SI);
792 Instruction *foldSelectExtConst(SelectInst &Sel);
793 Instruction *foldSelectEqualityTest(SelectInst &SI);
794 Instruction *foldSelectOpOp(SelectInst &SI, Instruction *TI, Instruction *FI);
795 Instruction *foldSelectIntrinsic(SelectInst &SI);
796 Instruction *foldSelectIntoOp(SelectInst &SI, Value *, Value *);
797 Instruction *foldSPFofSPF(Instruction *Inner, SelectPatternFlavor SPF1,
798 Value *A, Value *B, Instruction &Outer,
799 SelectPatternFlavor SPF2, Value *C);
800 Instruction *foldSelectInstWithICmp(SelectInst &SI, ICmpInst *ICI);
801 Value *foldSelectWithConstOpToBinOp(ICmpInst *Cmp, Value *TrueVal,
802 Value *FalseVal);
803 Instruction *foldSelectValueEquivalence(SelectInst &SI, CmpInst &CI);
804 bool replaceInInstruction(Value *V, Value *Old, Value *New,
805 unsigned Depth = 0);
806
807 Value *insertRangeTest(Value *V, const APInt &Lo, const APInt &Hi,
808 bool isSigned, bool Inside);
809 bool mergeStoreIntoSuccessor(StoreInst &SI);
810
811 /// Given an initial instruction, check to see if it is the root of a
812 /// bswap/bitreverse idiom. If so, return the equivalent bswap/bitreverse
813 /// intrinsic.
814 Instruction *matchBSwapOrBitReverse(Instruction &I, bool MatchBSwaps,
815 bool MatchBitReversals);
816
817 Instruction *SimplifyAnyMemTransfer(AnyMemTransferInst *MI);
818 Instruction *SimplifyAnyMemSet(AnyMemSetInst *MI);
819
820 Value *EvaluateInDifferentType(Value *V, Type *Ty, bool isSigned);
821
822 bool tryToSinkInstruction(Instruction *I, BasicBlock *DestBlock);
823 void tryToSinkInstructionDbgVariableRecords(
824 Instruction *I, BasicBlock::iterator InsertPos, BasicBlock *SrcBlock,
825 BasicBlock *DestBlock, SmallVectorImpl<DbgVariableRecord *> &DPUsers);
826
827 bool removeInstructionsBeforeUnreachable(Instruction &I);
828 void addDeadEdge(BasicBlock *From, BasicBlock *To,
829 SmallVectorImpl<BasicBlock *> &Worklist);
830 void handleUnreachableFrom(Instruction *I,
831 SmallVectorImpl<BasicBlock *> &Worklist);
832 void handlePotentiallyDeadBlocks(SmallVectorImpl<BasicBlock *> &Worklist);
833 void handlePotentiallyDeadSuccessors(BasicBlock *BB, BasicBlock *LiveSucc);
834 void freelyInvertAllUsersOf(Value *V, Value *IgnoredUser = nullptr);
835
836 /// Take the exact integer log2 of the value. If DoFold is true, create the
837 /// actual instructions, otherwise return a non-null dummy value. Return
838 /// nullptr on failure. Note, if DoFold is true the caller must ensure that
839 /// takeLog2 will succeed, otherwise it may create stray instructions.
840 Value *takeLog2(Value *Op, unsigned Depth, bool AssumeNonZero, bool DoFold);
841
842 Value *tryGetLog2(Value *Op, bool AssumeNonZero) {
843 if (takeLog2(Op, /*Depth=*/Depth: 0, AssumeNonZero, /*DoFold=*/DoFold: false))
844 return takeLog2(Op, /*Depth=*/Depth: 0, AssumeNonZero, /*DoFold=*/DoFold: true);
845 return nullptr;
846 }
847};
848
849class Negator final {
850 /// Top-to-bottom, def-to-use negated instruction tree we produced.
851 SmallVector<Instruction *, NegatorMaxNodesSSO> NewInstructions;
852
853 using BuilderTy = IRBuilder<TargetFolder, IRBuilderCallbackInserter>;
854 BuilderTy Builder;
855
856 const DominatorTree &DT;
857
858 const bool IsTrulyNegation;
859
860 SmallDenseMap<Value *, Value *> NegationsCache;
861
862 Negator(LLVMContext &C, const DataLayout &DL, const DominatorTree &DT,
863 bool IsTrulyNegation);
864
865#if LLVM_ENABLE_STATS
866 unsigned NumValuesVisitedInThisNegator = 0;
867 ~Negator();
868#endif
869
870 using Result = std::pair<ArrayRef<Instruction *> /*NewInstructions*/,
871 Value * /*NegatedRoot*/>;
872
873 std::array<Value *, 2> getSortedOperandsOfBinOp(Instruction *I);
874
875 [[nodiscard]] Value *visitImpl(Value *V, bool IsNSW, unsigned Depth);
876
877 [[nodiscard]] Value *negate(Value *V, bool IsNSW, unsigned Depth);
878
879 /// Recurse depth-first and attempt to sink the negation.
880 /// FIXME: use worklist?
881 [[nodiscard]] std::optional<Result> run(Value *Root, bool IsNSW);
882
883 Negator(const Negator &) = delete;
884 Negator(Negator &&) = delete;
885 Negator &operator=(const Negator &) = delete;
886 Negator &operator=(Negator &&) = delete;
887
888public:
889 /// Attempt to negate \p Root. Retuns nullptr if negation can't be performed,
890 /// otherwise returns negated value.
891 [[nodiscard]] static Value *Negate(bool LHSIsZero, bool IsNSW, Value *Root,
892 InstCombinerImpl &IC);
893};
894
895struct CommonPointerBase {
896 /// Common base pointer.
897 Value *Ptr = nullptr;
898 /// LHS GEPs until common base.
899 SmallVector<GEPOperator *> LHSGEPs;
900 /// RHS GEPs until common base.
901 SmallVector<GEPOperator *> RHSGEPs;
902 /// LHS GEP NoWrapFlags until common base.
903 GEPNoWrapFlags LHSNW = GEPNoWrapFlags::all();
904 /// RHS GEP NoWrapFlags until common base.
905 GEPNoWrapFlags RHSNW = GEPNoWrapFlags::all();
906
907 static CommonPointerBase compute(Value *LHS, Value *RHS);
908
909 /// Whether expanding the GEP chains is expensive.
910 bool isExpensive() const;
911};
912
913} // end namespace llvm
914
915#undef DEBUG_TYPE
916
917#endif // LLVM_LIB_TRANSFORMS_INSTCOMBINE_INSTCOMBINEINTERNAL_H
918