1 //===- InstCombine.h - Main InstCombine pass definition -------------------===//
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/IRBuilder.h"
15 #include "llvm/InstVisitor.h"
16 #include "llvm/IntrinsicInst.h"
17 #include "llvm/Operator.h"
18 #include "llvm/Pass.h"
19 #include "llvm/Analysis/ValueTracking.h"
20 #include "llvm/Support/TargetFolder.h"
21 #include "llvm/Transforms/Utils/SimplifyLibCalls.h"
26 class TargetLibraryInfo;
31 /// SelectPatternFlavor - We can match a variety of different patterns for
32 /// select operations.
33 enum SelectPatternFlavor {
40 /// getComplexity: Assign a complexity or rank value to LLVM Values...
41 /// 0 -> undef, 1 -> Const, 2 -> Other, 3 -> Arg, 3 -> Unary, 4 -> OtherInst
42 static inline unsigned getComplexity(Value *V) {
43 if (isa<Instruction>(V)) {
44 if (BinaryOperator::isNeg(V) ||
45 BinaryOperator::isFNeg(V) ||
46 BinaryOperator::isNot(V))
50 if (isa<Argument>(V)) return 3;
51 return isa<Constant>(V) ? (isa<UndefValue>(V) ? 0 : 1) : 2;
55 /// InstCombineIRInserter - This is an IRBuilder insertion helper that works
56 /// just like the normal insertion helper, but also adds any new instructions
57 /// to the instcombine worklist.
58 class LLVM_LIBRARY_VISIBILITY InstCombineIRInserter
59 : public IRBuilderDefaultInserter<true> {
60 InstCombineWorklist &Worklist;
62 InstCombineIRInserter(InstCombineWorklist &WL) : Worklist(WL) {}
64 void InsertHelper(Instruction *I, const Twine &Name,
65 BasicBlock *BB, BasicBlock::iterator InsertPt) const {
66 IRBuilderDefaultInserter<true>::InsertHelper(I, Name, BB, InsertPt);
71 /// InstCombiner - The -instcombine pass.
72 class LLVM_LIBRARY_VISIBILITY InstCombiner
73 : public FunctionPass,
74 public InstVisitor<InstCombiner, Instruction*> {
76 TargetLibraryInfo *TLI;
78 LibCallSimplifier *Simplifier;
80 /// Worklist - All of the instructions that need to be simplified.
81 InstCombineWorklist Worklist;
83 /// Builder - This is an IRBuilder that automatically inserts new
84 /// instructions into the worklist when they are created.
85 typedef IRBuilder<true, TargetFolder, InstCombineIRInserter> BuilderTy;
88 static char ID; // Pass identification, replacement for typeid
89 InstCombiner() : FunctionPass(ID), TD(0), Builder(0) {
90 initializeInstCombinerPass(*PassRegistry::getPassRegistry());
94 virtual bool runOnFunction(Function &F);
96 bool DoOneIteration(Function &F, unsigned ItNum);
98 virtual void getAnalysisUsage(AnalysisUsage &AU) const;
100 DataLayout *getDataLayout() const { return TD; }
102 TargetLibraryInfo *getTargetLibraryInfo() const { return TLI; }
104 // Visitation implementation - Implement instruction combining for different
105 // instruction types. The semantics are as follows:
107 // null - No change was made
108 // I - Change was made, I is still valid, I may be dead though
109 // otherwise - Change was made, replace I with returned instruction
111 Instruction *visitAdd(BinaryOperator &I);
112 Instruction *visitFAdd(BinaryOperator &I);
113 Value *OptimizePointerDifference(Value *LHS, Value *RHS, Type *Ty);
114 Instruction *visitSub(BinaryOperator &I);
115 Instruction *visitFSub(BinaryOperator &I);
116 Instruction *visitMul(BinaryOperator &I);
117 Instruction *visitFMul(BinaryOperator &I);
118 Instruction *visitURem(BinaryOperator &I);
119 Instruction *visitSRem(BinaryOperator &I);
120 Instruction *visitFRem(BinaryOperator &I);
121 bool SimplifyDivRemOfSelect(BinaryOperator &I);
122 Instruction *commonRemTransforms(BinaryOperator &I);
123 Instruction *commonIRemTransforms(BinaryOperator &I);
124 Instruction *commonDivTransforms(BinaryOperator &I);
125 Instruction *commonIDivTransforms(BinaryOperator &I);
126 Instruction *visitUDiv(BinaryOperator &I);
127 Instruction *visitSDiv(BinaryOperator &I);
128 Instruction *visitFDiv(BinaryOperator &I);
129 Value *FoldAndOfICmps(ICmpInst *LHS, ICmpInst *RHS);
130 Value *FoldAndOfFCmps(FCmpInst *LHS, FCmpInst *RHS);
131 Instruction *visitAnd(BinaryOperator &I);
132 Value *FoldOrOfICmps(ICmpInst *LHS, ICmpInst *RHS);
133 Value *FoldOrOfFCmps(FCmpInst *LHS, FCmpInst *RHS);
134 Instruction *FoldOrWithConstants(BinaryOperator &I, Value *Op,
135 Value *A, Value *B, Value *C);
136 Instruction *visitOr (BinaryOperator &I);
137 Instruction *visitXor(BinaryOperator &I);
138 Instruction *visitShl(BinaryOperator &I);
139 Instruction *visitAShr(BinaryOperator &I);
140 Instruction *visitLShr(BinaryOperator &I);
141 Instruction *commonShiftTransforms(BinaryOperator &I);
142 Instruction *FoldFCmp_IntToFP_Cst(FCmpInst &I, Instruction *LHSI,
144 Instruction *FoldCmpLoadFromIndexedGlobal(GetElementPtrInst *GEP,
145 GlobalVariable *GV, CmpInst &ICI,
146 ConstantInt *AndCst = 0);
147 Instruction *visitFCmpInst(FCmpInst &I);
148 Instruction *visitICmpInst(ICmpInst &I);
149 Instruction *visitICmpInstWithCastAndCast(ICmpInst &ICI);
150 Instruction *visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
153 Instruction *FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
154 ConstantInt *DivRHS);
155 Instruction *FoldICmpShrCst(ICmpInst &ICI, BinaryOperator *DivI,
156 ConstantInt *DivRHS);
157 Instruction *FoldICmpAddOpCst(ICmpInst &ICI, Value *X, ConstantInt *CI,
158 ICmpInst::Predicate Pred, Value *TheAdd);
159 Instruction *FoldGEPICmp(GEPOperator *GEPLHS, Value *RHS,
160 ICmpInst::Predicate Cond, Instruction &I);
161 Instruction *FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
163 Instruction *commonCastTransforms(CastInst &CI);
164 Instruction *commonPointerCastTransforms(CastInst &CI);
165 Instruction *visitTrunc(TruncInst &CI);
166 Instruction *visitZExt(ZExtInst &CI);
167 Instruction *visitSExt(SExtInst &CI);
168 Instruction *visitFPTrunc(FPTruncInst &CI);
169 Instruction *visitFPExt(CastInst &CI);
170 Instruction *visitFPToUI(FPToUIInst &FI);
171 Instruction *visitFPToSI(FPToSIInst &FI);
172 Instruction *visitUIToFP(CastInst &CI);
173 Instruction *visitSIToFP(CastInst &CI);
174 Instruction *visitPtrToInt(PtrToIntInst &CI);
175 Instruction *visitIntToPtr(IntToPtrInst &CI);
176 Instruction *visitBitCast(BitCastInst &CI);
177 Instruction *FoldSelectOpOp(SelectInst &SI, Instruction *TI,
179 Instruction *FoldSelectIntoOp(SelectInst &SI, Value*, Value*);
180 Instruction *FoldSPFofSPF(Instruction *Inner, SelectPatternFlavor SPF1,
181 Value *A, Value *B, Instruction &Outer,
182 SelectPatternFlavor SPF2, Value *C);
183 Instruction *visitSelectInst(SelectInst &SI);
184 Instruction *visitSelectInstWithICmp(SelectInst &SI, ICmpInst *ICI);
185 Instruction *visitCallInst(CallInst &CI);
186 Instruction *visitInvokeInst(InvokeInst &II);
188 Instruction *SliceUpIllegalIntegerPHI(PHINode &PN);
189 Instruction *visitPHINode(PHINode &PN);
190 Instruction *visitGetElementPtrInst(GetElementPtrInst &GEP);
191 Instruction *visitAllocaInst(AllocaInst &AI);
192 Instruction *visitAllocSite(Instruction &FI);
193 Instruction *visitFree(CallInst &FI);
194 Instruction *visitLoadInst(LoadInst &LI);
195 Instruction *visitStoreInst(StoreInst &SI);
196 Instruction *visitBranchInst(BranchInst &BI);
197 Instruction *visitSwitchInst(SwitchInst &SI);
198 Instruction *visitInsertElementInst(InsertElementInst &IE);
199 Instruction *visitExtractElementInst(ExtractElementInst &EI);
200 Instruction *visitShuffleVectorInst(ShuffleVectorInst &SVI);
201 Instruction *visitExtractValueInst(ExtractValueInst &EV);
202 Instruction *visitLandingPadInst(LandingPadInst &LI);
204 // visitInstruction - Specify what to return for unhandled instructions...
205 Instruction *visitInstruction(Instruction &I) { return 0; }
208 bool ShouldChangeType(Type *From, Type *To) const;
209 Value *dyn_castNegVal(Value *V) const;
210 Value *dyn_castFNegVal(Value *V) const;
211 Type *FindElementAtOffset(Type *Ty, int64_t Offset,
212 SmallVectorImpl<Value*> &NewIndices);
213 Instruction *FoldOpIntoSelect(Instruction &Op, SelectInst *SI);
215 /// ShouldOptimizeCast - Return true if the cast from "V to Ty" actually
216 /// results in any code being generated and is interesting to optimize out. If
217 /// the cast can be eliminated by some other simple transformation, we prefer
218 /// to do the simplification first.
219 bool ShouldOptimizeCast(Instruction::CastOps opcode,const Value *V,
222 Instruction *visitCallSite(CallSite CS);
223 Instruction *tryOptimizeCall(CallInst *CI, const DataLayout *TD);
224 bool transformConstExprCastCall(CallSite CS);
225 Instruction *transformCallThroughTrampoline(CallSite CS,
226 IntrinsicInst *Tramp);
227 Instruction *transformZExtICmp(ICmpInst *ICI, Instruction &CI,
228 bool DoXform = true);
229 Instruction *transformSExtICmp(ICmpInst *ICI, Instruction &CI);
230 bool WillNotOverflowSignedAdd(Value *LHS, Value *RHS);
231 Value *EmitGEPOffset(User *GEP);
234 // InsertNewInstBefore - insert an instruction New before instruction Old
235 // in the program. Add the new instruction to the worklist.
237 Instruction *InsertNewInstBefore(Instruction *New, Instruction &Old) {
238 assert(New && New->getParent() == 0 &&
239 "New instruction already inserted into a basic block!");
240 BasicBlock *BB = Old.getParent();
241 BB->getInstList().insert(&Old, New); // Insert inst
246 // InsertNewInstWith - same as InsertNewInstBefore, but also sets the
249 Instruction *InsertNewInstWith(Instruction *New, Instruction &Old) {
250 New->setDebugLoc(Old.getDebugLoc());
251 return InsertNewInstBefore(New, Old);
254 // ReplaceInstUsesWith - This method is to be used when an instruction is
255 // found to be dead, replacable with another preexisting expression. Here
256 // we add all uses of I to the worklist, replace all uses of I with the new
257 // value, then return I, so that the inst combiner will know that I was
260 Instruction *ReplaceInstUsesWith(Instruction &I, Value *V) {
261 Worklist.AddUsersToWorkList(I); // Add all modified instrs to worklist.
263 // If we are replacing the instruction with itself, this must be in a
264 // segment of unreachable code, so just clobber the instruction.
266 V = UndefValue::get(I.getType());
268 DEBUG(errs() << "IC: Replacing " << I << "\n"
269 " with " << *V << '\n');
271 I.replaceAllUsesWith(V);
275 // EraseInstFromFunction - When dealing with an instruction that has side
276 // effects or produces a void value, we can't rely on DCE to delete the
277 // instruction. Instead, visit methods should return the value returned by
279 Instruction *EraseInstFromFunction(Instruction &I) {
280 DEBUG(errs() << "IC: ERASE " << I << '\n');
282 assert(I.use_empty() && "Cannot erase instruction that is used!");
283 // Make sure that we reprocess all operands now that we reduced their
285 if (I.getNumOperands() < 8) {
286 for (User::op_iterator i = I.op_begin(), e = I.op_end(); i != e; ++i)
287 if (Instruction *Op = dyn_cast<Instruction>(*i))
293 return 0; // Don't do anything with FI
296 void ComputeMaskedBits(Value *V, APInt &KnownZero,
297 APInt &KnownOne, unsigned Depth = 0) const {
298 return llvm::ComputeMaskedBits(V, KnownZero, KnownOne, TD, Depth);
301 bool MaskedValueIsZero(Value *V, const APInt &Mask,
302 unsigned Depth = 0) const {
303 return llvm::MaskedValueIsZero(V, Mask, TD, Depth);
305 unsigned ComputeNumSignBits(Value *Op, unsigned Depth = 0) const {
306 return llvm::ComputeNumSignBits(Op, TD, Depth);
311 /// SimplifyAssociativeOrCommutative - This performs a few simplifications for
312 /// operators which are associative or commutative.
313 bool SimplifyAssociativeOrCommutative(BinaryOperator &I);
315 /// SimplifyUsingDistributiveLaws - This tries to simplify binary operations
316 /// which some other binary operation distributes over either by factorizing
317 /// out common terms (eg "(A*B)+(A*C)" -> "A*(B+C)") or expanding out if this
318 /// results in simplifications (eg: "A & (B | C) -> (A&B) | (A&C)" if this is
319 /// a win). Returns the simplified value, or null if it didn't simplify.
320 Value *SimplifyUsingDistributiveLaws(BinaryOperator &I);
322 /// SimplifyDemandedUseBits - Attempts to replace V with a simpler value
323 /// based on the demanded bits.
324 Value *SimplifyDemandedUseBits(Value *V, APInt DemandedMask,
325 APInt& KnownZero, APInt& KnownOne,
327 bool SimplifyDemandedBits(Use &U, APInt DemandedMask,
328 APInt& KnownZero, APInt& KnownOne,
331 /// SimplifyDemandedInstructionBits - Inst is an integer instruction that
332 /// SimplifyDemandedBits knows about. See if the instruction has any
333 /// properties that allow us to simplify its operands.
334 bool SimplifyDemandedInstructionBits(Instruction &Inst);
336 Value *SimplifyDemandedVectorElts(Value *V, APInt DemandedElts,
337 APInt& UndefElts, unsigned Depth = 0);
339 // FoldOpIntoPhi - Given a binary operator, cast instruction, or select
340 // which has a PHI node as operand #0, see if we can fold the instruction
341 // into the PHI (which is only possible if all operands to the PHI are
344 Instruction *FoldOpIntoPhi(Instruction &I);
346 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
347 // operator and they all are only used by the PHI, PHI together their
348 // inputs, and do the operation once, to the result of the PHI.
349 Instruction *FoldPHIArgOpIntoPHI(PHINode &PN);
350 Instruction *FoldPHIArgBinOpIntoPHI(PHINode &PN);
351 Instruction *FoldPHIArgGEPIntoPHI(PHINode &PN);
352 Instruction *FoldPHIArgLoadIntoPHI(PHINode &PN);
355 Instruction *OptAndOp(Instruction *Op, ConstantInt *OpRHS,
356 ConstantInt *AndRHS, BinaryOperator &TheAnd);
358 Value *FoldLogicalPlusAnd(Value *LHS, Value *RHS, ConstantInt *Mask,
359 bool isSub, Instruction &I);
360 Value *InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
361 bool isSigned, bool Inside);
362 Instruction *PromoteCastOfAllocation(BitCastInst &CI, AllocaInst &AI);
363 Instruction *MatchBSwap(BinaryOperator &I);
364 bool SimplifyStoreAtEndOfBlock(StoreInst &SI);
365 Instruction *SimplifyMemTransfer(MemIntrinsic *MI);
366 Instruction *SimplifyMemSet(MemSetInst *MI);
369 Value *EvaluateInDifferentType(Value *V, Type *Ty, bool isSigned);
371 /// Descale - Return a value X such that Val = X * Scale, or null if none. If
372 /// the multiplication is known not to overflow then NoSignedWrap is set.
373 Value *Descale(Value *Val, APInt Scale, bool &NoSignedWrap);
378 } // end namespace llvm.