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 INSTCOMBINE_INSTCOMBINE_H
11 #define 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 {
42 /// getComplexity: Assign a complexity or rank value to LLVM Values...
43 /// 0 -> undef, 1 -> Const, 2 -> Other, 3 -> Arg, 3 -> Unary, 4 -> OtherInst
44 static inline unsigned getComplexity(Value *V) {
45 if (isa<Instruction>(V)) {
46 if (BinaryOperator::isNeg(V) ||
47 BinaryOperator::isFNeg(V) ||
48 BinaryOperator::isNot(V))
52 if (isa<Argument>(V)) return 3;
53 return isa<Constant>(V) ? (isa<UndefValue>(V) ? 0 : 1) : 2;
56 /// AddOne - Add one to a Constant
57 static inline Constant *AddOne(Constant *C) {
58 return ConstantExpr::getAdd(C, ConstantInt::get(C->getType(), 1));
60 /// SubOne - Subtract one from a Constant
61 static inline Constant *SubOne(Constant *C) {
62 return ConstantExpr::getSub(C, ConstantInt::get(C->getType(), 1));
66 /// InstCombineIRInserter - This is an IRBuilder insertion helper that works
67 /// just like the normal insertion helper, but also adds any new instructions
68 /// to the instcombine worklist.
69 class LLVM_LIBRARY_VISIBILITY InstCombineIRInserter
70 : public IRBuilderDefaultInserter<true> {
71 InstCombineWorklist &Worklist;
73 InstCombineIRInserter(InstCombineWorklist &WL) : Worklist(WL) {}
75 void InsertHelper(Instruction *I, const Twine &Name,
76 BasicBlock *BB, BasicBlock::iterator InsertPt) const {
77 IRBuilderDefaultInserter<true>::InsertHelper(I, Name, BB, InsertPt);
82 /// InstCombiner - The -instcombine pass.
83 class LLVM_LIBRARY_VISIBILITY InstCombiner
84 : public FunctionPass,
85 public InstVisitor<InstCombiner, Instruction*> {
87 TargetLibraryInfo *TLI;
89 LibCallSimplifier *Simplifier;
92 /// Worklist - All of the instructions that need to be simplified.
93 InstCombineWorklist Worklist;
95 /// Builder - This is an IRBuilder that automatically inserts new
96 /// instructions into the worklist when they are created.
97 typedef IRBuilder<true, TargetFolder, InstCombineIRInserter> BuilderTy;
100 static char ID; // Pass identification, replacement for typeid
101 InstCombiner() : FunctionPass(ID), DL(0), Builder(0) {
102 MinimizeSize = false;
103 initializeInstCombinerPass(*PassRegistry::getPassRegistry());
107 bool runOnFunction(Function &F) override;
109 bool DoOneIteration(Function &F, unsigned ItNum);
111 void getAnalysisUsage(AnalysisUsage &AU) const override;
113 const DataLayout *getDataLayout() const { return DL; }
115 TargetLibraryInfo *getTargetLibraryInfo() const { return TLI; }
117 // Visitation implementation - Implement instruction combining for different
118 // instruction types. The semantics are as follows:
120 // null - No change was made
121 // I - Change was made, I is still valid, I may be dead though
122 // otherwise - Change was made, replace I with returned instruction
124 Instruction *visitAdd(BinaryOperator &I);
125 Instruction *visitFAdd(BinaryOperator &I);
126 Value *OptimizePointerDifference(Value *LHS, Value *RHS, Type *Ty);
127 Instruction *visitSub(BinaryOperator &I);
128 Instruction *visitFSub(BinaryOperator &I);
129 Instruction *visitMul(BinaryOperator &I);
130 Value *foldFMulConst(Instruction *FMulOrDiv, Constant *C,
131 Instruction *InsertBefore);
132 Instruction *visitFMul(BinaryOperator &I);
133 Instruction *visitURem(BinaryOperator &I);
134 Instruction *visitSRem(BinaryOperator &I);
135 Instruction *visitFRem(BinaryOperator &I);
136 bool SimplifyDivRemOfSelect(BinaryOperator &I);
137 Instruction *commonRemTransforms(BinaryOperator &I);
138 Instruction *commonIRemTransforms(BinaryOperator &I);
139 Instruction *commonDivTransforms(BinaryOperator &I);
140 Instruction *commonIDivTransforms(BinaryOperator &I);
141 Instruction *visitUDiv(BinaryOperator &I);
142 Instruction *visitSDiv(BinaryOperator &I);
143 Instruction *visitFDiv(BinaryOperator &I);
144 Value *FoldAndOfICmps(ICmpInst *LHS, ICmpInst *RHS);
145 Value *FoldAndOfFCmps(FCmpInst *LHS, FCmpInst *RHS);
146 Instruction *visitAnd(BinaryOperator &I);
147 Value *FoldOrOfICmps(ICmpInst *LHS, ICmpInst *RHS);
148 Value *FoldOrOfFCmps(FCmpInst *LHS, FCmpInst *RHS);
149 Instruction *FoldOrWithConstants(BinaryOperator &I, Value *Op,
150 Value *A, Value *B, Value *C);
151 Instruction *visitOr (BinaryOperator &I);
152 Instruction *visitXor(BinaryOperator &I);
153 Instruction *visitShl(BinaryOperator &I);
154 Instruction *visitAShr(BinaryOperator &I);
155 Instruction *visitLShr(BinaryOperator &I);
156 Instruction *commonShiftTransforms(BinaryOperator &I);
157 Instruction *FoldFCmp_IntToFP_Cst(FCmpInst &I, Instruction *LHSI,
159 Instruction *FoldCmpLoadFromIndexedGlobal(GetElementPtrInst *GEP,
160 GlobalVariable *GV, CmpInst &ICI,
161 ConstantInt *AndCst = 0);
162 Instruction *visitFCmpInst(FCmpInst &I);
163 Instruction *visitICmpInst(ICmpInst &I);
164 Instruction *visitICmpInstWithCastAndCast(ICmpInst &ICI);
165 Instruction *visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
168 Instruction *FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
169 ConstantInt *DivRHS);
170 Instruction *FoldICmpShrCst(ICmpInst &ICI, BinaryOperator *DivI,
171 ConstantInt *DivRHS);
172 Instruction *FoldICmpAddOpCst(Instruction &ICI, Value *X, ConstantInt *CI,
173 ICmpInst::Predicate Pred);
174 Instruction *FoldGEPICmp(GEPOperator *GEPLHS, Value *RHS,
175 ICmpInst::Predicate Cond, Instruction &I);
176 Instruction *FoldShiftByConstant(Value *Op0, Constant *Op1,
178 Instruction *commonCastTransforms(CastInst &CI);
179 Instruction *commonPointerCastTransforms(CastInst &CI);
180 Instruction *visitTrunc(TruncInst &CI);
181 Instruction *visitZExt(ZExtInst &CI);
182 Instruction *visitSExt(SExtInst &CI);
183 Instruction *visitFPTrunc(FPTruncInst &CI);
184 Instruction *visitFPExt(CastInst &CI);
185 Instruction *visitFPToUI(FPToUIInst &FI);
186 Instruction *visitFPToSI(FPToSIInst &FI);
187 Instruction *visitUIToFP(CastInst &CI);
188 Instruction *visitSIToFP(CastInst &CI);
189 Instruction *visitPtrToInt(PtrToIntInst &CI);
190 Instruction *visitIntToPtr(IntToPtrInst &CI);
191 Instruction *visitBitCast(BitCastInst &CI);
192 Instruction *visitAddrSpaceCast(AddrSpaceCastInst &CI);
193 Instruction *FoldSelectOpOp(SelectInst &SI, Instruction *TI,
195 Instruction *FoldSelectIntoOp(SelectInst &SI, Value*, Value*);
196 Instruction *FoldSPFofSPF(Instruction *Inner, SelectPatternFlavor SPF1,
197 Value *A, Value *B, Instruction &Outer,
198 SelectPatternFlavor SPF2, Value *C);
199 Instruction *visitSelectInst(SelectInst &SI);
200 Instruction *visitSelectInstWithICmp(SelectInst &SI, ICmpInst *ICI);
201 Instruction *visitCallInst(CallInst &CI);
202 Instruction *visitInvokeInst(InvokeInst &II);
204 Instruction *SliceUpIllegalIntegerPHI(PHINode &PN);
205 Instruction *visitPHINode(PHINode &PN);
206 Instruction *visitGetElementPtrInst(GetElementPtrInst &GEP);
207 Instruction *visitAllocaInst(AllocaInst &AI);
208 Instruction *visitAllocSite(Instruction &FI);
209 Instruction *visitFree(CallInst &FI);
210 Instruction *visitLoadInst(LoadInst &LI);
211 Instruction *visitStoreInst(StoreInst &SI);
212 Instruction *visitBranchInst(BranchInst &BI);
213 Instruction *visitSwitchInst(SwitchInst &SI);
214 Instruction *visitInsertElementInst(InsertElementInst &IE);
215 Instruction *visitExtractElementInst(ExtractElementInst &EI);
216 Instruction *visitShuffleVectorInst(ShuffleVectorInst &SVI);
217 Instruction *visitExtractValueInst(ExtractValueInst &EV);
218 Instruction *visitLandingPadInst(LandingPadInst &LI);
220 // visitInstruction - Specify what to return for unhandled instructions...
221 Instruction *visitInstruction(Instruction &I) { return 0; }
224 bool ShouldChangeType(Type *From, Type *To) const;
225 Value *dyn_castNegVal(Value *V) const;
226 Value *dyn_castFNegVal(Value *V, bool NoSignedZero=false) const;
227 Type *FindElementAtOffset(Type *PtrTy, int64_t Offset,
228 SmallVectorImpl<Value*> &NewIndices);
229 Instruction *FoldOpIntoSelect(Instruction &Op, SelectInst *SI);
231 /// ShouldOptimizeCast - Return true if the cast from "V to Ty" actually
232 /// results in any code being generated and is interesting to optimize out. If
233 /// the cast can be eliminated by some other simple transformation, we prefer
234 /// to do the simplification first.
235 bool ShouldOptimizeCast(Instruction::CastOps opcode,const Value *V,
238 Instruction *visitCallSite(CallSite CS);
239 Instruction *tryOptimizeCall(CallInst *CI, const DataLayout *DL);
240 bool transformConstExprCastCall(CallSite CS);
241 Instruction *transformCallThroughTrampoline(CallSite CS,
242 IntrinsicInst *Tramp);
243 Instruction *transformZExtICmp(ICmpInst *ICI, Instruction &CI,
244 bool DoXform = true);
245 Instruction *transformSExtICmp(ICmpInst *ICI, Instruction &CI);
246 bool WillNotOverflowSignedAdd(Value *LHS, Value *RHS);
247 Value *EmitGEPOffset(User *GEP);
248 Instruction *scalarizePHI(ExtractElementInst &EI, PHINode *PN);
249 Value *EvaluateInDifferentElementOrder(Value *V, ArrayRef<int> Mask);
252 // InsertNewInstBefore - insert an instruction New before instruction Old
253 // in the program. Add the new instruction to the worklist.
255 Instruction *InsertNewInstBefore(Instruction *New, Instruction &Old) {
256 assert(New && New->getParent() == 0 &&
257 "New instruction already inserted into a basic block!");
258 BasicBlock *BB = Old.getParent();
259 BB->getInstList().insert(&Old, New); // Insert inst
264 // InsertNewInstWith - same as InsertNewInstBefore, but also sets the
267 Instruction *InsertNewInstWith(Instruction *New, Instruction &Old) {
268 New->setDebugLoc(Old.getDebugLoc());
269 return InsertNewInstBefore(New, Old);
272 // ReplaceInstUsesWith - This method is to be used when an instruction is
273 // found to be dead, replacable with another preexisting expression. Here
274 // we add all uses of I to the worklist, replace all uses of I with the new
275 // value, then return I, so that the inst combiner will know that I was
278 Instruction *ReplaceInstUsesWith(Instruction &I, Value *V) {
279 Worklist.AddUsersToWorkList(I); // Add all modified instrs to worklist.
281 // If we are replacing the instruction with itself, this must be in a
282 // segment of unreachable code, so just clobber the instruction.
284 V = UndefValue::get(I.getType());
286 DEBUG(dbgs() << "IC: Replacing " << I << "\n"
287 " with " << *V << '\n');
289 I.replaceAllUsesWith(V);
293 // EraseInstFromFunction - When dealing with an instruction that has side
294 // effects or produces a void value, we can't rely on DCE to delete the
295 // instruction. Instead, visit methods should return the value returned by
297 Instruction *EraseInstFromFunction(Instruction &I) {
298 DEBUG(dbgs() << "IC: ERASE " << I << '\n');
300 assert(I.use_empty() && "Cannot erase instruction that is used!");
301 // Make sure that we reprocess all operands now that we reduced their
303 if (I.getNumOperands() < 8) {
304 for (User::op_iterator i = I.op_begin(), e = I.op_end(); i != e; ++i)
305 if (Instruction *Op = dyn_cast<Instruction>(*i))
311 return 0; // Don't do anything with FI
314 void ComputeMaskedBits(Value *V, APInt &KnownZero,
315 APInt &KnownOne, unsigned Depth = 0) const {
316 return llvm::ComputeMaskedBits(V, KnownZero, KnownOne, DL, Depth);
319 bool MaskedValueIsZero(Value *V, const APInt &Mask,
320 unsigned Depth = 0) const {
321 return llvm::MaskedValueIsZero(V, Mask, DL, Depth);
323 unsigned ComputeNumSignBits(Value *Op, unsigned Depth = 0) const {
324 return llvm::ComputeNumSignBits(Op, DL, Depth);
329 /// SimplifyAssociativeOrCommutative - This performs a few simplifications for
330 /// operators which are associative or commutative.
331 bool SimplifyAssociativeOrCommutative(BinaryOperator &I);
333 /// SimplifyUsingDistributiveLaws - This tries to simplify binary operations
334 /// which some other binary operation distributes over either by factorizing
335 /// out common terms (eg "(A*B)+(A*C)" -> "A*(B+C)") or expanding out if this
336 /// results in simplifications (eg: "A & (B | C) -> (A&B) | (A&C)" if this is
337 /// a win). Returns the simplified value, or null if it didn't simplify.
338 Value *SimplifyUsingDistributiveLaws(BinaryOperator &I);
340 /// SimplifyDemandedUseBits - Attempts to replace V with a simpler value
341 /// based on the demanded bits.
342 Value *SimplifyDemandedUseBits(Value *V, APInt DemandedMask,
343 APInt& KnownZero, APInt& KnownOne,
345 bool SimplifyDemandedBits(Use &U, APInt DemandedMask,
346 APInt& KnownZero, APInt& KnownOne,
348 /// Helper routine of SimplifyDemandedUseBits. It tries to simplify demanded
349 /// bit for "r1 = shr x, c1; r2 = shl r1, c2" instruction sequence.
350 Value *SimplifyShrShlDemandedBits(Instruction *Lsr, Instruction *Sftl,
351 APInt DemandedMask, APInt &KnownZero,
354 /// SimplifyDemandedInstructionBits - Inst is an integer instruction that
355 /// SimplifyDemandedBits knows about. See if the instruction has any
356 /// properties that allow us to simplify its operands.
357 bool SimplifyDemandedInstructionBits(Instruction &Inst);
359 Value *SimplifyDemandedVectorElts(Value *V, APInt DemandedElts,
360 APInt& UndefElts, unsigned Depth = 0);
362 // FoldOpIntoPhi - Given a binary operator, cast instruction, or select
363 // which has a PHI node as operand #0, see if we can fold the instruction
364 // into the PHI (which is only possible if all operands to the PHI are
367 Instruction *FoldOpIntoPhi(Instruction &I);
369 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
370 // operator and they all are only used by the PHI, PHI together their
371 // inputs, and do the operation once, to the result of the PHI.
372 Instruction *FoldPHIArgOpIntoPHI(PHINode &PN);
373 Instruction *FoldPHIArgBinOpIntoPHI(PHINode &PN);
374 Instruction *FoldPHIArgGEPIntoPHI(PHINode &PN);
375 Instruction *FoldPHIArgLoadIntoPHI(PHINode &PN);
378 Instruction *OptAndOp(Instruction *Op, ConstantInt *OpRHS,
379 ConstantInt *AndRHS, BinaryOperator &TheAnd);
381 Value *FoldLogicalPlusAnd(Value *LHS, Value *RHS, ConstantInt *Mask,
382 bool isSub, Instruction &I);
383 Value *InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
384 bool isSigned, bool Inside);
385 Instruction *PromoteCastOfAllocation(BitCastInst &CI, AllocaInst &AI);
386 Instruction *MatchBSwap(BinaryOperator &I);
387 bool SimplifyStoreAtEndOfBlock(StoreInst &SI);
388 Instruction *SimplifyMemTransfer(MemIntrinsic *MI);
389 Instruction *SimplifyMemSet(MemSetInst *MI);
392 Value *EvaluateInDifferentType(Value *V, Type *Ty, bool isSigned);
394 /// Descale - Return a value X such that Val = X * Scale, or null if none. If
395 /// the multiplication is known not to overflow then NoSignedWrap is set.
396 Value *Descale(Value *Val, APInt Scale, bool &NoSignedWrap);
399 } // end namespace llvm.