1 //===- InstCombine.h - Main InstCombine pass definition ---------*- 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 //===----------------------------------------------------------------------===//
10 #ifndef LLVM_LIB_TRANSFORMS_INSTCOMBINE_INSTCOMBINE_H
11 #define LLVM_LIB_TRANSFORMS_INSTCOMBINE_INSTCOMBINE_H
13 #include "InstCombineWorklist.h"
14 #include "llvm/Analysis/TargetFolder.h"
15 #include "llvm/Analysis/ValueTracking.h"
16 #include "llvm/IR/IRBuilder.h"
17 #include "llvm/IR/InstVisitor.h"
18 #include "llvm/IR/IntrinsicInst.h"
19 #include "llvm/IR/Operator.h"
20 #include "llvm/Pass.h"
21 #include "llvm/Transforms/Utils/SimplifyLibCalls.h"
23 #define DEBUG_TYPE "instcombine"
28 class TargetLibraryInfo;
33 /// SelectPatternFlavor - We can match a variety of different patterns for
34 /// select operations.
35 enum SelectPatternFlavor {
45 /// getComplexity: Assign a complexity or rank value to LLVM Values...
46 /// 0 -> undef, 1 -> Const, 2 -> Other, 3 -> Arg, 3 -> Unary, 4 -> OtherInst
47 static inline unsigned getComplexity(Value *V) {
48 if (isa<Instruction>(V)) {
49 if (BinaryOperator::isNeg(V) || BinaryOperator::isFNeg(V) ||
50 BinaryOperator::isNot(V))
56 return isa<Constant>(V) ? (isa<UndefValue>(V) ? 0 : 1) : 2;
59 /// AddOne - Add one to a Constant
60 static inline Constant *AddOne(Constant *C) {
61 return ConstantExpr::getAdd(C, ConstantInt::get(C->getType(), 1));
63 /// SubOne - Subtract one from a Constant
64 static inline Constant *SubOne(Constant *C) {
65 return ConstantExpr::getSub(C, ConstantInt::get(C->getType(), 1));
68 /// InstCombineIRInserter - This is an IRBuilder insertion helper that works
69 /// just like the normal insertion helper, but also adds any new instructions
70 /// to the instcombine worklist.
71 class LLVM_LIBRARY_VISIBILITY InstCombineIRInserter
72 : public IRBuilderDefaultInserter<true> {
73 InstCombineWorklist &Worklist;
76 InstCombineIRInserter(InstCombineWorklist &WL) : Worklist(WL) {}
78 void InsertHelper(Instruction *I, const Twine &Name, BasicBlock *BB,
79 BasicBlock::iterator InsertPt) const {
80 IRBuilderDefaultInserter<true>::InsertHelper(I, Name, BB, InsertPt);
85 /// InstCombiner - The -instcombine pass.
86 class LLVM_LIBRARY_VISIBILITY InstCombiner
87 : public FunctionPass,
88 public InstVisitor<InstCombiner, Instruction *> {
90 TargetLibraryInfo *TLI;
92 LibCallSimplifier *Simplifier;
96 /// Worklist - All of the instructions that need to be simplified.
97 InstCombineWorklist Worklist;
99 /// Builder - This is an IRBuilder that automatically inserts new
100 /// instructions into the worklist when they are created.
101 typedef IRBuilder<true, TargetFolder, InstCombineIRInserter> BuilderTy;
104 static char ID; // Pass identification, replacement for typeid
105 InstCombiner() : FunctionPass(ID), DL(nullptr), Builder(nullptr) {
106 MinimizeSize = false;
107 initializeInstCombinerPass(*PassRegistry::getPassRegistry());
111 bool runOnFunction(Function &F) override;
113 bool DoOneIteration(Function &F, unsigned ItNum);
115 void getAnalysisUsage(AnalysisUsage &AU) const override;
117 const DataLayout *getDataLayout() const { return DL; }
119 TargetLibraryInfo *getTargetLibraryInfo() const { return TLI; }
121 // Visitation implementation - Implement instruction combining for different
122 // instruction types. The semantics are as follows:
124 // null - No change was made
125 // I - Change was made, I is still valid, I may be dead though
126 // otherwise - Change was made, replace I with returned instruction
128 Instruction *visitAdd(BinaryOperator &I);
129 Instruction *visitFAdd(BinaryOperator &I);
130 Value *OptimizePointerDifference(Value *LHS, Value *RHS, Type *Ty);
131 Instruction *visitSub(BinaryOperator &I);
132 Instruction *visitFSub(BinaryOperator &I);
133 Instruction *visitMul(BinaryOperator &I);
134 Value *foldFMulConst(Instruction *FMulOrDiv, Constant *C,
135 Instruction *InsertBefore);
136 Instruction *visitFMul(BinaryOperator &I);
137 Instruction *visitURem(BinaryOperator &I);
138 Instruction *visitSRem(BinaryOperator &I);
139 Instruction *visitFRem(BinaryOperator &I);
140 bool SimplifyDivRemOfSelect(BinaryOperator &I);
141 Instruction *commonRemTransforms(BinaryOperator &I);
142 Instruction *commonIRemTransforms(BinaryOperator &I);
143 Instruction *commonDivTransforms(BinaryOperator &I);
144 Instruction *commonIDivTransforms(BinaryOperator &I);
145 Instruction *visitUDiv(BinaryOperator &I);
146 Instruction *visitSDiv(BinaryOperator &I);
147 Instruction *visitFDiv(BinaryOperator &I);
148 Value *FoldAndOfICmps(ICmpInst *LHS, ICmpInst *RHS);
149 Value *FoldAndOfFCmps(FCmpInst *LHS, FCmpInst *RHS);
150 Instruction *visitAnd(BinaryOperator &I);
151 Value *FoldOrOfICmps(ICmpInst *LHS, ICmpInst *RHS);
152 Value *FoldOrOfFCmps(FCmpInst *LHS, FCmpInst *RHS);
153 Instruction *FoldOrWithConstants(BinaryOperator &I, Value *Op, Value *A,
155 Instruction *visitOr(BinaryOperator &I);
156 Instruction *visitXor(BinaryOperator &I);
157 Instruction *visitShl(BinaryOperator &I);
158 Instruction *visitAShr(BinaryOperator &I);
159 Instruction *visitLShr(BinaryOperator &I);
160 Instruction *commonShiftTransforms(BinaryOperator &I);
161 Instruction *FoldFCmp_IntToFP_Cst(FCmpInst &I, Instruction *LHSI,
163 Instruction *FoldCmpLoadFromIndexedGlobal(GetElementPtrInst *GEP,
164 GlobalVariable *GV, CmpInst &ICI,
165 ConstantInt *AndCst = nullptr);
166 Instruction *visitFCmpInst(FCmpInst &I);
167 Instruction *visitICmpInst(ICmpInst &I);
168 Instruction *visitICmpInstWithCastAndCast(ICmpInst &ICI);
169 Instruction *visitICmpInstWithInstAndIntCst(ICmpInst &ICI, Instruction *LHS,
171 Instruction *FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
172 ConstantInt *DivRHS);
173 Instruction *FoldICmpShrCst(ICmpInst &ICI, BinaryOperator *DivI,
174 ConstantInt *DivRHS);
175 Instruction *FoldICmpCstShrCst(ICmpInst &I, Value *Op, Value *A,
176 ConstantInt *CI1, ConstantInt *CI2);
177 Instruction *FoldICmpAddOpCst(Instruction &ICI, Value *X, ConstantInt *CI,
178 ICmpInst::Predicate Pred);
179 Instruction *FoldGEPICmp(GEPOperator *GEPLHS, Value *RHS,
180 ICmpInst::Predicate Cond, Instruction &I);
181 Instruction *FoldShiftByConstant(Value *Op0, Constant *Op1,
183 Instruction *commonCastTransforms(CastInst &CI);
184 Instruction *commonPointerCastTransforms(CastInst &CI);
185 Instruction *visitTrunc(TruncInst &CI);
186 Instruction *visitZExt(ZExtInst &CI);
187 Instruction *visitSExt(SExtInst &CI);
188 Instruction *visitFPTrunc(FPTruncInst &CI);
189 Instruction *visitFPExt(CastInst &CI);
190 Instruction *visitFPToUI(FPToUIInst &FI);
191 Instruction *visitFPToSI(FPToSIInst &FI);
192 Instruction *visitUIToFP(CastInst &CI);
193 Instruction *visitSIToFP(CastInst &CI);
194 Instruction *visitPtrToInt(PtrToIntInst &CI);
195 Instruction *visitIntToPtr(IntToPtrInst &CI);
196 Instruction *visitBitCast(BitCastInst &CI);
197 Instruction *visitAddrSpaceCast(AddrSpaceCastInst &CI);
198 Instruction *FoldSelectOpOp(SelectInst &SI, Instruction *TI, Instruction *FI);
199 Instruction *FoldSelectIntoOp(SelectInst &SI, Value *, Value *);
200 Instruction *FoldSPFofSPF(Instruction *Inner, SelectPatternFlavor SPF1,
201 Value *A, Value *B, Instruction &Outer,
202 SelectPatternFlavor SPF2, Value *C);
203 Instruction *visitSelectInst(SelectInst &SI);
204 Instruction *visitSelectInstWithICmp(SelectInst &SI, ICmpInst *ICI);
205 Instruction *visitCallInst(CallInst &CI);
206 Instruction *visitInvokeInst(InvokeInst &II);
208 Instruction *SliceUpIllegalIntegerPHI(PHINode &PN);
209 Instruction *visitPHINode(PHINode &PN);
210 Instruction *visitGetElementPtrInst(GetElementPtrInst &GEP);
211 Instruction *visitAllocaInst(AllocaInst &AI);
212 Instruction *visitAllocSite(Instruction &FI);
213 Instruction *visitFree(CallInst &FI);
214 Instruction *visitLoadInst(LoadInst &LI);
215 Instruction *visitStoreInst(StoreInst &SI);
216 Instruction *visitBranchInst(BranchInst &BI);
217 Instruction *visitSwitchInst(SwitchInst &SI);
218 Instruction *visitInsertValueInst(InsertValueInst &IV);
219 Instruction *visitInsertElementInst(InsertElementInst &IE);
220 Instruction *visitExtractElementInst(ExtractElementInst &EI);
221 Instruction *visitShuffleVectorInst(ShuffleVectorInst &SVI);
222 Instruction *visitExtractValueInst(ExtractValueInst &EV);
223 Instruction *visitLandingPadInst(LandingPadInst &LI);
225 // visitInstruction - Specify what to return for unhandled instructions...
226 Instruction *visitInstruction(Instruction &I) { return nullptr; }
229 bool ShouldChangeType(Type *From, Type *To) const;
230 Value *dyn_castNegVal(Value *V) const;
231 Value *dyn_castFNegVal(Value *V, bool NoSignedZero = false) const;
232 Type *FindElementAtOffset(Type *PtrTy, int64_t Offset,
233 SmallVectorImpl<Value *> &NewIndices);
234 Instruction *FoldOpIntoSelect(Instruction &Op, SelectInst *SI);
236 /// ShouldOptimizeCast - Return true if the cast from "V to Ty" actually
237 /// results in any code being generated and is interesting to optimize out. If
238 /// the cast can be eliminated by some other simple transformation, we prefer
239 /// to do the simplification first.
240 bool ShouldOptimizeCast(Instruction::CastOps opcode, const Value *V,
243 Instruction *visitCallSite(CallSite CS);
244 Instruction *tryOptimizeCall(CallInst *CI, const DataLayout *DL);
245 bool transformConstExprCastCall(CallSite CS);
246 Instruction *transformCallThroughTrampoline(CallSite CS,
247 IntrinsicInst *Tramp);
248 Instruction *transformZExtICmp(ICmpInst *ICI, Instruction &CI,
249 bool DoXform = true);
250 Instruction *transformSExtICmp(ICmpInst *ICI, Instruction &CI);
251 bool WillNotOverflowSignedAdd(Value *LHS, Value *RHS);
252 bool WillNotOverflowUnsignedAdd(Value *LHS, Value *RHS);
253 Value *EmitGEPOffset(User *GEP);
254 Instruction *scalarizePHI(ExtractElementInst &EI, PHINode *PN);
255 Value *EvaluateInDifferentElementOrder(Value *V, ArrayRef<int> Mask);
258 // InsertNewInstBefore - insert an instruction New before instruction Old
259 // in the program. Add the new instruction to the worklist.
261 Instruction *InsertNewInstBefore(Instruction *New, Instruction &Old) {
262 assert(New && !New->getParent() &&
263 "New instruction already inserted into a basic block!");
264 BasicBlock *BB = Old.getParent();
265 BB->getInstList().insert(&Old, New); // Insert inst
270 // InsertNewInstWith - same as InsertNewInstBefore, but also sets the
273 Instruction *InsertNewInstWith(Instruction *New, Instruction &Old) {
274 New->setDebugLoc(Old.getDebugLoc());
275 return InsertNewInstBefore(New, Old);
278 // ReplaceInstUsesWith - This method is to be used when an instruction is
279 // found to be dead, replacable with another preexisting expression. Here
280 // we add all uses of I to the worklist, replace all uses of I with the new
281 // value, then return I, so that the inst combiner will know that I was
284 Instruction *ReplaceInstUsesWith(Instruction &I, Value *V) {
285 Worklist.AddUsersToWorkList(I); // Add all modified instrs to worklist.
287 // If we are replacing the instruction with itself, this must be in a
288 // segment of unreachable code, so just clobber the instruction.
290 V = UndefValue::get(I.getType());
292 DEBUG(dbgs() << "IC: Replacing " << I << "\n"
293 " with " << *V << '\n');
295 I.replaceAllUsesWith(V);
299 // EraseInstFromFunction - When dealing with an instruction that has side
300 // effects or produces a void value, we can't rely on DCE to delete the
301 // instruction. Instead, visit methods should return the value returned by
303 Instruction *EraseInstFromFunction(Instruction &I) {
304 DEBUG(dbgs() << "IC: ERASE " << I << '\n');
306 assert(I.use_empty() && "Cannot erase instruction that is used!");
307 // Make sure that we reprocess all operands now that we reduced their
309 if (I.getNumOperands() < 8) {
310 for (User::op_iterator i = I.op_begin(), e = I.op_end(); i != e; ++i)
311 if (Instruction *Op = dyn_cast<Instruction>(*i))
317 return nullptr; // Don't do anything with FI
320 void computeKnownBits(Value *V, APInt &KnownZero, APInt &KnownOne,
321 unsigned Depth = 0) const {
322 return llvm::computeKnownBits(V, KnownZero, KnownOne, DL, Depth);
325 bool MaskedValueIsZero(Value *V, const APInt &Mask,
326 unsigned Depth = 0) const {
327 return llvm::MaskedValueIsZero(V, Mask, DL, Depth);
329 unsigned ComputeNumSignBits(Value *Op, unsigned Depth = 0) const {
330 return llvm::ComputeNumSignBits(Op, DL, Depth);
334 /// SimplifyAssociativeOrCommutative - This performs a few simplifications for
335 /// operators which are associative or commutative.
336 bool SimplifyAssociativeOrCommutative(BinaryOperator &I);
338 /// SimplifyUsingDistributiveLaws - This tries to simplify binary operations
339 /// which some other binary operation distributes over either by factorizing
340 /// out common terms (eg "(A*B)+(A*C)" -> "A*(B+C)") or expanding out if this
341 /// results in simplifications (eg: "A & (B | C) -> (A&B) | (A&C)" if this is
342 /// a win). Returns the simplified value, or null if it didn't simplify.
343 Value *SimplifyUsingDistributiveLaws(BinaryOperator &I);
345 /// SimplifyDemandedUseBits - Attempts to replace V with a simpler value
346 /// based on the demanded bits.
347 Value *SimplifyDemandedUseBits(Value *V, APInt DemandedMask, APInt &KnownZero,
348 APInt &KnownOne, unsigned Depth);
349 bool SimplifyDemandedBits(Use &U, APInt DemandedMask, APInt &KnownZero,
350 APInt &KnownOne, unsigned Depth = 0);
351 /// Helper routine of SimplifyDemandedUseBits. It tries to simplify demanded
352 /// bit for "r1 = shr x, c1; r2 = shl r1, c2" instruction sequence.
353 Value *SimplifyShrShlDemandedBits(Instruction *Lsr, Instruction *Sftl,
354 APInt DemandedMask, APInt &KnownZero,
357 /// SimplifyDemandedInstructionBits - Inst is an integer instruction that
358 /// SimplifyDemandedBits knows about. See if the instruction has any
359 /// properties that allow us to simplify its operands.
360 bool SimplifyDemandedInstructionBits(Instruction &Inst);
362 Value *SimplifyDemandedVectorElts(Value *V, APInt DemandedElts,
363 APInt &UndefElts, unsigned Depth = 0);
365 Value *SimplifyVectorOp(BinaryOperator &Inst);
367 // FoldOpIntoPhi - Given a binary operator, cast instruction, or select
368 // which has a PHI node as operand #0, see if we can fold the instruction
369 // into the PHI (which is only possible if all operands to the PHI are
372 Instruction *FoldOpIntoPhi(Instruction &I);
374 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
375 // operator and they all are only used by the PHI, PHI together their
376 // inputs, and do the operation once, to the result of the PHI.
377 Instruction *FoldPHIArgOpIntoPHI(PHINode &PN);
378 Instruction *FoldPHIArgBinOpIntoPHI(PHINode &PN);
379 Instruction *FoldPHIArgGEPIntoPHI(PHINode &PN);
380 Instruction *FoldPHIArgLoadIntoPHI(PHINode &PN);
382 Instruction *OptAndOp(Instruction *Op, ConstantInt *OpRHS,
383 ConstantInt *AndRHS, BinaryOperator &TheAnd);
385 Value *FoldLogicalPlusAnd(Value *LHS, Value *RHS, ConstantInt *Mask,
386 bool isSub, Instruction &I);
387 Value *InsertRangeTest(Value *V, Constant *Lo, Constant *Hi, bool isSigned,
389 Instruction *PromoteCastOfAllocation(BitCastInst &CI, AllocaInst &AI);
390 Instruction *MatchBSwap(BinaryOperator &I);
391 bool SimplifyStoreAtEndOfBlock(StoreInst &SI);
392 Instruction *SimplifyMemTransfer(MemIntrinsic *MI);
393 Instruction *SimplifyMemSet(MemSetInst *MI);
395 Value *EvaluateInDifferentType(Value *V, Type *Ty, bool isSigned);
397 /// Descale - Return a value X such that Val = X * Scale, or null if none. If
398 /// the multiplication is known not to overflow then NoSignedWrap is set.
399 Value *Descale(Value *Val, APInt Scale, bool &NoSignedWrap);
402 } // end namespace llvm.