xref: /freebsd/contrib/llvm-project/llvm/lib/Transforms/InstCombine/InstCombineInternal.h (revision 0fca6ea1d4eea4c934cfff25ac9ee8ad6fe95583)
1 //===- InstCombineInternal.h - InstCombine pass internals -------*- C++ -*-===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 /// \file
10 ///
11 /// This file provides internal interfaces used to implement the InstCombine.
12 //
13 //===----------------------------------------------------------------------===//
14 
15 #ifndef LLVM_LIB_TRANSFORMS_INSTCOMBINE_INSTCOMBINEINTERNAL_H
16 #define LLVM_LIB_TRANSFORMS_INSTCOMBINE_INSTCOMBINEINTERNAL_H
17 
18 #include "llvm/ADT/Statistic.h"
19 #include "llvm/ADT/PostOrderIterator.h"
20 #include "llvm/Analysis/InstructionSimplify.h"
21 #include "llvm/Analysis/TargetFolder.h"
22 #include "llvm/Analysis/ValueTracking.h"
23 #include "llvm/IR/IRBuilder.h"
24 #include "llvm/IR/InstVisitor.h"
25 #include "llvm/IR/PatternMatch.h"
26 #include "llvm/IR/Value.h"
27 #include "llvm/Support/Debug.h"
28 #include "llvm/Support/KnownBits.h"
29 #include "llvm/Transforms/InstCombine/InstCombiner.h"
30 #include "llvm/Transforms/Utils/Local.h"
31 #include <cassert>
32 
33 #define DEBUG_TYPE "instcombine"
34 #include "llvm/Transforms/Utils/InstructionWorklist.h"
35 
36 using namespace llvm::PatternMatch;
37 
38 // As a default, let's assume that we want to be aggressive,
39 // and attempt to traverse with no limits in attempt to sink negation.
40 static constexpr unsigned NegatorDefaultMaxDepth = ~0U;
41 
42 // Let's guesstimate that most often we will end up visiting/producing
43 // fairly small number of new instructions.
44 static constexpr unsigned NegatorMaxNodesSSO = 16;
45 
46 namespace llvm {
47 
48 class AAResults;
49 class APInt;
50 class AssumptionCache;
51 class BlockFrequencyInfo;
52 class DataLayout;
53 class DominatorTree;
54 class GEPOperator;
55 class GlobalVariable;
56 class LoopInfo;
57 class OptimizationRemarkEmitter;
58 class ProfileSummaryInfo;
59 class TargetLibraryInfo;
60 class User;
61 
62 class LLVM_LIBRARY_VISIBILITY InstCombinerImpl final
63     : public InstCombiner,
64       public InstVisitor<InstCombinerImpl, Instruction *> {
65 public:
InstCombinerImpl(InstructionWorklist & Worklist,BuilderTy & Builder,bool MinimizeSize,AAResults * AA,AssumptionCache & AC,TargetLibraryInfo & TLI,TargetTransformInfo & TTI,DominatorTree & DT,OptimizationRemarkEmitter & ORE,BlockFrequencyInfo * BFI,BranchProbabilityInfo * BPI,ProfileSummaryInfo * PSI,const DataLayout & DL,LoopInfo * LI)66   InstCombinerImpl(InstructionWorklist &Worklist, BuilderTy &Builder,
67                    bool MinimizeSize, AAResults *AA, AssumptionCache &AC,
68                    TargetLibraryInfo &TLI, TargetTransformInfo &TTI,
69                    DominatorTree &DT, OptimizationRemarkEmitter &ORE,
70                    BlockFrequencyInfo *BFI, BranchProbabilityInfo *BPI,
71                    ProfileSummaryInfo *PSI, const DataLayout &DL, LoopInfo *LI)
72       : InstCombiner(Worklist, Builder, MinimizeSize, AA, AC, TLI, TTI, DT, ORE,
73                      BFI, BPI, PSI, DL, LI) {}
74 
75   virtual ~InstCombinerImpl() = default;
76 
77   /// Perform early cleanup and prepare the InstCombine worklist.
78   bool prepareWorklist(Function &F,
79                        ReversePostOrderTraversal<BasicBlock *> &RPOT);
80 
81   /// Run the combiner over the entire worklist until it is empty.
82   ///
83   /// \returns true if the IR is changed.
84   bool run();
85 
86   // Visitation implementation - Implement instruction combining for different
87   // instruction types.  The semantics are as follows:
88   // Return Value:
89   //    null        - No change was made
90   //     I          - Change was made, I is still valid, I may be dead though
91   //   otherwise    - Change was made, replace I with returned instruction
92   //
93   Instruction *visitFNeg(UnaryOperator &I);
94   Instruction *visitAdd(BinaryOperator &I);
95   Instruction *visitFAdd(BinaryOperator &I);
96   Value *OptimizePointerDifference(
97       Value *LHS, Value *RHS, Type *Ty, bool isNUW);
98   Instruction *visitSub(BinaryOperator &I);
99   Instruction *visitFSub(BinaryOperator &I);
100   Instruction *visitMul(BinaryOperator &I);
101   Instruction *foldPowiReassoc(BinaryOperator &I);
102   Instruction *foldFMulReassoc(BinaryOperator &I);
103   Instruction *visitFMul(BinaryOperator &I);
104   Instruction *visitURem(BinaryOperator &I);
105   Instruction *visitSRem(BinaryOperator &I);
106   Instruction *visitFRem(BinaryOperator &I);
107   bool simplifyDivRemOfSelectWithZeroOp(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 *visitAnd(BinaryOperator &I);
115   Instruction *visitOr(BinaryOperator &I);
116   bool sinkNotIntoLogicalOp(Instruction &I);
117   bool sinkNotIntoOtherHandOfLogicalOp(Instruction &I);
118   Instruction *visitXor(BinaryOperator &I);
119   Instruction *visitShl(BinaryOperator &I);
120   Value *reassociateShiftAmtsOfTwoSameDirectionShifts(
121       BinaryOperator *Sh0, const SimplifyQuery &SQ,
122       bool AnalyzeForSignBitExtraction = false);
123   Instruction *canonicalizeCondSignextOfHighBitExtractToSignextHighBitExtract(
124       BinaryOperator &I);
125   Instruction *foldVariableSignZeroExtensionOfVariableHighBitExtract(
126       BinaryOperator &OldAShr);
127   Instruction *visitAShr(BinaryOperator &I);
128   Instruction *visitLShr(BinaryOperator &I);
129   Instruction *commonShiftTransforms(BinaryOperator &I);
130   Instruction *visitFCmpInst(FCmpInst &I);
131   CmpInst *canonicalizeICmpPredicate(CmpInst &I);
132   Instruction *visitICmpInst(ICmpInst &I);
133   Instruction *FoldShiftByConstant(Value *Op0, Constant *Op1,
134                                    BinaryOperator &I);
135   Instruction *commonCastTransforms(CastInst &CI);
136   Instruction *visitTrunc(TruncInst &CI);
137   Instruction *visitZExt(ZExtInst &Zext);
138   Instruction *visitSExt(SExtInst &Sext);
139   Instruction *visitFPTrunc(FPTruncInst &CI);
140   Instruction *visitFPExt(CastInst &CI);
141   Instruction *visitFPToUI(FPToUIInst &FI);
142   Instruction *visitFPToSI(FPToSIInst &FI);
143   Instruction *visitUIToFP(CastInst &CI);
144   Instruction *visitSIToFP(CastInst &CI);
145   Instruction *visitPtrToInt(PtrToIntInst &CI);
146   Instruction *visitIntToPtr(IntToPtrInst &CI);
147   Instruction *visitBitCast(BitCastInst &CI);
148   Instruction *visitAddrSpaceCast(AddrSpaceCastInst &CI);
149   Instruction *foldItoFPtoI(CastInst &FI);
150   Instruction *visitSelectInst(SelectInst &SI);
151   Instruction *visitCallInst(CallInst &CI);
152   Instruction *visitInvokeInst(InvokeInst &II);
153   Instruction *visitCallBrInst(CallBrInst &CBI);
154 
155   Instruction *SliceUpIllegalIntegerPHI(PHINode &PN);
156   Instruction *visitPHINode(PHINode &PN);
157   Instruction *visitGetElementPtrInst(GetElementPtrInst &GEP);
158   Instruction *visitGEPOfGEP(GetElementPtrInst &GEP, GEPOperator *Src);
159   Instruction *visitAllocaInst(AllocaInst &AI);
160   Instruction *visitAllocSite(Instruction &FI);
161   Instruction *visitFree(CallInst &FI, Value *FreedOp);
162   Instruction *visitLoadInst(LoadInst &LI);
163   Instruction *visitStoreInst(StoreInst &SI);
164   Instruction *visitAtomicRMWInst(AtomicRMWInst &SI);
165   Instruction *visitUnconditionalBranchInst(BranchInst &BI);
166   Instruction *visitBranchInst(BranchInst &BI);
167   Instruction *visitFenceInst(FenceInst &FI);
168   Instruction *visitSwitchInst(SwitchInst &SI);
169   Instruction *visitReturnInst(ReturnInst &RI);
170   Instruction *visitUnreachableInst(UnreachableInst &I);
171   Instruction *
172   foldAggregateConstructionIntoAggregateReuse(InsertValueInst &OrigIVI);
173   Instruction *visitInsertValueInst(InsertValueInst &IV);
174   Instruction *visitInsertElementInst(InsertElementInst &IE);
175   Instruction *visitExtractElementInst(ExtractElementInst &EI);
176   Instruction *simplifyBinOpSplats(ShuffleVectorInst &SVI);
177   Instruction *visitShuffleVectorInst(ShuffleVectorInst &SVI);
178   Instruction *visitExtractValueInst(ExtractValueInst &EV);
179   Instruction *visitLandingPadInst(LandingPadInst &LI);
180   Instruction *visitVAEndInst(VAEndInst &I);
181   Value *pushFreezeToPreventPoisonFromPropagating(FreezeInst &FI);
182   bool freezeOtherUses(FreezeInst &FI);
183   Instruction *foldFreezeIntoRecurrence(FreezeInst &I, PHINode *PN);
184   Instruction *visitFreeze(FreezeInst &I);
185 
186   /// Specify what to return for unhandled instructions.
visitInstruction(Instruction & I)187   Instruction *visitInstruction(Instruction &I) { return nullptr; }
188 
189   /// True when DB dominates all uses of DI except UI.
190   /// UI must be in the same block as DI.
191   /// The routine checks that the DI parent and DB are different.
192   bool dominatesAllUses(const Instruction *DI, const Instruction *UI,
193                         const BasicBlock *DB) const;
194 
195   /// Try to replace select with select operand SIOpd in SI-ICmp sequence.
196   bool replacedSelectWithOperand(SelectInst *SI, const ICmpInst *Icmp,
197                                  const unsigned SIOpd);
198 
199   LoadInst *combineLoadToNewType(LoadInst &LI, Type *NewTy,
200                                  const Twine &Suffix = "");
201 
202   KnownFPClass computeKnownFPClass(Value *Val, FastMathFlags FMF,
203                                    FPClassTest Interested = fcAllFlags,
204                                    const Instruction *CtxI = nullptr,
205                                    unsigned Depth = 0) const {
206     return llvm::computeKnownFPClass(
207         Val, FMF, Interested, Depth,
208         getSimplifyQuery().getWithInstruction(CtxI));
209   }
210 
211   KnownFPClass computeKnownFPClass(Value *Val,
212                                    FPClassTest Interested = fcAllFlags,
213                                    const Instruction *CtxI = nullptr,
214                                    unsigned Depth = 0) const {
215     return llvm::computeKnownFPClass(
216         Val, Interested, Depth, getSimplifyQuery().getWithInstruction(CtxI));
217   }
218 
219   /// Check if fmul \p MulVal, +0.0 will yield +0.0 (or signed zero is
220   /// ignorable).
221   bool fmulByZeroIsZero(Value *MulVal, FastMathFlags FMF,
222                         const Instruction *CtxI) const;
223 
getLosslessTrunc(Constant * C,Type * TruncTy,unsigned ExtOp)224   Constant *getLosslessTrunc(Constant *C, Type *TruncTy, unsigned ExtOp) {
225     Constant *TruncC = ConstantExpr::getTrunc(C, TruncTy);
226     Constant *ExtTruncC =
227         ConstantFoldCastOperand(ExtOp, TruncC, C->getType(), DL);
228     if (ExtTruncC && ExtTruncC == C)
229       return TruncC;
230     return nullptr;
231   }
232 
getLosslessUnsignedTrunc(Constant * C,Type * TruncTy)233   Constant *getLosslessUnsignedTrunc(Constant *C, Type *TruncTy) {
234     return getLosslessTrunc(C, TruncTy, Instruction::ZExt);
235   }
236 
getLosslessSignedTrunc(Constant * C,Type * TruncTy)237   Constant *getLosslessSignedTrunc(Constant *C, Type *TruncTy) {
238     return getLosslessTrunc(C, TruncTy, Instruction::SExt);
239   }
240 
241   std::optional<std::pair<Intrinsic::ID, SmallVector<Value *, 3>>>
242   convertOrOfShiftsToFunnelShift(Instruction &Or);
243 
244 private:
245   bool annotateAnyAllocSite(CallBase &Call, const TargetLibraryInfo *TLI);
246   bool isDesirableIntType(unsigned BitWidth) const;
247   bool shouldChangeType(unsigned FromBitWidth, unsigned ToBitWidth) const;
248   bool shouldChangeType(Type *From, Type *To) const;
249   Value *dyn_castNegVal(Value *V) const;
250 
251   /// Classify whether a cast is worth optimizing.
252   ///
253   /// This is a helper to decide whether the simplification of
254   /// logic(cast(A), cast(B)) to cast(logic(A, B)) should be performed.
255   ///
256   /// \param CI The cast we are interested in.
257   ///
258   /// \return true if this cast actually results in any code being generated and
259   /// if it cannot already be eliminated by some other transformation.
260   bool shouldOptimizeCast(CastInst *CI);
261 
262   /// Try to optimize a sequence of instructions checking if an operation
263   /// on LHS and RHS overflows.
264   ///
265   /// If this overflow check is done via one of the overflow check intrinsics,
266   /// then CtxI has to be the call instruction calling that intrinsic.  If this
267   /// overflow check is done by arithmetic followed by a compare, then CtxI has
268   /// to be the arithmetic instruction.
269   ///
270   /// If a simplification is possible, stores the simplified result of the
271   /// operation in OperationResult and result of the overflow check in
272   /// OverflowResult, and return true.  If no simplification is possible,
273   /// returns false.
274   bool OptimizeOverflowCheck(Instruction::BinaryOps BinaryOp, bool IsSigned,
275                              Value *LHS, Value *RHS,
276                              Instruction &CtxI, Value *&OperationResult,
277                              Constant *&OverflowResult);
278 
279   Instruction *visitCallBase(CallBase &Call);
280   Instruction *tryOptimizeCall(CallInst *CI);
281   bool transformConstExprCastCall(CallBase &Call);
282   Instruction *transformCallThroughTrampoline(CallBase &Call,
283                                               IntrinsicInst &Tramp);
284 
285   // Return (a, b) if (LHS, RHS) is known to be (a, b) or (b, a).
286   // Otherwise, return std::nullopt
287   // Currently it matches:
288   // - LHS = (select c, a, b), RHS = (select c, b, a)
289   // - LHS = (phi [a, BB0], [b, BB1]), RHS = (phi [b, BB0], [a, BB1])
290   // - LHS = min(a, b), RHS = max(a, b)
291   std::optional<std::pair<Value *, Value *>> matchSymmetricPair(Value *LHS,
292                                                                 Value *RHS);
293 
294   Value *simplifyMaskedLoad(IntrinsicInst &II);
295   Instruction *simplifyMaskedStore(IntrinsicInst &II);
296   Instruction *simplifyMaskedGather(IntrinsicInst &II);
297   Instruction *simplifyMaskedScatter(IntrinsicInst &II);
298 
299   /// Transform (zext icmp) to bitwise / integer operations in order to
300   /// eliminate it.
301   ///
302   /// \param ICI The icmp of the (zext icmp) pair we are interested in.
303   /// \parem CI The zext of the (zext icmp) pair we are interested in.
304   ///
305   /// \return null if the transformation cannot be performed. If the
306   /// transformation can be performed the new instruction that replaces the
307   /// (zext icmp) pair will be returned.
308   Instruction *transformZExtICmp(ICmpInst *Cmp, ZExtInst &Zext);
309 
310   Instruction *transformSExtICmp(ICmpInst *Cmp, SExtInst &Sext);
311 
willNotOverflowSignedAdd(const WithCache<const Value * > & LHS,const WithCache<const Value * > & RHS,const Instruction & CxtI)312   bool willNotOverflowSignedAdd(const WithCache<const Value *> &LHS,
313                                 const WithCache<const Value *> &RHS,
314                                 const Instruction &CxtI) const {
315     return computeOverflowForSignedAdd(LHS, RHS, &CxtI) ==
316            OverflowResult::NeverOverflows;
317   }
318 
willNotOverflowUnsignedAdd(const WithCache<const Value * > & LHS,const WithCache<const Value * > & RHS,const Instruction & CxtI)319   bool willNotOverflowUnsignedAdd(const WithCache<const Value *> &LHS,
320                                   const WithCache<const Value *> &RHS,
321                                   const Instruction &CxtI) const {
322     return computeOverflowForUnsignedAdd(LHS, RHS, &CxtI) ==
323            OverflowResult::NeverOverflows;
324   }
325 
willNotOverflowAdd(const Value * LHS,const Value * RHS,const Instruction & CxtI,bool IsSigned)326   bool willNotOverflowAdd(const Value *LHS, const Value *RHS,
327                           const Instruction &CxtI, bool IsSigned) const {
328     return IsSigned ? willNotOverflowSignedAdd(LHS, RHS, CxtI)
329                     : willNotOverflowUnsignedAdd(LHS, RHS, CxtI);
330   }
331 
willNotOverflowSignedSub(const Value * LHS,const Value * RHS,const Instruction & CxtI)332   bool willNotOverflowSignedSub(const Value *LHS, const Value *RHS,
333                                 const Instruction &CxtI) const {
334     return computeOverflowForSignedSub(LHS, RHS, &CxtI) ==
335            OverflowResult::NeverOverflows;
336   }
337 
willNotOverflowUnsignedSub(const Value * LHS,const Value * RHS,const Instruction & CxtI)338   bool willNotOverflowUnsignedSub(const Value *LHS, const Value *RHS,
339                                   const Instruction &CxtI) const {
340     return computeOverflowForUnsignedSub(LHS, RHS, &CxtI) ==
341            OverflowResult::NeverOverflows;
342   }
343 
willNotOverflowSub(const Value * LHS,const Value * RHS,const Instruction & CxtI,bool IsSigned)344   bool willNotOverflowSub(const Value *LHS, const Value *RHS,
345                           const Instruction &CxtI, bool IsSigned) const {
346     return IsSigned ? willNotOverflowSignedSub(LHS, RHS, CxtI)
347                     : willNotOverflowUnsignedSub(LHS, RHS, CxtI);
348   }
349 
willNotOverflowSignedMul(const Value * LHS,const Value * RHS,const Instruction & CxtI)350   bool willNotOverflowSignedMul(const Value *LHS, const Value *RHS,
351                                 const Instruction &CxtI) const {
352     return computeOverflowForSignedMul(LHS, RHS, &CxtI) ==
353            OverflowResult::NeverOverflows;
354   }
355 
356   bool willNotOverflowUnsignedMul(const Value *LHS, const Value *RHS,
357                                   const Instruction &CxtI,
358                                   bool IsNSW = false) const {
359     return computeOverflowForUnsignedMul(LHS, RHS, &CxtI, IsNSW) ==
360            OverflowResult::NeverOverflows;
361   }
362 
willNotOverflowMul(const Value * LHS,const Value * RHS,const Instruction & CxtI,bool IsSigned)363   bool willNotOverflowMul(const Value *LHS, const Value *RHS,
364                           const Instruction &CxtI, bool IsSigned) const {
365     return IsSigned ? willNotOverflowSignedMul(LHS, RHS, CxtI)
366                     : willNotOverflowUnsignedMul(LHS, RHS, CxtI);
367   }
368 
willNotOverflow(BinaryOperator::BinaryOps Opcode,const Value * LHS,const Value * RHS,const Instruction & CxtI,bool IsSigned)369   bool willNotOverflow(BinaryOperator::BinaryOps Opcode, const Value *LHS,
370                        const Value *RHS, const Instruction &CxtI,
371                        bool IsSigned) const {
372     switch (Opcode) {
373     case Instruction::Add: return willNotOverflowAdd(LHS, RHS, CxtI, IsSigned);
374     case Instruction::Sub: return willNotOverflowSub(LHS, RHS, CxtI, IsSigned);
375     case Instruction::Mul: return willNotOverflowMul(LHS, RHS, CxtI, IsSigned);
376     default: llvm_unreachable("Unexpected opcode for overflow query");
377     }
378   }
379 
380   Value *EmitGEPOffset(GEPOperator *GEP, bool RewriteGEP = false);
381   Instruction *scalarizePHI(ExtractElementInst &EI, PHINode *PN);
382   Instruction *foldBitcastExtElt(ExtractElementInst &ExtElt);
383   Instruction *foldCastedBitwiseLogic(BinaryOperator &I);
384   Instruction *foldFBinOpOfIntCasts(BinaryOperator &I);
385   // Should only be called by `foldFBinOpOfIntCasts`.
386   Instruction *foldFBinOpOfIntCastsFromSign(
387       BinaryOperator &BO, bool OpsFromSigned, std::array<Value *, 2> IntOps,
388       Constant *Op1FpC, SmallVectorImpl<WithCache<const Value *>> &OpsKnown);
389   Instruction *foldBinopOfSextBoolToSelect(BinaryOperator &I);
390   Instruction *narrowBinOp(TruncInst &Trunc);
391   Instruction *narrowMaskedBinOp(BinaryOperator &And);
392   Instruction *narrowMathIfNoOverflow(BinaryOperator &I);
393   Instruction *narrowFunnelShift(TruncInst &Trunc);
394   Instruction *optimizeBitCastFromPhi(CastInst &CI, PHINode *PN);
395   Instruction *matchSAddSubSat(IntrinsicInst &MinMax1);
396   Instruction *foldNot(BinaryOperator &I);
397   Instruction *foldBinOpOfDisplacedShifts(BinaryOperator &I);
398 
399   /// Determine if a pair of casts can be replaced by a single cast.
400   ///
401   /// \param CI1 The first of a pair of casts.
402   /// \param CI2 The second of a pair of casts.
403   ///
404   /// \return 0 if the cast pair cannot be eliminated, otherwise returns an
405   /// Instruction::CastOps value for a cast that can replace the pair, casting
406   /// CI1->getSrcTy() to CI2->getDstTy().
407   ///
408   /// \see CastInst::isEliminableCastPair
409   Instruction::CastOps isEliminableCastPair(const CastInst *CI1,
410                                             const CastInst *CI2);
411   Value *simplifyIntToPtrRoundTripCast(Value *Val);
412 
413   Value *foldAndOrOfICmps(ICmpInst *LHS, ICmpInst *RHS, Instruction &I,
414                           bool IsAnd, bool IsLogical = false);
415   Value *foldXorOfICmps(ICmpInst *LHS, ICmpInst *RHS, BinaryOperator &Xor);
416 
417   Value *foldEqOfParts(ICmpInst *Cmp0, ICmpInst *Cmp1, bool IsAnd);
418 
419   Value *foldAndOrOfICmpsUsingRanges(ICmpInst *ICmp1, ICmpInst *ICmp2,
420                                      bool IsAnd);
421 
422   /// Optimize (fcmp)&(fcmp) or (fcmp)|(fcmp).
423   /// NOTE: Unlike most of instcombine, this returns a Value which should
424   /// already be inserted into the function.
425   Value *foldLogicOfFCmps(FCmpInst *LHS, FCmpInst *RHS, bool IsAnd,
426                           bool IsLogicalSelect = false);
427 
428   Instruction *foldLogicOfIsFPClass(BinaryOperator &Operator, Value *LHS,
429                                     Value *RHS);
430 
431   Instruction *
432   canonicalizeConditionalNegationViaMathToSelect(BinaryOperator &i);
433 
434   Value *foldAndOrOfICmpsOfAndWithPow2(ICmpInst *LHS, ICmpInst *RHS,
435                                        Instruction *CxtI, bool IsAnd,
436                                        bool IsLogical = false);
437   Value *matchSelectFromAndOr(Value *A, Value *B, Value *C, Value *D,
438                               bool InvertFalseVal = false);
439   Value *getSelectCondition(Value *A, Value *B, bool ABIsTheSame);
440 
441   Instruction *foldLShrOverflowBit(BinaryOperator &I);
442   Instruction *foldExtractOfOverflowIntrinsic(ExtractValueInst &EV);
443   Instruction *foldIntrinsicWithOverflowCommon(IntrinsicInst *II);
444   Instruction *foldIntrinsicIsFPClass(IntrinsicInst &II);
445   Instruction *foldFPSignBitOps(BinaryOperator &I);
446   Instruction *foldFDivConstantDivisor(BinaryOperator &I);
447 
448   // Optimize one of these forms:
449   //   and i1 Op, SI / select i1 Op, i1 SI, i1 false (if IsAnd = true)
450   //   or i1 Op, SI  / select i1 Op, i1 true, i1 SI  (if IsAnd = false)
451   // into simplier select instruction using isImpliedCondition.
452   Instruction *foldAndOrOfSelectUsingImpliedCond(Value *Op, SelectInst &SI,
453                                                  bool IsAnd);
454 
455   Instruction *hoistFNegAboveFMulFDiv(Value *FNegOp, Instruction &FMFSource);
456 
457 public:
458   /// Create and insert the idiom we use to indicate a block is unreachable
459   /// without having to rewrite the CFG from within InstCombine.
CreateNonTerminatorUnreachable(Instruction * InsertAt)460   void CreateNonTerminatorUnreachable(Instruction *InsertAt) {
461     auto &Ctx = InsertAt->getContext();
462     auto *SI = new StoreInst(ConstantInt::getTrue(Ctx),
463                              PoisonValue::get(PointerType::getUnqual(Ctx)),
464                              /*isVolatile*/ false, Align(1));
465     InsertNewInstWith(SI, InsertAt->getIterator());
466   }
467 
468   /// Combiner aware instruction erasure.
469   ///
470   /// When dealing with an instruction that has side effects or produces a void
471   /// value, we can't rely on DCE to delete the instruction. Instead, visit
472   /// methods should return the value returned by this function.
eraseInstFromFunction(Instruction & I)473   Instruction *eraseInstFromFunction(Instruction &I) override {
474     LLVM_DEBUG(dbgs() << "IC: ERASE " << I << '\n');
475     assert(I.use_empty() && "Cannot erase instruction that is used!");
476     salvageDebugInfo(I);
477 
478     // Make sure that we reprocess all operands now that we reduced their
479     // use counts.
480     SmallVector<Value *> Ops(I.operands());
481     Worklist.remove(&I);
482     DC.removeValue(&I);
483     I.eraseFromParent();
484     for (Value *Op : Ops)
485       Worklist.handleUseCountDecrement(Op);
486     MadeIRChange = true;
487     return nullptr; // Don't do anything with FI
488   }
489 
490   OverflowResult computeOverflow(
491       Instruction::BinaryOps BinaryOp, bool IsSigned,
492       Value *LHS, Value *RHS, Instruction *CxtI) const;
493 
494   /// Performs a few simplifications for operators which are associative
495   /// or commutative.
496   bool SimplifyAssociativeOrCommutative(BinaryOperator &I);
497 
498   /// Tries to simplify binary operations which some other binary
499   /// operation distributes over.
500   ///
501   /// It does this by either by factorizing out common terms (eg "(A*B)+(A*C)"
502   /// -> "A*(B+C)") or expanding out if this results in simplifications (eg: "A
503   /// & (B | C) -> (A&B) | (A&C)" if this is a win).  Returns the simplified
504   /// value, or null if it didn't simplify.
505   Value *foldUsingDistributiveLaws(BinaryOperator &I);
506 
507   /// Tries to simplify add operations using the definition of remainder.
508   ///
509   /// The definition of remainder is X % C = X - (X / C ) * C. The add
510   /// expression X % C0 + (( X / C0 ) % C1) * C0 can be simplified to
511   /// X % (C0 * C1)
512   Value *SimplifyAddWithRemainder(BinaryOperator &I);
513 
514   // Binary Op helper for select operations where the expression can be
515   // efficiently reorganized.
516   Value *SimplifySelectsFeedingBinaryOp(BinaryOperator &I, Value *LHS,
517                                         Value *RHS);
518 
519   // If `I` has operand `(ctpop (not x))`, fold `I` with `(sub nuw nsw
520   // BitWidth(x), (ctpop x))`.
521   Instruction *tryFoldInstWithCtpopWithNot(Instruction *I);
522 
523   // (Binop1 (Binop2 (logic_shift X, C), C1), (logic_shift Y, C))
524   //    -> (logic_shift (Binop1 (Binop2 X, inv_logic_shift(C1, C)), Y), C)
525   // (Binop1 (Binop2 (logic_shift X, Amt), Mask), (logic_shift Y, Amt))
526   //    -> (BinOp (logic_shift (BinOp X, Y)), Mask)
527   Instruction *foldBinOpShiftWithShift(BinaryOperator &I);
528 
529   /// Tries to simplify binops of select and cast of the select condition.
530   ///
531   /// (Binop (cast C), (select C, T, F))
532   ///    -> (select C, C0, C1)
533   Instruction *foldBinOpOfSelectAndCastOfSelectCondition(BinaryOperator &I);
534 
535   /// This tries to simplify binary operations by factorizing out common terms
536   /// (e. g. "(A*B)+(A*C)" -> "A*(B+C)").
537   Value *tryFactorizationFolds(BinaryOperator &I);
538 
539   /// Match a select chain which produces one of three values based on whether
540   /// the LHS is less than, equal to, or greater than RHS respectively.
541   /// Return true if we matched a three way compare idiom. The LHS, RHS, Less,
542   /// Equal and Greater values are saved in the matching process and returned to
543   /// the caller.
544   bool matchThreeWayIntCompare(SelectInst *SI, Value *&LHS, Value *&RHS,
545                                ConstantInt *&Less, ConstantInt *&Equal,
546                                ConstantInt *&Greater);
547 
548   /// Attempts to replace I with a simpler value based on the demanded
549   /// bits.
550   Value *SimplifyDemandedUseBits(Instruction *I, const APInt &DemandedMask,
551                                  KnownBits &Known, unsigned Depth,
552                                  const SimplifyQuery &Q);
553   using InstCombiner::SimplifyDemandedBits;
554   bool SimplifyDemandedBits(Instruction *I, unsigned Op,
555                             const APInt &DemandedMask, KnownBits &Known,
556                             unsigned Depth, const SimplifyQuery &Q) override;
557 
558   /// Helper routine of SimplifyDemandedUseBits. It computes KnownZero/KnownOne
559   /// bits. It also tries to handle simplifications that can be done based on
560   /// DemandedMask, but without modifying the Instruction.
561   Value *SimplifyMultipleUseDemandedBits(Instruction *I,
562                                          const APInt &DemandedMask,
563                                          KnownBits &Known, unsigned Depth,
564                                          const SimplifyQuery &Q);
565 
566   /// Helper routine of SimplifyDemandedUseBits. It tries to simplify demanded
567   /// bit for "r1 = shr x, c1; r2 = shl r1, c2" instruction sequence.
568   Value *simplifyShrShlDemandedBits(
569       Instruction *Shr, const APInt &ShrOp1, Instruction *Shl,
570       const APInt &ShlOp1, const APInt &DemandedMask, KnownBits &Known);
571 
572   /// Tries to simplify operands to an integer instruction based on its
573   /// demanded bits.
574   bool SimplifyDemandedInstructionBits(Instruction &Inst);
575   bool SimplifyDemandedInstructionBits(Instruction &Inst, KnownBits &Known);
576 
577   Value *SimplifyDemandedVectorElts(Value *V, APInt DemandedElts,
578                                     APInt &PoisonElts, unsigned Depth = 0,
579                                     bool AllowMultipleUsers = false) override;
580 
581   /// Attempts to replace V with a simpler value based on the demanded
582   /// floating-point classes
583   Value *SimplifyDemandedUseFPClass(Value *V, FPClassTest DemandedMask,
584                                     KnownFPClass &Known, unsigned Depth,
585                                     Instruction *CxtI);
586   bool SimplifyDemandedFPClass(Instruction *I, unsigned Op,
587                                FPClassTest DemandedMask, KnownFPClass &Known,
588                                unsigned Depth = 0);
589 
590   /// Canonicalize the position of binops relative to shufflevector.
591   Instruction *foldVectorBinop(BinaryOperator &Inst);
592   Instruction *foldVectorSelect(SelectInst &Sel);
593   Instruction *foldSelectShuffle(ShuffleVectorInst &Shuf);
594 
595   /// Given a binary operator, cast instruction, or select which has a PHI node
596   /// as operand #0, see if we can fold the instruction into the PHI (which is
597   /// only possible if all operands to the PHI are constants).
598   Instruction *foldOpIntoPhi(Instruction &I, PHINode *PN);
599 
600   /// For a binary operator with 2 phi operands, try to hoist the binary
601   /// operation before the phi. This can result in fewer instructions in
602   /// patterns where at least one set of phi operands simplifies.
603   /// Example:
604   /// BB3: binop (phi [X, BB1], [C1, BB2]), (phi [Y, BB1], [C2, BB2])
605   /// -->
606   /// BB1: BO = binop X, Y
607   /// BB3: phi [BO, BB1], [(binop C1, C2), BB2]
608   Instruction *foldBinopWithPhiOperands(BinaryOperator &BO);
609 
610   /// Given an instruction with a select as one operand and a constant as the
611   /// other operand, try to fold the binary operator into the select arguments.
612   /// This also works for Cast instructions, which obviously do not have a
613   /// second operand.
614   Instruction *FoldOpIntoSelect(Instruction &Op, SelectInst *SI,
615                                 bool FoldWithMultiUse = false);
616 
617   /// This is a convenience wrapper function for the above two functions.
618   Instruction *foldBinOpIntoSelectOrPhi(BinaryOperator &I);
619 
620   Instruction *foldAddWithConstant(BinaryOperator &Add);
621 
622   Instruction *foldSquareSumInt(BinaryOperator &I);
623   Instruction *foldSquareSumFP(BinaryOperator &I);
624 
625   /// Try to rotate an operation below a PHI node, using PHI nodes for
626   /// its operands.
627   Instruction *foldPHIArgOpIntoPHI(PHINode &PN);
628   Instruction *foldPHIArgBinOpIntoPHI(PHINode &PN);
629   Instruction *foldPHIArgInsertValueInstructionIntoPHI(PHINode &PN);
630   Instruction *foldPHIArgExtractValueInstructionIntoPHI(PHINode &PN);
631   Instruction *foldPHIArgGEPIntoPHI(PHINode &PN);
632   Instruction *foldPHIArgLoadIntoPHI(PHINode &PN);
633   Instruction *foldPHIArgZextsIntoPHI(PHINode &PN);
634   Instruction *foldPHIArgIntToPtrToPHI(PHINode &PN);
635 
636   /// If an integer typed PHI has only one use which is an IntToPtr operation,
637   /// replace the PHI with an existing pointer typed PHI if it exists. Otherwise
638   /// insert a new pointer typed PHI and replace the original one.
639   bool foldIntegerTypedPHI(PHINode &PN);
640 
641   /// Helper function for FoldPHIArgXIntoPHI() to set debug location for the
642   /// folded operation.
643   void PHIArgMergedDebugLoc(Instruction *Inst, PHINode &PN);
644 
645   Instruction *foldGEPICmp(GEPOperator *GEPLHS, Value *RHS,
646                            ICmpInst::Predicate Cond, Instruction &I);
647   Instruction *foldSelectICmp(ICmpInst::Predicate Pred, SelectInst *SI,
648                               Value *RHS, const ICmpInst &I);
649   bool foldAllocaCmp(AllocaInst *Alloca);
650   Instruction *foldCmpLoadFromIndexedGlobal(LoadInst *LI,
651                                             GetElementPtrInst *GEP,
652                                             GlobalVariable *GV, CmpInst &ICI,
653                                             ConstantInt *AndCst = nullptr);
654   Instruction *foldFCmpIntToFPConst(FCmpInst &I, Instruction *LHSI,
655                                     Constant *RHSC);
656   Instruction *foldICmpAddOpConst(Value *X, const APInt &C,
657                                   ICmpInst::Predicate Pred);
658   Instruction *foldICmpWithCastOp(ICmpInst &ICmp);
659   Instruction *foldICmpWithZextOrSext(ICmpInst &ICmp);
660 
661   Instruction *foldICmpUsingKnownBits(ICmpInst &Cmp);
662   Instruction *foldICmpWithDominatingICmp(ICmpInst &Cmp);
663   Instruction *foldICmpWithConstant(ICmpInst &Cmp);
664   Instruction *foldICmpUsingBoolRange(ICmpInst &I);
665   Instruction *foldICmpInstWithConstant(ICmpInst &Cmp);
666   Instruction *foldICmpInstWithConstantNotInt(ICmpInst &Cmp);
667   Instruction *foldICmpInstWithConstantAllowPoison(ICmpInst &Cmp,
668                                                    const APInt &C);
669   Instruction *foldICmpBinOp(ICmpInst &Cmp, const SimplifyQuery &SQ);
670   Instruction *foldICmpWithMinMax(Instruction &I, MinMaxIntrinsic *MinMax,
671                                   Value *Z, ICmpInst::Predicate Pred);
672   Instruction *foldICmpEquality(ICmpInst &Cmp);
673   Instruction *foldIRemByPowerOfTwoToBitTest(ICmpInst &I);
674   Instruction *foldSignBitTest(ICmpInst &I);
675   Instruction *foldICmpWithZero(ICmpInst &Cmp);
676 
677   Value *foldMultiplicationOverflowCheck(ICmpInst &Cmp);
678 
679   Instruction *foldICmpBinOpWithConstant(ICmpInst &Cmp, BinaryOperator *BO,
680                                          const APInt &C);
681   Instruction *foldICmpSelectConstant(ICmpInst &Cmp, SelectInst *Select,
682                                       ConstantInt *C);
683   Instruction *foldICmpTruncConstant(ICmpInst &Cmp, TruncInst *Trunc,
684                                      const APInt &C);
685   Instruction *foldICmpTruncWithTruncOrExt(ICmpInst &Cmp,
686                                            const SimplifyQuery &Q);
687   Instruction *foldICmpAndConstant(ICmpInst &Cmp, BinaryOperator *And,
688                                    const APInt &C);
689   Instruction *foldICmpXorConstant(ICmpInst &Cmp, BinaryOperator *Xor,
690                                    const APInt &C);
691   Instruction *foldICmpOrConstant(ICmpInst &Cmp, BinaryOperator *Or,
692                                   const APInt &C);
693   Instruction *foldICmpMulConstant(ICmpInst &Cmp, BinaryOperator *Mul,
694                                    const APInt &C);
695   Instruction *foldICmpShlConstant(ICmpInst &Cmp, BinaryOperator *Shl,
696                                    const APInt &C);
697   Instruction *foldICmpShrConstant(ICmpInst &Cmp, BinaryOperator *Shr,
698                                    const APInt &C);
699   Instruction *foldICmpSRemConstant(ICmpInst &Cmp, BinaryOperator *UDiv,
700                                     const APInt &C);
701   Instruction *foldICmpUDivConstant(ICmpInst &Cmp, BinaryOperator *UDiv,
702                                     const APInt &C);
703   Instruction *foldICmpDivConstant(ICmpInst &Cmp, BinaryOperator *Div,
704                                    const APInt &C);
705   Instruction *foldICmpSubConstant(ICmpInst &Cmp, BinaryOperator *Sub,
706                                    const APInt &C);
707   Instruction *foldICmpAddConstant(ICmpInst &Cmp, BinaryOperator *Add,
708                                    const APInt &C);
709   Instruction *foldICmpAndConstConst(ICmpInst &Cmp, BinaryOperator *And,
710                                      const APInt &C1);
711   Instruction *foldICmpAndShift(ICmpInst &Cmp, BinaryOperator *And,
712                                 const APInt &C1, const APInt &C2);
713   Instruction *foldICmpXorShiftConst(ICmpInst &Cmp, BinaryOperator *Xor,
714                                      const APInt &C);
715   Instruction *foldICmpShrConstConst(ICmpInst &I, Value *ShAmt, const APInt &C1,
716                                      const APInt &C2);
717   Instruction *foldICmpShlConstConst(ICmpInst &I, Value *ShAmt, const APInt &C1,
718                                      const APInt &C2);
719 
720   Instruction *foldICmpBinOpEqualityWithConstant(ICmpInst &Cmp,
721                                                  BinaryOperator *BO,
722                                                  const APInt &C);
723   Instruction *foldICmpIntrinsicWithConstant(ICmpInst &ICI, IntrinsicInst *II,
724                                              const APInt &C);
725   Instruction *foldICmpEqIntrinsicWithConstant(ICmpInst &ICI, IntrinsicInst *II,
726                                                const APInt &C);
727   Instruction *foldICmpBitCast(ICmpInst &Cmp);
728   Instruction *foldICmpWithTrunc(ICmpInst &Cmp);
729   Instruction *foldICmpCommutative(ICmpInst::Predicate Pred, Value *Op0,
730                                    Value *Op1, ICmpInst &CxtI);
731 
732   // Helpers of visitSelectInst().
733   Instruction *foldSelectOfBools(SelectInst &SI);
734   Instruction *foldSelectExtConst(SelectInst &Sel);
735   Instruction *foldSelectOpOp(SelectInst &SI, Instruction *TI, Instruction *FI);
736   Instruction *foldSelectIntoOp(SelectInst &SI, Value *, Value *);
737   Instruction *foldSPFofSPF(Instruction *Inner, SelectPatternFlavor SPF1,
738                             Value *A, Value *B, Instruction &Outer,
739                             SelectPatternFlavor SPF2, Value *C);
740   Instruction *foldSelectInstWithICmp(SelectInst &SI, ICmpInst *ICI);
741   Instruction *foldSelectValueEquivalence(SelectInst &SI, ICmpInst &ICI);
742   bool replaceInInstruction(Value *V, Value *Old, Value *New,
743                             unsigned Depth = 0);
744 
745   Value *insertRangeTest(Value *V, const APInt &Lo, const APInt &Hi,
746                          bool isSigned, bool Inside);
747   bool mergeStoreIntoSuccessor(StoreInst &SI);
748 
749   /// Given an initial instruction, check to see if it is the root of a
750   /// bswap/bitreverse idiom. If so, return the equivalent bswap/bitreverse
751   /// intrinsic.
752   Instruction *matchBSwapOrBitReverse(Instruction &I, bool MatchBSwaps,
753                                       bool MatchBitReversals);
754 
755   Instruction *SimplifyAnyMemTransfer(AnyMemTransferInst *MI);
756   Instruction *SimplifyAnyMemSet(AnyMemSetInst *MI);
757 
758   Value *EvaluateInDifferentType(Value *V, Type *Ty, bool isSigned);
759 
760   bool tryToSinkInstruction(Instruction *I, BasicBlock *DestBlock);
761   void tryToSinkInstructionDbgValues(
762       Instruction *I, BasicBlock::iterator InsertPos, BasicBlock *SrcBlock,
763       BasicBlock *DestBlock, SmallVectorImpl<DbgVariableIntrinsic *> &DbgUsers);
764   void tryToSinkInstructionDbgVariableRecords(
765       Instruction *I, BasicBlock::iterator InsertPos, BasicBlock *SrcBlock,
766       BasicBlock *DestBlock, SmallVectorImpl<DbgVariableRecord *> &DPUsers);
767 
768   bool removeInstructionsBeforeUnreachable(Instruction &I);
769   void addDeadEdge(BasicBlock *From, BasicBlock *To,
770                    SmallVectorImpl<BasicBlock *> &Worklist);
771   void handleUnreachableFrom(Instruction *I,
772                              SmallVectorImpl<BasicBlock *> &Worklist);
773   void handlePotentiallyDeadBlocks(SmallVectorImpl<BasicBlock *> &Worklist);
774   void handlePotentiallyDeadSuccessors(BasicBlock *BB, BasicBlock *LiveSucc);
775   void freelyInvertAllUsersOf(Value *V, Value *IgnoredUser = nullptr);
776 };
777 
778 class Negator final {
779   /// Top-to-bottom, def-to-use negated instruction tree we produced.
780   SmallVector<Instruction *, NegatorMaxNodesSSO> NewInstructions;
781 
782   using BuilderTy = IRBuilder<TargetFolder, IRBuilderCallbackInserter>;
783   BuilderTy Builder;
784 
785   const bool IsTrulyNegation;
786 
787   SmallDenseMap<Value *, Value *> NegationsCache;
788 
789   Negator(LLVMContext &C, const DataLayout &DL, bool IsTrulyNegation);
790 
791 #if LLVM_ENABLE_STATS
792   unsigned NumValuesVisitedInThisNegator = 0;
793   ~Negator();
794 #endif
795 
796   using Result = std::pair<ArrayRef<Instruction *> /*NewInstructions*/,
797                            Value * /*NegatedRoot*/>;
798 
799   std::array<Value *, 2> getSortedOperandsOfBinOp(Instruction *I);
800 
801   [[nodiscard]] Value *visitImpl(Value *V, bool IsNSW, unsigned Depth);
802 
803   [[nodiscard]] Value *negate(Value *V, bool IsNSW, unsigned Depth);
804 
805   /// Recurse depth-first and attempt to sink the negation.
806   /// FIXME: use worklist?
807   [[nodiscard]] std::optional<Result> run(Value *Root, bool IsNSW);
808 
809   Negator(const Negator &) = delete;
810   Negator(Negator &&) = delete;
811   Negator &operator=(const Negator &) = delete;
812   Negator &operator=(Negator &&) = delete;
813 
814 public:
815   /// Attempt to negate \p Root. Retuns nullptr if negation can't be performed,
816   /// otherwise returns negated value.
817   [[nodiscard]] static Value *Negate(bool LHSIsZero, bool IsNSW, Value *Root,
818                                      InstCombinerImpl &IC);
819 };
820 
821 } // end namespace llvm
822 
823 #undef DEBUG_TYPE
824 
825 #endif // LLVM_LIB_TRANSFORMS_INSTCOMBINE_INSTCOMBINEINTERNAL_H
826