1 //===- InstCombine.h - Main InstCombine pass definition -------------------===// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 10 #ifndef INSTCOMBINE_INSTCOMBINE_H 11 #define INSTCOMBINE_INSTCOMBINE_H 12 13 #include "InstCombineWorklist.h" 14 #include "llvm/IntrinsicInst.h" 15 #include "llvm/Operator.h" 16 #include "llvm/Pass.h" 17 #include "llvm/Analysis/ValueTracking.h" 18 #include "llvm/Support/IRBuilder.h" 19 #include "llvm/Support/InstVisitor.h" 20 #include "llvm/Support/TargetFolder.h" 21 22 namespace llvm { 23 class CallSite; 24 class TargetData; 25 class TargetLibraryInfo; 26 class DbgDeclareInst; 27 class MemIntrinsic; 28 class MemSetInst; 29 30 /// SelectPatternFlavor - We can match a variety of different patterns for 31 /// select operations. 32 enum SelectPatternFlavor { 33 SPF_UNKNOWN = 0, 34 SPF_SMIN, SPF_UMIN, 35 SPF_SMAX, SPF_UMAX 36 //SPF_ABS - TODO. 37 }; 38 39 /// getComplexity: Assign a complexity or rank value to LLVM Values... 40 /// 0 -> undef, 1 -> Const, 2 -> Other, 3 -> Arg, 3 -> Unary, 4 -> OtherInst 41 static inline unsigned getComplexity(Value *V) { 42 if (isa<Instruction>(V)) { 43 if (BinaryOperator::isNeg(V) || 44 BinaryOperator::isFNeg(V) || 45 BinaryOperator::isNot(V)) 46 return 3; 47 return 4; 48 } 49 if (isa<Argument>(V)) return 3; 50 return isa<Constant>(V) ? (isa<UndefValue>(V) ? 0 : 1) : 2; 51 } 52 53 54 /// InstCombineIRInserter - This is an IRBuilder insertion helper that works 55 /// just like the normal insertion helper, but also adds any new instructions 56 /// to the instcombine worklist. 57 class LLVM_LIBRARY_VISIBILITY InstCombineIRInserter 58 : public IRBuilderDefaultInserter<true> { 59 InstCombineWorklist &Worklist; 60 public: 61 InstCombineIRInserter(InstCombineWorklist &WL) : Worklist(WL) {} 62 63 void InsertHelper(Instruction *I, const Twine &Name, 64 BasicBlock *BB, BasicBlock::iterator InsertPt) const { 65 IRBuilderDefaultInserter<true>::InsertHelper(I, Name, BB, InsertPt); 66 Worklist.Add(I); 67 } 68 }; 69 70 /// InstCombiner - The -instcombine pass. 71 class LLVM_LIBRARY_VISIBILITY InstCombiner 72 : public FunctionPass, 73 public InstVisitor<InstCombiner, Instruction*> { 74 TargetData *TD; 75 TargetLibraryInfo *TLI; 76 bool MadeIRChange; 77 public: 78 /// Worklist - All of the instructions that need to be simplified. 79 InstCombineWorklist Worklist; 80 81 /// Builder - This is an IRBuilder that automatically inserts new 82 /// instructions into the worklist when they are created. 83 typedef IRBuilder<true, TargetFolder, InstCombineIRInserter> BuilderTy; 84 BuilderTy *Builder; 85 86 static char ID; // Pass identification, replacement for typeid 87 InstCombiner() : FunctionPass(ID), TD(0), Builder(0) { 88 initializeInstCombinerPass(*PassRegistry::getPassRegistry()); 89 } 90 91 public: 92 virtual bool runOnFunction(Function &F); 93 94 bool DoOneIteration(Function &F, unsigned ItNum); 95 96 virtual void getAnalysisUsage(AnalysisUsage &AU) const; 97 98 TargetData *getTargetData() const { return TD; } 99 100 TargetLibraryInfo *getTargetLibraryInfo() const { return TLI; } 101 102 // Visitation implementation - Implement instruction combining for different 103 // instruction types. The semantics are as follows: 104 // Return Value: 105 // null - No change was made 106 // I - Change was made, I is still valid, I may be dead though 107 // otherwise - Change was made, replace I with returned instruction 108 // 109 Instruction *visitAdd(BinaryOperator &I); 110 Instruction *visitFAdd(BinaryOperator &I); 111 Value *OptimizePointerDifference(Value *LHS, Value *RHS, Type *Ty); 112 Instruction *visitSub(BinaryOperator &I); 113 Instruction *visitFSub(BinaryOperator &I); 114 Instruction *visitMul(BinaryOperator &I); 115 Instruction *visitFMul(BinaryOperator &I); 116 Instruction *visitURem(BinaryOperator &I); 117 Instruction *visitSRem(BinaryOperator &I); 118 Instruction *visitFRem(BinaryOperator &I); 119 bool SimplifyDivRemOfSelect(BinaryOperator &I); 120 Instruction *commonRemTransforms(BinaryOperator &I); 121 Instruction *commonIRemTransforms(BinaryOperator &I); 122 Instruction *commonDivTransforms(BinaryOperator &I); 123 Instruction *commonIDivTransforms(BinaryOperator &I); 124 Instruction *visitUDiv(BinaryOperator &I); 125 Instruction *visitSDiv(BinaryOperator &I); 126 Instruction *visitFDiv(BinaryOperator &I); 127 Value *FoldAndOfICmps(ICmpInst *LHS, ICmpInst *RHS); 128 Value *FoldAndOfFCmps(FCmpInst *LHS, FCmpInst *RHS); 129 Instruction *visitAnd(BinaryOperator &I); 130 Value *FoldOrOfICmps(ICmpInst *LHS, ICmpInst *RHS); 131 Value *FoldOrOfFCmps(FCmpInst *LHS, FCmpInst *RHS); 132 Instruction *FoldOrWithConstants(BinaryOperator &I, Value *Op, 133 Value *A, Value *B, Value *C); 134 Instruction *visitOr (BinaryOperator &I); 135 Instruction *visitXor(BinaryOperator &I); 136 Instruction *visitShl(BinaryOperator &I); 137 Instruction *visitAShr(BinaryOperator &I); 138 Instruction *visitLShr(BinaryOperator &I); 139 Instruction *commonShiftTransforms(BinaryOperator &I); 140 Instruction *FoldFCmp_IntToFP_Cst(FCmpInst &I, Instruction *LHSI, 141 Constant *RHSC); 142 Instruction *FoldCmpLoadFromIndexedGlobal(GetElementPtrInst *GEP, 143 GlobalVariable *GV, CmpInst &ICI, 144 ConstantInt *AndCst = 0); 145 Instruction *visitFCmpInst(FCmpInst &I); 146 Instruction *visitICmpInst(ICmpInst &I); 147 Instruction *visitICmpInstWithCastAndCast(ICmpInst &ICI); 148 Instruction *visitICmpInstWithInstAndIntCst(ICmpInst &ICI, 149 Instruction *LHS, 150 ConstantInt *RHS); 151 Instruction *FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI, 152 ConstantInt *DivRHS); 153 Instruction *FoldICmpShrCst(ICmpInst &ICI, BinaryOperator *DivI, 154 ConstantInt *DivRHS); 155 Instruction *FoldICmpAddOpCst(ICmpInst &ICI, Value *X, ConstantInt *CI, 156 ICmpInst::Predicate Pred, Value *TheAdd); 157 Instruction *FoldGEPICmp(GEPOperator *GEPLHS, Value *RHS, 158 ICmpInst::Predicate Cond, Instruction &I); 159 Instruction *FoldShiftByConstant(Value *Op0, ConstantInt *Op1, 160 BinaryOperator &I); 161 Instruction *commonCastTransforms(CastInst &CI); 162 Instruction *commonPointerCastTransforms(CastInst &CI); 163 Instruction *visitTrunc(TruncInst &CI); 164 Instruction *visitZExt(ZExtInst &CI); 165 Instruction *visitSExt(SExtInst &CI); 166 Instruction *visitFPTrunc(FPTruncInst &CI); 167 Instruction *visitFPExt(CastInst &CI); 168 Instruction *visitFPToUI(FPToUIInst &FI); 169 Instruction *visitFPToSI(FPToSIInst &FI); 170 Instruction *visitUIToFP(CastInst &CI); 171 Instruction *visitSIToFP(CastInst &CI); 172 Instruction *visitPtrToInt(PtrToIntInst &CI); 173 Instruction *visitIntToPtr(IntToPtrInst &CI); 174 Instruction *visitBitCast(BitCastInst &CI); 175 Instruction *FoldSelectOpOp(SelectInst &SI, Instruction *TI, 176 Instruction *FI); 177 Instruction *FoldSelectIntoOp(SelectInst &SI, Value*, Value*); 178 Instruction *FoldSPFofSPF(Instruction *Inner, SelectPatternFlavor SPF1, 179 Value *A, Value *B, Instruction &Outer, 180 SelectPatternFlavor SPF2, Value *C); 181 Instruction *visitSelectInst(SelectInst &SI); 182 Instruction *visitSelectInstWithICmp(SelectInst &SI, ICmpInst *ICI); 183 Instruction *visitCallInst(CallInst &CI); 184 Instruction *visitInvokeInst(InvokeInst &II); 185 186 Instruction *SliceUpIllegalIntegerPHI(PHINode &PN); 187 Instruction *visitPHINode(PHINode &PN); 188 Instruction *visitGetElementPtrInst(GetElementPtrInst &GEP); 189 Instruction *visitAllocaInst(AllocaInst &AI); 190 Instruction *visitMalloc(Instruction &FI); 191 Instruction *visitFree(CallInst &FI); 192 Instruction *visitLoadInst(LoadInst &LI); 193 Instruction *visitStoreInst(StoreInst &SI); 194 Instruction *visitBranchInst(BranchInst &BI); 195 Instruction *visitSwitchInst(SwitchInst &SI); 196 Instruction *visitInsertElementInst(InsertElementInst &IE); 197 Instruction *visitExtractElementInst(ExtractElementInst &EI); 198 Instruction *visitShuffleVectorInst(ShuffleVectorInst &SVI); 199 Instruction *visitExtractValueInst(ExtractValueInst &EV); 200 Instruction *visitLandingPadInst(LandingPadInst &LI); 201 202 // visitInstruction - Specify what to return for unhandled instructions... 203 Instruction *visitInstruction(Instruction &I) { return 0; } 204 205 private: 206 bool ShouldChangeType(Type *From, Type *To) const; 207 Value *dyn_castNegVal(Value *V) const; 208 Value *dyn_castFNegVal(Value *V) const; 209 Type *FindElementAtOffset(Type *Ty, int64_t Offset, 210 SmallVectorImpl<Value*> &NewIndices); 211 Instruction *FoldOpIntoSelect(Instruction &Op, SelectInst *SI); 212 213 /// ShouldOptimizeCast - Return true if the cast from "V to Ty" actually 214 /// results in any code being generated and is interesting to optimize out. If 215 /// the cast can be eliminated by some other simple transformation, we prefer 216 /// to do the simplification first. 217 bool ShouldOptimizeCast(Instruction::CastOps opcode,const Value *V, 218 Type *Ty); 219 220 Instruction *visitCallSite(CallSite CS); 221 Instruction *tryOptimizeCall(CallInst *CI, const TargetData *TD); 222 bool transformConstExprCastCall(CallSite CS); 223 Instruction *transformCallThroughTrampoline(CallSite CS, 224 IntrinsicInst *Tramp); 225 Instruction *transformZExtICmp(ICmpInst *ICI, Instruction &CI, 226 bool DoXform = true); 227 Instruction *transformSExtICmp(ICmpInst *ICI, Instruction &CI); 228 bool WillNotOverflowSignedAdd(Value *LHS, Value *RHS); 229 Value *EmitGEPOffset(User *GEP); 230 231 public: 232 // InsertNewInstBefore - insert an instruction New before instruction Old 233 // in the program. Add the new instruction to the worklist. 234 // 235 Instruction *InsertNewInstBefore(Instruction *New, Instruction &Old) { 236 assert(New && New->getParent() == 0 && 237 "New instruction already inserted into a basic block!"); 238 BasicBlock *BB = Old.getParent(); 239 BB->getInstList().insert(&Old, New); // Insert inst 240 Worklist.Add(New); 241 return New; 242 } 243 244 // InsertNewInstWith - same as InsertNewInstBefore, but also sets the 245 // debug loc. 246 // 247 Instruction *InsertNewInstWith(Instruction *New, Instruction &Old) { 248 New->setDebugLoc(Old.getDebugLoc()); 249 return InsertNewInstBefore(New, Old); 250 } 251 252 // ReplaceInstUsesWith - This method is to be used when an instruction is 253 // found to be dead, replacable with another preexisting expression. Here 254 // we add all uses of I to the worklist, replace all uses of I with the new 255 // value, then return I, so that the inst combiner will know that I was 256 // modified. 257 // 258 Instruction *ReplaceInstUsesWith(Instruction &I, Value *V) { 259 Worklist.AddUsersToWorkList(I); // Add all modified instrs to worklist. 260 261 // If we are replacing the instruction with itself, this must be in a 262 // segment of unreachable code, so just clobber the instruction. 263 if (&I == V) 264 V = UndefValue::get(I.getType()); 265 266 DEBUG(errs() << "IC: Replacing " << I << "\n" 267 " with " << *V << '\n'); 268 269 I.replaceAllUsesWith(V); 270 return &I; 271 } 272 273 // EraseInstFromFunction - When dealing with an instruction that has side 274 // effects or produces a void value, we can't rely on DCE to delete the 275 // instruction. Instead, visit methods should return the value returned by 276 // this function. 277 Instruction *EraseInstFromFunction(Instruction &I) { 278 DEBUG(errs() << "IC: ERASE " << I << '\n'); 279 280 assert(I.use_empty() && "Cannot erase instruction that is used!"); 281 // Make sure that we reprocess all operands now that we reduced their 282 // use counts. 283 if (I.getNumOperands() < 8) { 284 for (User::op_iterator i = I.op_begin(), e = I.op_end(); i != e; ++i) 285 if (Instruction *Op = dyn_cast<Instruction>(*i)) 286 Worklist.Add(Op); 287 } 288 Worklist.Remove(&I); 289 I.eraseFromParent(); 290 MadeIRChange = true; 291 return 0; // Don't do anything with FI 292 } 293 294 void ComputeMaskedBits(Value *V, APInt &KnownZero, 295 APInt &KnownOne, unsigned Depth = 0) const { 296 return llvm::ComputeMaskedBits(V, KnownZero, KnownOne, TD, Depth); 297 } 298 299 bool MaskedValueIsZero(Value *V, const APInt &Mask, 300 unsigned Depth = 0) const { 301 return llvm::MaskedValueIsZero(V, Mask, TD, Depth); 302 } 303 unsigned ComputeNumSignBits(Value *Op, unsigned Depth = 0) const { 304 return llvm::ComputeNumSignBits(Op, TD, Depth); 305 } 306 307 private: 308 309 /// SimplifyAssociativeOrCommutative - This performs a few simplifications for 310 /// operators which are associative or commutative. 311 bool SimplifyAssociativeOrCommutative(BinaryOperator &I); 312 313 /// SimplifyUsingDistributiveLaws - This tries to simplify binary operations 314 /// which some other binary operation distributes over either by factorizing 315 /// out common terms (eg "(A*B)+(A*C)" -> "A*(B+C)") or expanding out if this 316 /// results in simplifications (eg: "A & (B | C) -> (A&B) | (A&C)" if this is 317 /// a win). Returns the simplified value, or null if it didn't simplify. 318 Value *SimplifyUsingDistributiveLaws(BinaryOperator &I); 319 320 /// SimplifyDemandedUseBits - Attempts to replace V with a simpler value 321 /// based on the demanded bits. 322 Value *SimplifyDemandedUseBits(Value *V, APInt DemandedMask, 323 APInt& KnownZero, APInt& KnownOne, 324 unsigned Depth); 325 bool SimplifyDemandedBits(Use &U, APInt DemandedMask, 326 APInt& KnownZero, APInt& KnownOne, 327 unsigned Depth=0); 328 329 /// SimplifyDemandedInstructionBits - Inst is an integer instruction that 330 /// SimplifyDemandedBits knows about. See if the instruction has any 331 /// properties that allow us to simplify its operands. 332 bool SimplifyDemandedInstructionBits(Instruction &Inst); 333 334 Value *SimplifyDemandedVectorElts(Value *V, APInt DemandedElts, 335 APInt& UndefElts, unsigned Depth = 0); 336 337 // FoldOpIntoPhi - Given a binary operator, cast instruction, or select 338 // which has a PHI node as operand #0, see if we can fold the instruction 339 // into the PHI (which is only possible if all operands to the PHI are 340 // constants). 341 // 342 Instruction *FoldOpIntoPhi(Instruction &I); 343 344 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary" 345 // operator and they all are only used by the PHI, PHI together their 346 // inputs, and do the operation once, to the result of the PHI. 347 Instruction *FoldPHIArgOpIntoPHI(PHINode &PN); 348 Instruction *FoldPHIArgBinOpIntoPHI(PHINode &PN); 349 Instruction *FoldPHIArgGEPIntoPHI(PHINode &PN); 350 Instruction *FoldPHIArgLoadIntoPHI(PHINode &PN); 351 352 353 Instruction *OptAndOp(Instruction *Op, ConstantInt *OpRHS, 354 ConstantInt *AndRHS, BinaryOperator &TheAnd); 355 356 Value *FoldLogicalPlusAnd(Value *LHS, Value *RHS, ConstantInt *Mask, 357 bool isSub, Instruction &I); 358 Value *InsertRangeTest(Value *V, Constant *Lo, Constant *Hi, 359 bool isSigned, bool Inside); 360 Instruction *PromoteCastOfAllocation(BitCastInst &CI, AllocaInst &AI); 361 Instruction *MatchBSwap(BinaryOperator &I); 362 bool SimplifyStoreAtEndOfBlock(StoreInst &SI); 363 Instruction *SimplifyMemTransfer(MemIntrinsic *MI); 364 Instruction *SimplifyMemSet(MemSetInst *MI); 365 366 367 Value *EvaluateInDifferentType(Value *V, Type *Ty, bool isSigned); 368 }; 369 370 371 372 } // end namespace llvm. 373 374 #endif 375
