1 //===- InstructionCombining.cpp - Combine multiple instructions -----------===//
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 // InstructionCombining - Combine instructions to form fewer, simple
11 // instructions. This pass does not modify the CFG. This pass is where
12 // algebraic simplification happens.
14 // This pass combines things like:
20 // This is a simple worklist driven algorithm.
22 // This pass guarantees that the following canonicalizations are performed on
24 // 1. If a binary operator has a constant operand, it is moved to the RHS
25 // 2. Bitwise operators with constant operands are always grouped so that
26 // shifts are performed first, then or's, then and's, then xor's.
27 // 3. Compare instructions are converted from <,>,<=,>= to ==,!= if possible
28 // 4. All cmp instructions on boolean values are replaced with logical ops
29 // 5. add X, X is represented as (X*2) => (X << 1)
30 // 6. Multiplies with a power-of-two constant argument are transformed into
34 //===----------------------------------------------------------------------===//
36 #define DEBUG_TYPE "instcombine"
37 #include "llvm/Transforms/Scalar.h"
38 #include "llvm/IntrinsicInst.h"
39 #include "llvm/Pass.h"
40 #include "llvm/DerivedTypes.h"
41 #include "llvm/GlobalVariable.h"
42 #include "llvm/Analysis/ConstantFolding.h"
43 #include "llvm/Analysis/ValueTracking.h"
44 #include "llvm/Target/TargetData.h"
45 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
46 #include "llvm/Transforms/Utils/Local.h"
47 #include "llvm/Support/CallSite.h"
48 #include "llvm/Support/ConstantRange.h"
49 #include "llvm/Support/Debug.h"
50 #include "llvm/Support/GetElementPtrTypeIterator.h"
51 #include "llvm/Support/InstVisitor.h"
52 #include "llvm/Support/MathExtras.h"
53 #include "llvm/Support/PatternMatch.h"
54 #include "llvm/Support/Compiler.h"
55 #include "llvm/ADT/DenseMap.h"
56 #include "llvm/ADT/SmallVector.h"
57 #include "llvm/ADT/SmallPtrSet.h"
58 #include "llvm/ADT/Statistic.h"
59 #include "llvm/ADT/STLExtras.h"
64 using namespace llvm::PatternMatch;
66 STATISTIC(NumCombined , "Number of insts combined");
67 STATISTIC(NumConstProp, "Number of constant folds");
68 STATISTIC(NumDeadInst , "Number of dead inst eliminated");
69 STATISTIC(NumDeadStore, "Number of dead stores eliminated");
70 STATISTIC(NumSunkInst , "Number of instructions sunk");
73 class VISIBILITY_HIDDEN InstCombiner
74 : public FunctionPass,
75 public InstVisitor<InstCombiner, Instruction*> {
76 // Worklist of all of the instructions that need to be simplified.
77 SmallVector<Instruction*, 256> Worklist;
78 DenseMap<Instruction*, unsigned> WorklistMap;
80 bool MustPreserveLCSSA;
82 static char ID; // Pass identification, replacement for typeid
83 InstCombiner() : FunctionPass(&ID) {}
85 /// AddToWorkList - Add the specified instruction to the worklist if it
86 /// isn't already in it.
87 void AddToWorkList(Instruction *I) {
88 if (WorklistMap.insert(std::make_pair(I, Worklist.size())).second)
89 Worklist.push_back(I);
92 // RemoveFromWorkList - remove I from the worklist if it exists.
93 void RemoveFromWorkList(Instruction *I) {
94 DenseMap<Instruction*, unsigned>::iterator It = WorklistMap.find(I);
95 if (It == WorklistMap.end()) return; // Not in worklist.
97 // Don't bother moving everything down, just null out the slot.
98 Worklist[It->second] = 0;
100 WorklistMap.erase(It);
103 Instruction *RemoveOneFromWorkList() {
104 Instruction *I = Worklist.back();
106 WorklistMap.erase(I);
111 /// AddUsersToWorkList - When an instruction is simplified, add all users of
112 /// the instruction to the work lists because they might get more simplified
115 void AddUsersToWorkList(Value &I) {
116 for (Value::use_iterator UI = I.use_begin(), UE = I.use_end();
118 AddToWorkList(cast<Instruction>(*UI));
121 /// AddUsesToWorkList - When an instruction is simplified, add operands to
122 /// the work lists because they might get more simplified now.
124 void AddUsesToWorkList(Instruction &I) {
125 for (User::op_iterator i = I.op_begin(), e = I.op_end(); i != e; ++i)
126 if (Instruction *Op = dyn_cast<Instruction>(*i))
130 /// AddSoonDeadInstToWorklist - The specified instruction is about to become
131 /// dead. Add all of its operands to the worklist, turning them into
132 /// undef's to reduce the number of uses of those instructions.
134 /// Return the specified operand before it is turned into an undef.
136 Value *AddSoonDeadInstToWorklist(Instruction &I, unsigned op) {
137 Value *R = I.getOperand(op);
139 for (User::op_iterator i = I.op_begin(), e = I.op_end(); i != e; ++i)
140 if (Instruction *Op = dyn_cast<Instruction>(*i)) {
142 // Set the operand to undef to drop the use.
143 *i = UndefValue::get(Op->getType());
150 virtual bool runOnFunction(Function &F);
152 bool DoOneIteration(Function &F, unsigned ItNum);
154 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
155 AU.addRequired<TargetData>();
156 AU.addPreservedID(LCSSAID);
157 AU.setPreservesCFG();
160 TargetData &getTargetData() const { return *TD; }
162 // Visitation implementation - Implement instruction combining for different
163 // instruction types. The semantics are as follows:
165 // null - No change was made
166 // I - Change was made, I is still valid, I may be dead though
167 // otherwise - Change was made, replace I with returned instruction
169 Instruction *visitAdd(BinaryOperator &I);
170 Instruction *visitSub(BinaryOperator &I);
171 Instruction *visitMul(BinaryOperator &I);
172 Instruction *visitURem(BinaryOperator &I);
173 Instruction *visitSRem(BinaryOperator &I);
174 Instruction *visitFRem(BinaryOperator &I);
175 bool SimplifyDivRemOfSelect(BinaryOperator &I);
176 Instruction *commonRemTransforms(BinaryOperator &I);
177 Instruction *commonIRemTransforms(BinaryOperator &I);
178 Instruction *commonDivTransforms(BinaryOperator &I);
179 Instruction *commonIDivTransforms(BinaryOperator &I);
180 Instruction *visitUDiv(BinaryOperator &I);
181 Instruction *visitSDiv(BinaryOperator &I);
182 Instruction *visitFDiv(BinaryOperator &I);
183 Instruction *FoldAndOfICmps(Instruction &I, ICmpInst *LHS, ICmpInst *RHS);
184 Instruction *visitAnd(BinaryOperator &I);
185 Instruction *FoldOrOfICmps(Instruction &I, ICmpInst *LHS, ICmpInst *RHS);
186 Instruction *FoldOrWithConstants(BinaryOperator &I, Value *Op,
187 Value *A, Value *B, Value *C);
188 Instruction *visitOr (BinaryOperator &I);
189 Instruction *visitXor(BinaryOperator &I);
190 Instruction *visitShl(BinaryOperator &I);
191 Instruction *visitAShr(BinaryOperator &I);
192 Instruction *visitLShr(BinaryOperator &I);
193 Instruction *commonShiftTransforms(BinaryOperator &I);
194 Instruction *FoldFCmp_IntToFP_Cst(FCmpInst &I, Instruction *LHSI,
196 Instruction *visitFCmpInst(FCmpInst &I);
197 Instruction *visitICmpInst(ICmpInst &I);
198 Instruction *visitICmpInstWithCastAndCast(ICmpInst &ICI);
199 Instruction *visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
202 Instruction *FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
203 ConstantInt *DivRHS);
205 Instruction *FoldGEPICmp(User *GEPLHS, Value *RHS,
206 ICmpInst::Predicate Cond, Instruction &I);
207 Instruction *FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
209 Instruction *commonCastTransforms(CastInst &CI);
210 Instruction *commonIntCastTransforms(CastInst &CI);
211 Instruction *commonPointerCastTransforms(CastInst &CI);
212 Instruction *visitTrunc(TruncInst &CI);
213 Instruction *visitZExt(ZExtInst &CI);
214 Instruction *visitSExt(SExtInst &CI);
215 Instruction *visitFPTrunc(FPTruncInst &CI);
216 Instruction *visitFPExt(CastInst &CI);
217 Instruction *visitFPToUI(FPToUIInst &FI);
218 Instruction *visitFPToSI(FPToSIInst &FI);
219 Instruction *visitUIToFP(CastInst &CI);
220 Instruction *visitSIToFP(CastInst &CI);
221 Instruction *visitPtrToInt(PtrToIntInst &CI);
222 Instruction *visitIntToPtr(IntToPtrInst &CI);
223 Instruction *visitBitCast(BitCastInst &CI);
224 Instruction *FoldSelectOpOp(SelectInst &SI, Instruction *TI,
226 Instruction *visitSelectInst(SelectInst &SI);
227 Instruction *visitSelectInstWithICmp(SelectInst &SI, ICmpInst *ICI);
228 Instruction *visitCallInst(CallInst &CI);
229 Instruction *visitInvokeInst(InvokeInst &II);
230 Instruction *visitPHINode(PHINode &PN);
231 Instruction *visitGetElementPtrInst(GetElementPtrInst &GEP);
232 Instruction *visitAllocationInst(AllocationInst &AI);
233 Instruction *visitFreeInst(FreeInst &FI);
234 Instruction *visitLoadInst(LoadInst &LI);
235 Instruction *visitStoreInst(StoreInst &SI);
236 Instruction *visitBranchInst(BranchInst &BI);
237 Instruction *visitSwitchInst(SwitchInst &SI);
238 Instruction *visitInsertElementInst(InsertElementInst &IE);
239 Instruction *visitExtractElementInst(ExtractElementInst &EI);
240 Instruction *visitShuffleVectorInst(ShuffleVectorInst &SVI);
241 Instruction *visitExtractValueInst(ExtractValueInst &EV);
243 // visitInstruction - Specify what to return for unhandled instructions...
244 Instruction *visitInstruction(Instruction &I) { return 0; }
247 Instruction *visitCallSite(CallSite CS);
248 bool transformConstExprCastCall(CallSite CS);
249 Instruction *transformCallThroughTrampoline(CallSite CS);
250 Instruction *transformZExtICmp(ICmpInst *ICI, Instruction &CI,
251 bool DoXform = true);
252 bool WillNotOverflowSignedAdd(Value *LHS, Value *RHS);
253 DbgDeclareInst *hasOneUsePlusDeclare(Value *V);
257 // InsertNewInstBefore - insert an instruction New before instruction Old
258 // in the program. Add the new instruction to the worklist.
260 Instruction *InsertNewInstBefore(Instruction *New, Instruction &Old) {
261 assert(New && New->getParent() == 0 &&
262 "New instruction already inserted into a basic block!");
263 BasicBlock *BB = Old.getParent();
264 BB->getInstList().insert(&Old, New); // Insert inst
269 /// InsertCastBefore - Insert a cast of V to TY before the instruction POS.
270 /// This also adds the cast to the worklist. Finally, this returns the
272 Value *InsertCastBefore(Instruction::CastOps opc, Value *V, const Type *Ty,
274 if (V->getType() == Ty) return V;
276 if (Constant *CV = dyn_cast<Constant>(V))
277 return ConstantExpr::getCast(opc, CV, Ty);
279 Instruction *C = CastInst::Create(opc, V, Ty, V->getName(), &Pos);
284 Value *InsertBitCastBefore(Value *V, const Type *Ty, Instruction &Pos) {
285 return InsertCastBefore(Instruction::BitCast, V, Ty, Pos);
289 // ReplaceInstUsesWith - This method is to be used when an instruction is
290 // found to be dead, replacable with another preexisting expression. Here
291 // we add all uses of I to the worklist, replace all uses of I with the new
292 // value, then return I, so that the inst combiner will know that I was
295 Instruction *ReplaceInstUsesWith(Instruction &I, Value *V) {
296 AddUsersToWorkList(I); // Add all modified instrs to worklist
298 I.replaceAllUsesWith(V);
301 // If we are replacing the instruction with itself, this must be in a
302 // segment of unreachable code, so just clobber the instruction.
303 I.replaceAllUsesWith(UndefValue::get(I.getType()));
308 // EraseInstFromFunction - When dealing with an instruction that has side
309 // effects or produces a void value, we can't rely on DCE to delete the
310 // instruction. Instead, visit methods should return the value returned by
312 Instruction *EraseInstFromFunction(Instruction &I) {
313 assert(I.use_empty() && "Cannot erase instruction that is used!");
314 AddUsesToWorkList(I);
315 RemoveFromWorkList(&I);
317 return 0; // Don't do anything with FI
320 void ComputeMaskedBits(Value *V, const APInt &Mask, APInt &KnownZero,
321 APInt &KnownOne, unsigned Depth = 0) const {
322 return llvm::ComputeMaskedBits(V, Mask, KnownZero, KnownOne, TD, Depth);
325 bool MaskedValueIsZero(Value *V, const APInt &Mask,
326 unsigned Depth = 0) const {
327 return llvm::MaskedValueIsZero(V, Mask, TD, Depth);
329 unsigned ComputeNumSignBits(Value *Op, unsigned Depth = 0) const {
330 return llvm::ComputeNumSignBits(Op, TD, Depth);
335 /// SimplifyCommutative - This performs a few simplifications for
336 /// commutative operators.
337 bool SimplifyCommutative(BinaryOperator &I);
339 /// SimplifyCompare - This reorders the operands of a CmpInst to get them in
340 /// most-complex to least-complex order.
341 bool SimplifyCompare(CmpInst &I);
343 /// SimplifyDemandedUseBits - Attempts to replace V with a simpler value
344 /// based on the demanded bits.
345 Value *SimplifyDemandedUseBits(Value *V, APInt DemandedMask,
346 APInt& KnownZero, APInt& KnownOne,
348 bool SimplifyDemandedBits(Use &U, APInt DemandedMask,
349 APInt& KnownZero, APInt& KnownOne,
352 /// SimplifyDemandedInstructionBits - Inst is an integer instruction that
353 /// SimplifyDemandedBits knows about. See if the instruction has any
354 /// properties that allow us to simplify its operands.
355 bool SimplifyDemandedInstructionBits(Instruction &Inst);
357 Value *SimplifyDemandedVectorElts(Value *V, APInt DemandedElts,
358 APInt& UndefElts, unsigned Depth = 0);
360 // FoldOpIntoPhi - Given a binary operator or cast instruction which has a
361 // PHI node as operand #0, see if we can fold the instruction into the PHI
362 // (which is only possible if all operands to the PHI are constants).
363 Instruction *FoldOpIntoPhi(Instruction &I);
365 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
366 // operator and they all are only used by the PHI, PHI together their
367 // inputs, and do the operation once, to the result of the PHI.
368 Instruction *FoldPHIArgOpIntoPHI(PHINode &PN);
369 Instruction *FoldPHIArgBinOpIntoPHI(PHINode &PN);
370 Instruction *FoldPHIArgGEPIntoPHI(PHINode &PN);
373 Instruction *OptAndOp(Instruction *Op, ConstantInt *OpRHS,
374 ConstantInt *AndRHS, BinaryOperator &TheAnd);
376 Value *FoldLogicalPlusAnd(Value *LHS, Value *RHS, ConstantInt *Mask,
377 bool isSub, Instruction &I);
378 Instruction *InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
379 bool isSigned, bool Inside, Instruction &IB);
380 Instruction *PromoteCastOfAllocation(BitCastInst &CI, AllocationInst &AI);
381 Instruction *MatchBSwap(BinaryOperator &I);
382 bool SimplifyStoreAtEndOfBlock(StoreInst &SI);
383 Instruction *SimplifyMemTransfer(MemIntrinsic *MI);
384 Instruction *SimplifyMemSet(MemSetInst *MI);
387 Value *EvaluateInDifferentType(Value *V, const Type *Ty, bool isSigned);
389 bool CanEvaluateInDifferentType(Value *V, const IntegerType *Ty,
390 unsigned CastOpc, int &NumCastsRemoved);
391 unsigned GetOrEnforceKnownAlignment(Value *V,
392 unsigned PrefAlign = 0);
397 char InstCombiner::ID = 0;
398 static RegisterPass<InstCombiner>
399 X("instcombine", "Combine redundant instructions");
401 // getComplexity: Assign a complexity or rank value to LLVM Values...
402 // 0 -> undef, 1 -> Const, 2 -> Other, 3 -> Arg, 3 -> Unary, 4 -> OtherInst
403 static unsigned getComplexity(Value *V) {
404 if (isa<Instruction>(V)) {
405 if (BinaryOperator::isNeg(V) || BinaryOperator::isNot(V))
409 if (isa<Argument>(V)) return 3;
410 return isa<Constant>(V) ? (isa<UndefValue>(V) ? 0 : 1) : 2;
413 // isOnlyUse - Return true if this instruction will be deleted if we stop using
415 static bool isOnlyUse(Value *V) {
416 return V->hasOneUse() || isa<Constant>(V);
419 // getPromotedType - Return the specified type promoted as it would be to pass
420 // though a va_arg area...
421 static const Type *getPromotedType(const Type *Ty) {
422 if (const IntegerType* ITy = dyn_cast<IntegerType>(Ty)) {
423 if (ITy->getBitWidth() < 32)
424 return Type::Int32Ty;
429 /// getBitCastOperand - If the specified operand is a CastInst, a constant
430 /// expression bitcast, or a GetElementPtrInst with all zero indices, return the
431 /// operand value, otherwise return null.
432 static Value *getBitCastOperand(Value *V) {
433 if (BitCastInst *I = dyn_cast<BitCastInst>(V))
435 return I->getOperand(0);
436 else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(V)) {
437 // GetElementPtrInst?
438 if (GEP->hasAllZeroIndices())
439 return GEP->getOperand(0);
440 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) {
441 if (CE->getOpcode() == Instruction::BitCast)
442 // BitCast ConstantExp?
443 return CE->getOperand(0);
444 else if (CE->getOpcode() == Instruction::GetElementPtr) {
445 // GetElementPtr ConstantExp?
446 for (User::op_iterator I = CE->op_begin() + 1, E = CE->op_end();
448 ConstantInt *CI = dyn_cast<ConstantInt>(I);
449 if (!CI || !CI->isZero())
450 // Any non-zero indices? Not cast-like.
453 // All-zero indices? This is just like casting.
454 return CE->getOperand(0);
460 /// This function is a wrapper around CastInst::isEliminableCastPair. It
461 /// simply extracts arguments and returns what that function returns.
462 static Instruction::CastOps
463 isEliminableCastPair(
464 const CastInst *CI, ///< The first cast instruction
465 unsigned opcode, ///< The opcode of the second cast instruction
466 const Type *DstTy, ///< The target type for the second cast instruction
467 TargetData *TD ///< The target data for pointer size
470 const Type *SrcTy = CI->getOperand(0)->getType(); // A from above
471 const Type *MidTy = CI->getType(); // B from above
473 // Get the opcodes of the two Cast instructions
474 Instruction::CastOps firstOp = Instruction::CastOps(CI->getOpcode());
475 Instruction::CastOps secondOp = Instruction::CastOps(opcode);
477 unsigned Res = CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy,
478 DstTy, TD->getIntPtrType());
480 // We don't want to form an inttoptr or ptrtoint that converts to an integer
481 // type that differs from the pointer size.
482 if ((Res == Instruction::IntToPtr && SrcTy != TD->getIntPtrType()) ||
483 (Res == Instruction::PtrToInt && DstTy != TD->getIntPtrType()))
486 return Instruction::CastOps(Res);
489 /// ValueRequiresCast - Return true if the cast from "V to Ty" actually results
490 /// in any code being generated. It does not require codegen if V is simple
491 /// enough or if the cast can be folded into other casts.
492 static bool ValueRequiresCast(Instruction::CastOps opcode, const Value *V,
493 const Type *Ty, TargetData *TD) {
494 if (V->getType() == Ty || isa<Constant>(V)) return false;
496 // If this is another cast that can be eliminated, it isn't codegen either.
497 if (const CastInst *CI = dyn_cast<CastInst>(V))
498 if (isEliminableCastPair(CI, opcode, Ty, TD))
503 // SimplifyCommutative - This performs a few simplifications for commutative
506 // 1. Order operands such that they are listed from right (least complex) to
507 // left (most complex). This puts constants before unary operators before
510 // 2. Transform: (op (op V, C1), C2) ==> (op V, (op C1, C2))
511 // 3. Transform: (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
513 bool InstCombiner::SimplifyCommutative(BinaryOperator &I) {
514 bool Changed = false;
515 if (getComplexity(I.getOperand(0)) < getComplexity(I.getOperand(1)))
516 Changed = !I.swapOperands();
518 if (!I.isAssociative()) return Changed;
519 Instruction::BinaryOps Opcode = I.getOpcode();
520 if (BinaryOperator *Op = dyn_cast<BinaryOperator>(I.getOperand(0)))
521 if (Op->getOpcode() == Opcode && isa<Constant>(Op->getOperand(1))) {
522 if (isa<Constant>(I.getOperand(1))) {
523 Constant *Folded = ConstantExpr::get(I.getOpcode(),
524 cast<Constant>(I.getOperand(1)),
525 cast<Constant>(Op->getOperand(1)));
526 I.setOperand(0, Op->getOperand(0));
527 I.setOperand(1, Folded);
529 } else if (BinaryOperator *Op1=dyn_cast<BinaryOperator>(I.getOperand(1)))
530 if (Op1->getOpcode() == Opcode && isa<Constant>(Op1->getOperand(1)) &&
531 isOnlyUse(Op) && isOnlyUse(Op1)) {
532 Constant *C1 = cast<Constant>(Op->getOperand(1));
533 Constant *C2 = cast<Constant>(Op1->getOperand(1));
535 // Fold (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
536 Constant *Folded = ConstantExpr::get(I.getOpcode(), C1, C2);
537 Instruction *New = BinaryOperator::Create(Opcode, Op->getOperand(0),
541 I.setOperand(0, New);
542 I.setOperand(1, Folded);
549 /// SimplifyCompare - For a CmpInst this function just orders the operands
550 /// so that theyare listed from right (least complex) to left (most complex).
551 /// This puts constants before unary operators before binary operators.
552 bool InstCombiner::SimplifyCompare(CmpInst &I) {
553 if (getComplexity(I.getOperand(0)) >= getComplexity(I.getOperand(1)))
556 // Compare instructions are not associative so there's nothing else we can do.
560 // dyn_castNegVal - Given a 'sub' instruction, return the RHS of the instruction
561 // if the LHS is a constant zero (which is the 'negate' form).
563 static inline Value *dyn_castNegVal(Value *V) {
564 if (BinaryOperator::isNeg(V))
565 return BinaryOperator::getNegArgument(V);
567 // Constants can be considered to be negated values if they can be folded.
568 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
569 return ConstantExpr::getNeg(C);
571 if (ConstantVector *C = dyn_cast<ConstantVector>(V))
572 if (C->getType()->getElementType()->isInteger())
573 return ConstantExpr::getNeg(C);
578 static inline Value *dyn_castNotVal(Value *V) {
579 if (BinaryOperator::isNot(V))
580 return BinaryOperator::getNotArgument(V);
582 // Constants can be considered to be not'ed values...
583 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
584 return ConstantInt::get(~C->getValue());
588 // dyn_castFoldableMul - If this value is a multiply that can be folded into
589 // other computations (because it has a constant operand), return the
590 // non-constant operand of the multiply, and set CST to point to the multiplier.
591 // Otherwise, return null.
593 static inline Value *dyn_castFoldableMul(Value *V, ConstantInt *&CST) {
594 if (V->hasOneUse() && V->getType()->isInteger())
595 if (Instruction *I = dyn_cast<Instruction>(V)) {
596 if (I->getOpcode() == Instruction::Mul)
597 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1))))
598 return I->getOperand(0);
599 if (I->getOpcode() == Instruction::Shl)
600 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1)))) {
601 // The multiplier is really 1 << CST.
602 uint32_t BitWidth = cast<IntegerType>(V->getType())->getBitWidth();
603 uint32_t CSTVal = CST->getLimitedValue(BitWidth);
604 CST = ConstantInt::get(APInt(BitWidth, 1).shl(CSTVal));
605 return I->getOperand(0);
611 /// dyn_castGetElementPtr - If this is a getelementptr instruction or constant
612 /// expression, return it.
613 static User *dyn_castGetElementPtr(Value *V) {
614 if (isa<GetElementPtrInst>(V)) return cast<User>(V);
615 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
616 if (CE->getOpcode() == Instruction::GetElementPtr)
617 return cast<User>(V);
621 /// getOpcode - If this is an Instruction or a ConstantExpr, return the
622 /// opcode value. Otherwise return UserOp1.
623 static unsigned getOpcode(const Value *V) {
624 if (const Instruction *I = dyn_cast<Instruction>(V))
625 return I->getOpcode();
626 if (const ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
627 return CE->getOpcode();
628 // Use UserOp1 to mean there's no opcode.
629 return Instruction::UserOp1;
632 /// AddOne - Add one to a ConstantInt
633 static ConstantInt *AddOne(ConstantInt *C) {
634 APInt Val(C->getValue());
635 return ConstantInt::get(++Val);
637 /// SubOne - Subtract one from a ConstantInt
638 static ConstantInt *SubOne(ConstantInt *C) {
639 APInt Val(C->getValue());
640 return ConstantInt::get(--Val);
642 /// Add - Add two ConstantInts together
643 static ConstantInt *Add(ConstantInt *C1, ConstantInt *C2) {
644 return ConstantInt::get(C1->getValue() + C2->getValue());
646 /// And - Bitwise AND two ConstantInts together
647 static ConstantInt *And(ConstantInt *C1, ConstantInt *C2) {
648 return ConstantInt::get(C1->getValue() & C2->getValue());
650 /// Subtract - Subtract one ConstantInt from another
651 static ConstantInt *Subtract(ConstantInt *C1, ConstantInt *C2) {
652 return ConstantInt::get(C1->getValue() - C2->getValue());
654 /// Multiply - Multiply two ConstantInts together
655 static ConstantInt *Multiply(ConstantInt *C1, ConstantInt *C2) {
656 return ConstantInt::get(C1->getValue() * C2->getValue());
658 /// MultiplyOverflows - True if the multiply can not be expressed in an int
660 static bool MultiplyOverflows(ConstantInt *C1, ConstantInt *C2, bool sign) {
661 uint32_t W = C1->getBitWidth();
662 APInt LHSExt = C1->getValue(), RHSExt = C2->getValue();
671 APInt MulExt = LHSExt * RHSExt;
674 APInt Min = APInt::getSignedMinValue(W).sext(W * 2);
675 APInt Max = APInt::getSignedMaxValue(W).sext(W * 2);
676 return MulExt.slt(Min) || MulExt.sgt(Max);
678 return MulExt.ugt(APInt::getLowBitsSet(W * 2, W));
682 /// ShrinkDemandedConstant - Check to see if the specified operand of the
683 /// specified instruction is a constant integer. If so, check to see if there
684 /// are any bits set in the constant that are not demanded. If so, shrink the
685 /// constant and return true.
686 static bool ShrinkDemandedConstant(Instruction *I, unsigned OpNo,
688 assert(I && "No instruction?");
689 assert(OpNo < I->getNumOperands() && "Operand index too large");
691 // If the operand is not a constant integer, nothing to do.
692 ConstantInt *OpC = dyn_cast<ConstantInt>(I->getOperand(OpNo));
693 if (!OpC) return false;
695 // If there are no bits set that aren't demanded, nothing to do.
696 Demanded.zextOrTrunc(OpC->getValue().getBitWidth());
697 if ((~Demanded & OpC->getValue()) == 0)
700 // This instruction is producing bits that are not demanded. Shrink the RHS.
701 Demanded &= OpC->getValue();
702 I->setOperand(OpNo, ConstantInt::get(Demanded));
706 // ComputeSignedMinMaxValuesFromKnownBits - Given a signed integer type and a
707 // set of known zero and one bits, compute the maximum and minimum values that
708 // could have the specified known zero and known one bits, returning them in
710 static void ComputeSignedMinMaxValuesFromKnownBits(const Type *Ty,
711 const APInt& KnownZero,
712 const APInt& KnownOne,
713 APInt& Min, APInt& Max) {
714 uint32_t BitWidth = cast<IntegerType>(Ty)->getBitWidth();
715 assert(KnownZero.getBitWidth() == BitWidth &&
716 KnownOne.getBitWidth() == BitWidth &&
717 Min.getBitWidth() == BitWidth && Max.getBitWidth() == BitWidth &&
718 "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth.");
719 APInt UnknownBits = ~(KnownZero|KnownOne);
721 // The minimum value is when all unknown bits are zeros, EXCEPT for the sign
722 // bit if it is unknown.
724 Max = KnownOne|UnknownBits;
726 if (UnknownBits[BitWidth-1]) { // Sign bit is unknown
728 Max.clear(BitWidth-1);
732 // ComputeUnsignedMinMaxValuesFromKnownBits - Given an unsigned integer type and
733 // a set of known zero and one bits, compute the maximum and minimum values that
734 // could have the specified known zero and known one bits, returning them in
736 static void ComputeUnsignedMinMaxValuesFromKnownBits(const Type *Ty,
737 const APInt &KnownZero,
738 const APInt &KnownOne,
739 APInt &Min, APInt &Max) {
740 uint32_t BitWidth = cast<IntegerType>(Ty)->getBitWidth(); BitWidth = BitWidth;
741 assert(KnownZero.getBitWidth() == BitWidth &&
742 KnownOne.getBitWidth() == BitWidth &&
743 Min.getBitWidth() == BitWidth && Max.getBitWidth() &&
744 "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth.");
745 APInt UnknownBits = ~(KnownZero|KnownOne);
747 // The minimum value is when the unknown bits are all zeros.
749 // The maximum value is when the unknown bits are all ones.
750 Max = KnownOne|UnknownBits;
753 /// SimplifyDemandedInstructionBits - Inst is an integer instruction that
754 /// SimplifyDemandedBits knows about. See if the instruction has any
755 /// properties that allow us to simplify its operands.
756 bool InstCombiner::SimplifyDemandedInstructionBits(Instruction &Inst) {
757 unsigned BitWidth = cast<IntegerType>(Inst.getType())->getBitWidth();
758 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
759 APInt DemandedMask(APInt::getAllOnesValue(BitWidth));
761 Value *V = SimplifyDemandedUseBits(&Inst, DemandedMask,
762 KnownZero, KnownOne, 0);
763 if (V == 0) return false;
764 if (V == &Inst) return true;
765 ReplaceInstUsesWith(Inst, V);
769 /// SimplifyDemandedBits - This form of SimplifyDemandedBits simplifies the
770 /// specified instruction operand if possible, updating it in place. It returns
771 /// true if it made any change and false otherwise.
772 bool InstCombiner::SimplifyDemandedBits(Use &U, APInt DemandedMask,
773 APInt &KnownZero, APInt &KnownOne,
775 Value *NewVal = SimplifyDemandedUseBits(U.get(), DemandedMask,
776 KnownZero, KnownOne, Depth);
777 if (NewVal == 0) return false;
783 /// SimplifyDemandedUseBits - This function attempts to replace V with a simpler
784 /// value based on the demanded bits. When this function is called, it is known
785 /// that only the bits set in DemandedMask of the result of V are ever used
786 /// downstream. Consequently, depending on the mask and V, it may be possible
787 /// to replace V with a constant or one of its operands. In such cases, this
788 /// function does the replacement and returns true. In all other cases, it
789 /// returns false after analyzing the expression and setting KnownOne and known
790 /// to be one in the expression. KnownZero contains all the bits that are known
791 /// to be zero in the expression. These are provided to potentially allow the
792 /// caller (which might recursively be SimplifyDemandedBits itself) to simplify
793 /// the expression. KnownOne and KnownZero always follow the invariant that
794 /// KnownOne & KnownZero == 0. That is, a bit can't be both 1 and 0. Note that
795 /// the bits in KnownOne and KnownZero may only be accurate for those bits set
796 /// in DemandedMask. Note also that the bitwidth of V, DemandedMask, KnownZero
797 /// and KnownOne must all be the same.
799 /// This returns null if it did not change anything and it permits no
800 /// simplification. This returns V itself if it did some simplification of V's
801 /// operands based on the information about what bits are demanded. This returns
802 /// some other non-null value if it found out that V is equal to another value
803 /// in the context where the specified bits are demanded, but not for all users.
804 Value *InstCombiner::SimplifyDemandedUseBits(Value *V, APInt DemandedMask,
805 APInt &KnownZero, APInt &KnownOne,
807 assert(V != 0 && "Null pointer of Value???");
808 assert(Depth <= 6 && "Limit Search Depth");
809 uint32_t BitWidth = DemandedMask.getBitWidth();
810 const IntegerType *VTy = cast<IntegerType>(V->getType());
811 assert(VTy->getBitWidth() == BitWidth &&
812 KnownZero.getBitWidth() == BitWidth &&
813 KnownOne.getBitWidth() == BitWidth &&
814 "Value *V, DemandedMask, KnownZero and KnownOne \
815 must have same BitWidth");
816 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
817 // We know all of the bits for a constant!
818 KnownOne = CI->getValue() & DemandedMask;
819 KnownZero = ~KnownOne & DemandedMask;
825 if (DemandedMask == 0) { // Not demanding any bits from V.
826 if (isa<UndefValue>(V))
828 return UndefValue::get(VTy);
831 if (Depth == 6) // Limit search depth.
834 Instruction *I = dyn_cast<Instruction>(V);
835 if (!I) return 0; // Only analyze instructions.
837 APInt LHSKnownZero(BitWidth, 0), LHSKnownOne(BitWidth, 0);
838 APInt &RHSKnownZero = KnownZero, &RHSKnownOne = KnownOne;
840 // If there are multiple uses of this value and we aren't at the root, then
841 // we can't do any simplifications of the operands, because DemandedMask
842 // only reflects the bits demanded by *one* of the users.
843 if (Depth != 0 && !I->hasOneUse()) {
844 // Despite the fact that we can't simplify this instruction in all User's
845 // context, we can at least compute the knownzero/knownone bits, and we can
846 // do simplifications that apply to *just* the one user if we know that
847 // this instruction has a simpler value in that context.
848 if (I->getOpcode() == Instruction::And) {
849 // If either the LHS or the RHS are Zero, the result is zero.
850 ComputeMaskedBits(I->getOperand(1), DemandedMask,
851 RHSKnownZero, RHSKnownOne, Depth+1);
852 ComputeMaskedBits(I->getOperand(0), DemandedMask & ~RHSKnownZero,
853 LHSKnownZero, LHSKnownOne, Depth+1);
855 // If all of the demanded bits are known 1 on one side, return the other.
856 // These bits cannot contribute to the result of the 'and' in this
858 if ((DemandedMask & ~LHSKnownZero & RHSKnownOne) ==
859 (DemandedMask & ~LHSKnownZero))
860 return I->getOperand(0);
861 if ((DemandedMask & ~RHSKnownZero & LHSKnownOne) ==
862 (DemandedMask & ~RHSKnownZero))
863 return I->getOperand(1);
865 // If all of the demanded bits in the inputs are known zeros, return zero.
866 if ((DemandedMask & (RHSKnownZero|LHSKnownZero)) == DemandedMask)
867 return Constant::getNullValue(VTy);
869 } else if (I->getOpcode() == Instruction::Or) {
870 // We can simplify (X|Y) -> X or Y in the user's context if we know that
871 // only bits from X or Y are demanded.
873 // If either the LHS or the RHS are One, the result is One.
874 ComputeMaskedBits(I->getOperand(1), DemandedMask,
875 RHSKnownZero, RHSKnownOne, Depth+1);
876 ComputeMaskedBits(I->getOperand(0), DemandedMask & ~RHSKnownOne,
877 LHSKnownZero, LHSKnownOne, Depth+1);
879 // If all of the demanded bits are known zero on one side, return the
880 // other. These bits cannot contribute to the result of the 'or' in this
882 if ((DemandedMask & ~LHSKnownOne & RHSKnownZero) ==
883 (DemandedMask & ~LHSKnownOne))
884 return I->getOperand(0);
885 if ((DemandedMask & ~RHSKnownOne & LHSKnownZero) ==
886 (DemandedMask & ~RHSKnownOne))
887 return I->getOperand(1);
889 // If all of the potentially set bits on one side are known to be set on
890 // the other side, just use the 'other' side.
891 if ((DemandedMask & (~RHSKnownZero) & LHSKnownOne) ==
892 (DemandedMask & (~RHSKnownZero)))
893 return I->getOperand(0);
894 if ((DemandedMask & (~LHSKnownZero) & RHSKnownOne) ==
895 (DemandedMask & (~LHSKnownZero)))
896 return I->getOperand(1);
899 // Compute the KnownZero/KnownOne bits to simplify things downstream.
900 ComputeMaskedBits(I, DemandedMask, KnownZero, KnownOne, Depth);
904 // If this is the root being simplified, allow it to have multiple uses,
905 // just set the DemandedMask to all bits so that we can try to simplify the
906 // operands. This allows visitTruncInst (for example) to simplify the
907 // operand of a trunc without duplicating all the logic below.
908 if (Depth == 0 && !V->hasOneUse())
909 DemandedMask = APInt::getAllOnesValue(BitWidth);
911 switch (I->getOpcode()) {
913 ComputeMaskedBits(I, DemandedMask, RHSKnownZero, RHSKnownOne, Depth);
915 case Instruction::And:
916 // If either the LHS or the RHS are Zero, the result is zero.
917 if (SimplifyDemandedBits(I->getOperandUse(1), DemandedMask,
918 RHSKnownZero, RHSKnownOne, Depth+1) ||
919 SimplifyDemandedBits(I->getOperandUse(0), DemandedMask & ~RHSKnownZero,
920 LHSKnownZero, LHSKnownOne, Depth+1))
922 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
923 assert(!(LHSKnownZero & LHSKnownOne) && "Bits known to be one AND zero?");
925 // If all of the demanded bits are known 1 on one side, return the other.
926 // These bits cannot contribute to the result of the 'and'.
927 if ((DemandedMask & ~LHSKnownZero & RHSKnownOne) ==
928 (DemandedMask & ~LHSKnownZero))
929 return I->getOperand(0);
930 if ((DemandedMask & ~RHSKnownZero & LHSKnownOne) ==
931 (DemandedMask & ~RHSKnownZero))
932 return I->getOperand(1);
934 // If all of the demanded bits in the inputs are known zeros, return zero.
935 if ((DemandedMask & (RHSKnownZero|LHSKnownZero)) == DemandedMask)
936 return Constant::getNullValue(VTy);
938 // If the RHS is a constant, see if we can simplify it.
939 if (ShrinkDemandedConstant(I, 1, DemandedMask & ~LHSKnownZero))
942 // Output known-1 bits are only known if set in both the LHS & RHS.
943 RHSKnownOne &= LHSKnownOne;
944 // Output known-0 are known to be clear if zero in either the LHS | RHS.
945 RHSKnownZero |= LHSKnownZero;
947 case Instruction::Or:
948 // If either the LHS or the RHS are One, the result is One.
949 if (SimplifyDemandedBits(I->getOperandUse(1), DemandedMask,
950 RHSKnownZero, RHSKnownOne, Depth+1) ||
951 SimplifyDemandedBits(I->getOperandUse(0), DemandedMask & ~RHSKnownOne,
952 LHSKnownZero, LHSKnownOne, Depth+1))
954 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
955 assert(!(LHSKnownZero & LHSKnownOne) && "Bits known to be one AND zero?");
957 // If all of the demanded bits are known zero on one side, return the other.
958 // These bits cannot contribute to the result of the 'or'.
959 if ((DemandedMask & ~LHSKnownOne & RHSKnownZero) ==
960 (DemandedMask & ~LHSKnownOne))
961 return I->getOperand(0);
962 if ((DemandedMask & ~RHSKnownOne & LHSKnownZero) ==
963 (DemandedMask & ~RHSKnownOne))
964 return I->getOperand(1);
966 // If all of the potentially set bits on one side are known to be set on
967 // the other side, just use the 'other' side.
968 if ((DemandedMask & (~RHSKnownZero) & LHSKnownOne) ==
969 (DemandedMask & (~RHSKnownZero)))
970 return I->getOperand(0);
971 if ((DemandedMask & (~LHSKnownZero) & RHSKnownOne) ==
972 (DemandedMask & (~LHSKnownZero)))
973 return I->getOperand(1);
975 // If the RHS is a constant, see if we can simplify it.
976 if (ShrinkDemandedConstant(I, 1, DemandedMask))
979 // Output known-0 bits are only known if clear in both the LHS & RHS.
980 RHSKnownZero &= LHSKnownZero;
981 // Output known-1 are known to be set if set in either the LHS | RHS.
982 RHSKnownOne |= LHSKnownOne;
984 case Instruction::Xor: {
985 if (SimplifyDemandedBits(I->getOperandUse(1), DemandedMask,
986 RHSKnownZero, RHSKnownOne, Depth+1) ||
987 SimplifyDemandedBits(I->getOperandUse(0), DemandedMask,
988 LHSKnownZero, LHSKnownOne, Depth+1))
990 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
991 assert(!(LHSKnownZero & LHSKnownOne) && "Bits known to be one AND zero?");
993 // If all of the demanded bits are known zero on one side, return the other.
994 // These bits cannot contribute to the result of the 'xor'.
995 if ((DemandedMask & RHSKnownZero) == DemandedMask)
996 return I->getOperand(0);
997 if ((DemandedMask & LHSKnownZero) == DemandedMask)
998 return I->getOperand(1);
1000 // Output known-0 bits are known if clear or set in both the LHS & RHS.
1001 APInt KnownZeroOut = (RHSKnownZero & LHSKnownZero) |
1002 (RHSKnownOne & LHSKnownOne);
1003 // Output known-1 are known to be set if set in only one of the LHS, RHS.
1004 APInt KnownOneOut = (RHSKnownZero & LHSKnownOne) |
1005 (RHSKnownOne & LHSKnownZero);
1007 // If all of the demanded bits are known to be zero on one side or the
1008 // other, turn this into an *inclusive* or.
1009 // e.g. (A & C1)^(B & C2) -> (A & C1)|(B & C2) iff C1&C2 == 0
1010 if ((DemandedMask & ~RHSKnownZero & ~LHSKnownZero) == 0) {
1012 BinaryOperator::CreateOr(I->getOperand(0), I->getOperand(1),
1014 return InsertNewInstBefore(Or, *I);
1017 // If all of the demanded bits on one side are known, and all of the set
1018 // bits on that side are also known to be set on the other side, turn this
1019 // into an AND, as we know the bits will be cleared.
1020 // e.g. (X | C1) ^ C2 --> (X | C1) & ~C2 iff (C1&C2) == C2
1021 if ((DemandedMask & (RHSKnownZero|RHSKnownOne)) == DemandedMask) {
1023 if ((RHSKnownOne & LHSKnownOne) == RHSKnownOne) {
1024 Constant *AndC = ConstantInt::get(~RHSKnownOne & DemandedMask);
1026 BinaryOperator::CreateAnd(I->getOperand(0), AndC, "tmp");
1027 return InsertNewInstBefore(And, *I);
1031 // If the RHS is a constant, see if we can simplify it.
1032 // FIXME: for XOR, we prefer to force bits to 1 if they will make a -1.
1033 if (ShrinkDemandedConstant(I, 1, DemandedMask))
1036 RHSKnownZero = KnownZeroOut;
1037 RHSKnownOne = KnownOneOut;
1040 case Instruction::Select:
1041 if (SimplifyDemandedBits(I->getOperandUse(2), DemandedMask,
1042 RHSKnownZero, RHSKnownOne, Depth+1) ||
1043 SimplifyDemandedBits(I->getOperandUse(1), DemandedMask,
1044 LHSKnownZero, LHSKnownOne, Depth+1))
1046 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1047 assert(!(LHSKnownZero & LHSKnownOne) && "Bits known to be one AND zero?");
1049 // If the operands are constants, see if we can simplify them.
1050 if (ShrinkDemandedConstant(I, 1, DemandedMask) ||
1051 ShrinkDemandedConstant(I, 2, DemandedMask))
1054 // Only known if known in both the LHS and RHS.
1055 RHSKnownOne &= LHSKnownOne;
1056 RHSKnownZero &= LHSKnownZero;
1058 case Instruction::Trunc: {
1059 unsigned truncBf = I->getOperand(0)->getType()->getPrimitiveSizeInBits();
1060 DemandedMask.zext(truncBf);
1061 RHSKnownZero.zext(truncBf);
1062 RHSKnownOne.zext(truncBf);
1063 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedMask,
1064 RHSKnownZero, RHSKnownOne, Depth+1))
1066 DemandedMask.trunc(BitWidth);
1067 RHSKnownZero.trunc(BitWidth);
1068 RHSKnownOne.trunc(BitWidth);
1069 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1072 case Instruction::BitCast:
1073 if (!I->getOperand(0)->getType()->isInteger())
1074 return false; // vector->int or fp->int?
1075 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedMask,
1076 RHSKnownZero, RHSKnownOne, Depth+1))
1078 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1080 case Instruction::ZExt: {
1081 // Compute the bits in the result that are not present in the input.
1082 unsigned SrcBitWidth =I->getOperand(0)->getType()->getPrimitiveSizeInBits();
1084 DemandedMask.trunc(SrcBitWidth);
1085 RHSKnownZero.trunc(SrcBitWidth);
1086 RHSKnownOne.trunc(SrcBitWidth);
1087 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedMask,
1088 RHSKnownZero, RHSKnownOne, Depth+1))
1090 DemandedMask.zext(BitWidth);
1091 RHSKnownZero.zext(BitWidth);
1092 RHSKnownOne.zext(BitWidth);
1093 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1094 // The top bits are known to be zero.
1095 RHSKnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
1098 case Instruction::SExt: {
1099 // Compute the bits in the result that are not present in the input.
1100 unsigned SrcBitWidth =I->getOperand(0)->getType()->getPrimitiveSizeInBits();
1102 APInt InputDemandedBits = DemandedMask &
1103 APInt::getLowBitsSet(BitWidth, SrcBitWidth);
1105 APInt NewBits(APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth));
1106 // If any of the sign extended bits are demanded, we know that the sign
1108 if ((NewBits & DemandedMask) != 0)
1109 InputDemandedBits.set(SrcBitWidth-1);
1111 InputDemandedBits.trunc(SrcBitWidth);
1112 RHSKnownZero.trunc(SrcBitWidth);
1113 RHSKnownOne.trunc(SrcBitWidth);
1114 if (SimplifyDemandedBits(I->getOperandUse(0), InputDemandedBits,
1115 RHSKnownZero, RHSKnownOne, Depth+1))
1117 InputDemandedBits.zext(BitWidth);
1118 RHSKnownZero.zext(BitWidth);
1119 RHSKnownOne.zext(BitWidth);
1120 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1122 // If the sign bit of the input is known set or clear, then we know the
1123 // top bits of the result.
1125 // If the input sign bit is known zero, or if the NewBits are not demanded
1126 // convert this into a zero extension.
1127 if (RHSKnownZero[SrcBitWidth-1] || (NewBits & ~DemandedMask) == NewBits) {
1128 // Convert to ZExt cast
1129 CastInst *NewCast = new ZExtInst(I->getOperand(0), VTy, I->getName());
1130 return InsertNewInstBefore(NewCast, *I);
1131 } else if (RHSKnownOne[SrcBitWidth-1]) { // Input sign bit known set
1132 RHSKnownOne |= NewBits;
1136 case Instruction::Add: {
1137 // Figure out what the input bits are. If the top bits of the and result
1138 // are not demanded, then the add doesn't demand them from its input
1140 unsigned NLZ = DemandedMask.countLeadingZeros();
1142 // If there is a constant on the RHS, there are a variety of xformations
1144 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
1145 // If null, this should be simplified elsewhere. Some of the xforms here
1146 // won't work if the RHS is zero.
1150 // If the top bit of the output is demanded, demand everything from the
1151 // input. Otherwise, we demand all the input bits except NLZ top bits.
1152 APInt InDemandedBits(APInt::getLowBitsSet(BitWidth, BitWidth - NLZ));
1154 // Find information about known zero/one bits in the input.
1155 if (SimplifyDemandedBits(I->getOperandUse(0), InDemandedBits,
1156 LHSKnownZero, LHSKnownOne, Depth+1))
1159 // If the RHS of the add has bits set that can't affect the input, reduce
1161 if (ShrinkDemandedConstant(I, 1, InDemandedBits))
1164 // Avoid excess work.
1165 if (LHSKnownZero == 0 && LHSKnownOne == 0)
1168 // Turn it into OR if input bits are zero.
1169 if ((LHSKnownZero & RHS->getValue()) == RHS->getValue()) {
1171 BinaryOperator::CreateOr(I->getOperand(0), I->getOperand(1),
1173 return InsertNewInstBefore(Or, *I);
1176 // We can say something about the output known-zero and known-one bits,
1177 // depending on potential carries from the input constant and the
1178 // unknowns. For example if the LHS is known to have at most the 0x0F0F0
1179 // bits set and the RHS constant is 0x01001, then we know we have a known
1180 // one mask of 0x00001 and a known zero mask of 0xE0F0E.
1182 // To compute this, we first compute the potential carry bits. These are
1183 // the bits which may be modified. I'm not aware of a better way to do
1185 const APInt &RHSVal = RHS->getValue();
1186 APInt CarryBits((~LHSKnownZero + RHSVal) ^ (~LHSKnownZero ^ RHSVal));
1188 // Now that we know which bits have carries, compute the known-1/0 sets.
1190 // Bits are known one if they are known zero in one operand and one in the
1191 // other, and there is no input carry.
1192 RHSKnownOne = ((LHSKnownZero & RHSVal) |
1193 (LHSKnownOne & ~RHSVal)) & ~CarryBits;
1195 // Bits are known zero if they are known zero in both operands and there
1196 // is no input carry.
1197 RHSKnownZero = LHSKnownZero & ~RHSVal & ~CarryBits;
1199 // If the high-bits of this ADD are not demanded, then it does not demand
1200 // the high bits of its LHS or RHS.
1201 if (DemandedMask[BitWidth-1] == 0) {
1202 // Right fill the mask of bits for this ADD to demand the most
1203 // significant bit and all those below it.
1204 APInt DemandedFromOps(APInt::getLowBitsSet(BitWidth, BitWidth-NLZ));
1205 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedFromOps,
1206 LHSKnownZero, LHSKnownOne, Depth+1) ||
1207 SimplifyDemandedBits(I->getOperandUse(1), DemandedFromOps,
1208 LHSKnownZero, LHSKnownOne, Depth+1))
1214 case Instruction::Sub:
1215 // If the high-bits of this SUB are not demanded, then it does not demand
1216 // the high bits of its LHS or RHS.
1217 if (DemandedMask[BitWidth-1] == 0) {
1218 // Right fill the mask of bits for this SUB to demand the most
1219 // significant bit and all those below it.
1220 uint32_t NLZ = DemandedMask.countLeadingZeros();
1221 APInt DemandedFromOps(APInt::getLowBitsSet(BitWidth, BitWidth-NLZ));
1222 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedFromOps,
1223 LHSKnownZero, LHSKnownOne, Depth+1) ||
1224 SimplifyDemandedBits(I->getOperandUse(1), DemandedFromOps,
1225 LHSKnownZero, LHSKnownOne, Depth+1))
1228 // Otherwise just hand the sub off to ComputeMaskedBits to fill in
1229 // the known zeros and ones.
1230 ComputeMaskedBits(V, DemandedMask, RHSKnownZero, RHSKnownOne, Depth);
1232 case Instruction::Shl:
1233 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1234 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
1235 APInt DemandedMaskIn(DemandedMask.lshr(ShiftAmt));
1236 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedMaskIn,
1237 RHSKnownZero, RHSKnownOne, Depth+1))
1239 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1240 RHSKnownZero <<= ShiftAmt;
1241 RHSKnownOne <<= ShiftAmt;
1242 // low bits known zero.
1244 RHSKnownZero |= APInt::getLowBitsSet(BitWidth, ShiftAmt);
1247 case Instruction::LShr:
1248 // For a logical shift right
1249 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1250 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
1252 // Unsigned shift right.
1253 APInt DemandedMaskIn(DemandedMask.shl(ShiftAmt));
1254 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedMaskIn,
1255 RHSKnownZero, RHSKnownOne, Depth+1))
1257 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1258 RHSKnownZero = APIntOps::lshr(RHSKnownZero, ShiftAmt);
1259 RHSKnownOne = APIntOps::lshr(RHSKnownOne, ShiftAmt);
1261 // Compute the new bits that are at the top now.
1262 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
1263 RHSKnownZero |= HighBits; // high bits known zero.
1267 case Instruction::AShr:
1268 // If this is an arithmetic shift right and only the low-bit is set, we can
1269 // always convert this into a logical shr, even if the shift amount is
1270 // variable. The low bit of the shift cannot be an input sign bit unless
1271 // the shift amount is >= the size of the datatype, which is undefined.
1272 if (DemandedMask == 1) {
1273 // Perform the logical shift right.
1274 Instruction *NewVal = BinaryOperator::CreateLShr(
1275 I->getOperand(0), I->getOperand(1), I->getName());
1276 return InsertNewInstBefore(NewVal, *I);
1279 // If the sign bit is the only bit demanded by this ashr, then there is no
1280 // need to do it, the shift doesn't change the high bit.
1281 if (DemandedMask.isSignBit())
1282 return I->getOperand(0);
1284 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1285 uint32_t ShiftAmt = SA->getLimitedValue(BitWidth);
1287 // Signed shift right.
1288 APInt DemandedMaskIn(DemandedMask.shl(ShiftAmt));
1289 // If any of the "high bits" are demanded, we should set the sign bit as
1291 if (DemandedMask.countLeadingZeros() <= ShiftAmt)
1292 DemandedMaskIn.set(BitWidth-1);
1293 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedMaskIn,
1294 RHSKnownZero, RHSKnownOne, Depth+1))
1296 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1297 // Compute the new bits that are at the top now.
1298 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
1299 RHSKnownZero = APIntOps::lshr(RHSKnownZero, ShiftAmt);
1300 RHSKnownOne = APIntOps::lshr(RHSKnownOne, ShiftAmt);
1302 // Handle the sign bits.
1303 APInt SignBit(APInt::getSignBit(BitWidth));
1304 // Adjust to where it is now in the mask.
1305 SignBit = APIntOps::lshr(SignBit, ShiftAmt);
1307 // If the input sign bit is known to be zero, or if none of the top bits
1308 // are demanded, turn this into an unsigned shift right.
1309 if (BitWidth <= ShiftAmt || RHSKnownZero[BitWidth-ShiftAmt-1] ||
1310 (HighBits & ~DemandedMask) == HighBits) {
1311 // Perform the logical shift right.
1312 Instruction *NewVal = BinaryOperator::CreateLShr(
1313 I->getOperand(0), SA, I->getName());
1314 return InsertNewInstBefore(NewVal, *I);
1315 } else if ((RHSKnownOne & SignBit) != 0) { // New bits are known one.
1316 RHSKnownOne |= HighBits;
1320 case Instruction::SRem:
1321 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
1322 APInt RA = Rem->getValue().abs();
1323 if (RA.isPowerOf2()) {
1324 if (DemandedMask.ule(RA)) // srem won't affect demanded bits
1325 return I->getOperand(0);
1327 APInt LowBits = RA - 1;
1328 APInt Mask2 = LowBits | APInt::getSignBit(BitWidth);
1329 if (SimplifyDemandedBits(I->getOperandUse(0), Mask2,
1330 LHSKnownZero, LHSKnownOne, Depth+1))
1333 if (LHSKnownZero[BitWidth-1] || ((LHSKnownZero & LowBits) == LowBits))
1334 LHSKnownZero |= ~LowBits;
1336 KnownZero |= LHSKnownZero & DemandedMask;
1338 assert(!(KnownZero & KnownOne) && "Bits known to be one AND zero?");
1342 case Instruction::URem: {
1343 APInt KnownZero2(BitWidth, 0), KnownOne2(BitWidth, 0);
1344 APInt AllOnes = APInt::getAllOnesValue(BitWidth);
1345 if (SimplifyDemandedBits(I->getOperandUse(0), AllOnes,
1346 KnownZero2, KnownOne2, Depth+1) ||
1347 SimplifyDemandedBits(I->getOperandUse(1), AllOnes,
1348 KnownZero2, KnownOne2, Depth+1))
1351 unsigned Leaders = KnownZero2.countLeadingOnes();
1352 Leaders = std::max(Leaders,
1353 KnownZero2.countLeadingOnes());
1354 KnownZero = APInt::getHighBitsSet(BitWidth, Leaders) & DemandedMask;
1357 case Instruction::Call:
1358 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
1359 switch (II->getIntrinsicID()) {
1361 case Intrinsic::bswap: {
1362 // If the only bits demanded come from one byte of the bswap result,
1363 // just shift the input byte into position to eliminate the bswap.
1364 unsigned NLZ = DemandedMask.countLeadingZeros();
1365 unsigned NTZ = DemandedMask.countTrailingZeros();
1367 // Round NTZ down to the next byte. If we have 11 trailing zeros, then
1368 // we need all the bits down to bit 8. Likewise, round NLZ. If we
1369 // have 14 leading zeros, round to 8.
1372 // If we need exactly one byte, we can do this transformation.
1373 if (BitWidth-NLZ-NTZ == 8) {
1374 unsigned ResultBit = NTZ;
1375 unsigned InputBit = BitWidth-NTZ-8;
1377 // Replace this with either a left or right shift to get the byte into
1379 Instruction *NewVal;
1380 if (InputBit > ResultBit)
1381 NewVal = BinaryOperator::CreateLShr(I->getOperand(1),
1382 ConstantInt::get(I->getType(), InputBit-ResultBit));
1384 NewVal = BinaryOperator::CreateShl(I->getOperand(1),
1385 ConstantInt::get(I->getType(), ResultBit-InputBit));
1386 NewVal->takeName(I);
1387 return InsertNewInstBefore(NewVal, *I);
1390 // TODO: Could compute known zero/one bits based on the input.
1395 ComputeMaskedBits(V, DemandedMask, RHSKnownZero, RHSKnownOne, Depth);
1399 // If the client is only demanding bits that we know, return the known
1401 if ((DemandedMask & (RHSKnownZero|RHSKnownOne)) == DemandedMask)
1402 return ConstantInt::get(RHSKnownOne);
1407 /// SimplifyDemandedVectorElts - The specified value produces a vector with
1408 /// any number of elements. DemandedElts contains the set of elements that are
1409 /// actually used by the caller. This method analyzes which elements of the
1410 /// operand are undef and returns that information in UndefElts.
1412 /// If the information about demanded elements can be used to simplify the
1413 /// operation, the operation is simplified, then the resultant value is
1414 /// returned. This returns null if no change was made.
1415 Value *InstCombiner::SimplifyDemandedVectorElts(Value *V, APInt DemandedElts,
1418 unsigned VWidth = cast<VectorType>(V->getType())->getNumElements();
1419 APInt EltMask(APInt::getAllOnesValue(VWidth));
1420 assert((DemandedElts & ~EltMask) == 0 && "Invalid DemandedElts!");
1422 if (isa<UndefValue>(V)) {
1423 // If the entire vector is undefined, just return this info.
1424 UndefElts = EltMask;
1426 } else if (DemandedElts == 0) { // If nothing is demanded, provide undef.
1427 UndefElts = EltMask;
1428 return UndefValue::get(V->getType());
1432 if (ConstantVector *CP = dyn_cast<ConstantVector>(V)) {
1433 const Type *EltTy = cast<VectorType>(V->getType())->getElementType();
1434 Constant *Undef = UndefValue::get(EltTy);
1436 std::vector<Constant*> Elts;
1437 for (unsigned i = 0; i != VWidth; ++i)
1438 if (!DemandedElts[i]) { // If not demanded, set to undef.
1439 Elts.push_back(Undef);
1441 } else if (isa<UndefValue>(CP->getOperand(i))) { // Already undef.
1442 Elts.push_back(Undef);
1444 } else { // Otherwise, defined.
1445 Elts.push_back(CP->getOperand(i));
1448 // If we changed the constant, return it.
1449 Constant *NewCP = ConstantVector::get(Elts);
1450 return NewCP != CP ? NewCP : 0;
1451 } else if (isa<ConstantAggregateZero>(V)) {
1452 // Simplify the CAZ to a ConstantVector where the non-demanded elements are
1455 // Check if this is identity. If so, return 0 since we are not simplifying
1457 if (DemandedElts == ((1ULL << VWidth) -1))
1460 const Type *EltTy = cast<VectorType>(V->getType())->getElementType();
1461 Constant *Zero = Constant::getNullValue(EltTy);
1462 Constant *Undef = UndefValue::get(EltTy);
1463 std::vector<Constant*> Elts;
1464 for (unsigned i = 0; i != VWidth; ++i) {
1465 Constant *Elt = DemandedElts[i] ? Zero : Undef;
1466 Elts.push_back(Elt);
1468 UndefElts = DemandedElts ^ EltMask;
1469 return ConstantVector::get(Elts);
1472 // Limit search depth.
1476 // If multiple users are using the root value, procede with
1477 // simplification conservatively assuming that all elements
1479 if (!V->hasOneUse()) {
1480 // Quit if we find multiple users of a non-root value though.
1481 // They'll be handled when it's their turn to be visited by
1482 // the main instcombine process.
1484 // TODO: Just compute the UndefElts information recursively.
1487 // Conservatively assume that all elements are needed.
1488 DemandedElts = EltMask;
1491 Instruction *I = dyn_cast<Instruction>(V);
1492 if (!I) return false; // Only analyze instructions.
1494 bool MadeChange = false;
1495 APInt UndefElts2(VWidth, 0);
1497 switch (I->getOpcode()) {
1500 case Instruction::InsertElement: {
1501 // If this is a variable index, we don't know which element it overwrites.
1502 // demand exactly the same input as we produce.
1503 ConstantInt *Idx = dyn_cast<ConstantInt>(I->getOperand(2));
1505 // Note that we can't propagate undef elt info, because we don't know
1506 // which elt is getting updated.
1507 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
1508 UndefElts2, Depth+1);
1509 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1513 // If this is inserting an element that isn't demanded, remove this
1515 unsigned IdxNo = Idx->getZExtValue();
1516 if (IdxNo >= VWidth || !DemandedElts[IdxNo])
1517 return AddSoonDeadInstToWorklist(*I, 0);
1519 // Otherwise, the element inserted overwrites whatever was there, so the
1520 // input demanded set is simpler than the output set.
1521 APInt DemandedElts2 = DemandedElts;
1522 DemandedElts2.clear(IdxNo);
1523 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts2,
1524 UndefElts, Depth+1);
1525 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1527 // The inserted element is defined.
1528 UndefElts.clear(IdxNo);
1531 case Instruction::ShuffleVector: {
1532 ShuffleVectorInst *Shuffle = cast<ShuffleVectorInst>(I);
1533 uint64_t LHSVWidth =
1534 cast<VectorType>(Shuffle->getOperand(0)->getType())->getNumElements();
1535 APInt LeftDemanded(LHSVWidth, 0), RightDemanded(LHSVWidth, 0);
1536 for (unsigned i = 0; i < VWidth; i++) {
1537 if (DemandedElts[i]) {
1538 unsigned MaskVal = Shuffle->getMaskValue(i);
1539 if (MaskVal != -1u) {
1540 assert(MaskVal < LHSVWidth * 2 &&
1541 "shufflevector mask index out of range!");
1542 if (MaskVal < LHSVWidth)
1543 LeftDemanded.set(MaskVal);
1545 RightDemanded.set(MaskVal - LHSVWidth);
1550 APInt UndefElts4(LHSVWidth, 0);
1551 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), LeftDemanded,
1552 UndefElts4, Depth+1);
1553 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1555 APInt UndefElts3(LHSVWidth, 0);
1556 TmpV = SimplifyDemandedVectorElts(I->getOperand(1), RightDemanded,
1557 UndefElts3, Depth+1);
1558 if (TmpV) { I->setOperand(1, TmpV); MadeChange = true; }
1560 bool NewUndefElts = false;
1561 for (unsigned i = 0; i < VWidth; i++) {
1562 unsigned MaskVal = Shuffle->getMaskValue(i);
1563 if (MaskVal == -1u) {
1565 } else if (MaskVal < LHSVWidth) {
1566 if (UndefElts4[MaskVal]) {
1567 NewUndefElts = true;
1571 if (UndefElts3[MaskVal - LHSVWidth]) {
1572 NewUndefElts = true;
1579 // Add additional discovered undefs.
1580 std::vector<Constant*> Elts;
1581 for (unsigned i = 0; i < VWidth; ++i) {
1583 Elts.push_back(UndefValue::get(Type::Int32Ty));
1585 Elts.push_back(ConstantInt::get(Type::Int32Ty,
1586 Shuffle->getMaskValue(i)));
1588 I->setOperand(2, ConstantVector::get(Elts));
1593 case Instruction::BitCast: {
1594 // Vector->vector casts only.
1595 const VectorType *VTy = dyn_cast<VectorType>(I->getOperand(0)->getType());
1597 unsigned InVWidth = VTy->getNumElements();
1598 APInt InputDemandedElts(InVWidth, 0);
1601 if (VWidth == InVWidth) {
1602 // If we are converting from <4 x i32> -> <4 x f32>, we demand the same
1603 // elements as are demanded of us.
1605 InputDemandedElts = DemandedElts;
1606 } else if (VWidth > InVWidth) {
1610 // If there are more elements in the result than there are in the source,
1611 // then an input element is live if any of the corresponding output
1612 // elements are live.
1613 Ratio = VWidth/InVWidth;
1614 for (unsigned OutIdx = 0; OutIdx != VWidth; ++OutIdx) {
1615 if (DemandedElts[OutIdx])
1616 InputDemandedElts.set(OutIdx/Ratio);
1622 // If there are more elements in the source than there are in the result,
1623 // then an input element is live if the corresponding output element is
1625 Ratio = InVWidth/VWidth;
1626 for (unsigned InIdx = 0; InIdx != InVWidth; ++InIdx)
1627 if (DemandedElts[InIdx/Ratio])
1628 InputDemandedElts.set(InIdx);
1631 // div/rem demand all inputs, because they don't want divide by zero.
1632 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), InputDemandedElts,
1633 UndefElts2, Depth+1);
1635 I->setOperand(0, TmpV);
1639 UndefElts = UndefElts2;
1640 if (VWidth > InVWidth) {
1641 assert(0 && "Unimp");
1642 // If there are more elements in the result than there are in the source,
1643 // then an output element is undef if the corresponding input element is
1645 for (unsigned OutIdx = 0; OutIdx != VWidth; ++OutIdx)
1646 if (UndefElts2[OutIdx/Ratio])
1647 UndefElts.set(OutIdx);
1648 } else if (VWidth < InVWidth) {
1649 assert(0 && "Unimp");
1650 // If there are more elements in the source than there are in the result,
1651 // then a result element is undef if all of the corresponding input
1652 // elements are undef.
1653 UndefElts = ~0ULL >> (64-VWidth); // Start out all undef.
1654 for (unsigned InIdx = 0; InIdx != InVWidth; ++InIdx)
1655 if (!UndefElts2[InIdx]) // Not undef?
1656 UndefElts.clear(InIdx/Ratio); // Clear undef bit.
1660 case Instruction::And:
1661 case Instruction::Or:
1662 case Instruction::Xor:
1663 case Instruction::Add:
1664 case Instruction::Sub:
1665 case Instruction::Mul:
1666 // div/rem demand all inputs, because they don't want divide by zero.
1667 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
1668 UndefElts, Depth+1);
1669 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1670 TmpV = SimplifyDemandedVectorElts(I->getOperand(1), DemandedElts,
1671 UndefElts2, Depth+1);
1672 if (TmpV) { I->setOperand(1, TmpV); MadeChange = true; }
1674 // Output elements are undefined if both are undefined. Consider things
1675 // like undef&0. The result is known zero, not undef.
1676 UndefElts &= UndefElts2;
1679 case Instruction::Call: {
1680 IntrinsicInst *II = dyn_cast<IntrinsicInst>(I);
1682 switch (II->getIntrinsicID()) {
1685 // Binary vector operations that work column-wise. A dest element is a
1686 // function of the corresponding input elements from the two inputs.
1687 case Intrinsic::x86_sse_sub_ss:
1688 case Intrinsic::x86_sse_mul_ss:
1689 case Intrinsic::x86_sse_min_ss:
1690 case Intrinsic::x86_sse_max_ss:
1691 case Intrinsic::x86_sse2_sub_sd:
1692 case Intrinsic::x86_sse2_mul_sd:
1693 case Intrinsic::x86_sse2_min_sd:
1694 case Intrinsic::x86_sse2_max_sd:
1695 TmpV = SimplifyDemandedVectorElts(II->getOperand(1), DemandedElts,
1696 UndefElts, Depth+1);
1697 if (TmpV) { II->setOperand(1, TmpV); MadeChange = true; }
1698 TmpV = SimplifyDemandedVectorElts(II->getOperand(2), DemandedElts,
1699 UndefElts2, Depth+1);
1700 if (TmpV) { II->setOperand(2, TmpV); MadeChange = true; }
1702 // If only the low elt is demanded and this is a scalarizable intrinsic,
1703 // scalarize it now.
1704 if (DemandedElts == 1) {
1705 switch (II->getIntrinsicID()) {
1707 case Intrinsic::x86_sse_sub_ss:
1708 case Intrinsic::x86_sse_mul_ss:
1709 case Intrinsic::x86_sse2_sub_sd:
1710 case Intrinsic::x86_sse2_mul_sd:
1711 // TODO: Lower MIN/MAX/ABS/etc
1712 Value *LHS = II->getOperand(1);
1713 Value *RHS = II->getOperand(2);
1714 // Extract the element as scalars.
1715 LHS = InsertNewInstBefore(new ExtractElementInst(LHS, 0U,"tmp"), *II);
1716 RHS = InsertNewInstBefore(new ExtractElementInst(RHS, 0U,"tmp"), *II);
1718 switch (II->getIntrinsicID()) {
1719 default: assert(0 && "Case stmts out of sync!");
1720 case Intrinsic::x86_sse_sub_ss:
1721 case Intrinsic::x86_sse2_sub_sd:
1722 TmpV = InsertNewInstBefore(BinaryOperator::CreateSub(LHS, RHS,
1723 II->getName()), *II);
1725 case Intrinsic::x86_sse_mul_ss:
1726 case Intrinsic::x86_sse2_mul_sd:
1727 TmpV = InsertNewInstBefore(BinaryOperator::CreateMul(LHS, RHS,
1728 II->getName()), *II);
1733 InsertElementInst::Create(UndefValue::get(II->getType()), TmpV, 0U,
1735 InsertNewInstBefore(New, *II);
1736 AddSoonDeadInstToWorklist(*II, 0);
1741 // Output elements are undefined if both are undefined. Consider things
1742 // like undef&0. The result is known zero, not undef.
1743 UndefElts &= UndefElts2;
1749 return MadeChange ? I : 0;
1753 /// AssociativeOpt - Perform an optimization on an associative operator. This
1754 /// function is designed to check a chain of associative operators for a
1755 /// potential to apply a certain optimization. Since the optimization may be
1756 /// applicable if the expression was reassociated, this checks the chain, then
1757 /// reassociates the expression as necessary to expose the optimization
1758 /// opportunity. This makes use of a special Functor, which must define
1759 /// 'shouldApply' and 'apply' methods.
1761 template<typename Functor>
1762 static Instruction *AssociativeOpt(BinaryOperator &Root, const Functor &F) {
1763 unsigned Opcode = Root.getOpcode();
1764 Value *LHS = Root.getOperand(0);
1766 // Quick check, see if the immediate LHS matches...
1767 if (F.shouldApply(LHS))
1768 return F.apply(Root);
1770 // Otherwise, if the LHS is not of the same opcode as the root, return.
1771 Instruction *LHSI = dyn_cast<Instruction>(LHS);
1772 while (LHSI && LHSI->getOpcode() == Opcode && LHSI->hasOneUse()) {
1773 // Should we apply this transform to the RHS?
1774 bool ShouldApply = F.shouldApply(LHSI->getOperand(1));
1776 // If not to the RHS, check to see if we should apply to the LHS...
1777 if (!ShouldApply && F.shouldApply(LHSI->getOperand(0))) {
1778 cast<BinaryOperator>(LHSI)->swapOperands(); // Make the LHS the RHS
1782 // If the functor wants to apply the optimization to the RHS of LHSI,
1783 // reassociate the expression from ((? op A) op B) to (? op (A op B))
1785 // Now all of the instructions are in the current basic block, go ahead
1786 // and perform the reassociation.
1787 Instruction *TmpLHSI = cast<Instruction>(Root.getOperand(0));
1789 // First move the selected RHS to the LHS of the root...
1790 Root.setOperand(0, LHSI->getOperand(1));
1792 // Make what used to be the LHS of the root be the user of the root...
1793 Value *ExtraOperand = TmpLHSI->getOperand(1);
1794 if (&Root == TmpLHSI) {
1795 Root.replaceAllUsesWith(Constant::getNullValue(TmpLHSI->getType()));
1798 Root.replaceAllUsesWith(TmpLHSI); // Users now use TmpLHSI
1799 TmpLHSI->setOperand(1, &Root); // TmpLHSI now uses the root
1800 BasicBlock::iterator ARI = &Root; ++ARI;
1801 TmpLHSI->moveBefore(ARI); // Move TmpLHSI to after Root
1804 // Now propagate the ExtraOperand down the chain of instructions until we
1806 while (TmpLHSI != LHSI) {
1807 Instruction *NextLHSI = cast<Instruction>(TmpLHSI->getOperand(0));
1808 // Move the instruction to immediately before the chain we are
1809 // constructing to avoid breaking dominance properties.
1810 NextLHSI->moveBefore(ARI);
1813 Value *NextOp = NextLHSI->getOperand(1);
1814 NextLHSI->setOperand(1, ExtraOperand);
1816 ExtraOperand = NextOp;
1819 // Now that the instructions are reassociated, have the functor perform
1820 // the transformation...
1821 return F.apply(Root);
1824 LHSI = dyn_cast<Instruction>(LHSI->getOperand(0));
1831 // AddRHS - Implements: X + X --> X << 1
1834 AddRHS(Value *rhs) : RHS(rhs) {}
1835 bool shouldApply(Value *LHS) const { return LHS == RHS; }
1836 Instruction *apply(BinaryOperator &Add) const {
1837 return BinaryOperator::CreateShl(Add.getOperand(0),
1838 ConstantInt::get(Add.getType(), 1));
1842 // AddMaskingAnd - Implements (A & C1)+(B & C2) --> (A & C1)|(B & C2)
1844 struct AddMaskingAnd {
1846 AddMaskingAnd(Constant *c) : C2(c) {}
1847 bool shouldApply(Value *LHS) const {
1849 return match(LHS, m_And(m_Value(), m_ConstantInt(C1))) &&
1850 ConstantExpr::getAnd(C1, C2)->isNullValue();
1852 Instruction *apply(BinaryOperator &Add) const {
1853 return BinaryOperator::CreateOr(Add.getOperand(0), Add.getOperand(1));
1859 static Value *FoldOperationIntoSelectOperand(Instruction &I, Value *SO,
1861 if (CastInst *CI = dyn_cast<CastInst>(&I)) {
1862 return IC->InsertCastBefore(CI->getOpcode(), SO, I.getType(), I);
1865 // Figure out if the constant is the left or the right argument.
1866 bool ConstIsRHS = isa<Constant>(I.getOperand(1));
1867 Constant *ConstOperand = cast<Constant>(I.getOperand(ConstIsRHS));
1869 if (Constant *SOC = dyn_cast<Constant>(SO)) {
1871 return ConstantExpr::get(I.getOpcode(), SOC, ConstOperand);
1872 return ConstantExpr::get(I.getOpcode(), ConstOperand, SOC);
1875 Value *Op0 = SO, *Op1 = ConstOperand;
1877 std::swap(Op0, Op1);
1879 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
1880 New = BinaryOperator::Create(BO->getOpcode(), Op0, Op1,SO->getName()+".op");
1881 else if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1882 New = CmpInst::Create(CI->getOpcode(), CI->getPredicate(), Op0, Op1,
1883 SO->getName()+".cmp");
1885 assert(0 && "Unknown binary instruction type!");
1888 return IC->InsertNewInstBefore(New, I);
1891 // FoldOpIntoSelect - Given an instruction with a select as one operand and a
1892 // constant as the other operand, try to fold the binary operator into the
1893 // select arguments. This also works for Cast instructions, which obviously do
1894 // not have a second operand.
1895 static Instruction *FoldOpIntoSelect(Instruction &Op, SelectInst *SI,
1897 // Don't modify shared select instructions
1898 if (!SI->hasOneUse()) return 0;
1899 Value *TV = SI->getOperand(1);
1900 Value *FV = SI->getOperand(2);
1902 if (isa<Constant>(TV) || isa<Constant>(FV)) {
1903 // Bool selects with constant operands can be folded to logical ops.
1904 if (SI->getType() == Type::Int1Ty) return 0;
1906 Value *SelectTrueVal = FoldOperationIntoSelectOperand(Op, TV, IC);
1907 Value *SelectFalseVal = FoldOperationIntoSelectOperand(Op, FV, IC);
1909 return SelectInst::Create(SI->getCondition(), SelectTrueVal,
1916 /// FoldOpIntoPhi - Given a binary operator or cast instruction which has a PHI
1917 /// node as operand #0, see if we can fold the instruction into the PHI (which
1918 /// is only possible if all operands to the PHI are constants).
1919 Instruction *InstCombiner::FoldOpIntoPhi(Instruction &I) {
1920 PHINode *PN = cast<PHINode>(I.getOperand(0));
1921 unsigned NumPHIValues = PN->getNumIncomingValues();
1922 if (!PN->hasOneUse() || NumPHIValues == 0) return 0;
1924 // Check to see if all of the operands of the PHI are constants. If there is
1925 // one non-constant value, remember the BB it is. If there is more than one
1926 // or if *it* is a PHI, bail out.
1927 BasicBlock *NonConstBB = 0;
1928 for (unsigned i = 0; i != NumPHIValues; ++i)
1929 if (!isa<Constant>(PN->getIncomingValue(i))) {
1930 if (NonConstBB) return 0; // More than one non-const value.
1931 if (isa<PHINode>(PN->getIncomingValue(i))) return 0; // Itself a phi.
1932 NonConstBB = PN->getIncomingBlock(i);
1934 // If the incoming non-constant value is in I's block, we have an infinite
1936 if (NonConstBB == I.getParent())
1940 // If there is exactly one non-constant value, we can insert a copy of the
1941 // operation in that block. However, if this is a critical edge, we would be
1942 // inserting the computation one some other paths (e.g. inside a loop). Only
1943 // do this if the pred block is unconditionally branching into the phi block.
1945 BranchInst *BI = dyn_cast<BranchInst>(NonConstBB->getTerminator());
1946 if (!BI || !BI->isUnconditional()) return 0;
1949 // Okay, we can do the transformation: create the new PHI node.
1950 PHINode *NewPN = PHINode::Create(I.getType(), "");
1951 NewPN->reserveOperandSpace(PN->getNumOperands()/2);
1952 InsertNewInstBefore(NewPN, *PN);
1953 NewPN->takeName(PN);
1955 // Next, add all of the operands to the PHI.
1956 if (I.getNumOperands() == 2) {
1957 Constant *C = cast<Constant>(I.getOperand(1));
1958 for (unsigned i = 0; i != NumPHIValues; ++i) {
1960 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
1961 if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1962 InV = ConstantExpr::getCompare(CI->getPredicate(), InC, C);
1964 InV = ConstantExpr::get(I.getOpcode(), InC, C);
1966 assert(PN->getIncomingBlock(i) == NonConstBB);
1967 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
1968 InV = BinaryOperator::Create(BO->getOpcode(),
1969 PN->getIncomingValue(i), C, "phitmp",
1970 NonConstBB->getTerminator());
1971 else if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1972 InV = CmpInst::Create(CI->getOpcode(),
1974 PN->getIncomingValue(i), C, "phitmp",
1975 NonConstBB->getTerminator());
1977 assert(0 && "Unknown binop!");
1979 AddToWorkList(cast<Instruction>(InV));
1981 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
1984 CastInst *CI = cast<CastInst>(&I);
1985 const Type *RetTy = CI->getType();
1986 for (unsigned i = 0; i != NumPHIValues; ++i) {
1988 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
1989 InV = ConstantExpr::getCast(CI->getOpcode(), InC, RetTy);
1991 assert(PN->getIncomingBlock(i) == NonConstBB);
1992 InV = CastInst::Create(CI->getOpcode(), PN->getIncomingValue(i),
1993 I.getType(), "phitmp",
1994 NonConstBB->getTerminator());
1995 AddToWorkList(cast<Instruction>(InV));
1997 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
2000 return ReplaceInstUsesWith(I, NewPN);
2004 /// WillNotOverflowSignedAdd - Return true if we can prove that:
2005 /// (sext (add LHS, RHS)) === (add (sext LHS), (sext RHS))
2006 /// This basically requires proving that the add in the original type would not
2007 /// overflow to change the sign bit or have a carry out.
2008 bool InstCombiner::WillNotOverflowSignedAdd(Value *LHS, Value *RHS) {
2009 // There are different heuristics we can use for this. Here are some simple
2012 // Add has the property that adding any two 2's complement numbers can only
2013 // have one carry bit which can change a sign. As such, if LHS and RHS each
2014 // have at least two sign bits, we know that the addition of the two values will
2015 // sign extend fine.
2016 if (ComputeNumSignBits(LHS) > 1 && ComputeNumSignBits(RHS) > 1)
2020 // If one of the operands only has one non-zero bit, and if the other operand
2021 // has a known-zero bit in a more significant place than it (not including the
2022 // sign bit) the ripple may go up to and fill the zero, but won't change the
2023 // sign. For example, (X & ~4) + 1.
2031 Instruction *InstCombiner::visitAdd(BinaryOperator &I) {
2032 bool Changed = SimplifyCommutative(I);
2033 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
2035 if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
2036 // X + undef -> undef
2037 if (isa<UndefValue>(RHS))
2038 return ReplaceInstUsesWith(I, RHS);
2041 if (!I.getType()->isFPOrFPVector()) { // NOTE: -0 + +0 = +0.
2042 if (RHSC->isNullValue())
2043 return ReplaceInstUsesWith(I, LHS);
2044 } else if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHSC)) {
2045 if (CFP->isExactlyValue(ConstantFP::getNegativeZero
2046 (I.getType())->getValueAPF()))
2047 return ReplaceInstUsesWith(I, LHS);
2050 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHSC)) {
2051 // X + (signbit) --> X ^ signbit
2052 const APInt& Val = CI->getValue();
2053 uint32_t BitWidth = Val.getBitWidth();
2054 if (Val == APInt::getSignBit(BitWidth))
2055 return BinaryOperator::CreateXor(LHS, RHS);
2057 // See if SimplifyDemandedBits can simplify this. This handles stuff like
2058 // (X & 254)+1 -> (X&254)|1
2059 if (!isa<VectorType>(I.getType()) && SimplifyDemandedInstructionBits(I))
2062 // zext(i1) - 1 -> select i1, 0, -1
2063 if (ZExtInst *ZI = dyn_cast<ZExtInst>(LHS))
2064 if (CI->isAllOnesValue() &&
2065 ZI->getOperand(0)->getType() == Type::Int1Ty)
2066 return SelectInst::Create(ZI->getOperand(0),
2067 Constant::getNullValue(I.getType()),
2068 ConstantInt::getAllOnesValue(I.getType()));
2071 if (isa<PHINode>(LHS))
2072 if (Instruction *NV = FoldOpIntoPhi(I))
2075 ConstantInt *XorRHS = 0;
2077 if (isa<ConstantInt>(RHSC) &&
2078 match(LHS, m_Xor(m_Value(XorLHS), m_ConstantInt(XorRHS)))) {
2079 uint32_t TySizeBits = I.getType()->getPrimitiveSizeInBits();
2080 const APInt& RHSVal = cast<ConstantInt>(RHSC)->getValue();
2082 uint32_t Size = TySizeBits / 2;
2083 APInt C0080Val(APInt(TySizeBits, 1ULL).shl(Size - 1));
2084 APInt CFF80Val(-C0080Val);
2086 if (TySizeBits > Size) {
2087 // If we have ADD(XOR(AND(X, 0xFF), 0x80), 0xF..F80), it's a sext.
2088 // If we have ADD(XOR(AND(X, 0xFF), 0xF..F80), 0x80), it's a sext.
2089 if ((RHSVal == CFF80Val && XorRHS->getValue() == C0080Val) ||
2090 (RHSVal == C0080Val && XorRHS->getValue() == CFF80Val)) {
2091 // This is a sign extend if the top bits are known zero.
2092 if (!MaskedValueIsZero(XorLHS,
2093 APInt::getHighBitsSet(TySizeBits, TySizeBits - Size)))
2094 Size = 0; // Not a sign ext, but can't be any others either.
2099 C0080Val = APIntOps::lshr(C0080Val, Size);
2100 CFF80Val = APIntOps::ashr(CFF80Val, Size);
2101 } while (Size >= 1);
2103 // FIXME: This shouldn't be necessary. When the backends can handle types
2104 // with funny bit widths then this switch statement should be removed. It
2105 // is just here to get the size of the "middle" type back up to something
2106 // that the back ends can handle.
2107 const Type *MiddleType = 0;
2110 case 32: MiddleType = Type::Int32Ty; break;
2111 case 16: MiddleType = Type::Int16Ty; break;
2112 case 8: MiddleType = Type::Int8Ty; break;
2115 Instruction *NewTrunc = new TruncInst(XorLHS, MiddleType, "sext");
2116 InsertNewInstBefore(NewTrunc, I);
2117 return new SExtInst(NewTrunc, I.getType(), I.getName());
2122 if (I.getType() == Type::Int1Ty)
2123 return BinaryOperator::CreateXor(LHS, RHS);
2126 if (I.getType()->isInteger()) {
2127 if (Instruction *Result = AssociativeOpt(I, AddRHS(RHS))) return Result;
2129 if (Instruction *RHSI = dyn_cast<Instruction>(RHS)) {
2130 if (RHSI->getOpcode() == Instruction::Sub)
2131 if (LHS == RHSI->getOperand(1)) // A + (B - A) --> B
2132 return ReplaceInstUsesWith(I, RHSI->getOperand(0));
2134 if (Instruction *LHSI = dyn_cast<Instruction>(LHS)) {
2135 if (LHSI->getOpcode() == Instruction::Sub)
2136 if (RHS == LHSI->getOperand(1)) // (B - A) + A --> B
2137 return ReplaceInstUsesWith(I, LHSI->getOperand(0));
2142 // -A + -B --> -(A + B)
2143 if (Value *LHSV = dyn_castNegVal(LHS)) {
2144 if (LHS->getType()->isIntOrIntVector()) {
2145 if (Value *RHSV = dyn_castNegVal(RHS)) {
2146 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSV, RHSV, "sum");
2147 InsertNewInstBefore(NewAdd, I);
2148 return BinaryOperator::CreateNeg(NewAdd);
2152 return BinaryOperator::CreateSub(RHS, LHSV);
2156 if (!isa<Constant>(RHS))
2157 if (Value *V = dyn_castNegVal(RHS))
2158 return BinaryOperator::CreateSub(LHS, V);
2162 if (Value *X = dyn_castFoldableMul(LHS, C2)) {
2163 if (X == RHS) // X*C + X --> X * (C+1)
2164 return BinaryOperator::CreateMul(RHS, AddOne(C2));
2166 // X*C1 + X*C2 --> X * (C1+C2)
2168 if (X == dyn_castFoldableMul(RHS, C1))
2169 return BinaryOperator::CreateMul(X, Add(C1, C2));
2172 // X + X*C --> X * (C+1)
2173 if (dyn_castFoldableMul(RHS, C2) == LHS)
2174 return BinaryOperator::CreateMul(LHS, AddOne(C2));
2176 // X + ~X --> -1 since ~X = -X-1
2177 if (dyn_castNotVal(LHS) == RHS || dyn_castNotVal(RHS) == LHS)
2178 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
2181 // (A & C1)+(B & C2) --> (A & C1)|(B & C2) iff C1&C2 == 0
2182 if (match(RHS, m_And(m_Value(), m_ConstantInt(C2))))
2183 if (Instruction *R = AssociativeOpt(I, AddMaskingAnd(C2)))
2186 // A+B --> A|B iff A and B have no bits set in common.
2187 if (const IntegerType *IT = dyn_cast<IntegerType>(I.getType())) {
2188 APInt Mask = APInt::getAllOnesValue(IT->getBitWidth());
2189 APInt LHSKnownOne(IT->getBitWidth(), 0);
2190 APInt LHSKnownZero(IT->getBitWidth(), 0);
2191 ComputeMaskedBits(LHS, Mask, LHSKnownZero, LHSKnownOne);
2192 if (LHSKnownZero != 0) {
2193 APInt RHSKnownOne(IT->getBitWidth(), 0);
2194 APInt RHSKnownZero(IT->getBitWidth(), 0);
2195 ComputeMaskedBits(RHS, Mask, RHSKnownZero, RHSKnownOne);
2197 // No bits in common -> bitwise or.
2198 if ((LHSKnownZero|RHSKnownZero).isAllOnesValue())
2199 return BinaryOperator::CreateOr(LHS, RHS);
2203 // W*X + Y*Z --> W * (X+Z) iff W == Y
2204 if (I.getType()->isIntOrIntVector()) {
2205 Value *W, *X, *Y, *Z;
2206 if (match(LHS, m_Mul(m_Value(W), m_Value(X))) &&
2207 match(RHS, m_Mul(m_Value(Y), m_Value(Z)))) {
2211 } else if (Y == X) {
2213 } else if (X == Z) {
2220 Value *NewAdd = InsertNewInstBefore(BinaryOperator::CreateAdd(X, Z,
2221 LHS->getName()), I);
2222 return BinaryOperator::CreateMul(W, NewAdd);
2227 if (ConstantInt *CRHS = dyn_cast<ConstantInt>(RHS)) {
2229 if (match(LHS, m_Not(m_Value(X)))) // ~X + C --> (C-1) - X
2230 return BinaryOperator::CreateSub(SubOne(CRHS), X);
2232 // (X & FF00) + xx00 -> (X+xx00) & FF00
2233 if (LHS->hasOneUse() && match(LHS, m_And(m_Value(X), m_ConstantInt(C2)))) {
2234 Constant *Anded = And(CRHS, C2);
2235 if (Anded == CRHS) {
2236 // See if all bits from the first bit set in the Add RHS up are included
2237 // in the mask. First, get the rightmost bit.
2238 const APInt& AddRHSV = CRHS->getValue();
2240 // Form a mask of all bits from the lowest bit added through the top.
2241 APInt AddRHSHighBits(~((AddRHSV & -AddRHSV)-1));
2243 // See if the and mask includes all of these bits.
2244 APInt AddRHSHighBitsAnd(AddRHSHighBits & C2->getValue());
2246 if (AddRHSHighBits == AddRHSHighBitsAnd) {
2247 // Okay, the xform is safe. Insert the new add pronto.
2248 Value *NewAdd = InsertNewInstBefore(BinaryOperator::CreateAdd(X, CRHS,
2249 LHS->getName()), I);
2250 return BinaryOperator::CreateAnd(NewAdd, C2);
2255 // Try to fold constant add into select arguments.
2256 if (SelectInst *SI = dyn_cast<SelectInst>(LHS))
2257 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2261 // add (cast *A to intptrtype) B ->
2262 // cast (GEP (cast *A to sbyte*) B) --> intptrtype
2264 CastInst *CI = dyn_cast<CastInst>(LHS);
2267 CI = dyn_cast<CastInst>(RHS);
2270 if (CI && CI->getType()->isSized() &&
2271 (CI->getType()->getPrimitiveSizeInBits() ==
2272 TD->getIntPtrType()->getPrimitiveSizeInBits())
2273 && isa<PointerType>(CI->getOperand(0)->getType())) {
2275 cast<PointerType>(CI->getOperand(0)->getType())->getAddressSpace();
2276 Value *I2 = InsertBitCastBefore(CI->getOperand(0),
2277 PointerType::get(Type::Int8Ty, AS), I);
2278 I2 = InsertNewInstBefore(GetElementPtrInst::Create(I2, Other, "ctg2"), I);
2279 return new PtrToIntInst(I2, CI->getType());
2283 // add (select X 0 (sub n A)) A --> select X A n
2285 SelectInst *SI = dyn_cast<SelectInst>(LHS);
2288 SI = dyn_cast<SelectInst>(RHS);
2291 if (SI && SI->hasOneUse()) {
2292 Value *TV = SI->getTrueValue();
2293 Value *FV = SI->getFalseValue();
2296 // Can we fold the add into the argument of the select?
2297 // We check both true and false select arguments for a matching subtract.
2298 if (match(FV, m_Zero()) && match(TV, m_Sub(m_Value(N), m_Specific(A))))
2299 // Fold the add into the true select value.
2300 return SelectInst::Create(SI->getCondition(), N, A);
2301 if (match(TV, m_Zero()) && match(FV, m_Sub(m_Value(N), m_Specific(A))))
2302 // Fold the add into the false select value.
2303 return SelectInst::Create(SI->getCondition(), A, N);
2307 // Check for X+0.0. Simplify it to X if we know X is not -0.0.
2308 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHS))
2309 if (CFP->getValueAPF().isPosZero() && CannotBeNegativeZero(LHS))
2310 return ReplaceInstUsesWith(I, LHS);
2312 // Check for (add (sext x), y), see if we can merge this into an
2313 // integer add followed by a sext.
2314 if (SExtInst *LHSConv = dyn_cast<SExtInst>(LHS)) {
2315 // (add (sext x), cst) --> (sext (add x, cst'))
2316 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS)) {
2318 ConstantExpr::getTrunc(RHSC, LHSConv->getOperand(0)->getType());
2319 if (LHSConv->hasOneUse() &&
2320 ConstantExpr::getSExt(CI, I.getType()) == RHSC &&
2321 WillNotOverflowSignedAdd(LHSConv->getOperand(0), CI)) {
2322 // Insert the new, smaller add.
2323 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSConv->getOperand(0),
2325 InsertNewInstBefore(NewAdd, I);
2326 return new SExtInst(NewAdd, I.getType());
2330 // (add (sext x), (sext y)) --> (sext (add int x, y))
2331 if (SExtInst *RHSConv = dyn_cast<SExtInst>(RHS)) {
2332 // Only do this if x/y have the same type, if at last one of them has a
2333 // single use (so we don't increase the number of sexts), and if the
2334 // integer add will not overflow.
2335 if (LHSConv->getOperand(0)->getType()==RHSConv->getOperand(0)->getType()&&
2336 (LHSConv->hasOneUse() || RHSConv->hasOneUse()) &&
2337 WillNotOverflowSignedAdd(LHSConv->getOperand(0),
2338 RHSConv->getOperand(0))) {
2339 // Insert the new integer add.
2340 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSConv->getOperand(0),
2341 RHSConv->getOperand(0),
2343 InsertNewInstBefore(NewAdd, I);
2344 return new SExtInst(NewAdd, I.getType());
2349 // Check for (add double (sitofp x), y), see if we can merge this into an
2350 // integer add followed by a promotion.
2351 if (SIToFPInst *LHSConv = dyn_cast<SIToFPInst>(LHS)) {
2352 // (add double (sitofp x), fpcst) --> (sitofp (add int x, intcst))
2353 // ... if the constant fits in the integer value. This is useful for things
2354 // like (double)(x & 1234) + 4.0 -> (double)((X & 1234)+4) which no longer
2355 // requires a constant pool load, and generally allows the add to be better
2357 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHS)) {
2359 ConstantExpr::getFPToSI(CFP, LHSConv->getOperand(0)->getType());
2360 if (LHSConv->hasOneUse() &&
2361 ConstantExpr::getSIToFP(CI, I.getType()) == CFP &&
2362 WillNotOverflowSignedAdd(LHSConv->getOperand(0), CI)) {
2363 // Insert the new integer add.
2364 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSConv->getOperand(0),
2366 InsertNewInstBefore(NewAdd, I);
2367 return new SIToFPInst(NewAdd, I.getType());
2371 // (add double (sitofp x), (sitofp y)) --> (sitofp (add int x, y))
2372 if (SIToFPInst *RHSConv = dyn_cast<SIToFPInst>(RHS)) {
2373 // Only do this if x/y have the same type, if at last one of them has a
2374 // single use (so we don't increase the number of int->fp conversions),
2375 // and if the integer add will not overflow.
2376 if (LHSConv->getOperand(0)->getType()==RHSConv->getOperand(0)->getType()&&
2377 (LHSConv->hasOneUse() || RHSConv->hasOneUse()) &&
2378 WillNotOverflowSignedAdd(LHSConv->getOperand(0),
2379 RHSConv->getOperand(0))) {
2380 // Insert the new integer add.
2381 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSConv->getOperand(0),
2382 RHSConv->getOperand(0),
2384 InsertNewInstBefore(NewAdd, I);
2385 return new SIToFPInst(NewAdd, I.getType());
2390 return Changed ? &I : 0;
2393 Instruction *InstCombiner::visitSub(BinaryOperator &I) {
2394 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2396 if (Op0 == Op1 && // sub X, X -> 0
2397 !I.getType()->isFPOrFPVector())
2398 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2400 // If this is a 'B = x-(-A)', change to B = x+A...
2401 if (Value *V = dyn_castNegVal(Op1))
2402 return BinaryOperator::CreateAdd(Op0, V);
2404 if (isa<UndefValue>(Op0))
2405 return ReplaceInstUsesWith(I, Op0); // undef - X -> undef
2406 if (isa<UndefValue>(Op1))
2407 return ReplaceInstUsesWith(I, Op1); // X - undef -> undef
2409 if (ConstantInt *C = dyn_cast<ConstantInt>(Op0)) {
2410 // Replace (-1 - A) with (~A)...
2411 if (C->isAllOnesValue())
2412 return BinaryOperator::CreateNot(Op1);
2414 // C - ~X == X + (1+C)
2416 if (match(Op1, m_Not(m_Value(X))))
2417 return BinaryOperator::CreateAdd(X, AddOne(C));
2419 // -(X >>u 31) -> (X >>s 31)
2420 // -(X >>s 31) -> (X >>u 31)
2422 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op1)) {
2423 if (SI->getOpcode() == Instruction::LShr) {
2424 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
2425 // Check to see if we are shifting out everything but the sign bit.
2426 if (CU->getLimitedValue(SI->getType()->getPrimitiveSizeInBits()) ==
2427 SI->getType()->getPrimitiveSizeInBits()-1) {
2428 // Ok, the transformation is safe. Insert AShr.
2429 return BinaryOperator::Create(Instruction::AShr,
2430 SI->getOperand(0), CU, SI->getName());
2434 else if (SI->getOpcode() == Instruction::AShr) {
2435 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
2436 // Check to see if we are shifting out everything but the sign bit.
2437 if (CU->getLimitedValue(SI->getType()->getPrimitiveSizeInBits()) ==
2438 SI->getType()->getPrimitiveSizeInBits()-1) {
2439 // Ok, the transformation is safe. Insert LShr.
2440 return BinaryOperator::CreateLShr(
2441 SI->getOperand(0), CU, SI->getName());
2448 // Try to fold constant sub into select arguments.
2449 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
2450 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2454 if (I.getType() == Type::Int1Ty)
2455 return BinaryOperator::CreateXor(Op0, Op1);
2457 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
2458 if (Op1I->getOpcode() == Instruction::Add &&
2459 !Op0->getType()->isFPOrFPVector()) {
2460 if (Op1I->getOperand(0) == Op0) // X-(X+Y) == -Y
2461 return BinaryOperator::CreateNeg(Op1I->getOperand(1), I.getName());
2462 else if (Op1I->getOperand(1) == Op0) // X-(Y+X) == -Y
2463 return BinaryOperator::CreateNeg(Op1I->getOperand(0), I.getName());
2464 else if (ConstantInt *CI1 = dyn_cast<ConstantInt>(I.getOperand(0))) {
2465 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(Op1I->getOperand(1)))
2466 // C1-(X+C2) --> (C1-C2)-X
2467 return BinaryOperator::CreateSub(Subtract(CI1, CI2),
2468 Op1I->getOperand(0));
2472 if (Op1I->hasOneUse()) {
2473 // Replace (x - (y - z)) with (x + (z - y)) if the (y - z) subexpression
2474 // is not used by anyone else...
2476 if (Op1I->getOpcode() == Instruction::Sub &&
2477 !Op1I->getType()->isFPOrFPVector()) {
2478 // Swap the two operands of the subexpr...
2479 Value *IIOp0 = Op1I->getOperand(0), *IIOp1 = Op1I->getOperand(1);
2480 Op1I->setOperand(0, IIOp1);
2481 Op1I->setOperand(1, IIOp0);
2483 // Create the new top level add instruction...
2484 return BinaryOperator::CreateAdd(Op0, Op1);
2487 // Replace (A - (A & B)) with (A & ~B) if this is the only use of (A&B)...
2489 if (Op1I->getOpcode() == Instruction::And &&
2490 (Op1I->getOperand(0) == Op0 || Op1I->getOperand(1) == Op0)) {
2491 Value *OtherOp = Op1I->getOperand(Op1I->getOperand(0) == Op0);
2494 InsertNewInstBefore(BinaryOperator::CreateNot(OtherOp, "B.not"), I);
2495 return BinaryOperator::CreateAnd(Op0, NewNot);
2498 // 0 - (X sdiv C) -> (X sdiv -C)
2499 if (Op1I->getOpcode() == Instruction::SDiv)
2500 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
2502 if (Constant *DivRHS = dyn_cast<Constant>(Op1I->getOperand(1)))
2503 return BinaryOperator::CreateSDiv(Op1I->getOperand(0),
2504 ConstantExpr::getNeg(DivRHS));
2506 // X - X*C --> X * (1-C)
2507 ConstantInt *C2 = 0;
2508 if (dyn_castFoldableMul(Op1I, C2) == Op0) {
2509 Constant *CP1 = Subtract(ConstantInt::get(I.getType(), 1), C2);
2510 return BinaryOperator::CreateMul(Op0, CP1);
2515 if (!Op0->getType()->isFPOrFPVector())
2516 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
2517 if (Op0I->getOpcode() == Instruction::Add) {
2518 if (Op0I->getOperand(0) == Op1) // (Y+X)-Y == X
2519 return ReplaceInstUsesWith(I, Op0I->getOperand(1));
2520 else if (Op0I->getOperand(1) == Op1) // (X+Y)-Y == X
2521 return ReplaceInstUsesWith(I, Op0I->getOperand(0));
2522 } else if (Op0I->getOpcode() == Instruction::Sub) {
2523 if (Op0I->getOperand(0) == Op1) // (X-Y)-X == -Y
2524 return BinaryOperator::CreateNeg(Op0I->getOperand(1), I.getName());
2529 if (Value *X = dyn_castFoldableMul(Op0, C1)) {
2530 if (X == Op1) // X*C - X --> X * (C-1)
2531 return BinaryOperator::CreateMul(Op1, SubOne(C1));
2533 ConstantInt *C2; // X*C1 - X*C2 -> X * (C1-C2)
2534 if (X == dyn_castFoldableMul(Op1, C2))
2535 return BinaryOperator::CreateMul(X, Subtract(C1, C2));
2540 /// isSignBitCheck - Given an exploded icmp instruction, return true if the
2541 /// comparison only checks the sign bit. If it only checks the sign bit, set
2542 /// TrueIfSigned if the result of the comparison is true when the input value is
2544 static bool isSignBitCheck(ICmpInst::Predicate pred, ConstantInt *RHS,
2545 bool &TrueIfSigned) {
2547 case ICmpInst::ICMP_SLT: // True if LHS s< 0
2548 TrueIfSigned = true;
2549 return RHS->isZero();
2550 case ICmpInst::ICMP_SLE: // True if LHS s<= RHS and RHS == -1
2551 TrueIfSigned = true;
2552 return RHS->isAllOnesValue();
2553 case ICmpInst::ICMP_SGT: // True if LHS s> -1
2554 TrueIfSigned = false;
2555 return RHS->isAllOnesValue();
2556 case ICmpInst::ICMP_UGT:
2557 // True if LHS u> RHS and RHS == high-bit-mask - 1
2558 TrueIfSigned = true;
2559 return RHS->getValue() ==
2560 APInt::getSignedMaxValue(RHS->getType()->getPrimitiveSizeInBits());
2561 case ICmpInst::ICMP_UGE:
2562 // True if LHS u>= RHS and RHS == high-bit-mask (2^7, 2^15, 2^31, etc)
2563 TrueIfSigned = true;
2564 return RHS->getValue().isSignBit();
2570 Instruction *InstCombiner::visitMul(BinaryOperator &I) {
2571 bool Changed = SimplifyCommutative(I);
2572 Value *Op0 = I.getOperand(0);
2574 if (isa<UndefValue>(I.getOperand(1))) // undef * X -> 0
2575 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2577 // Simplify mul instructions with a constant RHS...
2578 if (Constant *Op1 = dyn_cast<Constant>(I.getOperand(1))) {
2579 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2581 // ((X << C1)*C2) == (X * (C2 << C1))
2582 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op0))
2583 if (SI->getOpcode() == Instruction::Shl)
2584 if (Constant *ShOp = dyn_cast<Constant>(SI->getOperand(1)))
2585 return BinaryOperator::CreateMul(SI->getOperand(0),
2586 ConstantExpr::getShl(CI, ShOp));
2589 return ReplaceInstUsesWith(I, Op1); // X * 0 == 0
2590 if (CI->equalsInt(1)) // X * 1 == X
2591 return ReplaceInstUsesWith(I, Op0);
2592 if (CI->isAllOnesValue()) // X * -1 == 0 - X
2593 return BinaryOperator::CreateNeg(Op0, I.getName());
2595 const APInt& Val = cast<ConstantInt>(CI)->getValue();
2596 if (Val.isPowerOf2()) { // Replace X*(2^C) with X << C
2597 return BinaryOperator::CreateShl(Op0,
2598 ConstantInt::get(Op0->getType(), Val.logBase2()));
2600 } else if (ConstantFP *Op1F = dyn_cast<ConstantFP>(Op1)) {
2601 if (Op1F->isNullValue())
2602 return ReplaceInstUsesWith(I, Op1);
2604 // "In IEEE floating point, x*1 is not equivalent to x for nans. However,
2605 // ANSI says we can drop signals, so we can do this anyway." (from GCC)
2606 if (Op1F->isExactlyValue(1.0))
2607 return ReplaceInstUsesWith(I, Op0); // Eliminate 'mul double %X, 1.0'
2608 } else if (isa<VectorType>(Op1->getType())) {
2609 if (isa<ConstantAggregateZero>(Op1))
2610 return ReplaceInstUsesWith(I, Op1);
2612 if (ConstantVector *Op1V = dyn_cast<ConstantVector>(Op1)) {
2613 if (Op1V->isAllOnesValue()) // X * -1 == 0 - X
2614 return BinaryOperator::CreateNeg(Op0, I.getName());
2616 // As above, vector X*splat(1.0) -> X in all defined cases.
2617 if (Constant *Splat = Op1V->getSplatValue()) {
2618 if (ConstantFP *F = dyn_cast<ConstantFP>(Splat))
2619 if (F->isExactlyValue(1.0))
2620 return ReplaceInstUsesWith(I, Op0);
2621 if (ConstantInt *CI = dyn_cast<ConstantInt>(Splat))
2622 if (CI->equalsInt(1))
2623 return ReplaceInstUsesWith(I, Op0);
2628 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
2629 if (Op0I->getOpcode() == Instruction::Add && Op0I->hasOneUse() &&
2630 isa<ConstantInt>(Op0I->getOperand(1)) && isa<ConstantInt>(Op1)) {
2631 // Canonicalize (X+C1)*C2 -> X*C2+C1*C2.
2632 Instruction *Add = BinaryOperator::CreateMul(Op0I->getOperand(0),
2634 InsertNewInstBefore(Add, I);
2635 Value *C1C2 = ConstantExpr::getMul(Op1,
2636 cast<Constant>(Op0I->getOperand(1)));
2637 return BinaryOperator::CreateAdd(Add, C1C2);
2641 // Try to fold constant mul into select arguments.
2642 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2643 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2646 if (isa<PHINode>(Op0))
2647 if (Instruction *NV = FoldOpIntoPhi(I))
2651 if (Value *Op0v = dyn_castNegVal(Op0)) // -X * -Y = X*Y
2652 if (Value *Op1v = dyn_castNegVal(I.getOperand(1)))
2653 return BinaryOperator::CreateMul(Op0v, Op1v);
2655 // (X / Y) * Y = X - (X % Y)
2656 // (X / Y) * -Y = (X % Y) - X
2658 Value *Op1 = I.getOperand(1);
2659 BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0);
2661 (BO->getOpcode() != Instruction::UDiv &&
2662 BO->getOpcode() != Instruction::SDiv)) {
2664 BO = dyn_cast<BinaryOperator>(I.getOperand(1));
2666 Value *Neg = dyn_castNegVal(Op1);
2667 if (BO && BO->hasOneUse() &&
2668 (BO->getOperand(1) == Op1 || BO->getOperand(1) == Neg) &&
2669 (BO->getOpcode() == Instruction::UDiv ||
2670 BO->getOpcode() == Instruction::SDiv)) {
2671 Value *Op0BO = BO->getOperand(0), *Op1BO = BO->getOperand(1);
2674 if (BO->getOpcode() == Instruction::UDiv)
2675 Rem = BinaryOperator::CreateURem(Op0BO, Op1BO);
2677 Rem = BinaryOperator::CreateSRem(Op0BO, Op1BO);
2679 InsertNewInstBefore(Rem, I);
2683 return BinaryOperator::CreateSub(Op0BO, Rem);
2685 return BinaryOperator::CreateSub(Rem, Op0BO);
2689 if (I.getType() == Type::Int1Ty)
2690 return BinaryOperator::CreateAnd(Op0, I.getOperand(1));
2692 // If one of the operands of the multiply is a cast from a boolean value, then
2693 // we know the bool is either zero or one, so this is a 'masking' multiply.
2694 // See if we can simplify things based on how the boolean was originally
2696 CastInst *BoolCast = 0;
2697 if (ZExtInst *CI = dyn_cast<ZExtInst>(Op0))
2698 if (CI->getOperand(0)->getType() == Type::Int1Ty)
2701 if (ZExtInst *CI = dyn_cast<ZExtInst>(I.getOperand(1)))
2702 if (CI->getOperand(0)->getType() == Type::Int1Ty)
2705 if (ICmpInst *SCI = dyn_cast<ICmpInst>(BoolCast->getOperand(0))) {
2706 Value *SCIOp0 = SCI->getOperand(0), *SCIOp1 = SCI->getOperand(1);
2707 const Type *SCOpTy = SCIOp0->getType();
2710 // If the icmp is true iff the sign bit of X is set, then convert this
2711 // multiply into a shift/and combination.
2712 if (isa<ConstantInt>(SCIOp1) &&
2713 isSignBitCheck(SCI->getPredicate(), cast<ConstantInt>(SCIOp1), TIS) &&
2715 // Shift the X value right to turn it into "all signbits".
2716 Constant *Amt = ConstantInt::get(SCIOp0->getType(),
2717 SCOpTy->getPrimitiveSizeInBits()-1);
2719 InsertNewInstBefore(
2720 BinaryOperator::Create(Instruction::AShr, SCIOp0, Amt,
2721 BoolCast->getOperand(0)->getName()+
2724 // If the multiply type is not the same as the source type, sign extend
2725 // or truncate to the multiply type.
2726 if (I.getType() != V->getType()) {
2727 uint32_t SrcBits = V->getType()->getPrimitiveSizeInBits();
2728 uint32_t DstBits = I.getType()->getPrimitiveSizeInBits();
2729 Instruction::CastOps opcode =
2730 (SrcBits == DstBits ? Instruction::BitCast :
2731 (SrcBits < DstBits ? Instruction::SExt : Instruction::Trunc));
2732 V = InsertCastBefore(opcode, V, I.getType(), I);
2735 Value *OtherOp = Op0 == BoolCast ? I.getOperand(1) : Op0;
2736 return BinaryOperator::CreateAnd(V, OtherOp);
2741 return Changed ? &I : 0;
2744 /// SimplifyDivRemOfSelect - Try to fold a divide or remainder of a select
2746 bool InstCombiner::SimplifyDivRemOfSelect(BinaryOperator &I) {
2747 SelectInst *SI = cast<SelectInst>(I.getOperand(1));
2749 // div/rem X, (Cond ? 0 : Y) -> div/rem X, Y
2750 int NonNullOperand = -1;
2751 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(1)))
2752 if (ST->isNullValue())
2754 // div/rem X, (Cond ? Y : 0) -> div/rem X, Y
2755 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(2)))
2756 if (ST->isNullValue())
2759 if (NonNullOperand == -1)
2762 Value *SelectCond = SI->getOperand(0);
2764 // Change the div/rem to use 'Y' instead of the select.
2765 I.setOperand(1, SI->getOperand(NonNullOperand));
2767 // Okay, we know we replace the operand of the div/rem with 'Y' with no
2768 // problem. However, the select, or the condition of the select may have
2769 // multiple uses. Based on our knowledge that the operand must be non-zero,
2770 // propagate the known value for the select into other uses of it, and
2771 // propagate a known value of the condition into its other users.
2773 // If the select and condition only have a single use, don't bother with this,
2775 if (SI->use_empty() && SelectCond->hasOneUse())
2778 // Scan the current block backward, looking for other uses of SI.
2779 BasicBlock::iterator BBI = &I, BBFront = I.getParent()->begin();
2781 while (BBI != BBFront) {
2783 // If we found a call to a function, we can't assume it will return, so
2784 // information from below it cannot be propagated above it.
2785 if (isa<CallInst>(BBI) && !isa<IntrinsicInst>(BBI))
2788 // Replace uses of the select or its condition with the known values.
2789 for (Instruction::op_iterator I = BBI->op_begin(), E = BBI->op_end();
2792 *I = SI->getOperand(NonNullOperand);
2794 } else if (*I == SelectCond) {
2795 *I = NonNullOperand == 1 ? ConstantInt::getTrue() :
2796 ConstantInt::getFalse();
2801 // If we past the instruction, quit looking for it.
2804 if (&*BBI == SelectCond)
2807 // If we ran out of things to eliminate, break out of the loop.
2808 if (SelectCond == 0 && SI == 0)
2816 /// This function implements the transforms on div instructions that work
2817 /// regardless of the kind of div instruction it is (udiv, sdiv, or fdiv). It is
2818 /// used by the visitors to those instructions.
2819 /// @brief Transforms common to all three div instructions
2820 Instruction *InstCombiner::commonDivTransforms(BinaryOperator &I) {
2821 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2823 // undef / X -> 0 for integer.
2824 // undef / X -> undef for FP (the undef could be a snan).
2825 if (isa<UndefValue>(Op0)) {
2826 if (Op0->getType()->isFPOrFPVector())
2827 return ReplaceInstUsesWith(I, Op0);
2828 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2831 // X / undef -> undef
2832 if (isa<UndefValue>(Op1))
2833 return ReplaceInstUsesWith(I, Op1);
2838 /// This function implements the transforms common to both integer division
2839 /// instructions (udiv and sdiv). It is called by the visitors to those integer
2840 /// division instructions.
2841 /// @brief Common integer divide transforms
2842 Instruction *InstCombiner::commonIDivTransforms(BinaryOperator &I) {
2843 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2845 // (sdiv X, X) --> 1 (udiv X, X) --> 1
2847 if (const VectorType *Ty = dyn_cast<VectorType>(I.getType())) {
2848 ConstantInt *CI = ConstantInt::get(Ty->getElementType(), 1);
2849 std::vector<Constant*> Elts(Ty->getNumElements(), CI);
2850 return ReplaceInstUsesWith(I, ConstantVector::get(Elts));
2853 ConstantInt *CI = ConstantInt::get(I.getType(), 1);
2854 return ReplaceInstUsesWith(I, CI);
2857 if (Instruction *Common = commonDivTransforms(I))
2860 // Handle cases involving: [su]div X, (select Cond, Y, Z)
2861 // This does not apply for fdiv.
2862 if (isa<SelectInst>(Op1) && SimplifyDivRemOfSelect(I))
2865 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2867 if (RHS->equalsInt(1))
2868 return ReplaceInstUsesWith(I, Op0);
2870 // (X / C1) / C2 -> X / (C1*C2)
2871 if (Instruction *LHS = dyn_cast<Instruction>(Op0))
2872 if (Instruction::BinaryOps(LHS->getOpcode()) == I.getOpcode())
2873 if (ConstantInt *LHSRHS = dyn_cast<ConstantInt>(LHS->getOperand(1))) {
2874 if (MultiplyOverflows(RHS, LHSRHS, I.getOpcode()==Instruction::SDiv))
2875 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2877 return BinaryOperator::Create(I.getOpcode(), LHS->getOperand(0),
2878 Multiply(RHS, LHSRHS));
2881 if (!RHS->isZero()) { // avoid X udiv 0
2882 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2883 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2885 if (isa<PHINode>(Op0))
2886 if (Instruction *NV = FoldOpIntoPhi(I))
2891 // 0 / X == 0, we don't need to preserve faults!
2892 if (ConstantInt *LHS = dyn_cast<ConstantInt>(Op0))
2893 if (LHS->equalsInt(0))
2894 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2896 // It can't be division by zero, hence it must be division by one.
2897 if (I.getType() == Type::Int1Ty)
2898 return ReplaceInstUsesWith(I, Op0);
2900 if (ConstantVector *Op1V = dyn_cast<ConstantVector>(Op1)) {
2901 if (ConstantInt *X = cast_or_null<ConstantInt>(Op1V->getSplatValue()))
2904 return ReplaceInstUsesWith(I, Op0);
2910 Instruction *InstCombiner::visitUDiv(BinaryOperator &I) {
2911 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2913 // Handle the integer div common cases
2914 if (Instruction *Common = commonIDivTransforms(I))
2917 if (ConstantInt *C = dyn_cast<ConstantInt>(Op1)) {
2918 // X udiv C^2 -> X >> C
2919 // Check to see if this is an unsigned division with an exact power of 2,
2920 // if so, convert to a right shift.
2921 if (C->getValue().isPowerOf2()) // 0 not included in isPowerOf2
2922 return BinaryOperator::CreateLShr(Op0,
2923 ConstantInt::get(Op0->getType(), C->getValue().logBase2()));
2925 // X udiv C, where C >= signbit
2926 if (C->getValue().isNegative()) {
2927 Value *IC = InsertNewInstBefore(new ICmpInst(ICmpInst::ICMP_ULT, Op0, C),
2929 return SelectInst::Create(IC, Constant::getNullValue(I.getType()),
2930 ConstantInt::get(I.getType(), 1));
2934 // X udiv (C1 << N), where C1 is "1<<C2" --> X >> (N+C2)
2935 if (BinaryOperator *RHSI = dyn_cast<BinaryOperator>(I.getOperand(1))) {
2936 if (RHSI->getOpcode() == Instruction::Shl &&
2937 isa<ConstantInt>(RHSI->getOperand(0))) {
2938 const APInt& C1 = cast<ConstantInt>(RHSI->getOperand(0))->getValue();
2939 if (C1.isPowerOf2()) {
2940 Value *N = RHSI->getOperand(1);
2941 const Type *NTy = N->getType();
2942 if (uint32_t C2 = C1.logBase2()) {
2943 Constant *C2V = ConstantInt::get(NTy, C2);
2944 N = InsertNewInstBefore(BinaryOperator::CreateAdd(N, C2V, "tmp"), I);
2946 return BinaryOperator::CreateLShr(Op0, N);
2951 // udiv X, (Select Cond, C1, C2) --> Select Cond, (shr X, C1), (shr X, C2)
2952 // where C1&C2 are powers of two.
2953 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
2954 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
2955 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
2956 const APInt &TVA = STO->getValue(), &FVA = SFO->getValue();
2957 if (TVA.isPowerOf2() && FVA.isPowerOf2()) {
2958 // Compute the shift amounts
2959 uint32_t TSA = TVA.logBase2(), FSA = FVA.logBase2();
2960 // Construct the "on true" case of the select
2961 Constant *TC = ConstantInt::get(Op0->getType(), TSA);
2962 Instruction *TSI = BinaryOperator::CreateLShr(
2963 Op0, TC, SI->getName()+".t");
2964 TSI = InsertNewInstBefore(TSI, I);
2966 // Construct the "on false" case of the select
2967 Constant *FC = ConstantInt::get(Op0->getType(), FSA);
2968 Instruction *FSI = BinaryOperator::CreateLShr(
2969 Op0, FC, SI->getName()+".f");
2970 FSI = InsertNewInstBefore(FSI, I);
2972 // construct the select instruction and return it.
2973 return SelectInst::Create(SI->getOperand(0), TSI, FSI, SI->getName());
2979 Instruction *InstCombiner::visitSDiv(BinaryOperator &I) {
2980 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2982 // Handle the integer div common cases
2983 if (Instruction *Common = commonIDivTransforms(I))
2986 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2988 if (RHS->isAllOnesValue())
2989 return BinaryOperator::CreateNeg(Op0);
2992 // If the sign bits of both operands are zero (i.e. we can prove they are
2993 // unsigned inputs), turn this into a udiv.
2994 if (I.getType()->isInteger()) {
2995 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()));
2996 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
2997 // X sdiv Y -> X udiv Y, iff X and Y don't have sign bit set
2998 return BinaryOperator::CreateUDiv(Op0, Op1, I.getName());
3005 Instruction *InstCombiner::visitFDiv(BinaryOperator &I) {
3006 return commonDivTransforms(I);
3009 /// This function implements the transforms on rem instructions that work
3010 /// regardless of the kind of rem instruction it is (urem, srem, or frem). It
3011 /// is used by the visitors to those instructions.
3012 /// @brief Transforms common to all three rem instructions
3013 Instruction *InstCombiner::commonRemTransforms(BinaryOperator &I) {
3014 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3016 if (isa<UndefValue>(Op0)) { // undef % X -> 0
3017 if (I.getType()->isFPOrFPVector())
3018 return ReplaceInstUsesWith(I, Op0); // X % undef -> undef (could be SNaN)
3019 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3021 if (isa<UndefValue>(Op1))
3022 return ReplaceInstUsesWith(I, Op1); // X % undef -> undef
3024 // Handle cases involving: rem X, (select Cond, Y, Z)
3025 if (isa<SelectInst>(Op1) && SimplifyDivRemOfSelect(I))
3031 /// This function implements the transforms common to both integer remainder
3032 /// instructions (urem and srem). It is called by the visitors to those integer
3033 /// remainder instructions.
3034 /// @brief Common integer remainder transforms
3035 Instruction *InstCombiner::commonIRemTransforms(BinaryOperator &I) {
3036 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3038 if (Instruction *common = commonRemTransforms(I))
3041 // 0 % X == 0 for integer, we don't need to preserve faults!
3042 if (Constant *LHS = dyn_cast<Constant>(Op0))
3043 if (LHS->isNullValue())
3044 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3046 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3047 // X % 0 == undef, we don't need to preserve faults!
3048 if (RHS->equalsInt(0))
3049 return ReplaceInstUsesWith(I, UndefValue::get(I.getType()));
3051 if (RHS->equalsInt(1)) // X % 1 == 0
3052 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3054 if (Instruction *Op0I = dyn_cast<Instruction>(Op0)) {
3055 if (SelectInst *SI = dyn_cast<SelectInst>(Op0I)) {
3056 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3058 } else if (isa<PHINode>(Op0I)) {
3059 if (Instruction *NV = FoldOpIntoPhi(I))
3063 // See if we can fold away this rem instruction.
3064 if (SimplifyDemandedInstructionBits(I))
3072 Instruction *InstCombiner::visitURem(BinaryOperator &I) {
3073 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3075 if (Instruction *common = commonIRemTransforms(I))
3078 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3079 // X urem C^2 -> X and C
3080 // Check to see if this is an unsigned remainder with an exact power of 2,
3081 // if so, convert to a bitwise and.
3082 if (ConstantInt *C = dyn_cast<ConstantInt>(RHS))
3083 if (C->getValue().isPowerOf2())
3084 return BinaryOperator::CreateAnd(Op0, SubOne(C));
3087 if (Instruction *RHSI = dyn_cast<Instruction>(I.getOperand(1))) {
3088 // Turn A % (C << N), where C is 2^k, into A & ((C << N)-1)
3089 if (RHSI->getOpcode() == Instruction::Shl &&
3090 isa<ConstantInt>(RHSI->getOperand(0))) {
3091 if (cast<ConstantInt>(RHSI->getOperand(0))->getValue().isPowerOf2()) {
3092 Constant *N1 = ConstantInt::getAllOnesValue(I.getType());
3093 Value *Add = InsertNewInstBefore(BinaryOperator::CreateAdd(RHSI, N1,
3095 return BinaryOperator::CreateAnd(Op0, Add);
3100 // urem X, (select Cond, 2^C1, 2^C2) --> select Cond, (and X, C1), (and X, C2)
3101 // where C1&C2 are powers of two.
3102 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
3103 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
3104 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
3105 // STO == 0 and SFO == 0 handled above.
3106 if ((STO->getValue().isPowerOf2()) &&
3107 (SFO->getValue().isPowerOf2())) {
3108 Value *TrueAnd = InsertNewInstBefore(
3109 BinaryOperator::CreateAnd(Op0, SubOne(STO), SI->getName()+".t"), I);
3110 Value *FalseAnd = InsertNewInstBefore(
3111 BinaryOperator::CreateAnd(Op0, SubOne(SFO), SI->getName()+".f"), I);
3112 return SelectInst::Create(SI->getOperand(0), TrueAnd, FalseAnd);
3120 Instruction *InstCombiner::visitSRem(BinaryOperator &I) {
3121 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3123 // Handle the integer rem common cases
3124 if (Instruction *common = commonIRemTransforms(I))
3127 if (Value *RHSNeg = dyn_castNegVal(Op1))
3128 if (!isa<Constant>(RHSNeg) ||
3129 (isa<ConstantInt>(RHSNeg) &&
3130 cast<ConstantInt>(RHSNeg)->getValue().isStrictlyPositive())) {
3132 AddUsesToWorkList(I);
3133 I.setOperand(1, RHSNeg);
3137 // If the sign bits of both operands are zero (i.e. we can prove they are
3138 // unsigned inputs), turn this into a urem.
3139 if (I.getType()->isInteger()) {
3140 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()));
3141 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
3142 // X srem Y -> X urem Y, iff X and Y don't have sign bit set
3143 return BinaryOperator::CreateURem(Op0, Op1, I.getName());
3147 // If it's a constant vector, flip any negative values positive.
3148 if (ConstantVector *RHSV = dyn_cast<ConstantVector>(Op1)) {
3149 unsigned VWidth = RHSV->getNumOperands();
3151 bool hasNegative = false;
3152 for (unsigned i = 0; !hasNegative && i != VWidth; ++i)
3153 if (ConstantInt *RHS = dyn_cast<ConstantInt>(RHSV->getOperand(i)))
3154 if (RHS->getValue().isNegative())
3158 std::vector<Constant *> Elts(VWidth);
3159 for (unsigned i = 0; i != VWidth; ++i) {
3160 if (ConstantInt *RHS = dyn_cast<ConstantInt>(RHSV->getOperand(i))) {
3161 if (RHS->getValue().isNegative())
3162 Elts[i] = cast<ConstantInt>(ConstantExpr::getNeg(RHS));
3168 Constant *NewRHSV = ConstantVector::get(Elts);
3169 if (NewRHSV != RHSV) {
3170 AddUsesToWorkList(I);
3171 I.setOperand(1, NewRHSV);
3180 Instruction *InstCombiner::visitFRem(BinaryOperator &I) {
3181 return commonRemTransforms(I);
3184 // isOneBitSet - Return true if there is exactly one bit set in the specified
3186 static bool isOneBitSet(const ConstantInt *CI) {
3187 return CI->getValue().isPowerOf2();
3190 // isHighOnes - Return true if the constant is of the form 1+0+.
3191 // This is the same as lowones(~X).
3192 static bool isHighOnes(const ConstantInt *CI) {
3193 return (~CI->getValue() + 1).isPowerOf2();
3196 /// getICmpCode - Encode a icmp predicate into a three bit mask. These bits
3197 /// are carefully arranged to allow folding of expressions such as:
3199 /// (A < B) | (A > B) --> (A != B)
3201 /// Note that this is only valid if the first and second predicates have the
3202 /// same sign. Is illegal to do: (A u< B) | (A s> B)
3204 /// Three bits are used to represent the condition, as follows:
3209 /// <=> Value Definition
3210 /// 000 0 Always false
3217 /// 111 7 Always true
3219 static unsigned getICmpCode(const ICmpInst *ICI) {
3220 switch (ICI->getPredicate()) {
3222 case ICmpInst::ICMP_UGT: return 1; // 001
3223 case ICmpInst::ICMP_SGT: return 1; // 001
3224 case ICmpInst::ICMP_EQ: return 2; // 010
3225 case ICmpInst::ICMP_UGE: return 3; // 011
3226 case ICmpInst::ICMP_SGE: return 3; // 011
3227 case ICmpInst::ICMP_ULT: return 4; // 100
3228 case ICmpInst::ICMP_SLT: return 4; // 100
3229 case ICmpInst::ICMP_NE: return 5; // 101
3230 case ICmpInst::ICMP_ULE: return 6; // 110
3231 case ICmpInst::ICMP_SLE: return 6; // 110
3234 assert(0 && "Invalid ICmp predicate!");
3239 /// getFCmpCode - Similar to getICmpCode but for FCmpInst. This encodes a fcmp
3240 /// predicate into a three bit mask. It also returns whether it is an ordered
3241 /// predicate by reference.
3242 static unsigned getFCmpCode(FCmpInst::Predicate CC, bool &isOrdered) {
3245 case FCmpInst::FCMP_ORD: isOrdered = true; return 0; // 000
3246 case FCmpInst::FCMP_UNO: return 0; // 000
3247 case FCmpInst::FCMP_OGT: isOrdered = true; return 1; // 001
3248 case FCmpInst::FCMP_UGT: return 1; // 001
3249 case FCmpInst::FCMP_OEQ: isOrdered = true; return 2; // 010
3250 case FCmpInst::FCMP_UEQ: return 2; // 010
3251 case FCmpInst::FCMP_OGE: isOrdered = true; return 3; // 011
3252 case FCmpInst::FCMP_UGE: return 3; // 011
3253 case FCmpInst::FCMP_OLT: isOrdered = true; return 4; // 100
3254 case FCmpInst::FCMP_ULT: return 4; // 100
3255 case FCmpInst::FCMP_ONE: isOrdered = true; return 5; // 101
3256 case FCmpInst::FCMP_UNE: return 5; // 101
3257 case FCmpInst::FCMP_OLE: isOrdered = true; return 6; // 110
3258 case FCmpInst::FCMP_ULE: return 6; // 110
3261 // Not expecting FCMP_FALSE and FCMP_TRUE;
3262 assert(0 && "Unexpected FCmp predicate!");
3267 /// getICmpValue - This is the complement of getICmpCode, which turns an
3268 /// opcode and two operands into either a constant true or false, or a brand
3269 /// new ICmp instruction. The sign is passed in to determine which kind
3270 /// of predicate to use in the new icmp instruction.
3271 static Value *getICmpValue(bool sign, unsigned code, Value *LHS, Value *RHS) {
3273 default: assert(0 && "Illegal ICmp code!");
3274 case 0: return ConstantInt::getFalse();
3277 return new ICmpInst(ICmpInst::ICMP_SGT, LHS, RHS);
3279 return new ICmpInst(ICmpInst::ICMP_UGT, LHS, RHS);
3280 case 2: return new ICmpInst(ICmpInst::ICMP_EQ, LHS, RHS);
3283 return new ICmpInst(ICmpInst::ICMP_SGE, LHS, RHS);
3285 return new ICmpInst(ICmpInst::ICMP_UGE, LHS, RHS);
3288 return new ICmpInst(ICmpInst::ICMP_SLT, LHS, RHS);
3290 return new ICmpInst(ICmpInst::ICMP_ULT, LHS, RHS);
3291 case 5: return new ICmpInst(ICmpInst::ICMP_NE, LHS, RHS);
3294 return new ICmpInst(ICmpInst::ICMP_SLE, LHS, RHS);
3296 return new ICmpInst(ICmpInst::ICMP_ULE, LHS, RHS);
3297 case 7: return ConstantInt::getTrue();
3301 /// getFCmpValue - This is the complement of getFCmpCode, which turns an
3302 /// opcode and two operands into either a FCmp instruction. isordered is passed
3303 /// in to determine which kind of predicate to use in the new fcmp instruction.
3304 static Value *getFCmpValue(bool isordered, unsigned code,
3305 Value *LHS, Value *RHS) {
3307 default: assert(0 && "Illegal FCmp code!");
3310 return new FCmpInst(FCmpInst::FCMP_ORD, LHS, RHS);
3312 return new FCmpInst(FCmpInst::FCMP_UNO, LHS, RHS);
3315 return new FCmpInst(FCmpInst::FCMP_OGT, LHS, RHS);
3317 return new FCmpInst(FCmpInst::FCMP_UGT, LHS, RHS);
3320 return new FCmpInst(FCmpInst::FCMP_OEQ, LHS, RHS);
3322 return new FCmpInst(FCmpInst::FCMP_UEQ, LHS, RHS);
3325 return new FCmpInst(FCmpInst::FCMP_OGE, LHS, RHS);
3327 return new FCmpInst(FCmpInst::FCMP_UGE, LHS, RHS);
3330 return new FCmpInst(FCmpInst::FCMP_OLT, LHS, RHS);
3332 return new FCmpInst(FCmpInst::FCMP_ULT, LHS, RHS);
3335 return new FCmpInst(FCmpInst::FCMP_ONE, LHS, RHS);
3337 return new FCmpInst(FCmpInst::FCMP_UNE, LHS, RHS);
3340 return new FCmpInst(FCmpInst::FCMP_OLE, LHS, RHS);
3342 return new FCmpInst(FCmpInst::FCMP_ULE, LHS, RHS);
3343 case 7: return ConstantInt::getTrue();
3347 /// PredicatesFoldable - Return true if both predicates match sign or if at
3348 /// least one of them is an equality comparison (which is signless).
3349 static bool PredicatesFoldable(ICmpInst::Predicate p1, ICmpInst::Predicate p2) {
3350 return (ICmpInst::isSignedPredicate(p1) == ICmpInst::isSignedPredicate(p2)) ||
3351 (ICmpInst::isSignedPredicate(p1) && ICmpInst::isEquality(p2)) ||
3352 (ICmpInst::isSignedPredicate(p2) && ICmpInst::isEquality(p1));
3356 // FoldICmpLogical - Implements (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
3357 struct FoldICmpLogical {
3360 ICmpInst::Predicate pred;
3361 FoldICmpLogical(InstCombiner &ic, ICmpInst *ICI)
3362 : IC(ic), LHS(ICI->getOperand(0)), RHS(ICI->getOperand(1)),
3363 pred(ICI->getPredicate()) {}
3364 bool shouldApply(Value *V) const {
3365 if (ICmpInst *ICI = dyn_cast<ICmpInst>(V))
3366 if (PredicatesFoldable(pred, ICI->getPredicate()))
3367 return ((ICI->getOperand(0) == LHS && ICI->getOperand(1) == RHS) ||
3368 (ICI->getOperand(0) == RHS && ICI->getOperand(1) == LHS));
3371 Instruction *apply(Instruction &Log) const {
3372 ICmpInst *ICI = cast<ICmpInst>(Log.getOperand(0));
3373 if (ICI->getOperand(0) != LHS) {
3374 assert(ICI->getOperand(1) == LHS);
3375 ICI->swapOperands(); // Swap the LHS and RHS of the ICmp
3378 ICmpInst *RHSICI = cast<ICmpInst>(Log.getOperand(1));
3379 unsigned LHSCode = getICmpCode(ICI);
3380 unsigned RHSCode = getICmpCode(RHSICI);
3382 switch (Log.getOpcode()) {
3383 case Instruction::And: Code = LHSCode & RHSCode; break;
3384 case Instruction::Or: Code = LHSCode | RHSCode; break;
3385 case Instruction::Xor: Code = LHSCode ^ RHSCode; break;
3386 default: assert(0 && "Illegal logical opcode!"); return 0;
3389 bool isSigned = ICmpInst::isSignedPredicate(RHSICI->getPredicate()) ||
3390 ICmpInst::isSignedPredicate(ICI->getPredicate());
3392 Value *RV = getICmpValue(isSigned, Code, LHS, RHS);
3393 if (Instruction *I = dyn_cast<Instruction>(RV))
3395 // Otherwise, it's a constant boolean value...
3396 return IC.ReplaceInstUsesWith(Log, RV);
3399 } // end anonymous namespace
3401 // OptAndOp - This handles expressions of the form ((val OP C1) & C2). Where
3402 // the Op parameter is 'OP', OpRHS is 'C1', and AndRHS is 'C2'. Op is
3403 // guaranteed to be a binary operator.
3404 Instruction *InstCombiner::OptAndOp(Instruction *Op,
3406 ConstantInt *AndRHS,
3407 BinaryOperator &TheAnd) {
3408 Value *X = Op->getOperand(0);
3409 Constant *Together = 0;
3411 Together = And(AndRHS, OpRHS);
3413 switch (Op->getOpcode()) {
3414 case Instruction::Xor:
3415 if (Op->hasOneUse()) {
3416 // (X ^ C1) & C2 --> (X & C2) ^ (C1&C2)
3417 Instruction *And = BinaryOperator::CreateAnd(X, AndRHS);
3418 InsertNewInstBefore(And, TheAnd);
3420 return BinaryOperator::CreateXor(And, Together);
3423 case Instruction::Or:
3424 if (Together == AndRHS) // (X | C) & C --> C
3425 return ReplaceInstUsesWith(TheAnd, AndRHS);
3427 if (Op->hasOneUse() && Together != OpRHS) {
3428 // (X | C1) & C2 --> (X | (C1&C2)) & C2
3429 Instruction *Or = BinaryOperator::CreateOr(X, Together);
3430 InsertNewInstBefore(Or, TheAnd);
3432 return BinaryOperator::CreateAnd(Or, AndRHS);
3435 case Instruction::Add:
3436 if (Op->hasOneUse()) {
3437 // Adding a one to a single bit bit-field should be turned into an XOR
3438 // of the bit. First thing to check is to see if this AND is with a
3439 // single bit constant.
3440 const APInt& AndRHSV = cast<ConstantInt>(AndRHS)->getValue();
3442 // If there is only one bit set...
3443 if (isOneBitSet(cast<ConstantInt>(AndRHS))) {
3444 // Ok, at this point, we know that we are masking the result of the
3445 // ADD down to exactly one bit. If the constant we are adding has
3446 // no bits set below this bit, then we can eliminate the ADD.
3447 const APInt& AddRHS = cast<ConstantInt>(OpRHS)->getValue();
3449 // Check to see if any bits below the one bit set in AndRHSV are set.
3450 if ((AddRHS & (AndRHSV-1)) == 0) {
3451 // If not, the only thing that can effect the output of the AND is
3452 // the bit specified by AndRHSV. If that bit is set, the effect of
3453 // the XOR is to toggle the bit. If it is clear, then the ADD has
3455 if ((AddRHS & AndRHSV) == 0) { // Bit is not set, noop
3456 TheAnd.setOperand(0, X);
3459 // Pull the XOR out of the AND.
3460 Instruction *NewAnd = BinaryOperator::CreateAnd(X, AndRHS);
3461 InsertNewInstBefore(NewAnd, TheAnd);
3462 NewAnd->takeName(Op);
3463 return BinaryOperator::CreateXor(NewAnd, AndRHS);
3470 case Instruction::Shl: {
3471 // We know that the AND will not produce any of the bits shifted in, so if
3472 // the anded constant includes them, clear them now!
3474 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3475 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3476 APInt ShlMask(APInt::getHighBitsSet(BitWidth, BitWidth-OpRHSVal));
3477 ConstantInt *CI = ConstantInt::get(AndRHS->getValue() & ShlMask);
3479 if (CI->getValue() == ShlMask) {
3480 // Masking out bits that the shift already masks
3481 return ReplaceInstUsesWith(TheAnd, Op); // No need for the and.
3482 } else if (CI != AndRHS) { // Reducing bits set in and.
3483 TheAnd.setOperand(1, CI);
3488 case Instruction::LShr:
3490 // We know that the AND will not produce any of the bits shifted in, so if
3491 // the anded constant includes them, clear them now! This only applies to
3492 // unsigned shifts, because a signed shr may bring in set bits!
3494 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3495 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3496 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
3497 ConstantInt *CI = ConstantInt::get(AndRHS->getValue() & ShrMask);
3499 if (CI->getValue() == ShrMask) {
3500 // Masking out bits that the shift already masks.
3501 return ReplaceInstUsesWith(TheAnd, Op);
3502 } else if (CI != AndRHS) {
3503 TheAnd.setOperand(1, CI); // Reduce bits set in and cst.
3508 case Instruction::AShr:
3510 // See if this is shifting in some sign extension, then masking it out
3512 if (Op->hasOneUse()) {
3513 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3514 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3515 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
3516 Constant *C = ConstantInt::get(AndRHS->getValue() & ShrMask);
3517 if (C == AndRHS) { // Masking out bits shifted in.
3518 // (Val ashr C1) & C2 -> (Val lshr C1) & C2
3519 // Make the argument unsigned.
3520 Value *ShVal = Op->getOperand(0);
3521 ShVal = InsertNewInstBefore(
3522 BinaryOperator::CreateLShr(ShVal, OpRHS,
3523 Op->getName()), TheAnd);
3524 return BinaryOperator::CreateAnd(ShVal, AndRHS, TheAnd.getName());
3533 /// InsertRangeTest - Emit a computation of: (V >= Lo && V < Hi) if Inside is
3534 /// true, otherwise (V < Lo || V >= Hi). In pratice, we emit the more efficient
3535 /// (V-Lo) <u Hi-Lo. This method expects that Lo <= Hi. isSigned indicates
3536 /// whether to treat the V, Lo and HI as signed or not. IB is the location to
3537 /// insert new instructions.
3538 Instruction *InstCombiner::InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
3539 bool isSigned, bool Inside,
3541 assert(cast<ConstantInt>(ConstantExpr::getICmp((isSigned ?
3542 ICmpInst::ICMP_SLE:ICmpInst::ICMP_ULE), Lo, Hi))->getZExtValue() &&
3543 "Lo is not <= Hi in range emission code!");
3546 if (Lo == Hi) // Trivially false.
3547 return new ICmpInst(ICmpInst::ICMP_NE, V, V);
3549 // V >= Min && V < Hi --> V < Hi
3550 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
3551 ICmpInst::Predicate pred = (isSigned ?
3552 ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT);
3553 return new ICmpInst(pred, V, Hi);
3556 // Emit V-Lo <u Hi-Lo
3557 Constant *NegLo = ConstantExpr::getNeg(Lo);
3558 Instruction *Add = BinaryOperator::CreateAdd(V, NegLo, V->getName()+".off");
3559 InsertNewInstBefore(Add, IB);
3560 Constant *UpperBound = ConstantExpr::getAdd(NegLo, Hi);
3561 return new ICmpInst(ICmpInst::ICMP_ULT, Add, UpperBound);
3564 if (Lo == Hi) // Trivially true.
3565 return new ICmpInst(ICmpInst::ICMP_EQ, V, V);
3567 // V < Min || V >= Hi -> V > Hi-1
3568 Hi = SubOne(cast<ConstantInt>(Hi));
3569 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
3570 ICmpInst::Predicate pred = (isSigned ?
3571 ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT);
3572 return new ICmpInst(pred, V, Hi);
3575 // Emit V-Lo >u Hi-1-Lo
3576 // Note that Hi has already had one subtracted from it, above.
3577 ConstantInt *NegLo = cast<ConstantInt>(ConstantExpr::getNeg(Lo));
3578 Instruction *Add = BinaryOperator::CreateAdd(V, NegLo, V->getName()+".off");
3579 InsertNewInstBefore(Add, IB);
3580 Constant *LowerBound = ConstantExpr::getAdd(NegLo, Hi);
3581 return new ICmpInst(ICmpInst::ICMP_UGT, Add, LowerBound);
3584 // isRunOfOnes - Returns true iff Val consists of one contiguous run of 1s with
3585 // any number of 0s on either side. The 1s are allowed to wrap from LSB to
3586 // MSB, so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs. 0x0F0F0000 is
3587 // not, since all 1s are not contiguous.
3588 static bool isRunOfOnes(ConstantInt *Val, uint32_t &MB, uint32_t &ME) {
3589 const APInt& V = Val->getValue();
3590 uint32_t BitWidth = Val->getType()->getBitWidth();
3591 if (!APIntOps::isShiftedMask(BitWidth, V)) return false;
3593 // look for the first zero bit after the run of ones
3594 MB = BitWidth - ((V - 1) ^ V).countLeadingZeros();
3595 // look for the first non-zero bit
3596 ME = V.getActiveBits();
3600 /// FoldLogicalPlusAnd - This is part of an expression (LHS +/- RHS) & Mask,
3601 /// where isSub determines whether the operator is a sub. If we can fold one of
3602 /// the following xforms:
3604 /// ((A & N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == Mask
3605 /// ((A | N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3606 /// ((A ^ N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3608 /// return (A +/- B).
3610 Value *InstCombiner::FoldLogicalPlusAnd(Value *LHS, Value *RHS,
3611 ConstantInt *Mask, bool isSub,
3613 Instruction *LHSI = dyn_cast<Instruction>(LHS);
3614 if (!LHSI || LHSI->getNumOperands() != 2 ||
3615 !isa<ConstantInt>(LHSI->getOperand(1))) return 0;
3617 ConstantInt *N = cast<ConstantInt>(LHSI->getOperand(1));
3619 switch (LHSI->getOpcode()) {
3621 case Instruction::And:
3622 if (And(N, Mask) == Mask) {
3623 // If the AndRHS is a power of two minus one (0+1+), this is simple.
3624 if ((Mask->getValue().countLeadingZeros() +
3625 Mask->getValue().countPopulation()) ==
3626 Mask->getValue().getBitWidth())
3629 // Otherwise, if Mask is 0+1+0+, and if B is known to have the low 0+
3630 // part, we don't need any explicit masks to take them out of A. If that
3631 // is all N is, ignore it.
3632 uint32_t MB = 0, ME = 0;
3633 if (isRunOfOnes(Mask, MB, ME)) { // begin/end bit of run, inclusive
3634 uint32_t BitWidth = cast<IntegerType>(RHS->getType())->getBitWidth();
3635 APInt Mask(APInt::getLowBitsSet(BitWidth, MB-1));
3636 if (MaskedValueIsZero(RHS, Mask))
3641 case Instruction::Or:
3642 case Instruction::Xor:
3643 // If the AndRHS is a power of two minus one (0+1+), and N&Mask == 0
3644 if ((Mask->getValue().countLeadingZeros() +
3645 Mask->getValue().countPopulation()) == Mask->getValue().getBitWidth()
3646 && And(N, Mask)->isZero())
3653 New = BinaryOperator::CreateSub(LHSI->getOperand(0), RHS, "fold");
3655 New = BinaryOperator::CreateAdd(LHSI->getOperand(0), RHS, "fold");
3656 return InsertNewInstBefore(New, I);
3659 /// FoldAndOfICmps - Fold (icmp)&(icmp) if possible.
3660 Instruction *InstCombiner::FoldAndOfICmps(Instruction &I,
3661 ICmpInst *LHS, ICmpInst *RHS) {
3663 ConstantInt *LHSCst, *RHSCst;
3664 ICmpInst::Predicate LHSCC, RHSCC;
3666 // This only handles icmp of constants: (icmp1 A, C1) & (icmp2 B, C2).
3667 if (!match(LHS, m_ICmp(LHSCC, m_Value(Val), m_ConstantInt(LHSCst))) ||
3668 !match(RHS, m_ICmp(RHSCC, m_Value(Val2), m_ConstantInt(RHSCst))))
3671 // (icmp ult A, C) & (icmp ult B, C) --> (icmp ult (A|B), C)
3672 // where C is a power of 2
3673 if (LHSCst == RHSCst && LHSCC == RHSCC && LHSCC == ICmpInst::ICMP_ULT &&
3674 LHSCst->getValue().isPowerOf2()) {
3675 Instruction *NewOr = BinaryOperator::CreateOr(Val, Val2);
3676 InsertNewInstBefore(NewOr, I);
3677 return new ICmpInst(LHSCC, NewOr, LHSCst);
3680 // From here on, we only handle:
3681 // (icmp1 A, C1) & (icmp2 A, C2) --> something simpler.
3682 if (Val != Val2) return 0;
3684 // ICMP_[US][GL]E X, CST is folded to ICMP_[US][GL]T elsewhere.
3685 if (LHSCC == ICmpInst::ICMP_UGE || LHSCC == ICmpInst::ICMP_ULE ||
3686 RHSCC == ICmpInst::ICMP_UGE || RHSCC == ICmpInst::ICMP_ULE ||
3687 LHSCC == ICmpInst::ICMP_SGE || LHSCC == ICmpInst::ICMP_SLE ||
3688 RHSCC == ICmpInst::ICMP_SGE || RHSCC == ICmpInst::ICMP_SLE)
3691 // We can't fold (ugt x, C) & (sgt x, C2).
3692 if (!PredicatesFoldable(LHSCC, RHSCC))
3695 // Ensure that the larger constant is on the RHS.
3697 if (ICmpInst::isSignedPredicate(LHSCC) ||
3698 (ICmpInst::isEquality(LHSCC) &&
3699 ICmpInst::isSignedPredicate(RHSCC)))
3700 ShouldSwap = LHSCst->getValue().sgt(RHSCst->getValue());
3702 ShouldSwap = LHSCst->getValue().ugt(RHSCst->getValue());
3705 std::swap(LHS, RHS);
3706 std::swap(LHSCst, RHSCst);
3707 std::swap(LHSCC, RHSCC);
3710 // At this point, we know we have have two icmp instructions
3711 // comparing a value against two constants and and'ing the result
3712 // together. Because of the above check, we know that we only have
3713 // icmp eq, icmp ne, icmp [su]lt, and icmp [SU]gt here. We also know
3714 // (from the FoldICmpLogical check above), that the two constants
3715 // are not equal and that the larger constant is on the RHS
3716 assert(LHSCst != RHSCst && "Compares not folded above?");
3719 default: assert(0 && "Unknown integer condition code!");
3720 case ICmpInst::ICMP_EQ:
3722 default: assert(0 && "Unknown integer condition code!");
3723 case ICmpInst::ICMP_EQ: // (X == 13 & X == 15) -> false
3724 case ICmpInst::ICMP_UGT: // (X == 13 & X > 15) -> false
3725 case ICmpInst::ICMP_SGT: // (X == 13 & X > 15) -> false
3726 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3727 case ICmpInst::ICMP_NE: // (X == 13 & X != 15) -> X == 13
3728 case ICmpInst::ICMP_ULT: // (X == 13 & X < 15) -> X == 13
3729 case ICmpInst::ICMP_SLT: // (X == 13 & X < 15) -> X == 13
3730 return ReplaceInstUsesWith(I, LHS);
3732 case ICmpInst::ICMP_NE:
3734 default: assert(0 && "Unknown integer condition code!");
3735 case ICmpInst::ICMP_ULT:
3736 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X u< 14) -> X < 13
3737 return new ICmpInst(ICmpInst::ICMP_ULT, Val, LHSCst);
3738 break; // (X != 13 & X u< 15) -> no change
3739 case ICmpInst::ICMP_SLT:
3740 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X s< 14) -> X < 13
3741 return new ICmpInst(ICmpInst::ICMP_SLT, Val, LHSCst);
3742 break; // (X != 13 & X s< 15) -> no change
3743 case ICmpInst::ICMP_EQ: // (X != 13 & X == 15) -> X == 15
3744 case ICmpInst::ICMP_UGT: // (X != 13 & X u> 15) -> X u> 15
3745 case ICmpInst::ICMP_SGT: // (X != 13 & X s> 15) -> X s> 15
3746 return ReplaceInstUsesWith(I, RHS);
3747 case ICmpInst::ICMP_NE:
3748 if (LHSCst == SubOne(RHSCst)){// (X != 13 & X != 14) -> X-13 >u 1
3749 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
3750 Instruction *Add = BinaryOperator::CreateAdd(Val, AddCST,
3751 Val->getName()+".off");
3752 InsertNewInstBefore(Add, I);
3753 return new ICmpInst(ICmpInst::ICMP_UGT, Add,
3754 ConstantInt::get(Add->getType(), 1));
3756 break; // (X != 13 & X != 15) -> no change
3759 case ICmpInst::ICMP_ULT:
3761 default: assert(0 && "Unknown integer condition code!");
3762 case ICmpInst::ICMP_EQ: // (X u< 13 & X == 15) -> false
3763 case ICmpInst::ICMP_UGT: // (X u< 13 & X u> 15) -> false
3764 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3765 case ICmpInst::ICMP_SGT: // (X u< 13 & X s> 15) -> no change
3767 case ICmpInst::ICMP_NE: // (X u< 13 & X != 15) -> X u< 13
3768 case ICmpInst::ICMP_ULT: // (X u< 13 & X u< 15) -> X u< 13
3769 return ReplaceInstUsesWith(I, LHS);
3770 case ICmpInst::ICMP_SLT: // (X u< 13 & X s< 15) -> no change
3774 case ICmpInst::ICMP_SLT:
3776 default: assert(0 && "Unknown integer condition code!");
3777 case ICmpInst::ICMP_EQ: // (X s< 13 & X == 15) -> false
3778 case ICmpInst::ICMP_SGT: // (X s< 13 & X s> 15) -> false
3779 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3780 case ICmpInst::ICMP_UGT: // (X s< 13 & X u> 15) -> no change
3782 case ICmpInst::ICMP_NE: // (X s< 13 & X != 15) -> X < 13
3783 case ICmpInst::ICMP_SLT: // (X s< 13 & X s< 15) -> X < 13
3784 return ReplaceInstUsesWith(I, LHS);
3785 case ICmpInst::ICMP_ULT: // (X s< 13 & X u< 15) -> no change
3789 case ICmpInst::ICMP_UGT:
3791 default: assert(0 && "Unknown integer condition code!");
3792 case ICmpInst::ICMP_EQ: // (X u> 13 & X == 15) -> X == 15
3793 case ICmpInst::ICMP_UGT: // (X u> 13 & X u> 15) -> X u> 15
3794 return ReplaceInstUsesWith(I, RHS);
3795 case ICmpInst::ICMP_SGT: // (X u> 13 & X s> 15) -> no change
3797 case ICmpInst::ICMP_NE:
3798 if (RHSCst == AddOne(LHSCst)) // (X u> 13 & X != 14) -> X u> 14
3799 return new ICmpInst(LHSCC, Val, RHSCst);
3800 break; // (X u> 13 & X != 15) -> no change
3801 case ICmpInst::ICMP_ULT: // (X u> 13 & X u< 15) -> (X-14) <u 1
3802 return InsertRangeTest(Val, AddOne(LHSCst), RHSCst, false, true, I);
3803 case ICmpInst::ICMP_SLT: // (X u> 13 & X s< 15) -> no change
3807 case ICmpInst::ICMP_SGT:
3809 default: assert(0 && "Unknown integer condition code!");
3810 case ICmpInst::ICMP_EQ: // (X s> 13 & X == 15) -> X == 15
3811 case ICmpInst::ICMP_SGT: // (X s> 13 & X s> 15) -> X s> 15
3812 return ReplaceInstUsesWith(I, RHS);
3813 case ICmpInst::ICMP_UGT: // (X s> 13 & X u> 15) -> no change
3815 case ICmpInst::ICMP_NE:
3816 if (RHSCst == AddOne(LHSCst)) // (X s> 13 & X != 14) -> X s> 14
3817 return new ICmpInst(LHSCC, Val, RHSCst);
3818 break; // (X s> 13 & X != 15) -> no change
3819 case ICmpInst::ICMP_SLT: // (X s> 13 & X s< 15) -> (X-14) s< 1
3820 return InsertRangeTest(Val, AddOne(LHSCst), RHSCst, true, true, I);
3821 case ICmpInst::ICMP_ULT: // (X s> 13 & X u< 15) -> no change
3831 Instruction *InstCombiner::visitAnd(BinaryOperator &I) {
3832 bool Changed = SimplifyCommutative(I);
3833 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3835 if (isa<UndefValue>(Op1)) // X & undef -> 0
3836 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3840 return ReplaceInstUsesWith(I, Op1);
3842 // See if we can simplify any instructions used by the instruction whose sole
3843 // purpose is to compute bits we don't care about.
3844 if (!isa<VectorType>(I.getType())) {
3845 if (SimplifyDemandedInstructionBits(I))
3848 if (ConstantVector *CP = dyn_cast<ConstantVector>(Op1)) {
3849 if (CP->isAllOnesValue()) // X & <-1,-1> -> X
3850 return ReplaceInstUsesWith(I, I.getOperand(0));
3851 } else if (isa<ConstantAggregateZero>(Op1)) {
3852 return ReplaceInstUsesWith(I, Op1); // X & <0,0> -> <0,0>
3856 if (ConstantInt *AndRHS = dyn_cast<ConstantInt>(Op1)) {
3857 const APInt& AndRHSMask = AndRHS->getValue();
3858 APInt NotAndRHS(~AndRHSMask);
3860 // Optimize a variety of ((val OP C1) & C2) combinations...
3861 if (isa<BinaryOperator>(Op0)) {
3862 Instruction *Op0I = cast<Instruction>(Op0);
3863 Value *Op0LHS = Op0I->getOperand(0);
3864 Value *Op0RHS = Op0I->getOperand(1);
3865 switch (Op0I->getOpcode()) {
3866 case Instruction::Xor:
3867 case Instruction::Or:
3868 // If the mask is only needed on one incoming arm, push it up.
3869 if (Op0I->hasOneUse()) {
3870 if (MaskedValueIsZero(Op0LHS, NotAndRHS)) {
3871 // Not masking anything out for the LHS, move to RHS.
3872 Instruction *NewRHS = BinaryOperator::CreateAnd(Op0RHS, AndRHS,
3873 Op0RHS->getName()+".masked");
3874 InsertNewInstBefore(NewRHS, I);
3875 return BinaryOperator::Create(
3876 cast<BinaryOperator>(Op0I)->getOpcode(), Op0LHS, NewRHS);
3878 if (!isa<Constant>(Op0RHS) &&
3879 MaskedValueIsZero(Op0RHS, NotAndRHS)) {
3880 // Not masking anything out for the RHS, move to LHS.
3881 Instruction *NewLHS = BinaryOperator::CreateAnd(Op0LHS, AndRHS,
3882 Op0LHS->getName()+".masked");
3883 InsertNewInstBefore(NewLHS, I);
3884 return BinaryOperator::Create(
3885 cast<BinaryOperator>(Op0I)->getOpcode(), NewLHS, Op0RHS);
3890 case Instruction::Add:
3891 // ((A & N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == AndRHS.
3892 // ((A | N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
3893 // ((A ^ N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
3894 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, false, I))
3895 return BinaryOperator::CreateAnd(V, AndRHS);
3896 if (Value *V = FoldLogicalPlusAnd(Op0RHS, Op0LHS, AndRHS, false, I))
3897 return BinaryOperator::CreateAnd(V, AndRHS); // Add commutes
3900 case Instruction::Sub:
3901 // ((A & N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == AndRHS.
3902 // ((A | N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
3903 // ((A ^ N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
3904 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, true, I))
3905 return BinaryOperator::CreateAnd(V, AndRHS);
3907 // (A - N) & AndRHS -> -N & AndRHS iff A&AndRHS==0 and AndRHS
3908 // has 1's for all bits that the subtraction with A might affect.
3909 if (Op0I->hasOneUse()) {
3910 uint32_t BitWidth = AndRHSMask.getBitWidth();
3911 uint32_t Zeros = AndRHSMask.countLeadingZeros();
3912 APInt Mask = APInt::getLowBitsSet(BitWidth, BitWidth - Zeros);
3914 ConstantInt *A = dyn_cast<ConstantInt>(Op0LHS);
3915 if (!(A && A->isZero()) && // avoid infinite recursion.
3916 MaskedValueIsZero(Op0LHS, Mask)) {
3917 Instruction *NewNeg = BinaryOperator::CreateNeg(Op0RHS);
3918 InsertNewInstBefore(NewNeg, I);
3919 return BinaryOperator::CreateAnd(NewNeg, AndRHS);
3924 case Instruction::Shl:
3925 case Instruction::LShr:
3926 // (1 << x) & 1 --> zext(x == 0)
3927 // (1 >> x) & 1 --> zext(x == 0)
3928 if (AndRHSMask == 1 && Op0LHS == AndRHS) {
3929 Instruction *NewICmp = new ICmpInst(ICmpInst::ICMP_EQ, Op0RHS,
3930 Constant::getNullValue(I.getType()));
3931 InsertNewInstBefore(NewICmp, I);
3932 return new ZExtInst(NewICmp, I.getType());
3937 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
3938 if (Instruction *Res = OptAndOp(Op0I, Op0CI, AndRHS, I))
3940 } else if (CastInst *CI = dyn_cast<CastInst>(Op0)) {
3941 // If this is an integer truncation or change from signed-to-unsigned, and
3942 // if the source is an and/or with immediate, transform it. This
3943 // frequently occurs for bitfield accesses.
3944 if (Instruction *CastOp = dyn_cast<Instruction>(CI->getOperand(0))) {
3945 if ((isa<TruncInst>(CI) || isa<BitCastInst>(CI)) &&
3946 CastOp->getNumOperands() == 2)
3947 if (ConstantInt *AndCI = dyn_cast<ConstantInt>(CastOp->getOperand(1))) {
3948 if (CastOp->getOpcode() == Instruction::And) {
3949 // Change: and (cast (and X, C1) to T), C2
3950 // into : and (cast X to T), trunc_or_bitcast(C1)&C2
3951 // This will fold the two constants together, which may allow
3952 // other simplifications.
3953 Instruction *NewCast = CastInst::CreateTruncOrBitCast(
3954 CastOp->getOperand(0), I.getType(),
3955 CastOp->getName()+".shrunk");
3956 NewCast = InsertNewInstBefore(NewCast, I);
3957 // trunc_or_bitcast(C1)&C2
3958 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
3959 C3 = ConstantExpr::getAnd(C3, AndRHS);
3960 return BinaryOperator::CreateAnd(NewCast, C3);
3961 } else if (CastOp->getOpcode() == Instruction::Or) {
3962 // Change: and (cast (or X, C1) to T), C2
3963 // into : trunc(C1)&C2 iff trunc(C1)&C2 == C2
3964 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
3965 if (ConstantExpr::getAnd(C3, AndRHS) == AndRHS) // trunc(C1)&C2
3966 return ReplaceInstUsesWith(I, AndRHS);
3972 // Try to fold constant and into select arguments.
3973 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3974 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3976 if (isa<PHINode>(Op0))
3977 if (Instruction *NV = FoldOpIntoPhi(I))
3981 Value *Op0NotVal = dyn_castNotVal(Op0);
3982 Value *Op1NotVal = dyn_castNotVal(Op1);
3984 if (Op0NotVal == Op1 || Op1NotVal == Op0) // A & ~A == ~A & A == 0
3985 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3987 // (~A & ~B) == (~(A | B)) - De Morgan's Law
3988 if (Op0NotVal && Op1NotVal && isOnlyUse(Op0) && isOnlyUse(Op1)) {
3989 Instruction *Or = BinaryOperator::CreateOr(Op0NotVal, Op1NotVal,
3990 I.getName()+".demorgan");
3991 InsertNewInstBefore(Or, I);
3992 return BinaryOperator::CreateNot(Or);
3996 Value *A = 0, *B = 0, *C = 0, *D = 0;
3997 if (match(Op0, m_Or(m_Value(A), m_Value(B)))) {
3998 if (A == Op1 || B == Op1) // (A | ?) & A --> A
3999 return ReplaceInstUsesWith(I, Op1);
4001 // (A|B) & ~(A&B) -> A^B
4002 if (match(Op1, m_Not(m_And(m_Value(C), m_Value(D))))) {
4003 if ((A == C && B == D) || (A == D && B == C))
4004 return BinaryOperator::CreateXor(A, B);
4008 if (match(Op1, m_Or(m_Value(A), m_Value(B)))) {
4009 if (A == Op0 || B == Op0) // A & (A | ?) --> A
4010 return ReplaceInstUsesWith(I, Op0);
4012 // ~(A&B) & (A|B) -> A^B
4013 if (match(Op0, m_Not(m_And(m_Value(C), m_Value(D))))) {
4014 if ((A == C && B == D) || (A == D && B == C))
4015 return BinaryOperator::CreateXor(A, B);
4019 if (Op0->hasOneUse() &&
4020 match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
4021 if (A == Op1) { // (A^B)&A -> A&(A^B)
4022 I.swapOperands(); // Simplify below
4023 std::swap(Op0, Op1);
4024 } else if (B == Op1) { // (A^B)&B -> B&(B^A)
4025 cast<BinaryOperator>(Op0)->swapOperands();
4026 I.swapOperands(); // Simplify below
4027 std::swap(Op0, Op1);
4031 if (Op1->hasOneUse() &&
4032 match(Op1, m_Xor(m_Value(A), m_Value(B)))) {
4033 if (B == Op0) { // B&(A^B) -> B&(B^A)
4034 cast<BinaryOperator>(Op1)->swapOperands();
4037 if (A == Op0) { // A&(A^B) -> A & ~B
4038 Instruction *NotB = BinaryOperator::CreateNot(B, "tmp");
4039 InsertNewInstBefore(NotB, I);
4040 return BinaryOperator::CreateAnd(A, NotB);
4044 // (A&((~A)|B)) -> A&B
4045 if (match(Op0, m_Or(m_Not(m_Specific(Op1)), m_Value(A))) ||
4046 match(Op0, m_Or(m_Value(A), m_Not(m_Specific(Op1)))))
4047 return BinaryOperator::CreateAnd(A, Op1);
4048 if (match(Op1, m_Or(m_Not(m_Specific(Op0)), m_Value(A))) ||
4049 match(Op1, m_Or(m_Value(A), m_Not(m_Specific(Op0)))))
4050 return BinaryOperator::CreateAnd(A, Op0);
4053 if (ICmpInst *RHS = dyn_cast<ICmpInst>(Op1)) {
4054 // (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
4055 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
4058 if (ICmpInst *LHS = dyn_cast<ICmpInst>(Op0))
4059 if (Instruction *Res = FoldAndOfICmps(I, LHS, RHS))
4063 // fold (and (cast A), (cast B)) -> (cast (and A, B))
4064 if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
4065 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4066 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind ?
4067 const Type *SrcTy = Op0C->getOperand(0)->getType();
4068 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
4069 // Only do this if the casts both really cause code to be generated.
4070 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4072 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4074 Instruction *NewOp = BinaryOperator::CreateAnd(Op0C->getOperand(0),
4075 Op1C->getOperand(0),
4077 InsertNewInstBefore(NewOp, I);
4078 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
4082 // (X >> Z) & (Y >> Z) -> (X&Y) >> Z for all shifts.
4083 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
4084 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
4085 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
4086 SI0->getOperand(1) == SI1->getOperand(1) &&
4087 (SI0->hasOneUse() || SI1->hasOneUse())) {
4088 Instruction *NewOp =
4089 InsertNewInstBefore(BinaryOperator::CreateAnd(SI0->getOperand(0),
4091 SI0->getName()), I);
4092 return BinaryOperator::Create(SI1->getOpcode(), NewOp,
4093 SI1->getOperand(1));
4097 // If and'ing two fcmp, try combine them into one.
4098 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0))) {
4099 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1))) {
4100 if (LHS->getPredicate() == FCmpInst::FCMP_ORD &&
4101 RHS->getPredicate() == FCmpInst::FCMP_ORD) {
4102 // (fcmp ord x, c) & (fcmp ord y, c) -> (fcmp ord x, y)
4103 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
4104 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
4105 // If either of the constants are nans, then the whole thing returns
4107 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
4108 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4109 return new FCmpInst(FCmpInst::FCMP_ORD, LHS->getOperand(0),
4110 RHS->getOperand(0));
4113 Value *Op0LHS, *Op0RHS, *Op1LHS, *Op1RHS;
4114 FCmpInst::Predicate Op0CC, Op1CC;
4115 if (match(Op0, m_FCmp(Op0CC, m_Value(Op0LHS), m_Value(Op0RHS))) &&
4116 match(Op1, m_FCmp(Op1CC, m_Value(Op1LHS), m_Value(Op1RHS)))) {
4117 if (Op0LHS == Op1RHS && Op0RHS == Op1LHS) {
4118 // Swap RHS operands to match LHS.
4119 Op1CC = FCmpInst::getSwappedPredicate(Op1CC);
4120 std::swap(Op1LHS, Op1RHS);
4122 if (Op0LHS == Op1LHS && Op0RHS == Op1RHS) {
4123 // Simplify (fcmp cc0 x, y) & (fcmp cc1 x, y).
4125 return new FCmpInst((FCmpInst::Predicate)Op0CC, Op0LHS, Op0RHS);
4126 else if (Op0CC == FCmpInst::FCMP_FALSE ||
4127 Op1CC == FCmpInst::FCMP_FALSE)
4128 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4129 else if (Op0CC == FCmpInst::FCMP_TRUE)
4130 return ReplaceInstUsesWith(I, Op1);
4131 else if (Op1CC == FCmpInst::FCMP_TRUE)
4132 return ReplaceInstUsesWith(I, Op0);
4135 unsigned Op0Pred = getFCmpCode(Op0CC, Op0Ordered);
4136 unsigned Op1Pred = getFCmpCode(Op1CC, Op1Ordered);
4138 std::swap(Op0, Op1);
4139 std::swap(Op0Pred, Op1Pred);
4140 std::swap(Op0Ordered, Op1Ordered);
4143 // uno && ueq -> uno && (uno || eq) -> ueq
4144 // ord && olt -> ord && (ord && lt) -> olt
4145 if (Op0Ordered == Op1Ordered)
4146 return ReplaceInstUsesWith(I, Op1);
4147 // uno && oeq -> uno && (ord && eq) -> false
4148 // uno && ord -> false
4150 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4151 // ord && ueq -> ord && (uno || eq) -> oeq
4152 return cast<Instruction>(getFCmpValue(true, Op1Pred,
4161 return Changed ? &I : 0;
4164 /// CollectBSwapParts - Analyze the specified subexpression and see if it is
4165 /// capable of providing pieces of a bswap. The subexpression provides pieces
4166 /// of a bswap if it is proven that each of the non-zero bytes in the output of
4167 /// the expression came from the corresponding "byte swapped" byte in some other
4168 /// value. For example, if the current subexpression is "(shl i32 %X, 24)" then
4169 /// we know that the expression deposits the low byte of %X into the high byte
4170 /// of the bswap result and that all other bytes are zero. This expression is
4171 /// accepted, the high byte of ByteValues is set to X to indicate a correct
4174 /// This function returns true if the match was unsuccessful and false if so.
4175 /// On entry to the function the "OverallLeftShift" is a signed integer value
4176 /// indicating the number of bytes that the subexpression is later shifted. For
4177 /// example, if the expression is later right shifted by 16 bits, the
4178 /// OverallLeftShift value would be -2 on entry. This is used to specify which
4179 /// byte of ByteValues is actually being set.
4181 /// Similarly, ByteMask is a bitmask where a bit is clear if its corresponding
4182 /// byte is masked to zero by a user. For example, in (X & 255), X will be
4183 /// processed with a bytemask of 1. Because bytemask is 32-bits, this limits
4184 /// this function to working on up to 32-byte (256 bit) values. ByteMask is
4185 /// always in the local (OverallLeftShift) coordinate space.
4187 static bool CollectBSwapParts(Value *V, int OverallLeftShift, uint32_t ByteMask,
4188 SmallVector<Value*, 8> &ByteValues) {
4189 if (Instruction *I = dyn_cast<Instruction>(V)) {
4190 // If this is an or instruction, it may be an inner node of the bswap.
4191 if (I->getOpcode() == Instruction::Or) {
4192 return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
4194 CollectBSwapParts(I->getOperand(1), OverallLeftShift, ByteMask,
4198 // If this is a logical shift by a constant multiple of 8, recurse with
4199 // OverallLeftShift and ByteMask adjusted.
4200 if (I->isLogicalShift() && isa<ConstantInt>(I->getOperand(1))) {
4202 cast<ConstantInt>(I->getOperand(1))->getLimitedValue(~0U);
4203 // Ensure the shift amount is defined and of a byte value.
4204 if ((ShAmt & 7) || (ShAmt > 8*ByteValues.size()))
4207 unsigned ByteShift = ShAmt >> 3;
4208 if (I->getOpcode() == Instruction::Shl) {
4209 // X << 2 -> collect(X, +2)
4210 OverallLeftShift += ByteShift;
4211 ByteMask >>= ByteShift;
4213 // X >>u 2 -> collect(X, -2)
4214 OverallLeftShift -= ByteShift;
4215 ByteMask <<= ByteShift;
4216 ByteMask &= (~0U >> (32-ByteValues.size()));
4219 if (OverallLeftShift >= (int)ByteValues.size()) return true;
4220 if (OverallLeftShift <= -(int)ByteValues.size()) return true;
4222 return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
4226 // If this is a logical 'and' with a mask that clears bytes, clear the
4227 // corresponding bytes in ByteMask.
4228 if (I->getOpcode() == Instruction::And &&
4229 isa<ConstantInt>(I->getOperand(1))) {
4230 // Scan every byte of the and mask, seeing if the byte is either 0 or 255.
4231 unsigned NumBytes = ByteValues.size();
4232 APInt Byte(I->getType()->getPrimitiveSizeInBits(), 255);
4233 const APInt &AndMask = cast<ConstantInt>(I->getOperand(1))->getValue();
4235 for (unsigned i = 0; i != NumBytes; ++i, Byte <<= 8) {
4236 // If this byte is masked out by a later operation, we don't care what
4238 if ((ByteMask & (1 << i)) == 0)
4241 // If the AndMask is all zeros for this byte, clear the bit.
4242 APInt MaskB = AndMask & Byte;
4244 ByteMask &= ~(1U << i);
4248 // If the AndMask is not all ones for this byte, it's not a bytezap.
4252 // Otherwise, this byte is kept.
4255 return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
4260 // Okay, we got to something that isn't a shift, 'or' or 'and'. This must be
4261 // the input value to the bswap. Some observations: 1) if more than one byte
4262 // is demanded from this input, then it could not be successfully assembled
4263 // into a byteswap. At least one of the two bytes would not be aligned with
4264 // their ultimate destination.
4265 if (!isPowerOf2_32(ByteMask)) return true;
4266 unsigned InputByteNo = CountTrailingZeros_32(ByteMask);
4268 // 2) The input and ultimate destinations must line up: if byte 3 of an i32
4269 // is demanded, it needs to go into byte 0 of the result. This means that the
4270 // byte needs to be shifted until it lands in the right byte bucket. The
4271 // shift amount depends on the position: if the byte is coming from the high
4272 // part of the value (e.g. byte 3) then it must be shifted right. If from the
4273 // low part, it must be shifted left.
4274 unsigned DestByteNo = InputByteNo + OverallLeftShift;
4275 if (InputByteNo < ByteValues.size()/2) {
4276 if (ByteValues.size()-1-DestByteNo != InputByteNo)
4279 if (ByteValues.size()-1-DestByteNo != InputByteNo)
4283 // If the destination byte value is already defined, the values are or'd
4284 // together, which isn't a bswap (unless it's an or of the same bits).
4285 if (ByteValues[DestByteNo] && ByteValues[DestByteNo] != V)
4287 ByteValues[DestByteNo] = V;
4291 /// MatchBSwap - Given an OR instruction, check to see if this is a bswap idiom.
4292 /// If so, insert the new bswap intrinsic and return it.
4293 Instruction *InstCombiner::MatchBSwap(BinaryOperator &I) {
4294 const IntegerType *ITy = dyn_cast<IntegerType>(I.getType());
4295 if (!ITy || ITy->getBitWidth() % 16 ||
4296 // ByteMask only allows up to 32-byte values.
4297 ITy->getBitWidth() > 32*8)
4298 return 0; // Can only bswap pairs of bytes. Can't do vectors.
4300 /// ByteValues - For each byte of the result, we keep track of which value
4301 /// defines each byte.
4302 SmallVector<Value*, 8> ByteValues;
4303 ByteValues.resize(ITy->getBitWidth()/8);
4305 // Try to find all the pieces corresponding to the bswap.
4306 uint32_t ByteMask = ~0U >> (32-ByteValues.size());
4307 if (CollectBSwapParts(&I, 0, ByteMask, ByteValues))
4310 // Check to see if all of the bytes come from the same value.
4311 Value *V = ByteValues[0];
4312 if (V == 0) return 0; // Didn't find a byte? Must be zero.
4314 // Check to make sure that all of the bytes come from the same value.
4315 for (unsigned i = 1, e = ByteValues.size(); i != e; ++i)
4316 if (ByteValues[i] != V)
4318 const Type *Tys[] = { ITy };
4319 Module *M = I.getParent()->getParent()->getParent();
4320 Function *F = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 1);
4321 return CallInst::Create(F, V);
4324 /// MatchSelectFromAndOr - We have an expression of the form (A&C)|(B&D). Check
4325 /// If A is (cond?-1:0) and either B or D is ~(cond?-1,0) or (cond?0,-1), then
4326 /// we can simplify this expression to "cond ? C : D or B".
4327 static Instruction *MatchSelectFromAndOr(Value *A, Value *B,
4328 Value *C, Value *D) {
4329 // If A is not a select of -1/0, this cannot match.
4331 if (!match(A, m_SelectCst<-1, 0>(m_Value(Cond))))
4334 // ((cond?-1:0)&C) | (B&(cond?0:-1)) -> cond ? C : B.
4335 if (match(D, m_SelectCst<0, -1>(m_Specific(Cond))))
4336 return SelectInst::Create(Cond, C, B);
4337 if (match(D, m_Not(m_SelectCst<-1, 0>(m_Specific(Cond)))))
4338 return SelectInst::Create(Cond, C, B);
4339 // ((cond?-1:0)&C) | ((cond?0:-1)&D) -> cond ? C : D.
4340 if (match(B, m_SelectCst<0, -1>(m_Specific(Cond))))
4341 return SelectInst::Create(Cond, C, D);
4342 if (match(B, m_Not(m_SelectCst<-1, 0>(m_Specific(Cond)))))
4343 return SelectInst::Create(Cond, C, D);
4347 /// FoldOrOfICmps - Fold (icmp)|(icmp) if possible.
4348 Instruction *InstCombiner::FoldOrOfICmps(Instruction &I,
4349 ICmpInst *LHS, ICmpInst *RHS) {
4351 ConstantInt *LHSCst, *RHSCst;
4352 ICmpInst::Predicate LHSCC, RHSCC;
4354 // This only handles icmp of constants: (icmp1 A, C1) | (icmp2 B, C2).
4355 if (!match(LHS, m_ICmp(LHSCC, m_Value(Val), m_ConstantInt(LHSCst))) ||
4356 !match(RHS, m_ICmp(RHSCC, m_Value(Val2), m_ConstantInt(RHSCst))))
4359 // From here on, we only handle:
4360 // (icmp1 A, C1) | (icmp2 A, C2) --> something simpler.
4361 if (Val != Val2) return 0;
4363 // ICMP_[US][GL]E X, CST is folded to ICMP_[US][GL]T elsewhere.
4364 if (LHSCC == ICmpInst::ICMP_UGE || LHSCC == ICmpInst::ICMP_ULE ||
4365 RHSCC == ICmpInst::ICMP_UGE || RHSCC == ICmpInst::ICMP_ULE ||
4366 LHSCC == ICmpInst::ICMP_SGE || LHSCC == ICmpInst::ICMP_SLE ||
4367 RHSCC == ICmpInst::ICMP_SGE || RHSCC == ICmpInst::ICMP_SLE)
4370 // We can't fold (ugt x, C) | (sgt x, C2).
4371 if (!PredicatesFoldable(LHSCC, RHSCC))
4374 // Ensure that the larger constant is on the RHS.
4376 if (ICmpInst::isSignedPredicate(LHSCC) ||
4377 (ICmpInst::isEquality(LHSCC) &&
4378 ICmpInst::isSignedPredicate(RHSCC)))
4379 ShouldSwap = LHSCst->getValue().sgt(RHSCst->getValue());
4381 ShouldSwap = LHSCst->getValue().ugt(RHSCst->getValue());
4384 std::swap(LHS, RHS);
4385 std::swap(LHSCst, RHSCst);
4386 std::swap(LHSCC, RHSCC);
4389 // At this point, we know we have have two icmp instructions
4390 // comparing a value against two constants and or'ing the result
4391 // together. Because of the above check, we know that we only have
4392 // ICMP_EQ, ICMP_NE, ICMP_LT, and ICMP_GT here. We also know (from the
4393 // FoldICmpLogical check above), that the two constants are not
4395 assert(LHSCst != RHSCst && "Compares not folded above?");
4398 default: assert(0 && "Unknown integer condition code!");
4399 case ICmpInst::ICMP_EQ:
4401 default: assert(0 && "Unknown integer condition code!");
4402 case ICmpInst::ICMP_EQ:
4403 if (LHSCst == SubOne(RHSCst)) { // (X == 13 | X == 14) -> X-13 <u 2
4404 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
4405 Instruction *Add = BinaryOperator::CreateAdd(Val, AddCST,
4406 Val->getName()+".off");
4407 InsertNewInstBefore(Add, I);
4408 AddCST = Subtract(AddOne(RHSCst), LHSCst);
4409 return new ICmpInst(ICmpInst::ICMP_ULT, Add, AddCST);
4411 break; // (X == 13 | X == 15) -> no change
4412 case ICmpInst::ICMP_UGT: // (X == 13 | X u> 14) -> no change
4413 case ICmpInst::ICMP_SGT: // (X == 13 | X s> 14) -> no change
4415 case ICmpInst::ICMP_NE: // (X == 13 | X != 15) -> X != 15
4416 case ICmpInst::ICMP_ULT: // (X == 13 | X u< 15) -> X u< 15
4417 case ICmpInst::ICMP_SLT: // (X == 13 | X s< 15) -> X s< 15
4418 return ReplaceInstUsesWith(I, RHS);
4421 case ICmpInst::ICMP_NE:
4423 default: assert(0 && "Unknown integer condition code!");
4424 case ICmpInst::ICMP_EQ: // (X != 13 | X == 15) -> X != 13
4425 case ICmpInst::ICMP_UGT: // (X != 13 | X u> 15) -> X != 13
4426 case ICmpInst::ICMP_SGT: // (X != 13 | X s> 15) -> X != 13
4427 return ReplaceInstUsesWith(I, LHS);
4428 case ICmpInst::ICMP_NE: // (X != 13 | X != 15) -> true
4429 case ICmpInst::ICMP_ULT: // (X != 13 | X u< 15) -> true
4430 case ICmpInst::ICMP_SLT: // (X != 13 | X s< 15) -> true
4431 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4434 case ICmpInst::ICMP_ULT:
4436 default: assert(0 && "Unknown integer condition code!");
4437 case ICmpInst::ICMP_EQ: // (X u< 13 | X == 14) -> no change
4439 case ICmpInst::ICMP_UGT: // (X u< 13 | X u> 15) -> (X-13) u> 2
4440 // If RHSCst is [us]MAXINT, it is always false. Not handling
4441 // this can cause overflow.
4442 if (RHSCst->isMaxValue(false))
4443 return ReplaceInstUsesWith(I, LHS);
4444 return InsertRangeTest(Val, LHSCst, AddOne(RHSCst), false, false, I);
4445 case ICmpInst::ICMP_SGT: // (X u< 13 | X s> 15) -> no change
4447 case ICmpInst::ICMP_NE: // (X u< 13 | X != 15) -> X != 15
4448 case ICmpInst::ICMP_ULT: // (X u< 13 | X u< 15) -> X u< 15
4449 return ReplaceInstUsesWith(I, RHS);
4450 case ICmpInst::ICMP_SLT: // (X u< 13 | X s< 15) -> no change
4454 case ICmpInst::ICMP_SLT:
4456 default: assert(0 && "Unknown integer condition code!");
4457 case ICmpInst::ICMP_EQ: // (X s< 13 | X == 14) -> no change
4459 case ICmpInst::ICMP_SGT: // (X s< 13 | X s> 15) -> (X-13) s> 2
4460 // If RHSCst is [us]MAXINT, it is always false. Not handling
4461 // this can cause overflow.
4462 if (RHSCst->isMaxValue(true))
4463 return ReplaceInstUsesWith(I, LHS);
4464 return InsertRangeTest(Val, LHSCst, AddOne(RHSCst), true, false, I);
4465 case ICmpInst::ICMP_UGT: // (X s< 13 | X u> 15) -> no change
4467 case ICmpInst::ICMP_NE: // (X s< 13 | X != 15) -> X != 15
4468 case ICmpInst::ICMP_SLT: // (X s< 13 | X s< 15) -> X s< 15
4469 return ReplaceInstUsesWith(I, RHS);
4470 case ICmpInst::ICMP_ULT: // (X s< 13 | X u< 15) -> no change
4474 case ICmpInst::ICMP_UGT:
4476 default: assert(0 && "Unknown integer condition code!");
4477 case ICmpInst::ICMP_EQ: // (X u> 13 | X == 15) -> X u> 13
4478 case ICmpInst::ICMP_UGT: // (X u> 13 | X u> 15) -> X u> 13
4479 return ReplaceInstUsesWith(I, LHS);
4480 case ICmpInst::ICMP_SGT: // (X u> 13 | X s> 15) -> no change
4482 case ICmpInst::ICMP_NE: // (X u> 13 | X != 15) -> true
4483 case ICmpInst::ICMP_ULT: // (X u> 13 | X u< 15) -> true
4484 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4485 case ICmpInst::ICMP_SLT: // (X u> 13 | X s< 15) -> no change
4489 case ICmpInst::ICMP_SGT:
4491 default: assert(0 && "Unknown integer condition code!");
4492 case ICmpInst::ICMP_EQ: // (X s> 13 | X == 15) -> X > 13
4493 case ICmpInst::ICMP_SGT: // (X s> 13 | X s> 15) -> X > 13
4494 return ReplaceInstUsesWith(I, LHS);
4495 case ICmpInst::ICMP_UGT: // (X s> 13 | X u> 15) -> no change
4497 case ICmpInst::ICMP_NE: // (X s> 13 | X != 15) -> true
4498 case ICmpInst::ICMP_SLT: // (X s> 13 | X s< 15) -> true
4499 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4500 case ICmpInst::ICMP_ULT: // (X s> 13 | X u< 15) -> no change
4508 /// FoldOrWithConstants - This helper function folds:
4510 /// ((A | B) & C1) | (B & C2)
4516 /// when the XOR of the two constants is "all ones" (-1).
4517 Instruction *InstCombiner::FoldOrWithConstants(BinaryOperator &I, Value *Op,
4518 Value *A, Value *B, Value *C) {
4519 ConstantInt *CI1 = dyn_cast<ConstantInt>(C);
4523 ConstantInt *CI2 = 0;
4524 if (!match(Op, m_And(m_Value(V1), m_ConstantInt(CI2)))) return 0;
4526 APInt Xor = CI1->getValue() ^ CI2->getValue();
4527 if (!Xor.isAllOnesValue()) return 0;
4529 if (V1 == A || V1 == B) {
4530 Instruction *NewOp =
4531 InsertNewInstBefore(BinaryOperator::CreateAnd((V1 == A) ? B : A, CI1), I);
4532 return BinaryOperator::CreateOr(NewOp, V1);
4538 Instruction *InstCombiner::visitOr(BinaryOperator &I) {
4539 bool Changed = SimplifyCommutative(I);
4540 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4542 if (isa<UndefValue>(Op1)) // X | undef -> -1
4543 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4547 return ReplaceInstUsesWith(I, Op0);
4549 // See if we can simplify any instructions used by the instruction whose sole
4550 // purpose is to compute bits we don't care about.
4551 if (!isa<VectorType>(I.getType())) {
4552 if (SimplifyDemandedInstructionBits(I))
4554 } else if (isa<ConstantAggregateZero>(Op1)) {
4555 return ReplaceInstUsesWith(I, Op0); // X | <0,0> -> X
4556 } else if (ConstantVector *CP = dyn_cast<ConstantVector>(Op1)) {
4557 if (CP->isAllOnesValue()) // X | <-1,-1> -> <-1,-1>
4558 return ReplaceInstUsesWith(I, I.getOperand(1));
4564 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
4565 ConstantInt *C1 = 0; Value *X = 0;
4566 // (X & C1) | C2 --> (X | C2) & (C1|C2)
4567 if (match(Op0, m_And(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
4568 Instruction *Or = BinaryOperator::CreateOr(X, RHS);
4569 InsertNewInstBefore(Or, I);
4571 return BinaryOperator::CreateAnd(Or,
4572 ConstantInt::get(RHS->getValue() | C1->getValue()));
4575 // (X ^ C1) | C2 --> (X | C2) ^ (C1&~C2)
4576 if (match(Op0, m_Xor(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
4577 Instruction *Or = BinaryOperator::CreateOr(X, RHS);
4578 InsertNewInstBefore(Or, I);
4580 return BinaryOperator::CreateXor(Or,
4581 ConstantInt::get(C1->getValue() & ~RHS->getValue()));
4584 // Try to fold constant and into select arguments.
4585 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
4586 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
4588 if (isa<PHINode>(Op0))
4589 if (Instruction *NV = FoldOpIntoPhi(I))
4593 Value *A = 0, *B = 0;
4594 ConstantInt *C1 = 0, *C2 = 0;
4596 if (match(Op0, m_And(m_Value(A), m_Value(B))))
4597 if (A == Op1 || B == Op1) // (A & ?) | A --> A
4598 return ReplaceInstUsesWith(I, Op1);
4599 if (match(Op1, m_And(m_Value(A), m_Value(B))))
4600 if (A == Op0 || B == Op0) // A | (A & ?) --> A
4601 return ReplaceInstUsesWith(I, Op0);
4603 // (A | B) | C and A | (B | C) -> bswap if possible.
4604 // (A >> B) | (C << D) and (A << B) | (B >> C) -> bswap if possible.
4605 if (match(Op0, m_Or(m_Value(), m_Value())) ||
4606 match(Op1, m_Or(m_Value(), m_Value())) ||
4607 (match(Op0, m_Shift(m_Value(), m_Value())) &&
4608 match(Op1, m_Shift(m_Value(), m_Value())))) {
4609 if (Instruction *BSwap = MatchBSwap(I))
4613 // (X^C)|Y -> (X|Y)^C iff Y&C == 0
4614 if (Op0->hasOneUse() && match(Op0, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
4615 MaskedValueIsZero(Op1, C1->getValue())) {
4616 Instruction *NOr = BinaryOperator::CreateOr(A, Op1);
4617 InsertNewInstBefore(NOr, I);
4619 return BinaryOperator::CreateXor(NOr, C1);
4622 // Y|(X^C) -> (X|Y)^C iff Y&C == 0
4623 if (Op1->hasOneUse() && match(Op1, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
4624 MaskedValueIsZero(Op0, C1->getValue())) {
4625 Instruction *NOr = BinaryOperator::CreateOr(A, Op0);
4626 InsertNewInstBefore(NOr, I);
4628 return BinaryOperator::CreateXor(NOr, C1);
4632 Value *C = 0, *D = 0;
4633 if (match(Op0, m_And(m_Value(A), m_Value(C))) &&
4634 match(Op1, m_And(m_Value(B), m_Value(D)))) {
4635 Value *V1 = 0, *V2 = 0, *V3 = 0;
4636 C1 = dyn_cast<ConstantInt>(C);
4637 C2 = dyn_cast<ConstantInt>(D);
4638 if (C1 && C2) { // (A & C1)|(B & C2)
4639 // If we have: ((V + N) & C1) | (V & C2)
4640 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
4641 // replace with V+N.
4642 if (C1->getValue() == ~C2->getValue()) {
4643 if ((C2->getValue() & (C2->getValue()+1)) == 0 && // C2 == 0+1+
4644 match(A, m_Add(m_Value(V1), m_Value(V2)))) {
4645 // Add commutes, try both ways.
4646 if (V1 == B && MaskedValueIsZero(V2, C2->getValue()))
4647 return ReplaceInstUsesWith(I, A);
4648 if (V2 == B && MaskedValueIsZero(V1, C2->getValue()))
4649 return ReplaceInstUsesWith(I, A);
4651 // Or commutes, try both ways.
4652 if ((C1->getValue() & (C1->getValue()+1)) == 0 &&
4653 match(B, m_Add(m_Value(V1), m_Value(V2)))) {
4654 // Add commutes, try both ways.
4655 if (V1 == A && MaskedValueIsZero(V2, C1->getValue()))
4656 return ReplaceInstUsesWith(I, B);
4657 if (V2 == A && MaskedValueIsZero(V1, C1->getValue()))
4658 return ReplaceInstUsesWith(I, B);
4661 V1 = 0; V2 = 0; V3 = 0;
4664 // Check to see if we have any common things being and'ed. If so, find the
4665 // terms for V1 & (V2|V3).
4666 if (isOnlyUse(Op0) || isOnlyUse(Op1)) {
4667 if (A == B) // (A & C)|(A & D) == A & (C|D)
4668 V1 = A, V2 = C, V3 = D;
4669 else if (A == D) // (A & C)|(B & A) == A & (B|C)
4670 V1 = A, V2 = B, V3 = C;
4671 else if (C == B) // (A & C)|(C & D) == C & (A|D)
4672 V1 = C, V2 = A, V3 = D;
4673 else if (C == D) // (A & C)|(B & C) == C & (A|B)
4674 V1 = C, V2 = A, V3 = B;
4678 InsertNewInstBefore(BinaryOperator::CreateOr(V2, V3, "tmp"), I);
4679 return BinaryOperator::CreateAnd(V1, Or);
4683 // (A & (C0?-1:0)) | (B & ~(C0?-1:0)) -> C0 ? A : B, and commuted variants
4684 if (Instruction *Match = MatchSelectFromAndOr(A, B, C, D))
4686 if (Instruction *Match = MatchSelectFromAndOr(B, A, D, C))
4688 if (Instruction *Match = MatchSelectFromAndOr(C, B, A, D))
4690 if (Instruction *Match = MatchSelectFromAndOr(D, A, B, C))
4693 // ((A&~B)|(~A&B)) -> A^B
4694 if ((match(C, m_Not(m_Specific(D))) &&
4695 match(B, m_Not(m_Specific(A)))))
4696 return BinaryOperator::CreateXor(A, D);
4697 // ((~B&A)|(~A&B)) -> A^B
4698 if ((match(A, m_Not(m_Specific(D))) &&
4699 match(B, m_Not(m_Specific(C)))))
4700 return BinaryOperator::CreateXor(C, D);
4701 // ((A&~B)|(B&~A)) -> A^B
4702 if ((match(C, m_Not(m_Specific(B))) &&
4703 match(D, m_Not(m_Specific(A)))))
4704 return BinaryOperator::CreateXor(A, B);
4705 // ((~B&A)|(B&~A)) -> A^B
4706 if ((match(A, m_Not(m_Specific(B))) &&
4707 match(D, m_Not(m_Specific(C)))))
4708 return BinaryOperator::CreateXor(C, B);
4711 // (X >> Z) | (Y >> Z) -> (X|Y) >> Z for all shifts.
4712 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
4713 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
4714 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
4715 SI0->getOperand(1) == SI1->getOperand(1) &&
4716 (SI0->hasOneUse() || SI1->hasOneUse())) {
4717 Instruction *NewOp =
4718 InsertNewInstBefore(BinaryOperator::CreateOr(SI0->getOperand(0),
4720 SI0->getName()), I);
4721 return BinaryOperator::Create(SI1->getOpcode(), NewOp,
4722 SI1->getOperand(1));
4726 // ((A|B)&1)|(B&-2) -> (A&1) | B
4727 if (match(Op0, m_And(m_Or(m_Value(A), m_Value(B)), m_Value(C))) ||
4728 match(Op0, m_And(m_Value(C), m_Or(m_Value(A), m_Value(B))))) {
4729 Instruction *Ret = FoldOrWithConstants(I, Op1, A, B, C);
4730 if (Ret) return Ret;
4732 // (B&-2)|((A|B)&1) -> (A&1) | B
4733 if (match(Op1, m_And(m_Or(m_Value(A), m_Value(B)), m_Value(C))) ||
4734 match(Op1, m_And(m_Value(C), m_Or(m_Value(A), m_Value(B))))) {
4735 Instruction *Ret = FoldOrWithConstants(I, Op0, A, B, C);
4736 if (Ret) return Ret;
4739 if (match(Op0, m_Not(m_Value(A)))) { // ~A | Op1
4740 if (A == Op1) // ~A | A == -1
4741 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4745 // Note, A is still live here!
4746 if (match(Op1, m_Not(m_Value(B)))) { // Op0 | ~B
4748 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4750 // (~A | ~B) == (~(A & B)) - De Morgan's Law
4751 if (A && isOnlyUse(Op0) && isOnlyUse(Op1)) {
4752 Value *And = InsertNewInstBefore(BinaryOperator::CreateAnd(A, B,
4753 I.getName()+".demorgan"), I);
4754 return BinaryOperator::CreateNot(And);
4758 // (icmp1 A, B) | (icmp2 A, B) --> (icmp3 A, B)
4759 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1))) {
4760 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
4763 if (ICmpInst *LHS = dyn_cast<ICmpInst>(I.getOperand(0)))
4764 if (Instruction *Res = FoldOrOfICmps(I, LHS, RHS))
4768 // fold (or (cast A), (cast B)) -> (cast (or A, B))
4769 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
4770 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4771 if (Op0C->getOpcode() == Op1C->getOpcode()) {// same cast kind ?
4772 if (!isa<ICmpInst>(Op0C->getOperand(0)) ||
4773 !isa<ICmpInst>(Op1C->getOperand(0))) {
4774 const Type *SrcTy = Op0C->getOperand(0)->getType();
4775 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
4776 // Only do this if the casts both really cause code to be
4778 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4780 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4782 Instruction *NewOp = BinaryOperator::CreateOr(Op0C->getOperand(0),
4783 Op1C->getOperand(0),
4785 InsertNewInstBefore(NewOp, I);
4786 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
4793 // (fcmp uno x, c) | (fcmp uno y, c) -> (fcmp uno x, y)
4794 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0))) {
4795 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1))) {
4796 if (LHS->getPredicate() == FCmpInst::FCMP_UNO &&
4797 RHS->getPredicate() == FCmpInst::FCMP_UNO &&
4798 LHS->getOperand(0)->getType() == RHS->getOperand(0)->getType()) {
4799 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
4800 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
4801 // If either of the constants are nans, then the whole thing returns
4803 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
4804 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4806 // Otherwise, no need to compare the two constants, compare the
4808 return new FCmpInst(FCmpInst::FCMP_UNO, LHS->getOperand(0),
4809 RHS->getOperand(0));
4812 Value *Op0LHS, *Op0RHS, *Op1LHS, *Op1RHS;
4813 FCmpInst::Predicate Op0CC, Op1CC;
4814 if (match(Op0, m_FCmp(Op0CC, m_Value(Op0LHS), m_Value(Op0RHS))) &&
4815 match(Op1, m_FCmp(Op1CC, m_Value(Op1LHS), m_Value(Op1RHS)))) {
4816 if (Op0LHS == Op1RHS && Op0RHS == Op1LHS) {
4817 // Swap RHS operands to match LHS.
4818 Op1CC = FCmpInst::getSwappedPredicate(Op1CC);
4819 std::swap(Op1LHS, Op1RHS);
4821 if (Op0LHS == Op1LHS && Op0RHS == Op1RHS) {
4822 // Simplify (fcmp cc0 x, y) | (fcmp cc1 x, y).
4824 return new FCmpInst((FCmpInst::Predicate)Op0CC, Op0LHS, Op0RHS);
4825 else if (Op0CC == FCmpInst::FCMP_TRUE ||
4826 Op1CC == FCmpInst::FCMP_TRUE)
4827 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4828 else if (Op0CC == FCmpInst::FCMP_FALSE)
4829 return ReplaceInstUsesWith(I, Op1);
4830 else if (Op1CC == FCmpInst::FCMP_FALSE)
4831 return ReplaceInstUsesWith(I, Op0);
4834 unsigned Op0Pred = getFCmpCode(Op0CC, Op0Ordered);
4835 unsigned Op1Pred = getFCmpCode(Op1CC, Op1Ordered);
4836 if (Op0Ordered == Op1Ordered) {
4837 // If both are ordered or unordered, return a new fcmp with
4838 // or'ed predicates.
4839 Value *RV = getFCmpValue(Op0Ordered, Op0Pred|Op1Pred,
4841 if (Instruction *I = dyn_cast<Instruction>(RV))
4843 // Otherwise, it's a constant boolean value...
4844 return ReplaceInstUsesWith(I, RV);
4852 return Changed ? &I : 0;
4857 // XorSelf - Implements: X ^ X --> 0
4860 XorSelf(Value *rhs) : RHS(rhs) {}
4861 bool shouldApply(Value *LHS) const { return LHS == RHS; }
4862 Instruction *apply(BinaryOperator &Xor) const {
4869 Instruction *InstCombiner::visitXor(BinaryOperator &I) {
4870 bool Changed = SimplifyCommutative(I);
4871 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4873 if (isa<UndefValue>(Op1)) {
4874 if (isa<UndefValue>(Op0))
4875 // Handle undef ^ undef -> 0 special case. This is a common
4877 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
4878 return ReplaceInstUsesWith(I, Op1); // X ^ undef -> undef
4881 // xor X, X = 0, even if X is nested in a sequence of Xor's.
4882 if (Instruction *Result = AssociativeOpt(I, XorSelf(Op1))) {
4883 assert(Result == &I && "AssociativeOpt didn't work?"); Result=Result;
4884 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
4887 // See if we can simplify any instructions used by the instruction whose sole
4888 // purpose is to compute bits we don't care about.
4889 if (!isa<VectorType>(I.getType())) {
4890 if (SimplifyDemandedInstructionBits(I))
4892 } else if (isa<ConstantAggregateZero>(Op1)) {
4893 return ReplaceInstUsesWith(I, Op0); // X ^ <0,0> -> X
4896 // Is this a ~ operation?
4897 if (Value *NotOp = dyn_castNotVal(&I)) {
4898 // ~(~X & Y) --> (X | ~Y) - De Morgan's Law
4899 // ~(~X | Y) === (X & ~Y) - De Morgan's Law
4900 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(NotOp)) {
4901 if (Op0I->getOpcode() == Instruction::And ||
4902 Op0I->getOpcode() == Instruction::Or) {
4903 if (dyn_castNotVal(Op0I->getOperand(1))) Op0I->swapOperands();
4904 if (Value *Op0NotVal = dyn_castNotVal(Op0I->getOperand(0))) {
4906 BinaryOperator::CreateNot(Op0I->getOperand(1),
4907 Op0I->getOperand(1)->getName()+".not");
4908 InsertNewInstBefore(NotY, I);
4909 if (Op0I->getOpcode() == Instruction::And)
4910 return BinaryOperator::CreateOr(Op0NotVal, NotY);
4912 return BinaryOperator::CreateAnd(Op0NotVal, NotY);
4919 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
4920 if (RHS == ConstantInt::getTrue() && Op0->hasOneUse()) {
4921 // xor (cmp A, B), true = not (cmp A, B) = !cmp A, B
4922 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Op0))
4923 return new ICmpInst(ICI->getInversePredicate(),
4924 ICI->getOperand(0), ICI->getOperand(1));
4926 if (FCmpInst *FCI = dyn_cast<FCmpInst>(Op0))
4927 return new FCmpInst(FCI->getInversePredicate(),
4928 FCI->getOperand(0), FCI->getOperand(1));
4931 // fold (xor(zext(cmp)), 1) and (xor(sext(cmp)), -1) to ext(!cmp).
4932 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
4933 if (CmpInst *CI = dyn_cast<CmpInst>(Op0C->getOperand(0))) {
4934 if (CI->hasOneUse() && Op0C->hasOneUse()) {
4935 Instruction::CastOps Opcode = Op0C->getOpcode();
4936 if (Opcode == Instruction::ZExt || Opcode == Instruction::SExt) {
4937 if (RHS == ConstantExpr::getCast(Opcode, ConstantInt::getTrue(),
4938 Op0C->getDestTy())) {
4939 Instruction *NewCI = InsertNewInstBefore(CmpInst::Create(
4940 CI->getOpcode(), CI->getInversePredicate(),
4941 CI->getOperand(0), CI->getOperand(1)), I);
4942 NewCI->takeName(CI);
4943 return CastInst::Create(Opcode, NewCI, Op0C->getType());
4950 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
4951 // ~(c-X) == X-c-1 == X+(-c-1)
4952 if (Op0I->getOpcode() == Instruction::Sub && RHS->isAllOnesValue())
4953 if (Constant *Op0I0C = dyn_cast<Constant>(Op0I->getOperand(0))) {
4954 Constant *NegOp0I0C = ConstantExpr::getNeg(Op0I0C);
4955 Constant *ConstantRHS = ConstantExpr::getSub(NegOp0I0C,
4956 ConstantInt::get(I.getType(), 1));
4957 return BinaryOperator::CreateAdd(Op0I->getOperand(1), ConstantRHS);
4960 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) {
4961 if (Op0I->getOpcode() == Instruction::Add) {
4962 // ~(X-c) --> (-c-1)-X
4963 if (RHS->isAllOnesValue()) {
4964 Constant *NegOp0CI = ConstantExpr::getNeg(Op0CI);
4965 return BinaryOperator::CreateSub(
4966 ConstantExpr::getSub(NegOp0CI,
4967 ConstantInt::get(I.getType(), 1)),
4968 Op0I->getOperand(0));
4969 } else if (RHS->getValue().isSignBit()) {
4970 // (X + C) ^ signbit -> (X + C + signbit)
4971 Constant *C = ConstantInt::get(RHS->getValue() + Op0CI->getValue());
4972 return BinaryOperator::CreateAdd(Op0I->getOperand(0), C);
4975 } else if (Op0I->getOpcode() == Instruction::Or) {
4976 // (X|C1)^C2 -> X^(C1|C2) iff X&~C1 == 0
4977 if (MaskedValueIsZero(Op0I->getOperand(0), Op0CI->getValue())) {
4978 Constant *NewRHS = ConstantExpr::getOr(Op0CI, RHS);
4979 // Anything in both C1 and C2 is known to be zero, remove it from
4981 Constant *CommonBits = And(Op0CI, RHS);
4982 NewRHS = ConstantExpr::getAnd(NewRHS,
4983 ConstantExpr::getNot(CommonBits));
4984 AddToWorkList(Op0I);
4985 I.setOperand(0, Op0I->getOperand(0));
4986 I.setOperand(1, NewRHS);
4993 // Try to fold constant and into select arguments.
4994 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
4995 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
4997 if (isa<PHINode>(Op0))
4998 if (Instruction *NV = FoldOpIntoPhi(I))
5002 if (Value *X = dyn_castNotVal(Op0)) // ~A ^ A == -1
5004 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
5006 if (Value *X = dyn_castNotVal(Op1)) // A ^ ~A == -1
5008 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
5011 BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1);
5014 if (match(Op1I, m_Or(m_Value(A), m_Value(B)))) {
5015 if (A == Op0) { // B^(B|A) == (A|B)^B
5016 Op1I->swapOperands();
5018 std::swap(Op0, Op1);
5019 } else if (B == Op0) { // B^(A|B) == (A|B)^B
5020 I.swapOperands(); // Simplified below.
5021 std::swap(Op0, Op1);
5023 } else if (match(Op1I, m_Xor(m_Specific(Op0), m_Value(B)))) {
5024 return ReplaceInstUsesWith(I, B); // A^(A^B) == B
5025 } else if (match(Op1I, m_Xor(m_Value(A), m_Specific(Op0)))) {
5026 return ReplaceInstUsesWith(I, A); // A^(B^A) == B
5027 } else if (match(Op1I, m_And(m_Value(A), m_Value(B))) && Op1I->hasOneUse()){
5028 if (A == Op0) { // A^(A&B) -> A^(B&A)
5029 Op1I->swapOperands();
5032 if (B == Op0) { // A^(B&A) -> (B&A)^A
5033 I.swapOperands(); // Simplified below.
5034 std::swap(Op0, Op1);
5039 BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0);
5042 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) && Op0I->hasOneUse()) {
5043 if (A == Op1) // (B|A)^B == (A|B)^B
5045 if (B == Op1) { // (A|B)^B == A & ~B
5047 InsertNewInstBefore(BinaryOperator::CreateNot(Op1, "tmp"), I);
5048 return BinaryOperator::CreateAnd(A, NotB);
5050 } else if (match(Op0I, m_Xor(m_Specific(Op1), m_Value(B)))) {
5051 return ReplaceInstUsesWith(I, B); // (A^B)^A == B
5052 } else if (match(Op0I, m_Xor(m_Value(A), m_Specific(Op1)))) {
5053 return ReplaceInstUsesWith(I, A); // (B^A)^A == B
5054 } else if (match(Op0I, m_And(m_Value(A), m_Value(B))) && Op0I->hasOneUse()){
5055 if (A == Op1) // (A&B)^A -> (B&A)^A
5057 if (B == Op1 && // (B&A)^A == ~B & A
5058 !isa<ConstantInt>(Op1)) { // Canonical form is (B&C)^C
5060 InsertNewInstBefore(BinaryOperator::CreateNot(A, "tmp"), I);
5061 return BinaryOperator::CreateAnd(N, Op1);
5066 // (X >> Z) ^ (Y >> Z) -> (X^Y) >> Z for all shifts.
5067 if (Op0I && Op1I && Op0I->isShift() &&
5068 Op0I->getOpcode() == Op1I->getOpcode() &&
5069 Op0I->getOperand(1) == Op1I->getOperand(1) &&
5070 (Op1I->hasOneUse() || Op1I->hasOneUse())) {
5071 Instruction *NewOp =
5072 InsertNewInstBefore(BinaryOperator::CreateXor(Op0I->getOperand(0),
5073 Op1I->getOperand(0),
5074 Op0I->getName()), I);
5075 return BinaryOperator::Create(Op1I->getOpcode(), NewOp,
5076 Op1I->getOperand(1));
5080 Value *A, *B, *C, *D;
5081 // (A & B)^(A | B) -> A ^ B
5082 if (match(Op0I, m_And(m_Value(A), m_Value(B))) &&
5083 match(Op1I, m_Or(m_Value(C), m_Value(D)))) {
5084 if ((A == C && B == D) || (A == D && B == C))
5085 return BinaryOperator::CreateXor(A, B);
5087 // (A | B)^(A & B) -> A ^ B
5088 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) &&
5089 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
5090 if ((A == C && B == D) || (A == D && B == C))
5091 return BinaryOperator::CreateXor(A, B);
5095 if ((Op0I->hasOneUse() || Op1I->hasOneUse()) &&
5096 match(Op0I, m_And(m_Value(A), m_Value(B))) &&
5097 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
5098 // (X & Y)^(X & Y) -> (Y^Z) & X
5099 Value *X = 0, *Y = 0, *Z = 0;
5101 X = A, Y = B, Z = D;
5103 X = A, Y = B, Z = C;
5105 X = B, Y = A, Z = D;
5107 X = B, Y = A, Z = C;
5110 Instruction *NewOp =
5111 InsertNewInstBefore(BinaryOperator::CreateXor(Y, Z, Op0->getName()), I);
5112 return BinaryOperator::CreateAnd(NewOp, X);
5117 // (icmp1 A, B) ^ (icmp2 A, B) --> (icmp3 A, B)
5118 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1)))
5119 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
5122 // fold (xor (cast A), (cast B)) -> (cast (xor A, B))
5123 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
5124 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
5125 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind?
5126 const Type *SrcTy = Op0C->getOperand(0)->getType();
5127 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
5128 // Only do this if the casts both really cause code to be generated.
5129 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
5131 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
5133 Instruction *NewOp = BinaryOperator::CreateXor(Op0C->getOperand(0),
5134 Op1C->getOperand(0),
5136 InsertNewInstBefore(NewOp, I);
5137 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
5142 return Changed ? &I : 0;
5145 /// AddWithOverflow - Compute Result = In1+In2, returning true if the result
5146 /// overflowed for this type.
5147 static bool AddWithOverflow(ConstantInt *&Result, ConstantInt *In1,
5148 ConstantInt *In2, bool IsSigned = false) {
5149 Result = cast<ConstantInt>(Add(In1, In2));
5152 if (In2->getValue().isNegative())
5153 return Result->getValue().sgt(In1->getValue());
5155 return Result->getValue().slt(In1->getValue());
5157 return Result->getValue().ult(In1->getValue());
5160 /// SubWithOverflow - Compute Result = In1-In2, returning true if the result
5161 /// overflowed for this type.
5162 static bool SubWithOverflow(ConstantInt *&Result, ConstantInt *In1,
5163 ConstantInt *In2, bool IsSigned = false) {
5164 Result = cast<ConstantInt>(Subtract(In1, In2));
5167 if (In2->getValue().isNegative())
5168 return Result->getValue().slt(In1->getValue());
5170 return Result->getValue().sgt(In1->getValue());
5172 return Result->getValue().ugt(In1->getValue());
5175 /// EmitGEPOffset - Given a getelementptr instruction/constantexpr, emit the
5176 /// code necessary to compute the offset from the base pointer (without adding
5177 /// in the base pointer). Return the result as a signed integer of intptr size.
5178 static Value *EmitGEPOffset(User *GEP, Instruction &I, InstCombiner &IC) {
5179 TargetData &TD = IC.getTargetData();
5180 gep_type_iterator GTI = gep_type_begin(GEP);
5181 const Type *IntPtrTy = TD.getIntPtrType();
5182 Value *Result = Constant::getNullValue(IntPtrTy);
5184 // Build a mask for high order bits.
5185 unsigned IntPtrWidth = TD.getPointerSizeInBits();
5186 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth);
5188 for (User::op_iterator i = GEP->op_begin() + 1, e = GEP->op_end(); i != e;
5191 uint64_t Size = TD.getTypePaddedSize(GTI.getIndexedType()) & PtrSizeMask;
5192 if (ConstantInt *OpC = dyn_cast<ConstantInt>(Op)) {
5193 if (OpC->isZero()) continue;
5195 // Handle a struct index, which adds its field offset to the pointer.
5196 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
5197 Size = TD.getStructLayout(STy)->getElementOffset(OpC->getZExtValue());
5199 if (ConstantInt *RC = dyn_cast<ConstantInt>(Result))
5200 Result = ConstantInt::get(RC->getValue() + APInt(IntPtrWidth, Size));
5202 Result = IC.InsertNewInstBefore(
5203 BinaryOperator::CreateAdd(Result,
5204 ConstantInt::get(IntPtrTy, Size),
5205 GEP->getName()+".offs"), I);
5209 Constant *Scale = ConstantInt::get(IntPtrTy, Size);
5210 Constant *OC = ConstantExpr::getIntegerCast(OpC, IntPtrTy, true /*SExt*/);
5211 Scale = ConstantExpr::getMul(OC, Scale);
5212 if (Constant *RC = dyn_cast<Constant>(Result))
5213 Result = ConstantExpr::getAdd(RC, Scale);
5215 // Emit an add instruction.
5216 Result = IC.InsertNewInstBefore(
5217 BinaryOperator::CreateAdd(Result, Scale,
5218 GEP->getName()+".offs"), I);
5222 // Convert to correct type.
5223 if (Op->getType() != IntPtrTy) {
5224 if (Constant *OpC = dyn_cast<Constant>(Op))
5225 Op = ConstantExpr::getSExt(OpC, IntPtrTy);
5227 Op = IC.InsertNewInstBefore(new SExtInst(Op, IntPtrTy,
5228 Op->getName()+".c"), I);
5231 Constant *Scale = ConstantInt::get(IntPtrTy, Size);
5232 if (Constant *OpC = dyn_cast<Constant>(Op))
5233 Op = ConstantExpr::getMul(OpC, Scale);
5234 else // We'll let instcombine(mul) convert this to a shl if possible.
5235 Op = IC.InsertNewInstBefore(BinaryOperator::CreateMul(Op, Scale,
5236 GEP->getName()+".idx"), I);
5239 // Emit an add instruction.
5240 if (isa<Constant>(Op) && isa<Constant>(Result))
5241 Result = ConstantExpr::getAdd(cast<Constant>(Op),
5242 cast<Constant>(Result));
5244 Result = IC.InsertNewInstBefore(BinaryOperator::CreateAdd(Op, Result,
5245 GEP->getName()+".offs"), I);
5251 /// EvaluateGEPOffsetExpression - Return an value that can be used to compare of
5252 /// the *offset* implied by GEP to zero. For example, if we have &A[i], we want
5253 /// to return 'i' for "icmp ne i, 0". Note that, in general, indices can be
5254 /// complex, and scales are involved. The above expression would also be legal
5255 /// to codegen as "icmp ne (i*4), 0" (assuming A is a pointer to i32). This
5256 /// later form is less amenable to optimization though, and we are allowed to
5257 /// generate the first by knowing that pointer arithmetic doesn't overflow.
5259 /// If we can't emit an optimized form for this expression, this returns null.
5261 static Value *EvaluateGEPOffsetExpression(User *GEP, Instruction &I,
5263 TargetData &TD = IC.getTargetData();
5264 gep_type_iterator GTI = gep_type_begin(GEP);
5266 // Check to see if this gep only has a single variable index. If so, and if
5267 // any constant indices are a multiple of its scale, then we can compute this
5268 // in terms of the scale of the variable index. For example, if the GEP
5269 // implies an offset of "12 + i*4", then we can codegen this as "3 + i",
5270 // because the expression will cross zero at the same point.
5271 unsigned i, e = GEP->getNumOperands();
5273 for (i = 1; i != e; ++i, ++GTI) {
5274 if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i))) {
5275 // Compute the aggregate offset of constant indices.
5276 if (CI->isZero()) continue;
5278 // Handle a struct index, which adds its field offset to the pointer.
5279 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
5280 Offset += TD.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
5282 uint64_t Size = TD.getTypePaddedSize(GTI.getIndexedType());
5283 Offset += Size*CI->getSExtValue();
5286 // Found our variable index.
5291 // If there are no variable indices, we must have a constant offset, just
5292 // evaluate it the general way.
5293 if (i == e) return 0;
5295 Value *VariableIdx = GEP->getOperand(i);
5296 // Determine the scale factor of the variable element. For example, this is
5297 // 4 if the variable index is into an array of i32.
5298 uint64_t VariableScale = TD.getTypePaddedSize(GTI.getIndexedType());
5300 // Verify that there are no other variable indices. If so, emit the hard way.
5301 for (++i, ++GTI; i != e; ++i, ++GTI) {
5302 ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i));
5305 // Compute the aggregate offset of constant indices.
5306 if (CI->isZero()) continue;
5308 // Handle a struct index, which adds its field offset to the pointer.
5309 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
5310 Offset += TD.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
5312 uint64_t Size = TD.getTypePaddedSize(GTI.getIndexedType());
5313 Offset += Size*CI->getSExtValue();
5317 // Okay, we know we have a single variable index, which must be a
5318 // pointer/array/vector index. If there is no offset, life is simple, return
5320 unsigned IntPtrWidth = TD.getPointerSizeInBits();
5322 // Cast to intptrty in case a truncation occurs. If an extension is needed,
5323 // we don't need to bother extending: the extension won't affect where the
5324 // computation crosses zero.
5325 if (VariableIdx->getType()->getPrimitiveSizeInBits() > IntPtrWidth)
5326 VariableIdx = new TruncInst(VariableIdx, TD.getIntPtrType(),
5327 VariableIdx->getNameStart(), &I);
5331 // Otherwise, there is an index. The computation we will do will be modulo
5332 // the pointer size, so get it.
5333 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth);
5335 Offset &= PtrSizeMask;
5336 VariableScale &= PtrSizeMask;
5338 // To do this transformation, any constant index must be a multiple of the
5339 // variable scale factor. For example, we can evaluate "12 + 4*i" as "3 + i",
5340 // but we can't evaluate "10 + 3*i" in terms of i. Check that the offset is a
5341 // multiple of the variable scale.
5342 int64_t NewOffs = Offset / (int64_t)VariableScale;
5343 if (Offset != NewOffs*(int64_t)VariableScale)
5346 // Okay, we can do this evaluation. Start by converting the index to intptr.
5347 const Type *IntPtrTy = TD.getIntPtrType();
5348 if (VariableIdx->getType() != IntPtrTy)
5349 VariableIdx = CastInst::CreateIntegerCast(VariableIdx, IntPtrTy,
5351 VariableIdx->getNameStart(), &I);
5352 Constant *OffsetVal = ConstantInt::get(IntPtrTy, NewOffs);
5353 return BinaryOperator::CreateAdd(VariableIdx, OffsetVal, "offset", &I);
5357 /// FoldGEPICmp - Fold comparisons between a GEP instruction and something
5358 /// else. At this point we know that the GEP is on the LHS of the comparison.
5359 Instruction *InstCombiner::FoldGEPICmp(User *GEPLHS, Value *RHS,
5360 ICmpInst::Predicate Cond,
5362 assert(dyn_castGetElementPtr(GEPLHS) && "LHS is not a getelementptr!");
5364 // Look through bitcasts.
5365 if (BitCastInst *BCI = dyn_cast<BitCastInst>(RHS))
5366 RHS = BCI->getOperand(0);
5368 Value *PtrBase = GEPLHS->getOperand(0);
5369 if (PtrBase == RHS) {
5370 // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
5371 // This transformation (ignoring the base and scales) is valid because we
5372 // know pointers can't overflow. See if we can output an optimized form.
5373 Value *Offset = EvaluateGEPOffsetExpression(GEPLHS, I, *this);
5375 // If not, synthesize the offset the hard way.
5377 Offset = EmitGEPOffset(GEPLHS, I, *this);
5378 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Offset,
5379 Constant::getNullValue(Offset->getType()));
5380 } else if (User *GEPRHS = dyn_castGetElementPtr(RHS)) {
5381 // If the base pointers are different, but the indices are the same, just
5382 // compare the base pointer.
5383 if (PtrBase != GEPRHS->getOperand(0)) {
5384 bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands();
5385 IndicesTheSame &= GEPLHS->getOperand(0)->getType() ==
5386 GEPRHS->getOperand(0)->getType();
5388 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
5389 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
5390 IndicesTheSame = false;
5394 // If all indices are the same, just compare the base pointers.
5396 return new ICmpInst(ICmpInst::getSignedPredicate(Cond),
5397 GEPLHS->getOperand(0), GEPRHS->getOperand(0));
5399 // Otherwise, the base pointers are different and the indices are
5400 // different, bail out.
5404 // If one of the GEPs has all zero indices, recurse.
5405 bool AllZeros = true;
5406 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
5407 if (!isa<Constant>(GEPLHS->getOperand(i)) ||
5408 !cast<Constant>(GEPLHS->getOperand(i))->isNullValue()) {
5413 return FoldGEPICmp(GEPRHS, GEPLHS->getOperand(0),
5414 ICmpInst::getSwappedPredicate(Cond), I);
5416 // If the other GEP has all zero indices, recurse.
5418 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
5419 if (!isa<Constant>(GEPRHS->getOperand(i)) ||
5420 !cast<Constant>(GEPRHS->getOperand(i))->isNullValue()) {
5425 return FoldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I);
5427 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
5428 // If the GEPs only differ by one index, compare it.
5429 unsigned NumDifferences = 0; // Keep track of # differences.
5430 unsigned DiffOperand = 0; // The operand that differs.
5431 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
5432 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
5433 if (GEPLHS->getOperand(i)->getType()->getPrimitiveSizeInBits() !=
5434 GEPRHS->getOperand(i)->getType()->getPrimitiveSizeInBits()) {
5435 // Irreconcilable differences.
5439 if (NumDifferences++) break;
5444 if (NumDifferences == 0) // SAME GEP?
5445 return ReplaceInstUsesWith(I, // No comparison is needed here.
5446 ConstantInt::get(Type::Int1Ty,
5447 ICmpInst::isTrueWhenEqual(Cond)));
5449 else if (NumDifferences == 1) {
5450 Value *LHSV = GEPLHS->getOperand(DiffOperand);
5451 Value *RHSV = GEPRHS->getOperand(DiffOperand);
5452 // Make sure we do a signed comparison here.
5453 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), LHSV, RHSV);
5457 // Only lower this if the icmp is the only user of the GEP or if we expect
5458 // the result to fold to a constant!
5459 if ((isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
5460 (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
5461 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
5462 Value *L = EmitGEPOffset(GEPLHS, I, *this);
5463 Value *R = EmitGEPOffset(GEPRHS, I, *this);
5464 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), L, R);
5470 /// FoldFCmp_IntToFP_Cst - Fold fcmp ([us]itofp x, cst) if possible.
5472 Instruction *InstCombiner::FoldFCmp_IntToFP_Cst(FCmpInst &I,
5475 if (!isa<ConstantFP>(RHSC)) return 0;
5476 const APFloat &RHS = cast<ConstantFP>(RHSC)->getValueAPF();
5478 // Get the width of the mantissa. We don't want to hack on conversions that
5479 // might lose information from the integer, e.g. "i64 -> float"
5480 int MantissaWidth = LHSI->getType()->getFPMantissaWidth();
5481 if (MantissaWidth == -1) return 0; // Unknown.
5483 // Check to see that the input is converted from an integer type that is small
5484 // enough that preserves all bits. TODO: check here for "known" sign bits.
5485 // This would allow us to handle (fptosi (x >>s 62) to float) if x is i64 f.e.
5486 unsigned InputSize = LHSI->getOperand(0)->getType()->getPrimitiveSizeInBits();
5488 // If this is a uitofp instruction, we need an extra bit to hold the sign.
5489 bool LHSUnsigned = isa<UIToFPInst>(LHSI);
5493 // If the conversion would lose info, don't hack on this.
5494 if ((int)InputSize > MantissaWidth)
5497 // Otherwise, we can potentially simplify the comparison. We know that it
5498 // will always come through as an integer value and we know the constant is
5499 // not a NAN (it would have been previously simplified).
5500 assert(!RHS.isNaN() && "NaN comparison not already folded!");
5502 ICmpInst::Predicate Pred;
5503 switch (I.getPredicate()) {
5504 default: assert(0 && "Unexpected predicate!");
5505 case FCmpInst::FCMP_UEQ:
5506 case FCmpInst::FCMP_OEQ:
5507 Pred = ICmpInst::ICMP_EQ;
5509 case FCmpInst::FCMP_UGT:
5510 case FCmpInst::FCMP_OGT:
5511 Pred = LHSUnsigned ? ICmpInst::ICMP_UGT : ICmpInst::ICMP_SGT;
5513 case FCmpInst::FCMP_UGE:
5514 case FCmpInst::FCMP_OGE:
5515 Pred = LHSUnsigned ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_SGE;
5517 case FCmpInst::FCMP_ULT:
5518 case FCmpInst::FCMP_OLT:
5519 Pred = LHSUnsigned ? ICmpInst::ICMP_ULT : ICmpInst::ICMP_SLT;
5521 case FCmpInst::FCMP_ULE:
5522 case FCmpInst::FCMP_OLE:
5523 Pred = LHSUnsigned ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_SLE;
5525 case FCmpInst::FCMP_UNE:
5526 case FCmpInst::FCMP_ONE:
5527 Pred = ICmpInst::ICMP_NE;
5529 case FCmpInst::FCMP_ORD:
5530 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5531 case FCmpInst::FCMP_UNO:
5532 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5535 const IntegerType *IntTy = cast<IntegerType>(LHSI->getOperand(0)->getType());
5537 // Now we know that the APFloat is a normal number, zero or inf.
5539 // See if the FP constant is too large for the integer. For example,
5540 // comparing an i8 to 300.0.
5541 unsigned IntWidth = IntTy->getPrimitiveSizeInBits();
5544 // If the RHS value is > SignedMax, fold the comparison. This handles +INF
5545 // and large values.
5546 APFloat SMax(RHS.getSemantics(), APFloat::fcZero, false);
5547 SMax.convertFromAPInt(APInt::getSignedMaxValue(IntWidth), true,
5548 APFloat::rmNearestTiesToEven);
5549 if (SMax.compare(RHS) == APFloat::cmpLessThan) { // smax < 13123.0
5550 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SLT ||
5551 Pred == ICmpInst::ICMP_SLE)
5552 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5553 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5556 // If the RHS value is > UnsignedMax, fold the comparison. This handles
5557 // +INF and large values.
5558 APFloat UMax(RHS.getSemantics(), APFloat::fcZero, false);
5559 UMax.convertFromAPInt(APInt::getMaxValue(IntWidth), false,
5560 APFloat::rmNearestTiesToEven);
5561 if (UMax.compare(RHS) == APFloat::cmpLessThan) { // umax < 13123.0
5562 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_ULT ||
5563 Pred == ICmpInst::ICMP_ULE)
5564 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5565 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5570 // See if the RHS value is < SignedMin.
5571 APFloat SMin(RHS.getSemantics(), APFloat::fcZero, false);
5572 SMin.convertFromAPInt(APInt::getSignedMinValue(IntWidth), true,
5573 APFloat::rmNearestTiesToEven);
5574 if (SMin.compare(RHS) == APFloat::cmpGreaterThan) { // smin > 12312.0
5575 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SGT ||
5576 Pred == ICmpInst::ICMP_SGE)
5577 return ReplaceInstUsesWith(I,ConstantInt::getTrue());
5578 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5582 // Okay, now we know that the FP constant fits in the range [SMIN, SMAX] or
5583 // [0, UMAX], but it may still be fractional. See if it is fractional by
5584 // casting the FP value to the integer value and back, checking for equality.
5585 // Don't do this for zero, because -0.0 is not fractional.
5586 Constant *RHSInt = ConstantExpr::getFPToSI(RHSC, IntTy);
5587 if (!RHS.isZero() &&
5588 ConstantExpr::getSIToFP(RHSInt, RHSC->getType()) != RHSC) {
5589 // If we had a comparison against a fractional value, we have to adjust the
5590 // compare predicate and sometimes the value. RHSC is rounded towards zero
5593 default: assert(0 && "Unexpected integer comparison!");
5594 case ICmpInst::ICMP_NE: // (float)int != 4.4 --> true
5595 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5596 case ICmpInst::ICMP_EQ: // (float)int == 4.4 --> false
5597 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5598 case ICmpInst::ICMP_ULE:
5599 // (float)int <= 4.4 --> int <= 4
5600 // (float)int <= -4.4 --> false
5601 if (RHS.isNegative())
5602 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5604 case ICmpInst::ICMP_SLE:
5605 // (float)int <= 4.4 --> int <= 4
5606 // (float)int <= -4.4 --> int < -4
5607 if (RHS.isNegative())
5608 Pred = ICmpInst::ICMP_SLT;
5610 case ICmpInst::ICMP_ULT:
5611 // (float)int < -4.4 --> false
5612 // (float)int < 4.4 --> int <= 4
5613 if (RHS.isNegative())
5614 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5615 Pred = ICmpInst::ICMP_ULE;
5617 case ICmpInst::ICMP_SLT:
5618 // (float)int < -4.4 --> int < -4
5619 // (float)int < 4.4 --> int <= 4
5620 if (!RHS.isNegative())
5621 Pred = ICmpInst::ICMP_SLE;
5623 case ICmpInst::ICMP_UGT:
5624 // (float)int > 4.4 --> int > 4
5625 // (float)int > -4.4 --> true
5626 if (RHS.isNegative())
5627 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5629 case ICmpInst::ICMP_SGT:
5630 // (float)int > 4.4 --> int > 4
5631 // (float)int > -4.4 --> int >= -4
5632 if (RHS.isNegative())
5633 Pred = ICmpInst::ICMP_SGE;
5635 case ICmpInst::ICMP_UGE:
5636 // (float)int >= -4.4 --> true
5637 // (float)int >= 4.4 --> int > 4
5638 if (!RHS.isNegative())
5639 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5640 Pred = ICmpInst::ICMP_UGT;
5642 case ICmpInst::ICMP_SGE:
5643 // (float)int >= -4.4 --> int >= -4
5644 // (float)int >= 4.4 --> int > 4
5645 if (!RHS.isNegative())
5646 Pred = ICmpInst::ICMP_SGT;
5651 // Lower this FP comparison into an appropriate integer version of the
5653 return new ICmpInst(Pred, LHSI->getOperand(0), RHSInt);
5656 Instruction *InstCombiner::visitFCmpInst(FCmpInst &I) {
5657 bool Changed = SimplifyCompare(I);
5658 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5660 // Fold trivial predicates.
5661 if (I.getPredicate() == FCmpInst::FCMP_FALSE)
5662 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5663 if (I.getPredicate() == FCmpInst::FCMP_TRUE)
5664 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5666 // Simplify 'fcmp pred X, X'
5668 switch (I.getPredicate()) {
5669 default: assert(0 && "Unknown predicate!");
5670 case FCmpInst::FCMP_UEQ: // True if unordered or equal
5671 case FCmpInst::FCMP_UGE: // True if unordered, greater than, or equal
5672 case FCmpInst::FCMP_ULE: // True if unordered, less than, or equal
5673 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5674 case FCmpInst::FCMP_OGT: // True if ordered and greater than
5675 case FCmpInst::FCMP_OLT: // True if ordered and less than
5676 case FCmpInst::FCMP_ONE: // True if ordered and operands are unequal
5677 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5679 case FCmpInst::FCMP_UNO: // True if unordered: isnan(X) | isnan(Y)
5680 case FCmpInst::FCMP_ULT: // True if unordered or less than
5681 case FCmpInst::FCMP_UGT: // True if unordered or greater than
5682 case FCmpInst::FCMP_UNE: // True if unordered or not equal
5683 // Canonicalize these to be 'fcmp uno %X, 0.0'.
5684 I.setPredicate(FCmpInst::FCMP_UNO);
5685 I.setOperand(1, Constant::getNullValue(Op0->getType()));
5688 case FCmpInst::FCMP_ORD: // True if ordered (no nans)
5689 case FCmpInst::FCMP_OEQ: // True if ordered and equal
5690 case FCmpInst::FCMP_OGE: // True if ordered and greater than or equal
5691 case FCmpInst::FCMP_OLE: // True if ordered and less than or equal
5692 // Canonicalize these to be 'fcmp ord %X, 0.0'.
5693 I.setPredicate(FCmpInst::FCMP_ORD);
5694 I.setOperand(1, Constant::getNullValue(Op0->getType()));
5699 if (isa<UndefValue>(Op1)) // fcmp pred X, undef -> undef
5700 return ReplaceInstUsesWith(I, UndefValue::get(Type::Int1Ty));
5702 // Handle fcmp with constant RHS
5703 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
5704 // If the constant is a nan, see if we can fold the comparison based on it.
5705 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHSC)) {
5706 if (CFP->getValueAPF().isNaN()) {
5707 if (FCmpInst::isOrdered(I.getPredicate())) // True if ordered and...
5708 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5709 assert(FCmpInst::isUnordered(I.getPredicate()) &&
5710 "Comparison must be either ordered or unordered!");
5711 // True if unordered.
5712 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5716 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
5717 switch (LHSI->getOpcode()) {
5718 case Instruction::PHI:
5719 // Only fold fcmp into the PHI if the phi and fcmp are in the same
5720 // block. If in the same block, we're encouraging jump threading. If
5721 // not, we are just pessimizing the code by making an i1 phi.
5722 if (LHSI->getParent() == I.getParent())
5723 if (Instruction *NV = FoldOpIntoPhi(I))
5726 case Instruction::SIToFP:
5727 case Instruction::UIToFP:
5728 if (Instruction *NV = FoldFCmp_IntToFP_Cst(I, LHSI, RHSC))
5731 case Instruction::Select:
5732 // If either operand of the select is a constant, we can fold the
5733 // comparison into the select arms, which will cause one to be
5734 // constant folded and the select turned into a bitwise or.
5735 Value *Op1 = 0, *Op2 = 0;
5736 if (LHSI->hasOneUse()) {
5737 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
5738 // Fold the known value into the constant operand.
5739 Op1 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
5740 // Insert a new FCmp of the other select operand.
5741 Op2 = InsertNewInstBefore(new FCmpInst(I.getPredicate(),
5742 LHSI->getOperand(2), RHSC,
5744 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
5745 // Fold the known value into the constant operand.
5746 Op2 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
5747 // Insert a new FCmp of the other select operand.
5748 Op1 = InsertNewInstBefore(new FCmpInst(I.getPredicate(),
5749 LHSI->getOperand(1), RHSC,
5755 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
5760 return Changed ? &I : 0;
5763 Instruction *InstCombiner::visitICmpInst(ICmpInst &I) {
5764 bool Changed = SimplifyCompare(I);
5765 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5766 const Type *Ty = Op0->getType();
5770 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
5771 I.isTrueWhenEqual()));
5773 if (isa<UndefValue>(Op1)) // X icmp undef -> undef
5774 return ReplaceInstUsesWith(I, UndefValue::get(Type::Int1Ty));
5776 // icmp <global/alloca*/null>, <global/alloca*/null> - Global/Stack value
5777 // addresses never equal each other! We already know that Op0 != Op1.
5778 if ((isa<GlobalValue>(Op0) || isa<AllocaInst>(Op0) ||
5779 isa<ConstantPointerNull>(Op0)) &&
5780 (isa<GlobalValue>(Op1) || isa<AllocaInst>(Op1) ||
5781 isa<ConstantPointerNull>(Op1)))
5782 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
5783 !I.isTrueWhenEqual()));
5785 // icmp's with boolean values can always be turned into bitwise operations
5786 if (Ty == Type::Int1Ty) {
5787 switch (I.getPredicate()) {
5788 default: assert(0 && "Invalid icmp instruction!");
5789 case ICmpInst::ICMP_EQ: { // icmp eq i1 A, B -> ~(A^B)
5790 Instruction *Xor = BinaryOperator::CreateXor(Op0, Op1, I.getName()+"tmp");
5791 InsertNewInstBefore(Xor, I);
5792 return BinaryOperator::CreateNot(Xor);
5794 case ICmpInst::ICMP_NE: // icmp eq i1 A, B -> A^B
5795 return BinaryOperator::CreateXor(Op0, Op1);
5797 case ICmpInst::ICMP_UGT:
5798 std::swap(Op0, Op1); // Change icmp ugt -> icmp ult
5800 case ICmpInst::ICMP_ULT:{ // icmp ult i1 A, B -> ~A & B
5801 Instruction *Not = BinaryOperator::CreateNot(Op0, I.getName()+"tmp");
5802 InsertNewInstBefore(Not, I);
5803 return BinaryOperator::CreateAnd(Not, Op1);
5805 case ICmpInst::ICMP_SGT:
5806 std::swap(Op0, Op1); // Change icmp sgt -> icmp slt
5808 case ICmpInst::ICMP_SLT: { // icmp slt i1 A, B -> A & ~B
5809 Instruction *Not = BinaryOperator::CreateNot(Op1, I.getName()+"tmp");
5810 InsertNewInstBefore(Not, I);
5811 return BinaryOperator::CreateAnd(Not, Op0);
5813 case ICmpInst::ICMP_UGE:
5814 std::swap(Op0, Op1); // Change icmp uge -> icmp ule
5816 case ICmpInst::ICMP_ULE: { // icmp ule i1 A, B -> ~A | B
5817 Instruction *Not = BinaryOperator::CreateNot(Op0, I.getName()+"tmp");
5818 InsertNewInstBefore(Not, I);
5819 return BinaryOperator::CreateOr(Not, Op1);
5821 case ICmpInst::ICMP_SGE:
5822 std::swap(Op0, Op1); // Change icmp sge -> icmp sle
5824 case ICmpInst::ICMP_SLE: { // icmp sle i1 A, B -> A | ~B
5825 Instruction *Not = BinaryOperator::CreateNot(Op1, I.getName()+"tmp");
5826 InsertNewInstBefore(Not, I);
5827 return BinaryOperator::CreateOr(Not, Op0);
5832 // See if we are doing a comparison with a constant.
5833 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
5834 Value *A = 0, *B = 0;
5836 // (icmp ne/eq (sub A B) 0) -> (icmp ne/eq A, B)
5837 if (I.isEquality() && CI->isNullValue() &&
5838 match(Op0, m_Sub(m_Value(A), m_Value(B)))) {
5839 // (icmp cond A B) if cond is equality
5840 return new ICmpInst(I.getPredicate(), A, B);
5843 // If we have an icmp le or icmp ge instruction, turn it into the
5844 // appropriate icmp lt or icmp gt instruction. This allows us to rely on
5845 // them being folded in the code below.
5846 switch (I.getPredicate()) {
5848 case ICmpInst::ICMP_ULE:
5849 if (CI->isMaxValue(false)) // A <=u MAX -> TRUE
5850 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5851 return new ICmpInst(ICmpInst::ICMP_ULT, Op0, AddOne(CI));
5852 case ICmpInst::ICMP_SLE:
5853 if (CI->isMaxValue(true)) // A <=s MAX -> TRUE
5854 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5855 return new ICmpInst(ICmpInst::ICMP_SLT, Op0, AddOne(CI));
5856 case ICmpInst::ICMP_UGE:
5857 if (CI->isMinValue(false)) // A >=u MIN -> TRUE
5858 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5859 return new ICmpInst( ICmpInst::ICMP_UGT, Op0, SubOne(CI));
5860 case ICmpInst::ICMP_SGE:
5861 if (CI->isMinValue(true)) // A >=s MIN -> TRUE
5862 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5863 return new ICmpInst(ICmpInst::ICMP_SGT, Op0, SubOne(CI));
5866 // See if we can fold the comparison based on range information we can get
5867 // by checking whether bits are known to be zero or one in the input.
5868 uint32_t BitWidth = cast<IntegerType>(Ty)->getBitWidth();
5869 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
5871 // If this comparison is a normal comparison, it demands all
5872 // bits, if it is a sign bit comparison, it only demands the sign bit.
5874 bool isSignBit = isSignBitCheck(I.getPredicate(), CI, UnusedBit);
5876 if (SimplifyDemandedBits(I.getOperandUse(0),
5877 isSignBit ? APInt::getSignBit(BitWidth)
5878 : APInt::getAllOnesValue(BitWidth),
5879 KnownZero, KnownOne, 0))
5882 // Given the known and unknown bits, compute a range that the LHS could be
5883 // in. Compute the Min, Max and RHS values based on the known bits. For the
5884 // EQ and NE we use unsigned values.
5885 APInt Min(BitWidth, 0), Max(BitWidth, 0);
5886 if (ICmpInst::isSignedPredicate(I.getPredicate()))
5887 ComputeSignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne, Min, Max);
5889 ComputeUnsignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne,Min,Max);
5891 // If Min and Max are known to be the same, then SimplifyDemandedBits
5892 // figured out that the LHS is a constant. Just constant fold this now so
5893 // that code below can assume that Min != Max.
5895 return ReplaceInstUsesWith(I, ConstantExpr::getICmp(I.getPredicate(),
5896 ConstantInt::get(Min),
5899 // Based on the range information we know about the LHS, see if we can
5900 // simplify this comparison. For example, (x&4) < 8 is always true.
5901 const APInt &RHSVal = CI->getValue();
5902 switch (I.getPredicate()) { // LE/GE have been folded already.
5903 default: assert(0 && "Unknown icmp opcode!");
5904 case ICmpInst::ICMP_EQ:
5905 if (Max.ult(RHSVal) || Min.ugt(RHSVal))
5906 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5908 case ICmpInst::ICMP_NE:
5909 if (Max.ult(RHSVal) || Min.ugt(RHSVal))
5910 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5912 case ICmpInst::ICMP_ULT:
5913 if (Max.ult(RHSVal)) // A <u C -> true iff max(A) < C
5914 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5915 if (Min.uge(RHSVal)) // A <u C -> false iff min(A) >= C
5916 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5917 if (RHSVal == Max) // A <u MAX -> A != MAX
5918 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5919 if (RHSVal == Min+1) // A <u MIN+1 -> A == MIN
5920 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, SubOne(CI));
5922 // (x <u 2147483648) -> (x >s -1) -> true if sign bit clear
5923 if (CI->isMinValue(true))
5924 return new ICmpInst(ICmpInst::ICMP_SGT, Op0,
5925 ConstantInt::getAllOnesValue(Op0->getType()));
5927 case ICmpInst::ICMP_UGT:
5928 if (Min.ugt(RHSVal)) // A >u C -> true iff min(A) > C
5929 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5930 if (Max.ule(RHSVal)) // A >u C -> false iff max(A) <= C
5931 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5933 if (RHSVal == Min) // A >u MIN -> A != MIN
5934 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5935 if (RHSVal == Max-1) // A >u MAX-1 -> A == MAX
5936 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, AddOne(CI));
5938 // (x >u 2147483647) -> (x <s 0) -> true if sign bit set
5939 if (CI->isMaxValue(true))
5940 return new ICmpInst(ICmpInst::ICMP_SLT, Op0,
5941 ConstantInt::getNullValue(Op0->getType()));
5943 case ICmpInst::ICMP_SLT:
5944 if (Max.slt(RHSVal)) // A <s C -> true iff max(A) < C
5945 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5946 if (Min.sge(RHSVal)) // A <s C -> false iff min(A) >= C
5947 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5948 if (RHSVal == Max) // A <s MAX -> A != MAX
5949 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5950 if (RHSVal == Min+1) // A <s MIN+1 -> A == MIN
5951 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, SubOne(CI));
5953 case ICmpInst::ICMP_SGT:
5954 if (Min.sgt(RHSVal)) // A >s C -> true iff min(A) > C
5955 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5956 if (Max.sle(RHSVal)) // A >s C -> false iff max(A) <= C
5957 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5959 if (RHSVal == Min) // A >s MIN -> A != MIN
5960 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5961 if (RHSVal == Max-1) // A >s MAX-1 -> A == MAX
5962 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, AddOne(CI));
5967 // Test if the ICmpInst instruction is used exclusively by a select as
5968 // part of a minimum or maximum operation. If so, refrain from doing
5969 // any other folding. This helps out other analyses which understand
5970 // non-obfuscated minimum and maximum idioms, such as ScalarEvolution
5971 // and CodeGen. And in this case, at least one of the comparison
5972 // operands has at least one user besides the compare (the select),
5973 // which would often largely negate the benefit of folding anyway.
5975 if (SelectInst *SI = dyn_cast<SelectInst>(*I.use_begin()))
5976 if ((SI->getOperand(1) == Op0 && SI->getOperand(2) == Op1) ||
5977 (SI->getOperand(2) == Op0 && SI->getOperand(1) == Op1))
5980 // See if we are doing a comparison between a constant and an instruction that
5981 // can be folded into the comparison.
5982 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
5983 // Since the RHS is a ConstantInt (CI), if the left hand side is an
5984 // instruction, see if that instruction also has constants so that the
5985 // instruction can be folded into the icmp
5986 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
5987 if (Instruction *Res = visitICmpInstWithInstAndIntCst(I, LHSI, CI))
5991 // Handle icmp with constant (but not simple integer constant) RHS
5992 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
5993 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
5994 switch (LHSI->getOpcode()) {
5995 case Instruction::GetElementPtr:
5996 if (RHSC->isNullValue()) {
5997 // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
5998 bool isAllZeros = true;
5999 for (unsigned i = 1, e = LHSI->getNumOperands(); i != e; ++i)
6000 if (!isa<Constant>(LHSI->getOperand(i)) ||
6001 !cast<Constant>(LHSI->getOperand(i))->isNullValue()) {
6006 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
6007 Constant::getNullValue(LHSI->getOperand(0)->getType()));
6011 case Instruction::PHI:
6012 // Only fold icmp into the PHI if the phi and fcmp are in the same
6013 // block. If in the same block, we're encouraging jump threading. If
6014 // not, we are just pessimizing the code by making an i1 phi.
6015 if (LHSI->getParent() == I.getParent())
6016 if (Instruction *NV = FoldOpIntoPhi(I))
6019 case Instruction::Select: {
6020 // If either operand of the select is a constant, we can fold the
6021 // comparison into the select arms, which will cause one to be
6022 // constant folded and the select turned into a bitwise or.
6023 Value *Op1 = 0, *Op2 = 0;
6024 if (LHSI->hasOneUse()) {
6025 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
6026 // Fold the known value into the constant operand.
6027 Op1 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
6028 // Insert a new ICmp of the other select operand.
6029 Op2 = InsertNewInstBefore(new ICmpInst(I.getPredicate(),
6030 LHSI->getOperand(2), RHSC,
6032 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
6033 // Fold the known value into the constant operand.
6034 Op2 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
6035 // Insert a new ICmp of the other select operand.
6036 Op1 = InsertNewInstBefore(new ICmpInst(I.getPredicate(),
6037 LHSI->getOperand(1), RHSC,
6043 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
6046 case Instruction::Malloc:
6047 // If we have (malloc != null), and if the malloc has a single use, we
6048 // can assume it is successful and remove the malloc.
6049 if (LHSI->hasOneUse() && isa<ConstantPointerNull>(RHSC)) {
6050 AddToWorkList(LHSI);
6051 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
6052 !I.isTrueWhenEqual()));
6058 // If we can optimize a 'icmp GEP, P' or 'icmp P, GEP', do so now.
6059 if (User *GEP = dyn_castGetElementPtr(Op0))
6060 if (Instruction *NI = FoldGEPICmp(GEP, Op1, I.getPredicate(), I))
6062 if (User *GEP = dyn_castGetElementPtr(Op1))
6063 if (Instruction *NI = FoldGEPICmp(GEP, Op0,
6064 ICmpInst::getSwappedPredicate(I.getPredicate()), I))
6067 // Test to see if the operands of the icmp are casted versions of other
6068 // values. If the ptr->ptr cast can be stripped off both arguments, we do so
6070 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op0)) {
6071 if (isa<PointerType>(Op0->getType()) &&
6072 (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
6073 // We keep moving the cast from the left operand over to the right
6074 // operand, where it can often be eliminated completely.
6075 Op0 = CI->getOperand(0);
6077 // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast
6078 // so eliminate it as well.
6079 if (BitCastInst *CI2 = dyn_cast<BitCastInst>(Op1))
6080 Op1 = CI2->getOperand(0);
6082 // If Op1 is a constant, we can fold the cast into the constant.
6083 if (Op0->getType() != Op1->getType()) {
6084 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
6085 Op1 = ConstantExpr::getBitCast(Op1C, Op0->getType());
6087 // Otherwise, cast the RHS right before the icmp
6088 Op1 = InsertBitCastBefore(Op1, Op0->getType(), I);
6091 return new ICmpInst(I.getPredicate(), Op0, Op1);
6095 if (isa<CastInst>(Op0)) {
6096 // Handle the special case of: icmp (cast bool to X), <cst>
6097 // This comes up when you have code like
6100 // For generality, we handle any zero-extension of any operand comparison
6101 // with a constant or another cast from the same type.
6102 if (isa<ConstantInt>(Op1) || isa<CastInst>(Op1))
6103 if (Instruction *R = visitICmpInstWithCastAndCast(I))
6107 // See if it's the same type of instruction on the left and right.
6108 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
6109 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
6110 if (Op0I->getOpcode() == Op1I->getOpcode() && Op0I->hasOneUse() &&
6111 Op1I->hasOneUse() && Op0I->getOperand(1) == Op1I->getOperand(1)) {
6112 switch (Op0I->getOpcode()) {
6114 case Instruction::Add:
6115 case Instruction::Sub:
6116 case Instruction::Xor:
6117 if (I.isEquality()) // a+x icmp eq/ne b+x --> a icmp b
6118 return new ICmpInst(I.getPredicate(), Op0I->getOperand(0),
6119 Op1I->getOperand(0));
6120 // icmp u/s (a ^ signbit), (b ^ signbit) --> icmp s/u a, b
6121 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) {
6122 if (CI->getValue().isSignBit()) {
6123 ICmpInst::Predicate Pred = I.isSignedPredicate()
6124 ? I.getUnsignedPredicate()
6125 : I.getSignedPredicate();
6126 return new ICmpInst(Pred, Op0I->getOperand(0),
6127 Op1I->getOperand(0));
6130 if (CI->getValue().isMaxSignedValue()) {
6131 ICmpInst::Predicate Pred = I.isSignedPredicate()
6132 ? I.getUnsignedPredicate()
6133 : I.getSignedPredicate();
6134 Pred = I.getSwappedPredicate(Pred);
6135 return new ICmpInst(Pred, Op0I->getOperand(0),
6136 Op1I->getOperand(0));
6140 case Instruction::Mul:
6141 if (!I.isEquality())
6144 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) {
6145 // a * Cst icmp eq/ne b * Cst --> a & Mask icmp b & Mask
6146 // Mask = -1 >> count-trailing-zeros(Cst).
6147 if (!CI->isZero() && !CI->isOne()) {
6148 const APInt &AP = CI->getValue();
6149 ConstantInt *Mask = ConstantInt::get(
6150 APInt::getLowBitsSet(AP.getBitWidth(),
6152 AP.countTrailingZeros()));
6153 Instruction *And1 = BinaryOperator::CreateAnd(Op0I->getOperand(0),
6155 Instruction *And2 = BinaryOperator::CreateAnd(Op1I->getOperand(0),
6157 InsertNewInstBefore(And1, I);
6158 InsertNewInstBefore(And2, I);
6159 return new ICmpInst(I.getPredicate(), And1, And2);
6168 // ~x < ~y --> y < x
6170 if (match(Op0, m_Not(m_Value(A))) &&
6171 match(Op1, m_Not(m_Value(B))))
6172 return new ICmpInst(I.getPredicate(), B, A);
6175 if (I.isEquality()) {
6176 Value *A, *B, *C, *D;
6178 // -x == -y --> x == y
6179 if (match(Op0, m_Neg(m_Value(A))) &&
6180 match(Op1, m_Neg(m_Value(B))))
6181 return new ICmpInst(I.getPredicate(), A, B);
6183 if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
6184 if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0
6185 Value *OtherVal = A == Op1 ? B : A;
6186 return new ICmpInst(I.getPredicate(), OtherVal,
6187 Constant::getNullValue(A->getType()));
6190 if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) {
6191 // A^c1 == C^c2 --> A == C^(c1^c2)
6192 ConstantInt *C1, *C2;
6193 if (match(B, m_ConstantInt(C1)) &&
6194 match(D, m_ConstantInt(C2)) && Op1->hasOneUse()) {
6195 Constant *NC = ConstantInt::get(C1->getValue() ^ C2->getValue());
6196 Instruction *Xor = BinaryOperator::CreateXor(C, NC, "tmp");
6197 return new ICmpInst(I.getPredicate(), A,
6198 InsertNewInstBefore(Xor, I));
6201 // A^B == A^D -> B == D
6202 if (A == C) return new ICmpInst(I.getPredicate(), B, D);
6203 if (A == D) return new ICmpInst(I.getPredicate(), B, C);
6204 if (B == C) return new ICmpInst(I.getPredicate(), A, D);
6205 if (B == D) return new ICmpInst(I.getPredicate(), A, C);
6209 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
6210 (A == Op0 || B == Op0)) {
6211 // A == (A^B) -> B == 0
6212 Value *OtherVal = A == Op0 ? B : A;
6213 return new ICmpInst(I.getPredicate(), OtherVal,
6214 Constant::getNullValue(A->getType()));
6217 // (A-B) == A -> B == 0
6218 if (match(Op0, m_Sub(m_Specific(Op1), m_Value(B))))
6219 return new ICmpInst(I.getPredicate(), B,
6220 Constant::getNullValue(B->getType()));
6222 // A == (A-B) -> B == 0
6223 if (match(Op1, m_Sub(m_Specific(Op0), m_Value(B))))
6224 return new ICmpInst(I.getPredicate(), B,
6225 Constant::getNullValue(B->getType()));
6227 // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
6228 if (Op0->hasOneUse() && Op1->hasOneUse() &&
6229 match(Op0, m_And(m_Value(A), m_Value(B))) &&
6230 match(Op1, m_And(m_Value(C), m_Value(D)))) {
6231 Value *X = 0, *Y = 0, *Z = 0;
6234 X = B; Y = D; Z = A;
6235 } else if (A == D) {
6236 X = B; Y = C; Z = A;
6237 } else if (B == C) {
6238 X = A; Y = D; Z = B;
6239 } else if (B == D) {
6240 X = A; Y = C; Z = B;
6243 if (X) { // Build (X^Y) & Z
6244 Op1 = InsertNewInstBefore(BinaryOperator::CreateXor(X, Y, "tmp"), I);
6245 Op1 = InsertNewInstBefore(BinaryOperator::CreateAnd(Op1, Z, "tmp"), I);
6246 I.setOperand(0, Op1);
6247 I.setOperand(1, Constant::getNullValue(Op1->getType()));
6252 return Changed ? &I : 0;
6256 /// FoldICmpDivCst - Fold "icmp pred, ([su]div X, DivRHS), CmpRHS" where DivRHS
6257 /// and CmpRHS are both known to be integer constants.
6258 Instruction *InstCombiner::FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
6259 ConstantInt *DivRHS) {
6260 ConstantInt *CmpRHS = cast<ConstantInt>(ICI.getOperand(1));
6261 const APInt &CmpRHSV = CmpRHS->getValue();
6263 // FIXME: If the operand types don't match the type of the divide
6264 // then don't attempt this transform. The code below doesn't have the
6265 // logic to deal with a signed divide and an unsigned compare (and
6266 // vice versa). This is because (x /s C1) <s C2 produces different
6267 // results than (x /s C1) <u C2 or (x /u C1) <s C2 or even
6268 // (x /u C1) <u C2. Simply casting the operands and result won't
6269 // work. :( The if statement below tests that condition and bails
6271 bool DivIsSigned = DivI->getOpcode() == Instruction::SDiv;
6272 if (!ICI.isEquality() && DivIsSigned != ICI.isSignedPredicate())
6274 if (DivRHS->isZero())
6275 return 0; // The ProdOV computation fails on divide by zero.
6276 if (DivIsSigned && DivRHS->isAllOnesValue())
6277 return 0; // The overflow computation also screws up here
6278 if (DivRHS->isOne())
6279 return 0; // Not worth bothering, and eliminates some funny cases
6282 // Compute Prod = CI * DivRHS. We are essentially solving an equation
6283 // of form X/C1=C2. We solve for X by multiplying C1 (DivRHS) and
6284 // C2 (CI). By solving for X we can turn this into a range check
6285 // instead of computing a divide.
6286 ConstantInt *Prod = Multiply(CmpRHS, DivRHS);
6288 // Determine if the product overflows by seeing if the product is
6289 // not equal to the divide. Make sure we do the same kind of divide
6290 // as in the LHS instruction that we're folding.
6291 bool ProdOV = (DivIsSigned ? ConstantExpr::getSDiv(Prod, DivRHS) :
6292 ConstantExpr::getUDiv(Prod, DivRHS)) != CmpRHS;
6294 // Get the ICmp opcode
6295 ICmpInst::Predicate Pred = ICI.getPredicate();
6297 // Figure out the interval that is being checked. For example, a comparison
6298 // like "X /u 5 == 0" is really checking that X is in the interval [0, 5).
6299 // Compute this interval based on the constants involved and the signedness of
6300 // the compare/divide. This computes a half-open interval, keeping track of
6301 // whether either value in the interval overflows. After analysis each
6302 // overflow variable is set to 0 if it's corresponding bound variable is valid
6303 // -1 if overflowed off the bottom end, or +1 if overflowed off the top end.
6304 int LoOverflow = 0, HiOverflow = 0;
6305 ConstantInt *LoBound = 0, *HiBound = 0;
6307 if (!DivIsSigned) { // udiv
6308 // e.g. X/5 op 3 --> [15, 20)
6310 HiOverflow = LoOverflow = ProdOV;
6312 HiOverflow = AddWithOverflow(HiBound, LoBound, DivRHS, false);
6313 } else if (DivRHS->getValue().isStrictlyPositive()) { // Divisor is > 0.
6314 if (CmpRHSV == 0) { // (X / pos) op 0
6315 // Can't overflow. e.g. X/2 op 0 --> [-1, 2)
6316 LoBound = cast<ConstantInt>(ConstantExpr::getNeg(SubOne(DivRHS)));
6318 } else if (CmpRHSV.isStrictlyPositive()) { // (X / pos) op pos
6319 LoBound = Prod; // e.g. X/5 op 3 --> [15, 20)
6320 HiOverflow = LoOverflow = ProdOV;
6322 HiOverflow = AddWithOverflow(HiBound, Prod, DivRHS, true);
6323 } else { // (X / pos) op neg
6324 // e.g. X/5 op -3 --> [-15-4, -15+1) --> [-19, -14)
6325 HiBound = AddOne(Prod);
6326 LoOverflow = HiOverflow = ProdOV ? -1 : 0;
6328 ConstantInt* DivNeg = cast<ConstantInt>(ConstantExpr::getNeg(DivRHS));
6329 LoOverflow = AddWithOverflow(LoBound, HiBound, DivNeg,
6333 } else if (DivRHS->getValue().isNegative()) { // Divisor is < 0.
6334 if (CmpRHSV == 0) { // (X / neg) op 0
6335 // e.g. X/-5 op 0 --> [-4, 5)
6336 LoBound = AddOne(DivRHS);
6337 HiBound = cast<ConstantInt>(ConstantExpr::getNeg(DivRHS));
6338 if (HiBound == DivRHS) { // -INTMIN = INTMIN
6339 HiOverflow = 1; // [INTMIN+1, overflow)
6340 HiBound = 0; // e.g. X/INTMIN = 0 --> X > INTMIN
6342 } else if (CmpRHSV.isStrictlyPositive()) { // (X / neg) op pos
6343 // e.g. X/-5 op 3 --> [-19, -14)
6344 HiBound = AddOne(Prod);
6345 HiOverflow = LoOverflow = ProdOV ? -1 : 0;
6347 LoOverflow = AddWithOverflow(LoBound, HiBound, DivRHS, true) ? -1 : 0;
6348 } else { // (X / neg) op neg
6349 LoBound = Prod; // e.g. X/-5 op -3 --> [15, 20)
6350 LoOverflow = HiOverflow = ProdOV;
6352 HiOverflow = SubWithOverflow(HiBound, Prod, DivRHS, true);
6355 // Dividing by a negative swaps the condition. LT <-> GT
6356 Pred = ICmpInst::getSwappedPredicate(Pred);
6359 Value *X = DivI->getOperand(0);
6361 default: assert(0 && "Unhandled icmp opcode!");
6362 case ICmpInst::ICMP_EQ:
6363 if (LoOverflow && HiOverflow)
6364 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
6365 else if (HiOverflow)
6366 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
6367 ICmpInst::ICMP_UGE, X, LoBound);
6368 else if (LoOverflow)
6369 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
6370 ICmpInst::ICMP_ULT, X, HiBound);
6372 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned, true, ICI);
6373 case ICmpInst::ICMP_NE:
6374 if (LoOverflow && HiOverflow)
6375 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
6376 else if (HiOverflow)
6377 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
6378 ICmpInst::ICMP_ULT, X, LoBound);
6379 else if (LoOverflow)
6380 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
6381 ICmpInst::ICMP_UGE, X, HiBound);
6383 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned, false, ICI);
6384 case ICmpInst::ICMP_ULT:
6385 case ICmpInst::ICMP_SLT:
6386 if (LoOverflow == +1) // Low bound is greater than input range.
6387 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
6388 if (LoOverflow == -1) // Low bound is less than input range.
6389 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
6390 return new ICmpInst(Pred, X, LoBound);
6391 case ICmpInst::ICMP_UGT:
6392 case ICmpInst::ICMP_SGT:
6393 if (HiOverflow == +1) // High bound greater than input range.
6394 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
6395 else if (HiOverflow == -1) // High bound less than input range.
6396 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
6397 if (Pred == ICmpInst::ICMP_UGT)
6398 return new ICmpInst(ICmpInst::ICMP_UGE, X, HiBound);
6400 return new ICmpInst(ICmpInst::ICMP_SGE, X, HiBound);
6405 /// visitICmpInstWithInstAndIntCst - Handle "icmp (instr, intcst)".
6407 Instruction *InstCombiner::visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
6410 const APInt &RHSV = RHS->getValue();
6412 switch (LHSI->getOpcode()) {
6413 case Instruction::Trunc:
6414 if (ICI.isEquality() && LHSI->hasOneUse()) {
6415 // Simplify icmp eq (trunc x to i8), 42 -> icmp eq x, 42|highbits if all
6416 // of the high bits truncated out of x are known.
6417 unsigned DstBits = LHSI->getType()->getPrimitiveSizeInBits(),
6418 SrcBits = LHSI->getOperand(0)->getType()->getPrimitiveSizeInBits();
6419 APInt Mask(APInt::getHighBitsSet(SrcBits, SrcBits-DstBits));
6420 APInt KnownZero(SrcBits, 0), KnownOne(SrcBits, 0);
6421 ComputeMaskedBits(LHSI->getOperand(0), Mask, KnownZero, KnownOne);
6423 // If all the high bits are known, we can do this xform.
6424 if ((KnownZero|KnownOne).countLeadingOnes() >= SrcBits-DstBits) {
6425 // Pull in the high bits from known-ones set.
6426 APInt NewRHS(RHS->getValue());
6427 NewRHS.zext(SrcBits);
6429 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
6430 ConstantInt::get(NewRHS));
6435 case Instruction::Xor: // (icmp pred (xor X, XorCST), CI)
6436 if (ConstantInt *XorCST = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
6437 // If this is a comparison that tests the signbit (X < 0) or (x > -1),
6439 if ((ICI.getPredicate() == ICmpInst::ICMP_SLT && RHSV == 0) ||
6440 (ICI.getPredicate() == ICmpInst::ICMP_SGT && RHSV.isAllOnesValue())) {
6441 Value *CompareVal = LHSI->getOperand(0);
6443 // If the sign bit of the XorCST is not set, there is no change to
6444 // the operation, just stop using the Xor.
6445 if (!XorCST->getValue().isNegative()) {
6446 ICI.setOperand(0, CompareVal);
6447 AddToWorkList(LHSI);
6451 // Was the old condition true if the operand is positive?
6452 bool isTrueIfPositive = ICI.getPredicate() == ICmpInst::ICMP_SGT;
6454 // If so, the new one isn't.
6455 isTrueIfPositive ^= true;
6457 if (isTrueIfPositive)
6458 return new ICmpInst(ICmpInst::ICMP_SGT, CompareVal, SubOne(RHS));
6460 return new ICmpInst(ICmpInst::ICMP_SLT, CompareVal, AddOne(RHS));
6463 if (LHSI->hasOneUse()) {
6464 // (icmp u/s (xor A SignBit), C) -> (icmp s/u A, (xor C SignBit))
6465 if (!ICI.isEquality() && XorCST->getValue().isSignBit()) {
6466 const APInt &SignBit = XorCST->getValue();
6467 ICmpInst::Predicate Pred = ICI.isSignedPredicate()
6468 ? ICI.getUnsignedPredicate()
6469 : ICI.getSignedPredicate();
6470 return new ICmpInst(Pred, LHSI->getOperand(0),
6471 ConstantInt::get(RHSV ^ SignBit));
6474 // (icmp u/s (xor A ~SignBit), C) -> (icmp s/u (xor C ~SignBit), A)
6475 if (!ICI.isEquality() && XorCST->getValue().isMaxSignedValue()) {
6476 const APInt &NotSignBit = XorCST->getValue();
6477 ICmpInst::Predicate Pred = ICI.isSignedPredicate()
6478 ? ICI.getUnsignedPredicate()
6479 : ICI.getSignedPredicate();
6480 Pred = ICI.getSwappedPredicate(Pred);
6481 return new ICmpInst(Pred, LHSI->getOperand(0),
6482 ConstantInt::get(RHSV ^ NotSignBit));
6487 case Instruction::And: // (icmp pred (and X, AndCST), RHS)
6488 if (LHSI->hasOneUse() && isa<ConstantInt>(LHSI->getOperand(1)) &&
6489 LHSI->getOperand(0)->hasOneUse()) {
6490 ConstantInt *AndCST = cast<ConstantInt>(LHSI->getOperand(1));
6492 // If the LHS is an AND of a truncating cast, we can widen the
6493 // and/compare to be the input width without changing the value
6494 // produced, eliminating a cast.
6495 if (TruncInst *Cast = dyn_cast<TruncInst>(LHSI->getOperand(0))) {
6496 // We can do this transformation if either the AND constant does not
6497 // have its sign bit set or if it is an equality comparison.
6498 // Extending a relational comparison when we're checking the sign
6499 // bit would not work.
6500 if (Cast->hasOneUse() &&
6501 (ICI.isEquality() ||
6502 (AndCST->getValue().isNonNegative() && RHSV.isNonNegative()))) {
6504 cast<IntegerType>(Cast->getOperand(0)->getType())->getBitWidth();
6505 APInt NewCST = AndCST->getValue();
6506 NewCST.zext(BitWidth);
6508 NewCI.zext(BitWidth);
6509 Instruction *NewAnd =
6510 BinaryOperator::CreateAnd(Cast->getOperand(0),
6511 ConstantInt::get(NewCST),LHSI->getName());
6512 InsertNewInstBefore(NewAnd, ICI);
6513 return new ICmpInst(ICI.getPredicate(), NewAnd,
6514 ConstantInt::get(NewCI));
6518 // If this is: (X >> C1) & C2 != C3 (where any shift and any compare
6519 // could exist), turn it into (X & (C2 << C1)) != (C3 << C1). This
6520 // happens a LOT in code produced by the C front-end, for bitfield
6522 BinaryOperator *Shift = dyn_cast<BinaryOperator>(LHSI->getOperand(0));
6523 if (Shift && !Shift->isShift())
6527 ShAmt = Shift ? dyn_cast<ConstantInt>(Shift->getOperand(1)) : 0;
6528 const Type *Ty = Shift ? Shift->getType() : 0; // Type of the shift.
6529 const Type *AndTy = AndCST->getType(); // Type of the and.
6531 // We can fold this as long as we can't shift unknown bits
6532 // into the mask. This can only happen with signed shift
6533 // rights, as they sign-extend.
6535 bool CanFold = Shift->isLogicalShift();
6537 // To test for the bad case of the signed shr, see if any
6538 // of the bits shifted in could be tested after the mask.
6539 uint32_t TyBits = Ty->getPrimitiveSizeInBits();
6540 int ShAmtVal = TyBits - ShAmt->getLimitedValue(TyBits);
6542 uint32_t BitWidth = AndTy->getPrimitiveSizeInBits();
6543 if ((APInt::getHighBitsSet(BitWidth, BitWidth-ShAmtVal) &
6544 AndCST->getValue()) == 0)
6550 if (Shift->getOpcode() == Instruction::Shl)
6551 NewCst = ConstantExpr::getLShr(RHS, ShAmt);
6553 NewCst = ConstantExpr::getShl(RHS, ShAmt);
6555 // Check to see if we are shifting out any of the bits being
6557 if (ConstantExpr::get(Shift->getOpcode(), NewCst, ShAmt) != RHS) {
6558 // If we shifted bits out, the fold is not going to work out.
6559 // As a special case, check to see if this means that the
6560 // result is always true or false now.
6561 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
6562 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
6563 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
6564 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
6566 ICI.setOperand(1, NewCst);
6567 Constant *NewAndCST;
6568 if (Shift->getOpcode() == Instruction::Shl)
6569 NewAndCST = ConstantExpr::getLShr(AndCST, ShAmt);
6571 NewAndCST = ConstantExpr::getShl(AndCST, ShAmt);
6572 LHSI->setOperand(1, NewAndCST);
6573 LHSI->setOperand(0, Shift->getOperand(0));
6574 AddToWorkList(Shift); // Shift is dead.
6575 AddUsesToWorkList(ICI);
6581 // Turn ((X >> Y) & C) == 0 into (X & (C << Y)) == 0. The later is
6582 // preferable because it allows the C<<Y expression to be hoisted out
6583 // of a loop if Y is invariant and X is not.
6584 if (Shift && Shift->hasOneUse() && RHSV == 0 &&
6585 ICI.isEquality() && !Shift->isArithmeticShift()/* &&
6586 isa<Instruction>(Shift->getOperand(0))*/) {
6589 if (Shift->getOpcode() == Instruction::LShr) {
6590 NS = BinaryOperator::CreateShl(AndCST,
6591 Shift->getOperand(1), "tmp");
6593 // Insert a logical shift.
6594 NS = BinaryOperator::CreateLShr(AndCST,
6595 Shift->getOperand(1), "tmp");
6597 InsertNewInstBefore(cast<Instruction>(NS), ICI);
6599 // Compute X & (C << Y).
6600 Instruction *NewAnd =
6601 BinaryOperator::CreateAnd(Shift->getOperand(0), NS, LHSI->getName());
6602 InsertNewInstBefore(NewAnd, ICI);
6604 ICI.setOperand(0, NewAnd);
6610 case Instruction::Shl: { // (icmp pred (shl X, ShAmt), CI)
6611 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
6614 uint32_t TypeBits = RHSV.getBitWidth();
6616 // Check that the shift amount is in range. If not, don't perform
6617 // undefined shifts. When the shift is visited it will be
6619 if (ShAmt->uge(TypeBits))
6622 if (ICI.isEquality()) {
6623 // If we are comparing against bits always shifted out, the
6624 // comparison cannot succeed.
6626 ConstantExpr::getShl(ConstantExpr::getLShr(RHS, ShAmt), ShAmt);
6627 if (Comp != RHS) {// Comparing against a bit that we know is zero.
6628 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
6629 Constant *Cst = ConstantInt::get(Type::Int1Ty, IsICMP_NE);
6630 return ReplaceInstUsesWith(ICI, Cst);
6633 if (LHSI->hasOneUse()) {
6634 // Otherwise strength reduce the shift into an and.
6635 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
6637 ConstantInt::get(APInt::getLowBitsSet(TypeBits, TypeBits-ShAmtVal));
6640 BinaryOperator::CreateAnd(LHSI->getOperand(0),
6641 Mask, LHSI->getName()+".mask");
6642 Value *And = InsertNewInstBefore(AndI, ICI);
6643 return new ICmpInst(ICI.getPredicate(), And,
6644 ConstantInt::get(RHSV.lshr(ShAmtVal)));
6648 // Otherwise, if this is a comparison of the sign bit, simplify to and/test.
6649 bool TrueIfSigned = false;
6650 if (LHSI->hasOneUse() &&
6651 isSignBitCheck(ICI.getPredicate(), RHS, TrueIfSigned)) {
6652 // (X << 31) <s 0 --> (X&1) != 0
6653 Constant *Mask = ConstantInt::get(APInt(TypeBits, 1) <<
6654 (TypeBits-ShAmt->getZExtValue()-1));
6656 BinaryOperator::CreateAnd(LHSI->getOperand(0),
6657 Mask, LHSI->getName()+".mask");
6658 Value *And = InsertNewInstBefore(AndI, ICI);
6660 return new ICmpInst(TrueIfSigned ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ,
6661 And, Constant::getNullValue(And->getType()));
6666 case Instruction::LShr: // (icmp pred (shr X, ShAmt), CI)
6667 case Instruction::AShr: {
6668 // Only handle equality comparisons of shift-by-constant.
6669 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
6670 if (!ShAmt || !ICI.isEquality()) break;
6672 // Check that the shift amount is in range. If not, don't perform
6673 // undefined shifts. When the shift is visited it will be
6675 uint32_t TypeBits = RHSV.getBitWidth();
6676 if (ShAmt->uge(TypeBits))
6679 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
6681 // If we are comparing against bits always shifted out, the
6682 // comparison cannot succeed.
6683 APInt Comp = RHSV << ShAmtVal;
6684 if (LHSI->getOpcode() == Instruction::LShr)
6685 Comp = Comp.lshr(ShAmtVal);
6687 Comp = Comp.ashr(ShAmtVal);
6689 if (Comp != RHSV) { // Comparing against a bit that we know is zero.
6690 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
6691 Constant *Cst = ConstantInt::get(Type::Int1Ty, IsICMP_NE);
6692 return ReplaceInstUsesWith(ICI, Cst);
6695 // Otherwise, check to see if the bits shifted out are known to be zero.
6696 // If so, we can compare against the unshifted value:
6697 // (X & 4) >> 1 == 2 --> (X & 4) == 4.
6698 if (LHSI->hasOneUse() &&
6699 MaskedValueIsZero(LHSI->getOperand(0),
6700 APInt::getLowBitsSet(Comp.getBitWidth(), ShAmtVal))) {
6701 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
6702 ConstantExpr::getShl(RHS, ShAmt));
6705 if (LHSI->hasOneUse()) {
6706 // Otherwise strength reduce the shift into an and.
6707 APInt Val(APInt::getHighBitsSet(TypeBits, TypeBits - ShAmtVal));
6708 Constant *Mask = ConstantInt::get(Val);
6711 BinaryOperator::CreateAnd(LHSI->getOperand(0),
6712 Mask, LHSI->getName()+".mask");
6713 Value *And = InsertNewInstBefore(AndI, ICI);
6714 return new ICmpInst(ICI.getPredicate(), And,
6715 ConstantExpr::getShl(RHS, ShAmt));
6720 case Instruction::SDiv:
6721 case Instruction::UDiv:
6722 // Fold: icmp pred ([us]div X, C1), C2 -> range test
6723 // Fold this div into the comparison, producing a range check.
6724 // Determine, based on the divide type, what the range is being
6725 // checked. If there is an overflow on the low or high side, remember
6726 // it, otherwise compute the range [low, hi) bounding the new value.
6727 // See: InsertRangeTest above for the kinds of replacements possible.
6728 if (ConstantInt *DivRHS = dyn_cast<ConstantInt>(LHSI->getOperand(1)))
6729 if (Instruction *R = FoldICmpDivCst(ICI, cast<BinaryOperator>(LHSI),
6734 case Instruction::Add:
6735 // Fold: icmp pred (add, X, C1), C2
6737 if (!ICI.isEquality()) {
6738 ConstantInt *LHSC = dyn_cast<ConstantInt>(LHSI->getOperand(1));
6740 const APInt &LHSV = LHSC->getValue();
6742 ConstantRange CR = ICI.makeConstantRange(ICI.getPredicate(), RHSV)
6745 if (ICI.isSignedPredicate()) {
6746 if (CR.getLower().isSignBit()) {
6747 return new ICmpInst(ICmpInst::ICMP_SLT, LHSI->getOperand(0),
6748 ConstantInt::get(CR.getUpper()));
6749 } else if (CR.getUpper().isSignBit()) {
6750 return new ICmpInst(ICmpInst::ICMP_SGE, LHSI->getOperand(0),
6751 ConstantInt::get(CR.getLower()));
6754 if (CR.getLower().isMinValue()) {
6755 return new ICmpInst(ICmpInst::ICMP_ULT, LHSI->getOperand(0),
6756 ConstantInt::get(CR.getUpper()));
6757 } else if (CR.getUpper().isMinValue()) {
6758 return new ICmpInst(ICmpInst::ICMP_UGE, LHSI->getOperand(0),
6759 ConstantInt::get(CR.getLower()));
6766 // Simplify icmp_eq and icmp_ne instructions with integer constant RHS.
6767 if (ICI.isEquality()) {
6768 bool isICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
6770 // If the first operand is (add|sub|and|or|xor|rem) with a constant, and
6771 // the second operand is a constant, simplify a bit.
6772 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(LHSI)) {
6773 switch (BO->getOpcode()) {
6774 case Instruction::SRem:
6775 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
6776 if (RHSV == 0 && isa<ConstantInt>(BO->getOperand(1)) &&BO->hasOneUse()){
6777 const APInt &V = cast<ConstantInt>(BO->getOperand(1))->getValue();
6778 if (V.sgt(APInt(V.getBitWidth(), 1)) && V.isPowerOf2()) {
6779 Instruction *NewRem =
6780 BinaryOperator::CreateURem(BO->getOperand(0), BO->getOperand(1),
6782 InsertNewInstBefore(NewRem, ICI);
6783 return new ICmpInst(ICI.getPredicate(), NewRem,
6784 Constant::getNullValue(BO->getType()));
6788 case Instruction::Add:
6789 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
6790 if (ConstantInt *BOp1C = dyn_cast<ConstantInt>(BO->getOperand(1))) {
6791 if (BO->hasOneUse())
6792 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
6793 Subtract(RHS, BOp1C));
6794 } else if (RHSV == 0) {
6795 // Replace ((add A, B) != 0) with (A != -B) if A or B is
6796 // efficiently invertible, or if the add has just this one use.
6797 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
6799 if (Value *NegVal = dyn_castNegVal(BOp1))
6800 return new ICmpInst(ICI.getPredicate(), BOp0, NegVal);
6801 else if (Value *NegVal = dyn_castNegVal(BOp0))
6802 return new ICmpInst(ICI.getPredicate(), NegVal, BOp1);
6803 else if (BO->hasOneUse()) {
6804 Instruction *Neg = BinaryOperator::CreateNeg(BOp1);
6805 InsertNewInstBefore(Neg, ICI);
6807 return new ICmpInst(ICI.getPredicate(), BOp0, Neg);
6811 case Instruction::Xor:
6812 // For the xor case, we can xor two constants together, eliminating
6813 // the explicit xor.
6814 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1)))
6815 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
6816 ConstantExpr::getXor(RHS, BOC));
6819 case Instruction::Sub:
6820 // Replace (([sub|xor] A, B) != 0) with (A != B)
6822 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
6826 case Instruction::Or:
6827 // If bits are being or'd in that are not present in the constant we
6828 // are comparing against, then the comparison could never succeed!
6829 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1))) {
6830 Constant *NotCI = ConstantExpr::getNot(RHS);
6831 if (!ConstantExpr::getAnd(BOC, NotCI)->isNullValue())
6832 return ReplaceInstUsesWith(ICI, ConstantInt::get(Type::Int1Ty,
6837 case Instruction::And:
6838 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
6839 // If bits are being compared against that are and'd out, then the
6840 // comparison can never succeed!
6841 if ((RHSV & ~BOC->getValue()) != 0)
6842 return ReplaceInstUsesWith(ICI, ConstantInt::get(Type::Int1Ty,
6845 // If we have ((X & C) == C), turn it into ((X & C) != 0).
6846 if (RHS == BOC && RHSV.isPowerOf2())
6847 return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ :
6848 ICmpInst::ICMP_NE, LHSI,
6849 Constant::getNullValue(RHS->getType()));
6851 // Replace (and X, (1 << size(X)-1) != 0) with x s< 0
6852 if (BOC->getValue().isSignBit()) {
6853 Value *X = BO->getOperand(0);
6854 Constant *Zero = Constant::getNullValue(X->getType());
6855 ICmpInst::Predicate pred = isICMP_NE ?
6856 ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
6857 return new ICmpInst(pred, X, Zero);
6860 // ((X & ~7) == 0) --> X < 8
6861 if (RHSV == 0 && isHighOnes(BOC)) {
6862 Value *X = BO->getOperand(0);
6863 Constant *NegX = ConstantExpr::getNeg(BOC);
6864 ICmpInst::Predicate pred = isICMP_NE ?
6865 ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
6866 return new ICmpInst(pred, X, NegX);
6871 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(LHSI)) {
6872 // Handle icmp {eq|ne} <intrinsic>, intcst.
6873 if (II->getIntrinsicID() == Intrinsic::bswap) {
6875 ICI.setOperand(0, II->getOperand(1));
6876 ICI.setOperand(1, ConstantInt::get(RHSV.byteSwap()));
6884 /// visitICmpInstWithCastAndCast - Handle icmp (cast x to y), (cast/cst).
6885 /// We only handle extending casts so far.
6887 Instruction *InstCombiner::visitICmpInstWithCastAndCast(ICmpInst &ICI) {
6888 const CastInst *LHSCI = cast<CastInst>(ICI.getOperand(0));
6889 Value *LHSCIOp = LHSCI->getOperand(0);
6890 const Type *SrcTy = LHSCIOp->getType();
6891 const Type *DestTy = LHSCI->getType();
6894 // Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the
6895 // integer type is the same size as the pointer type.
6896 if (LHSCI->getOpcode() == Instruction::PtrToInt &&
6897 getTargetData().getPointerSizeInBits() ==
6898 cast<IntegerType>(DestTy)->getBitWidth()) {
6900 if (Constant *RHSC = dyn_cast<Constant>(ICI.getOperand(1))) {
6901 RHSOp = ConstantExpr::getIntToPtr(RHSC, SrcTy);
6902 } else if (PtrToIntInst *RHSC = dyn_cast<PtrToIntInst>(ICI.getOperand(1))) {
6903 RHSOp = RHSC->getOperand(0);
6904 // If the pointer types don't match, insert a bitcast.
6905 if (LHSCIOp->getType() != RHSOp->getType())
6906 RHSOp = InsertBitCastBefore(RHSOp, LHSCIOp->getType(), ICI);
6910 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSOp);
6913 // The code below only handles extension cast instructions, so far.
6915 if (LHSCI->getOpcode() != Instruction::ZExt &&
6916 LHSCI->getOpcode() != Instruction::SExt)
6919 bool isSignedExt = LHSCI->getOpcode() == Instruction::SExt;
6920 bool isSignedCmp = ICI.isSignedPredicate();
6922 if (CastInst *CI = dyn_cast<CastInst>(ICI.getOperand(1))) {
6923 // Not an extension from the same type?
6924 RHSCIOp = CI->getOperand(0);
6925 if (RHSCIOp->getType() != LHSCIOp->getType())
6928 // If the signedness of the two casts doesn't agree (i.e. one is a sext
6929 // and the other is a zext), then we can't handle this.
6930 if (CI->getOpcode() != LHSCI->getOpcode())
6933 // Deal with equality cases early.
6934 if (ICI.isEquality())
6935 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
6937 // A signed comparison of sign extended values simplifies into a
6938 // signed comparison.
6939 if (isSignedCmp && isSignedExt)
6940 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
6942 // The other three cases all fold into an unsigned comparison.
6943 return new ICmpInst(ICI.getUnsignedPredicate(), LHSCIOp, RHSCIOp);
6946 // If we aren't dealing with a constant on the RHS, exit early
6947 ConstantInt *CI = dyn_cast<ConstantInt>(ICI.getOperand(1));
6951 // Compute the constant that would happen if we truncated to SrcTy then
6952 // reextended to DestTy.
6953 Constant *Res1 = ConstantExpr::getTrunc(CI, SrcTy);
6954 Constant *Res2 = ConstantExpr::getCast(LHSCI->getOpcode(), Res1, DestTy);
6956 // If the re-extended constant didn't change...
6958 // Make sure that sign of the Cmp and the sign of the Cast are the same.
6959 // For example, we might have:
6960 // %A = sext short %X to uint
6961 // %B = icmp ugt uint %A, 1330
6962 // It is incorrect to transform this into
6963 // %B = icmp ugt short %X, 1330
6964 // because %A may have negative value.
6966 // However, we allow this when the compare is EQ/NE, because they are
6968 if (isSignedExt == isSignedCmp || ICI.isEquality())
6969 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
6973 // The re-extended constant changed so the constant cannot be represented
6974 // in the shorter type. Consequently, we cannot emit a simple comparison.
6976 // First, handle some easy cases. We know the result cannot be equal at this
6977 // point so handle the ICI.isEquality() cases
6978 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
6979 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
6980 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
6981 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
6983 // Evaluate the comparison for LT (we invert for GT below). LE and GE cases
6984 // should have been folded away previously and not enter in here.
6987 // We're performing a signed comparison.
6988 if (cast<ConstantInt>(CI)->getValue().isNegative())
6989 Result = ConstantInt::getFalse(); // X < (small) --> false
6991 Result = ConstantInt::getTrue(); // X < (large) --> true
6993 // We're performing an unsigned comparison.
6995 // We're performing an unsigned comp with a sign extended value.
6996 // This is true if the input is >= 0. [aka >s -1]
6997 Constant *NegOne = ConstantInt::getAllOnesValue(SrcTy);
6998 Result = InsertNewInstBefore(new ICmpInst(ICmpInst::ICMP_SGT, LHSCIOp,
6999 NegOne, ICI.getName()), ICI);
7001 // Unsigned extend & unsigned compare -> always true.
7002 Result = ConstantInt::getTrue();
7006 // Finally, return the value computed.
7007 if (ICI.getPredicate() == ICmpInst::ICMP_ULT ||
7008 ICI.getPredicate() == ICmpInst::ICMP_SLT)
7009 return ReplaceInstUsesWith(ICI, Result);
7011 assert((ICI.getPredicate()==ICmpInst::ICMP_UGT ||
7012 ICI.getPredicate()==ICmpInst::ICMP_SGT) &&
7013 "ICmp should be folded!");
7014 if (Constant *CI = dyn_cast<Constant>(Result))
7015 return ReplaceInstUsesWith(ICI, ConstantExpr::getNot(CI));
7016 return BinaryOperator::CreateNot(Result);
7019 Instruction *InstCombiner::visitShl(BinaryOperator &I) {
7020 return commonShiftTransforms(I);
7023 Instruction *InstCombiner::visitLShr(BinaryOperator &I) {
7024 return commonShiftTransforms(I);
7027 Instruction *InstCombiner::visitAShr(BinaryOperator &I) {
7028 if (Instruction *R = commonShiftTransforms(I))
7031 Value *Op0 = I.getOperand(0);
7033 // ashr int -1, X = -1 (for any arithmetic shift rights of ~0)
7034 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
7035 if (CSI->isAllOnesValue())
7036 return ReplaceInstUsesWith(I, CSI);
7038 // See if we can turn a signed shr into an unsigned shr.
7039 if (!isa<VectorType>(I.getType())) {
7040 if (MaskedValueIsZero(Op0,
7041 APInt::getSignBit(I.getType()->getPrimitiveSizeInBits())))
7042 return BinaryOperator::CreateLShr(Op0, I.getOperand(1));
7044 // Arithmetic shifting an all-sign-bit value is a no-op.
7045 unsigned NumSignBits = ComputeNumSignBits(Op0);
7046 if (NumSignBits == Op0->getType()->getPrimitiveSizeInBits())
7047 return ReplaceInstUsesWith(I, Op0);
7053 Instruction *InstCombiner::commonShiftTransforms(BinaryOperator &I) {
7054 assert(I.getOperand(1)->getType() == I.getOperand(0)->getType());
7055 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
7057 // shl X, 0 == X and shr X, 0 == X
7058 // shl 0, X == 0 and shr 0, X == 0
7059 if (Op1 == Constant::getNullValue(Op1->getType()) ||
7060 Op0 == Constant::getNullValue(Op0->getType()))
7061 return ReplaceInstUsesWith(I, Op0);
7063 if (isa<UndefValue>(Op0)) {
7064 if (I.getOpcode() == Instruction::AShr) // undef >>s X -> undef
7065 return ReplaceInstUsesWith(I, Op0);
7066 else // undef << X -> 0, undef >>u X -> 0
7067 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
7069 if (isa<UndefValue>(Op1)) {
7070 if (I.getOpcode() == Instruction::AShr) // X >>s undef -> X
7071 return ReplaceInstUsesWith(I, Op0);
7072 else // X << undef, X >>u undef -> 0
7073 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
7076 // Try to fold constant and into select arguments.
7077 if (isa<Constant>(Op0))
7078 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
7079 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
7082 if (ConstantInt *CUI = dyn_cast<ConstantInt>(Op1))
7083 if (Instruction *Res = FoldShiftByConstant(Op0, CUI, I))
7088 Instruction *InstCombiner::FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
7089 BinaryOperator &I) {
7090 bool isLeftShift = I.getOpcode() == Instruction::Shl;
7092 // See if we can simplify any instructions used by the instruction whose sole
7093 // purpose is to compute bits we don't care about.
7094 uint32_t TypeBits = Op0->getType()->getPrimitiveSizeInBits();
7095 if (SimplifyDemandedInstructionBits(I))
7098 // shl uint X, 32 = 0 and shr ubyte Y, 9 = 0, ... just don't eliminate shr
7099 // of a signed value.
7101 if (Op1->uge(TypeBits)) {
7102 if (I.getOpcode() != Instruction::AShr)
7103 return ReplaceInstUsesWith(I, Constant::getNullValue(Op0->getType()));
7105 I.setOperand(1, ConstantInt::get(I.getType(), TypeBits-1));
7110 // ((X*C1) << C2) == (X * (C1 << C2))
7111 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0))
7112 if (BO->getOpcode() == Instruction::Mul && isLeftShift)
7113 if (Constant *BOOp = dyn_cast<Constant>(BO->getOperand(1)))
7114 return BinaryOperator::CreateMul(BO->getOperand(0),
7115 ConstantExpr::getShl(BOOp, Op1));
7117 // Try to fold constant and into select arguments.
7118 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
7119 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
7121 if (isa<PHINode>(Op0))
7122 if (Instruction *NV = FoldOpIntoPhi(I))
7125 // Fold shift2(trunc(shift1(x,c1)), c2) -> trunc(shift2(shift1(x,c1),c2))
7126 if (TruncInst *TI = dyn_cast<TruncInst>(Op0)) {
7127 Instruction *TrOp = dyn_cast<Instruction>(TI->getOperand(0));
7128 // If 'shift2' is an ashr, we would have to get the sign bit into a funny
7129 // place. Don't try to do this transformation in this case. Also, we
7130 // require that the input operand is a shift-by-constant so that we have
7131 // confidence that the shifts will get folded together. We could do this
7132 // xform in more cases, but it is unlikely to be profitable.
7133 if (TrOp && I.isLogicalShift() && TrOp->isShift() &&
7134 isa<ConstantInt>(TrOp->getOperand(1))) {
7135 // Okay, we'll do this xform. Make the shift of shift.
7136 Constant *ShAmt = ConstantExpr::getZExt(Op1, TrOp->getType());
7137 Instruction *NSh = BinaryOperator::Create(I.getOpcode(), TrOp, ShAmt,
7139 InsertNewInstBefore(NSh, I); // (shift2 (shift1 & 0x00FF), c2)
7141 // For logical shifts, the truncation has the effect of making the high
7142 // part of the register be zeros. Emulate this by inserting an AND to
7143 // clear the top bits as needed. This 'and' will usually be zapped by
7144 // other xforms later if dead.
7145 unsigned SrcSize = TrOp->getType()->getPrimitiveSizeInBits();
7146 unsigned DstSize = TI->getType()->getPrimitiveSizeInBits();
7147 APInt MaskV(APInt::getLowBitsSet(SrcSize, DstSize));
7149 // The mask we constructed says what the trunc would do if occurring
7150 // between the shifts. We want to know the effect *after* the second
7151 // shift. We know that it is a logical shift by a constant, so adjust the
7152 // mask as appropriate.
7153 if (I.getOpcode() == Instruction::Shl)
7154 MaskV <<= Op1->getZExtValue();
7156 assert(I.getOpcode() == Instruction::LShr && "Unknown logical shift");
7157 MaskV = MaskV.lshr(Op1->getZExtValue());
7160 Instruction *And = BinaryOperator::CreateAnd(NSh, ConstantInt::get(MaskV),
7162 InsertNewInstBefore(And, I); // shift1 & 0x00FF
7164 // Return the value truncated to the interesting size.
7165 return new TruncInst(And, I.getType());
7169 if (Op0->hasOneUse()) {
7170 if (BinaryOperator *Op0BO = dyn_cast<BinaryOperator>(Op0)) {
7171 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
7174 switch (Op0BO->getOpcode()) {
7176 case Instruction::Add:
7177 case Instruction::And:
7178 case Instruction::Or:
7179 case Instruction::Xor: {
7180 // These operators commute.
7181 // Turn (Y + (X >> C)) << C -> (X + (Y << C)) & (~0 << C)
7182 if (isLeftShift && Op0BO->getOperand(1)->hasOneUse() &&
7183 match(Op0BO->getOperand(1), m_Shr(m_Value(V1), m_Specific(Op1)))){
7184 Instruction *YS = BinaryOperator::CreateShl(
7185 Op0BO->getOperand(0), Op1,
7187 InsertNewInstBefore(YS, I); // (Y << C)
7189 BinaryOperator::Create(Op0BO->getOpcode(), YS, V1,
7190 Op0BO->getOperand(1)->getName());
7191 InsertNewInstBefore(X, I); // (X + (Y << C))
7192 uint32_t Op1Val = Op1->getLimitedValue(TypeBits);
7193 return BinaryOperator::CreateAnd(X, ConstantInt::get(
7194 APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
7197 // Turn (Y + ((X >> C) & CC)) << C -> ((X & (CC << C)) + (Y << C))
7198 Value *Op0BOOp1 = Op0BO->getOperand(1);
7199 if (isLeftShift && Op0BOOp1->hasOneUse() &&
7201 m_And(m_Shr(m_Value(V1), m_Specific(Op1)),
7202 m_ConstantInt(CC))) &&
7203 cast<BinaryOperator>(Op0BOOp1)->getOperand(0)->hasOneUse()) {
7204 Instruction *YS = BinaryOperator::CreateShl(
7205 Op0BO->getOperand(0), Op1,
7207 InsertNewInstBefore(YS, I); // (Y << C)
7209 BinaryOperator::CreateAnd(V1, ConstantExpr::getShl(CC, Op1),
7210 V1->getName()+".mask");
7211 InsertNewInstBefore(XM, I); // X & (CC << C)
7213 return BinaryOperator::Create(Op0BO->getOpcode(), YS, XM);
7218 case Instruction::Sub: {
7219 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
7220 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
7221 match(Op0BO->getOperand(0), m_Shr(m_Value(V1), m_Specific(Op1)))){
7222 Instruction *YS = BinaryOperator::CreateShl(
7223 Op0BO->getOperand(1), Op1,
7225 InsertNewInstBefore(YS, I); // (Y << C)
7227 BinaryOperator::Create(Op0BO->getOpcode(), V1, YS,
7228 Op0BO->getOperand(0)->getName());
7229 InsertNewInstBefore(X, I); // (X + (Y << C))
7230 uint32_t Op1Val = Op1->getLimitedValue(TypeBits);
7231 return BinaryOperator::CreateAnd(X, ConstantInt::get(
7232 APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
7235 // Turn (((X >> C)&CC) + Y) << C -> (X + (Y << C)) & (CC << C)
7236 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
7237 match(Op0BO->getOperand(0),
7238 m_And(m_Shr(m_Value(V1), m_Value(V2)),
7239 m_ConstantInt(CC))) && V2 == Op1 &&
7240 cast<BinaryOperator>(Op0BO->getOperand(0))
7241 ->getOperand(0)->hasOneUse()) {
7242 Instruction *YS = BinaryOperator::CreateShl(
7243 Op0BO->getOperand(1), Op1,
7245 InsertNewInstBefore(YS, I); // (Y << C)
7247 BinaryOperator::CreateAnd(V1, ConstantExpr::getShl(CC, Op1),
7248 V1->getName()+".mask");
7249 InsertNewInstBefore(XM, I); // X & (CC << C)
7251 return BinaryOperator::Create(Op0BO->getOpcode(), XM, YS);
7259 // If the operand is an bitwise operator with a constant RHS, and the
7260 // shift is the only use, we can pull it out of the shift.
7261 if (ConstantInt *Op0C = dyn_cast<ConstantInt>(Op0BO->getOperand(1))) {
7262 bool isValid = true; // Valid only for And, Or, Xor
7263 bool highBitSet = false; // Transform if high bit of constant set?
7265 switch (Op0BO->getOpcode()) {
7266 default: isValid = false; break; // Do not perform transform!
7267 case Instruction::Add:
7268 isValid = isLeftShift;
7270 case Instruction::Or:
7271 case Instruction::Xor:
7274 case Instruction::And:
7279 // If this is a signed shift right, and the high bit is modified
7280 // by the logical operation, do not perform the transformation.
7281 // The highBitSet boolean indicates the value of the high bit of
7282 // the constant which would cause it to be modified for this
7285 if (isValid && I.getOpcode() == Instruction::AShr)
7286 isValid = Op0C->getValue()[TypeBits-1] == highBitSet;
7289 Constant *NewRHS = ConstantExpr::get(I.getOpcode(), Op0C, Op1);
7291 Instruction *NewShift =
7292 BinaryOperator::Create(I.getOpcode(), Op0BO->getOperand(0), Op1);
7293 InsertNewInstBefore(NewShift, I);
7294 NewShift->takeName(Op0BO);
7296 return BinaryOperator::Create(Op0BO->getOpcode(), NewShift,
7303 // Find out if this is a shift of a shift by a constant.
7304 BinaryOperator *ShiftOp = dyn_cast<BinaryOperator>(Op0);
7305 if (ShiftOp && !ShiftOp->isShift())
7308 if (ShiftOp && isa<ConstantInt>(ShiftOp->getOperand(1))) {
7309 ConstantInt *ShiftAmt1C = cast<ConstantInt>(ShiftOp->getOperand(1));
7310 uint32_t ShiftAmt1 = ShiftAmt1C->getLimitedValue(TypeBits);
7311 uint32_t ShiftAmt2 = Op1->getLimitedValue(TypeBits);
7312 assert(ShiftAmt2 != 0 && "Should have been simplified earlier");
7313 if (ShiftAmt1 == 0) return 0; // Will be simplified in the future.
7314 Value *X = ShiftOp->getOperand(0);
7316 uint32_t AmtSum = ShiftAmt1+ShiftAmt2; // Fold into one big shift.
7318 const IntegerType *Ty = cast<IntegerType>(I.getType());
7320 // Check for (X << c1) << c2 and (X >> c1) >> c2
7321 if (I.getOpcode() == ShiftOp->getOpcode()) {
7322 // If this is oversized composite shift, then unsigned shifts get 0, ashr
7324 if (AmtSum >= TypeBits) {
7325 if (I.getOpcode() != Instruction::AShr)
7326 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
7327 AmtSum = TypeBits-1; // Saturate to 31 for i32 ashr.
7330 return BinaryOperator::Create(I.getOpcode(), X,
7331 ConstantInt::get(Ty, AmtSum));
7332 } else if (ShiftOp->getOpcode() == Instruction::LShr &&
7333 I.getOpcode() == Instruction::AShr) {
7334 if (AmtSum >= TypeBits)
7335 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
7337 // ((X >>u C1) >>s C2) -> (X >>u (C1+C2)) since C1 != 0.
7338 return BinaryOperator::CreateLShr(X, ConstantInt::get(Ty, AmtSum));
7339 } else if (ShiftOp->getOpcode() == Instruction::AShr &&
7340 I.getOpcode() == Instruction::LShr) {
7341 // ((X >>s C1) >>u C2) -> ((X >>s (C1+C2)) & mask) since C1 != 0.
7342 if (AmtSum >= TypeBits)
7343 AmtSum = TypeBits-1;
7345 Instruction *Shift =
7346 BinaryOperator::CreateAShr(X, ConstantInt::get(Ty, AmtSum));
7347 InsertNewInstBefore(Shift, I);
7349 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
7350 return BinaryOperator::CreateAnd(Shift, ConstantInt::get(Mask));
7353 // Okay, if we get here, one shift must be left, and the other shift must be
7354 // right. See if the amounts are equal.
7355 if (ShiftAmt1 == ShiftAmt2) {
7356 // If we have ((X >>? C) << C), turn this into X & (-1 << C).
7357 if (I.getOpcode() == Instruction::Shl) {
7358 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt1));
7359 return BinaryOperator::CreateAnd(X, ConstantInt::get(Mask));
7361 // If we have ((X << C) >>u C), turn this into X & (-1 >>u C).
7362 if (I.getOpcode() == Instruction::LShr) {
7363 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt1));
7364 return BinaryOperator::CreateAnd(X, ConstantInt::get(Mask));
7366 // We can simplify ((X << C) >>s C) into a trunc + sext.
7367 // NOTE: we could do this for any C, but that would make 'unusual' integer
7368 // types. For now, just stick to ones well-supported by the code
7370 const Type *SExtType = 0;
7371 switch (Ty->getBitWidth() - ShiftAmt1) {
7378 SExtType = IntegerType::get(Ty->getBitWidth() - ShiftAmt1);
7383 Instruction *NewTrunc = new TruncInst(X, SExtType, "sext");
7384 InsertNewInstBefore(NewTrunc, I);
7385 return new SExtInst(NewTrunc, Ty);
7387 // Otherwise, we can't handle it yet.
7388 } else if (ShiftAmt1 < ShiftAmt2) {
7389 uint32_t ShiftDiff = ShiftAmt2-ShiftAmt1;
7391 // (X >>? C1) << C2 --> X << (C2-C1) & (-1 << C2)
7392 if (I.getOpcode() == Instruction::Shl) {
7393 assert(ShiftOp->getOpcode() == Instruction::LShr ||
7394 ShiftOp->getOpcode() == Instruction::AShr);
7395 Instruction *Shift =
7396 BinaryOperator::CreateShl(X, ConstantInt::get(Ty, ShiftDiff));
7397 InsertNewInstBefore(Shift, I);
7399 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
7400 return BinaryOperator::CreateAnd(Shift, ConstantInt::get(Mask));
7403 // (X << C1) >>u C2 --> X >>u (C2-C1) & (-1 >> C2)
7404 if (I.getOpcode() == Instruction::LShr) {
7405 assert(ShiftOp->getOpcode() == Instruction::Shl);
7406 Instruction *Shift =
7407 BinaryOperator::CreateLShr(X, ConstantInt::get(Ty, ShiftDiff));
7408 InsertNewInstBefore(Shift, I);
7410 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
7411 return BinaryOperator::CreateAnd(Shift, ConstantInt::get(Mask));
7414 // We can't handle (X << C1) >>s C2, it shifts arbitrary bits in.
7416 assert(ShiftAmt2 < ShiftAmt1);
7417 uint32_t ShiftDiff = ShiftAmt1-ShiftAmt2;
7419 // (X >>? C1) << C2 --> X >>? (C1-C2) & (-1 << C2)
7420 if (I.getOpcode() == Instruction::Shl) {
7421 assert(ShiftOp->getOpcode() == Instruction::LShr ||
7422 ShiftOp->getOpcode() == Instruction::AShr);
7423 Instruction *Shift =
7424 BinaryOperator::Create(ShiftOp->getOpcode(), X,
7425 ConstantInt::get(Ty, ShiftDiff));
7426 InsertNewInstBefore(Shift, I);
7428 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
7429 return BinaryOperator::CreateAnd(Shift, ConstantInt::get(Mask));
7432 // (X << C1) >>u C2 --> X << (C1-C2) & (-1 >> C2)
7433 if (I.getOpcode() == Instruction::LShr) {
7434 assert(ShiftOp->getOpcode() == Instruction::Shl);
7435 Instruction *Shift =
7436 BinaryOperator::CreateShl(X, ConstantInt::get(Ty, ShiftDiff));
7437 InsertNewInstBefore(Shift, I);
7439 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
7440 return BinaryOperator::CreateAnd(Shift, ConstantInt::get(Mask));
7443 // We can't handle (X << C1) >>a C2, it shifts arbitrary bits in.
7450 /// DecomposeSimpleLinearExpr - Analyze 'Val', seeing if it is a simple linear
7451 /// expression. If so, decompose it, returning some value X, such that Val is
7454 static Value *DecomposeSimpleLinearExpr(Value *Val, unsigned &Scale,
7456 assert(Val->getType() == Type::Int32Ty && "Unexpected allocation size type!");
7457 if (ConstantInt *CI = dyn_cast<ConstantInt>(Val)) {
7458 Offset = CI->getZExtValue();
7460 return ConstantInt::get(Type::Int32Ty, 0);
7461 } else if (BinaryOperator *I = dyn_cast<BinaryOperator>(Val)) {
7462 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
7463 if (I->getOpcode() == Instruction::Shl) {
7464 // This is a value scaled by '1 << the shift amt'.
7465 Scale = 1U << RHS->getZExtValue();
7467 return I->getOperand(0);
7468 } else if (I->getOpcode() == Instruction::Mul) {
7469 // This value is scaled by 'RHS'.
7470 Scale = RHS->getZExtValue();
7472 return I->getOperand(0);
7473 } else if (I->getOpcode() == Instruction::Add) {
7474 // We have X+C. Check to see if we really have (X*C2)+C1,
7475 // where C1 is divisible by C2.
7478 DecomposeSimpleLinearExpr(I->getOperand(0), SubScale, Offset);
7479 Offset += RHS->getZExtValue();
7486 // Otherwise, we can't look past this.
7493 /// PromoteCastOfAllocation - If we find a cast of an allocation instruction,
7494 /// try to eliminate the cast by moving the type information into the alloc.
7495 Instruction *InstCombiner::PromoteCastOfAllocation(BitCastInst &CI,
7496 AllocationInst &AI) {
7497 const PointerType *PTy = cast<PointerType>(CI.getType());
7499 // Remove any uses of AI that are dead.
7500 assert(!CI.use_empty() && "Dead instructions should be removed earlier!");
7502 for (Value::use_iterator UI = AI.use_begin(), E = AI.use_end(); UI != E; ) {
7503 Instruction *User = cast<Instruction>(*UI++);
7504 if (isInstructionTriviallyDead(User)) {
7505 while (UI != E && *UI == User)
7506 ++UI; // If this instruction uses AI more than once, don't break UI.
7509 DOUT << "IC: DCE: " << *User;
7510 EraseInstFromFunction(*User);
7514 // Get the type really allocated and the type casted to.
7515 const Type *AllocElTy = AI.getAllocatedType();
7516 const Type *CastElTy = PTy->getElementType();
7517 if (!AllocElTy->isSized() || !CastElTy->isSized()) return 0;
7519 unsigned AllocElTyAlign = TD->getABITypeAlignment(AllocElTy);
7520 unsigned CastElTyAlign = TD->getABITypeAlignment(CastElTy);
7521 if (CastElTyAlign < AllocElTyAlign) return 0;
7523 // If the allocation has multiple uses, only promote it if we are strictly
7524 // increasing the alignment of the resultant allocation. If we keep it the
7525 // same, we open the door to infinite loops of various kinds. (A reference
7526 // from a dbg.declare doesn't count as a use for this purpose.)
7527 if (!AI.hasOneUse() && !hasOneUsePlusDeclare(&AI) &&
7528 CastElTyAlign == AllocElTyAlign) return 0;
7530 uint64_t AllocElTySize = TD->getTypePaddedSize(AllocElTy);
7531 uint64_t CastElTySize = TD->getTypePaddedSize(CastElTy);
7532 if (CastElTySize == 0 || AllocElTySize == 0) return 0;
7534 // See if we can satisfy the modulus by pulling a scale out of the array
7536 unsigned ArraySizeScale;
7538 Value *NumElements = // See if the array size is a decomposable linear expr.
7539 DecomposeSimpleLinearExpr(AI.getOperand(0), ArraySizeScale, ArrayOffset);
7541 // If we can now satisfy the modulus, by using a non-1 scale, we really can
7543 if ((AllocElTySize*ArraySizeScale) % CastElTySize != 0 ||
7544 (AllocElTySize*ArrayOffset ) % CastElTySize != 0) return 0;
7546 unsigned Scale = (AllocElTySize*ArraySizeScale)/CastElTySize;
7551 // If the allocation size is constant, form a constant mul expression
7552 Amt = ConstantInt::get(Type::Int32Ty, Scale);
7553 if (isa<ConstantInt>(NumElements))
7554 Amt = Multiply(cast<ConstantInt>(NumElements), cast<ConstantInt>(Amt));
7555 // otherwise multiply the amount and the number of elements
7557 Instruction *Tmp = BinaryOperator::CreateMul(Amt, NumElements, "tmp");
7558 Amt = InsertNewInstBefore(Tmp, AI);
7562 if (int Offset = (AllocElTySize*ArrayOffset)/CastElTySize) {
7563 Value *Off = ConstantInt::get(Type::Int32Ty, Offset, true);
7564 Instruction *Tmp = BinaryOperator::CreateAdd(Amt, Off, "tmp");
7565 Amt = InsertNewInstBefore(Tmp, AI);
7568 AllocationInst *New;
7569 if (isa<MallocInst>(AI))
7570 New = new MallocInst(CastElTy, Amt, AI.getAlignment());
7572 New = new AllocaInst(CastElTy, Amt, AI.getAlignment());
7573 InsertNewInstBefore(New, AI);
7576 // If the allocation has one real use plus a dbg.declare, just remove the
7578 if (DbgDeclareInst *DI = hasOneUsePlusDeclare(&AI)) {
7579 EraseInstFromFunction(*DI);
7581 // If the allocation has multiple real uses, insert a cast and change all
7582 // things that used it to use the new cast. This will also hack on CI, but it
7584 else if (!AI.hasOneUse()) {
7585 AddUsesToWorkList(AI);
7586 // New is the allocation instruction, pointer typed. AI is the original
7587 // allocation instruction, also pointer typed. Thus, cast to use is BitCast.
7588 CastInst *NewCast = new BitCastInst(New, AI.getType(), "tmpcast");
7589 InsertNewInstBefore(NewCast, AI);
7590 AI.replaceAllUsesWith(NewCast);
7592 return ReplaceInstUsesWith(CI, New);
7595 /// CanEvaluateInDifferentType - Return true if we can take the specified value
7596 /// and return it as type Ty without inserting any new casts and without
7597 /// changing the computed value. This is used by code that tries to decide
7598 /// whether promoting or shrinking integer operations to wider or smaller types
7599 /// will allow us to eliminate a truncate or extend.
7601 /// This is a truncation operation if Ty is smaller than V->getType(), or an
7602 /// extension operation if Ty is larger.
7604 /// If CastOpc is a truncation, then Ty will be a type smaller than V. We
7605 /// should return true if trunc(V) can be computed by computing V in the smaller
7606 /// type. If V is an instruction, then trunc(inst(x,y)) can be computed as
7607 /// inst(trunc(x),trunc(y)), which only makes sense if x and y can be
7608 /// efficiently truncated.
7610 /// If CastOpc is a sext or zext, we are asking if the low bits of the value can
7611 /// bit computed in a larger type, which is then and'd or sext_in_reg'd to get
7612 /// the final result.
7613 bool InstCombiner::CanEvaluateInDifferentType(Value *V, const IntegerType *Ty,
7615 int &NumCastsRemoved){
7616 // We can always evaluate constants in another type.
7617 if (isa<ConstantInt>(V))
7620 Instruction *I = dyn_cast<Instruction>(V);
7621 if (!I) return false;
7623 const IntegerType *OrigTy = cast<IntegerType>(V->getType());
7625 // If this is an extension or truncate, we can often eliminate it.
7626 if (isa<TruncInst>(I) || isa<ZExtInst>(I) || isa<SExtInst>(I)) {
7627 // If this is a cast from the destination type, we can trivially eliminate
7628 // it, and this will remove a cast overall.
7629 if (I->getOperand(0)->getType() == Ty) {
7630 // If the first operand is itself a cast, and is eliminable, do not count
7631 // this as an eliminable cast. We would prefer to eliminate those two
7633 if (!isa<CastInst>(I->getOperand(0)) && I->hasOneUse())
7639 // We can't extend or shrink something that has multiple uses: doing so would
7640 // require duplicating the instruction in general, which isn't profitable.
7641 if (!I->hasOneUse()) return false;
7643 unsigned Opc = I->getOpcode();
7645 case Instruction::Add:
7646 case Instruction::Sub:
7647 case Instruction::Mul:
7648 case Instruction::And:
7649 case Instruction::Or:
7650 case Instruction::Xor:
7651 // These operators can all arbitrarily be extended or truncated.
7652 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
7654 CanEvaluateInDifferentType(I->getOperand(1), Ty, CastOpc,
7657 case Instruction::Shl:
7658 // If we are truncating the result of this SHL, and if it's a shift of a
7659 // constant amount, we can always perform a SHL in a smaller type.
7660 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
7661 uint32_t BitWidth = Ty->getBitWidth();
7662 if (BitWidth < OrigTy->getBitWidth() &&
7663 CI->getLimitedValue(BitWidth) < BitWidth)
7664 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
7668 case Instruction::LShr:
7669 // If this is a truncate of a logical shr, we can truncate it to a smaller
7670 // lshr iff we know that the bits we would otherwise be shifting in are
7672 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
7673 uint32_t OrigBitWidth = OrigTy->getBitWidth();
7674 uint32_t BitWidth = Ty->getBitWidth();
7675 if (BitWidth < OrigBitWidth &&
7676 MaskedValueIsZero(I->getOperand(0),
7677 APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth)) &&
7678 CI->getLimitedValue(BitWidth) < BitWidth) {
7679 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
7684 case Instruction::ZExt:
7685 case Instruction::SExt:
7686 case Instruction::Trunc:
7687 // If this is the same kind of case as our original (e.g. zext+zext), we
7688 // can safely replace it. Note that replacing it does not reduce the number
7689 // of casts in the input.
7693 // sext (zext ty1), ty2 -> zext ty2
7694 if (CastOpc == Instruction::SExt && Opc == Instruction::ZExt)
7697 case Instruction::Select: {
7698 SelectInst *SI = cast<SelectInst>(I);
7699 return CanEvaluateInDifferentType(SI->getTrueValue(), Ty, CastOpc,
7701 CanEvaluateInDifferentType(SI->getFalseValue(), Ty, CastOpc,
7704 case Instruction::PHI: {
7705 // We can change a phi if we can change all operands.
7706 PHINode *PN = cast<PHINode>(I);
7707 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
7708 if (!CanEvaluateInDifferentType(PN->getIncomingValue(i), Ty, CastOpc,
7714 // TODO: Can handle more cases here.
7721 /// EvaluateInDifferentType - Given an expression that
7722 /// CanEvaluateInDifferentType returns true for, actually insert the code to
7723 /// evaluate the expression.
7724 Value *InstCombiner::EvaluateInDifferentType(Value *V, const Type *Ty,
7726 if (Constant *C = dyn_cast<Constant>(V))
7727 return ConstantExpr::getIntegerCast(C, Ty, isSigned /*Sext or ZExt*/);
7729 // Otherwise, it must be an instruction.
7730 Instruction *I = cast<Instruction>(V);
7731 Instruction *Res = 0;
7732 unsigned Opc = I->getOpcode();
7734 case Instruction::Add:
7735 case Instruction::Sub:
7736 case Instruction::Mul:
7737 case Instruction::And:
7738 case Instruction::Or:
7739 case Instruction::Xor:
7740 case Instruction::AShr:
7741 case Instruction::LShr:
7742 case Instruction::Shl: {
7743 Value *LHS = EvaluateInDifferentType(I->getOperand(0), Ty, isSigned);
7744 Value *RHS = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
7745 Res = BinaryOperator::Create((Instruction::BinaryOps)Opc, LHS, RHS);
7748 case Instruction::Trunc:
7749 case Instruction::ZExt:
7750 case Instruction::SExt:
7751 // If the source type of the cast is the type we're trying for then we can
7752 // just return the source. There's no need to insert it because it is not
7754 if (I->getOperand(0)->getType() == Ty)
7755 return I->getOperand(0);
7757 // Otherwise, must be the same type of cast, so just reinsert a new one.
7758 Res = CastInst::Create(cast<CastInst>(I)->getOpcode(), I->getOperand(0),
7761 case Instruction::Select: {
7762 Value *True = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
7763 Value *False = EvaluateInDifferentType(I->getOperand(2), Ty, isSigned);
7764 Res = SelectInst::Create(I->getOperand(0), True, False);
7767 case Instruction::PHI: {
7768 PHINode *OPN = cast<PHINode>(I);
7769 PHINode *NPN = PHINode::Create(Ty);
7770 for (unsigned i = 0, e = OPN->getNumIncomingValues(); i != e; ++i) {
7771 Value *V =EvaluateInDifferentType(OPN->getIncomingValue(i), Ty, isSigned);
7772 NPN->addIncoming(V, OPN->getIncomingBlock(i));
7778 // TODO: Can handle more cases here.
7779 assert(0 && "Unreachable!");
7784 return InsertNewInstBefore(Res, *I);
7787 /// @brief Implement the transforms common to all CastInst visitors.
7788 Instruction *InstCombiner::commonCastTransforms(CastInst &CI) {
7789 Value *Src = CI.getOperand(0);
7791 // Many cases of "cast of a cast" are eliminable. If it's eliminable we just
7792 // eliminate it now.
7793 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
7794 if (Instruction::CastOps opc =
7795 isEliminableCastPair(CSrc, CI.getOpcode(), CI.getType(), TD)) {
7796 // The first cast (CSrc) is eliminable so we need to fix up or replace
7797 // the second cast (CI). CSrc will then have a good chance of being dead.
7798 return CastInst::Create(opc, CSrc->getOperand(0), CI.getType());
7802 // If we are casting a select then fold the cast into the select
7803 if (SelectInst *SI = dyn_cast<SelectInst>(Src))
7804 if (Instruction *NV = FoldOpIntoSelect(CI, SI, this))
7807 // If we are casting a PHI then fold the cast into the PHI
7808 if (isa<PHINode>(Src))
7809 if (Instruction *NV = FoldOpIntoPhi(CI))
7815 /// FindElementAtOffset - Given a type and a constant offset, determine whether
7816 /// or not there is a sequence of GEP indices into the type that will land us at
7817 /// the specified offset. If so, fill them into NewIndices and return the
7818 /// resultant element type, otherwise return null.
7819 static const Type *FindElementAtOffset(const Type *Ty, int64_t Offset,
7820 SmallVectorImpl<Value*> &NewIndices,
7821 const TargetData *TD) {
7822 if (!Ty->isSized()) return 0;
7824 // Start with the index over the outer type. Note that the type size
7825 // might be zero (even if the offset isn't zero) if the indexed type
7826 // is something like [0 x {int, int}]
7827 const Type *IntPtrTy = TD->getIntPtrType();
7828 int64_t FirstIdx = 0;
7829 if (int64_t TySize = TD->getTypePaddedSize(Ty)) {
7830 FirstIdx = Offset/TySize;
7831 Offset -= FirstIdx*TySize;
7833 // Handle hosts where % returns negative instead of values [0..TySize).
7837 assert(Offset >= 0);
7839 assert((uint64_t)Offset < (uint64_t)TySize && "Out of range offset");
7842 NewIndices.push_back(ConstantInt::get(IntPtrTy, FirstIdx));
7844 // Index into the types. If we fail, set OrigBase to null.
7846 // Indexing into tail padding between struct/array elements.
7847 if (uint64_t(Offset*8) >= TD->getTypeSizeInBits(Ty))
7850 if (const StructType *STy = dyn_cast<StructType>(Ty)) {
7851 const StructLayout *SL = TD->getStructLayout(STy);
7852 assert(Offset < (int64_t)SL->getSizeInBytes() &&
7853 "Offset must stay within the indexed type");
7855 unsigned Elt = SL->getElementContainingOffset(Offset);
7856 NewIndices.push_back(ConstantInt::get(Type::Int32Ty, Elt));
7858 Offset -= SL->getElementOffset(Elt);
7859 Ty = STy->getElementType(Elt);
7860 } else if (const ArrayType *AT = dyn_cast<ArrayType>(Ty)) {
7861 uint64_t EltSize = TD->getTypePaddedSize(AT->getElementType());
7862 assert(EltSize && "Cannot index into a zero-sized array");
7863 NewIndices.push_back(ConstantInt::get(IntPtrTy,Offset/EltSize));
7865 Ty = AT->getElementType();
7867 // Otherwise, we can't index into the middle of this atomic type, bail.
7875 /// @brief Implement the transforms for cast of pointer (bitcast/ptrtoint)
7876 Instruction *InstCombiner::commonPointerCastTransforms(CastInst &CI) {
7877 Value *Src = CI.getOperand(0);
7879 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Src)) {
7880 // If casting the result of a getelementptr instruction with no offset, turn
7881 // this into a cast of the original pointer!
7882 if (GEP->hasAllZeroIndices()) {
7883 // Changing the cast operand is usually not a good idea but it is safe
7884 // here because the pointer operand is being replaced with another
7885 // pointer operand so the opcode doesn't need to change.
7887 CI.setOperand(0, GEP->getOperand(0));
7891 // If the GEP has a single use, and the base pointer is a bitcast, and the
7892 // GEP computes a constant offset, see if we can convert these three
7893 // instructions into fewer. This typically happens with unions and other
7894 // non-type-safe code.
7895 if (GEP->hasOneUse() && isa<BitCastInst>(GEP->getOperand(0))) {
7896 if (GEP->hasAllConstantIndices()) {
7897 // We are guaranteed to get a constant from EmitGEPOffset.
7898 ConstantInt *OffsetV = cast<ConstantInt>(EmitGEPOffset(GEP, CI, *this));
7899 int64_t Offset = OffsetV->getSExtValue();
7901 // Get the base pointer input of the bitcast, and the type it points to.
7902 Value *OrigBase = cast<BitCastInst>(GEP->getOperand(0))->getOperand(0);
7903 const Type *GEPIdxTy =
7904 cast<PointerType>(OrigBase->getType())->getElementType();
7905 SmallVector<Value*, 8> NewIndices;
7906 if (FindElementAtOffset(GEPIdxTy, Offset, NewIndices, TD)) {
7907 // If we were able to index down into an element, create the GEP
7908 // and bitcast the result. This eliminates one bitcast, potentially
7910 Instruction *NGEP = GetElementPtrInst::Create(OrigBase,
7912 NewIndices.end(), "");
7913 InsertNewInstBefore(NGEP, CI);
7914 NGEP->takeName(GEP);
7916 if (isa<BitCastInst>(CI))
7917 return new BitCastInst(NGEP, CI.getType());
7918 assert(isa<PtrToIntInst>(CI));
7919 return new PtrToIntInst(NGEP, CI.getType());
7925 return commonCastTransforms(CI);
7929 /// Only the TRUNC, ZEXT, SEXT, and BITCAST can both operand and result as
7930 /// integer types. This function implements the common transforms for all those
7932 /// @brief Implement the transforms common to CastInst with integer operands
7933 Instruction *InstCombiner::commonIntCastTransforms(CastInst &CI) {
7934 if (Instruction *Result = commonCastTransforms(CI))
7937 Value *Src = CI.getOperand(0);
7938 const Type *SrcTy = Src->getType();
7939 const Type *DestTy = CI.getType();
7940 uint32_t SrcBitSize = SrcTy->getPrimitiveSizeInBits();
7941 uint32_t DestBitSize = DestTy->getPrimitiveSizeInBits();
7943 // See if we can simplify any instructions used by the LHS whose sole
7944 // purpose is to compute bits we don't care about.
7945 if (SimplifyDemandedInstructionBits(CI))
7948 // If the source isn't an instruction or has more than one use then we
7949 // can't do anything more.
7950 Instruction *SrcI = dyn_cast<Instruction>(Src);
7951 if (!SrcI || !Src->hasOneUse())
7954 // Attempt to propagate the cast into the instruction for int->int casts.
7955 int NumCastsRemoved = 0;
7956 if (!isa<BitCastInst>(CI) &&
7957 CanEvaluateInDifferentType(SrcI, cast<IntegerType>(DestTy),
7958 CI.getOpcode(), NumCastsRemoved)) {
7959 // If this cast is a truncate, evaluting in a different type always
7960 // eliminates the cast, so it is always a win. If this is a zero-extension,
7961 // we need to do an AND to maintain the clear top-part of the computation,
7962 // so we require that the input have eliminated at least one cast. If this
7963 // is a sign extension, we insert two new casts (to do the extension) so we
7964 // require that two casts have been eliminated.
7965 bool DoXForm = false;
7966 bool JustReplace = false;
7967 switch (CI.getOpcode()) {
7969 // All the others use floating point so we shouldn't actually
7970 // get here because of the check above.
7971 assert(0 && "Unknown cast type");
7972 case Instruction::Trunc:
7975 case Instruction::ZExt: {
7976 DoXForm = NumCastsRemoved >= 1;
7977 if (!DoXForm && 0) {
7978 // If it's unnecessary to issue an AND to clear the high bits, it's
7979 // always profitable to do this xform.
7980 Value *TryRes = EvaluateInDifferentType(SrcI, DestTy, false);
7981 APInt Mask(APInt::getBitsSet(DestBitSize, SrcBitSize, DestBitSize));
7982 if (MaskedValueIsZero(TryRes, Mask))
7983 return ReplaceInstUsesWith(CI, TryRes);
7985 if (Instruction *TryI = dyn_cast<Instruction>(TryRes))
7986 if (TryI->use_empty())
7987 EraseInstFromFunction(*TryI);
7991 case Instruction::SExt: {
7992 DoXForm = NumCastsRemoved >= 2;
7993 if (!DoXForm && !isa<TruncInst>(SrcI) && 0) {
7994 // If we do not have to emit the truncate + sext pair, then it's always
7995 // profitable to do this xform.
7997 // It's not safe to eliminate the trunc + sext pair if one of the
7998 // eliminated cast is a truncate. e.g.
7999 // t2 = trunc i32 t1 to i16
8000 // t3 = sext i16 t2 to i32
8003 Value *TryRes = EvaluateInDifferentType(SrcI, DestTy, true);
8004 unsigned NumSignBits = ComputeNumSignBits(TryRes);
8005 if (NumSignBits > (DestBitSize - SrcBitSize))
8006 return ReplaceInstUsesWith(CI, TryRes);
8008 if (Instruction *TryI = dyn_cast<Instruction>(TryRes))
8009 if (TryI->use_empty())
8010 EraseInstFromFunction(*TryI);
8017 DOUT << "ICE: EvaluateInDifferentType converting expression type to avoid"
8019 Value *Res = EvaluateInDifferentType(SrcI, DestTy,
8020 CI.getOpcode() == Instruction::SExt);
8022 // Just replace this cast with the result.
8023 return ReplaceInstUsesWith(CI, Res);
8025 assert(Res->getType() == DestTy);
8026 switch (CI.getOpcode()) {
8027 default: assert(0 && "Unknown cast type!");
8028 case Instruction::Trunc:
8029 case Instruction::BitCast:
8030 // Just replace this cast with the result.
8031 return ReplaceInstUsesWith(CI, Res);
8032 case Instruction::ZExt: {
8033 assert(SrcBitSize < DestBitSize && "Not a zext?");
8035 // If the high bits are already zero, just replace this cast with the
8037 APInt Mask(APInt::getBitsSet(DestBitSize, SrcBitSize, DestBitSize));
8038 if (MaskedValueIsZero(Res, Mask))
8039 return ReplaceInstUsesWith(CI, Res);
8041 // We need to emit an AND to clear the high bits.
8042 Constant *C = ConstantInt::get(APInt::getLowBitsSet(DestBitSize,
8044 return BinaryOperator::CreateAnd(Res, C);
8046 case Instruction::SExt: {
8047 // If the high bits are already filled with sign bit, just replace this
8048 // cast with the result.
8049 unsigned NumSignBits = ComputeNumSignBits(Res);
8050 if (NumSignBits > (DestBitSize - SrcBitSize))
8051 return ReplaceInstUsesWith(CI, Res);
8053 // We need to emit a cast to truncate, then a cast to sext.
8054 return CastInst::Create(Instruction::SExt,
8055 InsertCastBefore(Instruction::Trunc, Res, Src->getType(),
8062 Value *Op0 = SrcI->getNumOperands() > 0 ? SrcI->getOperand(0) : 0;
8063 Value *Op1 = SrcI->getNumOperands() > 1 ? SrcI->getOperand(1) : 0;
8065 switch (SrcI->getOpcode()) {
8066 case Instruction::Add:
8067 case Instruction::Mul:
8068 case Instruction::And:
8069 case Instruction::Or:
8070 case Instruction::Xor:
8071 // If we are discarding information, rewrite.
8072 if (DestBitSize <= SrcBitSize && DestBitSize != 1) {
8073 // Don't insert two casts if they cannot be eliminated. We allow
8074 // two casts to be inserted if the sizes are the same. This could
8075 // only be converting signedness, which is a noop.
8076 if (DestBitSize == SrcBitSize ||
8077 !ValueRequiresCast(CI.getOpcode(), Op1, DestTy,TD) ||
8078 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
8079 Instruction::CastOps opcode = CI.getOpcode();
8080 Value *Op0c = InsertCastBefore(opcode, Op0, DestTy, *SrcI);
8081 Value *Op1c = InsertCastBefore(opcode, Op1, DestTy, *SrcI);
8082 return BinaryOperator::Create(
8083 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
8087 // cast (xor bool X, true) to int --> xor (cast bool X to int), 1
8088 if (isa<ZExtInst>(CI) && SrcBitSize == 1 &&
8089 SrcI->getOpcode() == Instruction::Xor &&
8090 Op1 == ConstantInt::getTrue() &&
8091 (!Op0->hasOneUse() || !isa<CmpInst>(Op0))) {
8092 Value *New = InsertCastBefore(Instruction::ZExt, Op0, DestTy, CI);
8093 return BinaryOperator::CreateXor(New, ConstantInt::get(CI.getType(), 1));
8096 case Instruction::SDiv:
8097 case Instruction::UDiv:
8098 case Instruction::SRem:
8099 case Instruction::URem:
8100 // If we are just changing the sign, rewrite.
8101 if (DestBitSize == SrcBitSize) {
8102 // Don't insert two casts if they cannot be eliminated. We allow
8103 // two casts to be inserted if the sizes are the same. This could
8104 // only be converting signedness, which is a noop.
8105 if (!ValueRequiresCast(CI.getOpcode(), Op1, DestTy, TD) ||
8106 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
8107 Value *Op0c = InsertCastBefore(Instruction::BitCast,
8108 Op0, DestTy, *SrcI);
8109 Value *Op1c = InsertCastBefore(Instruction::BitCast,
8110 Op1, DestTy, *SrcI);
8111 return BinaryOperator::Create(
8112 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
8117 case Instruction::Shl:
8118 // Allow changing the sign of the source operand. Do not allow
8119 // changing the size of the shift, UNLESS the shift amount is a
8120 // constant. We must not change variable sized shifts to a smaller
8121 // size, because it is undefined to shift more bits out than exist
8123 if (DestBitSize == SrcBitSize ||
8124 (DestBitSize < SrcBitSize && isa<Constant>(Op1))) {
8125 Instruction::CastOps opcode = (DestBitSize == SrcBitSize ?
8126 Instruction::BitCast : Instruction::Trunc);
8127 Value *Op0c = InsertCastBefore(opcode, Op0, DestTy, *SrcI);
8128 Value *Op1c = InsertCastBefore(opcode, Op1, DestTy, *SrcI);
8129 return BinaryOperator::CreateShl(Op0c, Op1c);
8132 case Instruction::AShr:
8133 // If this is a signed shr, and if all bits shifted in are about to be
8134 // truncated off, turn it into an unsigned shr to allow greater
8136 if (DestBitSize < SrcBitSize &&
8137 isa<ConstantInt>(Op1)) {
8138 uint32_t ShiftAmt = cast<ConstantInt>(Op1)->getLimitedValue(SrcBitSize);
8139 if (SrcBitSize > ShiftAmt && SrcBitSize-ShiftAmt >= DestBitSize) {
8140 // Insert the new logical shift right.
8141 return BinaryOperator::CreateLShr(Op0, Op1);
8149 Instruction *InstCombiner::visitTrunc(TruncInst &CI) {
8150 if (Instruction *Result = commonIntCastTransforms(CI))
8153 Value *Src = CI.getOperand(0);
8154 const Type *Ty = CI.getType();
8155 uint32_t DestBitWidth = Ty->getPrimitiveSizeInBits();
8156 uint32_t SrcBitWidth = cast<IntegerType>(Src->getType())->getBitWidth();
8158 // Canonicalize trunc x to i1 -> (icmp ne (and x, 1), 0)
8159 if (DestBitWidth == 1) {
8160 Constant *One = ConstantInt::get(Src->getType(), 1);
8161 Src = InsertNewInstBefore(BinaryOperator::CreateAnd(Src, One, "tmp"), CI);
8162 Value *Zero = Constant::getNullValue(Src->getType());
8163 return new ICmpInst(ICmpInst::ICMP_NE, Src, Zero);
8166 // Optimize trunc(lshr(), c) to pull the shift through the truncate.
8167 ConstantInt *ShAmtV = 0;
8169 if (Src->hasOneUse() &&
8170 match(Src, m_LShr(m_Value(ShiftOp), m_ConstantInt(ShAmtV)))) {
8171 uint32_t ShAmt = ShAmtV->getLimitedValue(SrcBitWidth);
8173 // Get a mask for the bits shifting in.
8174 APInt Mask(APInt::getLowBitsSet(SrcBitWidth, ShAmt).shl(DestBitWidth));
8175 if (MaskedValueIsZero(ShiftOp, Mask)) {
8176 if (ShAmt >= DestBitWidth) // All zeros.
8177 return ReplaceInstUsesWith(CI, Constant::getNullValue(Ty));
8179 // Okay, we can shrink this. Truncate the input, then return a new
8181 Value *V1 = InsertCastBefore(Instruction::Trunc, ShiftOp, Ty, CI);
8182 Value *V2 = ConstantExpr::getTrunc(ShAmtV, Ty);
8183 return BinaryOperator::CreateLShr(V1, V2);
8190 /// transformZExtICmp - Transform (zext icmp) to bitwise / integer operations
8191 /// in order to eliminate the icmp.
8192 Instruction *InstCombiner::transformZExtICmp(ICmpInst *ICI, Instruction &CI,
8194 // If we are just checking for a icmp eq of a single bit and zext'ing it
8195 // to an integer, then shift the bit to the appropriate place and then
8196 // cast to integer to avoid the comparison.
8197 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(ICI->getOperand(1))) {
8198 const APInt &Op1CV = Op1C->getValue();
8200 // zext (x <s 0) to i32 --> x>>u31 true if signbit set.
8201 // zext (x >s -1) to i32 --> (x>>u31)^1 true if signbit clear.
8202 if ((ICI->getPredicate() == ICmpInst::ICMP_SLT && Op1CV == 0) ||
8203 (ICI->getPredicate() == ICmpInst::ICMP_SGT &&Op1CV.isAllOnesValue())) {
8204 if (!DoXform) return ICI;
8206 Value *In = ICI->getOperand(0);
8207 Value *Sh = ConstantInt::get(In->getType(),
8208 In->getType()->getPrimitiveSizeInBits()-1);
8209 In = InsertNewInstBefore(BinaryOperator::CreateLShr(In, Sh,
8210 In->getName()+".lobit"),
8212 if (In->getType() != CI.getType())
8213 In = CastInst::CreateIntegerCast(In, CI.getType(),
8214 false/*ZExt*/, "tmp", &CI);
8216 if (ICI->getPredicate() == ICmpInst::ICMP_SGT) {
8217 Constant *One = ConstantInt::get(In->getType(), 1);
8218 In = InsertNewInstBefore(BinaryOperator::CreateXor(In, One,
8219 In->getName()+".not"),
8223 return ReplaceInstUsesWith(CI, In);
8228 // zext (X == 0) to i32 --> X^1 iff X has only the low bit set.
8229 // zext (X == 0) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
8230 // zext (X == 1) to i32 --> X iff X has only the low bit set.
8231 // zext (X == 2) to i32 --> X>>1 iff X has only the 2nd bit set.
8232 // zext (X != 0) to i32 --> X iff X has only the low bit set.
8233 // zext (X != 0) to i32 --> X>>1 iff X has only the 2nd bit set.
8234 // zext (X != 1) to i32 --> X^1 iff X has only the low bit set.
8235 // zext (X != 2) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
8236 if ((Op1CV == 0 || Op1CV.isPowerOf2()) &&
8237 // This only works for EQ and NE
8238 ICI->isEquality()) {
8239 // If Op1C some other power of two, convert:
8240 uint32_t BitWidth = Op1C->getType()->getBitWidth();
8241 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
8242 APInt TypeMask(APInt::getAllOnesValue(BitWidth));
8243 ComputeMaskedBits(ICI->getOperand(0), TypeMask, KnownZero, KnownOne);
8245 APInt KnownZeroMask(~KnownZero);
8246 if (KnownZeroMask.isPowerOf2()) { // Exactly 1 possible 1?
8247 if (!DoXform) return ICI;
8249 bool isNE = ICI->getPredicate() == ICmpInst::ICMP_NE;
8250 if (Op1CV != 0 && (Op1CV != KnownZeroMask)) {
8251 // (X&4) == 2 --> false
8252 // (X&4) != 2 --> true
8253 Constant *Res = ConstantInt::get(Type::Int1Ty, isNE);
8254 Res = ConstantExpr::getZExt(Res, CI.getType());
8255 return ReplaceInstUsesWith(CI, Res);
8258 uint32_t ShiftAmt = KnownZeroMask.logBase2();
8259 Value *In = ICI->getOperand(0);
8261 // Perform a logical shr by shiftamt.
8262 // Insert the shift to put the result in the low bit.
8263 In = InsertNewInstBefore(BinaryOperator::CreateLShr(In,
8264 ConstantInt::get(In->getType(), ShiftAmt),
8265 In->getName()+".lobit"), CI);
8268 if ((Op1CV != 0) == isNE) { // Toggle the low bit.
8269 Constant *One = ConstantInt::get(In->getType(), 1);
8270 In = BinaryOperator::CreateXor(In, One, "tmp");
8271 InsertNewInstBefore(cast<Instruction>(In), CI);
8274 if (CI.getType() == In->getType())
8275 return ReplaceInstUsesWith(CI, In);
8277 return CastInst::CreateIntegerCast(In, CI.getType(), false/*ZExt*/);
8285 Instruction *InstCombiner::visitZExt(ZExtInst &CI) {
8286 // If one of the common conversion will work ..
8287 if (Instruction *Result = commonIntCastTransforms(CI))
8290 Value *Src = CI.getOperand(0);
8292 // If this is a TRUNC followed by a ZEXT then we are dealing with integral
8293 // types and if the sizes are just right we can convert this into a logical
8294 // 'and' which will be much cheaper than the pair of casts.
8295 if (TruncInst *CSrc = dyn_cast<TruncInst>(Src)) { // A->B->C cast
8296 // Get the sizes of the types involved. We know that the intermediate type
8297 // will be smaller than A or C, but don't know the relation between A and C.
8298 Value *A = CSrc->getOperand(0);
8299 unsigned SrcSize = A->getType()->getPrimitiveSizeInBits();
8300 unsigned MidSize = CSrc->getType()->getPrimitiveSizeInBits();
8301 unsigned DstSize = CI.getType()->getPrimitiveSizeInBits();
8302 // If we're actually extending zero bits, then if
8303 // SrcSize < DstSize: zext(a & mask)
8304 // SrcSize == DstSize: a & mask
8305 // SrcSize > DstSize: trunc(a) & mask
8306 if (SrcSize < DstSize) {
8307 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
8308 Constant *AndConst = ConstantInt::get(AndValue);
8310 BinaryOperator::CreateAnd(A, AndConst, CSrc->getName()+".mask");
8311 InsertNewInstBefore(And, CI);
8312 return new ZExtInst(And, CI.getType());
8313 } else if (SrcSize == DstSize) {
8314 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
8315 return BinaryOperator::CreateAnd(A, ConstantInt::get(AndValue));
8316 } else if (SrcSize > DstSize) {
8317 Instruction *Trunc = new TruncInst(A, CI.getType(), "tmp");
8318 InsertNewInstBefore(Trunc, CI);
8319 APInt AndValue(APInt::getLowBitsSet(DstSize, MidSize));
8320 return BinaryOperator::CreateAnd(Trunc, ConstantInt::get(AndValue));
8324 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src))
8325 return transformZExtICmp(ICI, CI);
8327 BinaryOperator *SrcI = dyn_cast<BinaryOperator>(Src);
8328 if (SrcI && SrcI->getOpcode() == Instruction::Or) {
8329 // zext (or icmp, icmp) --> or (zext icmp), (zext icmp) if at least one
8330 // of the (zext icmp) will be transformed.
8331 ICmpInst *LHS = dyn_cast<ICmpInst>(SrcI->getOperand(0));
8332 ICmpInst *RHS = dyn_cast<ICmpInst>(SrcI->getOperand(1));
8333 if (LHS && RHS && LHS->hasOneUse() && RHS->hasOneUse() &&
8334 (transformZExtICmp(LHS, CI, false) ||
8335 transformZExtICmp(RHS, CI, false))) {
8336 Value *LCast = InsertCastBefore(Instruction::ZExt, LHS, CI.getType(), CI);
8337 Value *RCast = InsertCastBefore(Instruction::ZExt, RHS, CI.getType(), CI);
8338 return BinaryOperator::Create(Instruction::Or, LCast, RCast);
8345 Instruction *InstCombiner::visitSExt(SExtInst &CI) {
8346 if (Instruction *I = commonIntCastTransforms(CI))
8349 Value *Src = CI.getOperand(0);
8351 // Canonicalize sign-extend from i1 to a select.
8352 if (Src->getType() == Type::Int1Ty)
8353 return SelectInst::Create(Src,
8354 ConstantInt::getAllOnesValue(CI.getType()),
8355 Constant::getNullValue(CI.getType()));
8357 // See if the value being truncated is already sign extended. If so, just
8358 // eliminate the trunc/sext pair.
8359 if (getOpcode(Src) == Instruction::Trunc) {
8360 Value *Op = cast<User>(Src)->getOperand(0);
8361 unsigned OpBits = cast<IntegerType>(Op->getType())->getBitWidth();
8362 unsigned MidBits = cast<IntegerType>(Src->getType())->getBitWidth();
8363 unsigned DestBits = cast<IntegerType>(CI.getType())->getBitWidth();
8364 unsigned NumSignBits = ComputeNumSignBits(Op);
8366 if (OpBits == DestBits) {
8367 // Op is i32, Mid is i8, and Dest is i32. If Op has more than 24 sign
8368 // bits, it is already ready.
8369 if (NumSignBits > DestBits-MidBits)
8370 return ReplaceInstUsesWith(CI, Op);
8371 } else if (OpBits < DestBits) {
8372 // Op is i32, Mid is i8, and Dest is i64. If Op has more than 24 sign
8373 // bits, just sext from i32.
8374 if (NumSignBits > OpBits-MidBits)
8375 return new SExtInst(Op, CI.getType(), "tmp");
8377 // Op is i64, Mid is i8, and Dest is i32. If Op has more than 56 sign
8378 // bits, just truncate to i32.
8379 if (NumSignBits > OpBits-MidBits)
8380 return new TruncInst(Op, CI.getType(), "tmp");
8384 // If the input is a shl/ashr pair of a same constant, then this is a sign
8385 // extension from a smaller value. If we could trust arbitrary bitwidth
8386 // integers, we could turn this into a truncate to the smaller bit and then
8387 // use a sext for the whole extension. Since we don't, look deeper and check
8388 // for a truncate. If the source and dest are the same type, eliminate the
8389 // trunc and extend and just do shifts. For example, turn:
8390 // %a = trunc i32 %i to i8
8391 // %b = shl i8 %a, 6
8392 // %c = ashr i8 %b, 6
8393 // %d = sext i8 %c to i32
8395 // %a = shl i32 %i, 30
8396 // %d = ashr i32 %a, 30
8398 ConstantInt *BA = 0, *CA = 0;
8399 if (match(Src, m_AShr(m_Shl(m_Value(A), m_ConstantInt(BA)),
8400 m_ConstantInt(CA))) &&
8401 BA == CA && isa<TruncInst>(A)) {
8402 Value *I = cast<TruncInst>(A)->getOperand(0);
8403 if (I->getType() == CI.getType()) {
8404 unsigned MidSize = Src->getType()->getPrimitiveSizeInBits();
8405 unsigned SrcDstSize = CI.getType()->getPrimitiveSizeInBits();
8406 unsigned ShAmt = CA->getZExtValue()+SrcDstSize-MidSize;
8407 Constant *ShAmtV = ConstantInt::get(CI.getType(), ShAmt);
8408 I = InsertNewInstBefore(BinaryOperator::CreateShl(I, ShAmtV,
8410 return BinaryOperator::CreateAShr(I, ShAmtV);
8417 /// FitsInFPType - Return a Constant* for the specified FP constant if it fits
8418 /// in the specified FP type without changing its value.
8419 static Constant *FitsInFPType(ConstantFP *CFP, const fltSemantics &Sem) {
8421 APFloat F = CFP->getValueAPF();
8422 (void)F.convert(Sem, APFloat::rmNearestTiesToEven, &losesInfo);
8424 return ConstantFP::get(F);
8428 /// LookThroughFPExtensions - If this is an fp extension instruction, look
8429 /// through it until we get the source value.
8430 static Value *LookThroughFPExtensions(Value *V) {
8431 if (Instruction *I = dyn_cast<Instruction>(V))
8432 if (I->getOpcode() == Instruction::FPExt)
8433 return LookThroughFPExtensions(I->getOperand(0));
8435 // If this value is a constant, return the constant in the smallest FP type
8436 // that can accurately represent it. This allows us to turn
8437 // (float)((double)X+2.0) into x+2.0f.
8438 if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
8439 if (CFP->getType() == Type::PPC_FP128Ty)
8440 return V; // No constant folding of this.
8441 // See if the value can be truncated to float and then reextended.
8442 if (Value *V = FitsInFPType(CFP, APFloat::IEEEsingle))
8444 if (CFP->getType() == Type::DoubleTy)
8445 return V; // Won't shrink.
8446 if (Value *V = FitsInFPType(CFP, APFloat::IEEEdouble))
8448 // Don't try to shrink to various long double types.
8454 Instruction *InstCombiner::visitFPTrunc(FPTruncInst &CI) {
8455 if (Instruction *I = commonCastTransforms(CI))
8458 // If we have fptrunc(add (fpextend x), (fpextend y)), where x and y are
8459 // smaller than the destination type, we can eliminate the truncate by doing
8460 // the add as the smaller type. This applies to add/sub/mul/div as well as
8461 // many builtins (sqrt, etc).
8462 BinaryOperator *OpI = dyn_cast<BinaryOperator>(CI.getOperand(0));
8463 if (OpI && OpI->hasOneUse()) {
8464 switch (OpI->getOpcode()) {
8466 case Instruction::Add:
8467 case Instruction::Sub:
8468 case Instruction::Mul:
8469 case Instruction::FDiv:
8470 case Instruction::FRem:
8471 const Type *SrcTy = OpI->getType();
8472 Value *LHSTrunc = LookThroughFPExtensions(OpI->getOperand(0));
8473 Value *RHSTrunc = LookThroughFPExtensions(OpI->getOperand(1));
8474 if (LHSTrunc->getType() != SrcTy &&
8475 RHSTrunc->getType() != SrcTy) {
8476 unsigned DstSize = CI.getType()->getPrimitiveSizeInBits();
8477 // If the source types were both smaller than the destination type of
8478 // the cast, do this xform.
8479 if (LHSTrunc->getType()->getPrimitiveSizeInBits() <= DstSize &&
8480 RHSTrunc->getType()->getPrimitiveSizeInBits() <= DstSize) {
8481 LHSTrunc = InsertCastBefore(Instruction::FPExt, LHSTrunc,
8483 RHSTrunc = InsertCastBefore(Instruction::FPExt, RHSTrunc,
8485 return BinaryOperator::Create(OpI->getOpcode(), LHSTrunc, RHSTrunc);
8494 Instruction *InstCombiner::visitFPExt(CastInst &CI) {
8495 return commonCastTransforms(CI);
8498 Instruction *InstCombiner::visitFPToUI(FPToUIInst &FI) {
8499 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0));
8501 return commonCastTransforms(FI);
8503 // fptoui(uitofp(X)) --> X
8504 // fptoui(sitofp(X)) --> X
8505 // This is safe if the intermediate type has enough bits in its mantissa to
8506 // accurately represent all values of X. For example, do not do this with
8507 // i64->float->i64. This is also safe for sitofp case, because any negative
8508 // 'X' value would cause an undefined result for the fptoui.
8509 if ((isa<UIToFPInst>(OpI) || isa<SIToFPInst>(OpI)) &&
8510 OpI->getOperand(0)->getType() == FI.getType() &&
8511 (int)FI.getType()->getPrimitiveSizeInBits() < /*extra bit for sign */
8512 OpI->getType()->getFPMantissaWidth())
8513 return ReplaceInstUsesWith(FI, OpI->getOperand(0));
8515 return commonCastTransforms(FI);
8518 Instruction *InstCombiner::visitFPToSI(FPToSIInst &FI) {
8519 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0));
8521 return commonCastTransforms(FI);
8523 // fptosi(sitofp(X)) --> X
8524 // fptosi(uitofp(X)) --> X
8525 // This is safe if the intermediate type has enough bits in its mantissa to
8526 // accurately represent all values of X. For example, do not do this with
8527 // i64->float->i64. This is also safe for sitofp case, because any negative
8528 // 'X' value would cause an undefined result for the fptoui.
8529 if ((isa<UIToFPInst>(OpI) || isa<SIToFPInst>(OpI)) &&
8530 OpI->getOperand(0)->getType() == FI.getType() &&
8531 (int)FI.getType()->getPrimitiveSizeInBits() <=
8532 OpI->getType()->getFPMantissaWidth())
8533 return ReplaceInstUsesWith(FI, OpI->getOperand(0));
8535 return commonCastTransforms(FI);
8538 Instruction *InstCombiner::visitUIToFP(CastInst &CI) {
8539 return commonCastTransforms(CI);
8542 Instruction *InstCombiner::visitSIToFP(CastInst &CI) {
8543 return commonCastTransforms(CI);
8546 Instruction *InstCombiner::visitPtrToInt(PtrToIntInst &CI) {
8547 // If the destination integer type is smaller than the intptr_t type for
8548 // this target, do a ptrtoint to intptr_t then do a trunc. This allows the
8549 // trunc to be exposed to other transforms. Don't do this for extending
8550 // ptrtoint's, because we don't know if the target sign or zero extends its
8552 if (CI.getType()->getPrimitiveSizeInBits() < TD->getPointerSizeInBits()) {
8553 Value *P = InsertNewInstBefore(new PtrToIntInst(CI.getOperand(0),
8554 TD->getIntPtrType(),
8556 return new TruncInst(P, CI.getType());
8559 return commonPointerCastTransforms(CI);
8562 Instruction *InstCombiner::visitIntToPtr(IntToPtrInst &CI) {
8563 // If the source integer type is larger than the intptr_t type for
8564 // this target, do a trunc to the intptr_t type, then inttoptr of it. This
8565 // allows the trunc to be exposed to other transforms. Don't do this for
8566 // extending inttoptr's, because we don't know if the target sign or zero
8567 // extends to pointers.
8568 if (CI.getOperand(0)->getType()->getPrimitiveSizeInBits() >
8569 TD->getPointerSizeInBits()) {
8570 Value *P = InsertNewInstBefore(new TruncInst(CI.getOperand(0),
8571 TD->getIntPtrType(),
8573 return new IntToPtrInst(P, CI.getType());
8576 if (Instruction *I = commonCastTransforms(CI))
8579 const Type *DestPointee = cast<PointerType>(CI.getType())->getElementType();
8580 if (!DestPointee->isSized()) return 0;
8582 // If this is inttoptr(add (ptrtoint x), cst), try to turn this into a GEP.
8585 if (match(CI.getOperand(0), m_Add(m_Cast<PtrToIntInst>(m_Value(X)),
8586 m_ConstantInt(Cst)))) {
8587 // If the source and destination operands have the same type, see if this
8588 // is a single-index GEP.
8589 if (X->getType() == CI.getType()) {
8590 // Get the size of the pointee type.
8591 uint64_t Size = TD->getTypePaddedSize(DestPointee);
8593 // Convert the constant to intptr type.
8594 APInt Offset = Cst->getValue();
8595 Offset.sextOrTrunc(TD->getPointerSizeInBits());
8597 // If Offset is evenly divisible by Size, we can do this xform.
8598 if (Size && !APIntOps::srem(Offset, APInt(Offset.getBitWidth(), Size))){
8599 Offset = APIntOps::sdiv(Offset, APInt(Offset.getBitWidth(), Size));
8600 return GetElementPtrInst::Create(X, ConstantInt::get(Offset));
8603 // TODO: Could handle other cases, e.g. where add is indexing into field of
8605 } else if (CI.getOperand(0)->hasOneUse() &&
8606 match(CI.getOperand(0), m_Add(m_Value(X), m_ConstantInt(Cst)))) {
8607 // Otherwise, if this is inttoptr(add x, cst), try to turn this into an
8608 // "inttoptr+GEP" instead of "add+intptr".
8610 // Get the size of the pointee type.
8611 uint64_t Size = TD->getTypePaddedSize(DestPointee);
8613 // Convert the constant to intptr type.
8614 APInt Offset = Cst->getValue();
8615 Offset.sextOrTrunc(TD->getPointerSizeInBits());
8617 // If Offset is evenly divisible by Size, we can do this xform.
8618 if (Size && !APIntOps::srem(Offset, APInt(Offset.getBitWidth(), Size))){
8619 Offset = APIntOps::sdiv(Offset, APInt(Offset.getBitWidth(), Size));
8621 Instruction *P = InsertNewInstBefore(new IntToPtrInst(X, CI.getType(),
8623 return GetElementPtrInst::Create(P, ConstantInt::get(Offset), "tmp");
8629 Instruction *InstCombiner::visitBitCast(BitCastInst &CI) {
8630 // If the operands are integer typed then apply the integer transforms,
8631 // otherwise just apply the common ones.
8632 Value *Src = CI.getOperand(0);
8633 const Type *SrcTy = Src->getType();
8634 const Type *DestTy = CI.getType();
8636 if (SrcTy->isInteger() && DestTy->isInteger()) {
8637 if (Instruction *Result = commonIntCastTransforms(CI))
8639 } else if (isa<PointerType>(SrcTy)) {
8640 if (Instruction *I = commonPointerCastTransforms(CI))
8643 if (Instruction *Result = commonCastTransforms(CI))
8648 // Get rid of casts from one type to the same type. These are useless and can
8649 // be replaced by the operand.
8650 if (DestTy == Src->getType())
8651 return ReplaceInstUsesWith(CI, Src);
8653 if (const PointerType *DstPTy = dyn_cast<PointerType>(DestTy)) {
8654 const PointerType *SrcPTy = cast<PointerType>(SrcTy);
8655 const Type *DstElTy = DstPTy->getElementType();
8656 const Type *SrcElTy = SrcPTy->getElementType();
8658 // If the address spaces don't match, don't eliminate the bitcast, which is
8659 // required for changing types.
8660 if (SrcPTy->getAddressSpace() != DstPTy->getAddressSpace())
8663 // If we are casting a malloc or alloca to a pointer to a type of the same
8664 // size, rewrite the allocation instruction to allocate the "right" type.
8665 if (AllocationInst *AI = dyn_cast<AllocationInst>(Src))
8666 if (Instruction *V = PromoteCastOfAllocation(CI, *AI))
8669 // If the source and destination are pointers, and this cast is equivalent
8670 // to a getelementptr X, 0, 0, 0... turn it into the appropriate gep.
8671 // This can enhance SROA and other transforms that want type-safe pointers.
8672 Constant *ZeroUInt = Constant::getNullValue(Type::Int32Ty);
8673 unsigned NumZeros = 0;
8674 while (SrcElTy != DstElTy &&
8675 isa<CompositeType>(SrcElTy) && !isa<PointerType>(SrcElTy) &&
8676 SrcElTy->getNumContainedTypes() /* not "{}" */) {
8677 SrcElTy = cast<CompositeType>(SrcElTy)->getTypeAtIndex(ZeroUInt);
8681 // If we found a path from the src to dest, create the getelementptr now.
8682 if (SrcElTy == DstElTy) {
8683 SmallVector<Value*, 8> Idxs(NumZeros+1, ZeroUInt);
8684 return GetElementPtrInst::Create(Src, Idxs.begin(), Idxs.end(), "",
8685 ((Instruction*) NULL));
8689 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(Src)) {
8690 if (SVI->hasOneUse()) {
8691 // Okay, we have (bitconvert (shuffle ..)). Check to see if this is
8692 // a bitconvert to a vector with the same # elts.
8693 if (isa<VectorType>(DestTy) &&
8694 cast<VectorType>(DestTy)->getNumElements() ==
8695 SVI->getType()->getNumElements() &&
8696 SVI->getType()->getNumElements() ==
8697 cast<VectorType>(SVI->getOperand(0)->getType())->getNumElements()) {
8699 // If either of the operands is a cast from CI.getType(), then
8700 // evaluating the shuffle in the casted destination's type will allow
8701 // us to eliminate at least one cast.
8702 if (((Tmp = dyn_cast<CastInst>(SVI->getOperand(0))) &&
8703 Tmp->getOperand(0)->getType() == DestTy) ||
8704 ((Tmp = dyn_cast<CastInst>(SVI->getOperand(1))) &&
8705 Tmp->getOperand(0)->getType() == DestTy)) {
8706 Value *LHS = InsertCastBefore(Instruction::BitCast,
8707 SVI->getOperand(0), DestTy, CI);
8708 Value *RHS = InsertCastBefore(Instruction::BitCast,
8709 SVI->getOperand(1), DestTy, CI);
8710 // Return a new shuffle vector. Use the same element ID's, as we
8711 // know the vector types match #elts.
8712 return new ShuffleVectorInst(LHS, RHS, SVI->getOperand(2));
8720 /// GetSelectFoldableOperands - We want to turn code that looks like this:
8722 /// %D = select %cond, %C, %A
8724 /// %C = select %cond, %B, 0
8727 /// Assuming that the specified instruction is an operand to the select, return
8728 /// a bitmask indicating which operands of this instruction are foldable if they
8729 /// equal the other incoming value of the select.
8731 static unsigned GetSelectFoldableOperands(Instruction *I) {
8732 switch (I->getOpcode()) {
8733 case Instruction::Add:
8734 case Instruction::Mul:
8735 case Instruction::And:
8736 case Instruction::Or:
8737 case Instruction::Xor:
8738 return 3; // Can fold through either operand.
8739 case Instruction::Sub: // Can only fold on the amount subtracted.
8740 case Instruction::Shl: // Can only fold on the shift amount.
8741 case Instruction::LShr:
8742 case Instruction::AShr:
8745 return 0; // Cannot fold
8749 /// GetSelectFoldableConstant - For the same transformation as the previous
8750 /// function, return the identity constant that goes into the select.
8751 static Constant *GetSelectFoldableConstant(Instruction *I) {
8752 switch (I->getOpcode()) {
8753 default: assert(0 && "This cannot happen!"); abort();
8754 case Instruction::Add:
8755 case Instruction::Sub:
8756 case Instruction::Or:
8757 case Instruction::Xor:
8758 case Instruction::Shl:
8759 case Instruction::LShr:
8760 case Instruction::AShr:
8761 return Constant::getNullValue(I->getType());
8762 case Instruction::And:
8763 return Constant::getAllOnesValue(I->getType());
8764 case Instruction::Mul:
8765 return ConstantInt::get(I->getType(), 1);
8769 /// FoldSelectOpOp - Here we have (select c, TI, FI), and we know that TI and FI
8770 /// have the same opcode and only one use each. Try to simplify this.
8771 Instruction *InstCombiner::FoldSelectOpOp(SelectInst &SI, Instruction *TI,
8773 if (TI->getNumOperands() == 1) {
8774 // If this is a non-volatile load or a cast from the same type,
8777 if (TI->getOperand(0)->getType() != FI->getOperand(0)->getType())
8780 return 0; // unknown unary op.
8783 // Fold this by inserting a select from the input values.
8784 SelectInst *NewSI = SelectInst::Create(SI.getCondition(), TI->getOperand(0),
8785 FI->getOperand(0), SI.getName()+".v");
8786 InsertNewInstBefore(NewSI, SI);
8787 return CastInst::Create(Instruction::CastOps(TI->getOpcode()), NewSI,
8791 // Only handle binary operators here.
8792 if (!isa<BinaryOperator>(TI))
8795 // Figure out if the operations have any operands in common.
8796 Value *MatchOp, *OtherOpT, *OtherOpF;
8798 if (TI->getOperand(0) == FI->getOperand(0)) {
8799 MatchOp = TI->getOperand(0);
8800 OtherOpT = TI->getOperand(1);
8801 OtherOpF = FI->getOperand(1);
8802 MatchIsOpZero = true;
8803 } else if (TI->getOperand(1) == FI->getOperand(1)) {
8804 MatchOp = TI->getOperand(1);
8805 OtherOpT = TI->getOperand(0);
8806 OtherOpF = FI->getOperand(0);
8807 MatchIsOpZero = false;
8808 } else if (!TI->isCommutative()) {
8810 } else if (TI->getOperand(0) == FI->getOperand(1)) {
8811 MatchOp = TI->getOperand(0);
8812 OtherOpT = TI->getOperand(1);
8813 OtherOpF = FI->getOperand(0);
8814 MatchIsOpZero = true;
8815 } else if (TI->getOperand(1) == FI->getOperand(0)) {
8816 MatchOp = TI->getOperand(1);
8817 OtherOpT = TI->getOperand(0);
8818 OtherOpF = FI->getOperand(1);
8819 MatchIsOpZero = true;
8824 // If we reach here, they do have operations in common.
8825 SelectInst *NewSI = SelectInst::Create(SI.getCondition(), OtherOpT,
8826 OtherOpF, SI.getName()+".v");
8827 InsertNewInstBefore(NewSI, SI);
8829 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TI)) {
8831 return BinaryOperator::Create(BO->getOpcode(), MatchOp, NewSI);
8833 return BinaryOperator::Create(BO->getOpcode(), NewSI, MatchOp);
8835 assert(0 && "Shouldn't get here");
8839 /// visitSelectInstWithICmp - Visit a SelectInst that has an
8840 /// ICmpInst as its first operand.
8842 Instruction *InstCombiner::visitSelectInstWithICmp(SelectInst &SI,
8844 bool Changed = false;
8845 ICmpInst::Predicate Pred = ICI->getPredicate();
8846 Value *CmpLHS = ICI->getOperand(0);
8847 Value *CmpRHS = ICI->getOperand(1);
8848 Value *TrueVal = SI.getTrueValue();
8849 Value *FalseVal = SI.getFalseValue();
8851 // Check cases where the comparison is with a constant that
8852 // can be adjusted to fit the min/max idiom. We may edit ICI in
8853 // place here, so make sure the select is the only user.
8854 if (ICI->hasOneUse())
8855 if (ConstantInt *CI = dyn_cast<ConstantInt>(CmpRHS)) {
8858 case ICmpInst::ICMP_ULT:
8859 case ICmpInst::ICMP_SLT: {
8860 // X < MIN ? T : F --> F
8861 if (CI->isMinValue(Pred == ICmpInst::ICMP_SLT))
8862 return ReplaceInstUsesWith(SI, FalseVal);
8863 // X < C ? X : C-1 --> X > C-1 ? C-1 : X
8864 Constant *AdjustedRHS = SubOne(CI);
8865 if ((CmpLHS == TrueVal && AdjustedRHS == FalseVal) ||
8866 (CmpLHS == FalseVal && AdjustedRHS == TrueVal)) {
8867 Pred = ICmpInst::getSwappedPredicate(Pred);
8868 CmpRHS = AdjustedRHS;
8869 std::swap(FalseVal, TrueVal);
8870 ICI->setPredicate(Pred);
8871 ICI->setOperand(1, CmpRHS);
8872 SI.setOperand(1, TrueVal);
8873 SI.setOperand(2, FalseVal);
8878 case ICmpInst::ICMP_UGT:
8879 case ICmpInst::ICMP_SGT: {
8880 // X > MAX ? T : F --> F
8881 if (CI->isMaxValue(Pred == ICmpInst::ICMP_SGT))
8882 return ReplaceInstUsesWith(SI, FalseVal);
8883 // X > C ? X : C+1 --> X < C+1 ? C+1 : X
8884 Constant *AdjustedRHS = AddOne(CI);
8885 if ((CmpLHS == TrueVal && AdjustedRHS == FalseVal) ||
8886 (CmpLHS == FalseVal && AdjustedRHS == TrueVal)) {
8887 Pred = ICmpInst::getSwappedPredicate(Pred);
8888 CmpRHS = AdjustedRHS;
8889 std::swap(FalseVal, TrueVal);
8890 ICI->setPredicate(Pred);
8891 ICI->setOperand(1, CmpRHS);
8892 SI.setOperand(1, TrueVal);
8893 SI.setOperand(2, FalseVal);
8900 // (x <s 0) ? -1 : 0 -> ashr x, 31 -> all ones if signed
8901 // (x >s -1) ? -1 : 0 -> ashr x, 31 -> all ones if not signed
8902 CmpInst::Predicate Pred = CmpInst::BAD_ICMP_PREDICATE;
8903 if (match(TrueVal, m_ConstantInt<-1>()) &&
8904 match(FalseVal, m_ConstantInt<0>()))
8905 Pred = ICI->getPredicate();
8906 else if (match(TrueVal, m_ConstantInt<0>()) &&
8907 match(FalseVal, m_ConstantInt<-1>()))
8908 Pred = CmpInst::getInversePredicate(ICI->getPredicate());
8910 if (Pred != CmpInst::BAD_ICMP_PREDICATE) {
8911 // If we are just checking for a icmp eq of a single bit and zext'ing it
8912 // to an integer, then shift the bit to the appropriate place and then
8913 // cast to integer to avoid the comparison.
8914 const APInt &Op1CV = CI->getValue();
8916 // sext (x <s 0) to i32 --> x>>s31 true if signbit set.
8917 // sext (x >s -1) to i32 --> (x>>s31)^-1 true if signbit clear.
8918 if ((Pred == ICmpInst::ICMP_SLT && Op1CV == 0) ||
8919 (Pred == ICmpInst::ICMP_SGT && Op1CV.isAllOnesValue())) {
8920 Value *In = ICI->getOperand(0);
8921 Value *Sh = ConstantInt::get(In->getType(),
8922 In->getType()->getPrimitiveSizeInBits()-1);
8923 In = InsertNewInstBefore(BinaryOperator::CreateAShr(In, Sh,
8924 In->getName()+".lobit"),
8926 if (In->getType() != SI.getType())
8927 In = CastInst::CreateIntegerCast(In, SI.getType(),
8928 true/*SExt*/, "tmp", ICI);
8930 if (Pred == ICmpInst::ICMP_SGT)
8931 In = InsertNewInstBefore(BinaryOperator::CreateNot(In,
8932 In->getName()+".not"), *ICI);
8934 return ReplaceInstUsesWith(SI, In);
8939 if (CmpLHS == TrueVal && CmpRHS == FalseVal) {
8940 // Transform (X == Y) ? X : Y -> Y
8941 if (Pred == ICmpInst::ICMP_EQ)
8942 return ReplaceInstUsesWith(SI, FalseVal);
8943 // Transform (X != Y) ? X : Y -> X
8944 if (Pred == ICmpInst::ICMP_NE)
8945 return ReplaceInstUsesWith(SI, TrueVal);
8946 /// NOTE: if we wanted to, this is where to detect integer MIN/MAX
8948 } else if (CmpLHS == FalseVal && CmpRHS == TrueVal) {
8949 // Transform (X == Y) ? Y : X -> X
8950 if (Pred == ICmpInst::ICMP_EQ)
8951 return ReplaceInstUsesWith(SI, FalseVal);
8952 // Transform (X != Y) ? Y : X -> Y
8953 if (Pred == ICmpInst::ICMP_NE)
8954 return ReplaceInstUsesWith(SI, TrueVal);
8955 /// NOTE: if we wanted to, this is where to detect integer MIN/MAX
8958 /// NOTE: if we wanted to, this is where to detect integer ABS
8960 return Changed ? &SI : 0;
8963 Instruction *InstCombiner::visitSelectInst(SelectInst &SI) {
8964 Value *CondVal = SI.getCondition();
8965 Value *TrueVal = SI.getTrueValue();
8966 Value *FalseVal = SI.getFalseValue();
8968 // select true, X, Y -> X
8969 // select false, X, Y -> Y
8970 if (ConstantInt *C = dyn_cast<ConstantInt>(CondVal))
8971 return ReplaceInstUsesWith(SI, C->getZExtValue() ? TrueVal : FalseVal);
8973 // select C, X, X -> X
8974 if (TrueVal == FalseVal)
8975 return ReplaceInstUsesWith(SI, TrueVal);
8977 if (isa<UndefValue>(TrueVal)) // select C, undef, X -> X
8978 return ReplaceInstUsesWith(SI, FalseVal);
8979 if (isa<UndefValue>(FalseVal)) // select C, X, undef -> X
8980 return ReplaceInstUsesWith(SI, TrueVal);
8981 if (isa<UndefValue>(CondVal)) { // select undef, X, Y -> X or Y
8982 if (isa<Constant>(TrueVal))
8983 return ReplaceInstUsesWith(SI, TrueVal);
8985 return ReplaceInstUsesWith(SI, FalseVal);
8988 if (SI.getType() == Type::Int1Ty) {
8989 if (ConstantInt *C = dyn_cast<ConstantInt>(TrueVal)) {
8990 if (C->getZExtValue()) {
8991 // Change: A = select B, true, C --> A = or B, C
8992 return BinaryOperator::CreateOr(CondVal, FalseVal);
8994 // Change: A = select B, false, C --> A = and !B, C
8996 InsertNewInstBefore(BinaryOperator::CreateNot(CondVal,
8997 "not."+CondVal->getName()), SI);
8998 return BinaryOperator::CreateAnd(NotCond, FalseVal);
9000 } else if (ConstantInt *C = dyn_cast<ConstantInt>(FalseVal)) {
9001 if (C->getZExtValue() == false) {
9002 // Change: A = select B, C, false --> A = and B, C
9003 return BinaryOperator::CreateAnd(CondVal, TrueVal);
9005 // Change: A = select B, C, true --> A = or !B, C
9007 InsertNewInstBefore(BinaryOperator::CreateNot(CondVal,
9008 "not."+CondVal->getName()), SI);
9009 return BinaryOperator::CreateOr(NotCond, TrueVal);
9013 // select a, b, a -> a&b
9014 // select a, a, b -> a|b
9015 if (CondVal == TrueVal)
9016 return BinaryOperator::CreateOr(CondVal, FalseVal);
9017 else if (CondVal == FalseVal)
9018 return BinaryOperator::CreateAnd(CondVal, TrueVal);
9021 // Selecting between two integer constants?
9022 if (ConstantInt *TrueValC = dyn_cast<ConstantInt>(TrueVal))
9023 if (ConstantInt *FalseValC = dyn_cast<ConstantInt>(FalseVal)) {
9024 // select C, 1, 0 -> zext C to int
9025 if (FalseValC->isZero() && TrueValC->getValue() == 1) {
9026 return CastInst::Create(Instruction::ZExt, CondVal, SI.getType());
9027 } else if (TrueValC->isZero() && FalseValC->getValue() == 1) {
9028 // select C, 0, 1 -> zext !C to int
9030 InsertNewInstBefore(BinaryOperator::CreateNot(CondVal,
9031 "not."+CondVal->getName()), SI);
9032 return CastInst::Create(Instruction::ZExt, NotCond, SI.getType());
9035 if (ICmpInst *IC = dyn_cast<ICmpInst>(SI.getCondition())) {
9037 // (x <s 0) ? -1 : 0 -> ashr x, 31
9038 if (TrueValC->isAllOnesValue() && FalseValC->isZero())
9039 if (ConstantInt *CmpCst = dyn_cast<ConstantInt>(IC->getOperand(1))) {
9040 if (IC->getPredicate() == ICmpInst::ICMP_SLT && CmpCst->isZero()) {
9041 // The comparison constant and the result are not neccessarily the
9042 // same width. Make an all-ones value by inserting a AShr.
9043 Value *X = IC->getOperand(0);
9044 uint32_t Bits = X->getType()->getPrimitiveSizeInBits();
9045 Constant *ShAmt = ConstantInt::get(X->getType(), Bits-1);
9046 Instruction *SRA = BinaryOperator::Create(Instruction::AShr, X,
9048 InsertNewInstBefore(SRA, SI);
9050 // Then cast to the appropriate width.
9051 return CastInst::CreateIntegerCast(SRA, SI.getType(), true);
9056 // If one of the constants is zero (we know they can't both be) and we
9057 // have an icmp instruction with zero, and we have an 'and' with the
9058 // non-constant value, eliminate this whole mess. This corresponds to
9059 // cases like this: ((X & 27) ? 27 : 0)
9060 if (TrueValC->isZero() || FalseValC->isZero())
9061 if (IC->isEquality() && isa<ConstantInt>(IC->getOperand(1)) &&
9062 cast<Constant>(IC->getOperand(1))->isNullValue())
9063 if (Instruction *ICA = dyn_cast<Instruction>(IC->getOperand(0)))
9064 if (ICA->getOpcode() == Instruction::And &&
9065 isa<ConstantInt>(ICA->getOperand(1)) &&
9066 (ICA->getOperand(1) == TrueValC ||
9067 ICA->getOperand(1) == FalseValC) &&
9068 isOneBitSet(cast<ConstantInt>(ICA->getOperand(1)))) {
9069 // Okay, now we know that everything is set up, we just don't
9070 // know whether we have a icmp_ne or icmp_eq and whether the
9071 // true or false val is the zero.
9072 bool ShouldNotVal = !TrueValC->isZero();
9073 ShouldNotVal ^= IC->getPredicate() == ICmpInst::ICMP_NE;
9076 V = InsertNewInstBefore(BinaryOperator::Create(
9077 Instruction::Xor, V, ICA->getOperand(1)), SI);
9078 return ReplaceInstUsesWith(SI, V);
9083 // See if we are selecting two values based on a comparison of the two values.
9084 if (FCmpInst *FCI = dyn_cast<FCmpInst>(CondVal)) {
9085 if (FCI->getOperand(0) == TrueVal && FCI->getOperand(1) == FalseVal) {
9086 // Transform (X == Y) ? X : Y -> Y
9087 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ) {
9088 // This is not safe in general for floating point:
9089 // consider X== -0, Y== +0.
9090 // It becomes safe if either operand is a nonzero constant.
9091 ConstantFP *CFPt, *CFPf;
9092 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) &&
9093 !CFPt->getValueAPF().isZero()) ||
9094 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) &&
9095 !CFPf->getValueAPF().isZero()))
9096 return ReplaceInstUsesWith(SI, FalseVal);
9098 // Transform (X != Y) ? X : Y -> X
9099 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
9100 return ReplaceInstUsesWith(SI, TrueVal);
9101 // NOTE: if we wanted to, this is where to detect MIN/MAX
9103 } else if (FCI->getOperand(0) == FalseVal && FCI->getOperand(1) == TrueVal){
9104 // Transform (X == Y) ? Y : X -> X
9105 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ) {
9106 // This is not safe in general for floating point:
9107 // consider X== -0, Y== +0.
9108 // It becomes safe if either operand is a nonzero constant.
9109 ConstantFP *CFPt, *CFPf;
9110 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) &&
9111 !CFPt->getValueAPF().isZero()) ||
9112 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) &&
9113 !CFPf->getValueAPF().isZero()))
9114 return ReplaceInstUsesWith(SI, FalseVal);
9116 // Transform (X != Y) ? Y : X -> Y
9117 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
9118 return ReplaceInstUsesWith(SI, TrueVal);
9119 // NOTE: if we wanted to, this is where to detect MIN/MAX
9121 // NOTE: if we wanted to, this is where to detect ABS
9124 // See if we are selecting two values based on a comparison of the two values.
9125 if (ICmpInst *ICI = dyn_cast<ICmpInst>(CondVal))
9126 if (Instruction *Result = visitSelectInstWithICmp(SI, ICI))
9129 if (Instruction *TI = dyn_cast<Instruction>(TrueVal))
9130 if (Instruction *FI = dyn_cast<Instruction>(FalseVal))
9131 if (TI->hasOneUse() && FI->hasOneUse()) {
9132 Instruction *AddOp = 0, *SubOp = 0;
9134 // Turn (select C, (op X, Y), (op X, Z)) -> (op X, (select C, Y, Z))
9135 if (TI->getOpcode() == FI->getOpcode())
9136 if (Instruction *IV = FoldSelectOpOp(SI, TI, FI))
9139 // Turn select C, (X+Y), (X-Y) --> (X+(select C, Y, (-Y))). This is
9140 // even legal for FP.
9141 if (TI->getOpcode() == Instruction::Sub &&
9142 FI->getOpcode() == Instruction::Add) {
9143 AddOp = FI; SubOp = TI;
9144 } else if (FI->getOpcode() == Instruction::Sub &&
9145 TI->getOpcode() == Instruction::Add) {
9146 AddOp = TI; SubOp = FI;
9150 Value *OtherAddOp = 0;
9151 if (SubOp->getOperand(0) == AddOp->getOperand(0)) {
9152 OtherAddOp = AddOp->getOperand(1);
9153 } else if (SubOp->getOperand(0) == AddOp->getOperand(1)) {
9154 OtherAddOp = AddOp->getOperand(0);
9158 // So at this point we know we have (Y -> OtherAddOp):
9159 // select C, (add X, Y), (sub X, Z)
9160 Value *NegVal; // Compute -Z
9161 if (Constant *C = dyn_cast<Constant>(SubOp->getOperand(1))) {
9162 NegVal = ConstantExpr::getNeg(C);
9164 NegVal = InsertNewInstBefore(
9165 BinaryOperator::CreateNeg(SubOp->getOperand(1), "tmp"), SI);
9168 Value *NewTrueOp = OtherAddOp;
9169 Value *NewFalseOp = NegVal;
9171 std::swap(NewTrueOp, NewFalseOp);
9172 Instruction *NewSel =
9173 SelectInst::Create(CondVal, NewTrueOp,
9174 NewFalseOp, SI.getName() + ".p");
9176 NewSel = InsertNewInstBefore(NewSel, SI);
9177 return BinaryOperator::CreateAdd(SubOp->getOperand(0), NewSel);
9182 // See if we can fold the select into one of our operands.
9183 if (SI.getType()->isInteger()) {
9184 // See the comment above GetSelectFoldableOperands for a description of the
9185 // transformation we are doing here.
9186 if (Instruction *TVI = dyn_cast<Instruction>(TrueVal))
9187 if (TVI->hasOneUse() && TVI->getNumOperands() == 2 &&
9188 !isa<Constant>(FalseVal))
9189 if (unsigned SFO = GetSelectFoldableOperands(TVI)) {
9190 unsigned OpToFold = 0;
9191 if ((SFO & 1) && FalseVal == TVI->getOperand(0)) {
9193 } else if ((SFO & 2) && FalseVal == TVI->getOperand(1)) {
9198 Constant *C = GetSelectFoldableConstant(TVI);
9199 Instruction *NewSel =
9200 SelectInst::Create(SI.getCondition(),
9201 TVI->getOperand(2-OpToFold), C);
9202 InsertNewInstBefore(NewSel, SI);
9203 NewSel->takeName(TVI);
9204 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TVI))
9205 return BinaryOperator::Create(BO->getOpcode(), FalseVal, NewSel);
9207 assert(0 && "Unknown instruction!!");
9212 if (Instruction *FVI = dyn_cast<Instruction>(FalseVal))
9213 if (FVI->hasOneUse() && FVI->getNumOperands() == 2 &&
9214 !isa<Constant>(TrueVal))
9215 if (unsigned SFO = GetSelectFoldableOperands(FVI)) {
9216 unsigned OpToFold = 0;
9217 if ((SFO & 1) && TrueVal == FVI->getOperand(0)) {
9219 } else if ((SFO & 2) && TrueVal == FVI->getOperand(1)) {
9224 Constant *C = GetSelectFoldableConstant(FVI);
9225 Instruction *NewSel =
9226 SelectInst::Create(SI.getCondition(), C,
9227 FVI->getOperand(2-OpToFold));
9228 InsertNewInstBefore(NewSel, SI);
9229 NewSel->takeName(FVI);
9230 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(FVI))
9231 return BinaryOperator::Create(BO->getOpcode(), TrueVal, NewSel);
9233 assert(0 && "Unknown instruction!!");
9238 if (BinaryOperator::isNot(CondVal)) {
9239 SI.setOperand(0, BinaryOperator::getNotArgument(CondVal));
9240 SI.setOperand(1, FalseVal);
9241 SI.setOperand(2, TrueVal);
9248 /// EnforceKnownAlignment - If the specified pointer points to an object that
9249 /// we control, modify the object's alignment to PrefAlign. This isn't
9250 /// often possible though. If alignment is important, a more reliable approach
9251 /// is to simply align all global variables and allocation instructions to
9252 /// their preferred alignment from the beginning.
9254 static unsigned EnforceKnownAlignment(Value *V,
9255 unsigned Align, unsigned PrefAlign) {
9257 User *U = dyn_cast<User>(V);
9258 if (!U) return Align;
9260 switch (getOpcode(U)) {
9262 case Instruction::BitCast:
9263 return EnforceKnownAlignment(U->getOperand(0), Align, PrefAlign);
9264 case Instruction::GetElementPtr: {
9265 // If all indexes are zero, it is just the alignment of the base pointer.
9266 bool AllZeroOperands = true;
9267 for (User::op_iterator i = U->op_begin() + 1, e = U->op_end(); i != e; ++i)
9268 if (!isa<Constant>(*i) ||
9269 !cast<Constant>(*i)->isNullValue()) {
9270 AllZeroOperands = false;
9274 if (AllZeroOperands) {
9275 // Treat this like a bitcast.
9276 return EnforceKnownAlignment(U->getOperand(0), Align, PrefAlign);
9282 if (GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
9283 // If there is a large requested alignment and we can, bump up the alignment
9285 if (!GV->isDeclaration()) {
9286 if (GV->getAlignment() >= PrefAlign)
9287 Align = GV->getAlignment();
9289 GV->setAlignment(PrefAlign);
9293 } else if (AllocationInst *AI = dyn_cast<AllocationInst>(V)) {
9294 // If there is a requested alignment and if this is an alloca, round up. We
9295 // don't do this for malloc, because some systems can't respect the request.
9296 if (isa<AllocaInst>(AI)) {
9297 if (AI->getAlignment() >= PrefAlign)
9298 Align = AI->getAlignment();
9300 AI->setAlignment(PrefAlign);
9309 /// GetOrEnforceKnownAlignment - If the specified pointer has an alignment that
9310 /// we can determine, return it, otherwise return 0. If PrefAlign is specified,
9311 /// and it is more than the alignment of the ultimate object, see if we can
9312 /// increase the alignment of the ultimate object, making this check succeed.
9313 unsigned InstCombiner::GetOrEnforceKnownAlignment(Value *V,
9314 unsigned PrefAlign) {
9315 unsigned BitWidth = TD ? TD->getTypeSizeInBits(V->getType()) :
9316 sizeof(PrefAlign) * CHAR_BIT;
9317 APInt Mask = APInt::getAllOnesValue(BitWidth);
9318 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
9319 ComputeMaskedBits(V, Mask, KnownZero, KnownOne);
9320 unsigned TrailZ = KnownZero.countTrailingOnes();
9321 unsigned Align = 1u << std::min(BitWidth - 1, TrailZ);
9323 if (PrefAlign > Align)
9324 Align = EnforceKnownAlignment(V, Align, PrefAlign);
9326 // We don't need to make any adjustment.
9330 Instruction *InstCombiner::SimplifyMemTransfer(MemIntrinsic *MI) {
9331 unsigned DstAlign = GetOrEnforceKnownAlignment(MI->getOperand(1));
9332 unsigned SrcAlign = GetOrEnforceKnownAlignment(MI->getOperand(2));
9333 unsigned MinAlign = std::min(DstAlign, SrcAlign);
9334 unsigned CopyAlign = MI->getAlignment();
9336 if (CopyAlign < MinAlign) {
9337 MI->setAlignment(MinAlign);
9341 // If MemCpyInst length is 1/2/4/8 bytes then replace memcpy with
9343 ConstantInt *MemOpLength = dyn_cast<ConstantInt>(MI->getOperand(3));
9344 if (MemOpLength == 0) return 0;
9346 // Source and destination pointer types are always "i8*" for intrinsic. See
9347 // if the size is something we can handle with a single primitive load/store.
9348 // A single load+store correctly handles overlapping memory in the memmove
9350 unsigned Size = MemOpLength->getZExtValue();
9351 if (Size == 0) return MI; // Delete this mem transfer.
9353 if (Size > 8 || (Size&(Size-1)))
9354 return 0; // If not 1/2/4/8 bytes, exit.
9356 // Use an integer load+store unless we can find something better.
9357 Type *NewPtrTy = PointerType::getUnqual(IntegerType::get(Size<<3));
9359 // Memcpy forces the use of i8* for the source and destination. That means
9360 // that if you're using memcpy to move one double around, you'll get a cast
9361 // from double* to i8*. We'd much rather use a double load+store rather than
9362 // an i64 load+store, here because this improves the odds that the source or
9363 // dest address will be promotable. See if we can find a better type than the
9364 // integer datatype.
9365 if (Value *Op = getBitCastOperand(MI->getOperand(1))) {
9366 const Type *SrcETy = cast<PointerType>(Op->getType())->getElementType();
9367 if (SrcETy->isSized() && TD->getTypeStoreSize(SrcETy) == Size) {
9368 // The SrcETy might be something like {{{double}}} or [1 x double]. Rip
9369 // down through these levels if so.
9370 while (!SrcETy->isSingleValueType()) {
9371 if (const StructType *STy = dyn_cast<StructType>(SrcETy)) {
9372 if (STy->getNumElements() == 1)
9373 SrcETy = STy->getElementType(0);
9376 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(SrcETy)) {
9377 if (ATy->getNumElements() == 1)
9378 SrcETy = ATy->getElementType();
9385 if (SrcETy->isSingleValueType())
9386 NewPtrTy = PointerType::getUnqual(SrcETy);
9391 // If the memcpy/memmove provides better alignment info than we can
9393 SrcAlign = std::max(SrcAlign, CopyAlign);
9394 DstAlign = std::max(DstAlign, CopyAlign);
9396 Value *Src = InsertBitCastBefore(MI->getOperand(2), NewPtrTy, *MI);
9397 Value *Dest = InsertBitCastBefore(MI->getOperand(1), NewPtrTy, *MI);
9398 Instruction *L = new LoadInst(Src, "tmp", false, SrcAlign);
9399 InsertNewInstBefore(L, *MI);
9400 InsertNewInstBefore(new StoreInst(L, Dest, false, DstAlign), *MI);
9402 // Set the size of the copy to 0, it will be deleted on the next iteration.
9403 MI->setOperand(3, Constant::getNullValue(MemOpLength->getType()));
9407 Instruction *InstCombiner::SimplifyMemSet(MemSetInst *MI) {
9408 unsigned Alignment = GetOrEnforceKnownAlignment(MI->getDest());
9409 if (MI->getAlignment() < Alignment) {
9410 MI->setAlignment(Alignment);
9414 // Extract the length and alignment and fill if they are constant.
9415 ConstantInt *LenC = dyn_cast<ConstantInt>(MI->getLength());
9416 ConstantInt *FillC = dyn_cast<ConstantInt>(MI->getValue());
9417 if (!LenC || !FillC || FillC->getType() != Type::Int8Ty)
9419 uint64_t Len = LenC->getZExtValue();
9420 Alignment = MI->getAlignment();
9422 // If the length is zero, this is a no-op
9423 if (Len == 0) return MI; // memset(d,c,0,a) -> noop
9425 // memset(s,c,n) -> store s, c (for n=1,2,4,8)
9426 if (Len <= 8 && isPowerOf2_32((uint32_t)Len)) {
9427 const Type *ITy = IntegerType::get(Len*8); // n=1 -> i8.
9429 Value *Dest = MI->getDest();
9430 Dest = InsertBitCastBefore(Dest, PointerType::getUnqual(ITy), *MI);
9432 // Alignment 0 is identity for alignment 1 for memset, but not store.
9433 if (Alignment == 0) Alignment = 1;
9435 // Extract the fill value and store.
9436 uint64_t Fill = FillC->getZExtValue()*0x0101010101010101ULL;
9437 InsertNewInstBefore(new StoreInst(ConstantInt::get(ITy, Fill), Dest, false,
9440 // Set the size of the copy to 0, it will be deleted on the next iteration.
9441 MI->setLength(Constant::getNullValue(LenC->getType()));
9449 /// visitCallInst - CallInst simplification. This mostly only handles folding
9450 /// of intrinsic instructions. For normal calls, it allows visitCallSite to do
9451 /// the heavy lifting.
9453 Instruction *InstCombiner::visitCallInst(CallInst &CI) {
9454 IntrinsicInst *II = dyn_cast<IntrinsicInst>(&CI);
9455 if (!II) return visitCallSite(&CI);
9457 // Intrinsics cannot occur in an invoke, so handle them here instead of in
9459 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(II)) {
9460 bool Changed = false;
9462 // memmove/cpy/set of zero bytes is a noop.
9463 if (Constant *NumBytes = dyn_cast<Constant>(MI->getLength())) {
9464 if (NumBytes->isNullValue()) return EraseInstFromFunction(CI);
9466 if (ConstantInt *CI = dyn_cast<ConstantInt>(NumBytes))
9467 if (CI->getZExtValue() == 1) {
9468 // Replace the instruction with just byte operations. We would
9469 // transform other cases to loads/stores, but we don't know if
9470 // alignment is sufficient.
9474 // If we have a memmove and the source operation is a constant global,
9475 // then the source and dest pointers can't alias, so we can change this
9476 // into a call to memcpy.
9477 if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(MI)) {
9478 if (GlobalVariable *GVSrc = dyn_cast<GlobalVariable>(MMI->getSource()))
9479 if (GVSrc->isConstant()) {
9480 Module *M = CI.getParent()->getParent()->getParent();
9481 Intrinsic::ID MemCpyID = Intrinsic::memcpy;
9483 Tys[0] = CI.getOperand(3)->getType();
9485 Intrinsic::getDeclaration(M, MemCpyID, Tys, 1));
9489 // memmove(x,x,size) -> noop.
9490 if (MMI->getSource() == MMI->getDest())
9491 return EraseInstFromFunction(CI);
9494 // If we can determine a pointer alignment that is bigger than currently
9495 // set, update the alignment.
9496 if (isa<MemTransferInst>(MI)) {
9497 if (Instruction *I = SimplifyMemTransfer(MI))
9499 } else if (MemSetInst *MSI = dyn_cast<MemSetInst>(MI)) {
9500 if (Instruction *I = SimplifyMemSet(MSI))
9504 if (Changed) return II;
9507 switch (II->getIntrinsicID()) {
9509 case Intrinsic::bswap:
9510 // bswap(bswap(x)) -> x
9511 if (IntrinsicInst *Operand = dyn_cast<IntrinsicInst>(II->getOperand(1)))
9512 if (Operand->getIntrinsicID() == Intrinsic::bswap)
9513 return ReplaceInstUsesWith(CI, Operand->getOperand(1));
9515 case Intrinsic::ppc_altivec_lvx:
9516 case Intrinsic::ppc_altivec_lvxl:
9517 case Intrinsic::x86_sse_loadu_ps:
9518 case Intrinsic::x86_sse2_loadu_pd:
9519 case Intrinsic::x86_sse2_loadu_dq:
9520 // Turn PPC lvx -> load if the pointer is known aligned.
9521 // Turn X86 loadups -> load if the pointer is known aligned.
9522 if (GetOrEnforceKnownAlignment(II->getOperand(1), 16) >= 16) {
9523 Value *Ptr = InsertBitCastBefore(II->getOperand(1),
9524 PointerType::getUnqual(II->getType()),
9526 return new LoadInst(Ptr);
9529 case Intrinsic::ppc_altivec_stvx:
9530 case Intrinsic::ppc_altivec_stvxl:
9531 // Turn stvx -> store if the pointer is known aligned.
9532 if (GetOrEnforceKnownAlignment(II->getOperand(2), 16) >= 16) {
9533 const Type *OpPtrTy =
9534 PointerType::getUnqual(II->getOperand(1)->getType());
9535 Value *Ptr = InsertBitCastBefore(II->getOperand(2), OpPtrTy, CI);
9536 return new StoreInst(II->getOperand(1), Ptr);
9539 case Intrinsic::x86_sse_storeu_ps:
9540 case Intrinsic::x86_sse2_storeu_pd:
9541 case Intrinsic::x86_sse2_storeu_dq:
9542 // Turn X86 storeu -> store if the pointer is known aligned.
9543 if (GetOrEnforceKnownAlignment(II->getOperand(1), 16) >= 16) {
9544 const Type *OpPtrTy =
9545 PointerType::getUnqual(II->getOperand(2)->getType());
9546 Value *Ptr = InsertBitCastBefore(II->getOperand(1), OpPtrTy, CI);
9547 return new StoreInst(II->getOperand(2), Ptr);
9551 case Intrinsic::x86_sse_cvttss2si: {
9552 // These intrinsics only demands the 0th element of its input vector. If
9553 // we can simplify the input based on that, do so now.
9555 cast<VectorType>(II->getOperand(1)->getType())->getNumElements();
9556 APInt DemandedElts(VWidth, 1);
9557 APInt UndefElts(VWidth, 0);
9558 if (Value *V = SimplifyDemandedVectorElts(II->getOperand(1), DemandedElts,
9560 II->setOperand(1, V);
9566 case Intrinsic::ppc_altivec_vperm:
9567 // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant.
9568 if (ConstantVector *Mask = dyn_cast<ConstantVector>(II->getOperand(3))) {
9569 assert(Mask->getNumOperands() == 16 && "Bad type for intrinsic!");
9571 // Check that all of the elements are integer constants or undefs.
9572 bool AllEltsOk = true;
9573 for (unsigned i = 0; i != 16; ++i) {
9574 if (!isa<ConstantInt>(Mask->getOperand(i)) &&
9575 !isa<UndefValue>(Mask->getOperand(i))) {
9582 // Cast the input vectors to byte vectors.
9583 Value *Op0 =InsertBitCastBefore(II->getOperand(1),Mask->getType(),CI);
9584 Value *Op1 =InsertBitCastBefore(II->getOperand(2),Mask->getType(),CI);
9585 Value *Result = UndefValue::get(Op0->getType());
9587 // Only extract each element once.
9588 Value *ExtractedElts[32];
9589 memset(ExtractedElts, 0, sizeof(ExtractedElts));
9591 for (unsigned i = 0; i != 16; ++i) {
9592 if (isa<UndefValue>(Mask->getOperand(i)))
9594 unsigned Idx=cast<ConstantInt>(Mask->getOperand(i))->getZExtValue();
9595 Idx &= 31; // Match the hardware behavior.
9597 if (ExtractedElts[Idx] == 0) {
9599 new ExtractElementInst(Idx < 16 ? Op0 : Op1, Idx&15, "tmp");
9600 InsertNewInstBefore(Elt, CI);
9601 ExtractedElts[Idx] = Elt;
9604 // Insert this value into the result vector.
9605 Result = InsertElementInst::Create(Result, ExtractedElts[Idx],
9607 InsertNewInstBefore(cast<Instruction>(Result), CI);
9609 return CastInst::Create(Instruction::BitCast, Result, CI.getType());
9614 case Intrinsic::stackrestore: {
9615 // If the save is right next to the restore, remove the restore. This can
9616 // happen when variable allocas are DCE'd.
9617 if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getOperand(1))) {
9618 if (SS->getIntrinsicID() == Intrinsic::stacksave) {
9619 BasicBlock::iterator BI = SS;
9621 return EraseInstFromFunction(CI);
9625 // Scan down this block to see if there is another stack restore in the
9626 // same block without an intervening call/alloca.
9627 BasicBlock::iterator BI = II;
9628 TerminatorInst *TI = II->getParent()->getTerminator();
9629 bool CannotRemove = false;
9630 for (++BI; &*BI != TI; ++BI) {
9631 if (isa<AllocaInst>(BI)) {
9632 CannotRemove = true;
9635 if (CallInst *BCI = dyn_cast<CallInst>(BI)) {
9636 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(BCI)) {
9637 // If there is a stackrestore below this one, remove this one.
9638 if (II->getIntrinsicID() == Intrinsic::stackrestore)
9639 return EraseInstFromFunction(CI);
9640 // Otherwise, ignore the intrinsic.
9642 // If we found a non-intrinsic call, we can't remove the stack
9644 CannotRemove = true;
9650 // If the stack restore is in a return/unwind block and if there are no
9651 // allocas or calls between the restore and the return, nuke the restore.
9652 if (!CannotRemove && (isa<ReturnInst>(TI) || isa<UnwindInst>(TI)))
9653 return EraseInstFromFunction(CI);
9658 return visitCallSite(II);
9661 // InvokeInst simplification
9663 Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) {
9664 return visitCallSite(&II);
9667 /// isSafeToEliminateVarargsCast - If this cast does not affect the value
9668 /// passed through the varargs area, we can eliminate the use of the cast.
9669 static bool isSafeToEliminateVarargsCast(const CallSite CS,
9670 const CastInst * const CI,
9671 const TargetData * const TD,
9673 if (!CI->isLosslessCast())
9676 // The size of ByVal arguments is derived from the type, so we
9677 // can't change to a type with a different size. If the size were
9678 // passed explicitly we could avoid this check.
9679 if (!CS.paramHasAttr(ix, Attribute::ByVal))
9683 cast<PointerType>(CI->getOperand(0)->getType())->getElementType();
9684 const Type* DstTy = cast<PointerType>(CI->getType())->getElementType();
9685 if (!SrcTy->isSized() || !DstTy->isSized())
9687 if (TD->getTypePaddedSize(SrcTy) != TD->getTypePaddedSize(DstTy))
9692 // visitCallSite - Improvements for call and invoke instructions.
9694 Instruction *InstCombiner::visitCallSite(CallSite CS) {
9695 bool Changed = false;
9697 // If the callee is a constexpr cast of a function, attempt to move the cast
9698 // to the arguments of the call/invoke.
9699 if (transformConstExprCastCall(CS)) return 0;
9701 Value *Callee = CS.getCalledValue();
9703 if (Function *CalleeF = dyn_cast<Function>(Callee))
9704 if (CalleeF->getCallingConv() != CS.getCallingConv()) {
9705 Instruction *OldCall = CS.getInstruction();
9706 // If the call and callee calling conventions don't match, this call must
9707 // be unreachable, as the call is undefined.
9708 new StoreInst(ConstantInt::getTrue(),
9709 UndefValue::get(PointerType::getUnqual(Type::Int1Ty)),
9711 if (!OldCall->use_empty())
9712 OldCall->replaceAllUsesWith(UndefValue::get(OldCall->getType()));
9713 if (isa<CallInst>(OldCall)) // Not worth removing an invoke here.
9714 return EraseInstFromFunction(*OldCall);
9718 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) {
9719 // This instruction is not reachable, just remove it. We insert a store to
9720 // undef so that we know that this code is not reachable, despite the fact
9721 // that we can't modify the CFG here.
9722 new StoreInst(ConstantInt::getTrue(),
9723 UndefValue::get(PointerType::getUnqual(Type::Int1Ty)),
9724 CS.getInstruction());
9726 if (!CS.getInstruction()->use_empty())
9727 CS.getInstruction()->
9728 replaceAllUsesWith(UndefValue::get(CS.getInstruction()->getType()));
9730 if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) {
9731 // Don't break the CFG, insert a dummy cond branch.
9732 BranchInst::Create(II->getNormalDest(), II->getUnwindDest(),
9733 ConstantInt::getTrue(), II);
9735 return EraseInstFromFunction(*CS.getInstruction());
9738 if (BitCastInst *BC = dyn_cast<BitCastInst>(Callee))
9739 if (IntrinsicInst *In = dyn_cast<IntrinsicInst>(BC->getOperand(0)))
9740 if (In->getIntrinsicID() == Intrinsic::init_trampoline)
9741 return transformCallThroughTrampoline(CS);
9743 const PointerType *PTy = cast<PointerType>(Callee->getType());
9744 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
9745 if (FTy->isVarArg()) {
9746 int ix = FTy->getNumParams() + (isa<InvokeInst>(Callee) ? 3 : 1);
9747 // See if we can optimize any arguments passed through the varargs area of
9749 for (CallSite::arg_iterator I = CS.arg_begin()+FTy->getNumParams(),
9750 E = CS.arg_end(); I != E; ++I, ++ix) {
9751 CastInst *CI = dyn_cast<CastInst>(*I);
9752 if (CI && isSafeToEliminateVarargsCast(CS, CI, TD, ix)) {
9753 *I = CI->getOperand(0);
9759 if (isa<InlineAsm>(Callee) && !CS.doesNotThrow()) {
9760 // Inline asm calls cannot throw - mark them 'nounwind'.
9761 CS.setDoesNotThrow();
9765 return Changed ? CS.getInstruction() : 0;
9768 // transformConstExprCastCall - If the callee is a constexpr cast of a function,
9769 // attempt to move the cast to the arguments of the call/invoke.
9771 bool InstCombiner::transformConstExprCastCall(CallSite CS) {
9772 if (!isa<ConstantExpr>(CS.getCalledValue())) return false;
9773 ConstantExpr *CE = cast<ConstantExpr>(CS.getCalledValue());
9774 if (CE->getOpcode() != Instruction::BitCast ||
9775 !isa<Function>(CE->getOperand(0)))
9777 Function *Callee = cast<Function>(CE->getOperand(0));
9778 Instruction *Caller = CS.getInstruction();
9779 const AttrListPtr &CallerPAL = CS.getAttributes();
9781 // Okay, this is a cast from a function to a different type. Unless doing so
9782 // would cause a type conversion of one of our arguments, change this call to
9783 // be a direct call with arguments casted to the appropriate types.
9785 const FunctionType *FT = Callee->getFunctionType();
9786 const Type *OldRetTy = Caller->getType();
9787 const Type *NewRetTy = FT->getReturnType();
9789 if (isa<StructType>(NewRetTy))
9790 return false; // TODO: Handle multiple return values.
9792 // Check to see if we are changing the return type...
9793 if (OldRetTy != NewRetTy) {
9794 if (Callee->isDeclaration() &&
9795 // Conversion is ok if changing from one pointer type to another or from
9796 // a pointer to an integer of the same size.
9797 !((isa<PointerType>(OldRetTy) || OldRetTy == TD->getIntPtrType()) &&
9798 (isa<PointerType>(NewRetTy) || NewRetTy == TD->getIntPtrType())))
9799 return false; // Cannot transform this return value.
9801 if (!Caller->use_empty() &&
9802 // void -> non-void is handled specially
9803 NewRetTy != Type::VoidTy && !CastInst::isCastable(NewRetTy, OldRetTy))
9804 return false; // Cannot transform this return value.
9806 if (!CallerPAL.isEmpty() && !Caller->use_empty()) {
9807 Attributes RAttrs = CallerPAL.getRetAttributes();
9808 if (RAttrs & Attribute::typeIncompatible(NewRetTy))
9809 return false; // Attribute not compatible with transformed value.
9812 // If the callsite is an invoke instruction, and the return value is used by
9813 // a PHI node in a successor, we cannot change the return type of the call
9814 // because there is no place to put the cast instruction (without breaking
9815 // the critical edge). Bail out in this case.
9816 if (!Caller->use_empty())
9817 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller))
9818 for (Value::use_iterator UI = II->use_begin(), E = II->use_end();
9820 if (PHINode *PN = dyn_cast<PHINode>(*UI))
9821 if (PN->getParent() == II->getNormalDest() ||
9822 PN->getParent() == II->getUnwindDest())
9826 unsigned NumActualArgs = unsigned(CS.arg_end()-CS.arg_begin());
9827 unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs);
9829 CallSite::arg_iterator AI = CS.arg_begin();
9830 for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) {
9831 const Type *ParamTy = FT->getParamType(i);
9832 const Type *ActTy = (*AI)->getType();
9834 if (!CastInst::isCastable(ActTy, ParamTy))
9835 return false; // Cannot transform this parameter value.
9837 if (CallerPAL.getParamAttributes(i + 1)
9838 & Attribute::typeIncompatible(ParamTy))
9839 return false; // Attribute not compatible with transformed value.
9841 // Converting from one pointer type to another or between a pointer and an
9842 // integer of the same size is safe even if we do not have a body.
9843 bool isConvertible = ActTy == ParamTy ||
9844 ((isa<PointerType>(ParamTy) || ParamTy == TD->getIntPtrType()) &&
9845 (isa<PointerType>(ActTy) || ActTy == TD->getIntPtrType()));
9846 if (Callee->isDeclaration() && !isConvertible) return false;
9849 if (FT->getNumParams() < NumActualArgs && !FT->isVarArg() &&
9850 Callee->isDeclaration())
9851 return false; // Do not delete arguments unless we have a function body.
9853 if (FT->getNumParams() < NumActualArgs && FT->isVarArg() &&
9854 !CallerPAL.isEmpty())
9855 // In this case we have more arguments than the new function type, but we
9856 // won't be dropping them. Check that these extra arguments have attributes
9857 // that are compatible with being a vararg call argument.
9858 for (unsigned i = CallerPAL.getNumSlots(); i; --i) {
9859 if (CallerPAL.getSlot(i - 1).Index <= FT->getNumParams())
9861 Attributes PAttrs = CallerPAL.getSlot(i - 1).Attrs;
9862 if (PAttrs & Attribute::VarArgsIncompatible)
9866 // Okay, we decided that this is a safe thing to do: go ahead and start
9867 // inserting cast instructions as necessary...
9868 std::vector<Value*> Args;
9869 Args.reserve(NumActualArgs);
9870 SmallVector<AttributeWithIndex, 8> attrVec;
9871 attrVec.reserve(NumCommonArgs);
9873 // Get any return attributes.
9874 Attributes RAttrs = CallerPAL.getRetAttributes();
9876 // If the return value is not being used, the type may not be compatible
9877 // with the existing attributes. Wipe out any problematic attributes.
9878 RAttrs &= ~Attribute::typeIncompatible(NewRetTy);
9880 // Add the new return attributes.
9882 attrVec.push_back(AttributeWithIndex::get(0, RAttrs));
9884 AI = CS.arg_begin();
9885 for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) {
9886 const Type *ParamTy = FT->getParamType(i);
9887 if ((*AI)->getType() == ParamTy) {
9888 Args.push_back(*AI);
9890 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI,
9891 false, ParamTy, false);
9892 CastInst *NewCast = CastInst::Create(opcode, *AI, ParamTy, "tmp");
9893 Args.push_back(InsertNewInstBefore(NewCast, *Caller));
9896 // Add any parameter attributes.
9897 if (Attributes PAttrs = CallerPAL.getParamAttributes(i + 1))
9898 attrVec.push_back(AttributeWithIndex::get(i + 1, PAttrs));
9901 // If the function takes more arguments than the call was taking, add them
9903 for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i)
9904 Args.push_back(Constant::getNullValue(FT->getParamType(i)));
9906 // If we are removing arguments to the function, emit an obnoxious warning...
9907 if (FT->getNumParams() < NumActualArgs) {
9908 if (!FT->isVarArg()) {
9909 cerr << "WARNING: While resolving call to function '"
9910 << Callee->getName() << "' arguments were dropped!\n";
9912 // Add all of the arguments in their promoted form to the arg list...
9913 for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) {
9914 const Type *PTy = getPromotedType((*AI)->getType());
9915 if (PTy != (*AI)->getType()) {
9916 // Must promote to pass through va_arg area!
9917 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI, false,
9919 Instruction *Cast = CastInst::Create(opcode, *AI, PTy, "tmp");
9920 InsertNewInstBefore(Cast, *Caller);
9921 Args.push_back(Cast);
9923 Args.push_back(*AI);
9926 // Add any parameter attributes.
9927 if (Attributes PAttrs = CallerPAL.getParamAttributes(i + 1))
9928 attrVec.push_back(AttributeWithIndex::get(i + 1, PAttrs));
9933 if (Attributes FnAttrs = CallerPAL.getFnAttributes())
9934 attrVec.push_back(AttributeWithIndex::get(~0, FnAttrs));
9936 if (NewRetTy == Type::VoidTy)
9937 Caller->setName(""); // Void type should not have a name.
9939 const AttrListPtr &NewCallerPAL = AttrListPtr::get(attrVec.begin(),attrVec.end());
9942 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
9943 NC = InvokeInst::Create(Callee, II->getNormalDest(), II->getUnwindDest(),
9944 Args.begin(), Args.end(),
9945 Caller->getName(), Caller);
9946 cast<InvokeInst>(NC)->setCallingConv(II->getCallingConv());
9947 cast<InvokeInst>(NC)->setAttributes(NewCallerPAL);
9949 NC = CallInst::Create(Callee, Args.begin(), Args.end(),
9950 Caller->getName(), Caller);
9951 CallInst *CI = cast<CallInst>(Caller);
9952 if (CI->isTailCall())
9953 cast<CallInst>(NC)->setTailCall();
9954 cast<CallInst>(NC)->setCallingConv(CI->getCallingConv());
9955 cast<CallInst>(NC)->setAttributes(NewCallerPAL);
9958 // Insert a cast of the return type as necessary.
9960 if (OldRetTy != NV->getType() && !Caller->use_empty()) {
9961 if (NV->getType() != Type::VoidTy) {
9962 Instruction::CastOps opcode = CastInst::getCastOpcode(NC, false,
9964 NV = NC = CastInst::Create(opcode, NC, OldRetTy, "tmp");
9966 // If this is an invoke instruction, we should insert it after the first
9967 // non-phi, instruction in the normal successor block.
9968 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
9969 BasicBlock::iterator I = II->getNormalDest()->getFirstNonPHI();
9970 InsertNewInstBefore(NC, *I);
9972 // Otherwise, it's a call, just insert cast right after the call instr
9973 InsertNewInstBefore(NC, *Caller);
9975 AddUsersToWorkList(*Caller);
9977 NV = UndefValue::get(Caller->getType());
9981 if (Caller->getType() != Type::VoidTy && !Caller->use_empty())
9982 Caller->replaceAllUsesWith(NV);
9983 Caller->eraseFromParent();
9984 RemoveFromWorkList(Caller);
9988 // transformCallThroughTrampoline - Turn a call to a function created by the
9989 // init_trampoline intrinsic into a direct call to the underlying function.
9991 Instruction *InstCombiner::transformCallThroughTrampoline(CallSite CS) {
9992 Value *Callee = CS.getCalledValue();
9993 const PointerType *PTy = cast<PointerType>(Callee->getType());
9994 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
9995 const AttrListPtr &Attrs = CS.getAttributes();
9997 // If the call already has the 'nest' attribute somewhere then give up -
9998 // otherwise 'nest' would occur twice after splicing in the chain.
9999 if (Attrs.hasAttrSomewhere(Attribute::Nest))
10002 IntrinsicInst *Tramp =
10003 cast<IntrinsicInst>(cast<BitCastInst>(Callee)->getOperand(0));
10005 Function *NestF = cast<Function>(Tramp->getOperand(2)->stripPointerCasts());
10006 const PointerType *NestFPTy = cast<PointerType>(NestF->getType());
10007 const FunctionType *NestFTy = cast<FunctionType>(NestFPTy->getElementType());
10009 const AttrListPtr &NestAttrs = NestF->getAttributes();
10010 if (!NestAttrs.isEmpty()) {
10011 unsigned NestIdx = 1;
10012 const Type *NestTy = 0;
10013 Attributes NestAttr = Attribute::None;
10015 // Look for a parameter marked with the 'nest' attribute.
10016 for (FunctionType::param_iterator I = NestFTy->param_begin(),
10017 E = NestFTy->param_end(); I != E; ++NestIdx, ++I)
10018 if (NestAttrs.paramHasAttr(NestIdx, Attribute::Nest)) {
10019 // Record the parameter type and any other attributes.
10021 NestAttr = NestAttrs.getParamAttributes(NestIdx);
10026 Instruction *Caller = CS.getInstruction();
10027 std::vector<Value*> NewArgs;
10028 NewArgs.reserve(unsigned(CS.arg_end()-CS.arg_begin())+1);
10030 SmallVector<AttributeWithIndex, 8> NewAttrs;
10031 NewAttrs.reserve(Attrs.getNumSlots() + 1);
10033 // Insert the nest argument into the call argument list, which may
10034 // mean appending it. Likewise for attributes.
10036 // Add any result attributes.
10037 if (Attributes Attr = Attrs.getRetAttributes())
10038 NewAttrs.push_back(AttributeWithIndex::get(0, Attr));
10042 CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
10044 if (Idx == NestIdx) {
10045 // Add the chain argument and attributes.
10046 Value *NestVal = Tramp->getOperand(3);
10047 if (NestVal->getType() != NestTy)
10048 NestVal = new BitCastInst(NestVal, NestTy, "nest", Caller);
10049 NewArgs.push_back(NestVal);
10050 NewAttrs.push_back(AttributeWithIndex::get(NestIdx, NestAttr));
10056 // Add the original argument and attributes.
10057 NewArgs.push_back(*I);
10058 if (Attributes Attr = Attrs.getParamAttributes(Idx))
10060 (AttributeWithIndex::get(Idx + (Idx >= NestIdx), Attr));
10066 // Add any function attributes.
10067 if (Attributes Attr = Attrs.getFnAttributes())
10068 NewAttrs.push_back(AttributeWithIndex::get(~0, Attr));
10070 // The trampoline may have been bitcast to a bogus type (FTy).
10071 // Handle this by synthesizing a new function type, equal to FTy
10072 // with the chain parameter inserted.
10074 std::vector<const Type*> NewTypes;
10075 NewTypes.reserve(FTy->getNumParams()+1);
10077 // Insert the chain's type into the list of parameter types, which may
10078 // mean appending it.
10081 FunctionType::param_iterator I = FTy->param_begin(),
10082 E = FTy->param_end();
10085 if (Idx == NestIdx)
10086 // Add the chain's type.
10087 NewTypes.push_back(NestTy);
10092 // Add the original type.
10093 NewTypes.push_back(*I);
10099 // Replace the trampoline call with a direct call. Let the generic
10100 // code sort out any function type mismatches.
10101 FunctionType *NewFTy =
10102 FunctionType::get(FTy->getReturnType(), NewTypes, FTy->isVarArg());
10103 Constant *NewCallee = NestF->getType() == PointerType::getUnqual(NewFTy) ?
10104 NestF : ConstantExpr::getBitCast(NestF, PointerType::getUnqual(NewFTy));
10105 const AttrListPtr &NewPAL = AttrListPtr::get(NewAttrs.begin(),NewAttrs.end());
10107 Instruction *NewCaller;
10108 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
10109 NewCaller = InvokeInst::Create(NewCallee,
10110 II->getNormalDest(), II->getUnwindDest(),
10111 NewArgs.begin(), NewArgs.end(),
10112 Caller->getName(), Caller);
10113 cast<InvokeInst>(NewCaller)->setCallingConv(II->getCallingConv());
10114 cast<InvokeInst>(NewCaller)->setAttributes(NewPAL);
10116 NewCaller = CallInst::Create(NewCallee, NewArgs.begin(), NewArgs.end(),
10117 Caller->getName(), Caller);
10118 if (cast<CallInst>(Caller)->isTailCall())
10119 cast<CallInst>(NewCaller)->setTailCall();
10120 cast<CallInst>(NewCaller)->
10121 setCallingConv(cast<CallInst>(Caller)->getCallingConv());
10122 cast<CallInst>(NewCaller)->setAttributes(NewPAL);
10124 if (Caller->getType() != Type::VoidTy && !Caller->use_empty())
10125 Caller->replaceAllUsesWith(NewCaller);
10126 Caller->eraseFromParent();
10127 RemoveFromWorkList(Caller);
10132 // Replace the trampoline call with a direct call. Since there is no 'nest'
10133 // parameter, there is no need to adjust the argument list. Let the generic
10134 // code sort out any function type mismatches.
10135 Constant *NewCallee =
10136 NestF->getType() == PTy ? NestF : ConstantExpr::getBitCast(NestF, PTy);
10137 CS.setCalledFunction(NewCallee);
10138 return CS.getInstruction();
10141 /// FoldPHIArgBinOpIntoPHI - If we have something like phi [add (a,b), add(c,d)]
10142 /// and if a/b/c/d and the add's all have a single use, turn this into two phi's
10143 /// and a single binop.
10144 Instruction *InstCombiner::FoldPHIArgBinOpIntoPHI(PHINode &PN) {
10145 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
10146 assert(isa<BinaryOperator>(FirstInst) || isa<CmpInst>(FirstInst));
10147 unsigned Opc = FirstInst->getOpcode();
10148 Value *LHSVal = FirstInst->getOperand(0);
10149 Value *RHSVal = FirstInst->getOperand(1);
10151 const Type *LHSType = LHSVal->getType();
10152 const Type *RHSType = RHSVal->getType();
10154 // Scan to see if all operands are the same opcode, all have one use, and all
10155 // kill their operands (i.e. the operands have one use).
10156 for (unsigned i = 1; i != PN.getNumIncomingValues(); ++i) {
10157 Instruction *I = dyn_cast<Instruction>(PN.getIncomingValue(i));
10158 if (!I || I->getOpcode() != Opc || !I->hasOneUse() ||
10159 // Verify type of the LHS matches so we don't fold cmp's of different
10160 // types or GEP's with different index types.
10161 I->getOperand(0)->getType() != LHSType ||
10162 I->getOperand(1)->getType() != RHSType)
10165 // If they are CmpInst instructions, check their predicates
10166 if (Opc == Instruction::ICmp || Opc == Instruction::FCmp)
10167 if (cast<CmpInst>(I)->getPredicate() !=
10168 cast<CmpInst>(FirstInst)->getPredicate())
10171 // Keep track of which operand needs a phi node.
10172 if (I->getOperand(0) != LHSVal) LHSVal = 0;
10173 if (I->getOperand(1) != RHSVal) RHSVal = 0;
10176 // Otherwise, this is safe to transform!
10178 Value *InLHS = FirstInst->getOperand(0);
10179 Value *InRHS = FirstInst->getOperand(1);
10180 PHINode *NewLHS = 0, *NewRHS = 0;
10182 NewLHS = PHINode::Create(LHSType,
10183 FirstInst->getOperand(0)->getName() + ".pn");
10184 NewLHS->reserveOperandSpace(PN.getNumOperands()/2);
10185 NewLHS->addIncoming(InLHS, PN.getIncomingBlock(0));
10186 InsertNewInstBefore(NewLHS, PN);
10191 NewRHS = PHINode::Create(RHSType,
10192 FirstInst->getOperand(1)->getName() + ".pn");
10193 NewRHS->reserveOperandSpace(PN.getNumOperands()/2);
10194 NewRHS->addIncoming(InRHS, PN.getIncomingBlock(0));
10195 InsertNewInstBefore(NewRHS, PN);
10199 // Add all operands to the new PHIs.
10200 if (NewLHS || NewRHS) {
10201 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
10202 Instruction *InInst = cast<Instruction>(PN.getIncomingValue(i));
10204 Value *NewInLHS = InInst->getOperand(0);
10205 NewLHS->addIncoming(NewInLHS, PN.getIncomingBlock(i));
10208 Value *NewInRHS = InInst->getOperand(1);
10209 NewRHS->addIncoming(NewInRHS, PN.getIncomingBlock(i));
10214 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
10215 return BinaryOperator::Create(BinOp->getOpcode(), LHSVal, RHSVal);
10216 CmpInst *CIOp = cast<CmpInst>(FirstInst);
10217 return CmpInst::Create(CIOp->getOpcode(), CIOp->getPredicate(), LHSVal,
10221 Instruction *InstCombiner::FoldPHIArgGEPIntoPHI(PHINode &PN) {
10222 GetElementPtrInst *FirstInst =cast<GetElementPtrInst>(PN.getIncomingValue(0));
10224 SmallVector<Value*, 16> FixedOperands(FirstInst->op_begin(),
10225 FirstInst->op_end());
10226 // This is true if all GEP bases are allocas and if all indices into them are
10228 bool AllBasePointersAreAllocas = true;
10230 // Scan to see if all operands are the same opcode, all have one use, and all
10231 // kill their operands (i.e. the operands have one use).
10232 for (unsigned i = 1; i != PN.getNumIncomingValues(); ++i) {
10233 GetElementPtrInst *GEP= dyn_cast<GetElementPtrInst>(PN.getIncomingValue(i));
10234 if (!GEP || !GEP->hasOneUse() || GEP->getType() != FirstInst->getType() ||
10235 GEP->getNumOperands() != FirstInst->getNumOperands())
10238 // Keep track of whether or not all GEPs are of alloca pointers.
10239 if (AllBasePointersAreAllocas &&
10240 (!isa<AllocaInst>(GEP->getOperand(0)) ||
10241 !GEP->hasAllConstantIndices()))
10242 AllBasePointersAreAllocas = false;
10244 // Compare the operand lists.
10245 for (unsigned op = 0, e = FirstInst->getNumOperands(); op != e; ++op) {
10246 if (FirstInst->getOperand(op) == GEP->getOperand(op))
10249 // Don't merge two GEPs when two operands differ (introducing phi nodes)
10250 // if one of the PHIs has a constant for the index. The index may be
10251 // substantially cheaper to compute for the constants, so making it a
10252 // variable index could pessimize the path. This also handles the case
10253 // for struct indices, which must always be constant.
10254 if (isa<ConstantInt>(FirstInst->getOperand(op)) ||
10255 isa<ConstantInt>(GEP->getOperand(op)))
10258 if (FirstInst->getOperand(op)->getType() !=GEP->getOperand(op)->getType())
10260 FixedOperands[op] = 0; // Needs a PHI.
10264 // If all of the base pointers of the PHI'd GEPs are from allocas, don't
10265 // bother doing this transformation. At best, this will just save a bit of
10266 // offset calculation, but all the predecessors will have to materialize the
10267 // stack address into a register anyway. We'd actually rather *clone* the
10268 // load up into the predecessors so that we have a load of a gep of an alloca,
10269 // which can usually all be folded into the load.
10270 if (AllBasePointersAreAllocas)
10273 // Otherwise, this is safe to transform. Insert PHI nodes for each operand
10274 // that is variable.
10275 SmallVector<PHINode*, 16> OperandPhis(FixedOperands.size());
10277 bool HasAnyPHIs = false;
10278 for (unsigned i = 0, e = FixedOperands.size(); i != e; ++i) {
10279 if (FixedOperands[i]) continue; // operand doesn't need a phi.
10280 Value *FirstOp = FirstInst->getOperand(i);
10281 PHINode *NewPN = PHINode::Create(FirstOp->getType(),
10282 FirstOp->getName()+".pn");
10283 InsertNewInstBefore(NewPN, PN);
10285 NewPN->reserveOperandSpace(e);
10286 NewPN->addIncoming(FirstOp, PN.getIncomingBlock(0));
10287 OperandPhis[i] = NewPN;
10288 FixedOperands[i] = NewPN;
10293 // Add all operands to the new PHIs.
10295 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
10296 GetElementPtrInst *InGEP =cast<GetElementPtrInst>(PN.getIncomingValue(i));
10297 BasicBlock *InBB = PN.getIncomingBlock(i);
10299 for (unsigned op = 0, e = OperandPhis.size(); op != e; ++op)
10300 if (PHINode *OpPhi = OperandPhis[op])
10301 OpPhi->addIncoming(InGEP->getOperand(op), InBB);
10305 Value *Base = FixedOperands[0];
10306 return GetElementPtrInst::Create(Base, FixedOperands.begin()+1,
10307 FixedOperands.end());
10311 /// isSafeAndProfitableToSinkLoad - Return true if we know that it is safe to
10312 /// sink the load out of the block that defines it. This means that it must be
10313 /// obvious the value of the load is not changed from the point of the load to
10314 /// the end of the block it is in.
10316 /// Finally, it is safe, but not profitable, to sink a load targetting a
10317 /// non-address-taken alloca. Doing so will cause us to not promote the alloca
10319 static bool isSafeAndProfitableToSinkLoad(LoadInst *L) {
10320 BasicBlock::iterator BBI = L, E = L->getParent()->end();
10322 for (++BBI; BBI != E; ++BBI)
10323 if (BBI->mayWriteToMemory())
10326 // Check for non-address taken alloca. If not address-taken already, it isn't
10327 // profitable to do this xform.
10328 if (AllocaInst *AI = dyn_cast<AllocaInst>(L->getOperand(0))) {
10329 bool isAddressTaken = false;
10330 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
10332 if (isa<LoadInst>(UI)) continue;
10333 if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
10334 // If storing TO the alloca, then the address isn't taken.
10335 if (SI->getOperand(1) == AI) continue;
10337 isAddressTaken = true;
10341 if (!isAddressTaken && AI->isStaticAlloca())
10345 // If this load is a load from a GEP with a constant offset from an alloca,
10346 // then we don't want to sink it. In its present form, it will be
10347 // load [constant stack offset]. Sinking it will cause us to have to
10348 // materialize the stack addresses in each predecessor in a register only to
10349 // do a shared load from register in the successor.
10350 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(L->getOperand(0)))
10351 if (AllocaInst *AI = dyn_cast<AllocaInst>(GEP->getOperand(0)))
10352 if (AI->isStaticAlloca() && GEP->hasAllConstantIndices())
10359 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
10360 // operator and they all are only used by the PHI, PHI together their
10361 // inputs, and do the operation once, to the result of the PHI.
10362 Instruction *InstCombiner::FoldPHIArgOpIntoPHI(PHINode &PN) {
10363 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
10365 // Scan the instruction, looking for input operations that can be folded away.
10366 // If all input operands to the phi are the same instruction (e.g. a cast from
10367 // the same type or "+42") we can pull the operation through the PHI, reducing
10368 // code size and simplifying code.
10369 Constant *ConstantOp = 0;
10370 const Type *CastSrcTy = 0;
10371 bool isVolatile = false;
10372 if (isa<CastInst>(FirstInst)) {
10373 CastSrcTy = FirstInst->getOperand(0)->getType();
10374 } else if (isa<BinaryOperator>(FirstInst) || isa<CmpInst>(FirstInst)) {
10375 // Can fold binop, compare or shift here if the RHS is a constant,
10376 // otherwise call FoldPHIArgBinOpIntoPHI.
10377 ConstantOp = dyn_cast<Constant>(FirstInst->getOperand(1));
10378 if (ConstantOp == 0)
10379 return FoldPHIArgBinOpIntoPHI(PN);
10380 } else if (LoadInst *LI = dyn_cast<LoadInst>(FirstInst)) {
10381 isVolatile = LI->isVolatile();
10382 // We can't sink the load if the loaded value could be modified between the
10383 // load and the PHI.
10384 if (LI->getParent() != PN.getIncomingBlock(0) ||
10385 !isSafeAndProfitableToSinkLoad(LI))
10388 // If the PHI is of volatile loads and the load block has multiple
10389 // successors, sinking it would remove a load of the volatile value from
10390 // the path through the other successor.
10392 LI->getParent()->getTerminator()->getNumSuccessors() != 1)
10395 } else if (isa<GetElementPtrInst>(FirstInst)) {
10396 return FoldPHIArgGEPIntoPHI(PN);
10398 return 0; // Cannot fold this operation.
10401 // Check to see if all arguments are the same operation.
10402 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
10403 if (!isa<Instruction>(PN.getIncomingValue(i))) return 0;
10404 Instruction *I = cast<Instruction>(PN.getIncomingValue(i));
10405 if (!I->hasOneUse() || !I->isSameOperationAs(FirstInst))
10408 if (I->getOperand(0)->getType() != CastSrcTy)
10409 return 0; // Cast operation must match.
10410 } else if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
10411 // We can't sink the load if the loaded value could be modified between
10412 // the load and the PHI.
10413 if (LI->isVolatile() != isVolatile ||
10414 LI->getParent() != PN.getIncomingBlock(i) ||
10415 !isSafeAndProfitableToSinkLoad(LI))
10418 // If the PHI is of volatile loads and the load block has multiple
10419 // successors, sinking it would remove a load of the volatile value from
10420 // the path through the other successor.
10422 LI->getParent()->getTerminator()->getNumSuccessors() != 1)
10425 } else if (I->getOperand(1) != ConstantOp) {
10430 // Okay, they are all the same operation. Create a new PHI node of the
10431 // correct type, and PHI together all of the LHS's of the instructions.
10432 PHINode *NewPN = PHINode::Create(FirstInst->getOperand(0)->getType(),
10433 PN.getName()+".in");
10434 NewPN->reserveOperandSpace(PN.getNumOperands()/2);
10436 Value *InVal = FirstInst->getOperand(0);
10437 NewPN->addIncoming(InVal, PN.getIncomingBlock(0));
10439 // Add all operands to the new PHI.
10440 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
10441 Value *NewInVal = cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
10442 if (NewInVal != InVal)
10444 NewPN->addIncoming(NewInVal, PN.getIncomingBlock(i));
10449 // The new PHI unions all of the same values together. This is really
10450 // common, so we handle it intelligently here for compile-time speed.
10454 InsertNewInstBefore(NewPN, PN);
10458 // Insert and return the new operation.
10459 if (CastInst* FirstCI = dyn_cast<CastInst>(FirstInst))
10460 return CastInst::Create(FirstCI->getOpcode(), PhiVal, PN.getType());
10461 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
10462 return BinaryOperator::Create(BinOp->getOpcode(), PhiVal, ConstantOp);
10463 if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst))
10464 return CmpInst::Create(CIOp->getOpcode(), CIOp->getPredicate(),
10465 PhiVal, ConstantOp);
10466 assert(isa<LoadInst>(FirstInst) && "Unknown operation");
10468 // If this was a volatile load that we are merging, make sure to loop through
10469 // and mark all the input loads as non-volatile. If we don't do this, we will
10470 // insert a new volatile load and the old ones will not be deletable.
10472 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i)
10473 cast<LoadInst>(PN.getIncomingValue(i))->setVolatile(false);
10475 return new LoadInst(PhiVal, "", isVolatile);
10478 /// DeadPHICycle - Return true if this PHI node is only used by a PHI node cycle
10480 static bool DeadPHICycle(PHINode *PN,
10481 SmallPtrSet<PHINode*, 16> &PotentiallyDeadPHIs) {
10482 if (PN->use_empty()) return true;
10483 if (!PN->hasOneUse()) return false;
10485 // Remember this node, and if we find the cycle, return.
10486 if (!PotentiallyDeadPHIs.insert(PN))
10489 // Don't scan crazily complex things.
10490 if (PotentiallyDeadPHIs.size() == 16)
10493 if (PHINode *PU = dyn_cast<PHINode>(PN->use_back()))
10494 return DeadPHICycle(PU, PotentiallyDeadPHIs);
10499 /// PHIsEqualValue - Return true if this phi node is always equal to
10500 /// NonPhiInVal. This happens with mutually cyclic phi nodes like:
10501 /// z = some value; x = phi (y, z); y = phi (x, z)
10502 static bool PHIsEqualValue(PHINode *PN, Value *NonPhiInVal,
10503 SmallPtrSet<PHINode*, 16> &ValueEqualPHIs) {
10504 // See if we already saw this PHI node.
10505 if (!ValueEqualPHIs.insert(PN))
10508 // Don't scan crazily complex things.
10509 if (ValueEqualPHIs.size() == 16)
10512 // Scan the operands to see if they are either phi nodes or are equal to
10514 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
10515 Value *Op = PN->getIncomingValue(i);
10516 if (PHINode *OpPN = dyn_cast<PHINode>(Op)) {
10517 if (!PHIsEqualValue(OpPN, NonPhiInVal, ValueEqualPHIs))
10519 } else if (Op != NonPhiInVal)
10527 // PHINode simplification
10529 Instruction *InstCombiner::visitPHINode(PHINode &PN) {
10530 // If LCSSA is around, don't mess with Phi nodes
10531 if (MustPreserveLCSSA) return 0;
10533 if (Value *V = PN.hasConstantValue())
10534 return ReplaceInstUsesWith(PN, V);
10536 // If all PHI operands are the same operation, pull them through the PHI,
10537 // reducing code size.
10538 if (isa<Instruction>(PN.getIncomingValue(0)) &&
10539 isa<Instruction>(PN.getIncomingValue(1)) &&
10540 cast<Instruction>(PN.getIncomingValue(0))->getOpcode() ==
10541 cast<Instruction>(PN.getIncomingValue(1))->getOpcode() &&
10542 // FIXME: The hasOneUse check will fail for PHIs that use the value more
10543 // than themselves more than once.
10544 PN.getIncomingValue(0)->hasOneUse())
10545 if (Instruction *Result = FoldPHIArgOpIntoPHI(PN))
10548 // If this is a trivial cycle in the PHI node graph, remove it. Basically, if
10549 // this PHI only has a single use (a PHI), and if that PHI only has one use (a
10550 // PHI)... break the cycle.
10551 if (PN.hasOneUse()) {
10552 Instruction *PHIUser = cast<Instruction>(PN.use_back());
10553 if (PHINode *PU = dyn_cast<PHINode>(PHIUser)) {
10554 SmallPtrSet<PHINode*, 16> PotentiallyDeadPHIs;
10555 PotentiallyDeadPHIs.insert(&PN);
10556 if (DeadPHICycle(PU, PotentiallyDeadPHIs))
10557 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
10560 // If this phi has a single use, and if that use just computes a value for
10561 // the next iteration of a loop, delete the phi. This occurs with unused
10562 // induction variables, e.g. "for (int j = 0; ; ++j);". Detecting this
10563 // common case here is good because the only other things that catch this
10564 // are induction variable analysis (sometimes) and ADCE, which is only run
10566 if (PHIUser->hasOneUse() &&
10567 (isa<BinaryOperator>(PHIUser) || isa<GetElementPtrInst>(PHIUser)) &&
10568 PHIUser->use_back() == &PN) {
10569 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
10573 // We sometimes end up with phi cycles that non-obviously end up being the
10574 // same value, for example:
10575 // z = some value; x = phi (y, z); y = phi (x, z)
10576 // where the phi nodes don't necessarily need to be in the same block. Do a
10577 // quick check to see if the PHI node only contains a single non-phi value, if
10578 // so, scan to see if the phi cycle is actually equal to that value.
10580 unsigned InValNo = 0, NumOperandVals = PN.getNumIncomingValues();
10581 // Scan for the first non-phi operand.
10582 while (InValNo != NumOperandVals &&
10583 isa<PHINode>(PN.getIncomingValue(InValNo)))
10586 if (InValNo != NumOperandVals) {
10587 Value *NonPhiInVal = PN.getOperand(InValNo);
10589 // Scan the rest of the operands to see if there are any conflicts, if so
10590 // there is no need to recursively scan other phis.
10591 for (++InValNo; InValNo != NumOperandVals; ++InValNo) {
10592 Value *OpVal = PN.getIncomingValue(InValNo);
10593 if (OpVal != NonPhiInVal && !isa<PHINode>(OpVal))
10597 // If we scanned over all operands, then we have one unique value plus
10598 // phi values. Scan PHI nodes to see if they all merge in each other or
10600 if (InValNo == NumOperandVals) {
10601 SmallPtrSet<PHINode*, 16> ValueEqualPHIs;
10602 if (PHIsEqualValue(&PN, NonPhiInVal, ValueEqualPHIs))
10603 return ReplaceInstUsesWith(PN, NonPhiInVal);
10610 static Value *InsertCastToIntPtrTy(Value *V, const Type *DTy,
10611 Instruction *InsertPoint,
10612 InstCombiner *IC) {
10613 unsigned PtrSize = DTy->getPrimitiveSizeInBits();
10614 unsigned VTySize = V->getType()->getPrimitiveSizeInBits();
10615 // We must cast correctly to the pointer type. Ensure that we
10616 // sign extend the integer value if it is smaller as this is
10617 // used for address computation.
10618 Instruction::CastOps opcode =
10619 (VTySize < PtrSize ? Instruction::SExt :
10620 (VTySize == PtrSize ? Instruction::BitCast : Instruction::Trunc));
10621 return IC->InsertCastBefore(opcode, V, DTy, *InsertPoint);
10625 Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) {
10626 Value *PtrOp = GEP.getOperand(0);
10627 // Is it 'getelementptr %P, i32 0' or 'getelementptr %P'
10628 // If so, eliminate the noop.
10629 if (GEP.getNumOperands() == 1)
10630 return ReplaceInstUsesWith(GEP, PtrOp);
10632 if (isa<UndefValue>(GEP.getOperand(0)))
10633 return ReplaceInstUsesWith(GEP, UndefValue::get(GEP.getType()));
10635 bool HasZeroPointerIndex = false;
10636 if (Constant *C = dyn_cast<Constant>(GEP.getOperand(1)))
10637 HasZeroPointerIndex = C->isNullValue();
10639 if (GEP.getNumOperands() == 2 && HasZeroPointerIndex)
10640 return ReplaceInstUsesWith(GEP, PtrOp);
10642 // Eliminate unneeded casts for indices.
10643 bool MadeChange = false;
10645 gep_type_iterator GTI = gep_type_begin(GEP);
10646 for (User::op_iterator i = GEP.op_begin() + 1, e = GEP.op_end();
10647 i != e; ++i, ++GTI) {
10648 if (isa<SequentialType>(*GTI)) {
10649 if (CastInst *CI = dyn_cast<CastInst>(*i)) {
10650 if (CI->getOpcode() == Instruction::ZExt ||
10651 CI->getOpcode() == Instruction::SExt) {
10652 const Type *SrcTy = CI->getOperand(0)->getType();
10653 // We can eliminate a cast from i32 to i64 iff the target
10654 // is a 32-bit pointer target.
10655 if (SrcTy->getPrimitiveSizeInBits() >= TD->getPointerSizeInBits()) {
10657 *i = CI->getOperand(0);
10661 // If we are using a wider index than needed for this platform, shrink it
10662 // to what we need. If narrower, sign-extend it to what we need.
10663 // If the incoming value needs a cast instruction,
10664 // insert it. This explicit cast can make subsequent optimizations more
10667 if (TD->getTypeSizeInBits(Op->getType()) > TD->getPointerSizeInBits()) {
10668 if (Constant *C = dyn_cast<Constant>(Op)) {
10669 *i = ConstantExpr::getTrunc(C, TD->getIntPtrType());
10672 Op = InsertCastBefore(Instruction::Trunc, Op, TD->getIntPtrType(),
10677 } else if (TD->getTypeSizeInBits(Op->getType()) < TD->getPointerSizeInBits()) {
10678 if (Constant *C = dyn_cast<Constant>(Op)) {
10679 *i = ConstantExpr::getSExt(C, TD->getIntPtrType());
10682 Op = InsertCastBefore(Instruction::SExt, Op, TD->getIntPtrType(),
10690 if (MadeChange) return &GEP;
10692 // Combine Indices - If the source pointer to this getelementptr instruction
10693 // is a getelementptr instruction, combine the indices of the two
10694 // getelementptr instructions into a single instruction.
10696 SmallVector<Value*, 8> SrcGEPOperands;
10697 if (User *Src = dyn_castGetElementPtr(PtrOp))
10698 SrcGEPOperands.append(Src->op_begin(), Src->op_end());
10700 if (!SrcGEPOperands.empty()) {
10701 // Note that if our source is a gep chain itself that we wait for that
10702 // chain to be resolved before we perform this transformation. This
10703 // avoids us creating a TON of code in some cases.
10705 if (isa<GetElementPtrInst>(SrcGEPOperands[0]) &&
10706 cast<Instruction>(SrcGEPOperands[0])->getNumOperands() == 2)
10707 return 0; // Wait until our source is folded to completion.
10709 SmallVector<Value*, 8> Indices;
10711 // Find out whether the last index in the source GEP is a sequential idx.
10712 bool EndsWithSequential = false;
10713 for (gep_type_iterator I = gep_type_begin(*cast<User>(PtrOp)),
10714 E = gep_type_end(*cast<User>(PtrOp)); I != E; ++I)
10715 EndsWithSequential = !isa<StructType>(*I);
10717 // Can we combine the two pointer arithmetics offsets?
10718 if (EndsWithSequential) {
10719 // Replace: gep (gep %P, long B), long A, ...
10720 // With: T = long A+B; gep %P, T, ...
10722 Value *Sum, *SO1 = SrcGEPOperands.back(), *GO1 = GEP.getOperand(1);
10723 if (SO1 == Constant::getNullValue(SO1->getType())) {
10725 } else if (GO1 == Constant::getNullValue(GO1->getType())) {
10728 // If they aren't the same type, convert both to an integer of the
10729 // target's pointer size.
10730 if (SO1->getType() != GO1->getType()) {
10731 if (Constant *SO1C = dyn_cast<Constant>(SO1)) {
10732 SO1 = ConstantExpr::getIntegerCast(SO1C, GO1->getType(), true);
10733 } else if (Constant *GO1C = dyn_cast<Constant>(GO1)) {
10734 GO1 = ConstantExpr::getIntegerCast(GO1C, SO1->getType(), true);
10736 unsigned PS = TD->getPointerSizeInBits();
10737 if (TD->getTypeSizeInBits(SO1->getType()) == PS) {
10738 // Convert GO1 to SO1's type.
10739 GO1 = InsertCastToIntPtrTy(GO1, SO1->getType(), &GEP, this);
10741 } else if (TD->getTypeSizeInBits(GO1->getType()) == PS) {
10742 // Convert SO1 to GO1's type.
10743 SO1 = InsertCastToIntPtrTy(SO1, GO1->getType(), &GEP, this);
10745 const Type *PT = TD->getIntPtrType();
10746 SO1 = InsertCastToIntPtrTy(SO1, PT, &GEP, this);
10747 GO1 = InsertCastToIntPtrTy(GO1, PT, &GEP, this);
10751 if (isa<Constant>(SO1) && isa<Constant>(GO1))
10752 Sum = ConstantExpr::getAdd(cast<Constant>(SO1), cast<Constant>(GO1));
10754 Sum = BinaryOperator::CreateAdd(SO1, GO1, PtrOp->getName()+".sum");
10755 InsertNewInstBefore(cast<Instruction>(Sum), GEP);
10759 // Recycle the GEP we already have if possible.
10760 if (SrcGEPOperands.size() == 2) {
10761 GEP.setOperand(0, SrcGEPOperands[0]);
10762 GEP.setOperand(1, Sum);
10765 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
10766 SrcGEPOperands.end()-1);
10767 Indices.push_back(Sum);
10768 Indices.insert(Indices.end(), GEP.op_begin()+2, GEP.op_end());
10770 } else if (isa<Constant>(*GEP.idx_begin()) &&
10771 cast<Constant>(*GEP.idx_begin())->isNullValue() &&
10772 SrcGEPOperands.size() != 1) {
10773 // Otherwise we can do the fold if the first index of the GEP is a zero
10774 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
10775 SrcGEPOperands.end());
10776 Indices.insert(Indices.end(), GEP.idx_begin()+1, GEP.idx_end());
10779 if (!Indices.empty())
10780 return GetElementPtrInst::Create(SrcGEPOperands[0], Indices.begin(),
10781 Indices.end(), GEP.getName());
10783 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(PtrOp)) {
10784 // GEP of global variable. If all of the indices for this GEP are
10785 // constants, we can promote this to a constexpr instead of an instruction.
10787 // Scan for nonconstants...
10788 SmallVector<Constant*, 8> Indices;
10789 User::op_iterator I = GEP.idx_begin(), E = GEP.idx_end();
10790 for (; I != E && isa<Constant>(*I); ++I)
10791 Indices.push_back(cast<Constant>(*I));
10793 if (I == E) { // If they are all constants...
10794 Constant *CE = ConstantExpr::getGetElementPtr(GV,
10795 &Indices[0],Indices.size());
10797 // Replace all uses of the GEP with the new constexpr...
10798 return ReplaceInstUsesWith(GEP, CE);
10800 } else if (Value *X = getBitCastOperand(PtrOp)) { // Is the operand a cast?
10801 if (!isa<PointerType>(X->getType())) {
10802 // Not interesting. Source pointer must be a cast from pointer.
10803 } else if (HasZeroPointerIndex) {
10804 // transform: GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ...
10805 // into : GEP [10 x i8]* X, i32 0, ...
10807 // Likewise, transform: GEP (bitcast i8* X to [0 x i8]*), i32 0, ...
10808 // into : GEP i8* X, ...
10810 // This occurs when the program declares an array extern like "int X[];"
10811 const PointerType *CPTy = cast<PointerType>(PtrOp->getType());
10812 const PointerType *XTy = cast<PointerType>(X->getType());
10813 if (const ArrayType *CATy =
10814 dyn_cast<ArrayType>(CPTy->getElementType())) {
10815 // GEP (bitcast i8* X to [0 x i8]*), i32 0, ... ?
10816 if (CATy->getElementType() == XTy->getElementType()) {
10817 // -> GEP i8* X, ...
10818 SmallVector<Value*, 8> Indices(GEP.idx_begin()+1, GEP.idx_end());
10819 return GetElementPtrInst::Create(X, Indices.begin(), Indices.end(),
10821 } else if (const ArrayType *XATy =
10822 dyn_cast<ArrayType>(XTy->getElementType())) {
10823 // GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ... ?
10824 if (CATy->getElementType() == XATy->getElementType()) {
10825 // -> GEP [10 x i8]* X, i32 0, ...
10826 // At this point, we know that the cast source type is a pointer
10827 // to an array of the same type as the destination pointer
10828 // array. Because the array type is never stepped over (there
10829 // is a leading zero) we can fold the cast into this GEP.
10830 GEP.setOperand(0, X);
10835 } else if (GEP.getNumOperands() == 2) {
10836 // Transform things like:
10837 // %t = getelementptr i32* bitcast ([2 x i32]* %str to i32*), i32 %V
10838 // into: %t1 = getelementptr [2 x i32]* %str, i32 0, i32 %V; bitcast
10839 const Type *SrcElTy = cast<PointerType>(X->getType())->getElementType();
10840 const Type *ResElTy=cast<PointerType>(PtrOp->getType())->getElementType();
10841 if (isa<ArrayType>(SrcElTy) &&
10842 TD->getTypePaddedSize(cast<ArrayType>(SrcElTy)->getElementType()) ==
10843 TD->getTypePaddedSize(ResElTy)) {
10845 Idx[0] = Constant::getNullValue(Type::Int32Ty);
10846 Idx[1] = GEP.getOperand(1);
10847 Value *V = InsertNewInstBefore(
10848 GetElementPtrInst::Create(X, Idx, Idx + 2, GEP.getName()), GEP);
10849 // V and GEP are both pointer types --> BitCast
10850 return new BitCastInst(V, GEP.getType());
10853 // Transform things like:
10854 // getelementptr i8* bitcast ([100 x double]* X to i8*), i32 %tmp
10855 // (where tmp = 8*tmp2) into:
10856 // getelementptr [100 x double]* %arr, i32 0, i32 %tmp2; bitcast
10858 if (isa<ArrayType>(SrcElTy) && ResElTy == Type::Int8Ty) {
10859 uint64_t ArrayEltSize =
10860 TD->getTypePaddedSize(cast<ArrayType>(SrcElTy)->getElementType());
10862 // Check to see if "tmp" is a scale by a multiple of ArrayEltSize. We
10863 // allow either a mul, shift, or constant here.
10865 ConstantInt *Scale = 0;
10866 if (ArrayEltSize == 1) {
10867 NewIdx = GEP.getOperand(1);
10868 Scale = ConstantInt::get(NewIdx->getType(), 1);
10869 } else if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP.getOperand(1))) {
10870 NewIdx = ConstantInt::get(CI->getType(), 1);
10872 } else if (Instruction *Inst =dyn_cast<Instruction>(GEP.getOperand(1))){
10873 if (Inst->getOpcode() == Instruction::Shl &&
10874 isa<ConstantInt>(Inst->getOperand(1))) {
10875 ConstantInt *ShAmt = cast<ConstantInt>(Inst->getOperand(1));
10876 uint32_t ShAmtVal = ShAmt->getLimitedValue(64);
10877 Scale = ConstantInt::get(Inst->getType(), 1ULL << ShAmtVal);
10878 NewIdx = Inst->getOperand(0);
10879 } else if (Inst->getOpcode() == Instruction::Mul &&
10880 isa<ConstantInt>(Inst->getOperand(1))) {
10881 Scale = cast<ConstantInt>(Inst->getOperand(1));
10882 NewIdx = Inst->getOperand(0);
10886 // If the index will be to exactly the right offset with the scale taken
10887 // out, perform the transformation. Note, we don't know whether Scale is
10888 // signed or not. We'll use unsigned version of division/modulo
10889 // operation after making sure Scale doesn't have the sign bit set.
10890 if (ArrayEltSize && Scale && Scale->getSExtValue() >= 0LL &&
10891 Scale->getZExtValue() % ArrayEltSize == 0) {
10892 Scale = ConstantInt::get(Scale->getType(),
10893 Scale->getZExtValue() / ArrayEltSize);
10894 if (Scale->getZExtValue() != 1) {
10895 Constant *C = ConstantExpr::getIntegerCast(Scale, NewIdx->getType(),
10897 Instruction *Sc = BinaryOperator::CreateMul(NewIdx, C, "idxscale");
10898 NewIdx = InsertNewInstBefore(Sc, GEP);
10901 // Insert the new GEP instruction.
10903 Idx[0] = Constant::getNullValue(Type::Int32Ty);
10905 Instruction *NewGEP =
10906 GetElementPtrInst::Create(X, Idx, Idx + 2, GEP.getName());
10907 NewGEP = InsertNewInstBefore(NewGEP, GEP);
10908 // The NewGEP must be pointer typed, so must the old one -> BitCast
10909 return new BitCastInst(NewGEP, GEP.getType());
10915 /// See if we can simplify:
10916 /// X = bitcast A to B*
10917 /// Y = gep X, <...constant indices...>
10918 /// into a gep of the original struct. This is important for SROA and alias
10919 /// analysis of unions. If "A" is also a bitcast, wait for A/X to be merged.
10920 if (BitCastInst *BCI = dyn_cast<BitCastInst>(PtrOp)) {
10921 if (!isa<BitCastInst>(BCI->getOperand(0)) && GEP.hasAllConstantIndices()) {
10922 // Determine how much the GEP moves the pointer. We are guaranteed to get
10923 // a constant back from EmitGEPOffset.
10924 ConstantInt *OffsetV = cast<ConstantInt>(EmitGEPOffset(&GEP, GEP, *this));
10925 int64_t Offset = OffsetV->getSExtValue();
10927 // If this GEP instruction doesn't move the pointer, just replace the GEP
10928 // with a bitcast of the real input to the dest type.
10930 // If the bitcast is of an allocation, and the allocation will be
10931 // converted to match the type of the cast, don't touch this.
10932 if (isa<AllocationInst>(BCI->getOperand(0))) {
10933 // See if the bitcast simplifies, if so, don't nuke this GEP yet.
10934 if (Instruction *I = visitBitCast(*BCI)) {
10937 BCI->getParent()->getInstList().insert(BCI, I);
10938 ReplaceInstUsesWith(*BCI, I);
10943 return new BitCastInst(BCI->getOperand(0), GEP.getType());
10946 // Otherwise, if the offset is non-zero, we need to find out if there is a
10947 // field at Offset in 'A's type. If so, we can pull the cast through the
10949 SmallVector<Value*, 8> NewIndices;
10951 cast<PointerType>(BCI->getOperand(0)->getType())->getElementType();
10952 if (FindElementAtOffset(InTy, Offset, NewIndices, TD)) {
10953 Instruction *NGEP =
10954 GetElementPtrInst::Create(BCI->getOperand(0), NewIndices.begin(),
10956 if (NGEP->getType() == GEP.getType()) return NGEP;
10957 InsertNewInstBefore(NGEP, GEP);
10958 NGEP->takeName(&GEP);
10959 return new BitCastInst(NGEP, GEP.getType());
10967 Instruction *InstCombiner::visitAllocationInst(AllocationInst &AI) {
10968 // Convert: malloc Ty, C - where C is a constant != 1 into: malloc [C x Ty], 1
10969 if (AI.isArrayAllocation()) { // Check C != 1
10970 if (const ConstantInt *C = dyn_cast<ConstantInt>(AI.getArraySize())) {
10971 const Type *NewTy =
10972 ArrayType::get(AI.getAllocatedType(), C->getZExtValue());
10973 AllocationInst *New = 0;
10975 // Create and insert the replacement instruction...
10976 if (isa<MallocInst>(AI))
10977 New = new MallocInst(NewTy, 0, AI.getAlignment(), AI.getName());
10979 assert(isa<AllocaInst>(AI) && "Unknown type of allocation inst!");
10980 New = new AllocaInst(NewTy, 0, AI.getAlignment(), AI.getName());
10983 InsertNewInstBefore(New, AI);
10985 // Scan to the end of the allocation instructions, to skip over a block of
10986 // allocas if possible...also skip interleaved debug info
10988 BasicBlock::iterator It = New;
10989 while (isa<AllocationInst>(*It) || isa<DbgInfoIntrinsic>(*It)) ++It;
10991 // Now that I is pointing to the first non-allocation-inst in the block,
10992 // insert our getelementptr instruction...
10994 Value *NullIdx = Constant::getNullValue(Type::Int32Ty);
10998 Value *V = GetElementPtrInst::Create(New, Idx, Idx + 2,
10999 New->getName()+".sub", It);
11001 // Now make everything use the getelementptr instead of the original
11003 return ReplaceInstUsesWith(AI, V);
11004 } else if (isa<UndefValue>(AI.getArraySize())) {
11005 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
11009 if (isa<AllocaInst>(AI) && AI.getAllocatedType()->isSized()) {
11010 // If alloca'ing a zero byte object, replace the alloca with a null pointer.
11011 // Note that we only do this for alloca's, because malloc should allocate
11012 // and return a unique pointer, even for a zero byte allocation.
11013 if (TD->getTypePaddedSize(AI.getAllocatedType()) == 0)
11014 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
11016 // If the alignment is 0 (unspecified), assign it the preferred alignment.
11017 if (AI.getAlignment() == 0)
11018 AI.setAlignment(TD->getPrefTypeAlignment(AI.getAllocatedType()));
11024 Instruction *InstCombiner::visitFreeInst(FreeInst &FI) {
11025 Value *Op = FI.getOperand(0);
11027 // free undef -> unreachable.
11028 if (isa<UndefValue>(Op)) {
11029 // Insert a new store to null because we cannot modify the CFG here.
11030 new StoreInst(ConstantInt::getTrue(),
11031 UndefValue::get(PointerType::getUnqual(Type::Int1Ty)), &FI);
11032 return EraseInstFromFunction(FI);
11035 // If we have 'free null' delete the instruction. This can happen in stl code
11036 // when lots of inlining happens.
11037 if (isa<ConstantPointerNull>(Op))
11038 return EraseInstFromFunction(FI);
11040 // Change free <ty>* (cast <ty2>* X to <ty>*) into free <ty2>* X
11041 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op)) {
11042 FI.setOperand(0, CI->getOperand(0));
11046 // Change free (gep X, 0,0,0,0) into free(X)
11047 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op)) {
11048 if (GEPI->hasAllZeroIndices()) {
11049 AddToWorkList(GEPI);
11050 FI.setOperand(0, GEPI->getOperand(0));
11055 // Change free(malloc) into nothing, if the malloc has a single use.
11056 if (MallocInst *MI = dyn_cast<MallocInst>(Op))
11057 if (MI->hasOneUse()) {
11058 EraseInstFromFunction(FI);
11059 return EraseInstFromFunction(*MI);
11066 /// InstCombineLoadCast - Fold 'load (cast P)' -> cast (load P)' when possible.
11067 static Instruction *InstCombineLoadCast(InstCombiner &IC, LoadInst &LI,
11068 const TargetData *TD) {
11069 User *CI = cast<User>(LI.getOperand(0));
11070 Value *CastOp = CI->getOperand(0);
11072 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(CI)) {
11073 // Instead of loading constant c string, use corresponding integer value
11074 // directly if string length is small enough.
11076 if (GetConstantStringInfo(CE->getOperand(0), Str) && !Str.empty()) {
11077 unsigned len = Str.length();
11078 const Type *Ty = cast<PointerType>(CE->getType())->getElementType();
11079 unsigned numBits = Ty->getPrimitiveSizeInBits();
11080 // Replace LI with immediate integer store.
11081 if ((numBits >> 3) == len + 1) {
11082 APInt StrVal(numBits, 0);
11083 APInt SingleChar(numBits, 0);
11084 if (TD->isLittleEndian()) {
11085 for (signed i = len-1; i >= 0; i--) {
11086 SingleChar = (uint64_t) Str[i] & UCHAR_MAX;
11087 StrVal = (StrVal << 8) | SingleChar;
11090 for (unsigned i = 0; i < len; i++) {
11091 SingleChar = (uint64_t) Str[i] & UCHAR_MAX;
11092 StrVal = (StrVal << 8) | SingleChar;
11094 // Append NULL at the end.
11096 StrVal = (StrVal << 8) | SingleChar;
11098 Value *NL = ConstantInt::get(StrVal);
11099 return IC.ReplaceInstUsesWith(LI, NL);
11104 const PointerType *DestTy = cast<PointerType>(CI->getType());
11105 const Type *DestPTy = DestTy->getElementType();
11106 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
11108 // If the address spaces don't match, don't eliminate the cast.
11109 if (DestTy->getAddressSpace() != SrcTy->getAddressSpace())
11112 const Type *SrcPTy = SrcTy->getElementType();
11114 if (DestPTy->isInteger() || isa<PointerType>(DestPTy) ||
11115 isa<VectorType>(DestPTy)) {
11116 // If the source is an array, the code below will not succeed. Check to
11117 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
11119 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
11120 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
11121 if (ASrcTy->getNumElements() != 0) {
11123 Idxs[0] = Idxs[1] = Constant::getNullValue(Type::Int32Ty);
11124 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs, 2);
11125 SrcTy = cast<PointerType>(CastOp->getType());
11126 SrcPTy = SrcTy->getElementType();
11129 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy) ||
11130 isa<VectorType>(SrcPTy)) &&
11131 // Do not allow turning this into a load of an integer, which is then
11132 // casted to a pointer, this pessimizes pointer analysis a lot.
11133 (isa<PointerType>(SrcPTy) == isa<PointerType>(LI.getType())) &&
11134 IC.getTargetData().getTypeSizeInBits(SrcPTy) ==
11135 IC.getTargetData().getTypeSizeInBits(DestPTy)) {
11137 // Okay, we are casting from one integer or pointer type to another of
11138 // the same size. Instead of casting the pointer before the load, cast
11139 // the result of the loaded value.
11140 Value *NewLoad = IC.InsertNewInstBefore(new LoadInst(CastOp,
11142 LI.isVolatile()),LI);
11143 // Now cast the result of the load.
11144 return new BitCastInst(NewLoad, LI.getType());
11151 /// isSafeToLoadUnconditionally - Return true if we know that executing a load
11152 /// from this value cannot trap. If it is not obviously safe to load from the
11153 /// specified pointer, we do a quick local scan of the basic block containing
11154 /// ScanFrom, to determine if the address is already accessed.
11155 static bool isSafeToLoadUnconditionally(Value *V, Instruction *ScanFrom) {
11156 // If it is an alloca it is always safe to load from.
11157 if (isa<AllocaInst>(V)) return true;
11159 // If it is a global variable it is mostly safe to load from.
11160 if (const GlobalValue *GV = dyn_cast<GlobalVariable>(V))
11161 // Don't try to evaluate aliases. External weak GV can be null.
11162 return !isa<GlobalAlias>(GV) && !GV->hasExternalWeakLinkage();
11164 // Otherwise, be a little bit agressive by scanning the local block where we
11165 // want to check to see if the pointer is already being loaded or stored
11166 // from/to. If so, the previous load or store would have already trapped,
11167 // so there is no harm doing an extra load (also, CSE will later eliminate
11168 // the load entirely).
11169 BasicBlock::iterator BBI = ScanFrom, E = ScanFrom->getParent()->begin();
11174 // If we see a free or a call (which might do a free) the pointer could be
11176 if (isa<FreeInst>(BBI) ||
11177 (isa<CallInst>(BBI) && !isa<DbgInfoIntrinsic>(BBI)))
11180 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
11181 if (LI->getOperand(0) == V) return true;
11182 } else if (StoreInst *SI = dyn_cast<StoreInst>(BBI)) {
11183 if (SI->getOperand(1) == V) return true;
11190 Instruction *InstCombiner::visitLoadInst(LoadInst &LI) {
11191 Value *Op = LI.getOperand(0);
11193 // Attempt to improve the alignment.
11194 unsigned KnownAlign =
11195 GetOrEnforceKnownAlignment(Op, TD->getPrefTypeAlignment(LI.getType()));
11197 (LI.getAlignment() == 0 ? TD->getABITypeAlignment(LI.getType()) :
11198 LI.getAlignment()))
11199 LI.setAlignment(KnownAlign);
11201 // load (cast X) --> cast (load X) iff safe
11202 if (isa<CastInst>(Op))
11203 if (Instruction *Res = InstCombineLoadCast(*this, LI, TD))
11206 // None of the following transforms are legal for volatile loads.
11207 if (LI.isVolatile()) return 0;
11209 // Do really simple store-to-load forwarding and load CSE, to catch cases
11210 // where there are several consequtive memory accesses to the same location,
11211 // separated by a few arithmetic operations.
11212 BasicBlock::iterator BBI = &LI;
11213 if (Value *AvailableVal = FindAvailableLoadedValue(Op, LI.getParent(), BBI,6))
11214 return ReplaceInstUsesWith(LI, AvailableVal);
11216 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op)) {
11217 const Value *GEPI0 = GEPI->getOperand(0);
11218 // TODO: Consider a target hook for valid address spaces for this xform.
11219 if (isa<ConstantPointerNull>(GEPI0) &&
11220 cast<PointerType>(GEPI0->getType())->getAddressSpace() == 0) {
11221 // Insert a new store to null instruction before the load to indicate
11222 // that this code is not reachable. We do this instead of inserting
11223 // an unreachable instruction directly because we cannot modify the
11225 new StoreInst(UndefValue::get(LI.getType()),
11226 Constant::getNullValue(Op->getType()), &LI);
11227 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
11231 if (Constant *C = dyn_cast<Constant>(Op)) {
11232 // load null/undef -> undef
11233 // TODO: Consider a target hook for valid address spaces for this xform.
11234 if (isa<UndefValue>(C) || (C->isNullValue() &&
11235 cast<PointerType>(Op->getType())->getAddressSpace() == 0)) {
11236 // Insert a new store to null instruction before the load to indicate that
11237 // this code is not reachable. We do this instead of inserting an
11238 // unreachable instruction directly because we cannot modify the CFG.
11239 new StoreInst(UndefValue::get(LI.getType()),
11240 Constant::getNullValue(Op->getType()), &LI);
11241 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
11244 // Instcombine load (constant global) into the value loaded.
11245 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Op))
11246 if (GV->isConstant() && GV->hasDefinitiveInitializer())
11247 return ReplaceInstUsesWith(LI, GV->getInitializer());
11249 // Instcombine load (constantexpr_GEP global, 0, ...) into the value loaded.
11250 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Op)) {
11251 if (CE->getOpcode() == Instruction::GetElementPtr) {
11252 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(CE->getOperand(0)))
11253 if (GV->isConstant() && GV->hasDefinitiveInitializer())
11255 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE))
11256 return ReplaceInstUsesWith(LI, V);
11257 if (CE->getOperand(0)->isNullValue()) {
11258 // Insert a new store to null instruction before the load to indicate
11259 // that this code is not reachable. We do this instead of inserting
11260 // an unreachable instruction directly because we cannot modify the
11262 new StoreInst(UndefValue::get(LI.getType()),
11263 Constant::getNullValue(Op->getType()), &LI);
11264 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
11267 } else if (CE->isCast()) {
11268 if (Instruction *Res = InstCombineLoadCast(*this, LI, TD))
11274 // If this load comes from anywhere in a constant global, and if the global
11275 // is all undef or zero, we know what it loads.
11276 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Op->getUnderlyingObject())){
11277 if (GV->isConstant() && GV->hasDefinitiveInitializer()) {
11278 if (GV->getInitializer()->isNullValue())
11279 return ReplaceInstUsesWith(LI, Constant::getNullValue(LI.getType()));
11280 else if (isa<UndefValue>(GV->getInitializer()))
11281 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
11285 if (Op->hasOneUse()) {
11286 // Change select and PHI nodes to select values instead of addresses: this
11287 // helps alias analysis out a lot, allows many others simplifications, and
11288 // exposes redundancy in the code.
11290 // Note that we cannot do the transformation unless we know that the
11291 // introduced loads cannot trap! Something like this is valid as long as
11292 // the condition is always false: load (select bool %C, int* null, int* %G),
11293 // but it would not be valid if we transformed it to load from null
11294 // unconditionally.
11296 if (SelectInst *SI = dyn_cast<SelectInst>(Op)) {
11297 // load (select (Cond, &V1, &V2)) --> select(Cond, load &V1, load &V2).
11298 if (isSafeToLoadUnconditionally(SI->getOperand(1), SI) &&
11299 isSafeToLoadUnconditionally(SI->getOperand(2), SI)) {
11300 Value *V1 = InsertNewInstBefore(new LoadInst(SI->getOperand(1),
11301 SI->getOperand(1)->getName()+".val"), LI);
11302 Value *V2 = InsertNewInstBefore(new LoadInst(SI->getOperand(2),
11303 SI->getOperand(2)->getName()+".val"), LI);
11304 return SelectInst::Create(SI->getCondition(), V1, V2);
11307 // load (select (cond, null, P)) -> load P
11308 if (Constant *C = dyn_cast<Constant>(SI->getOperand(1)))
11309 if (C->isNullValue()) {
11310 LI.setOperand(0, SI->getOperand(2));
11314 // load (select (cond, P, null)) -> load P
11315 if (Constant *C = dyn_cast<Constant>(SI->getOperand(2)))
11316 if (C->isNullValue()) {
11317 LI.setOperand(0, SI->getOperand(1));
11325 /// InstCombineStoreToCast - Fold store V, (cast P) -> store (cast V), P
11326 /// when possible. This makes it generally easy to do alias analysis and/or
11327 /// SROA/mem2reg of the memory object.
11328 static Instruction *InstCombineStoreToCast(InstCombiner &IC, StoreInst &SI) {
11329 User *CI = cast<User>(SI.getOperand(1));
11330 Value *CastOp = CI->getOperand(0);
11332 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
11333 const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType());
11334 if (SrcTy == 0) return 0;
11336 const Type *SrcPTy = SrcTy->getElementType();
11338 if (!DestPTy->isInteger() && !isa<PointerType>(DestPTy))
11341 /// NewGEPIndices - If SrcPTy is an aggregate type, we can emit a "noop gep"
11342 /// to its first element. This allows us to handle things like:
11343 /// store i32 xxx, (bitcast {foo*, float}* %P to i32*)
11344 /// on 32-bit hosts.
11345 SmallVector<Value*, 4> NewGEPIndices;
11347 // If the source is an array, the code below will not succeed. Check to
11348 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
11350 if (isa<ArrayType>(SrcPTy) || isa<StructType>(SrcPTy)) {
11351 // Index through pointer.
11352 Constant *Zero = Constant::getNullValue(Type::Int32Ty);
11353 NewGEPIndices.push_back(Zero);
11356 if (const StructType *STy = dyn_cast<StructType>(SrcPTy)) {
11357 if (!STy->getNumElements()) /* Struct can be empty {} */
11359 NewGEPIndices.push_back(Zero);
11360 SrcPTy = STy->getElementType(0);
11361 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(SrcPTy)) {
11362 NewGEPIndices.push_back(Zero);
11363 SrcPTy = ATy->getElementType();
11369 SrcTy = PointerType::get(SrcPTy, SrcTy->getAddressSpace());
11372 if (!SrcPTy->isInteger() && !isa<PointerType>(SrcPTy))
11375 // If the pointers point into different address spaces or if they point to
11376 // values with different sizes, we can't do the transformation.
11377 if (SrcTy->getAddressSpace() !=
11378 cast<PointerType>(CI->getType())->getAddressSpace() ||
11379 IC.getTargetData().getTypeSizeInBits(SrcPTy) !=
11380 IC.getTargetData().getTypeSizeInBits(DestPTy))
11383 // Okay, we are casting from one integer or pointer type to another of
11384 // the same size. Instead of casting the pointer before
11385 // the store, cast the value to be stored.
11387 Value *SIOp0 = SI.getOperand(0);
11388 Instruction::CastOps opcode = Instruction::BitCast;
11389 const Type* CastSrcTy = SIOp0->getType();
11390 const Type* CastDstTy = SrcPTy;
11391 if (isa<PointerType>(CastDstTy)) {
11392 if (CastSrcTy->isInteger())
11393 opcode = Instruction::IntToPtr;
11394 } else if (isa<IntegerType>(CastDstTy)) {
11395 if (isa<PointerType>(SIOp0->getType()))
11396 opcode = Instruction::PtrToInt;
11399 // SIOp0 is a pointer to aggregate and this is a store to the first field,
11400 // emit a GEP to index into its first field.
11401 if (!NewGEPIndices.empty()) {
11402 if (Constant *C = dyn_cast<Constant>(CastOp))
11403 CastOp = ConstantExpr::getGetElementPtr(C, &NewGEPIndices[0],
11404 NewGEPIndices.size());
11406 CastOp = IC.InsertNewInstBefore(
11407 GetElementPtrInst::Create(CastOp, NewGEPIndices.begin(),
11408 NewGEPIndices.end()), SI);
11411 if (Constant *C = dyn_cast<Constant>(SIOp0))
11412 NewCast = ConstantExpr::getCast(opcode, C, CastDstTy);
11414 NewCast = IC.InsertNewInstBefore(
11415 CastInst::Create(opcode, SIOp0, CastDstTy, SIOp0->getName()+".c"),
11417 return new StoreInst(NewCast, CastOp);
11420 /// equivalentAddressValues - Test if A and B will obviously have the same
11421 /// value. This includes recognizing that %t0 and %t1 will have the same
11422 /// value in code like this:
11423 /// %t0 = getelementptr \@a, 0, 3
11424 /// store i32 0, i32* %t0
11425 /// %t1 = getelementptr \@a, 0, 3
11426 /// %t2 = load i32* %t1
11428 static bool equivalentAddressValues(Value *A, Value *B) {
11429 // Test if the values are trivially equivalent.
11430 if (A == B) return true;
11432 // Test if the values come form identical arithmetic instructions.
11433 if (isa<BinaryOperator>(A) ||
11434 isa<CastInst>(A) ||
11436 isa<GetElementPtrInst>(A))
11437 if (Instruction *BI = dyn_cast<Instruction>(B))
11438 if (cast<Instruction>(A)->isIdenticalTo(BI))
11441 // Otherwise they may not be equivalent.
11445 // If this instruction has two uses, one of which is a llvm.dbg.declare,
11446 // return the llvm.dbg.declare.
11447 DbgDeclareInst *InstCombiner::hasOneUsePlusDeclare(Value *V) {
11448 if (!V->hasNUses(2))
11450 for (Value::use_iterator UI = V->use_begin(), E = V->use_end();
11452 if (DbgDeclareInst *DI = dyn_cast<DbgDeclareInst>(UI))
11454 if (isa<BitCastInst>(UI) && UI->hasOneUse()) {
11455 if (DbgDeclareInst *DI = dyn_cast<DbgDeclareInst>(UI->use_begin()))
11462 Instruction *InstCombiner::visitStoreInst(StoreInst &SI) {
11463 Value *Val = SI.getOperand(0);
11464 Value *Ptr = SI.getOperand(1);
11466 if (isa<UndefValue>(Ptr)) { // store X, undef -> noop (even if volatile)
11467 EraseInstFromFunction(SI);
11472 // If the RHS is an alloca with a single use, zapify the store, making the
11474 // If the RHS is an alloca with a two uses, the other one being a
11475 // llvm.dbg.declare, zapify the store and the declare, making the
11476 // alloca dead. We must do this to prevent declare's from affecting
11478 if (!SI.isVolatile()) {
11479 if (Ptr->hasOneUse()) {
11480 if (isa<AllocaInst>(Ptr)) {
11481 EraseInstFromFunction(SI);
11485 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr)) {
11486 if (isa<AllocaInst>(GEP->getOperand(0))) {
11487 if (GEP->getOperand(0)->hasOneUse()) {
11488 EraseInstFromFunction(SI);
11492 if (DbgDeclareInst *DI = hasOneUsePlusDeclare(GEP->getOperand(0))) {
11493 EraseInstFromFunction(*DI);
11494 EraseInstFromFunction(SI);
11501 if (DbgDeclareInst *DI = hasOneUsePlusDeclare(Ptr)) {
11502 EraseInstFromFunction(*DI);
11503 EraseInstFromFunction(SI);
11509 // Attempt to improve the alignment.
11510 unsigned KnownAlign =
11511 GetOrEnforceKnownAlignment(Ptr, TD->getPrefTypeAlignment(Val->getType()));
11513 (SI.getAlignment() == 0 ? TD->getABITypeAlignment(Val->getType()) :
11514 SI.getAlignment()))
11515 SI.setAlignment(KnownAlign);
11517 // Do really simple DSE, to catch cases where there are several consecutive
11518 // stores to the same location, separated by a few arithmetic operations. This
11519 // situation often occurs with bitfield accesses.
11520 BasicBlock::iterator BBI = &SI;
11521 for (unsigned ScanInsts = 6; BBI != SI.getParent()->begin() && ScanInsts;
11524 // Don't count debug info directives, lest they affect codegen,
11525 // and we skip pointer-to-pointer bitcasts, which are NOPs.
11526 // It is necessary for correctness to skip those that feed into a
11527 // llvm.dbg.declare, as these are not present when debugging is off.
11528 if (isa<DbgInfoIntrinsic>(BBI) ||
11529 (isa<BitCastInst>(BBI) && isa<PointerType>(BBI->getType()))) {
11534 if (StoreInst *PrevSI = dyn_cast<StoreInst>(BBI)) {
11535 // Prev store isn't volatile, and stores to the same location?
11536 if (!PrevSI->isVolatile() &&equivalentAddressValues(PrevSI->getOperand(1),
11537 SI.getOperand(1))) {
11540 EraseInstFromFunction(*PrevSI);
11546 // If this is a load, we have to stop. However, if the loaded value is from
11547 // the pointer we're loading and is producing the pointer we're storing,
11548 // then *this* store is dead (X = load P; store X -> P).
11549 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
11550 if (LI == Val && equivalentAddressValues(LI->getOperand(0), Ptr) &&
11551 !SI.isVolatile()) {
11552 EraseInstFromFunction(SI);
11556 // Otherwise, this is a load from some other location. Stores before it
11557 // may not be dead.
11561 // Don't skip over loads or things that can modify memory.
11562 if (BBI->mayWriteToMemory() || BBI->mayReadFromMemory())
11567 if (SI.isVolatile()) return 0; // Don't hack volatile stores.
11569 // store X, null -> turns into 'unreachable' in SimplifyCFG
11570 if (isa<ConstantPointerNull>(Ptr)) {
11571 if (!isa<UndefValue>(Val)) {
11572 SI.setOperand(0, UndefValue::get(Val->getType()));
11573 if (Instruction *U = dyn_cast<Instruction>(Val))
11574 AddToWorkList(U); // Dropped a use.
11577 return 0; // Do not modify these!
11580 // store undef, Ptr -> noop
11581 if (isa<UndefValue>(Val)) {
11582 EraseInstFromFunction(SI);
11587 // If the pointer destination is a cast, see if we can fold the cast into the
11589 if (isa<CastInst>(Ptr))
11590 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
11592 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
11594 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
11598 // If this store is the last instruction in the basic block (possibly
11599 // excepting debug info instructions and the pointer bitcasts that feed
11600 // into them), and if the block ends with an unconditional branch, try
11601 // to move it to the successor block.
11605 } while (isa<DbgInfoIntrinsic>(BBI) ||
11606 (isa<BitCastInst>(BBI) && isa<PointerType>(BBI->getType())));
11607 if (BranchInst *BI = dyn_cast<BranchInst>(BBI))
11608 if (BI->isUnconditional())
11609 if (SimplifyStoreAtEndOfBlock(SI))
11610 return 0; // xform done!
11615 /// SimplifyStoreAtEndOfBlock - Turn things like:
11616 /// if () { *P = v1; } else { *P = v2 }
11617 /// into a phi node with a store in the successor.
11619 /// Simplify things like:
11620 /// *P = v1; if () { *P = v2; }
11621 /// into a phi node with a store in the successor.
11623 bool InstCombiner::SimplifyStoreAtEndOfBlock(StoreInst &SI) {
11624 BasicBlock *StoreBB = SI.getParent();
11626 // Check to see if the successor block has exactly two incoming edges. If
11627 // so, see if the other predecessor contains a store to the same location.
11628 // if so, insert a PHI node (if needed) and move the stores down.
11629 BasicBlock *DestBB = StoreBB->getTerminator()->getSuccessor(0);
11631 // Determine whether Dest has exactly two predecessors and, if so, compute
11632 // the other predecessor.
11633 pred_iterator PI = pred_begin(DestBB);
11634 BasicBlock *OtherBB = 0;
11635 if (*PI != StoreBB)
11638 if (PI == pred_end(DestBB))
11641 if (*PI != StoreBB) {
11646 if (++PI != pred_end(DestBB))
11649 // Bail out if all the relevant blocks aren't distinct (this can happen,
11650 // for example, if SI is in an infinite loop)
11651 if (StoreBB == DestBB || OtherBB == DestBB)
11654 // Verify that the other block ends in a branch and is not otherwise empty.
11655 BasicBlock::iterator BBI = OtherBB->getTerminator();
11656 BranchInst *OtherBr = dyn_cast<BranchInst>(BBI);
11657 if (!OtherBr || BBI == OtherBB->begin())
11660 // If the other block ends in an unconditional branch, check for the 'if then
11661 // else' case. there is an instruction before the branch.
11662 StoreInst *OtherStore = 0;
11663 if (OtherBr->isUnconditional()) {
11665 // Skip over debugging info.
11666 while (isa<DbgInfoIntrinsic>(BBI) ||
11667 (isa<BitCastInst>(BBI) && isa<PointerType>(BBI->getType()))) {
11668 if (BBI==OtherBB->begin())
11672 // If this isn't a store, or isn't a store to the same location, bail out.
11673 OtherStore = dyn_cast<StoreInst>(BBI);
11674 if (!OtherStore || OtherStore->getOperand(1) != SI.getOperand(1))
11677 // Otherwise, the other block ended with a conditional branch. If one of the
11678 // destinations is StoreBB, then we have the if/then case.
11679 if (OtherBr->getSuccessor(0) != StoreBB &&
11680 OtherBr->getSuccessor(1) != StoreBB)
11683 // Okay, we know that OtherBr now goes to Dest and StoreBB, so this is an
11684 // if/then triangle. See if there is a store to the same ptr as SI that
11685 // lives in OtherBB.
11687 // Check to see if we find the matching store.
11688 if ((OtherStore = dyn_cast<StoreInst>(BBI))) {
11689 if (OtherStore->getOperand(1) != SI.getOperand(1))
11693 // If we find something that may be using or overwriting the stored
11694 // value, or if we run out of instructions, we can't do the xform.
11695 if (BBI->mayReadFromMemory() || BBI->mayWriteToMemory() ||
11696 BBI == OtherBB->begin())
11700 // In order to eliminate the store in OtherBr, we have to
11701 // make sure nothing reads or overwrites the stored value in
11703 for (BasicBlock::iterator I = StoreBB->begin(); &*I != &SI; ++I) {
11704 // FIXME: This should really be AA driven.
11705 if (I->mayReadFromMemory() || I->mayWriteToMemory())
11710 // Insert a PHI node now if we need it.
11711 Value *MergedVal = OtherStore->getOperand(0);
11712 if (MergedVal != SI.getOperand(0)) {
11713 PHINode *PN = PHINode::Create(MergedVal->getType(), "storemerge");
11714 PN->reserveOperandSpace(2);
11715 PN->addIncoming(SI.getOperand(0), SI.getParent());
11716 PN->addIncoming(OtherStore->getOperand(0), OtherBB);
11717 MergedVal = InsertNewInstBefore(PN, DestBB->front());
11720 // Advance to a place where it is safe to insert the new store and
11722 BBI = DestBB->getFirstNonPHI();
11723 InsertNewInstBefore(new StoreInst(MergedVal, SI.getOperand(1),
11724 OtherStore->isVolatile()), *BBI);
11726 // Nuke the old stores.
11727 EraseInstFromFunction(SI);
11728 EraseInstFromFunction(*OtherStore);
11734 Instruction *InstCombiner::visitBranchInst(BranchInst &BI) {
11735 // Change br (not X), label True, label False to: br X, label False, True
11737 BasicBlock *TrueDest;
11738 BasicBlock *FalseDest;
11739 if (match(&BI, m_Br(m_Not(m_Value(X)), TrueDest, FalseDest)) &&
11740 !isa<Constant>(X)) {
11741 // Swap Destinations and condition...
11742 BI.setCondition(X);
11743 BI.setSuccessor(0, FalseDest);
11744 BI.setSuccessor(1, TrueDest);
11748 // Cannonicalize fcmp_one -> fcmp_oeq
11749 FCmpInst::Predicate FPred; Value *Y;
11750 if (match(&BI, m_Br(m_FCmp(FPred, m_Value(X), m_Value(Y)),
11751 TrueDest, FalseDest)))
11752 if ((FPred == FCmpInst::FCMP_ONE || FPred == FCmpInst::FCMP_OLE ||
11753 FPred == FCmpInst::FCMP_OGE) && BI.getCondition()->hasOneUse()) {
11754 FCmpInst *I = cast<FCmpInst>(BI.getCondition());
11755 FCmpInst::Predicate NewPred = FCmpInst::getInversePredicate(FPred);
11756 Instruction *NewSCC = new FCmpInst(NewPred, X, Y, "", I);
11757 NewSCC->takeName(I);
11758 // Swap Destinations and condition...
11759 BI.setCondition(NewSCC);
11760 BI.setSuccessor(0, FalseDest);
11761 BI.setSuccessor(1, TrueDest);
11762 RemoveFromWorkList(I);
11763 I->eraseFromParent();
11764 AddToWorkList(NewSCC);
11768 // Cannonicalize icmp_ne -> icmp_eq
11769 ICmpInst::Predicate IPred;
11770 if (match(&BI, m_Br(m_ICmp(IPred, m_Value(X), m_Value(Y)),
11771 TrueDest, FalseDest)))
11772 if ((IPred == ICmpInst::ICMP_NE || IPred == ICmpInst::ICMP_ULE ||
11773 IPred == ICmpInst::ICMP_SLE || IPred == ICmpInst::ICMP_UGE ||
11774 IPred == ICmpInst::ICMP_SGE) && BI.getCondition()->hasOneUse()) {
11775 ICmpInst *I = cast<ICmpInst>(BI.getCondition());
11776 ICmpInst::Predicate NewPred = ICmpInst::getInversePredicate(IPred);
11777 Instruction *NewSCC = new ICmpInst(NewPred, X, Y, "", I);
11778 NewSCC->takeName(I);
11779 // Swap Destinations and condition...
11780 BI.setCondition(NewSCC);
11781 BI.setSuccessor(0, FalseDest);
11782 BI.setSuccessor(1, TrueDest);
11783 RemoveFromWorkList(I);
11784 I->eraseFromParent();;
11785 AddToWorkList(NewSCC);
11792 Instruction *InstCombiner::visitSwitchInst(SwitchInst &SI) {
11793 Value *Cond = SI.getCondition();
11794 if (Instruction *I = dyn_cast<Instruction>(Cond)) {
11795 if (I->getOpcode() == Instruction::Add)
11796 if (ConstantInt *AddRHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
11797 // change 'switch (X+4) case 1:' into 'switch (X) case -3'
11798 for (unsigned i = 2, e = SI.getNumOperands(); i != e; i += 2)
11799 SI.setOperand(i,ConstantExpr::getSub(cast<Constant>(SI.getOperand(i)),
11801 SI.setOperand(0, I->getOperand(0));
11809 Instruction *InstCombiner::visitExtractValueInst(ExtractValueInst &EV) {
11810 Value *Agg = EV.getAggregateOperand();
11812 if (!EV.hasIndices())
11813 return ReplaceInstUsesWith(EV, Agg);
11815 if (Constant *C = dyn_cast<Constant>(Agg)) {
11816 if (isa<UndefValue>(C))
11817 return ReplaceInstUsesWith(EV, UndefValue::get(EV.getType()));
11819 if (isa<ConstantAggregateZero>(C))
11820 return ReplaceInstUsesWith(EV, Constant::getNullValue(EV.getType()));
11822 if (isa<ConstantArray>(C) || isa<ConstantStruct>(C)) {
11823 // Extract the element indexed by the first index out of the constant
11824 Value *V = C->getOperand(*EV.idx_begin());
11825 if (EV.getNumIndices() > 1)
11826 // Extract the remaining indices out of the constant indexed by the
11828 return ExtractValueInst::Create(V, EV.idx_begin() + 1, EV.idx_end());
11830 return ReplaceInstUsesWith(EV, V);
11832 return 0; // Can't handle other constants
11834 if (InsertValueInst *IV = dyn_cast<InsertValueInst>(Agg)) {
11835 // We're extracting from an insertvalue instruction, compare the indices
11836 const unsigned *exti, *exte, *insi, *inse;
11837 for (exti = EV.idx_begin(), insi = IV->idx_begin(),
11838 exte = EV.idx_end(), inse = IV->idx_end();
11839 exti != exte && insi != inse;
11841 if (*insi != *exti)
11842 // The insert and extract both reference distinctly different elements.
11843 // This means the extract is not influenced by the insert, and we can
11844 // replace the aggregate operand of the extract with the aggregate
11845 // operand of the insert. i.e., replace
11846 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1
11847 // %E = extractvalue { i32, { i32 } } %I, 0
11849 // %E = extractvalue { i32, { i32 } } %A, 0
11850 return ExtractValueInst::Create(IV->getAggregateOperand(),
11851 EV.idx_begin(), EV.idx_end());
11853 if (exti == exte && insi == inse)
11854 // Both iterators are at the end: Index lists are identical. Replace
11855 // %B = insertvalue { i32, { i32 } } %A, i32 42, 1, 0
11856 // %C = extractvalue { i32, { i32 } } %B, 1, 0
11858 return ReplaceInstUsesWith(EV, IV->getInsertedValueOperand());
11859 if (exti == exte) {
11860 // The extract list is a prefix of the insert list. i.e. replace
11861 // %I = insertvalue { i32, { i32 } } %A, i32 42, 1, 0
11862 // %E = extractvalue { i32, { i32 } } %I, 1
11864 // %X = extractvalue { i32, { i32 } } %A, 1
11865 // %E = insertvalue { i32 } %X, i32 42, 0
11866 // by switching the order of the insert and extract (though the
11867 // insertvalue should be left in, since it may have other uses).
11868 Value *NewEV = InsertNewInstBefore(
11869 ExtractValueInst::Create(IV->getAggregateOperand(),
11870 EV.idx_begin(), EV.idx_end()),
11872 return InsertValueInst::Create(NewEV, IV->getInsertedValueOperand(),
11876 // The insert list is a prefix of the extract list
11877 // We can simply remove the common indices from the extract and make it
11878 // operate on the inserted value instead of the insertvalue result.
11880 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1
11881 // %E = extractvalue { i32, { i32 } } %I, 1, 0
11883 // %E extractvalue { i32 } { i32 42 }, 0
11884 return ExtractValueInst::Create(IV->getInsertedValueOperand(),
11887 // Can't simplify extracts from other values. Note that nested extracts are
11888 // already simplified implicitely by the above (extract ( extract (insert) )
11889 // will be translated into extract ( insert ( extract ) ) first and then just
11890 // the value inserted, if appropriate).
11894 /// CheapToScalarize - Return true if the value is cheaper to scalarize than it
11895 /// is to leave as a vector operation.
11896 static bool CheapToScalarize(Value *V, bool isConstant) {
11897 if (isa<ConstantAggregateZero>(V))
11899 if (ConstantVector *C = dyn_cast<ConstantVector>(V)) {
11900 if (isConstant) return true;
11901 // If all elts are the same, we can extract.
11902 Constant *Op0 = C->getOperand(0);
11903 for (unsigned i = 1; i < C->getNumOperands(); ++i)
11904 if (C->getOperand(i) != Op0)
11908 Instruction *I = dyn_cast<Instruction>(V);
11909 if (!I) return false;
11911 // Insert element gets simplified to the inserted element or is deleted if
11912 // this is constant idx extract element and its a constant idx insertelt.
11913 if (I->getOpcode() == Instruction::InsertElement && isConstant &&
11914 isa<ConstantInt>(I->getOperand(2)))
11916 if (I->getOpcode() == Instruction::Load && I->hasOneUse())
11918 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I))
11919 if (BO->hasOneUse() &&
11920 (CheapToScalarize(BO->getOperand(0), isConstant) ||
11921 CheapToScalarize(BO->getOperand(1), isConstant)))
11923 if (CmpInst *CI = dyn_cast<CmpInst>(I))
11924 if (CI->hasOneUse() &&
11925 (CheapToScalarize(CI->getOperand(0), isConstant) ||
11926 CheapToScalarize(CI->getOperand(1), isConstant)))
11932 /// Read and decode a shufflevector mask.
11934 /// It turns undef elements into values that are larger than the number of
11935 /// elements in the input.
11936 static std::vector<unsigned> getShuffleMask(const ShuffleVectorInst *SVI) {
11937 unsigned NElts = SVI->getType()->getNumElements();
11938 if (isa<ConstantAggregateZero>(SVI->getOperand(2)))
11939 return std::vector<unsigned>(NElts, 0);
11940 if (isa<UndefValue>(SVI->getOperand(2)))
11941 return std::vector<unsigned>(NElts, 2*NElts);
11943 std::vector<unsigned> Result;
11944 const ConstantVector *CP = cast<ConstantVector>(SVI->getOperand(2));
11945 for (User::const_op_iterator i = CP->op_begin(), e = CP->op_end(); i!=e; ++i)
11946 if (isa<UndefValue>(*i))
11947 Result.push_back(NElts*2); // undef -> 8
11949 Result.push_back(cast<ConstantInt>(*i)->getZExtValue());
11953 /// FindScalarElement - Given a vector and an element number, see if the scalar
11954 /// value is already around as a register, for example if it were inserted then
11955 /// extracted from the vector.
11956 static Value *FindScalarElement(Value *V, unsigned EltNo) {
11957 assert(isa<VectorType>(V->getType()) && "Not looking at a vector?");
11958 const VectorType *PTy = cast<VectorType>(V->getType());
11959 unsigned Width = PTy->getNumElements();
11960 if (EltNo >= Width) // Out of range access.
11961 return UndefValue::get(PTy->getElementType());
11963 if (isa<UndefValue>(V))
11964 return UndefValue::get(PTy->getElementType());
11965 else if (isa<ConstantAggregateZero>(V))
11966 return Constant::getNullValue(PTy->getElementType());
11967 else if (ConstantVector *CP = dyn_cast<ConstantVector>(V))
11968 return CP->getOperand(EltNo);
11969 else if (InsertElementInst *III = dyn_cast<InsertElementInst>(V)) {
11970 // If this is an insert to a variable element, we don't know what it is.
11971 if (!isa<ConstantInt>(III->getOperand(2)))
11973 unsigned IIElt = cast<ConstantInt>(III->getOperand(2))->getZExtValue();
11975 // If this is an insert to the element we are looking for, return the
11977 if (EltNo == IIElt)
11978 return III->getOperand(1);
11980 // Otherwise, the insertelement doesn't modify the value, recurse on its
11982 return FindScalarElement(III->getOperand(0), EltNo);
11983 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(V)) {
11984 unsigned LHSWidth =
11985 cast<VectorType>(SVI->getOperand(0)->getType())->getNumElements();
11986 unsigned InEl = getShuffleMask(SVI)[EltNo];
11987 if (InEl < LHSWidth)
11988 return FindScalarElement(SVI->getOperand(0), InEl);
11989 else if (InEl < LHSWidth*2)
11990 return FindScalarElement(SVI->getOperand(1), InEl - LHSWidth);
11992 return UndefValue::get(PTy->getElementType());
11995 // Otherwise, we don't know.
11999 Instruction *InstCombiner::visitExtractElementInst(ExtractElementInst &EI) {
12000 // If vector val is undef, replace extract with scalar undef.
12001 if (isa<UndefValue>(EI.getOperand(0)))
12002 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
12004 // If vector val is constant 0, replace extract with scalar 0.
12005 if (isa<ConstantAggregateZero>(EI.getOperand(0)))
12006 return ReplaceInstUsesWith(EI, Constant::getNullValue(EI.getType()));
12008 if (ConstantVector *C = dyn_cast<ConstantVector>(EI.getOperand(0))) {
12009 // If vector val is constant with all elements the same, replace EI with
12010 // that element. When the elements are not identical, we cannot replace yet
12011 // (we do that below, but only when the index is constant).
12012 Constant *op0 = C->getOperand(0);
12013 for (unsigned i = 1; i < C->getNumOperands(); ++i)
12014 if (C->getOperand(i) != op0) {
12019 return ReplaceInstUsesWith(EI, op0);
12022 // If extracting a specified index from the vector, see if we can recursively
12023 // find a previously computed scalar that was inserted into the vector.
12024 if (ConstantInt *IdxC = dyn_cast<ConstantInt>(EI.getOperand(1))) {
12025 unsigned IndexVal = IdxC->getZExtValue();
12026 unsigned VectorWidth =
12027 cast<VectorType>(EI.getOperand(0)->getType())->getNumElements();
12029 // If this is extracting an invalid index, turn this into undef, to avoid
12030 // crashing the code below.
12031 if (IndexVal >= VectorWidth)
12032 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
12034 // This instruction only demands the single element from the input vector.
12035 // If the input vector has a single use, simplify it based on this use
12037 if (EI.getOperand(0)->hasOneUse() && VectorWidth != 1) {
12038 APInt UndefElts(VectorWidth, 0);
12039 APInt DemandedMask(VectorWidth, 1 << IndexVal);
12040 if (Value *V = SimplifyDemandedVectorElts(EI.getOperand(0),
12041 DemandedMask, UndefElts)) {
12042 EI.setOperand(0, V);
12047 if (Value *Elt = FindScalarElement(EI.getOperand(0), IndexVal))
12048 return ReplaceInstUsesWith(EI, Elt);
12050 // If the this extractelement is directly using a bitcast from a vector of
12051 // the same number of elements, see if we can find the source element from
12052 // it. In this case, we will end up needing to bitcast the scalars.
12053 if (BitCastInst *BCI = dyn_cast<BitCastInst>(EI.getOperand(0))) {
12054 if (const VectorType *VT =
12055 dyn_cast<VectorType>(BCI->getOperand(0)->getType()))
12056 if (VT->getNumElements() == VectorWidth)
12057 if (Value *Elt = FindScalarElement(BCI->getOperand(0), IndexVal))
12058 return new BitCastInst(Elt, EI.getType());
12062 if (Instruction *I = dyn_cast<Instruction>(EI.getOperand(0))) {
12063 if (I->hasOneUse()) {
12064 // Push extractelement into predecessor operation if legal and
12065 // profitable to do so
12066 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
12067 bool isConstantElt = isa<ConstantInt>(EI.getOperand(1));
12068 if (CheapToScalarize(BO, isConstantElt)) {
12069 ExtractElementInst *newEI0 =
12070 new ExtractElementInst(BO->getOperand(0), EI.getOperand(1),
12071 EI.getName()+".lhs");
12072 ExtractElementInst *newEI1 =
12073 new ExtractElementInst(BO->getOperand(1), EI.getOperand(1),
12074 EI.getName()+".rhs");
12075 InsertNewInstBefore(newEI0, EI);
12076 InsertNewInstBefore(newEI1, EI);
12077 return BinaryOperator::Create(BO->getOpcode(), newEI0, newEI1);
12079 } else if (isa<LoadInst>(I)) {
12081 cast<PointerType>(I->getOperand(0)->getType())->getAddressSpace();
12082 Value *Ptr = InsertBitCastBefore(I->getOperand(0),
12083 PointerType::get(EI.getType(), AS),EI);
12084 GetElementPtrInst *GEP =
12085 GetElementPtrInst::Create(Ptr, EI.getOperand(1), I->getName()+".gep");
12086 InsertNewInstBefore(GEP, EI);
12087 return new LoadInst(GEP);
12090 if (InsertElementInst *IE = dyn_cast<InsertElementInst>(I)) {
12091 // Extracting the inserted element?
12092 if (IE->getOperand(2) == EI.getOperand(1))
12093 return ReplaceInstUsesWith(EI, IE->getOperand(1));
12094 // If the inserted and extracted elements are constants, they must not
12095 // be the same value, extract from the pre-inserted value instead.
12096 if (isa<Constant>(IE->getOperand(2)) &&
12097 isa<Constant>(EI.getOperand(1))) {
12098 AddUsesToWorkList(EI);
12099 EI.setOperand(0, IE->getOperand(0));
12102 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I)) {
12103 // If this is extracting an element from a shufflevector, figure out where
12104 // it came from and extract from the appropriate input element instead.
12105 if (ConstantInt *Elt = dyn_cast<ConstantInt>(EI.getOperand(1))) {
12106 unsigned SrcIdx = getShuffleMask(SVI)[Elt->getZExtValue()];
12108 unsigned LHSWidth =
12109 cast<VectorType>(SVI->getOperand(0)->getType())->getNumElements();
12111 if (SrcIdx < LHSWidth)
12112 Src = SVI->getOperand(0);
12113 else if (SrcIdx < LHSWidth*2) {
12114 SrcIdx -= LHSWidth;
12115 Src = SVI->getOperand(1);
12117 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
12119 return new ExtractElementInst(Src, SrcIdx);
12126 /// CollectSingleShuffleElements - If V is a shuffle of values that ONLY returns
12127 /// elements from either LHS or RHS, return the shuffle mask and true.
12128 /// Otherwise, return false.
12129 static bool CollectSingleShuffleElements(Value *V, Value *LHS, Value *RHS,
12130 std::vector<Constant*> &Mask) {
12131 assert(V->getType() == LHS->getType() && V->getType() == RHS->getType() &&
12132 "Invalid CollectSingleShuffleElements");
12133 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
12135 if (isa<UndefValue>(V)) {
12136 Mask.assign(NumElts, UndefValue::get(Type::Int32Ty));
12138 } else if (V == LHS) {
12139 for (unsigned i = 0; i != NumElts; ++i)
12140 Mask.push_back(ConstantInt::get(Type::Int32Ty, i));
12142 } else if (V == RHS) {
12143 for (unsigned i = 0; i != NumElts; ++i)
12144 Mask.push_back(ConstantInt::get(Type::Int32Ty, i+NumElts));
12146 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
12147 // If this is an insert of an extract from some other vector, include it.
12148 Value *VecOp = IEI->getOperand(0);
12149 Value *ScalarOp = IEI->getOperand(1);
12150 Value *IdxOp = IEI->getOperand(2);
12152 if (!isa<ConstantInt>(IdxOp))
12154 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
12156 if (isa<UndefValue>(ScalarOp)) { // inserting undef into vector.
12157 // Okay, we can handle this if the vector we are insertinting into is
12158 // transitively ok.
12159 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
12160 // If so, update the mask to reflect the inserted undef.
12161 Mask[InsertedIdx] = UndefValue::get(Type::Int32Ty);
12164 } else if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)){
12165 if (isa<ConstantInt>(EI->getOperand(1)) &&
12166 EI->getOperand(0)->getType() == V->getType()) {
12167 unsigned ExtractedIdx =
12168 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
12170 // This must be extracting from either LHS or RHS.
12171 if (EI->getOperand(0) == LHS || EI->getOperand(0) == RHS) {
12172 // Okay, we can handle this if the vector we are insertinting into is
12173 // transitively ok.
12174 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
12175 // If so, update the mask to reflect the inserted value.
12176 if (EI->getOperand(0) == LHS) {
12177 Mask[InsertedIdx % NumElts] =
12178 ConstantInt::get(Type::Int32Ty, ExtractedIdx);
12180 assert(EI->getOperand(0) == RHS);
12181 Mask[InsertedIdx % NumElts] =
12182 ConstantInt::get(Type::Int32Ty, ExtractedIdx+NumElts);
12191 // TODO: Handle shufflevector here!
12196 /// CollectShuffleElements - We are building a shuffle of V, using RHS as the
12197 /// RHS of the shuffle instruction, if it is not null. Return a shuffle mask
12198 /// that computes V and the LHS value of the shuffle.
12199 static Value *CollectShuffleElements(Value *V, std::vector<Constant*> &Mask,
12201 assert(isa<VectorType>(V->getType()) &&
12202 (RHS == 0 || V->getType() == RHS->getType()) &&
12203 "Invalid shuffle!");
12204 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
12206 if (isa<UndefValue>(V)) {
12207 Mask.assign(NumElts, UndefValue::get(Type::Int32Ty));
12209 } else if (isa<ConstantAggregateZero>(V)) {
12210 Mask.assign(NumElts, ConstantInt::get(Type::Int32Ty, 0));
12212 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
12213 // If this is an insert of an extract from some other vector, include it.
12214 Value *VecOp = IEI->getOperand(0);
12215 Value *ScalarOp = IEI->getOperand(1);
12216 Value *IdxOp = IEI->getOperand(2);
12218 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
12219 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
12220 EI->getOperand(0)->getType() == V->getType()) {
12221 unsigned ExtractedIdx =
12222 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
12223 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
12225 // Either the extracted from or inserted into vector must be RHSVec,
12226 // otherwise we'd end up with a shuffle of three inputs.
12227 if (EI->getOperand(0) == RHS || RHS == 0) {
12228 RHS = EI->getOperand(0);
12229 Value *V = CollectShuffleElements(VecOp, Mask, RHS);
12230 Mask[InsertedIdx % NumElts] =
12231 ConstantInt::get(Type::Int32Ty, NumElts+ExtractedIdx);
12235 if (VecOp == RHS) {
12236 Value *V = CollectShuffleElements(EI->getOperand(0), Mask, RHS);
12237 // Everything but the extracted element is replaced with the RHS.
12238 for (unsigned i = 0; i != NumElts; ++i) {
12239 if (i != InsertedIdx)
12240 Mask[i] = ConstantInt::get(Type::Int32Ty, NumElts+i);
12245 // If this insertelement is a chain that comes from exactly these two
12246 // vectors, return the vector and the effective shuffle.
12247 if (CollectSingleShuffleElements(IEI, EI->getOperand(0), RHS, Mask))
12248 return EI->getOperand(0);
12253 // TODO: Handle shufflevector here!
12255 // Otherwise, can't do anything fancy. Return an identity vector.
12256 for (unsigned i = 0; i != NumElts; ++i)
12257 Mask.push_back(ConstantInt::get(Type::Int32Ty, i));
12261 Instruction *InstCombiner::visitInsertElementInst(InsertElementInst &IE) {
12262 Value *VecOp = IE.getOperand(0);
12263 Value *ScalarOp = IE.getOperand(1);
12264 Value *IdxOp = IE.getOperand(2);
12266 // Inserting an undef or into an undefined place, remove this.
12267 if (isa<UndefValue>(ScalarOp) || isa<UndefValue>(IdxOp))
12268 ReplaceInstUsesWith(IE, VecOp);
12270 // If the inserted element was extracted from some other vector, and if the
12271 // indexes are constant, try to turn this into a shufflevector operation.
12272 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
12273 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
12274 EI->getOperand(0)->getType() == IE.getType()) {
12275 unsigned NumVectorElts = IE.getType()->getNumElements();
12276 unsigned ExtractedIdx =
12277 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
12278 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
12280 if (ExtractedIdx >= NumVectorElts) // Out of range extract.
12281 return ReplaceInstUsesWith(IE, VecOp);
12283 if (InsertedIdx >= NumVectorElts) // Out of range insert.
12284 return ReplaceInstUsesWith(IE, UndefValue::get(IE.getType()));
12286 // If we are extracting a value from a vector, then inserting it right
12287 // back into the same place, just use the input vector.
12288 if (EI->getOperand(0) == VecOp && ExtractedIdx == InsertedIdx)
12289 return ReplaceInstUsesWith(IE, VecOp);
12291 // We could theoretically do this for ANY input. However, doing so could
12292 // turn chains of insertelement instructions into a chain of shufflevector
12293 // instructions, and right now we do not merge shufflevectors. As such,
12294 // only do this in a situation where it is clear that there is benefit.
12295 if (isa<UndefValue>(VecOp) || isa<ConstantAggregateZero>(VecOp)) {
12296 // Turn this into shuffle(EIOp0, VecOp, Mask). The result has all of
12297 // the values of VecOp, except then one read from EIOp0.
12298 // Build a new shuffle mask.
12299 std::vector<Constant*> Mask;
12300 if (isa<UndefValue>(VecOp))
12301 Mask.assign(NumVectorElts, UndefValue::get(Type::Int32Ty));
12303 assert(isa<ConstantAggregateZero>(VecOp) && "Unknown thing");
12304 Mask.assign(NumVectorElts, ConstantInt::get(Type::Int32Ty,
12307 Mask[InsertedIdx] = ConstantInt::get(Type::Int32Ty, ExtractedIdx);
12308 return new ShuffleVectorInst(EI->getOperand(0), VecOp,
12309 ConstantVector::get(Mask));
12312 // If this insertelement isn't used by some other insertelement, turn it
12313 // (and any insertelements it points to), into one big shuffle.
12314 if (!IE.hasOneUse() || !isa<InsertElementInst>(IE.use_back())) {
12315 std::vector<Constant*> Mask;
12317 Value *LHS = CollectShuffleElements(&IE, Mask, RHS);
12318 if (RHS == 0) RHS = UndefValue::get(LHS->getType());
12319 // We now have a shuffle of LHS, RHS, Mask.
12320 return new ShuffleVectorInst(LHS, RHS, ConstantVector::get(Mask));
12329 Instruction *InstCombiner::visitShuffleVectorInst(ShuffleVectorInst &SVI) {
12330 Value *LHS = SVI.getOperand(0);
12331 Value *RHS = SVI.getOperand(1);
12332 std::vector<unsigned> Mask = getShuffleMask(&SVI);
12334 bool MadeChange = false;
12336 // Undefined shuffle mask -> undefined value.
12337 if (isa<UndefValue>(SVI.getOperand(2)))
12338 return ReplaceInstUsesWith(SVI, UndefValue::get(SVI.getType()));
12340 unsigned VWidth = cast<VectorType>(SVI.getType())->getNumElements();
12342 if (VWidth != cast<VectorType>(LHS->getType())->getNumElements())
12345 APInt UndefElts(VWidth, 0);
12346 APInt AllOnesEltMask(APInt::getAllOnesValue(VWidth));
12347 if (SimplifyDemandedVectorElts(&SVI, AllOnesEltMask, UndefElts)) {
12348 LHS = SVI.getOperand(0);
12349 RHS = SVI.getOperand(1);
12353 // Canonicalize shuffle(x ,x,mask) -> shuffle(x, undef,mask')
12354 // Canonicalize shuffle(undef,x,mask) -> shuffle(x, undef,mask').
12355 if (LHS == RHS || isa<UndefValue>(LHS)) {
12356 if (isa<UndefValue>(LHS) && LHS == RHS) {
12357 // shuffle(undef,undef,mask) -> undef.
12358 return ReplaceInstUsesWith(SVI, LHS);
12361 // Remap any references to RHS to use LHS.
12362 std::vector<Constant*> Elts;
12363 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
12364 if (Mask[i] >= 2*e)
12365 Elts.push_back(UndefValue::get(Type::Int32Ty));
12367 if ((Mask[i] >= e && isa<UndefValue>(RHS)) ||
12368 (Mask[i] < e && isa<UndefValue>(LHS))) {
12369 Mask[i] = 2*e; // Turn into undef.
12370 Elts.push_back(UndefValue::get(Type::Int32Ty));
12372 Mask[i] = Mask[i] % e; // Force to LHS.
12373 Elts.push_back(ConstantInt::get(Type::Int32Ty, Mask[i]));
12377 SVI.setOperand(0, SVI.getOperand(1));
12378 SVI.setOperand(1, UndefValue::get(RHS->getType()));
12379 SVI.setOperand(2, ConstantVector::get(Elts));
12380 LHS = SVI.getOperand(0);
12381 RHS = SVI.getOperand(1);
12385 // Analyze the shuffle, are the LHS or RHS and identity shuffles?
12386 bool isLHSID = true, isRHSID = true;
12388 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
12389 if (Mask[i] >= e*2) continue; // Ignore undef values.
12390 // Is this an identity shuffle of the LHS value?
12391 isLHSID &= (Mask[i] == i);
12393 // Is this an identity shuffle of the RHS value?
12394 isRHSID &= (Mask[i]-e == i);
12397 // Eliminate identity shuffles.
12398 if (isLHSID) return ReplaceInstUsesWith(SVI, LHS);
12399 if (isRHSID) return ReplaceInstUsesWith(SVI, RHS);
12401 // If the LHS is a shufflevector itself, see if we can combine it with this
12402 // one without producing an unusual shuffle. Here we are really conservative:
12403 // we are absolutely afraid of producing a shuffle mask not in the input
12404 // program, because the code gen may not be smart enough to turn a merged
12405 // shuffle into two specific shuffles: it may produce worse code. As such,
12406 // we only merge two shuffles if the result is one of the two input shuffle
12407 // masks. In this case, merging the shuffles just removes one instruction,
12408 // which we know is safe. This is good for things like turning:
12409 // (splat(splat)) -> splat.
12410 if (ShuffleVectorInst *LHSSVI = dyn_cast<ShuffleVectorInst>(LHS)) {
12411 if (isa<UndefValue>(RHS)) {
12412 std::vector<unsigned> LHSMask = getShuffleMask(LHSSVI);
12414 std::vector<unsigned> NewMask;
12415 for (unsigned i = 0, e = Mask.size(); i != e; ++i)
12416 if (Mask[i] >= 2*e)
12417 NewMask.push_back(2*e);
12419 NewMask.push_back(LHSMask[Mask[i]]);
12421 // If the result mask is equal to the src shuffle or this shuffle mask, do
12422 // the replacement.
12423 if (NewMask == LHSMask || NewMask == Mask) {
12424 unsigned LHSInNElts =
12425 cast<VectorType>(LHSSVI->getOperand(0)->getType())->getNumElements();
12426 std::vector<Constant*> Elts;
12427 for (unsigned i = 0, e = NewMask.size(); i != e; ++i) {
12428 if (NewMask[i] >= LHSInNElts*2) {
12429 Elts.push_back(UndefValue::get(Type::Int32Ty));
12431 Elts.push_back(ConstantInt::get(Type::Int32Ty, NewMask[i]));
12434 return new ShuffleVectorInst(LHSSVI->getOperand(0),
12435 LHSSVI->getOperand(1),
12436 ConstantVector::get(Elts));
12441 return MadeChange ? &SVI : 0;
12447 /// TryToSinkInstruction - Try to move the specified instruction from its
12448 /// current block into the beginning of DestBlock, which can only happen if it's
12449 /// safe to move the instruction past all of the instructions between it and the
12450 /// end of its block.
12451 static bool TryToSinkInstruction(Instruction *I, BasicBlock *DestBlock) {
12452 assert(I->hasOneUse() && "Invariants didn't hold!");
12454 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
12455 if (isa<PHINode>(I) || I->mayWriteToMemory() || isa<TerminatorInst>(I))
12458 // Do not sink alloca instructions out of the entry block.
12459 if (isa<AllocaInst>(I) && I->getParent() ==
12460 &DestBlock->getParent()->getEntryBlock())
12463 // We can only sink load instructions if there is nothing between the load and
12464 // the end of block that could change the value.
12465 if (I->mayReadFromMemory()) {
12466 for (BasicBlock::iterator Scan = I, E = I->getParent()->end();
12468 if (Scan->mayWriteToMemory())
12472 BasicBlock::iterator InsertPos = DestBlock->getFirstNonPHI();
12474 CopyPrecedingStopPoint(I, InsertPos);
12475 I->moveBefore(InsertPos);
12481 /// AddReachableCodeToWorklist - Walk the function in depth-first order, adding
12482 /// all reachable code to the worklist.
12484 /// This has a couple of tricks to make the code faster and more powerful. In
12485 /// particular, we constant fold and DCE instructions as we go, to avoid adding
12486 /// them to the worklist (this significantly speeds up instcombine on code where
12487 /// many instructions are dead or constant). Additionally, if we find a branch
12488 /// whose condition is a known constant, we only visit the reachable successors.
12490 static void AddReachableCodeToWorklist(BasicBlock *BB,
12491 SmallPtrSet<BasicBlock*, 64> &Visited,
12493 const TargetData *TD) {
12494 SmallVector<BasicBlock*, 256> Worklist;
12495 Worklist.push_back(BB);
12497 while (!Worklist.empty()) {
12498 BB = Worklist.back();
12499 Worklist.pop_back();
12501 // We have now visited this block! If we've already been here, ignore it.
12502 if (!Visited.insert(BB)) continue;
12504 DbgInfoIntrinsic *DBI_Prev = NULL;
12505 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) {
12506 Instruction *Inst = BBI++;
12508 // DCE instruction if trivially dead.
12509 if (isInstructionTriviallyDead(Inst)) {
12511 DOUT << "IC: DCE: " << *Inst;
12512 Inst->eraseFromParent();
12516 // ConstantProp instruction if trivially constant.
12517 if (Constant *C = ConstantFoldInstruction(Inst, TD)) {
12518 DOUT << "IC: ConstFold to: " << *C << " from: " << *Inst;
12519 Inst->replaceAllUsesWith(C);
12521 Inst->eraseFromParent();
12525 // If there are two consecutive llvm.dbg.stoppoint calls then
12526 // it is likely that the optimizer deleted code in between these
12528 DbgInfoIntrinsic *DBI_Next = dyn_cast<DbgInfoIntrinsic>(Inst);
12531 && DBI_Prev->getIntrinsicID() == llvm::Intrinsic::dbg_stoppoint
12532 && DBI_Next->getIntrinsicID() == llvm::Intrinsic::dbg_stoppoint) {
12533 IC.RemoveFromWorkList(DBI_Prev);
12534 DBI_Prev->eraseFromParent();
12536 DBI_Prev = DBI_Next;
12541 IC.AddToWorkList(Inst);
12544 // Recursively visit successors. If this is a branch or switch on a
12545 // constant, only visit the reachable successor.
12546 TerminatorInst *TI = BB->getTerminator();
12547 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
12548 if (BI->isConditional() && isa<ConstantInt>(BI->getCondition())) {
12549 bool CondVal = cast<ConstantInt>(BI->getCondition())->getZExtValue();
12550 BasicBlock *ReachableBB = BI->getSuccessor(!CondVal);
12551 Worklist.push_back(ReachableBB);
12554 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
12555 if (ConstantInt *Cond = dyn_cast<ConstantInt>(SI->getCondition())) {
12556 // See if this is an explicit destination.
12557 for (unsigned i = 1, e = SI->getNumSuccessors(); i != e; ++i)
12558 if (SI->getCaseValue(i) == Cond) {
12559 BasicBlock *ReachableBB = SI->getSuccessor(i);
12560 Worklist.push_back(ReachableBB);
12564 // Otherwise it is the default destination.
12565 Worklist.push_back(SI->getSuccessor(0));
12570 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
12571 Worklist.push_back(TI->getSuccessor(i));
12575 bool InstCombiner::DoOneIteration(Function &F, unsigned Iteration) {
12576 bool Changed = false;
12577 TD = &getAnalysis<TargetData>();
12579 DEBUG(DOUT << "\n\nINSTCOMBINE ITERATION #" << Iteration << " on "
12580 << F.getNameStr() << "\n");
12583 // Do a depth-first traversal of the function, populate the worklist with
12584 // the reachable instructions. Ignore blocks that are not reachable. Keep
12585 // track of which blocks we visit.
12586 SmallPtrSet<BasicBlock*, 64> Visited;
12587 AddReachableCodeToWorklist(F.begin(), Visited, *this, TD);
12589 // Do a quick scan over the function. If we find any blocks that are
12590 // unreachable, remove any instructions inside of them. This prevents
12591 // the instcombine code from having to deal with some bad special cases.
12592 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
12593 if (!Visited.count(BB)) {
12594 Instruction *Term = BB->getTerminator();
12595 while (Term != BB->begin()) { // Remove instrs bottom-up
12596 BasicBlock::iterator I = Term; --I;
12598 DOUT << "IC: DCE: " << *I;
12599 // A debug intrinsic shouldn't force another iteration if we weren't
12600 // going to do one without it.
12601 if (!isa<DbgInfoIntrinsic>(I)) {
12605 if (!I->use_empty())
12606 I->replaceAllUsesWith(UndefValue::get(I->getType()));
12607 I->eraseFromParent();
12612 while (!Worklist.empty()) {
12613 Instruction *I = RemoveOneFromWorkList();
12614 if (I == 0) continue; // skip null values.
12616 // Check to see if we can DCE the instruction.
12617 if (isInstructionTriviallyDead(I)) {
12618 // Add operands to the worklist.
12619 if (I->getNumOperands() < 4)
12620 AddUsesToWorkList(*I);
12623 DOUT << "IC: DCE: " << *I;
12625 I->eraseFromParent();
12626 RemoveFromWorkList(I);
12631 // Instruction isn't dead, see if we can constant propagate it.
12632 if (Constant *C = ConstantFoldInstruction(I, TD)) {
12633 DOUT << "IC: ConstFold to: " << *C << " from: " << *I;
12635 // Add operands to the worklist.
12636 AddUsesToWorkList(*I);
12637 ReplaceInstUsesWith(*I, C);
12640 I->eraseFromParent();
12641 RemoveFromWorkList(I);
12646 if (TD && I->getType()->getTypeID() == Type::VoidTyID) {
12647 // See if we can constant fold its operands.
12648 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i)
12649 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(i))
12650 if (Constant *NewC = ConstantFoldConstantExpression(CE, TD))
12657 // See if we can trivially sink this instruction to a successor basic block.
12658 if (I->hasOneUse()) {
12659 BasicBlock *BB = I->getParent();
12660 BasicBlock *UserParent = cast<Instruction>(I->use_back())->getParent();
12661 if (UserParent != BB) {
12662 bool UserIsSuccessor = false;
12663 // See if the user is one of our successors.
12664 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI)
12665 if (*SI == UserParent) {
12666 UserIsSuccessor = true;
12670 // If the user is one of our immediate successors, and if that successor
12671 // only has us as a predecessors (we'd have to split the critical edge
12672 // otherwise), we can keep going.
12673 if (UserIsSuccessor && !isa<PHINode>(I->use_back()) &&
12674 next(pred_begin(UserParent)) == pred_end(UserParent))
12675 // Okay, the CFG is simple enough, try to sink this instruction.
12676 Changed |= TryToSinkInstruction(I, UserParent);
12680 // Now that we have an instruction, try combining it to simplify it...
12684 DEBUG(std::ostringstream SS; I->print(SS); OrigI = SS.str(););
12685 if (Instruction *Result = visit(*I)) {
12687 // Should we replace the old instruction with a new one?
12689 DOUT << "IC: Old = " << *I
12690 << " New = " << *Result;
12692 // Everything uses the new instruction now.
12693 I->replaceAllUsesWith(Result);
12695 // Push the new instruction and any users onto the worklist.
12696 AddToWorkList(Result);
12697 AddUsersToWorkList(*Result);
12699 // Move the name to the new instruction first.
12700 Result->takeName(I);
12702 // Insert the new instruction into the basic block...
12703 BasicBlock *InstParent = I->getParent();
12704 BasicBlock::iterator InsertPos = I;
12706 if (!isa<PHINode>(Result)) // If combining a PHI, don't insert
12707 while (isa<PHINode>(InsertPos)) // middle of a block of PHIs.
12710 InstParent->getInstList().insert(InsertPos, Result);
12712 // Make sure that we reprocess all operands now that we reduced their
12714 AddUsesToWorkList(*I);
12716 // Instructions can end up on the worklist more than once. Make sure
12717 // we do not process an instruction that has been deleted.
12718 RemoveFromWorkList(I);
12720 // Erase the old instruction.
12721 InstParent->getInstList().erase(I);
12724 DOUT << "IC: Mod = " << OrigI
12725 << " New = " << *I;
12728 // If the instruction was modified, it's possible that it is now dead.
12729 // if so, remove it.
12730 if (isInstructionTriviallyDead(I)) {
12731 // Make sure we process all operands now that we are reducing their
12733 AddUsesToWorkList(*I);
12735 // Instructions may end up in the worklist more than once. Erase all
12736 // occurrences of this instruction.
12737 RemoveFromWorkList(I);
12738 I->eraseFromParent();
12741 AddUsersToWorkList(*I);
12748 assert(WorklistMap.empty() && "Worklist empty, but map not?");
12750 // Do an explicit clear, this shrinks the map if needed.
12751 WorklistMap.clear();
12756 bool InstCombiner::runOnFunction(Function &F) {
12757 MustPreserveLCSSA = mustPreserveAnalysisID(LCSSAID);
12759 bool EverMadeChange = false;
12761 // Iterate while there is work to do.
12762 unsigned Iteration = 0;
12763 while (DoOneIteration(F, Iteration++))
12764 EverMadeChange = true;
12765 return EverMadeChange;
12768 FunctionPass *llvm::createInstructionCombiningPass() {
12769 return new InstCombiner();