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 *visitOr (BinaryOperator &I);
187 Instruction *visitXor(BinaryOperator &I);
188 Instruction *visitShl(BinaryOperator &I);
189 Instruction *visitAShr(BinaryOperator &I);
190 Instruction *visitLShr(BinaryOperator &I);
191 Instruction *commonShiftTransforms(BinaryOperator &I);
192 Instruction *FoldFCmp_IntToFP_Cst(FCmpInst &I, Instruction *LHSI,
194 Instruction *visitFCmpInst(FCmpInst &I);
195 Instruction *visitICmpInst(ICmpInst &I);
196 Instruction *visitICmpInstWithCastAndCast(ICmpInst &ICI);
197 Instruction *visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
200 Instruction *FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
201 ConstantInt *DivRHS);
203 Instruction *FoldGEPICmp(User *GEPLHS, Value *RHS,
204 ICmpInst::Predicate Cond, Instruction &I);
205 Instruction *FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
207 Instruction *commonCastTransforms(CastInst &CI);
208 Instruction *commonIntCastTransforms(CastInst &CI);
209 Instruction *commonPointerCastTransforms(CastInst &CI);
210 Instruction *visitTrunc(TruncInst &CI);
211 Instruction *visitZExt(ZExtInst &CI);
212 Instruction *visitSExt(SExtInst &CI);
213 Instruction *visitFPTrunc(FPTruncInst &CI);
214 Instruction *visitFPExt(CastInst &CI);
215 Instruction *visitFPToUI(FPToUIInst &FI);
216 Instruction *visitFPToSI(FPToSIInst &FI);
217 Instruction *visitUIToFP(CastInst &CI);
218 Instruction *visitSIToFP(CastInst &CI);
219 Instruction *visitPtrToInt(CastInst &CI);
220 Instruction *visitIntToPtr(IntToPtrInst &CI);
221 Instruction *visitBitCast(BitCastInst &CI);
222 Instruction *FoldSelectOpOp(SelectInst &SI, Instruction *TI,
224 Instruction *visitSelectInst(SelectInst &SI);
225 Instruction *visitSelectInstWithICmp(SelectInst &SI, ICmpInst *ICI);
226 Instruction *visitCallInst(CallInst &CI);
227 Instruction *visitInvokeInst(InvokeInst &II);
228 Instruction *visitPHINode(PHINode &PN);
229 Instruction *visitGetElementPtrInst(GetElementPtrInst &GEP);
230 Instruction *visitAllocationInst(AllocationInst &AI);
231 Instruction *visitFreeInst(FreeInst &FI);
232 Instruction *visitLoadInst(LoadInst &LI);
233 Instruction *visitStoreInst(StoreInst &SI);
234 Instruction *visitBranchInst(BranchInst &BI);
235 Instruction *visitSwitchInst(SwitchInst &SI);
236 Instruction *visitInsertElementInst(InsertElementInst &IE);
237 Instruction *visitExtractElementInst(ExtractElementInst &EI);
238 Instruction *visitShuffleVectorInst(ShuffleVectorInst &SVI);
239 Instruction *visitExtractValueInst(ExtractValueInst &EV);
241 // visitInstruction - Specify what to return for unhandled instructions...
242 Instruction *visitInstruction(Instruction &I) { return 0; }
245 Instruction *visitCallSite(CallSite CS);
246 bool transformConstExprCastCall(CallSite CS);
247 Instruction *transformCallThroughTrampoline(CallSite CS);
248 Instruction *transformZExtICmp(ICmpInst *ICI, Instruction &CI,
249 bool DoXform = true);
250 bool WillNotOverflowSignedAdd(Value *LHS, Value *RHS);
253 // InsertNewInstBefore - insert an instruction New before instruction Old
254 // in the program. Add the new instruction to the worklist.
256 Instruction *InsertNewInstBefore(Instruction *New, Instruction &Old) {
257 assert(New && New->getParent() == 0 &&
258 "New instruction already inserted into a basic block!");
259 BasicBlock *BB = Old.getParent();
260 BB->getInstList().insert(&Old, New); // Insert inst
265 /// InsertCastBefore - Insert a cast of V to TY before the instruction POS.
266 /// This also adds the cast to the worklist. Finally, this returns the
268 Value *InsertCastBefore(Instruction::CastOps opc, Value *V, const Type *Ty,
270 if (V->getType() == Ty) return V;
272 if (Constant *CV = dyn_cast<Constant>(V))
273 return ConstantExpr::getCast(opc, CV, Ty);
275 Instruction *C = CastInst::Create(opc, V, Ty, V->getName(), &Pos);
280 Value *InsertBitCastBefore(Value *V, const Type *Ty, Instruction &Pos) {
281 return InsertCastBefore(Instruction::BitCast, V, Ty, Pos);
285 // ReplaceInstUsesWith - This method is to be used when an instruction is
286 // found to be dead, replacable with another preexisting expression. Here
287 // we add all uses of I to the worklist, replace all uses of I with the new
288 // value, then return I, so that the inst combiner will know that I was
291 Instruction *ReplaceInstUsesWith(Instruction &I, Value *V) {
292 AddUsersToWorkList(I); // Add all modified instrs to worklist
294 I.replaceAllUsesWith(V);
297 // If we are replacing the instruction with itself, this must be in a
298 // segment of unreachable code, so just clobber the instruction.
299 I.replaceAllUsesWith(UndefValue::get(I.getType()));
304 // UpdateValueUsesWith - This method is to be used when an value is
305 // found to be replacable with another preexisting expression or was
306 // updated. Here we add all uses of I to the worklist, replace all uses of
307 // I with the new value (unless the instruction was just updated), then
308 // return true, so that the inst combiner will know that I was modified.
310 bool UpdateValueUsesWith(Value *Old, Value *New) {
311 AddUsersToWorkList(*Old); // Add all modified instrs to worklist
313 Old->replaceAllUsesWith(New);
314 if (Instruction *I = dyn_cast<Instruction>(Old))
316 if (Instruction *I = dyn_cast<Instruction>(New))
321 // EraseInstFromFunction - When dealing with an instruction that has side
322 // effects or produces a void value, we can't rely on DCE to delete the
323 // instruction. Instead, visit methods should return the value returned by
325 Instruction *EraseInstFromFunction(Instruction &I) {
326 assert(I.use_empty() && "Cannot erase instruction that is used!");
327 AddUsesToWorkList(I);
328 RemoveFromWorkList(&I);
330 return 0; // Don't do anything with FI
333 void ComputeMaskedBits(Value *V, const APInt &Mask, APInt &KnownZero,
334 APInt &KnownOne, unsigned Depth = 0) const {
335 return llvm::ComputeMaskedBits(V, Mask, KnownZero, KnownOne, TD, Depth);
338 bool MaskedValueIsZero(Value *V, const APInt &Mask,
339 unsigned Depth = 0) const {
340 return llvm::MaskedValueIsZero(V, Mask, TD, Depth);
342 unsigned ComputeNumSignBits(Value *Op, unsigned Depth = 0) const {
343 return llvm::ComputeNumSignBits(Op, TD, Depth);
348 /// SimplifyCommutative - This performs a few simplifications for
349 /// commutative operators.
350 bool SimplifyCommutative(BinaryOperator &I);
352 /// SimplifyCompare - This reorders the operands of a CmpInst to get them in
353 /// most-complex to least-complex order.
354 bool SimplifyCompare(CmpInst &I);
356 /// SimplifyDemandedBits - Attempts to replace V with a simpler value based
357 /// on the demanded bits.
358 bool SimplifyDemandedBits(Value *V, APInt DemandedMask,
359 APInt& KnownZero, APInt& KnownOne,
362 Value *SimplifyDemandedVectorElts(Value *V, uint64_t DemandedElts,
363 uint64_t &UndefElts, unsigned Depth = 0);
365 // FoldOpIntoPhi - Given a binary operator or cast instruction which has a
366 // PHI node as operand #0, see if we can fold the instruction into the PHI
367 // (which is only possible if all operands to the PHI are constants).
368 Instruction *FoldOpIntoPhi(Instruction &I);
370 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
371 // operator and they all are only used by the PHI, PHI together their
372 // inputs, and do the operation once, to the result of the PHI.
373 Instruction *FoldPHIArgOpIntoPHI(PHINode &PN);
374 Instruction *FoldPHIArgBinOpIntoPHI(PHINode &PN);
375 Instruction *FoldPHIArgGEPIntoPHI(PHINode &PN);
378 Instruction *OptAndOp(Instruction *Op, ConstantInt *OpRHS,
379 ConstantInt *AndRHS, BinaryOperator &TheAnd);
381 Value *FoldLogicalPlusAnd(Value *LHS, Value *RHS, ConstantInt *Mask,
382 bool isSub, Instruction &I);
383 Instruction *InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
384 bool isSigned, bool Inside, Instruction &IB);
385 Instruction *PromoteCastOfAllocation(BitCastInst &CI, AllocationInst &AI);
386 Instruction *MatchBSwap(BinaryOperator &I);
387 bool SimplifyStoreAtEndOfBlock(StoreInst &SI);
388 Instruction *SimplifyMemTransfer(MemIntrinsic *MI);
389 Instruction *SimplifyMemSet(MemSetInst *MI);
392 Value *EvaluateInDifferentType(Value *V, const Type *Ty, bool isSigned);
394 bool CanEvaluateInDifferentType(Value *V, const IntegerType *Ty,
396 int &NumCastsRemoved);
397 unsigned GetOrEnforceKnownAlignment(Value *V,
398 unsigned PrefAlign = 0);
403 char InstCombiner::ID = 0;
404 static RegisterPass<InstCombiner>
405 X("instcombine", "Combine redundant instructions");
407 // getComplexity: Assign a complexity or rank value to LLVM Values...
408 // 0 -> undef, 1 -> Const, 2 -> Other, 3 -> Arg, 3 -> Unary, 4 -> OtherInst
409 static unsigned getComplexity(Value *V) {
410 if (isa<Instruction>(V)) {
411 if (BinaryOperator::isNeg(V) || BinaryOperator::isNot(V))
415 if (isa<Argument>(V)) return 3;
416 return isa<Constant>(V) ? (isa<UndefValue>(V) ? 0 : 1) : 2;
419 // isOnlyUse - Return true if this instruction will be deleted if we stop using
421 static bool isOnlyUse(Value *V) {
422 return V->hasOneUse() || isa<Constant>(V);
425 // getPromotedType - Return the specified type promoted as it would be to pass
426 // though a va_arg area...
427 static const Type *getPromotedType(const Type *Ty) {
428 if (const IntegerType* ITy = dyn_cast<IntegerType>(Ty)) {
429 if (ITy->getBitWidth() < 32)
430 return Type::Int32Ty;
435 /// getBitCastOperand - If the specified operand is a CastInst, a constant
436 /// expression bitcast, or a GetElementPtrInst with all zero indices, return the
437 /// operand value, otherwise return null.
438 static Value *getBitCastOperand(Value *V) {
439 if (BitCastInst *I = dyn_cast<BitCastInst>(V))
441 return I->getOperand(0);
442 else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(V)) {
443 // GetElementPtrInst?
444 if (GEP->hasAllZeroIndices())
445 return GEP->getOperand(0);
446 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) {
447 if (CE->getOpcode() == Instruction::BitCast)
448 // BitCast ConstantExp?
449 return CE->getOperand(0);
450 else if (CE->getOpcode() == Instruction::GetElementPtr) {
451 // GetElementPtr ConstantExp?
452 for (User::op_iterator I = CE->op_begin() + 1, E = CE->op_end();
454 ConstantInt *CI = dyn_cast<ConstantInt>(I);
455 if (!CI || !CI->isZero())
456 // Any non-zero indices? Not cast-like.
459 // All-zero indices? This is just like casting.
460 return CE->getOperand(0);
466 /// This function is a wrapper around CastInst::isEliminableCastPair. It
467 /// simply extracts arguments and returns what that function returns.
468 static Instruction::CastOps
469 isEliminableCastPair(
470 const CastInst *CI, ///< The first cast instruction
471 unsigned opcode, ///< The opcode of the second cast instruction
472 const Type *DstTy, ///< The target type for the second cast instruction
473 TargetData *TD ///< The target data for pointer size
476 const Type *SrcTy = CI->getOperand(0)->getType(); // A from above
477 const Type *MidTy = CI->getType(); // B from above
479 // Get the opcodes of the two Cast instructions
480 Instruction::CastOps firstOp = Instruction::CastOps(CI->getOpcode());
481 Instruction::CastOps secondOp = Instruction::CastOps(opcode);
483 return Instruction::CastOps(
484 CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy,
485 DstTy, TD->getIntPtrType()));
488 /// ValueRequiresCast - Return true if the cast from "V to Ty" actually results
489 /// in any code being generated. It does not require codegen if V is simple
490 /// enough or if the cast can be folded into other casts.
491 static bool ValueRequiresCast(Instruction::CastOps opcode, const Value *V,
492 const Type *Ty, TargetData *TD) {
493 if (V->getType() == Ty || isa<Constant>(V)) return false;
495 // If this is another cast that can be eliminated, it isn't codegen either.
496 if (const CastInst *CI = dyn_cast<CastInst>(V))
497 if (isEliminableCastPair(CI, opcode, Ty, TD))
502 // SimplifyCommutative - This performs a few simplifications for commutative
505 // 1. Order operands such that they are listed from right (least complex) to
506 // left (most complex). This puts constants before unary operators before
509 // 2. Transform: (op (op V, C1), C2) ==> (op V, (op C1, C2))
510 // 3. Transform: (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
512 bool InstCombiner::SimplifyCommutative(BinaryOperator &I) {
513 bool Changed = false;
514 if (getComplexity(I.getOperand(0)) < getComplexity(I.getOperand(1)))
515 Changed = !I.swapOperands();
517 if (!I.isAssociative()) return Changed;
518 Instruction::BinaryOps Opcode = I.getOpcode();
519 if (BinaryOperator *Op = dyn_cast<BinaryOperator>(I.getOperand(0)))
520 if (Op->getOpcode() == Opcode && isa<Constant>(Op->getOperand(1))) {
521 if (isa<Constant>(I.getOperand(1))) {
522 Constant *Folded = ConstantExpr::get(I.getOpcode(),
523 cast<Constant>(I.getOperand(1)),
524 cast<Constant>(Op->getOperand(1)));
525 I.setOperand(0, Op->getOperand(0));
526 I.setOperand(1, Folded);
528 } else if (BinaryOperator *Op1=dyn_cast<BinaryOperator>(I.getOperand(1)))
529 if (Op1->getOpcode() == Opcode && isa<Constant>(Op1->getOperand(1)) &&
530 isOnlyUse(Op) && isOnlyUse(Op1)) {
531 Constant *C1 = cast<Constant>(Op->getOperand(1));
532 Constant *C2 = cast<Constant>(Op1->getOperand(1));
534 // Fold (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
535 Constant *Folded = ConstantExpr::get(I.getOpcode(), C1, C2);
536 Instruction *New = BinaryOperator::Create(Opcode, Op->getOperand(0),
540 I.setOperand(0, New);
541 I.setOperand(1, Folded);
548 /// SimplifyCompare - For a CmpInst this function just orders the operands
549 /// so that theyare listed from right (least complex) to left (most complex).
550 /// This puts constants before unary operators before binary operators.
551 bool InstCombiner::SimplifyCompare(CmpInst &I) {
552 if (getComplexity(I.getOperand(0)) >= getComplexity(I.getOperand(1)))
555 // Compare instructions are not associative so there's nothing else we can do.
559 // dyn_castNegVal - Given a 'sub' instruction, return the RHS of the instruction
560 // if the LHS is a constant zero (which is the 'negate' form).
562 static inline Value *dyn_castNegVal(Value *V) {
563 if (BinaryOperator::isNeg(V))
564 return BinaryOperator::getNegArgument(V);
566 // Constants can be considered to be negated values if they can be folded.
567 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
568 return ConstantExpr::getNeg(C);
570 if (ConstantVector *C = dyn_cast<ConstantVector>(V))
571 if (C->getType()->getElementType()->isInteger())
572 return ConstantExpr::getNeg(C);
577 static inline Value *dyn_castNotVal(Value *V) {
578 if (BinaryOperator::isNot(V))
579 return BinaryOperator::getNotArgument(V);
581 // Constants can be considered to be not'ed values...
582 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
583 return ConstantInt::get(~C->getValue());
587 // dyn_castFoldableMul - If this value is a multiply that can be folded into
588 // other computations (because it has a constant operand), return the
589 // non-constant operand of the multiply, and set CST to point to the multiplier.
590 // Otherwise, return null.
592 static inline Value *dyn_castFoldableMul(Value *V, ConstantInt *&CST) {
593 if (V->hasOneUse() && V->getType()->isInteger())
594 if (Instruction *I = dyn_cast<Instruction>(V)) {
595 if (I->getOpcode() == Instruction::Mul)
596 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1))))
597 return I->getOperand(0);
598 if (I->getOpcode() == Instruction::Shl)
599 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1)))) {
600 // The multiplier is really 1 << CST.
601 uint32_t BitWidth = cast<IntegerType>(V->getType())->getBitWidth();
602 uint32_t CSTVal = CST->getLimitedValue(BitWidth);
603 CST = ConstantInt::get(APInt(BitWidth, 1).shl(CSTVal));
604 return I->getOperand(0);
610 /// dyn_castGetElementPtr - If this is a getelementptr instruction or constant
611 /// expression, return it.
612 static User *dyn_castGetElementPtr(Value *V) {
613 if (isa<GetElementPtrInst>(V)) return cast<User>(V);
614 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
615 if (CE->getOpcode() == Instruction::GetElementPtr)
616 return cast<User>(V);
620 /// getOpcode - If this is an Instruction or a ConstantExpr, return the
621 /// opcode value. Otherwise return UserOp1.
622 static unsigned getOpcode(const Value *V) {
623 if (const Instruction *I = dyn_cast<Instruction>(V))
624 return I->getOpcode();
625 if (const ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
626 return CE->getOpcode();
627 // Use UserOp1 to mean there's no opcode.
628 return Instruction::UserOp1;
631 /// AddOne - Add one to a ConstantInt
632 static ConstantInt *AddOne(ConstantInt *C) {
633 APInt Val(C->getValue());
634 return ConstantInt::get(++Val);
636 /// SubOne - Subtract one from a ConstantInt
637 static ConstantInt *SubOne(ConstantInt *C) {
638 APInt Val(C->getValue());
639 return ConstantInt::get(--Val);
641 /// Add - Add two ConstantInts together
642 static ConstantInt *Add(ConstantInt *C1, ConstantInt *C2) {
643 return ConstantInt::get(C1->getValue() + C2->getValue());
645 /// And - Bitwise AND two ConstantInts together
646 static ConstantInt *And(ConstantInt *C1, ConstantInt *C2) {
647 return ConstantInt::get(C1->getValue() & C2->getValue());
649 /// Subtract - Subtract one ConstantInt from another
650 static ConstantInt *Subtract(ConstantInt *C1, ConstantInt *C2) {
651 return ConstantInt::get(C1->getValue() - C2->getValue());
653 /// Multiply - Multiply two ConstantInts together
654 static ConstantInt *Multiply(ConstantInt *C1, ConstantInt *C2) {
655 return ConstantInt::get(C1->getValue() * C2->getValue());
657 /// MultiplyOverflows - True if the multiply can not be expressed in an int
659 static bool MultiplyOverflows(ConstantInt *C1, ConstantInt *C2, bool sign) {
660 uint32_t W = C1->getBitWidth();
661 APInt LHSExt = C1->getValue(), RHSExt = C2->getValue();
670 APInt MulExt = LHSExt * RHSExt;
673 APInt Min = APInt::getSignedMinValue(W).sext(W * 2);
674 APInt Max = APInt::getSignedMaxValue(W).sext(W * 2);
675 return MulExt.slt(Min) || MulExt.sgt(Max);
677 return MulExt.ugt(APInt::getLowBitsSet(W * 2, W));
681 /// ShrinkDemandedConstant - Check to see if the specified operand of the
682 /// specified instruction is a constant integer. If so, check to see if there
683 /// are any bits set in the constant that are not demanded. If so, shrink the
684 /// constant and return true.
685 static bool ShrinkDemandedConstant(Instruction *I, unsigned OpNo,
687 assert(I && "No instruction?");
688 assert(OpNo < I->getNumOperands() && "Operand index too large");
690 // If the operand is not a constant integer, nothing to do.
691 ConstantInt *OpC = dyn_cast<ConstantInt>(I->getOperand(OpNo));
692 if (!OpC) return false;
694 // If there are no bits set that aren't demanded, nothing to do.
695 Demanded.zextOrTrunc(OpC->getValue().getBitWidth());
696 if ((~Demanded & OpC->getValue()) == 0)
699 // This instruction is producing bits that are not demanded. Shrink the RHS.
700 Demanded &= OpC->getValue();
701 I->setOperand(OpNo, ConstantInt::get(Demanded));
705 // ComputeSignedMinMaxValuesFromKnownBits - Given a signed integer type and a
706 // set of known zero and one bits, compute the maximum and minimum values that
707 // could have the specified known zero and known one bits, returning them in
709 static void ComputeSignedMinMaxValuesFromKnownBits(const Type *Ty,
710 const APInt& KnownZero,
711 const APInt& KnownOne,
712 APInt& Min, APInt& Max) {
713 uint32_t BitWidth = cast<IntegerType>(Ty)->getBitWidth();
714 assert(KnownZero.getBitWidth() == BitWidth &&
715 KnownOne.getBitWidth() == BitWidth &&
716 Min.getBitWidth() == BitWidth && Max.getBitWidth() == BitWidth &&
717 "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth.");
718 APInt UnknownBits = ~(KnownZero|KnownOne);
720 // The minimum value is when all unknown bits are zeros, EXCEPT for the sign
721 // bit if it is unknown.
723 Max = KnownOne|UnknownBits;
725 if (UnknownBits[BitWidth-1]) { // Sign bit is unknown
727 Max.clear(BitWidth-1);
731 // ComputeUnsignedMinMaxValuesFromKnownBits - Given an unsigned integer type and
732 // a set of known zero and one bits, compute the maximum and minimum values that
733 // could have the specified known zero and known one bits, returning them in
735 static void ComputeUnsignedMinMaxValuesFromKnownBits(const Type *Ty,
736 const APInt &KnownZero,
737 const APInt &KnownOne,
738 APInt &Min, APInt &Max) {
739 uint32_t BitWidth = cast<IntegerType>(Ty)->getBitWidth(); BitWidth = BitWidth;
740 assert(KnownZero.getBitWidth() == BitWidth &&
741 KnownOne.getBitWidth() == BitWidth &&
742 Min.getBitWidth() == BitWidth && Max.getBitWidth() &&
743 "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth.");
744 APInt UnknownBits = ~(KnownZero|KnownOne);
746 // The minimum value is when the unknown bits are all zeros.
748 // The maximum value is when the unknown bits are all ones.
749 Max = KnownOne|UnknownBits;
752 /// SimplifyDemandedBits - This function attempts to replace V with a simpler
753 /// value based on the demanded bits. When this function is called, it is known
754 /// that only the bits set in DemandedMask of the result of V are ever used
755 /// downstream. Consequently, depending on the mask and V, it may be possible
756 /// to replace V with a constant or one of its operands. In such cases, this
757 /// function does the replacement and returns true. In all other cases, it
758 /// returns false after analyzing the expression and setting KnownOne and known
759 /// to be one in the expression. KnownZero contains all the bits that are known
760 /// to be zero in the expression. These are provided to potentially allow the
761 /// caller (which might recursively be SimplifyDemandedBits itself) to simplify
762 /// the expression. KnownOne and KnownZero always follow the invariant that
763 /// KnownOne & KnownZero == 0. That is, a bit can't be both 1 and 0. Note that
764 /// the bits in KnownOne and KnownZero may only be accurate for those bits set
765 /// in DemandedMask. Note also that the bitwidth of V, DemandedMask, KnownZero
766 /// and KnownOne must all be the same.
767 bool InstCombiner::SimplifyDemandedBits(Value *V, APInt DemandedMask,
768 APInt& KnownZero, APInt& KnownOne,
770 assert(V != 0 && "Null pointer of Value???");
771 assert(Depth <= 6 && "Limit Search Depth");
772 uint32_t BitWidth = DemandedMask.getBitWidth();
773 const IntegerType *VTy = cast<IntegerType>(V->getType());
774 assert(VTy->getBitWidth() == BitWidth &&
775 KnownZero.getBitWidth() == BitWidth &&
776 KnownOne.getBitWidth() == BitWidth &&
777 "Value *V, DemandedMask, KnownZero and KnownOne \
778 must have same BitWidth");
779 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
780 // We know all of the bits for a constant!
781 KnownOne = CI->getValue() & DemandedMask;
782 KnownZero = ~KnownOne & DemandedMask;
788 if (!V->hasOneUse()) { // Other users may use these bits.
789 if (Depth != 0) { // Not at the root.
790 // Just compute the KnownZero/KnownOne bits to simplify things downstream.
791 ComputeMaskedBits(V, DemandedMask, KnownZero, KnownOne, Depth);
794 // If this is the root being simplified, allow it to have multiple uses,
795 // just set the DemandedMask to all bits.
796 DemandedMask = APInt::getAllOnesValue(BitWidth);
797 } else if (DemandedMask == 0) { // Not demanding any bits from V.
798 if (V != UndefValue::get(VTy))
799 return UpdateValueUsesWith(V, UndefValue::get(VTy));
801 } else if (Depth == 6) { // Limit search depth.
805 Instruction *I = dyn_cast<Instruction>(V);
806 if (!I) return false; // Only analyze instructions.
808 APInt LHSKnownZero(BitWidth, 0), LHSKnownOne(BitWidth, 0);
809 APInt &RHSKnownZero = KnownZero, &RHSKnownOne = KnownOne;
810 switch (I->getOpcode()) {
812 ComputeMaskedBits(V, DemandedMask, RHSKnownZero, RHSKnownOne, Depth);
814 case Instruction::And:
815 // If either the LHS or the RHS are Zero, the result is zero.
816 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
817 RHSKnownZero, RHSKnownOne, Depth+1))
819 assert((RHSKnownZero & RHSKnownOne) == 0 &&
820 "Bits known to be one AND zero?");
822 // If something is known zero on the RHS, the bits aren't demanded on the
824 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask & ~RHSKnownZero,
825 LHSKnownZero, LHSKnownOne, Depth+1))
827 assert((LHSKnownZero & LHSKnownOne) == 0 &&
828 "Bits known to be one AND zero?");
830 // If all of the demanded bits are known 1 on one side, return the other.
831 // These bits cannot contribute to the result of the 'and'.
832 if ((DemandedMask & ~LHSKnownZero & RHSKnownOne) ==
833 (DemandedMask & ~LHSKnownZero))
834 return UpdateValueUsesWith(I, I->getOperand(0));
835 if ((DemandedMask & ~RHSKnownZero & LHSKnownOne) ==
836 (DemandedMask & ~RHSKnownZero))
837 return UpdateValueUsesWith(I, I->getOperand(1));
839 // If all of the demanded bits in the inputs are known zeros, return zero.
840 if ((DemandedMask & (RHSKnownZero|LHSKnownZero)) == DemandedMask)
841 return UpdateValueUsesWith(I, Constant::getNullValue(VTy));
843 // If the RHS is a constant, see if we can simplify it.
844 if (ShrinkDemandedConstant(I, 1, DemandedMask & ~LHSKnownZero))
845 return UpdateValueUsesWith(I, I);
847 // Output known-1 bits are only known if set in both the LHS & RHS.
848 RHSKnownOne &= LHSKnownOne;
849 // Output known-0 are known to be clear if zero in either the LHS | RHS.
850 RHSKnownZero |= LHSKnownZero;
852 case Instruction::Or:
853 // If either the LHS or the RHS are One, the result is One.
854 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
855 RHSKnownZero, RHSKnownOne, Depth+1))
857 assert((RHSKnownZero & RHSKnownOne) == 0 &&
858 "Bits known to be one AND zero?");
859 // If something is known one on the RHS, the bits aren't demanded on the
861 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask & ~RHSKnownOne,
862 LHSKnownZero, LHSKnownOne, Depth+1))
864 assert((LHSKnownZero & LHSKnownOne) == 0 &&
865 "Bits known to be one AND zero?");
867 // If all of the demanded bits are known zero on one side, return the other.
868 // These bits cannot contribute to the result of the 'or'.
869 if ((DemandedMask & ~LHSKnownOne & RHSKnownZero) ==
870 (DemandedMask & ~LHSKnownOne))
871 return UpdateValueUsesWith(I, I->getOperand(0));
872 if ((DemandedMask & ~RHSKnownOne & LHSKnownZero) ==
873 (DemandedMask & ~RHSKnownOne))
874 return UpdateValueUsesWith(I, I->getOperand(1));
876 // If all of the potentially set bits on one side are known to be set on
877 // the other side, just use the 'other' side.
878 if ((DemandedMask & (~RHSKnownZero) & LHSKnownOne) ==
879 (DemandedMask & (~RHSKnownZero)))
880 return UpdateValueUsesWith(I, I->getOperand(0));
881 if ((DemandedMask & (~LHSKnownZero) & RHSKnownOne) ==
882 (DemandedMask & (~LHSKnownZero)))
883 return UpdateValueUsesWith(I, I->getOperand(1));
885 // If the RHS is a constant, see if we can simplify it.
886 if (ShrinkDemandedConstant(I, 1, DemandedMask))
887 return UpdateValueUsesWith(I, I);
889 // Output known-0 bits are only known if clear in both the LHS & RHS.
890 RHSKnownZero &= LHSKnownZero;
891 // Output known-1 are known to be set if set in either the LHS | RHS.
892 RHSKnownOne |= LHSKnownOne;
894 case Instruction::Xor: {
895 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
896 RHSKnownZero, RHSKnownOne, Depth+1))
898 assert((RHSKnownZero & RHSKnownOne) == 0 &&
899 "Bits known to be one AND zero?");
900 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
901 LHSKnownZero, LHSKnownOne, Depth+1))
903 assert((LHSKnownZero & LHSKnownOne) == 0 &&
904 "Bits known to be one AND zero?");
906 // If all of the demanded bits are known zero on one side, return the other.
907 // These bits cannot contribute to the result of the 'xor'.
908 if ((DemandedMask & RHSKnownZero) == DemandedMask)
909 return UpdateValueUsesWith(I, I->getOperand(0));
910 if ((DemandedMask & LHSKnownZero) == DemandedMask)
911 return UpdateValueUsesWith(I, I->getOperand(1));
913 // Output known-0 bits are known if clear or set in both the LHS & RHS.
914 APInt KnownZeroOut = (RHSKnownZero & LHSKnownZero) |
915 (RHSKnownOne & LHSKnownOne);
916 // Output known-1 are known to be set if set in only one of the LHS, RHS.
917 APInt KnownOneOut = (RHSKnownZero & LHSKnownOne) |
918 (RHSKnownOne & LHSKnownZero);
920 // If all of the demanded bits are known to be zero on one side or the
921 // other, turn this into an *inclusive* or.
922 // e.g. (A & C1)^(B & C2) -> (A & C1)|(B & C2) iff C1&C2 == 0
923 if ((DemandedMask & ~RHSKnownZero & ~LHSKnownZero) == 0) {
925 BinaryOperator::CreateOr(I->getOperand(0), I->getOperand(1),
927 InsertNewInstBefore(Or, *I);
928 return UpdateValueUsesWith(I, Or);
931 // If all of the demanded bits on one side are known, and all of the set
932 // bits on that side are also known to be set on the other side, turn this
933 // into an AND, as we know the bits will be cleared.
934 // e.g. (X | C1) ^ C2 --> (X | C1) & ~C2 iff (C1&C2) == C2
935 if ((DemandedMask & (RHSKnownZero|RHSKnownOne)) == DemandedMask) {
937 if ((RHSKnownOne & LHSKnownOne) == RHSKnownOne) {
938 Constant *AndC = ConstantInt::get(~RHSKnownOne & DemandedMask);
940 BinaryOperator::CreateAnd(I->getOperand(0), AndC, "tmp");
941 InsertNewInstBefore(And, *I);
942 return UpdateValueUsesWith(I, And);
946 // If the RHS is a constant, see if we can simplify it.
947 // FIXME: for XOR, we prefer to force bits to 1 if they will make a -1.
948 if (ShrinkDemandedConstant(I, 1, DemandedMask))
949 return UpdateValueUsesWith(I, I);
951 RHSKnownZero = KnownZeroOut;
952 RHSKnownOne = KnownOneOut;
955 case Instruction::Select:
956 if (SimplifyDemandedBits(I->getOperand(2), DemandedMask,
957 RHSKnownZero, RHSKnownOne, Depth+1))
959 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
960 LHSKnownZero, LHSKnownOne, Depth+1))
962 assert((RHSKnownZero & RHSKnownOne) == 0 &&
963 "Bits known to be one AND zero?");
964 assert((LHSKnownZero & LHSKnownOne) == 0 &&
965 "Bits known to be one AND zero?");
967 // If the operands are constants, see if we can simplify them.
968 if (ShrinkDemandedConstant(I, 1, DemandedMask))
969 return UpdateValueUsesWith(I, I);
970 if (ShrinkDemandedConstant(I, 2, DemandedMask))
971 return UpdateValueUsesWith(I, I);
973 // Only known if known in both the LHS and RHS.
974 RHSKnownOne &= LHSKnownOne;
975 RHSKnownZero &= LHSKnownZero;
977 case Instruction::Trunc: {
979 cast<IntegerType>(I->getOperand(0)->getType())->getBitWidth();
980 DemandedMask.zext(truncBf);
981 RHSKnownZero.zext(truncBf);
982 RHSKnownOne.zext(truncBf);
983 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
984 RHSKnownZero, RHSKnownOne, Depth+1))
986 DemandedMask.trunc(BitWidth);
987 RHSKnownZero.trunc(BitWidth);
988 RHSKnownOne.trunc(BitWidth);
989 assert((RHSKnownZero & RHSKnownOne) == 0 &&
990 "Bits known to be one AND zero?");
993 case Instruction::BitCast:
994 if (!I->getOperand(0)->getType()->isInteger())
997 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
998 RHSKnownZero, RHSKnownOne, Depth+1))
1000 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1001 "Bits known to be one AND zero?");
1003 case Instruction::ZExt: {
1004 // Compute the bits in the result that are not present in the input.
1005 const IntegerType *SrcTy = cast<IntegerType>(I->getOperand(0)->getType());
1006 uint32_t SrcBitWidth = SrcTy->getBitWidth();
1008 DemandedMask.trunc(SrcBitWidth);
1009 RHSKnownZero.trunc(SrcBitWidth);
1010 RHSKnownOne.trunc(SrcBitWidth);
1011 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1012 RHSKnownZero, RHSKnownOne, Depth+1))
1014 DemandedMask.zext(BitWidth);
1015 RHSKnownZero.zext(BitWidth);
1016 RHSKnownOne.zext(BitWidth);
1017 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1018 "Bits known to be one AND zero?");
1019 // The top bits are known to be zero.
1020 RHSKnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
1023 case Instruction::SExt: {
1024 // Compute the bits in the result that are not present in the input.
1025 const IntegerType *SrcTy = cast<IntegerType>(I->getOperand(0)->getType());
1026 uint32_t SrcBitWidth = SrcTy->getBitWidth();
1028 APInt InputDemandedBits = DemandedMask &
1029 APInt::getLowBitsSet(BitWidth, SrcBitWidth);
1031 APInt NewBits(APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth));
1032 // If any of the sign extended bits are demanded, we know that the sign
1034 if ((NewBits & DemandedMask) != 0)
1035 InputDemandedBits.set(SrcBitWidth-1);
1037 InputDemandedBits.trunc(SrcBitWidth);
1038 RHSKnownZero.trunc(SrcBitWidth);
1039 RHSKnownOne.trunc(SrcBitWidth);
1040 if (SimplifyDemandedBits(I->getOperand(0), InputDemandedBits,
1041 RHSKnownZero, RHSKnownOne, Depth+1))
1043 InputDemandedBits.zext(BitWidth);
1044 RHSKnownZero.zext(BitWidth);
1045 RHSKnownOne.zext(BitWidth);
1046 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1047 "Bits known to be one AND zero?");
1049 // If the sign bit of the input is known set or clear, then we know the
1050 // top bits of the result.
1052 // If the input sign bit is known zero, or if the NewBits are not demanded
1053 // convert this into a zero extension.
1054 if (RHSKnownZero[SrcBitWidth-1] || (NewBits & ~DemandedMask) == NewBits)
1056 // Convert to ZExt cast
1057 CastInst *NewCast = new ZExtInst(I->getOperand(0), VTy, I->getName(), I);
1058 return UpdateValueUsesWith(I, NewCast);
1059 } else if (RHSKnownOne[SrcBitWidth-1]) { // Input sign bit known set
1060 RHSKnownOne |= NewBits;
1064 case Instruction::Add: {
1065 // Figure out what the input bits are. If the top bits of the and result
1066 // are not demanded, then the add doesn't demand them from its input
1068 uint32_t NLZ = DemandedMask.countLeadingZeros();
1070 // If there is a constant on the RHS, there are a variety of xformations
1072 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
1073 // If null, this should be simplified elsewhere. Some of the xforms here
1074 // won't work if the RHS is zero.
1078 // If the top bit of the output is demanded, demand everything from the
1079 // input. Otherwise, we demand all the input bits except NLZ top bits.
1080 APInt InDemandedBits(APInt::getLowBitsSet(BitWidth, BitWidth - NLZ));
1082 // Find information about known zero/one bits in the input.
1083 if (SimplifyDemandedBits(I->getOperand(0), InDemandedBits,
1084 LHSKnownZero, LHSKnownOne, Depth+1))
1087 // If the RHS of the add has bits set that can't affect the input, reduce
1089 if (ShrinkDemandedConstant(I, 1, InDemandedBits))
1090 return UpdateValueUsesWith(I, I);
1092 // Avoid excess work.
1093 if (LHSKnownZero == 0 && LHSKnownOne == 0)
1096 // Turn it into OR if input bits are zero.
1097 if ((LHSKnownZero & RHS->getValue()) == RHS->getValue()) {
1099 BinaryOperator::CreateOr(I->getOperand(0), I->getOperand(1),
1101 InsertNewInstBefore(Or, *I);
1102 return UpdateValueUsesWith(I, Or);
1105 // We can say something about the output known-zero and known-one bits,
1106 // depending on potential carries from the input constant and the
1107 // unknowns. For example if the LHS is known to have at most the 0x0F0F0
1108 // bits set and the RHS constant is 0x01001, then we know we have a known
1109 // one mask of 0x00001 and a known zero mask of 0xE0F0E.
1111 // To compute this, we first compute the potential carry bits. These are
1112 // the bits which may be modified. I'm not aware of a better way to do
1114 const APInt& RHSVal = RHS->getValue();
1115 APInt CarryBits((~LHSKnownZero + RHSVal) ^ (~LHSKnownZero ^ RHSVal));
1117 // Now that we know which bits have carries, compute the known-1/0 sets.
1119 // Bits are known one if they are known zero in one operand and one in the
1120 // other, and there is no input carry.
1121 RHSKnownOne = ((LHSKnownZero & RHSVal) |
1122 (LHSKnownOne & ~RHSVal)) & ~CarryBits;
1124 // Bits are known zero if they are known zero in both operands and there
1125 // is no input carry.
1126 RHSKnownZero = LHSKnownZero & ~RHSVal & ~CarryBits;
1128 // If the high-bits of this ADD are not demanded, then it does not demand
1129 // the high bits of its LHS or RHS.
1130 if (DemandedMask[BitWidth-1] == 0) {
1131 // Right fill the mask of bits for this ADD to demand the most
1132 // significant bit and all those below it.
1133 APInt DemandedFromOps(APInt::getLowBitsSet(BitWidth, BitWidth-NLZ));
1134 if (SimplifyDemandedBits(I->getOperand(0), DemandedFromOps,
1135 LHSKnownZero, LHSKnownOne, Depth+1))
1137 if (SimplifyDemandedBits(I->getOperand(1), DemandedFromOps,
1138 LHSKnownZero, LHSKnownOne, Depth+1))
1144 case Instruction::Sub:
1145 // If the high-bits of this SUB are not demanded, then it does not demand
1146 // the high bits of its LHS or RHS.
1147 if (DemandedMask[BitWidth-1] == 0) {
1148 // Right fill the mask of bits for this SUB to demand the most
1149 // significant bit and all those below it.
1150 uint32_t NLZ = DemandedMask.countLeadingZeros();
1151 APInt DemandedFromOps(APInt::getLowBitsSet(BitWidth, BitWidth-NLZ));
1152 if (SimplifyDemandedBits(I->getOperand(0), DemandedFromOps,
1153 LHSKnownZero, LHSKnownOne, Depth+1))
1155 if (SimplifyDemandedBits(I->getOperand(1), DemandedFromOps,
1156 LHSKnownZero, LHSKnownOne, Depth+1))
1159 // Otherwise just hand the sub off to ComputeMaskedBits to fill in
1160 // the known zeros and ones.
1161 ComputeMaskedBits(V, DemandedMask, RHSKnownZero, RHSKnownOne, Depth);
1163 case Instruction::Shl:
1164 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1165 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
1166 APInt DemandedMaskIn(DemandedMask.lshr(ShiftAmt));
1167 if (SimplifyDemandedBits(I->getOperand(0), DemandedMaskIn,
1168 RHSKnownZero, RHSKnownOne, Depth+1))
1170 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1171 "Bits known to be one AND zero?");
1172 RHSKnownZero <<= ShiftAmt;
1173 RHSKnownOne <<= ShiftAmt;
1174 // low bits known zero.
1176 RHSKnownZero |= APInt::getLowBitsSet(BitWidth, ShiftAmt);
1179 case Instruction::LShr:
1180 // For a logical shift right
1181 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1182 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
1184 // Unsigned shift right.
1185 APInt DemandedMaskIn(DemandedMask.shl(ShiftAmt));
1186 if (SimplifyDemandedBits(I->getOperand(0), DemandedMaskIn,
1187 RHSKnownZero, RHSKnownOne, Depth+1))
1189 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1190 "Bits known to be one AND zero?");
1191 RHSKnownZero = APIntOps::lshr(RHSKnownZero, ShiftAmt);
1192 RHSKnownOne = APIntOps::lshr(RHSKnownOne, ShiftAmt);
1194 // Compute the new bits that are at the top now.
1195 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
1196 RHSKnownZero |= HighBits; // high bits known zero.
1200 case Instruction::AShr:
1201 // If this is an arithmetic shift right and only the low-bit is set, we can
1202 // always convert this into a logical shr, even if the shift amount is
1203 // variable. The low bit of the shift cannot be an input sign bit unless
1204 // the shift amount is >= the size of the datatype, which is undefined.
1205 if (DemandedMask == 1) {
1206 // Perform the logical shift right.
1207 Value *NewVal = BinaryOperator::CreateLShr(
1208 I->getOperand(0), I->getOperand(1), I->getName());
1209 InsertNewInstBefore(cast<Instruction>(NewVal), *I);
1210 return UpdateValueUsesWith(I, NewVal);
1213 // If the sign bit is the only bit demanded by this ashr, then there is no
1214 // need to do it, the shift doesn't change the high bit.
1215 if (DemandedMask.isSignBit())
1216 return UpdateValueUsesWith(I, I->getOperand(0));
1218 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1219 uint32_t ShiftAmt = SA->getLimitedValue(BitWidth);
1221 // Signed shift right.
1222 APInt DemandedMaskIn(DemandedMask.shl(ShiftAmt));
1223 // If any of the "high bits" are demanded, we should set the sign bit as
1225 if (DemandedMask.countLeadingZeros() <= ShiftAmt)
1226 DemandedMaskIn.set(BitWidth-1);
1227 if (SimplifyDemandedBits(I->getOperand(0),
1229 RHSKnownZero, RHSKnownOne, Depth+1))
1231 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1232 "Bits known to be one AND zero?");
1233 // Compute the new bits that are at the top now.
1234 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
1235 RHSKnownZero = APIntOps::lshr(RHSKnownZero, ShiftAmt);
1236 RHSKnownOne = APIntOps::lshr(RHSKnownOne, ShiftAmt);
1238 // Handle the sign bits.
1239 APInt SignBit(APInt::getSignBit(BitWidth));
1240 // Adjust to where it is now in the mask.
1241 SignBit = APIntOps::lshr(SignBit, ShiftAmt);
1243 // If the input sign bit is known to be zero, or if none of the top bits
1244 // are demanded, turn this into an unsigned shift right.
1245 if (BitWidth <= ShiftAmt || RHSKnownZero[BitWidth-ShiftAmt-1] ||
1246 (HighBits & ~DemandedMask) == HighBits) {
1247 // Perform the logical shift right.
1248 Value *NewVal = BinaryOperator::CreateLShr(
1249 I->getOperand(0), SA, I->getName());
1250 InsertNewInstBefore(cast<Instruction>(NewVal), *I);
1251 return UpdateValueUsesWith(I, NewVal);
1252 } else if ((RHSKnownOne & SignBit) != 0) { // New bits are known one.
1253 RHSKnownOne |= HighBits;
1257 case Instruction::SRem:
1258 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
1259 APInt RA = Rem->getValue().abs();
1260 if (RA.isPowerOf2()) {
1261 if (DemandedMask.ule(RA)) // srem won't affect demanded bits
1262 return UpdateValueUsesWith(I, I->getOperand(0));
1264 APInt LowBits = RA - 1;
1265 APInt Mask2 = LowBits | APInt::getSignBit(BitWidth);
1266 if (SimplifyDemandedBits(I->getOperand(0), Mask2,
1267 LHSKnownZero, LHSKnownOne, Depth+1))
1270 if (LHSKnownZero[BitWidth-1] || ((LHSKnownZero & LowBits) == LowBits))
1271 LHSKnownZero |= ~LowBits;
1273 KnownZero |= LHSKnownZero & DemandedMask;
1275 assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?");
1279 case Instruction::URem: {
1280 APInt KnownZero2(BitWidth, 0), KnownOne2(BitWidth, 0);
1281 APInt AllOnes = APInt::getAllOnesValue(BitWidth);
1282 if (SimplifyDemandedBits(I->getOperand(0), AllOnes,
1283 KnownZero2, KnownOne2, Depth+1))
1286 uint32_t Leaders = KnownZero2.countLeadingOnes();
1287 if (SimplifyDemandedBits(I->getOperand(1), AllOnes,
1288 KnownZero2, KnownOne2, Depth+1))
1291 Leaders = std::max(Leaders,
1292 KnownZero2.countLeadingOnes());
1293 KnownZero = APInt::getHighBitsSet(BitWidth, Leaders) & DemandedMask;
1296 case Instruction::Call:
1297 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
1298 switch (II->getIntrinsicID()) {
1300 case Intrinsic::bswap: {
1301 // If the only bits demanded come from one byte of the bswap result,
1302 // just shift the input byte into position to eliminate the bswap.
1303 unsigned NLZ = DemandedMask.countLeadingZeros();
1304 unsigned NTZ = DemandedMask.countTrailingZeros();
1306 // Round NTZ down to the next byte. If we have 11 trailing zeros, then
1307 // we need all the bits down to bit 8. Likewise, round NLZ. If we
1308 // have 14 leading zeros, round to 8.
1311 // If we need exactly one byte, we can do this transformation.
1312 if (BitWidth-NLZ-NTZ == 8) {
1313 unsigned ResultBit = NTZ;
1314 unsigned InputBit = BitWidth-NTZ-8;
1316 // Replace this with either a left or right shift to get the byte into
1318 Instruction *NewVal;
1319 if (InputBit > ResultBit)
1320 NewVal = BinaryOperator::CreateLShr(I->getOperand(1),
1321 ConstantInt::get(I->getType(), InputBit-ResultBit));
1323 NewVal = BinaryOperator::CreateShl(I->getOperand(1),
1324 ConstantInt::get(I->getType(), ResultBit-InputBit));
1325 NewVal->takeName(I);
1326 InsertNewInstBefore(NewVal, *I);
1327 return UpdateValueUsesWith(I, NewVal);
1330 // TODO: Could compute known zero/one bits based on the input.
1335 ComputeMaskedBits(V, DemandedMask, RHSKnownZero, RHSKnownOne, Depth);
1339 // If the client is only demanding bits that we know, return the known
1341 if ((DemandedMask & (RHSKnownZero|RHSKnownOne)) == DemandedMask)
1342 return UpdateValueUsesWith(I, ConstantInt::get(RHSKnownOne));
1347 /// SimplifyDemandedVectorElts - The specified value produces a vector with
1348 /// 64 or fewer elements. DemandedElts contains the set of elements that are
1349 /// actually used by the caller. This method analyzes which elements of the
1350 /// operand are undef and returns that information in UndefElts.
1352 /// If the information about demanded elements can be used to simplify the
1353 /// operation, the operation is simplified, then the resultant value is
1354 /// returned. This returns null if no change was made.
1355 Value *InstCombiner::SimplifyDemandedVectorElts(Value *V, uint64_t DemandedElts,
1356 uint64_t &UndefElts,
1358 unsigned VWidth = cast<VectorType>(V->getType())->getNumElements();
1359 assert(VWidth <= 64 && "Vector too wide to analyze!");
1360 uint64_t EltMask = ~0ULL >> (64-VWidth);
1361 assert((DemandedElts & ~EltMask) == 0 && "Invalid DemandedElts!");
1363 if (isa<UndefValue>(V)) {
1364 // If the entire vector is undefined, just return this info.
1365 UndefElts = EltMask;
1367 } else if (DemandedElts == 0) { // If nothing is demanded, provide undef.
1368 UndefElts = EltMask;
1369 return UndefValue::get(V->getType());
1373 if (ConstantVector *CP = dyn_cast<ConstantVector>(V)) {
1374 const Type *EltTy = cast<VectorType>(V->getType())->getElementType();
1375 Constant *Undef = UndefValue::get(EltTy);
1377 std::vector<Constant*> Elts;
1378 for (unsigned i = 0; i != VWidth; ++i)
1379 if (!(DemandedElts & (1ULL << i))) { // If not demanded, set to undef.
1380 Elts.push_back(Undef);
1381 UndefElts |= (1ULL << i);
1382 } else if (isa<UndefValue>(CP->getOperand(i))) { // Already undef.
1383 Elts.push_back(Undef);
1384 UndefElts |= (1ULL << i);
1385 } else { // Otherwise, defined.
1386 Elts.push_back(CP->getOperand(i));
1389 // If we changed the constant, return it.
1390 Constant *NewCP = ConstantVector::get(Elts);
1391 return NewCP != CP ? NewCP : 0;
1392 } else if (isa<ConstantAggregateZero>(V)) {
1393 // Simplify the CAZ to a ConstantVector where the non-demanded elements are
1396 // Check if this is identity. If so, return 0 since we are not simplifying
1398 if (DemandedElts == ((1ULL << VWidth) -1))
1401 const Type *EltTy = cast<VectorType>(V->getType())->getElementType();
1402 Constant *Zero = Constant::getNullValue(EltTy);
1403 Constant *Undef = UndefValue::get(EltTy);
1404 std::vector<Constant*> Elts;
1405 for (unsigned i = 0; i != VWidth; ++i)
1406 Elts.push_back((DemandedElts & (1ULL << i)) ? Zero : Undef);
1407 UndefElts = DemandedElts ^ EltMask;
1408 return ConstantVector::get(Elts);
1411 // Limit search depth.
1415 // If multiple users are using the root value, procede with
1416 // simplification conservatively assuming that all elements
1418 if (!V->hasOneUse()) {
1419 // Quit if we find multiple users of a non-root value though.
1420 // They'll be handled when it's their turn to be visited by
1421 // the main instcombine process.
1423 // TODO: Just compute the UndefElts information recursively.
1426 // Conservatively assume that all elements are needed.
1427 DemandedElts = EltMask;
1430 Instruction *I = dyn_cast<Instruction>(V);
1431 if (!I) return false; // Only analyze instructions.
1433 bool MadeChange = false;
1434 uint64_t UndefElts2;
1436 switch (I->getOpcode()) {
1439 case Instruction::InsertElement: {
1440 // If this is a variable index, we don't know which element it overwrites.
1441 // demand exactly the same input as we produce.
1442 ConstantInt *Idx = dyn_cast<ConstantInt>(I->getOperand(2));
1444 // Note that we can't propagate undef elt info, because we don't know
1445 // which elt is getting updated.
1446 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
1447 UndefElts2, Depth+1);
1448 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1452 // If this is inserting an element that isn't demanded, remove this
1454 unsigned IdxNo = Idx->getZExtValue();
1455 if (IdxNo >= VWidth || (DemandedElts & (1ULL << IdxNo)) == 0)
1456 return AddSoonDeadInstToWorklist(*I, 0);
1458 // Otherwise, the element inserted overwrites whatever was there, so the
1459 // input demanded set is simpler than the output set.
1460 TmpV = SimplifyDemandedVectorElts(I->getOperand(0),
1461 DemandedElts & ~(1ULL << IdxNo),
1462 UndefElts, Depth+1);
1463 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1465 // The inserted element is defined.
1466 UndefElts &= ~(1ULL << IdxNo);
1469 case Instruction::ShuffleVector: {
1470 ShuffleVectorInst *Shuffle = cast<ShuffleVectorInst>(I);
1471 uint64_t LHSVWidth =
1472 cast<VectorType>(Shuffle->getOperand(0)->getType())->getNumElements();
1473 uint64_t LeftDemanded = 0, RightDemanded = 0;
1474 for (unsigned i = 0; i < VWidth; i++) {
1475 if (DemandedElts & (1ULL << i)) {
1476 unsigned MaskVal = Shuffle->getMaskValue(i);
1477 if (MaskVal != -1u) {
1478 assert(MaskVal < LHSVWidth * 2 &&
1479 "shufflevector mask index out of range!");
1480 if (MaskVal < LHSVWidth)
1481 LeftDemanded |= 1ULL << MaskVal;
1483 RightDemanded |= 1ULL << (MaskVal - LHSVWidth);
1488 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), LeftDemanded,
1489 UndefElts2, Depth+1);
1490 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1492 uint64_t UndefElts3;
1493 TmpV = SimplifyDemandedVectorElts(I->getOperand(1), RightDemanded,
1494 UndefElts3, Depth+1);
1495 if (TmpV) { I->setOperand(1, TmpV); MadeChange = true; }
1497 bool NewUndefElts = false;
1498 for (unsigned i = 0; i < VWidth; i++) {
1499 unsigned MaskVal = Shuffle->getMaskValue(i);
1500 if (MaskVal == -1u) {
1501 uint64_t NewBit = 1ULL << i;
1502 UndefElts |= NewBit;
1503 } else if (MaskVal < LHSVWidth) {
1504 uint64_t NewBit = ((UndefElts2 >> MaskVal) & 1) << i;
1505 NewUndefElts |= NewBit;
1506 UndefElts |= NewBit;
1508 uint64_t NewBit = ((UndefElts3 >> (MaskVal - LHSVWidth)) & 1) << i;
1509 NewUndefElts |= NewBit;
1510 UndefElts |= NewBit;
1515 // Add additional discovered undefs.
1516 std::vector<Constant*> Elts;
1517 for (unsigned i = 0; i < VWidth; ++i) {
1518 if (UndefElts & (1ULL << i))
1519 Elts.push_back(UndefValue::get(Type::Int32Ty));
1521 Elts.push_back(ConstantInt::get(Type::Int32Ty,
1522 Shuffle->getMaskValue(i)));
1524 I->setOperand(2, ConstantVector::get(Elts));
1529 case Instruction::BitCast: {
1530 // Vector->vector casts only.
1531 const VectorType *VTy = dyn_cast<VectorType>(I->getOperand(0)->getType());
1533 unsigned InVWidth = VTy->getNumElements();
1534 uint64_t InputDemandedElts = 0;
1537 if (VWidth == InVWidth) {
1538 // If we are converting from <4 x i32> -> <4 x f32>, we demand the same
1539 // elements as are demanded of us.
1541 InputDemandedElts = DemandedElts;
1542 } else if (VWidth > InVWidth) {
1546 // If there are more elements in the result than there are in the source,
1547 // then an input element is live if any of the corresponding output
1548 // elements are live.
1549 Ratio = VWidth/InVWidth;
1550 for (unsigned OutIdx = 0; OutIdx != VWidth; ++OutIdx) {
1551 if (DemandedElts & (1ULL << OutIdx))
1552 InputDemandedElts |= 1ULL << (OutIdx/Ratio);
1558 // If there are more elements in the source than there are in the result,
1559 // then an input element is live if the corresponding output element is
1561 Ratio = InVWidth/VWidth;
1562 for (unsigned InIdx = 0; InIdx != InVWidth; ++InIdx)
1563 if (DemandedElts & (1ULL << InIdx/Ratio))
1564 InputDemandedElts |= 1ULL << InIdx;
1567 // div/rem demand all inputs, because they don't want divide by zero.
1568 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), InputDemandedElts,
1569 UndefElts2, Depth+1);
1571 I->setOperand(0, TmpV);
1575 UndefElts = UndefElts2;
1576 if (VWidth > InVWidth) {
1577 assert(0 && "Unimp");
1578 // If there are more elements in the result than there are in the source,
1579 // then an output element is undef if the corresponding input element is
1581 for (unsigned OutIdx = 0; OutIdx != VWidth; ++OutIdx)
1582 if (UndefElts2 & (1ULL << (OutIdx/Ratio)))
1583 UndefElts |= 1ULL << OutIdx;
1584 } else if (VWidth < InVWidth) {
1585 assert(0 && "Unimp");
1586 // If there are more elements in the source than there are in the result,
1587 // then a result element is undef if all of the corresponding input
1588 // elements are undef.
1589 UndefElts = ~0ULL >> (64-VWidth); // Start out all undef.
1590 for (unsigned InIdx = 0; InIdx != InVWidth; ++InIdx)
1591 if ((UndefElts2 & (1ULL << InIdx)) == 0) // Not undef?
1592 UndefElts &= ~(1ULL << (InIdx/Ratio)); // Clear undef bit.
1596 case Instruction::And:
1597 case Instruction::Or:
1598 case Instruction::Xor:
1599 case Instruction::Add:
1600 case Instruction::Sub:
1601 case Instruction::Mul:
1602 // div/rem demand all inputs, because they don't want divide by zero.
1603 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
1604 UndefElts, Depth+1);
1605 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1606 TmpV = SimplifyDemandedVectorElts(I->getOperand(1), DemandedElts,
1607 UndefElts2, Depth+1);
1608 if (TmpV) { I->setOperand(1, TmpV); MadeChange = true; }
1610 // Output elements are undefined if both are undefined. Consider things
1611 // like undef&0. The result is known zero, not undef.
1612 UndefElts &= UndefElts2;
1615 case Instruction::Call: {
1616 IntrinsicInst *II = dyn_cast<IntrinsicInst>(I);
1618 switch (II->getIntrinsicID()) {
1621 // Binary vector operations that work column-wise. A dest element is a
1622 // function of the corresponding input elements from the two inputs.
1623 case Intrinsic::x86_sse_sub_ss:
1624 case Intrinsic::x86_sse_mul_ss:
1625 case Intrinsic::x86_sse_min_ss:
1626 case Intrinsic::x86_sse_max_ss:
1627 case Intrinsic::x86_sse2_sub_sd:
1628 case Intrinsic::x86_sse2_mul_sd:
1629 case Intrinsic::x86_sse2_min_sd:
1630 case Intrinsic::x86_sse2_max_sd:
1631 TmpV = SimplifyDemandedVectorElts(II->getOperand(1), DemandedElts,
1632 UndefElts, Depth+1);
1633 if (TmpV) { II->setOperand(1, TmpV); MadeChange = true; }
1634 TmpV = SimplifyDemandedVectorElts(II->getOperand(2), DemandedElts,
1635 UndefElts2, Depth+1);
1636 if (TmpV) { II->setOperand(2, TmpV); MadeChange = true; }
1638 // If only the low elt is demanded and this is a scalarizable intrinsic,
1639 // scalarize it now.
1640 if (DemandedElts == 1) {
1641 switch (II->getIntrinsicID()) {
1643 case Intrinsic::x86_sse_sub_ss:
1644 case Intrinsic::x86_sse_mul_ss:
1645 case Intrinsic::x86_sse2_sub_sd:
1646 case Intrinsic::x86_sse2_mul_sd:
1647 // TODO: Lower MIN/MAX/ABS/etc
1648 Value *LHS = II->getOperand(1);
1649 Value *RHS = II->getOperand(2);
1650 // Extract the element as scalars.
1651 LHS = InsertNewInstBefore(new ExtractElementInst(LHS, 0U,"tmp"), *II);
1652 RHS = InsertNewInstBefore(new ExtractElementInst(RHS, 0U,"tmp"), *II);
1654 switch (II->getIntrinsicID()) {
1655 default: assert(0 && "Case stmts out of sync!");
1656 case Intrinsic::x86_sse_sub_ss:
1657 case Intrinsic::x86_sse2_sub_sd:
1658 TmpV = InsertNewInstBefore(BinaryOperator::CreateSub(LHS, RHS,
1659 II->getName()), *II);
1661 case Intrinsic::x86_sse_mul_ss:
1662 case Intrinsic::x86_sse2_mul_sd:
1663 TmpV = InsertNewInstBefore(BinaryOperator::CreateMul(LHS, RHS,
1664 II->getName()), *II);
1669 InsertElementInst::Create(UndefValue::get(II->getType()), TmpV, 0U,
1671 InsertNewInstBefore(New, *II);
1672 AddSoonDeadInstToWorklist(*II, 0);
1677 // Output elements are undefined if both are undefined. Consider things
1678 // like undef&0. The result is known zero, not undef.
1679 UndefElts &= UndefElts2;
1685 return MadeChange ? I : 0;
1689 /// AssociativeOpt - Perform an optimization on an associative operator. This
1690 /// function is designed to check a chain of associative operators for a
1691 /// potential to apply a certain optimization. Since the optimization may be
1692 /// applicable if the expression was reassociated, this checks the chain, then
1693 /// reassociates the expression as necessary to expose the optimization
1694 /// opportunity. This makes use of a special Functor, which must define
1695 /// 'shouldApply' and 'apply' methods.
1697 template<typename Functor>
1698 static Instruction *AssociativeOpt(BinaryOperator &Root, const Functor &F) {
1699 unsigned Opcode = Root.getOpcode();
1700 Value *LHS = Root.getOperand(0);
1702 // Quick check, see if the immediate LHS matches...
1703 if (F.shouldApply(LHS))
1704 return F.apply(Root);
1706 // Otherwise, if the LHS is not of the same opcode as the root, return.
1707 Instruction *LHSI = dyn_cast<Instruction>(LHS);
1708 while (LHSI && LHSI->getOpcode() == Opcode && LHSI->hasOneUse()) {
1709 // Should we apply this transform to the RHS?
1710 bool ShouldApply = F.shouldApply(LHSI->getOperand(1));
1712 // If not to the RHS, check to see if we should apply to the LHS...
1713 if (!ShouldApply && F.shouldApply(LHSI->getOperand(0))) {
1714 cast<BinaryOperator>(LHSI)->swapOperands(); // Make the LHS the RHS
1718 // If the functor wants to apply the optimization to the RHS of LHSI,
1719 // reassociate the expression from ((? op A) op B) to (? op (A op B))
1721 // Now all of the instructions are in the current basic block, go ahead
1722 // and perform the reassociation.
1723 Instruction *TmpLHSI = cast<Instruction>(Root.getOperand(0));
1725 // First move the selected RHS to the LHS of the root...
1726 Root.setOperand(0, LHSI->getOperand(1));
1728 // Make what used to be the LHS of the root be the user of the root...
1729 Value *ExtraOperand = TmpLHSI->getOperand(1);
1730 if (&Root == TmpLHSI) {
1731 Root.replaceAllUsesWith(Constant::getNullValue(TmpLHSI->getType()));
1734 Root.replaceAllUsesWith(TmpLHSI); // Users now use TmpLHSI
1735 TmpLHSI->setOperand(1, &Root); // TmpLHSI now uses the root
1736 BasicBlock::iterator ARI = &Root; ++ARI;
1737 TmpLHSI->moveBefore(ARI); // Move TmpLHSI to after Root
1740 // Now propagate the ExtraOperand down the chain of instructions until we
1742 while (TmpLHSI != LHSI) {
1743 Instruction *NextLHSI = cast<Instruction>(TmpLHSI->getOperand(0));
1744 // Move the instruction to immediately before the chain we are
1745 // constructing to avoid breaking dominance properties.
1746 NextLHSI->moveBefore(ARI);
1749 Value *NextOp = NextLHSI->getOperand(1);
1750 NextLHSI->setOperand(1, ExtraOperand);
1752 ExtraOperand = NextOp;
1755 // Now that the instructions are reassociated, have the functor perform
1756 // the transformation...
1757 return F.apply(Root);
1760 LHSI = dyn_cast<Instruction>(LHSI->getOperand(0));
1767 // AddRHS - Implements: X + X --> X << 1
1770 AddRHS(Value *rhs) : RHS(rhs) {}
1771 bool shouldApply(Value *LHS) const { return LHS == RHS; }
1772 Instruction *apply(BinaryOperator &Add) const {
1773 return BinaryOperator::CreateShl(Add.getOperand(0),
1774 ConstantInt::get(Add.getType(), 1));
1778 // AddMaskingAnd - Implements (A & C1)+(B & C2) --> (A & C1)|(B & C2)
1780 struct AddMaskingAnd {
1782 AddMaskingAnd(Constant *c) : C2(c) {}
1783 bool shouldApply(Value *LHS) const {
1785 return match(LHS, m_And(m_Value(), m_ConstantInt(C1))) &&
1786 ConstantExpr::getAnd(C1, C2)->isNullValue();
1788 Instruction *apply(BinaryOperator &Add) const {
1789 return BinaryOperator::CreateOr(Add.getOperand(0), Add.getOperand(1));
1795 static Value *FoldOperationIntoSelectOperand(Instruction &I, Value *SO,
1797 if (CastInst *CI = dyn_cast<CastInst>(&I)) {
1798 return IC->InsertCastBefore(CI->getOpcode(), SO, I.getType(), I);
1801 // Figure out if the constant is the left or the right argument.
1802 bool ConstIsRHS = isa<Constant>(I.getOperand(1));
1803 Constant *ConstOperand = cast<Constant>(I.getOperand(ConstIsRHS));
1805 if (Constant *SOC = dyn_cast<Constant>(SO)) {
1807 return ConstantExpr::get(I.getOpcode(), SOC, ConstOperand);
1808 return ConstantExpr::get(I.getOpcode(), ConstOperand, SOC);
1811 Value *Op0 = SO, *Op1 = ConstOperand;
1813 std::swap(Op0, Op1);
1815 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
1816 New = BinaryOperator::Create(BO->getOpcode(), Op0, Op1,SO->getName()+".op");
1817 else if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1818 New = CmpInst::Create(CI->getOpcode(), CI->getPredicate(), Op0, Op1,
1819 SO->getName()+".cmp");
1821 assert(0 && "Unknown binary instruction type!");
1824 return IC->InsertNewInstBefore(New, I);
1827 // FoldOpIntoSelect - Given an instruction with a select as one operand and a
1828 // constant as the other operand, try to fold the binary operator into the
1829 // select arguments. This also works for Cast instructions, which obviously do
1830 // not have a second operand.
1831 static Instruction *FoldOpIntoSelect(Instruction &Op, SelectInst *SI,
1833 // Don't modify shared select instructions
1834 if (!SI->hasOneUse()) return 0;
1835 Value *TV = SI->getOperand(1);
1836 Value *FV = SI->getOperand(2);
1838 if (isa<Constant>(TV) || isa<Constant>(FV)) {
1839 // Bool selects with constant operands can be folded to logical ops.
1840 if (SI->getType() == Type::Int1Ty) return 0;
1842 Value *SelectTrueVal = FoldOperationIntoSelectOperand(Op, TV, IC);
1843 Value *SelectFalseVal = FoldOperationIntoSelectOperand(Op, FV, IC);
1845 return SelectInst::Create(SI->getCondition(), SelectTrueVal,
1852 /// FoldOpIntoPhi - Given a binary operator or cast instruction which has a PHI
1853 /// node as operand #0, see if we can fold the instruction into the PHI (which
1854 /// is only possible if all operands to the PHI are constants).
1855 Instruction *InstCombiner::FoldOpIntoPhi(Instruction &I) {
1856 PHINode *PN = cast<PHINode>(I.getOperand(0));
1857 unsigned NumPHIValues = PN->getNumIncomingValues();
1858 if (!PN->hasOneUse() || NumPHIValues == 0) return 0;
1860 // Check to see if all of the operands of the PHI are constants. If there is
1861 // one non-constant value, remember the BB it is. If there is more than one
1862 // or if *it* is a PHI, bail out.
1863 BasicBlock *NonConstBB = 0;
1864 for (unsigned i = 0; i != NumPHIValues; ++i)
1865 if (!isa<Constant>(PN->getIncomingValue(i))) {
1866 if (NonConstBB) return 0; // More than one non-const value.
1867 if (isa<PHINode>(PN->getIncomingValue(i))) return 0; // Itself a phi.
1868 NonConstBB = PN->getIncomingBlock(i);
1870 // If the incoming non-constant value is in I's block, we have an infinite
1872 if (NonConstBB == I.getParent())
1876 // If there is exactly one non-constant value, we can insert a copy of the
1877 // operation in that block. However, if this is a critical edge, we would be
1878 // inserting the computation one some other paths (e.g. inside a loop). Only
1879 // do this if the pred block is unconditionally branching into the phi block.
1881 BranchInst *BI = dyn_cast<BranchInst>(NonConstBB->getTerminator());
1882 if (!BI || !BI->isUnconditional()) return 0;
1885 // Okay, we can do the transformation: create the new PHI node.
1886 PHINode *NewPN = PHINode::Create(I.getType(), "");
1887 NewPN->reserveOperandSpace(PN->getNumOperands()/2);
1888 InsertNewInstBefore(NewPN, *PN);
1889 NewPN->takeName(PN);
1891 // Next, add all of the operands to the PHI.
1892 if (I.getNumOperands() == 2) {
1893 Constant *C = cast<Constant>(I.getOperand(1));
1894 for (unsigned i = 0; i != NumPHIValues; ++i) {
1896 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
1897 if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1898 InV = ConstantExpr::getCompare(CI->getPredicate(), InC, C);
1900 InV = ConstantExpr::get(I.getOpcode(), InC, C);
1902 assert(PN->getIncomingBlock(i) == NonConstBB);
1903 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
1904 InV = BinaryOperator::Create(BO->getOpcode(),
1905 PN->getIncomingValue(i), C, "phitmp",
1906 NonConstBB->getTerminator());
1907 else if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1908 InV = CmpInst::Create(CI->getOpcode(),
1910 PN->getIncomingValue(i), C, "phitmp",
1911 NonConstBB->getTerminator());
1913 assert(0 && "Unknown binop!");
1915 AddToWorkList(cast<Instruction>(InV));
1917 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
1920 CastInst *CI = cast<CastInst>(&I);
1921 const Type *RetTy = CI->getType();
1922 for (unsigned i = 0; i != NumPHIValues; ++i) {
1924 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
1925 InV = ConstantExpr::getCast(CI->getOpcode(), InC, RetTy);
1927 assert(PN->getIncomingBlock(i) == NonConstBB);
1928 InV = CastInst::Create(CI->getOpcode(), PN->getIncomingValue(i),
1929 I.getType(), "phitmp",
1930 NonConstBB->getTerminator());
1931 AddToWorkList(cast<Instruction>(InV));
1933 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
1936 return ReplaceInstUsesWith(I, NewPN);
1940 /// WillNotOverflowSignedAdd - Return true if we can prove that:
1941 /// (sext (add LHS, RHS)) === (add (sext LHS), (sext RHS))
1942 /// This basically requires proving that the add in the original type would not
1943 /// overflow to change the sign bit or have a carry out.
1944 bool InstCombiner::WillNotOverflowSignedAdd(Value *LHS, Value *RHS) {
1945 // There are different heuristics we can use for this. Here are some simple
1948 // Add has the property that adding any two 2's complement numbers can only
1949 // have one carry bit which can change a sign. As such, if LHS and RHS each
1950 // have at least two sign bits, we know that the addition of the two values will
1951 // sign extend fine.
1952 if (ComputeNumSignBits(LHS) > 1 && ComputeNumSignBits(RHS) > 1)
1956 // If one of the operands only has one non-zero bit, and if the other operand
1957 // has a known-zero bit in a more significant place than it (not including the
1958 // sign bit) the ripple may go up to and fill the zero, but won't change the
1959 // sign. For example, (X & ~4) + 1.
1967 Instruction *InstCombiner::visitAdd(BinaryOperator &I) {
1968 bool Changed = SimplifyCommutative(I);
1969 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
1971 if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
1972 // X + undef -> undef
1973 if (isa<UndefValue>(RHS))
1974 return ReplaceInstUsesWith(I, RHS);
1977 if (!I.getType()->isFPOrFPVector()) { // NOTE: -0 + +0 = +0.
1978 if (RHSC->isNullValue())
1979 return ReplaceInstUsesWith(I, LHS);
1980 } else if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHSC)) {
1981 if (CFP->isExactlyValue(ConstantFP::getNegativeZero
1982 (I.getType())->getValueAPF()))
1983 return ReplaceInstUsesWith(I, LHS);
1986 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHSC)) {
1987 // X + (signbit) --> X ^ signbit
1988 const APInt& Val = CI->getValue();
1989 uint32_t BitWidth = Val.getBitWidth();
1990 if (Val == APInt::getSignBit(BitWidth))
1991 return BinaryOperator::CreateXor(LHS, RHS);
1993 // See if SimplifyDemandedBits can simplify this. This handles stuff like
1994 // (X & 254)+1 -> (X&254)|1
1995 if (!isa<VectorType>(I.getType())) {
1996 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
1997 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
1998 KnownZero, KnownOne))
2002 // zext(i1) - 1 -> select i1, 0, -1
2003 if (ZExtInst *ZI = dyn_cast<ZExtInst>(LHS))
2004 if (CI->isAllOnesValue() &&
2005 ZI->getOperand(0)->getType() == Type::Int1Ty)
2006 return SelectInst::Create(ZI->getOperand(0),
2007 Constant::getNullValue(I.getType()),
2008 ConstantInt::getAllOnesValue(I.getType()));
2011 if (isa<PHINode>(LHS))
2012 if (Instruction *NV = FoldOpIntoPhi(I))
2015 ConstantInt *XorRHS = 0;
2017 if (isa<ConstantInt>(RHSC) &&
2018 match(LHS, m_Xor(m_Value(XorLHS), m_ConstantInt(XorRHS)))) {
2019 uint32_t TySizeBits = I.getType()->getPrimitiveSizeInBits();
2020 const APInt& RHSVal = cast<ConstantInt>(RHSC)->getValue();
2022 uint32_t Size = TySizeBits / 2;
2023 APInt C0080Val(APInt(TySizeBits, 1ULL).shl(Size - 1));
2024 APInt CFF80Val(-C0080Val);
2026 if (TySizeBits > Size) {
2027 // If we have ADD(XOR(AND(X, 0xFF), 0x80), 0xF..F80), it's a sext.
2028 // If we have ADD(XOR(AND(X, 0xFF), 0xF..F80), 0x80), it's a sext.
2029 if ((RHSVal == CFF80Val && XorRHS->getValue() == C0080Val) ||
2030 (RHSVal == C0080Val && XorRHS->getValue() == CFF80Val)) {
2031 // This is a sign extend if the top bits are known zero.
2032 if (!MaskedValueIsZero(XorLHS,
2033 APInt::getHighBitsSet(TySizeBits, TySizeBits - Size)))
2034 Size = 0; // Not a sign ext, but can't be any others either.
2039 C0080Val = APIntOps::lshr(C0080Val, Size);
2040 CFF80Val = APIntOps::ashr(CFF80Val, Size);
2041 } while (Size >= 1);
2043 // FIXME: This shouldn't be necessary. When the backends can handle types
2044 // with funny bit widths then this switch statement should be removed. It
2045 // is just here to get the size of the "middle" type back up to something
2046 // that the back ends can handle.
2047 const Type *MiddleType = 0;
2050 case 32: MiddleType = Type::Int32Ty; break;
2051 case 16: MiddleType = Type::Int16Ty; break;
2052 case 8: MiddleType = Type::Int8Ty; break;
2055 Instruction *NewTrunc = new TruncInst(XorLHS, MiddleType, "sext");
2056 InsertNewInstBefore(NewTrunc, I);
2057 return new SExtInst(NewTrunc, I.getType(), I.getName());
2062 if (I.getType() == Type::Int1Ty)
2063 return BinaryOperator::CreateXor(LHS, RHS);
2066 if (I.getType()->isInteger()) {
2067 if (Instruction *Result = AssociativeOpt(I, AddRHS(RHS))) return Result;
2069 if (Instruction *RHSI = dyn_cast<Instruction>(RHS)) {
2070 if (RHSI->getOpcode() == Instruction::Sub)
2071 if (LHS == RHSI->getOperand(1)) // A + (B - A) --> B
2072 return ReplaceInstUsesWith(I, RHSI->getOperand(0));
2074 if (Instruction *LHSI = dyn_cast<Instruction>(LHS)) {
2075 if (LHSI->getOpcode() == Instruction::Sub)
2076 if (RHS == LHSI->getOperand(1)) // (B - A) + A --> B
2077 return ReplaceInstUsesWith(I, LHSI->getOperand(0));
2082 // -A + -B --> -(A + B)
2083 if (Value *LHSV = dyn_castNegVal(LHS)) {
2084 if (LHS->getType()->isIntOrIntVector()) {
2085 if (Value *RHSV = dyn_castNegVal(RHS)) {
2086 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSV, RHSV, "sum");
2087 InsertNewInstBefore(NewAdd, I);
2088 return BinaryOperator::CreateNeg(NewAdd);
2092 return BinaryOperator::CreateSub(RHS, LHSV);
2096 if (!isa<Constant>(RHS))
2097 if (Value *V = dyn_castNegVal(RHS))
2098 return BinaryOperator::CreateSub(LHS, V);
2102 if (Value *X = dyn_castFoldableMul(LHS, C2)) {
2103 if (X == RHS) // X*C + X --> X * (C+1)
2104 return BinaryOperator::CreateMul(RHS, AddOne(C2));
2106 // X*C1 + X*C2 --> X * (C1+C2)
2108 if (X == dyn_castFoldableMul(RHS, C1))
2109 return BinaryOperator::CreateMul(X, Add(C1, C2));
2112 // X + X*C --> X * (C+1)
2113 if (dyn_castFoldableMul(RHS, C2) == LHS)
2114 return BinaryOperator::CreateMul(LHS, AddOne(C2));
2116 // X + ~X --> -1 since ~X = -X-1
2117 if (dyn_castNotVal(LHS) == RHS || dyn_castNotVal(RHS) == LHS)
2118 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
2121 // (A & C1)+(B & C2) --> (A & C1)|(B & C2) iff C1&C2 == 0
2122 if (match(RHS, m_And(m_Value(), m_ConstantInt(C2))))
2123 if (Instruction *R = AssociativeOpt(I, AddMaskingAnd(C2)))
2126 // A+B --> A|B iff A and B have no bits set in common.
2127 if (const IntegerType *IT = dyn_cast<IntegerType>(I.getType())) {
2128 APInt Mask = APInt::getAllOnesValue(IT->getBitWidth());
2129 APInt LHSKnownOne(IT->getBitWidth(), 0);
2130 APInt LHSKnownZero(IT->getBitWidth(), 0);
2131 ComputeMaskedBits(LHS, Mask, LHSKnownZero, LHSKnownOne);
2132 if (LHSKnownZero != 0) {
2133 APInt RHSKnownOne(IT->getBitWidth(), 0);
2134 APInt RHSKnownZero(IT->getBitWidth(), 0);
2135 ComputeMaskedBits(RHS, Mask, RHSKnownZero, RHSKnownOne);
2137 // No bits in common -> bitwise or.
2138 if ((LHSKnownZero|RHSKnownZero).isAllOnesValue())
2139 return BinaryOperator::CreateOr(LHS, RHS);
2143 // W*X + Y*Z --> W * (X+Z) iff W == Y
2144 if (I.getType()->isIntOrIntVector()) {
2145 Value *W, *X, *Y, *Z;
2146 if (match(LHS, m_Mul(m_Value(W), m_Value(X))) &&
2147 match(RHS, m_Mul(m_Value(Y), m_Value(Z)))) {
2151 } else if (Y == X) {
2153 } else if (X == Z) {
2160 Value *NewAdd = InsertNewInstBefore(BinaryOperator::CreateAdd(X, Z,
2161 LHS->getName()), I);
2162 return BinaryOperator::CreateMul(W, NewAdd);
2167 if (ConstantInt *CRHS = dyn_cast<ConstantInt>(RHS)) {
2169 if (match(LHS, m_Not(m_Value(X)))) // ~X + C --> (C-1) - X
2170 return BinaryOperator::CreateSub(SubOne(CRHS), X);
2172 // (X & FF00) + xx00 -> (X+xx00) & FF00
2173 if (LHS->hasOneUse() && match(LHS, m_And(m_Value(X), m_ConstantInt(C2)))) {
2174 Constant *Anded = And(CRHS, C2);
2175 if (Anded == CRHS) {
2176 // See if all bits from the first bit set in the Add RHS up are included
2177 // in the mask. First, get the rightmost bit.
2178 const APInt& AddRHSV = CRHS->getValue();
2180 // Form a mask of all bits from the lowest bit added through the top.
2181 APInt AddRHSHighBits(~((AddRHSV & -AddRHSV)-1));
2183 // See if the and mask includes all of these bits.
2184 APInt AddRHSHighBitsAnd(AddRHSHighBits & C2->getValue());
2186 if (AddRHSHighBits == AddRHSHighBitsAnd) {
2187 // Okay, the xform is safe. Insert the new add pronto.
2188 Value *NewAdd = InsertNewInstBefore(BinaryOperator::CreateAdd(X, CRHS,
2189 LHS->getName()), I);
2190 return BinaryOperator::CreateAnd(NewAdd, C2);
2195 // Try to fold constant add into select arguments.
2196 if (SelectInst *SI = dyn_cast<SelectInst>(LHS))
2197 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2201 // add (cast *A to intptrtype) B ->
2202 // cast (GEP (cast *A to sbyte*) B) --> intptrtype
2204 CastInst *CI = dyn_cast<CastInst>(LHS);
2207 CI = dyn_cast<CastInst>(RHS);
2210 if (CI && CI->getType()->isSized() &&
2211 (CI->getType()->getPrimitiveSizeInBits() ==
2212 TD->getIntPtrType()->getPrimitiveSizeInBits())
2213 && isa<PointerType>(CI->getOperand(0)->getType())) {
2215 cast<PointerType>(CI->getOperand(0)->getType())->getAddressSpace();
2216 Value *I2 = InsertBitCastBefore(CI->getOperand(0),
2217 PointerType::get(Type::Int8Ty, AS), I);
2218 I2 = InsertNewInstBefore(GetElementPtrInst::Create(I2, Other, "ctg2"), I);
2219 return new PtrToIntInst(I2, CI->getType());
2223 // add (select X 0 (sub n A)) A --> select X A n
2225 SelectInst *SI = dyn_cast<SelectInst>(LHS);
2228 SI = dyn_cast<SelectInst>(RHS);
2231 if (SI && SI->hasOneUse()) {
2232 Value *TV = SI->getTrueValue();
2233 Value *FV = SI->getFalseValue();
2236 // Can we fold the add into the argument of the select?
2237 // We check both true and false select arguments for a matching subtract.
2238 if (match(FV, m_Zero()) && match(TV, m_Sub(m_Value(N), m_Specific(A))))
2239 // Fold the add into the true select value.
2240 return SelectInst::Create(SI->getCondition(), N, A);
2241 if (match(TV, m_Zero()) && match(FV, m_Sub(m_Value(N), m_Specific(A))))
2242 // Fold the add into the false select value.
2243 return SelectInst::Create(SI->getCondition(), A, N);
2247 // Check for X+0.0. Simplify it to X if we know X is not -0.0.
2248 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHS))
2249 if (CFP->getValueAPF().isPosZero() && CannotBeNegativeZero(LHS))
2250 return ReplaceInstUsesWith(I, LHS);
2252 // Check for (add (sext x), y), see if we can merge this into an
2253 // integer add followed by a sext.
2254 if (SExtInst *LHSConv = dyn_cast<SExtInst>(LHS)) {
2255 // (add (sext x), cst) --> (sext (add x, cst'))
2256 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS)) {
2258 ConstantExpr::getTrunc(RHSC, LHSConv->getOperand(0)->getType());
2259 if (LHSConv->hasOneUse() &&
2260 ConstantExpr::getSExt(CI, I.getType()) == RHSC &&
2261 WillNotOverflowSignedAdd(LHSConv->getOperand(0), CI)) {
2262 // Insert the new, smaller add.
2263 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSConv->getOperand(0),
2265 InsertNewInstBefore(NewAdd, I);
2266 return new SExtInst(NewAdd, I.getType());
2270 // (add (sext x), (sext y)) --> (sext (add int x, y))
2271 if (SExtInst *RHSConv = dyn_cast<SExtInst>(RHS)) {
2272 // Only do this if x/y have the same type, if at last one of them has a
2273 // single use (so we don't increase the number of sexts), and if the
2274 // integer add will not overflow.
2275 if (LHSConv->getOperand(0)->getType()==RHSConv->getOperand(0)->getType()&&
2276 (LHSConv->hasOneUse() || RHSConv->hasOneUse()) &&
2277 WillNotOverflowSignedAdd(LHSConv->getOperand(0),
2278 RHSConv->getOperand(0))) {
2279 // Insert the new integer add.
2280 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSConv->getOperand(0),
2281 RHSConv->getOperand(0),
2283 InsertNewInstBefore(NewAdd, I);
2284 return new SExtInst(NewAdd, I.getType());
2289 // Check for (add double (sitofp x), y), see if we can merge this into an
2290 // integer add followed by a promotion.
2291 if (SIToFPInst *LHSConv = dyn_cast<SIToFPInst>(LHS)) {
2292 // (add double (sitofp x), fpcst) --> (sitofp (add int x, intcst))
2293 // ... if the constant fits in the integer value. This is useful for things
2294 // like (double)(x & 1234) + 4.0 -> (double)((X & 1234)+4) which no longer
2295 // requires a constant pool load, and generally allows the add to be better
2297 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHS)) {
2299 ConstantExpr::getFPToSI(CFP, LHSConv->getOperand(0)->getType());
2300 if (LHSConv->hasOneUse() &&
2301 ConstantExpr::getSIToFP(CI, I.getType()) == CFP &&
2302 WillNotOverflowSignedAdd(LHSConv->getOperand(0), CI)) {
2303 // Insert the new integer add.
2304 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSConv->getOperand(0),
2306 InsertNewInstBefore(NewAdd, I);
2307 return new SIToFPInst(NewAdd, I.getType());
2311 // (add double (sitofp x), (sitofp y)) --> (sitofp (add int x, y))
2312 if (SIToFPInst *RHSConv = dyn_cast<SIToFPInst>(RHS)) {
2313 // Only do this if x/y have the same type, if at last one of them has a
2314 // single use (so we don't increase the number of int->fp conversions),
2315 // and if the integer add will not overflow.
2316 if (LHSConv->getOperand(0)->getType()==RHSConv->getOperand(0)->getType()&&
2317 (LHSConv->hasOneUse() || RHSConv->hasOneUse()) &&
2318 WillNotOverflowSignedAdd(LHSConv->getOperand(0),
2319 RHSConv->getOperand(0))) {
2320 // Insert the new integer add.
2321 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSConv->getOperand(0),
2322 RHSConv->getOperand(0),
2324 InsertNewInstBefore(NewAdd, I);
2325 return new SIToFPInst(NewAdd, I.getType());
2330 return Changed ? &I : 0;
2333 Instruction *InstCombiner::visitSub(BinaryOperator &I) {
2334 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2336 if (Op0 == Op1 && // sub X, X -> 0
2337 !I.getType()->isFPOrFPVector())
2338 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2340 // If this is a 'B = x-(-A)', change to B = x+A...
2341 if (Value *V = dyn_castNegVal(Op1))
2342 return BinaryOperator::CreateAdd(Op0, V);
2344 if (isa<UndefValue>(Op0))
2345 return ReplaceInstUsesWith(I, Op0); // undef - X -> undef
2346 if (isa<UndefValue>(Op1))
2347 return ReplaceInstUsesWith(I, Op1); // X - undef -> undef
2349 if (ConstantInt *C = dyn_cast<ConstantInt>(Op0)) {
2350 // Replace (-1 - A) with (~A)...
2351 if (C->isAllOnesValue())
2352 return BinaryOperator::CreateNot(Op1);
2354 // C - ~X == X + (1+C)
2356 if (match(Op1, m_Not(m_Value(X))))
2357 return BinaryOperator::CreateAdd(X, AddOne(C));
2359 // -(X >>u 31) -> (X >>s 31)
2360 // -(X >>s 31) -> (X >>u 31)
2362 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op1)) {
2363 if (SI->getOpcode() == Instruction::LShr) {
2364 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
2365 // Check to see if we are shifting out everything but the sign bit.
2366 if (CU->getLimitedValue(SI->getType()->getPrimitiveSizeInBits()) ==
2367 SI->getType()->getPrimitiveSizeInBits()-1) {
2368 // Ok, the transformation is safe. Insert AShr.
2369 return BinaryOperator::Create(Instruction::AShr,
2370 SI->getOperand(0), CU, SI->getName());
2374 else if (SI->getOpcode() == Instruction::AShr) {
2375 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
2376 // Check to see if we are shifting out everything but the sign bit.
2377 if (CU->getLimitedValue(SI->getType()->getPrimitiveSizeInBits()) ==
2378 SI->getType()->getPrimitiveSizeInBits()-1) {
2379 // Ok, the transformation is safe. Insert LShr.
2380 return BinaryOperator::CreateLShr(
2381 SI->getOperand(0), CU, SI->getName());
2388 // Try to fold constant sub into select arguments.
2389 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
2390 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2394 if (I.getType() == Type::Int1Ty)
2395 return BinaryOperator::CreateXor(Op0, Op1);
2397 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
2398 if (Op1I->getOpcode() == Instruction::Add &&
2399 !Op0->getType()->isFPOrFPVector()) {
2400 if (Op1I->getOperand(0) == Op0) // X-(X+Y) == -Y
2401 return BinaryOperator::CreateNeg(Op1I->getOperand(1), I.getName());
2402 else if (Op1I->getOperand(1) == Op0) // X-(Y+X) == -Y
2403 return BinaryOperator::CreateNeg(Op1I->getOperand(0), I.getName());
2404 else if (ConstantInt *CI1 = dyn_cast<ConstantInt>(I.getOperand(0))) {
2405 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(Op1I->getOperand(1)))
2406 // C1-(X+C2) --> (C1-C2)-X
2407 return BinaryOperator::CreateSub(Subtract(CI1, CI2),
2408 Op1I->getOperand(0));
2412 if (Op1I->hasOneUse()) {
2413 // Replace (x - (y - z)) with (x + (z - y)) if the (y - z) subexpression
2414 // is not used by anyone else...
2416 if (Op1I->getOpcode() == Instruction::Sub &&
2417 !Op1I->getType()->isFPOrFPVector()) {
2418 // Swap the two operands of the subexpr...
2419 Value *IIOp0 = Op1I->getOperand(0), *IIOp1 = Op1I->getOperand(1);
2420 Op1I->setOperand(0, IIOp1);
2421 Op1I->setOperand(1, IIOp0);
2423 // Create the new top level add instruction...
2424 return BinaryOperator::CreateAdd(Op0, Op1);
2427 // Replace (A - (A & B)) with (A & ~B) if this is the only use of (A&B)...
2429 if (Op1I->getOpcode() == Instruction::And &&
2430 (Op1I->getOperand(0) == Op0 || Op1I->getOperand(1) == Op0)) {
2431 Value *OtherOp = Op1I->getOperand(Op1I->getOperand(0) == Op0);
2434 InsertNewInstBefore(BinaryOperator::CreateNot(OtherOp, "B.not"), I);
2435 return BinaryOperator::CreateAnd(Op0, NewNot);
2438 // 0 - (X sdiv C) -> (X sdiv -C)
2439 if (Op1I->getOpcode() == Instruction::SDiv)
2440 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
2442 if (Constant *DivRHS = dyn_cast<Constant>(Op1I->getOperand(1)))
2443 return BinaryOperator::CreateSDiv(Op1I->getOperand(0),
2444 ConstantExpr::getNeg(DivRHS));
2446 // X - X*C --> X * (1-C)
2447 ConstantInt *C2 = 0;
2448 if (dyn_castFoldableMul(Op1I, C2) == Op0) {
2449 Constant *CP1 = Subtract(ConstantInt::get(I.getType(), 1), C2);
2450 return BinaryOperator::CreateMul(Op0, CP1);
2455 if (!Op0->getType()->isFPOrFPVector())
2456 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
2457 if (Op0I->getOpcode() == Instruction::Add) {
2458 if (Op0I->getOperand(0) == Op1) // (Y+X)-Y == X
2459 return ReplaceInstUsesWith(I, Op0I->getOperand(1));
2460 else if (Op0I->getOperand(1) == Op1) // (X+Y)-Y == X
2461 return ReplaceInstUsesWith(I, Op0I->getOperand(0));
2462 } else if (Op0I->getOpcode() == Instruction::Sub) {
2463 if (Op0I->getOperand(0) == Op1) // (X-Y)-X == -Y
2464 return BinaryOperator::CreateNeg(Op0I->getOperand(1), I.getName());
2469 if (Value *X = dyn_castFoldableMul(Op0, C1)) {
2470 if (X == Op1) // X*C - X --> X * (C-1)
2471 return BinaryOperator::CreateMul(Op1, SubOne(C1));
2473 ConstantInt *C2; // X*C1 - X*C2 -> X * (C1-C2)
2474 if (X == dyn_castFoldableMul(Op1, C2))
2475 return BinaryOperator::CreateMul(X, Subtract(C1, C2));
2480 /// isSignBitCheck - Given an exploded icmp instruction, return true if the
2481 /// comparison only checks the sign bit. If it only checks the sign bit, set
2482 /// TrueIfSigned if the result of the comparison is true when the input value is
2484 static bool isSignBitCheck(ICmpInst::Predicate pred, ConstantInt *RHS,
2485 bool &TrueIfSigned) {
2487 case ICmpInst::ICMP_SLT: // True if LHS s< 0
2488 TrueIfSigned = true;
2489 return RHS->isZero();
2490 case ICmpInst::ICMP_SLE: // True if LHS s<= RHS and RHS == -1
2491 TrueIfSigned = true;
2492 return RHS->isAllOnesValue();
2493 case ICmpInst::ICMP_SGT: // True if LHS s> -1
2494 TrueIfSigned = false;
2495 return RHS->isAllOnesValue();
2496 case ICmpInst::ICMP_UGT:
2497 // True if LHS u> RHS and RHS == high-bit-mask - 1
2498 TrueIfSigned = true;
2499 return RHS->getValue() ==
2500 APInt::getSignedMaxValue(RHS->getType()->getPrimitiveSizeInBits());
2501 case ICmpInst::ICMP_UGE:
2502 // True if LHS u>= RHS and RHS == high-bit-mask (2^7, 2^15, 2^31, etc)
2503 TrueIfSigned = true;
2504 return RHS->getValue().isSignBit();
2510 Instruction *InstCombiner::visitMul(BinaryOperator &I) {
2511 bool Changed = SimplifyCommutative(I);
2512 Value *Op0 = I.getOperand(0);
2514 if (isa<UndefValue>(I.getOperand(1))) // undef * X -> 0
2515 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2517 // Simplify mul instructions with a constant RHS...
2518 if (Constant *Op1 = dyn_cast<Constant>(I.getOperand(1))) {
2519 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2521 // ((X << C1)*C2) == (X * (C2 << C1))
2522 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op0))
2523 if (SI->getOpcode() == Instruction::Shl)
2524 if (Constant *ShOp = dyn_cast<Constant>(SI->getOperand(1)))
2525 return BinaryOperator::CreateMul(SI->getOperand(0),
2526 ConstantExpr::getShl(CI, ShOp));
2529 return ReplaceInstUsesWith(I, Op1); // X * 0 == 0
2530 if (CI->equalsInt(1)) // X * 1 == X
2531 return ReplaceInstUsesWith(I, Op0);
2532 if (CI->isAllOnesValue()) // X * -1 == 0 - X
2533 return BinaryOperator::CreateNeg(Op0, I.getName());
2535 const APInt& Val = cast<ConstantInt>(CI)->getValue();
2536 if (Val.isPowerOf2()) { // Replace X*(2^C) with X << C
2537 return BinaryOperator::CreateShl(Op0,
2538 ConstantInt::get(Op0->getType(), Val.logBase2()));
2540 } else if (ConstantFP *Op1F = dyn_cast<ConstantFP>(Op1)) {
2541 if (Op1F->isNullValue())
2542 return ReplaceInstUsesWith(I, Op1);
2544 // "In IEEE floating point, x*1 is not equivalent to x for nans. However,
2545 // ANSI says we can drop signals, so we can do this anyway." (from GCC)
2546 if (Op1F->isExactlyValue(1.0))
2547 return ReplaceInstUsesWith(I, Op0); // Eliminate 'mul double %X, 1.0'
2548 } else if (isa<VectorType>(Op1->getType())) {
2549 if (isa<ConstantAggregateZero>(Op1))
2550 return ReplaceInstUsesWith(I, Op1);
2552 if (ConstantVector *Op1V = dyn_cast<ConstantVector>(Op1)) {
2553 if (Op1V->isAllOnesValue()) // X * -1 == 0 - X
2554 return BinaryOperator::CreateNeg(Op0, I.getName());
2556 // As above, vector X*splat(1.0) -> X in all defined cases.
2557 if (Constant *Splat = Op1V->getSplatValue()) {
2558 if (ConstantFP *F = dyn_cast<ConstantFP>(Splat))
2559 if (F->isExactlyValue(1.0))
2560 return ReplaceInstUsesWith(I, Op0);
2561 if (ConstantInt *CI = dyn_cast<ConstantInt>(Splat))
2562 if (CI->equalsInt(1))
2563 return ReplaceInstUsesWith(I, Op0);
2568 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
2569 if (Op0I->getOpcode() == Instruction::Add && Op0I->hasOneUse() &&
2570 isa<ConstantInt>(Op0I->getOperand(1)) && isa<ConstantInt>(Op1)) {
2571 // Canonicalize (X+C1)*C2 -> X*C2+C1*C2.
2572 Instruction *Add = BinaryOperator::CreateMul(Op0I->getOperand(0),
2574 InsertNewInstBefore(Add, I);
2575 Value *C1C2 = ConstantExpr::getMul(Op1,
2576 cast<Constant>(Op0I->getOperand(1)));
2577 return BinaryOperator::CreateAdd(Add, C1C2);
2581 // Try to fold constant mul into select arguments.
2582 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2583 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2586 if (isa<PHINode>(Op0))
2587 if (Instruction *NV = FoldOpIntoPhi(I))
2591 if (Value *Op0v = dyn_castNegVal(Op0)) // -X * -Y = X*Y
2592 if (Value *Op1v = dyn_castNegVal(I.getOperand(1)))
2593 return BinaryOperator::CreateMul(Op0v, Op1v);
2595 // (X / Y) * Y = X - (X % Y)
2596 // (X / Y) * -Y = (X % Y) - X
2598 Value *Op1 = I.getOperand(1);
2599 BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0);
2601 (BO->getOpcode() != Instruction::UDiv &&
2602 BO->getOpcode() != Instruction::SDiv)) {
2604 BO = dyn_cast<BinaryOperator>(I.getOperand(1));
2606 Value *Neg = dyn_castNegVal(Op1);
2607 if (BO && BO->hasOneUse() &&
2608 (BO->getOperand(1) == Op1 || BO->getOperand(1) == Neg) &&
2609 (BO->getOpcode() == Instruction::UDiv ||
2610 BO->getOpcode() == Instruction::SDiv)) {
2611 Value *Op0BO = BO->getOperand(0), *Op1BO = BO->getOperand(1);
2614 if (BO->getOpcode() == Instruction::UDiv)
2615 Rem = BinaryOperator::CreateURem(Op0BO, Op1BO);
2617 Rem = BinaryOperator::CreateSRem(Op0BO, Op1BO);
2619 InsertNewInstBefore(Rem, I);
2623 return BinaryOperator::CreateSub(Op0BO, Rem);
2625 return BinaryOperator::CreateSub(Rem, Op0BO);
2629 if (I.getType() == Type::Int1Ty)
2630 return BinaryOperator::CreateAnd(Op0, I.getOperand(1));
2632 // If one of the operands of the multiply is a cast from a boolean value, then
2633 // we know the bool is either zero or one, so this is a 'masking' multiply.
2634 // See if we can simplify things based on how the boolean was originally
2636 CastInst *BoolCast = 0;
2637 if (ZExtInst *CI = dyn_cast<ZExtInst>(Op0))
2638 if (CI->getOperand(0)->getType() == Type::Int1Ty)
2641 if (ZExtInst *CI = dyn_cast<ZExtInst>(I.getOperand(1)))
2642 if (CI->getOperand(0)->getType() == Type::Int1Ty)
2645 if (ICmpInst *SCI = dyn_cast<ICmpInst>(BoolCast->getOperand(0))) {
2646 Value *SCIOp0 = SCI->getOperand(0), *SCIOp1 = SCI->getOperand(1);
2647 const Type *SCOpTy = SCIOp0->getType();
2650 // If the icmp is true iff the sign bit of X is set, then convert this
2651 // multiply into a shift/and combination.
2652 if (isa<ConstantInt>(SCIOp1) &&
2653 isSignBitCheck(SCI->getPredicate(), cast<ConstantInt>(SCIOp1), TIS) &&
2655 // Shift the X value right to turn it into "all signbits".
2656 Constant *Amt = ConstantInt::get(SCIOp0->getType(),
2657 SCOpTy->getPrimitiveSizeInBits()-1);
2659 InsertNewInstBefore(
2660 BinaryOperator::Create(Instruction::AShr, SCIOp0, Amt,
2661 BoolCast->getOperand(0)->getName()+
2664 // If the multiply type is not the same as the source type, sign extend
2665 // or truncate to the multiply type.
2666 if (I.getType() != V->getType()) {
2667 uint32_t SrcBits = V->getType()->getPrimitiveSizeInBits();
2668 uint32_t DstBits = I.getType()->getPrimitiveSizeInBits();
2669 Instruction::CastOps opcode =
2670 (SrcBits == DstBits ? Instruction::BitCast :
2671 (SrcBits < DstBits ? Instruction::SExt : Instruction::Trunc));
2672 V = InsertCastBefore(opcode, V, I.getType(), I);
2675 Value *OtherOp = Op0 == BoolCast ? I.getOperand(1) : Op0;
2676 return BinaryOperator::CreateAnd(V, OtherOp);
2681 return Changed ? &I : 0;
2684 /// SimplifyDivRemOfSelect - Try to fold a divide or remainder of a select
2686 bool InstCombiner::SimplifyDivRemOfSelect(BinaryOperator &I) {
2687 SelectInst *SI = cast<SelectInst>(I.getOperand(1));
2689 // div/rem X, (Cond ? 0 : Y) -> div/rem X, Y
2690 int NonNullOperand = -1;
2691 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(1)))
2692 if (ST->isNullValue())
2694 // div/rem X, (Cond ? Y : 0) -> div/rem X, Y
2695 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(2)))
2696 if (ST->isNullValue())
2699 if (NonNullOperand == -1)
2702 Value *SelectCond = SI->getOperand(0);
2704 // Change the div/rem to use 'Y' instead of the select.
2705 I.setOperand(1, SI->getOperand(NonNullOperand));
2707 // Okay, we know we replace the operand of the div/rem with 'Y' with no
2708 // problem. However, the select, or the condition of the select may have
2709 // multiple uses. Based on our knowledge that the operand must be non-zero,
2710 // propagate the known value for the select into other uses of it, and
2711 // propagate a known value of the condition into its other users.
2713 // If the select and condition only have a single use, don't bother with this,
2715 if (SI->use_empty() && SelectCond->hasOneUse())
2718 // Scan the current block backward, looking for other uses of SI.
2719 BasicBlock::iterator BBI = &I, BBFront = I.getParent()->begin();
2721 while (BBI != BBFront) {
2723 // If we found a call to a function, we can't assume it will return, so
2724 // information from below it cannot be propagated above it.
2725 if (isa<CallInst>(BBI) && !isa<IntrinsicInst>(BBI))
2728 // Replace uses of the select or its condition with the known values.
2729 for (Instruction::op_iterator I = BBI->op_begin(), E = BBI->op_end();
2732 *I = SI->getOperand(NonNullOperand);
2734 } else if (*I == SelectCond) {
2735 *I = NonNullOperand == 1 ? ConstantInt::getTrue() :
2736 ConstantInt::getFalse();
2741 // If we past the instruction, quit looking for it.
2744 if (&*BBI == SelectCond)
2747 // If we ran out of things to eliminate, break out of the loop.
2748 if (SelectCond == 0 && SI == 0)
2756 /// This function implements the transforms on div instructions that work
2757 /// regardless of the kind of div instruction it is (udiv, sdiv, or fdiv). It is
2758 /// used by the visitors to those instructions.
2759 /// @brief Transforms common to all three div instructions
2760 Instruction *InstCombiner::commonDivTransforms(BinaryOperator &I) {
2761 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2763 // undef / X -> 0 for integer.
2764 // undef / X -> undef for FP (the undef could be a snan).
2765 if (isa<UndefValue>(Op0)) {
2766 if (Op0->getType()->isFPOrFPVector())
2767 return ReplaceInstUsesWith(I, Op0);
2768 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2771 // X / undef -> undef
2772 if (isa<UndefValue>(Op1))
2773 return ReplaceInstUsesWith(I, Op1);
2778 /// This function implements the transforms common to both integer division
2779 /// instructions (udiv and sdiv). It is called by the visitors to those integer
2780 /// division instructions.
2781 /// @brief Common integer divide transforms
2782 Instruction *InstCombiner::commonIDivTransforms(BinaryOperator &I) {
2783 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2785 // (sdiv X, X) --> 1 (udiv X, X) --> 1
2787 if (const VectorType *Ty = dyn_cast<VectorType>(I.getType())) {
2788 ConstantInt *CI = ConstantInt::get(Ty->getElementType(), 1);
2789 std::vector<Constant*> Elts(Ty->getNumElements(), CI);
2790 return ReplaceInstUsesWith(I, ConstantVector::get(Elts));
2793 ConstantInt *CI = ConstantInt::get(I.getType(), 1);
2794 return ReplaceInstUsesWith(I, CI);
2797 if (Instruction *Common = commonDivTransforms(I))
2800 // Handle cases involving: [su]div X, (select Cond, Y, Z)
2801 // This does not apply for fdiv.
2802 if (isa<SelectInst>(Op1) && SimplifyDivRemOfSelect(I))
2805 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2807 if (RHS->equalsInt(1))
2808 return ReplaceInstUsesWith(I, Op0);
2810 // (X / C1) / C2 -> X / (C1*C2)
2811 if (Instruction *LHS = dyn_cast<Instruction>(Op0))
2812 if (Instruction::BinaryOps(LHS->getOpcode()) == I.getOpcode())
2813 if (ConstantInt *LHSRHS = dyn_cast<ConstantInt>(LHS->getOperand(1))) {
2814 if (MultiplyOverflows(RHS, LHSRHS, I.getOpcode()==Instruction::SDiv))
2815 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2817 return BinaryOperator::Create(I.getOpcode(), LHS->getOperand(0),
2818 Multiply(RHS, LHSRHS));
2821 if (!RHS->isZero()) { // avoid X udiv 0
2822 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2823 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2825 if (isa<PHINode>(Op0))
2826 if (Instruction *NV = FoldOpIntoPhi(I))
2831 // 0 / X == 0, we don't need to preserve faults!
2832 if (ConstantInt *LHS = dyn_cast<ConstantInt>(Op0))
2833 if (LHS->equalsInt(0))
2834 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2836 // It can't be division by zero, hence it must be division by one.
2837 if (I.getType() == Type::Int1Ty)
2838 return ReplaceInstUsesWith(I, Op0);
2840 if (ConstantVector *Op1V = dyn_cast<ConstantVector>(Op1)) {
2841 if (ConstantInt *X = cast_or_null<ConstantInt>(Op1V->getSplatValue()))
2844 return ReplaceInstUsesWith(I, Op0);
2850 Instruction *InstCombiner::visitUDiv(BinaryOperator &I) {
2851 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2853 // Handle the integer div common cases
2854 if (Instruction *Common = commonIDivTransforms(I))
2857 if (ConstantInt *C = dyn_cast<ConstantInt>(Op1)) {
2858 // X udiv C^2 -> X >> C
2859 // Check to see if this is an unsigned division with an exact power of 2,
2860 // if so, convert to a right shift.
2861 if (C->getValue().isPowerOf2()) // 0 not included in isPowerOf2
2862 return BinaryOperator::CreateLShr(Op0,
2863 ConstantInt::get(Op0->getType(), C->getValue().logBase2()));
2865 // X udiv C, where C >= signbit
2866 if (C->getValue().isNegative()) {
2867 Value *IC = InsertNewInstBefore(new ICmpInst(ICmpInst::ICMP_ULT, Op0, C),
2869 return SelectInst::Create(IC, Constant::getNullValue(I.getType()),
2870 ConstantInt::get(I.getType(), 1));
2874 // X udiv (C1 << N), where C1 is "1<<C2" --> X >> (N+C2)
2875 if (BinaryOperator *RHSI = dyn_cast<BinaryOperator>(I.getOperand(1))) {
2876 if (RHSI->getOpcode() == Instruction::Shl &&
2877 isa<ConstantInt>(RHSI->getOperand(0))) {
2878 const APInt& C1 = cast<ConstantInt>(RHSI->getOperand(0))->getValue();
2879 if (C1.isPowerOf2()) {
2880 Value *N = RHSI->getOperand(1);
2881 const Type *NTy = N->getType();
2882 if (uint32_t C2 = C1.logBase2()) {
2883 Constant *C2V = ConstantInt::get(NTy, C2);
2884 N = InsertNewInstBefore(BinaryOperator::CreateAdd(N, C2V, "tmp"), I);
2886 return BinaryOperator::CreateLShr(Op0, N);
2891 // udiv X, (Select Cond, C1, C2) --> Select Cond, (shr X, C1), (shr X, C2)
2892 // where C1&C2 are powers of two.
2893 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
2894 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
2895 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
2896 const APInt &TVA = STO->getValue(), &FVA = SFO->getValue();
2897 if (TVA.isPowerOf2() && FVA.isPowerOf2()) {
2898 // Compute the shift amounts
2899 uint32_t TSA = TVA.logBase2(), FSA = FVA.logBase2();
2900 // Construct the "on true" case of the select
2901 Constant *TC = ConstantInt::get(Op0->getType(), TSA);
2902 Instruction *TSI = BinaryOperator::CreateLShr(
2903 Op0, TC, SI->getName()+".t");
2904 TSI = InsertNewInstBefore(TSI, I);
2906 // Construct the "on false" case of the select
2907 Constant *FC = ConstantInt::get(Op0->getType(), FSA);
2908 Instruction *FSI = BinaryOperator::CreateLShr(
2909 Op0, FC, SI->getName()+".f");
2910 FSI = InsertNewInstBefore(FSI, I);
2912 // construct the select instruction and return it.
2913 return SelectInst::Create(SI->getOperand(0), TSI, FSI, SI->getName());
2919 Instruction *InstCombiner::visitSDiv(BinaryOperator &I) {
2920 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2922 // Handle the integer div common cases
2923 if (Instruction *Common = commonIDivTransforms(I))
2926 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2928 if (RHS->isAllOnesValue())
2929 return BinaryOperator::CreateNeg(Op0);
2931 ConstantInt *RHSNeg = cast<ConstantInt>(ConstantExpr::getNeg(RHS));
2932 APInt RHSNegAPI(RHSNeg->getValue());
2934 APInt NegOne = -APInt(RHSNeg->getBitWidth(), 1, true);
2935 APInt TwoToExp(RHSNeg->getBitWidth(), 1 << (RHSNeg->getBitWidth() - 1));
2937 // -X/C -> X/-C, if and only if negation doesn't overflow.
2938 if ((RHS->getValue().isNegative() && RHSNegAPI.slt(TwoToExp - 1)) ||
2939 (RHS->getValue().isNonNegative() && RHSNegAPI.sgt(TwoToExp * NegOne))) {
2940 if (Value *LHSNeg = dyn_castNegVal(Op0)) {
2941 if (ConstantInt *CI = dyn_cast<ConstantInt>(LHSNeg)) {
2942 ConstantInt *CINeg = cast<ConstantInt>(ConstantExpr::getNeg(CI));
2943 APInt CINegAPI(CINeg->getValue());
2945 if ((CI->getValue().isNegative() && CINegAPI.slt(TwoToExp - 1)) ||
2946 (CI->getValue().isNonNegative() && CINegAPI.sgt(TwoToExp*NegOne)))
2947 return BinaryOperator::CreateSDiv(LHSNeg,
2948 ConstantExpr::getNeg(RHS));
2954 // If the sign bits of both operands are zero (i.e. we can prove they are
2955 // unsigned inputs), turn this into a udiv.
2956 if (I.getType()->isInteger()) {
2957 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()));
2958 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
2959 // X sdiv Y -> X udiv Y, iff X and Y don't have sign bit set
2960 return BinaryOperator::CreateUDiv(Op0, Op1, I.getName());
2967 Instruction *InstCombiner::visitFDiv(BinaryOperator &I) {
2968 return commonDivTransforms(I);
2971 /// This function implements the transforms on rem instructions that work
2972 /// regardless of the kind of rem instruction it is (urem, srem, or frem). It
2973 /// is used by the visitors to those instructions.
2974 /// @brief Transforms common to all three rem instructions
2975 Instruction *InstCombiner::commonRemTransforms(BinaryOperator &I) {
2976 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2978 // 0 % X == 0 for integer, we don't need to preserve faults!
2979 if (Constant *LHS = dyn_cast<Constant>(Op0))
2980 if (LHS->isNullValue())
2981 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2983 if (isa<UndefValue>(Op0)) { // undef % X -> 0
2984 if (I.getType()->isFPOrFPVector())
2985 return ReplaceInstUsesWith(I, Op0); // X % undef -> undef (could be SNaN)
2986 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2988 if (isa<UndefValue>(Op1))
2989 return ReplaceInstUsesWith(I, Op1); // X % undef -> undef
2991 // Handle cases involving: rem X, (select Cond, Y, Z)
2992 if (isa<SelectInst>(Op1) && SimplifyDivRemOfSelect(I))
2998 /// This function implements the transforms common to both integer remainder
2999 /// instructions (urem and srem). It is called by the visitors to those integer
3000 /// remainder instructions.
3001 /// @brief Common integer remainder transforms
3002 Instruction *InstCombiner::commonIRemTransforms(BinaryOperator &I) {
3003 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3005 if (Instruction *common = commonRemTransforms(I))
3008 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3009 // X % 0 == undef, we don't need to preserve faults!
3010 if (RHS->equalsInt(0))
3011 return ReplaceInstUsesWith(I, UndefValue::get(I.getType()));
3013 if (RHS->equalsInt(1)) // X % 1 == 0
3014 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3016 if (Instruction *Op0I = dyn_cast<Instruction>(Op0)) {
3017 if (SelectInst *SI = dyn_cast<SelectInst>(Op0I)) {
3018 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3020 } else if (isa<PHINode>(Op0I)) {
3021 if (Instruction *NV = FoldOpIntoPhi(I))
3025 // See if we can fold away this rem instruction.
3026 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
3027 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
3028 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
3029 KnownZero, KnownOne))
3037 Instruction *InstCombiner::visitURem(BinaryOperator &I) {
3038 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3040 if (Instruction *common = commonIRemTransforms(I))
3043 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3044 // X urem C^2 -> X and C
3045 // Check to see if this is an unsigned remainder with an exact power of 2,
3046 // if so, convert to a bitwise and.
3047 if (ConstantInt *C = dyn_cast<ConstantInt>(RHS))
3048 if (C->getValue().isPowerOf2())
3049 return BinaryOperator::CreateAnd(Op0, SubOne(C));
3052 if (Instruction *RHSI = dyn_cast<Instruction>(I.getOperand(1))) {
3053 // Turn A % (C << N), where C is 2^k, into A & ((C << N)-1)
3054 if (RHSI->getOpcode() == Instruction::Shl &&
3055 isa<ConstantInt>(RHSI->getOperand(0))) {
3056 if (cast<ConstantInt>(RHSI->getOperand(0))->getValue().isPowerOf2()) {
3057 Constant *N1 = ConstantInt::getAllOnesValue(I.getType());
3058 Value *Add = InsertNewInstBefore(BinaryOperator::CreateAdd(RHSI, N1,
3060 return BinaryOperator::CreateAnd(Op0, Add);
3065 // urem X, (select Cond, 2^C1, 2^C2) --> select Cond, (and X, C1), (and X, C2)
3066 // where C1&C2 are powers of two.
3067 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
3068 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
3069 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
3070 // STO == 0 and SFO == 0 handled above.
3071 if ((STO->getValue().isPowerOf2()) &&
3072 (SFO->getValue().isPowerOf2())) {
3073 Value *TrueAnd = InsertNewInstBefore(
3074 BinaryOperator::CreateAnd(Op0, SubOne(STO), SI->getName()+".t"), I);
3075 Value *FalseAnd = InsertNewInstBefore(
3076 BinaryOperator::CreateAnd(Op0, SubOne(SFO), SI->getName()+".f"), I);
3077 return SelectInst::Create(SI->getOperand(0), TrueAnd, FalseAnd);
3085 Instruction *InstCombiner::visitSRem(BinaryOperator &I) {
3086 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3088 // Handle the integer rem common cases
3089 if (Instruction *common = commonIRemTransforms(I))
3092 if (Value *RHSNeg = dyn_castNegVal(Op1))
3093 if (!isa<Constant>(RHSNeg) ||
3094 (isa<ConstantInt>(RHSNeg) &&
3095 cast<ConstantInt>(RHSNeg)->getValue().isStrictlyPositive())) {
3097 AddUsesToWorkList(I);
3098 I.setOperand(1, RHSNeg);
3102 // If the sign bits of both operands are zero (i.e. we can prove they are
3103 // unsigned inputs), turn this into a urem.
3104 if (I.getType()->isInteger()) {
3105 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()));
3106 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
3107 // X srem Y -> X urem Y, iff X and Y don't have sign bit set
3108 return BinaryOperator::CreateURem(Op0, Op1, I.getName());
3115 Instruction *InstCombiner::visitFRem(BinaryOperator &I) {
3116 return commonRemTransforms(I);
3119 // isOneBitSet - Return true if there is exactly one bit set in the specified
3121 static bool isOneBitSet(const ConstantInt *CI) {
3122 return CI->getValue().isPowerOf2();
3125 // isHighOnes - Return true if the constant is of the form 1+0+.
3126 // This is the same as lowones(~X).
3127 static bool isHighOnes(const ConstantInt *CI) {
3128 return (~CI->getValue() + 1).isPowerOf2();
3131 /// getICmpCode - Encode a icmp predicate into a three bit mask. These bits
3132 /// are carefully arranged to allow folding of expressions such as:
3134 /// (A < B) | (A > B) --> (A != B)
3136 /// Note that this is only valid if the first and second predicates have the
3137 /// same sign. Is illegal to do: (A u< B) | (A s> B)
3139 /// Three bits are used to represent the condition, as follows:
3144 /// <=> Value Definition
3145 /// 000 0 Always false
3152 /// 111 7 Always true
3154 static unsigned getICmpCode(const ICmpInst *ICI) {
3155 switch (ICI->getPredicate()) {
3157 case ICmpInst::ICMP_UGT: return 1; // 001
3158 case ICmpInst::ICMP_SGT: return 1; // 001
3159 case ICmpInst::ICMP_EQ: return 2; // 010
3160 case ICmpInst::ICMP_UGE: return 3; // 011
3161 case ICmpInst::ICMP_SGE: return 3; // 011
3162 case ICmpInst::ICMP_ULT: return 4; // 100
3163 case ICmpInst::ICMP_SLT: return 4; // 100
3164 case ICmpInst::ICMP_NE: return 5; // 101
3165 case ICmpInst::ICMP_ULE: return 6; // 110
3166 case ICmpInst::ICMP_SLE: return 6; // 110
3169 assert(0 && "Invalid ICmp predicate!");
3174 /// getFCmpCode - Similar to getICmpCode but for FCmpInst. This encodes a fcmp
3175 /// predicate into a three bit mask. It also returns whether it is an ordered
3176 /// predicate by reference.
3177 static unsigned getFCmpCode(FCmpInst::Predicate CC, bool &isOrdered) {
3180 case FCmpInst::FCMP_ORD: isOrdered = true; return 0; // 000
3181 case FCmpInst::FCMP_UNO: return 0; // 000
3182 case FCmpInst::FCMP_OGT: isOrdered = true; return 1; // 001
3183 case FCmpInst::FCMP_UGT: return 1; // 001
3184 case FCmpInst::FCMP_OEQ: isOrdered = true; return 2; // 010
3185 case FCmpInst::FCMP_UEQ: return 2; // 010
3186 case FCmpInst::FCMP_OGE: isOrdered = true; return 3; // 011
3187 case FCmpInst::FCMP_UGE: return 3; // 011
3188 case FCmpInst::FCMP_OLT: isOrdered = true; return 4; // 100
3189 case FCmpInst::FCMP_ULT: return 4; // 100
3190 case FCmpInst::FCMP_ONE: isOrdered = true; return 5; // 101
3191 case FCmpInst::FCMP_UNE: return 5; // 101
3192 case FCmpInst::FCMP_OLE: isOrdered = true; return 6; // 110
3193 case FCmpInst::FCMP_ULE: return 6; // 110
3196 // Not expecting FCMP_FALSE and FCMP_TRUE;
3197 assert(0 && "Unexpected FCmp predicate!");
3202 /// getICmpValue - This is the complement of getICmpCode, which turns an
3203 /// opcode and two operands into either a constant true or false, or a brand
3204 /// new ICmp instruction. The sign is passed in to determine which kind
3205 /// of predicate to use in the new icmp instruction.
3206 static Value *getICmpValue(bool sign, unsigned code, Value *LHS, Value *RHS) {
3208 default: assert(0 && "Illegal ICmp code!");
3209 case 0: return ConstantInt::getFalse();
3212 return new ICmpInst(ICmpInst::ICMP_SGT, LHS, RHS);
3214 return new ICmpInst(ICmpInst::ICMP_UGT, LHS, RHS);
3215 case 2: return new ICmpInst(ICmpInst::ICMP_EQ, LHS, RHS);
3218 return new ICmpInst(ICmpInst::ICMP_SGE, LHS, RHS);
3220 return new ICmpInst(ICmpInst::ICMP_UGE, LHS, RHS);
3223 return new ICmpInst(ICmpInst::ICMP_SLT, LHS, RHS);
3225 return new ICmpInst(ICmpInst::ICMP_ULT, LHS, RHS);
3226 case 5: return new ICmpInst(ICmpInst::ICMP_NE, LHS, RHS);
3229 return new ICmpInst(ICmpInst::ICMP_SLE, LHS, RHS);
3231 return new ICmpInst(ICmpInst::ICMP_ULE, LHS, RHS);
3232 case 7: return ConstantInt::getTrue();
3236 /// getFCmpValue - This is the complement of getFCmpCode, which turns an
3237 /// opcode and two operands into either a FCmp instruction. isordered is passed
3238 /// in to determine which kind of predicate to use in the new fcmp instruction.
3239 static Value *getFCmpValue(bool isordered, unsigned code,
3240 Value *LHS, Value *RHS) {
3242 default: assert(0 && "Illegal FCmp code!");
3245 return new FCmpInst(FCmpInst::FCMP_ORD, LHS, RHS);
3247 return new FCmpInst(FCmpInst::FCMP_UNO, LHS, RHS);
3250 return new FCmpInst(FCmpInst::FCMP_OGT, LHS, RHS);
3252 return new FCmpInst(FCmpInst::FCMP_UGT, LHS, RHS);
3255 return new FCmpInst(FCmpInst::FCMP_OEQ, LHS, RHS);
3257 return new FCmpInst(FCmpInst::FCMP_UEQ, LHS, RHS);
3260 return new FCmpInst(FCmpInst::FCMP_OGE, LHS, RHS);
3262 return new FCmpInst(FCmpInst::FCMP_UGE, LHS, RHS);
3265 return new FCmpInst(FCmpInst::FCMP_OLT, LHS, RHS);
3267 return new FCmpInst(FCmpInst::FCMP_ULT, LHS, RHS);
3270 return new FCmpInst(FCmpInst::FCMP_ONE, LHS, RHS);
3272 return new FCmpInst(FCmpInst::FCMP_UNE, LHS, RHS);
3275 return new FCmpInst(FCmpInst::FCMP_OLE, LHS, RHS);
3277 return new FCmpInst(FCmpInst::FCMP_ULE, LHS, RHS);
3278 case 7: return ConstantInt::getTrue();
3282 /// PredicatesFoldable - Return true if both predicates match sign or if at
3283 /// least one of them is an equality comparison (which is signless).
3284 static bool PredicatesFoldable(ICmpInst::Predicate p1, ICmpInst::Predicate p2) {
3285 return (ICmpInst::isSignedPredicate(p1) == ICmpInst::isSignedPredicate(p2)) ||
3286 (ICmpInst::isSignedPredicate(p1) && ICmpInst::isEquality(p2)) ||
3287 (ICmpInst::isSignedPredicate(p2) && ICmpInst::isEquality(p1));
3291 // FoldICmpLogical - Implements (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
3292 struct FoldICmpLogical {
3295 ICmpInst::Predicate pred;
3296 FoldICmpLogical(InstCombiner &ic, ICmpInst *ICI)
3297 : IC(ic), LHS(ICI->getOperand(0)), RHS(ICI->getOperand(1)),
3298 pred(ICI->getPredicate()) {}
3299 bool shouldApply(Value *V) const {
3300 if (ICmpInst *ICI = dyn_cast<ICmpInst>(V))
3301 if (PredicatesFoldable(pred, ICI->getPredicate()))
3302 return ((ICI->getOperand(0) == LHS && ICI->getOperand(1) == RHS) ||
3303 (ICI->getOperand(0) == RHS && ICI->getOperand(1) == LHS));
3306 Instruction *apply(Instruction &Log) const {
3307 ICmpInst *ICI = cast<ICmpInst>(Log.getOperand(0));
3308 if (ICI->getOperand(0) != LHS) {
3309 assert(ICI->getOperand(1) == LHS);
3310 ICI->swapOperands(); // Swap the LHS and RHS of the ICmp
3313 ICmpInst *RHSICI = cast<ICmpInst>(Log.getOperand(1));
3314 unsigned LHSCode = getICmpCode(ICI);
3315 unsigned RHSCode = getICmpCode(RHSICI);
3317 switch (Log.getOpcode()) {
3318 case Instruction::And: Code = LHSCode & RHSCode; break;
3319 case Instruction::Or: Code = LHSCode | RHSCode; break;
3320 case Instruction::Xor: Code = LHSCode ^ RHSCode; break;
3321 default: assert(0 && "Illegal logical opcode!"); return 0;
3324 bool isSigned = ICmpInst::isSignedPredicate(RHSICI->getPredicate()) ||
3325 ICmpInst::isSignedPredicate(ICI->getPredicate());
3327 Value *RV = getICmpValue(isSigned, Code, LHS, RHS);
3328 if (Instruction *I = dyn_cast<Instruction>(RV))
3330 // Otherwise, it's a constant boolean value...
3331 return IC.ReplaceInstUsesWith(Log, RV);
3334 } // end anonymous namespace
3336 // OptAndOp - This handles expressions of the form ((val OP C1) & C2). Where
3337 // the Op parameter is 'OP', OpRHS is 'C1', and AndRHS is 'C2'. Op is
3338 // guaranteed to be a binary operator.
3339 Instruction *InstCombiner::OptAndOp(Instruction *Op,
3341 ConstantInt *AndRHS,
3342 BinaryOperator &TheAnd) {
3343 Value *X = Op->getOperand(0);
3344 Constant *Together = 0;
3346 Together = And(AndRHS, OpRHS);
3348 switch (Op->getOpcode()) {
3349 case Instruction::Xor:
3350 if (Op->hasOneUse()) {
3351 // (X ^ C1) & C2 --> (X & C2) ^ (C1&C2)
3352 Instruction *And = BinaryOperator::CreateAnd(X, AndRHS);
3353 InsertNewInstBefore(And, TheAnd);
3355 return BinaryOperator::CreateXor(And, Together);
3358 case Instruction::Or:
3359 if (Together == AndRHS) // (X | C) & C --> C
3360 return ReplaceInstUsesWith(TheAnd, AndRHS);
3362 if (Op->hasOneUse() && Together != OpRHS) {
3363 // (X | C1) & C2 --> (X | (C1&C2)) & C2
3364 Instruction *Or = BinaryOperator::CreateOr(X, Together);
3365 InsertNewInstBefore(Or, TheAnd);
3367 return BinaryOperator::CreateAnd(Or, AndRHS);
3370 case Instruction::Add:
3371 if (Op->hasOneUse()) {
3372 // Adding a one to a single bit bit-field should be turned into an XOR
3373 // of the bit. First thing to check is to see if this AND is with a
3374 // single bit constant.
3375 const APInt& AndRHSV = cast<ConstantInt>(AndRHS)->getValue();
3377 // If there is only one bit set...
3378 if (isOneBitSet(cast<ConstantInt>(AndRHS))) {
3379 // Ok, at this point, we know that we are masking the result of the
3380 // ADD down to exactly one bit. If the constant we are adding has
3381 // no bits set below this bit, then we can eliminate the ADD.
3382 const APInt& AddRHS = cast<ConstantInt>(OpRHS)->getValue();
3384 // Check to see if any bits below the one bit set in AndRHSV are set.
3385 if ((AddRHS & (AndRHSV-1)) == 0) {
3386 // If not, the only thing that can effect the output of the AND is
3387 // the bit specified by AndRHSV. If that bit is set, the effect of
3388 // the XOR is to toggle the bit. If it is clear, then the ADD has
3390 if ((AddRHS & AndRHSV) == 0) { // Bit is not set, noop
3391 TheAnd.setOperand(0, X);
3394 // Pull the XOR out of the AND.
3395 Instruction *NewAnd = BinaryOperator::CreateAnd(X, AndRHS);
3396 InsertNewInstBefore(NewAnd, TheAnd);
3397 NewAnd->takeName(Op);
3398 return BinaryOperator::CreateXor(NewAnd, AndRHS);
3405 case Instruction::Shl: {
3406 // We know that the AND will not produce any of the bits shifted in, so if
3407 // the anded constant includes them, clear them now!
3409 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3410 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3411 APInt ShlMask(APInt::getHighBitsSet(BitWidth, BitWidth-OpRHSVal));
3412 ConstantInt *CI = ConstantInt::get(AndRHS->getValue() & ShlMask);
3414 if (CI->getValue() == ShlMask) {
3415 // Masking out bits that the shift already masks
3416 return ReplaceInstUsesWith(TheAnd, Op); // No need for the and.
3417 } else if (CI != AndRHS) { // Reducing bits set in and.
3418 TheAnd.setOperand(1, CI);
3423 case Instruction::LShr:
3425 // We know that the AND will not produce any of the bits shifted in, so if
3426 // the anded constant includes them, clear them now! This only applies to
3427 // unsigned shifts, because a signed shr may bring in set bits!
3429 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3430 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3431 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
3432 ConstantInt *CI = ConstantInt::get(AndRHS->getValue() & ShrMask);
3434 if (CI->getValue() == ShrMask) {
3435 // Masking out bits that the shift already masks.
3436 return ReplaceInstUsesWith(TheAnd, Op);
3437 } else if (CI != AndRHS) {
3438 TheAnd.setOperand(1, CI); // Reduce bits set in and cst.
3443 case Instruction::AShr:
3445 // See if this is shifting in some sign extension, then masking it out
3447 if (Op->hasOneUse()) {
3448 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3449 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3450 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
3451 Constant *C = ConstantInt::get(AndRHS->getValue() & ShrMask);
3452 if (C == AndRHS) { // Masking out bits shifted in.
3453 // (Val ashr C1) & C2 -> (Val lshr C1) & C2
3454 // Make the argument unsigned.
3455 Value *ShVal = Op->getOperand(0);
3456 ShVal = InsertNewInstBefore(
3457 BinaryOperator::CreateLShr(ShVal, OpRHS,
3458 Op->getName()), TheAnd);
3459 return BinaryOperator::CreateAnd(ShVal, AndRHS, TheAnd.getName());
3468 /// InsertRangeTest - Emit a computation of: (V >= Lo && V < Hi) if Inside is
3469 /// true, otherwise (V < Lo || V >= Hi). In pratice, we emit the more efficient
3470 /// (V-Lo) <u Hi-Lo. This method expects that Lo <= Hi. isSigned indicates
3471 /// whether to treat the V, Lo and HI as signed or not. IB is the location to
3472 /// insert new instructions.
3473 Instruction *InstCombiner::InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
3474 bool isSigned, bool Inside,
3476 assert(cast<ConstantInt>(ConstantExpr::getICmp((isSigned ?
3477 ICmpInst::ICMP_SLE:ICmpInst::ICMP_ULE), Lo, Hi))->getZExtValue() &&
3478 "Lo is not <= Hi in range emission code!");
3481 if (Lo == Hi) // Trivially false.
3482 return new ICmpInst(ICmpInst::ICMP_NE, V, V);
3484 // V >= Min && V < Hi --> V < Hi
3485 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
3486 ICmpInst::Predicate pred = (isSigned ?
3487 ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT);
3488 return new ICmpInst(pred, V, Hi);
3491 // Emit V-Lo <u Hi-Lo
3492 Constant *NegLo = ConstantExpr::getNeg(Lo);
3493 Instruction *Add = BinaryOperator::CreateAdd(V, NegLo, V->getName()+".off");
3494 InsertNewInstBefore(Add, IB);
3495 Constant *UpperBound = ConstantExpr::getAdd(NegLo, Hi);
3496 return new ICmpInst(ICmpInst::ICMP_ULT, Add, UpperBound);
3499 if (Lo == Hi) // Trivially true.
3500 return new ICmpInst(ICmpInst::ICMP_EQ, V, V);
3502 // V < Min || V >= Hi -> V > Hi-1
3503 Hi = SubOne(cast<ConstantInt>(Hi));
3504 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
3505 ICmpInst::Predicate pred = (isSigned ?
3506 ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT);
3507 return new ICmpInst(pred, V, Hi);
3510 // Emit V-Lo >u Hi-1-Lo
3511 // Note that Hi has already had one subtracted from it, above.
3512 ConstantInt *NegLo = cast<ConstantInt>(ConstantExpr::getNeg(Lo));
3513 Instruction *Add = BinaryOperator::CreateAdd(V, NegLo, V->getName()+".off");
3514 InsertNewInstBefore(Add, IB);
3515 Constant *LowerBound = ConstantExpr::getAdd(NegLo, Hi);
3516 return new ICmpInst(ICmpInst::ICMP_UGT, Add, LowerBound);
3519 // isRunOfOnes - Returns true iff Val consists of one contiguous run of 1s with
3520 // any number of 0s on either side. The 1s are allowed to wrap from LSB to
3521 // MSB, so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs. 0x0F0F0000 is
3522 // not, since all 1s are not contiguous.
3523 static bool isRunOfOnes(ConstantInt *Val, uint32_t &MB, uint32_t &ME) {
3524 const APInt& V = Val->getValue();
3525 uint32_t BitWidth = Val->getType()->getBitWidth();
3526 if (!APIntOps::isShiftedMask(BitWidth, V)) return false;
3528 // look for the first zero bit after the run of ones
3529 MB = BitWidth - ((V - 1) ^ V).countLeadingZeros();
3530 // look for the first non-zero bit
3531 ME = V.getActiveBits();
3535 /// FoldLogicalPlusAnd - This is part of an expression (LHS +/- RHS) & Mask,
3536 /// where isSub determines whether the operator is a sub. If we can fold one of
3537 /// the following xforms:
3539 /// ((A & N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == Mask
3540 /// ((A | N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3541 /// ((A ^ N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3543 /// return (A +/- B).
3545 Value *InstCombiner::FoldLogicalPlusAnd(Value *LHS, Value *RHS,
3546 ConstantInt *Mask, bool isSub,
3548 Instruction *LHSI = dyn_cast<Instruction>(LHS);
3549 if (!LHSI || LHSI->getNumOperands() != 2 ||
3550 !isa<ConstantInt>(LHSI->getOperand(1))) return 0;
3552 ConstantInt *N = cast<ConstantInt>(LHSI->getOperand(1));
3554 switch (LHSI->getOpcode()) {
3556 case Instruction::And:
3557 if (And(N, Mask) == Mask) {
3558 // If the AndRHS is a power of two minus one (0+1+), this is simple.
3559 if ((Mask->getValue().countLeadingZeros() +
3560 Mask->getValue().countPopulation()) ==
3561 Mask->getValue().getBitWidth())
3564 // Otherwise, if Mask is 0+1+0+, and if B is known to have the low 0+
3565 // part, we don't need any explicit masks to take them out of A. If that
3566 // is all N is, ignore it.
3567 uint32_t MB = 0, ME = 0;
3568 if (isRunOfOnes(Mask, MB, ME)) { // begin/end bit of run, inclusive
3569 uint32_t BitWidth = cast<IntegerType>(RHS->getType())->getBitWidth();
3570 APInt Mask(APInt::getLowBitsSet(BitWidth, MB-1));
3571 if (MaskedValueIsZero(RHS, Mask))
3576 case Instruction::Or:
3577 case Instruction::Xor:
3578 // If the AndRHS is a power of two minus one (0+1+), and N&Mask == 0
3579 if ((Mask->getValue().countLeadingZeros() +
3580 Mask->getValue().countPopulation()) == Mask->getValue().getBitWidth()
3581 && And(N, Mask)->isZero())
3588 New = BinaryOperator::CreateSub(LHSI->getOperand(0), RHS, "fold");
3590 New = BinaryOperator::CreateAdd(LHSI->getOperand(0), RHS, "fold");
3591 return InsertNewInstBefore(New, I);
3594 /// FoldAndOfICmps - Fold (icmp)&(icmp) if possible.
3595 Instruction *InstCombiner::FoldAndOfICmps(Instruction &I,
3596 ICmpInst *LHS, ICmpInst *RHS) {
3598 ConstantInt *LHSCst, *RHSCst;
3599 ICmpInst::Predicate LHSCC, RHSCC;
3601 // This only handles icmp of constants: (icmp1 A, C1) & (icmp2 B, C2).
3602 if (!match(LHS, m_ICmp(LHSCC, m_Value(Val), m_ConstantInt(LHSCst))) ||
3603 !match(RHS, m_ICmp(RHSCC, m_Value(Val2), m_ConstantInt(RHSCst))))
3606 // (icmp ult A, C) & (icmp ult B, C) --> (icmp ult (A|B), C)
3607 // where C is a power of 2
3608 if (LHSCst == RHSCst && LHSCC == RHSCC && LHSCC == ICmpInst::ICMP_ULT &&
3609 LHSCst->getValue().isPowerOf2()) {
3610 Instruction *NewOr = BinaryOperator::CreateOr(Val, Val2);
3611 InsertNewInstBefore(NewOr, I);
3612 return new ICmpInst(LHSCC, NewOr, LHSCst);
3615 // From here on, we only handle:
3616 // (icmp1 A, C1) & (icmp2 A, C2) --> something simpler.
3617 if (Val != Val2) return 0;
3619 // ICMP_[US][GL]E X, CST is folded to ICMP_[US][GL]T elsewhere.
3620 if (LHSCC == ICmpInst::ICMP_UGE || LHSCC == ICmpInst::ICMP_ULE ||
3621 RHSCC == ICmpInst::ICMP_UGE || RHSCC == ICmpInst::ICMP_ULE ||
3622 LHSCC == ICmpInst::ICMP_SGE || LHSCC == ICmpInst::ICMP_SLE ||
3623 RHSCC == ICmpInst::ICMP_SGE || RHSCC == ICmpInst::ICMP_SLE)
3626 // We can't fold (ugt x, C) & (sgt x, C2).
3627 if (!PredicatesFoldable(LHSCC, RHSCC))
3630 // Ensure that the larger constant is on the RHS.
3632 if (ICmpInst::isSignedPredicate(LHSCC) ||
3633 (ICmpInst::isEquality(LHSCC) &&
3634 ICmpInst::isSignedPredicate(RHSCC)))
3635 ShouldSwap = LHSCst->getValue().sgt(RHSCst->getValue());
3637 ShouldSwap = LHSCst->getValue().ugt(RHSCst->getValue());
3640 std::swap(LHS, RHS);
3641 std::swap(LHSCst, RHSCst);
3642 std::swap(LHSCC, RHSCC);
3645 // At this point, we know we have have two icmp instructions
3646 // comparing a value against two constants and and'ing the result
3647 // together. Because of the above check, we know that we only have
3648 // icmp eq, icmp ne, icmp [su]lt, and icmp [SU]gt here. We also know
3649 // (from the FoldICmpLogical check above), that the two constants
3650 // are not equal and that the larger constant is on the RHS
3651 assert(LHSCst != RHSCst && "Compares not folded above?");
3654 default: assert(0 && "Unknown integer condition code!");
3655 case ICmpInst::ICMP_EQ:
3657 default: assert(0 && "Unknown integer condition code!");
3658 case ICmpInst::ICMP_EQ: // (X == 13 & X == 15) -> false
3659 case ICmpInst::ICMP_UGT: // (X == 13 & X > 15) -> false
3660 case ICmpInst::ICMP_SGT: // (X == 13 & X > 15) -> false
3661 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3662 case ICmpInst::ICMP_NE: // (X == 13 & X != 15) -> X == 13
3663 case ICmpInst::ICMP_ULT: // (X == 13 & X < 15) -> X == 13
3664 case ICmpInst::ICMP_SLT: // (X == 13 & X < 15) -> X == 13
3665 return ReplaceInstUsesWith(I, LHS);
3667 case ICmpInst::ICMP_NE:
3669 default: assert(0 && "Unknown integer condition code!");
3670 case ICmpInst::ICMP_ULT:
3671 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X u< 14) -> X < 13
3672 return new ICmpInst(ICmpInst::ICMP_ULT, Val, LHSCst);
3673 break; // (X != 13 & X u< 15) -> no change
3674 case ICmpInst::ICMP_SLT:
3675 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X s< 14) -> X < 13
3676 return new ICmpInst(ICmpInst::ICMP_SLT, Val, LHSCst);
3677 break; // (X != 13 & X s< 15) -> no change
3678 case ICmpInst::ICMP_EQ: // (X != 13 & X == 15) -> X == 15
3679 case ICmpInst::ICMP_UGT: // (X != 13 & X u> 15) -> X u> 15
3680 case ICmpInst::ICMP_SGT: // (X != 13 & X s> 15) -> X s> 15
3681 return ReplaceInstUsesWith(I, RHS);
3682 case ICmpInst::ICMP_NE:
3683 if (LHSCst == SubOne(RHSCst)){// (X != 13 & X != 14) -> X-13 >u 1
3684 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
3685 Instruction *Add = BinaryOperator::CreateAdd(Val, AddCST,
3686 Val->getName()+".off");
3687 InsertNewInstBefore(Add, I);
3688 return new ICmpInst(ICmpInst::ICMP_UGT, Add,
3689 ConstantInt::get(Add->getType(), 1));
3691 break; // (X != 13 & X != 15) -> no change
3694 case ICmpInst::ICMP_ULT:
3696 default: assert(0 && "Unknown integer condition code!");
3697 case ICmpInst::ICMP_EQ: // (X u< 13 & X == 15) -> false
3698 case ICmpInst::ICMP_UGT: // (X u< 13 & X u> 15) -> false
3699 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3700 case ICmpInst::ICMP_SGT: // (X u< 13 & X s> 15) -> no change
3702 case ICmpInst::ICMP_NE: // (X u< 13 & X != 15) -> X u< 13
3703 case ICmpInst::ICMP_ULT: // (X u< 13 & X u< 15) -> X u< 13
3704 return ReplaceInstUsesWith(I, LHS);
3705 case ICmpInst::ICMP_SLT: // (X u< 13 & X s< 15) -> no change
3709 case ICmpInst::ICMP_SLT:
3711 default: assert(0 && "Unknown integer condition code!");
3712 case ICmpInst::ICMP_EQ: // (X s< 13 & X == 15) -> false
3713 case ICmpInst::ICMP_SGT: // (X s< 13 & X s> 15) -> false
3714 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3715 case ICmpInst::ICMP_UGT: // (X s< 13 & X u> 15) -> no change
3717 case ICmpInst::ICMP_NE: // (X s< 13 & X != 15) -> X < 13
3718 case ICmpInst::ICMP_SLT: // (X s< 13 & X s< 15) -> X < 13
3719 return ReplaceInstUsesWith(I, LHS);
3720 case ICmpInst::ICMP_ULT: // (X s< 13 & X u< 15) -> no change
3724 case ICmpInst::ICMP_UGT:
3726 default: assert(0 && "Unknown integer condition code!");
3727 case ICmpInst::ICMP_EQ: // (X u> 13 & X == 15) -> X == 15
3728 case ICmpInst::ICMP_UGT: // (X u> 13 & X u> 15) -> X u> 15
3729 return ReplaceInstUsesWith(I, RHS);
3730 case ICmpInst::ICMP_SGT: // (X u> 13 & X s> 15) -> no change
3732 case ICmpInst::ICMP_NE:
3733 if (RHSCst == AddOne(LHSCst)) // (X u> 13 & X != 14) -> X u> 14
3734 return new ICmpInst(LHSCC, Val, RHSCst);
3735 break; // (X u> 13 & X != 15) -> no change
3736 case ICmpInst::ICMP_ULT: // (X u> 13 & X u< 15) -> (X-14) <u 1
3737 return InsertRangeTest(Val, AddOne(LHSCst), RHSCst, false, true, I);
3738 case ICmpInst::ICMP_SLT: // (X u> 13 & X s< 15) -> no change
3742 case ICmpInst::ICMP_SGT:
3744 default: assert(0 && "Unknown integer condition code!");
3745 case ICmpInst::ICMP_EQ: // (X s> 13 & X == 15) -> X == 15
3746 case ICmpInst::ICMP_SGT: // (X s> 13 & X s> 15) -> X s> 15
3747 return ReplaceInstUsesWith(I, RHS);
3748 case ICmpInst::ICMP_UGT: // (X s> 13 & X u> 15) -> no change
3750 case ICmpInst::ICMP_NE:
3751 if (RHSCst == AddOne(LHSCst)) // (X s> 13 & X != 14) -> X s> 14
3752 return new ICmpInst(LHSCC, Val, RHSCst);
3753 break; // (X s> 13 & X != 15) -> no change
3754 case ICmpInst::ICMP_SLT: // (X s> 13 & X s< 15) -> (X-14) s< 1
3755 return InsertRangeTest(Val, AddOne(LHSCst), RHSCst, true, true, I);
3756 case ICmpInst::ICMP_ULT: // (X s> 13 & X u< 15) -> no change
3766 Instruction *InstCombiner::visitAnd(BinaryOperator &I) {
3767 bool Changed = SimplifyCommutative(I);
3768 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3770 if (isa<UndefValue>(Op1)) // X & undef -> 0
3771 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3775 return ReplaceInstUsesWith(I, Op1);
3777 // See if we can simplify any instructions used by the instruction whose sole
3778 // purpose is to compute bits we don't care about.
3779 if (!isa<VectorType>(I.getType())) {
3780 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
3781 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
3782 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
3783 KnownZero, KnownOne))
3786 if (ConstantVector *CP = dyn_cast<ConstantVector>(Op1)) {
3787 if (CP->isAllOnesValue()) // X & <-1,-1> -> X
3788 return ReplaceInstUsesWith(I, I.getOperand(0));
3789 } else if (isa<ConstantAggregateZero>(Op1)) {
3790 return ReplaceInstUsesWith(I, Op1); // X & <0,0> -> <0,0>
3794 if (ConstantInt *AndRHS = dyn_cast<ConstantInt>(Op1)) {
3795 const APInt& AndRHSMask = AndRHS->getValue();
3796 APInt NotAndRHS(~AndRHSMask);
3798 // Optimize a variety of ((val OP C1) & C2) combinations...
3799 if (isa<BinaryOperator>(Op0)) {
3800 Instruction *Op0I = cast<Instruction>(Op0);
3801 Value *Op0LHS = Op0I->getOperand(0);
3802 Value *Op0RHS = Op0I->getOperand(1);
3803 switch (Op0I->getOpcode()) {
3804 case Instruction::Xor:
3805 case Instruction::Or:
3806 // If the mask is only needed on one incoming arm, push it up.
3807 if (Op0I->hasOneUse()) {
3808 if (MaskedValueIsZero(Op0LHS, NotAndRHS)) {
3809 // Not masking anything out for the LHS, move to RHS.
3810 Instruction *NewRHS = BinaryOperator::CreateAnd(Op0RHS, AndRHS,
3811 Op0RHS->getName()+".masked");
3812 InsertNewInstBefore(NewRHS, I);
3813 return BinaryOperator::Create(
3814 cast<BinaryOperator>(Op0I)->getOpcode(), Op0LHS, NewRHS);
3816 if (!isa<Constant>(Op0RHS) &&
3817 MaskedValueIsZero(Op0RHS, NotAndRHS)) {
3818 // Not masking anything out for the RHS, move to LHS.
3819 Instruction *NewLHS = BinaryOperator::CreateAnd(Op0LHS, AndRHS,
3820 Op0LHS->getName()+".masked");
3821 InsertNewInstBefore(NewLHS, I);
3822 return BinaryOperator::Create(
3823 cast<BinaryOperator>(Op0I)->getOpcode(), NewLHS, Op0RHS);
3828 case Instruction::Add:
3829 // ((A & N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == AndRHS.
3830 // ((A | N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
3831 // ((A ^ N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
3832 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, false, I))
3833 return BinaryOperator::CreateAnd(V, AndRHS);
3834 if (Value *V = FoldLogicalPlusAnd(Op0RHS, Op0LHS, AndRHS, false, I))
3835 return BinaryOperator::CreateAnd(V, AndRHS); // Add commutes
3838 case Instruction::Sub:
3839 // ((A & N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == AndRHS.
3840 // ((A | N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
3841 // ((A ^ N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
3842 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, true, I))
3843 return BinaryOperator::CreateAnd(V, AndRHS);
3845 // (A - N) & AndRHS -> -N & AndRHS iff A&AndRHS==0 and AndRHS
3846 // has 1's for all bits that the subtraction with A might affect.
3847 if (Op0I->hasOneUse()) {
3848 uint32_t BitWidth = AndRHSMask.getBitWidth();
3849 uint32_t Zeros = AndRHSMask.countLeadingZeros();
3850 APInt Mask = APInt::getLowBitsSet(BitWidth, BitWidth - Zeros);
3852 ConstantInt *A = dyn_cast<ConstantInt>(Op0LHS);
3853 if (!(A && A->isZero()) && // avoid infinite recursion.
3854 MaskedValueIsZero(Op0LHS, Mask)) {
3855 Instruction *NewNeg = BinaryOperator::CreateNeg(Op0RHS);
3856 InsertNewInstBefore(NewNeg, I);
3857 return BinaryOperator::CreateAnd(NewNeg, AndRHS);
3862 case Instruction::Shl:
3863 case Instruction::LShr:
3864 // (1 << x) & 1 --> zext(x == 0)
3865 // (1 >> x) & 1 --> zext(x == 0)
3866 if (AndRHSMask == 1 && Op0LHS == AndRHS) {
3867 Instruction *NewICmp = new ICmpInst(ICmpInst::ICMP_EQ, Op0RHS,
3868 Constant::getNullValue(I.getType()));
3869 InsertNewInstBefore(NewICmp, I);
3870 return new ZExtInst(NewICmp, I.getType());
3875 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
3876 if (Instruction *Res = OptAndOp(Op0I, Op0CI, AndRHS, I))
3878 } else if (CastInst *CI = dyn_cast<CastInst>(Op0)) {
3879 // If this is an integer truncation or change from signed-to-unsigned, and
3880 // if the source is an and/or with immediate, transform it. This
3881 // frequently occurs for bitfield accesses.
3882 if (Instruction *CastOp = dyn_cast<Instruction>(CI->getOperand(0))) {
3883 if ((isa<TruncInst>(CI) || isa<BitCastInst>(CI)) &&
3884 CastOp->getNumOperands() == 2)
3885 if (ConstantInt *AndCI = dyn_cast<ConstantInt>(CastOp->getOperand(1))) {
3886 if (CastOp->getOpcode() == Instruction::And) {
3887 // Change: and (cast (and X, C1) to T), C2
3888 // into : and (cast X to T), trunc_or_bitcast(C1)&C2
3889 // This will fold the two constants together, which may allow
3890 // other simplifications.
3891 Instruction *NewCast = CastInst::CreateTruncOrBitCast(
3892 CastOp->getOperand(0), I.getType(),
3893 CastOp->getName()+".shrunk");
3894 NewCast = InsertNewInstBefore(NewCast, I);
3895 // trunc_or_bitcast(C1)&C2
3896 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
3897 C3 = ConstantExpr::getAnd(C3, AndRHS);
3898 return BinaryOperator::CreateAnd(NewCast, C3);
3899 } else if (CastOp->getOpcode() == Instruction::Or) {
3900 // Change: and (cast (or X, C1) to T), C2
3901 // into : trunc(C1)&C2 iff trunc(C1)&C2 == C2
3902 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
3903 if (ConstantExpr::getAnd(C3, AndRHS) == AndRHS) // trunc(C1)&C2
3904 return ReplaceInstUsesWith(I, AndRHS);
3910 // Try to fold constant and into select arguments.
3911 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3912 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3914 if (isa<PHINode>(Op0))
3915 if (Instruction *NV = FoldOpIntoPhi(I))
3919 Value *Op0NotVal = dyn_castNotVal(Op0);
3920 Value *Op1NotVal = dyn_castNotVal(Op1);
3922 if (Op0NotVal == Op1 || Op1NotVal == Op0) // A & ~A == ~A & A == 0
3923 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3925 // (~A & ~B) == (~(A | B)) - De Morgan's Law
3926 if (Op0NotVal && Op1NotVal && isOnlyUse(Op0) && isOnlyUse(Op1)) {
3927 Instruction *Or = BinaryOperator::CreateOr(Op0NotVal, Op1NotVal,
3928 I.getName()+".demorgan");
3929 InsertNewInstBefore(Or, I);
3930 return BinaryOperator::CreateNot(Or);
3934 Value *A = 0, *B = 0, *C = 0, *D = 0;
3935 if (match(Op0, m_Or(m_Value(A), m_Value(B)))) {
3936 if (A == Op1 || B == Op1) // (A | ?) & A --> A
3937 return ReplaceInstUsesWith(I, Op1);
3939 // (A|B) & ~(A&B) -> A^B
3940 if (match(Op1, m_Not(m_And(m_Value(C), m_Value(D))))) {
3941 if ((A == C && B == D) || (A == D && B == C))
3942 return BinaryOperator::CreateXor(A, B);
3946 if (match(Op1, m_Or(m_Value(A), m_Value(B)))) {
3947 if (A == Op0 || B == Op0) // A & (A | ?) --> A
3948 return ReplaceInstUsesWith(I, Op0);
3950 // ~(A&B) & (A|B) -> A^B
3951 if (match(Op0, m_Not(m_And(m_Value(C), m_Value(D))))) {
3952 if ((A == C && B == D) || (A == D && B == C))
3953 return BinaryOperator::CreateXor(A, B);
3957 if (Op0->hasOneUse() &&
3958 match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
3959 if (A == Op1) { // (A^B)&A -> A&(A^B)
3960 I.swapOperands(); // Simplify below
3961 std::swap(Op0, Op1);
3962 } else if (B == Op1) { // (A^B)&B -> B&(B^A)
3963 cast<BinaryOperator>(Op0)->swapOperands();
3964 I.swapOperands(); // Simplify below
3965 std::swap(Op0, Op1);
3969 if (Op1->hasOneUse() &&
3970 match(Op1, m_Xor(m_Value(A), m_Value(B)))) {
3971 if (B == Op0) { // B&(A^B) -> B&(B^A)
3972 cast<BinaryOperator>(Op1)->swapOperands();
3975 if (A == Op0) { // A&(A^B) -> A & ~B
3976 Instruction *NotB = BinaryOperator::CreateNot(B, "tmp");
3977 InsertNewInstBefore(NotB, I);
3978 return BinaryOperator::CreateAnd(A, NotB);
3982 // (A&((~A)|B)) -> A&B
3983 if (match(Op0, m_Or(m_Not(m_Value(A)), m_Value(B)))) {
3985 return BinaryOperator::CreateAnd(A, B);
3987 if (match(Op0, m_Or(m_Value(A), m_Not(m_Value(B))))) {
3989 return BinaryOperator::CreateAnd(A, B);
3991 if (match(Op1, m_Or(m_Not(m_Value(A)), m_Value(B)))) {
3993 return BinaryOperator::CreateAnd(A, B);
3995 if (match(Op1, m_Or(m_Value(A), m_Not(m_Value(B))))) {
3997 return BinaryOperator::CreateAnd(A, B);
4001 if (ICmpInst *RHS = dyn_cast<ICmpInst>(Op1)) {
4002 // (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
4003 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
4006 if (ICmpInst *LHS = dyn_cast<ICmpInst>(Op0))
4007 if (Instruction *Res = FoldAndOfICmps(I, LHS, RHS))
4011 // fold (and (cast A), (cast B)) -> (cast (and A, B))
4012 if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
4013 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4014 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind ?
4015 const Type *SrcTy = Op0C->getOperand(0)->getType();
4016 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
4017 // Only do this if the casts both really cause code to be generated.
4018 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4020 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4022 Instruction *NewOp = BinaryOperator::CreateAnd(Op0C->getOperand(0),
4023 Op1C->getOperand(0),
4025 InsertNewInstBefore(NewOp, I);
4026 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
4030 // (X >> Z) & (Y >> Z) -> (X&Y) >> Z for all shifts.
4031 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
4032 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
4033 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
4034 SI0->getOperand(1) == SI1->getOperand(1) &&
4035 (SI0->hasOneUse() || SI1->hasOneUse())) {
4036 Instruction *NewOp =
4037 InsertNewInstBefore(BinaryOperator::CreateAnd(SI0->getOperand(0),
4039 SI0->getName()), I);
4040 return BinaryOperator::Create(SI1->getOpcode(), NewOp,
4041 SI1->getOperand(1));
4045 // If and'ing two fcmp, try combine them into one.
4046 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0))) {
4047 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1))) {
4048 if (LHS->getPredicate() == FCmpInst::FCMP_ORD &&
4049 RHS->getPredicate() == FCmpInst::FCMP_ORD) {
4050 // (fcmp ord x, c) & (fcmp ord y, c) -> (fcmp ord x, y)
4051 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
4052 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
4053 // If either of the constants are nans, then the whole thing returns
4055 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
4056 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4057 return new FCmpInst(FCmpInst::FCMP_ORD, LHS->getOperand(0),
4058 RHS->getOperand(0));
4061 Value *Op0LHS, *Op0RHS, *Op1LHS, *Op1RHS;
4062 FCmpInst::Predicate Op0CC, Op1CC;
4063 if (match(Op0, m_FCmp(Op0CC, m_Value(Op0LHS), m_Value(Op0RHS))) &&
4064 match(Op1, m_FCmp(Op1CC, m_Value(Op1LHS), m_Value(Op1RHS)))) {
4065 if (Op0LHS == Op1RHS && Op0RHS == Op1LHS) {
4066 // Swap RHS operands to match LHS.
4067 Op1CC = FCmpInst::getSwappedPredicate(Op1CC);
4068 std::swap(Op1LHS, Op1RHS);
4070 if (Op0LHS == Op1LHS && Op0RHS == Op1RHS) {
4071 // Simplify (fcmp cc0 x, y) & (fcmp cc1 x, y).
4073 return new FCmpInst((FCmpInst::Predicate)Op0CC, Op0LHS, Op0RHS);
4074 else if (Op0CC == FCmpInst::FCMP_FALSE ||
4075 Op1CC == FCmpInst::FCMP_FALSE)
4076 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4077 else if (Op0CC == FCmpInst::FCMP_TRUE)
4078 return ReplaceInstUsesWith(I, Op1);
4079 else if (Op1CC == FCmpInst::FCMP_TRUE)
4080 return ReplaceInstUsesWith(I, Op0);
4083 unsigned Op0Pred = getFCmpCode(Op0CC, Op0Ordered);
4084 unsigned Op1Pred = getFCmpCode(Op1CC, Op1Ordered);
4086 std::swap(Op0, Op1);
4087 std::swap(Op0Pred, Op1Pred);
4088 std::swap(Op0Ordered, Op1Ordered);
4091 // uno && ueq -> uno && (uno || eq) -> ueq
4092 // ord && olt -> ord && (ord && lt) -> olt
4093 if (Op0Ordered == Op1Ordered)
4094 return ReplaceInstUsesWith(I, Op1);
4095 // uno && oeq -> uno && (ord && eq) -> false
4096 // uno && ord -> false
4098 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4099 // ord && ueq -> ord && (uno || eq) -> oeq
4100 return cast<Instruction>(getFCmpValue(true, Op1Pred,
4109 return Changed ? &I : 0;
4112 /// CollectBSwapParts - Analyze the specified subexpression and see if it is
4113 /// capable of providing pieces of a bswap. The subexpression provides pieces
4114 /// of a bswap if it is proven that each of the non-zero bytes in the output of
4115 /// the expression came from the corresponding "byte swapped" byte in some other
4116 /// value. For example, if the current subexpression is "(shl i32 %X, 24)" then
4117 /// we know that the expression deposits the low byte of %X into the high byte
4118 /// of the bswap result and that all other bytes are zero. This expression is
4119 /// accepted, the high byte of ByteValues is set to X to indicate a correct
4122 /// This function returns true if the match was unsuccessful and false if so.
4123 /// On entry to the function the "OverallLeftShift" is a signed integer value
4124 /// indicating the number of bytes that the subexpression is later shifted. For
4125 /// example, if the expression is later right shifted by 16 bits, the
4126 /// OverallLeftShift value would be -2 on entry. This is used to specify which
4127 /// byte of ByteValues is actually being set.
4129 /// Similarly, ByteMask is a bitmask where a bit is clear if its corresponding
4130 /// byte is masked to zero by a user. For example, in (X & 255), X will be
4131 /// processed with a bytemask of 1. Because bytemask is 32-bits, this limits
4132 /// this function to working on up to 32-byte (256 bit) values. ByteMask is
4133 /// always in the local (OverallLeftShift) coordinate space.
4135 static bool CollectBSwapParts(Value *V, int OverallLeftShift, uint32_t ByteMask,
4136 SmallVector<Value*, 8> &ByteValues) {
4137 if (Instruction *I = dyn_cast<Instruction>(V)) {
4138 // If this is an or instruction, it may be an inner node of the bswap.
4139 if (I->getOpcode() == Instruction::Or) {
4140 return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
4142 CollectBSwapParts(I->getOperand(1), OverallLeftShift, ByteMask,
4146 // If this is a logical shift by a constant multiple of 8, recurse with
4147 // OverallLeftShift and ByteMask adjusted.
4148 if (I->isLogicalShift() && isa<ConstantInt>(I->getOperand(1))) {
4150 cast<ConstantInt>(I->getOperand(1))->getLimitedValue(~0U);
4151 // Ensure the shift amount is defined and of a byte value.
4152 if ((ShAmt & 7) || (ShAmt > 8*ByteValues.size()))
4155 unsigned ByteShift = ShAmt >> 3;
4156 if (I->getOpcode() == Instruction::Shl) {
4157 // X << 2 -> collect(X, +2)
4158 OverallLeftShift += ByteShift;
4159 ByteMask >>= ByteShift;
4161 // X >>u 2 -> collect(X, -2)
4162 OverallLeftShift -= ByteShift;
4163 ByteMask <<= ByteShift;
4164 ByteMask &= (~0U >> (32-ByteValues.size()));
4167 if (OverallLeftShift >= (int)ByteValues.size()) return true;
4168 if (OverallLeftShift <= -(int)ByteValues.size()) return true;
4170 return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
4174 // If this is a logical 'and' with a mask that clears bytes, clear the
4175 // corresponding bytes in ByteMask.
4176 if (I->getOpcode() == Instruction::And &&
4177 isa<ConstantInt>(I->getOperand(1))) {
4178 // Scan every byte of the and mask, seeing if the byte is either 0 or 255.
4179 unsigned NumBytes = ByteValues.size();
4180 APInt Byte(I->getType()->getPrimitiveSizeInBits(), 255);
4181 const APInt &AndMask = cast<ConstantInt>(I->getOperand(1))->getValue();
4183 for (unsigned i = 0; i != NumBytes; ++i, Byte <<= 8) {
4184 // If this byte is masked out by a later operation, we don't care what
4186 if ((ByteMask & (1 << i)) == 0)
4189 // If the AndMask is all zeros for this byte, clear the bit.
4190 APInt MaskB = AndMask & Byte;
4192 ByteMask &= ~(1U << i);
4196 // If the AndMask is not all ones for this byte, it's not a bytezap.
4200 // Otherwise, this byte is kept.
4203 return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
4208 // Okay, we got to something that isn't a shift, 'or' or 'and'. This must be
4209 // the input value to the bswap. Some observations: 1) if more than one byte
4210 // is demanded from this input, then it could not be successfully assembled
4211 // into a byteswap. At least one of the two bytes would not be aligned with
4212 // their ultimate destination.
4213 if (!isPowerOf2_32(ByteMask)) return true;
4214 unsigned InputByteNo = CountTrailingZeros_32(ByteMask);
4216 // 2) The input and ultimate destinations must line up: if byte 3 of an i32
4217 // is demanded, it needs to go into byte 0 of the result. This means that the
4218 // byte needs to be shifted until it lands in the right byte bucket. The
4219 // shift amount depends on the position: if the byte is coming from the high
4220 // part of the value (e.g. byte 3) then it must be shifted right. If from the
4221 // low part, it must be shifted left.
4222 unsigned DestByteNo = InputByteNo + OverallLeftShift;
4223 if (InputByteNo < ByteValues.size()/2) {
4224 if (ByteValues.size()-1-DestByteNo != InputByteNo)
4227 if (ByteValues.size()-1-DestByteNo != InputByteNo)
4231 // If the destination byte value is already defined, the values are or'd
4232 // together, which isn't a bswap (unless it's an or of the same bits).
4233 if (ByteValues[DestByteNo] && ByteValues[DestByteNo] != V)
4235 ByteValues[DestByteNo] = V;
4239 /// MatchBSwap - Given an OR instruction, check to see if this is a bswap idiom.
4240 /// If so, insert the new bswap intrinsic and return it.
4241 Instruction *InstCombiner::MatchBSwap(BinaryOperator &I) {
4242 const IntegerType *ITy = dyn_cast<IntegerType>(I.getType());
4243 if (!ITy || ITy->getBitWidth() % 16 ||
4244 // ByteMask only allows up to 32-byte values.
4245 ITy->getBitWidth() > 32*8)
4246 return 0; // Can only bswap pairs of bytes. Can't do vectors.
4248 /// ByteValues - For each byte of the result, we keep track of which value
4249 /// defines each byte.
4250 SmallVector<Value*, 8> ByteValues;
4251 ByteValues.resize(ITy->getBitWidth()/8);
4253 // Try to find all the pieces corresponding to the bswap.
4254 uint32_t ByteMask = ~0U >> (32-ByteValues.size());
4255 if (CollectBSwapParts(&I, 0, ByteMask, ByteValues))
4258 // Check to see if all of the bytes come from the same value.
4259 Value *V = ByteValues[0];
4260 if (V == 0) return 0; // Didn't find a byte? Must be zero.
4262 // Check to make sure that all of the bytes come from the same value.
4263 for (unsigned i = 1, e = ByteValues.size(); i != e; ++i)
4264 if (ByteValues[i] != V)
4266 const Type *Tys[] = { ITy };
4267 Module *M = I.getParent()->getParent()->getParent();
4268 Function *F = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 1);
4269 return CallInst::Create(F, V);
4272 /// MatchSelectFromAndOr - We have an expression of the form (A&C)|(B&D). Check
4273 /// If A is (cond?-1:0) and either B or D is ~(cond?-1,0) or (cond?0,-1), then
4274 /// we can simplify this expression to "cond ? C : D or B".
4275 static Instruction *MatchSelectFromAndOr(Value *A, Value *B,
4276 Value *C, Value *D) {
4277 // If A is not a select of -1/0, this cannot match.
4279 if (!match(A, m_SelectCst(m_Value(Cond), -1, 0)))
4282 // ((cond?-1:0)&C) | (B&(cond?0:-1)) -> cond ? C : B.
4283 if (match(D, m_SelectCst(m_Specific(Cond), 0, -1)))
4284 return SelectInst::Create(Cond, C, B);
4285 if (match(D, m_Not(m_SelectCst(m_Specific(Cond), -1, 0))))
4286 return SelectInst::Create(Cond, C, B);
4287 // ((cond?-1:0)&C) | ((cond?0:-1)&D) -> cond ? C : D.
4288 if (match(B, m_SelectCst(m_Specific(Cond), 0, -1)))
4289 return SelectInst::Create(Cond, C, D);
4290 if (match(B, m_Not(m_SelectCst(m_Specific(Cond), -1, 0))))
4291 return SelectInst::Create(Cond, C, D);
4295 /// FoldOrOfICmps - Fold (icmp)|(icmp) if possible.
4296 Instruction *InstCombiner::FoldOrOfICmps(Instruction &I,
4297 ICmpInst *LHS, ICmpInst *RHS) {
4299 ConstantInt *LHSCst, *RHSCst;
4300 ICmpInst::Predicate LHSCC, RHSCC;
4302 // This only handles icmp of constants: (icmp1 A, C1) | (icmp2 B, C2).
4303 if (!match(LHS, m_ICmp(LHSCC, m_Value(Val), m_ConstantInt(LHSCst))) ||
4304 !match(RHS, m_ICmp(RHSCC, m_Value(Val2), m_ConstantInt(RHSCst))))
4307 // From here on, we only handle:
4308 // (icmp1 A, C1) | (icmp2 A, C2) --> something simpler.
4309 if (Val != Val2) return 0;
4311 // ICMP_[US][GL]E X, CST is folded to ICMP_[US][GL]T elsewhere.
4312 if (LHSCC == ICmpInst::ICMP_UGE || LHSCC == ICmpInst::ICMP_ULE ||
4313 RHSCC == ICmpInst::ICMP_UGE || RHSCC == ICmpInst::ICMP_ULE ||
4314 LHSCC == ICmpInst::ICMP_SGE || LHSCC == ICmpInst::ICMP_SLE ||
4315 RHSCC == ICmpInst::ICMP_SGE || RHSCC == ICmpInst::ICMP_SLE)
4318 // We can't fold (ugt x, C) | (sgt x, C2).
4319 if (!PredicatesFoldable(LHSCC, RHSCC))
4322 // Ensure that the larger constant is on the RHS.
4324 if (ICmpInst::isSignedPredicate(LHSCC) ||
4325 (ICmpInst::isEquality(LHSCC) &&
4326 ICmpInst::isSignedPredicate(RHSCC)))
4327 ShouldSwap = LHSCst->getValue().sgt(RHSCst->getValue());
4329 ShouldSwap = LHSCst->getValue().ugt(RHSCst->getValue());
4332 std::swap(LHS, RHS);
4333 std::swap(LHSCst, RHSCst);
4334 std::swap(LHSCC, RHSCC);
4337 // At this point, we know we have have two icmp instructions
4338 // comparing a value against two constants and or'ing the result
4339 // together. Because of the above check, we know that we only have
4340 // ICMP_EQ, ICMP_NE, ICMP_LT, and ICMP_GT here. We also know (from the
4341 // FoldICmpLogical check above), that the two constants are not
4343 assert(LHSCst != RHSCst && "Compares not folded above?");
4346 default: assert(0 && "Unknown integer condition code!");
4347 case ICmpInst::ICMP_EQ:
4349 default: assert(0 && "Unknown integer condition code!");
4350 case ICmpInst::ICMP_EQ:
4351 if (LHSCst == SubOne(RHSCst)) { // (X == 13 | X == 14) -> X-13 <u 2
4352 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
4353 Instruction *Add = BinaryOperator::CreateAdd(Val, AddCST,
4354 Val->getName()+".off");
4355 InsertNewInstBefore(Add, I);
4356 AddCST = Subtract(AddOne(RHSCst), LHSCst);
4357 return new ICmpInst(ICmpInst::ICMP_ULT, Add, AddCST);
4359 break; // (X == 13 | X == 15) -> no change
4360 case ICmpInst::ICMP_UGT: // (X == 13 | X u> 14) -> no change
4361 case ICmpInst::ICMP_SGT: // (X == 13 | X s> 14) -> no change
4363 case ICmpInst::ICMP_NE: // (X == 13 | X != 15) -> X != 15
4364 case ICmpInst::ICMP_ULT: // (X == 13 | X u< 15) -> X u< 15
4365 case ICmpInst::ICMP_SLT: // (X == 13 | X s< 15) -> X s< 15
4366 return ReplaceInstUsesWith(I, RHS);
4369 case ICmpInst::ICMP_NE:
4371 default: assert(0 && "Unknown integer condition code!");
4372 case ICmpInst::ICMP_EQ: // (X != 13 | X == 15) -> X != 13
4373 case ICmpInst::ICMP_UGT: // (X != 13 | X u> 15) -> X != 13
4374 case ICmpInst::ICMP_SGT: // (X != 13 | X s> 15) -> X != 13
4375 return ReplaceInstUsesWith(I, LHS);
4376 case ICmpInst::ICMP_NE: // (X != 13 | X != 15) -> true
4377 case ICmpInst::ICMP_ULT: // (X != 13 | X u< 15) -> true
4378 case ICmpInst::ICMP_SLT: // (X != 13 | X s< 15) -> true
4379 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4382 case ICmpInst::ICMP_ULT:
4384 default: assert(0 && "Unknown integer condition code!");
4385 case ICmpInst::ICMP_EQ: // (X u< 13 | X == 14) -> no change
4387 case ICmpInst::ICMP_UGT: // (X u< 13 | X u> 15) -> (X-13) u> 2
4388 // If RHSCst is [us]MAXINT, it is always false. Not handling
4389 // this can cause overflow.
4390 if (RHSCst->isMaxValue(false))
4391 return ReplaceInstUsesWith(I, LHS);
4392 return InsertRangeTest(Val, LHSCst, AddOne(RHSCst), false, false, I);
4393 case ICmpInst::ICMP_SGT: // (X u< 13 | X s> 15) -> no change
4395 case ICmpInst::ICMP_NE: // (X u< 13 | X != 15) -> X != 15
4396 case ICmpInst::ICMP_ULT: // (X u< 13 | X u< 15) -> X u< 15
4397 return ReplaceInstUsesWith(I, RHS);
4398 case ICmpInst::ICMP_SLT: // (X u< 13 | X s< 15) -> no change
4402 case ICmpInst::ICMP_SLT:
4404 default: assert(0 && "Unknown integer condition code!");
4405 case ICmpInst::ICMP_EQ: // (X s< 13 | X == 14) -> no change
4407 case ICmpInst::ICMP_SGT: // (X s< 13 | X s> 15) -> (X-13) s> 2
4408 // If RHSCst is [us]MAXINT, it is always false. Not handling
4409 // this can cause overflow.
4410 if (RHSCst->isMaxValue(true))
4411 return ReplaceInstUsesWith(I, LHS);
4412 return InsertRangeTest(Val, LHSCst, AddOne(RHSCst), true, false, I);
4413 case ICmpInst::ICMP_UGT: // (X s< 13 | X u> 15) -> no change
4415 case ICmpInst::ICMP_NE: // (X s< 13 | X != 15) -> X != 15
4416 case ICmpInst::ICMP_SLT: // (X s< 13 | X s< 15) -> X s< 15
4417 return ReplaceInstUsesWith(I, RHS);
4418 case ICmpInst::ICMP_ULT: // (X s< 13 | X u< 15) -> no change
4422 case ICmpInst::ICMP_UGT:
4424 default: assert(0 && "Unknown integer condition code!");
4425 case ICmpInst::ICMP_EQ: // (X u> 13 | X == 15) -> X u> 13
4426 case ICmpInst::ICMP_UGT: // (X u> 13 | X u> 15) -> X u> 13
4427 return ReplaceInstUsesWith(I, LHS);
4428 case ICmpInst::ICMP_SGT: // (X u> 13 | X s> 15) -> no change
4430 case ICmpInst::ICMP_NE: // (X u> 13 | X != 15) -> true
4431 case ICmpInst::ICMP_ULT: // (X u> 13 | X u< 15) -> true
4432 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4433 case ICmpInst::ICMP_SLT: // (X u> 13 | X s< 15) -> no change
4437 case ICmpInst::ICMP_SGT:
4439 default: assert(0 && "Unknown integer condition code!");
4440 case ICmpInst::ICMP_EQ: // (X s> 13 | X == 15) -> X > 13
4441 case ICmpInst::ICMP_SGT: // (X s> 13 | X s> 15) -> X > 13
4442 return ReplaceInstUsesWith(I, LHS);
4443 case ICmpInst::ICMP_UGT: // (X s> 13 | X u> 15) -> no change
4445 case ICmpInst::ICMP_NE: // (X s> 13 | X != 15) -> true
4446 case ICmpInst::ICMP_SLT: // (X s> 13 | X s< 15) -> true
4447 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4448 case ICmpInst::ICMP_ULT: // (X s> 13 | X u< 15) -> no change
4456 Instruction *InstCombiner::visitOr(BinaryOperator &I) {
4457 bool Changed = SimplifyCommutative(I);
4458 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4460 if (isa<UndefValue>(Op1)) // X | undef -> -1
4461 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4465 return ReplaceInstUsesWith(I, Op0);
4467 // See if we can simplify any instructions used by the instruction whose sole
4468 // purpose is to compute bits we don't care about.
4469 if (!isa<VectorType>(I.getType())) {
4470 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
4471 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
4472 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
4473 KnownZero, KnownOne))
4475 } else if (isa<ConstantAggregateZero>(Op1)) {
4476 return ReplaceInstUsesWith(I, Op0); // X | <0,0> -> X
4477 } else if (ConstantVector *CP = dyn_cast<ConstantVector>(Op1)) {
4478 if (CP->isAllOnesValue()) // X | <-1,-1> -> <-1,-1>
4479 return ReplaceInstUsesWith(I, I.getOperand(1));
4485 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
4486 ConstantInt *C1 = 0; Value *X = 0;
4487 // (X & C1) | C2 --> (X | C2) & (C1|C2)
4488 if (match(Op0, m_And(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
4489 Instruction *Or = BinaryOperator::CreateOr(X, RHS);
4490 InsertNewInstBefore(Or, I);
4492 return BinaryOperator::CreateAnd(Or,
4493 ConstantInt::get(RHS->getValue() | C1->getValue()));
4496 // (X ^ C1) | C2 --> (X | C2) ^ (C1&~C2)
4497 if (match(Op0, m_Xor(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
4498 Instruction *Or = BinaryOperator::CreateOr(X, RHS);
4499 InsertNewInstBefore(Or, I);
4501 return BinaryOperator::CreateXor(Or,
4502 ConstantInt::get(C1->getValue() & ~RHS->getValue()));
4505 // Try to fold constant and into select arguments.
4506 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
4507 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
4509 if (isa<PHINode>(Op0))
4510 if (Instruction *NV = FoldOpIntoPhi(I))
4514 Value *A = 0, *B = 0;
4515 ConstantInt *C1 = 0, *C2 = 0;
4517 if (match(Op0, m_And(m_Value(A), m_Value(B))))
4518 if (A == Op1 || B == Op1) // (A & ?) | A --> A
4519 return ReplaceInstUsesWith(I, Op1);
4520 if (match(Op1, m_And(m_Value(A), m_Value(B))))
4521 if (A == Op0 || B == Op0) // A | (A & ?) --> A
4522 return ReplaceInstUsesWith(I, Op0);
4524 // (A | B) | C and A | (B | C) -> bswap if possible.
4525 // (A >> B) | (C << D) and (A << B) | (B >> C) -> bswap if possible.
4526 if (match(Op0, m_Or(m_Value(), m_Value())) ||
4527 match(Op1, m_Or(m_Value(), m_Value())) ||
4528 (match(Op0, m_Shift(m_Value(), m_Value())) &&
4529 match(Op1, m_Shift(m_Value(), m_Value())))) {
4530 if (Instruction *BSwap = MatchBSwap(I))
4534 // (X^C)|Y -> (X|Y)^C iff Y&C == 0
4535 if (Op0->hasOneUse() && match(Op0, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
4536 MaskedValueIsZero(Op1, C1->getValue())) {
4537 Instruction *NOr = BinaryOperator::CreateOr(A, Op1);
4538 InsertNewInstBefore(NOr, I);
4540 return BinaryOperator::CreateXor(NOr, C1);
4543 // Y|(X^C) -> (X|Y)^C iff Y&C == 0
4544 if (Op1->hasOneUse() && match(Op1, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
4545 MaskedValueIsZero(Op0, C1->getValue())) {
4546 Instruction *NOr = BinaryOperator::CreateOr(A, Op0);
4547 InsertNewInstBefore(NOr, I);
4549 return BinaryOperator::CreateXor(NOr, C1);
4553 Value *C = 0, *D = 0;
4554 if (match(Op0, m_And(m_Value(A), m_Value(C))) &&
4555 match(Op1, m_And(m_Value(B), m_Value(D)))) {
4556 Value *V1 = 0, *V2 = 0, *V3 = 0;
4557 C1 = dyn_cast<ConstantInt>(C);
4558 C2 = dyn_cast<ConstantInt>(D);
4559 if (C1 && C2) { // (A & C1)|(B & C2)
4560 // If we have: ((V + N) & C1) | (V & C2)
4561 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
4562 // replace with V+N.
4563 if (C1->getValue() == ~C2->getValue()) {
4564 if ((C2->getValue() & (C2->getValue()+1)) == 0 && // C2 == 0+1+
4565 match(A, m_Add(m_Value(V1), m_Value(V2)))) {
4566 // Add commutes, try both ways.
4567 if (V1 == B && MaskedValueIsZero(V2, C2->getValue()))
4568 return ReplaceInstUsesWith(I, A);
4569 if (V2 == B && MaskedValueIsZero(V1, C2->getValue()))
4570 return ReplaceInstUsesWith(I, A);
4572 // Or commutes, try both ways.
4573 if ((C1->getValue() & (C1->getValue()+1)) == 0 &&
4574 match(B, m_Add(m_Value(V1), m_Value(V2)))) {
4575 // Add commutes, try both ways.
4576 if (V1 == A && MaskedValueIsZero(V2, C1->getValue()))
4577 return ReplaceInstUsesWith(I, B);
4578 if (V2 == A && MaskedValueIsZero(V1, C1->getValue()))
4579 return ReplaceInstUsesWith(I, B);
4582 V1 = 0; V2 = 0; V3 = 0;
4585 // Check to see if we have any common things being and'ed. If so, find the
4586 // terms for V1 & (V2|V3).
4587 if (isOnlyUse(Op0) || isOnlyUse(Op1)) {
4588 if (A == B) // (A & C)|(A & D) == A & (C|D)
4589 V1 = A, V2 = C, V3 = D;
4590 else if (A == D) // (A & C)|(B & A) == A & (B|C)
4591 V1 = A, V2 = B, V3 = C;
4592 else if (C == B) // (A & C)|(C & D) == C & (A|D)
4593 V1 = C, V2 = A, V3 = D;
4594 else if (C == D) // (A & C)|(B & C) == C & (A|B)
4595 V1 = C, V2 = A, V3 = B;
4599 InsertNewInstBefore(BinaryOperator::CreateOr(V2, V3, "tmp"), I);
4600 return BinaryOperator::CreateAnd(V1, Or);
4604 // (A & (C0?-1:0)) | (B & ~(C0?-1:0)) -> C0 ? A : B, and commuted variants
4605 if (Instruction *Match = MatchSelectFromAndOr(A, B, C, D))
4607 if (Instruction *Match = MatchSelectFromAndOr(B, A, D, C))
4609 if (Instruction *Match = MatchSelectFromAndOr(C, B, A, D))
4611 if (Instruction *Match = MatchSelectFromAndOr(D, A, B, C))
4616 // ((A&~B)|(~A&B)) -> A^B
4617 if ((match(C, m_Not(m_Value(V1))) &&
4618 match(B, m_Not(m_Value(V2)))))
4619 if (V1 == D && V2 == A)
4620 return BinaryOperator::CreateXor(V1, V2);
4621 // ((~B&A)|(~A&B)) -> A^B
4622 if ((match(A, m_Not(m_Value(V1))) &&
4623 match(B, m_Not(m_Value(V2)))))
4624 if (V1 == D && V2 == C)
4625 return BinaryOperator::CreateXor(V1, V2);
4626 // ((A&~B)|(B&~A)) -> A^B
4627 if ((match(C, m_Not(m_Value(V1))) &&
4628 match(D, m_Not(m_Value(V2)))))
4629 if (V1 == B && V2 == A)
4630 return BinaryOperator::CreateXor(V1, V2);
4631 // ((~B&A)|(B&~A)) -> A^B
4632 if ((match(A, m_Not(m_Value(V1))) &&
4633 match(D, m_Not(m_Value(V2)))))
4634 if (V1 == B && V2 == C)
4635 return BinaryOperator::CreateXor(V1, V2);
4638 // (X >> Z) | (Y >> Z) -> (X|Y) >> Z for all shifts.
4639 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
4640 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
4641 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
4642 SI0->getOperand(1) == SI1->getOperand(1) &&
4643 (SI0->hasOneUse() || SI1->hasOneUse())) {
4644 Instruction *NewOp =
4645 InsertNewInstBefore(BinaryOperator::CreateOr(SI0->getOperand(0),
4647 SI0->getName()), I);
4648 return BinaryOperator::Create(SI1->getOpcode(), NewOp,
4649 SI1->getOperand(1));
4653 // ((A|B)&1)|(B&-2) -> (A&1) | B
4654 if (match(Op0, m_And(m_Or(m_Value(A), m_Value(B)), m_Value(C))) ||
4655 match(Op0, m_And(m_Value(C), m_Or(m_Value(A), m_Value(B))))) {
4656 if (ConstantInt *CI = dyn_cast<ConstantInt>(C)) {
4657 if (CI->getValue() == 1) {
4658 Value *V1 = 0, *C2 = 0;
4659 if (match(Op1, m_And(m_Value(V1), m_Value(C2)))) {
4660 ConstantInt *CI2 = dyn_cast<ConstantInt>(C2);
4664 CI2 = dyn_cast<ConstantInt>(C2);
4668 APInt NegTwo = -APInt(CI2->getValue().getBitWidth(), 2, true);
4669 if (CI2->getValue().eq(NegTwo)) {
4671 Instruction *NewOp =
4672 InsertNewInstBefore(BinaryOperator::CreateAnd(A, CI), I);
4673 return BinaryOperator::CreateOr(NewOp, B);
4676 Instruction *NewOp =
4677 InsertNewInstBefore(BinaryOperator::CreateAnd(B, CI), I);
4678 return BinaryOperator::CreateOr(NewOp, A);
4686 // (B&-2)|((A|B)&1) -> (A&1) | B
4687 if (match(Op1, m_And(m_Or(m_Value(A), m_Value(B)), m_Value(C))) ||
4688 match(Op1, m_And(m_Value(C), m_Or(m_Value(A), m_Value(B))))) {
4689 if (ConstantInt *CI = dyn_cast<ConstantInt>(C)) {
4690 if (CI->getValue() == 1) {
4691 Value *V1 = 0, *C2 = 0;
4692 if (match(Op0, m_And(m_Value(V1), m_Value(C2)))) {
4693 ConstantInt *CI2 = dyn_cast<ConstantInt>(C2);
4697 CI2 = dyn_cast<ConstantInt>(C2);
4701 APInt NegTwo = -APInt(CI2->getValue().getBitWidth(), 2, true);
4702 if (CI2->getValue().eq(NegTwo)) {
4704 Instruction *NewOp =
4705 InsertNewInstBefore(BinaryOperator::CreateAnd(A, CI), I);
4706 return BinaryOperator::CreateOr(NewOp, B);
4709 Instruction *NewOp =
4710 InsertNewInstBefore(BinaryOperator::CreateAnd(B, CI), I);
4711 return BinaryOperator::CreateOr(NewOp, A);
4720 if (match(Op0, m_Not(m_Value(A)))) { // ~A | Op1
4721 if (A == Op1) // ~A | A == -1
4722 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4726 // Note, A is still live here!
4727 if (match(Op1, m_Not(m_Value(B)))) { // Op0 | ~B
4729 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4731 // (~A | ~B) == (~(A & B)) - De Morgan's Law
4732 if (A && isOnlyUse(Op0) && isOnlyUse(Op1)) {
4733 Value *And = InsertNewInstBefore(BinaryOperator::CreateAnd(A, B,
4734 I.getName()+".demorgan"), I);
4735 return BinaryOperator::CreateNot(And);
4739 // (icmp1 A, B) | (icmp2 A, B) --> (icmp3 A, B)
4740 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1))) {
4741 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
4744 if (ICmpInst *LHS = dyn_cast<ICmpInst>(I.getOperand(0)))
4745 if (Instruction *Res = FoldOrOfICmps(I, LHS, RHS))
4749 // fold (or (cast A), (cast B)) -> (cast (or A, B))
4750 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
4751 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4752 if (Op0C->getOpcode() == Op1C->getOpcode()) {// same cast kind ?
4753 if (!isa<ICmpInst>(Op0C->getOperand(0)) ||
4754 !isa<ICmpInst>(Op1C->getOperand(0))) {
4755 const Type *SrcTy = Op0C->getOperand(0)->getType();
4756 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
4757 // Only do this if the casts both really cause code to be
4759 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4761 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4763 Instruction *NewOp = BinaryOperator::CreateOr(Op0C->getOperand(0),
4764 Op1C->getOperand(0),
4766 InsertNewInstBefore(NewOp, I);
4767 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
4774 // (fcmp uno x, c) | (fcmp uno y, c) -> (fcmp uno x, y)
4775 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0))) {
4776 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1))) {
4777 if (LHS->getPredicate() == FCmpInst::FCMP_UNO &&
4778 RHS->getPredicate() == FCmpInst::FCMP_UNO &&
4779 LHS->getOperand(0)->getType() == RHS->getOperand(0)->getType()) {
4780 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
4781 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
4782 // If either of the constants are nans, then the whole thing returns
4784 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
4785 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4787 // Otherwise, no need to compare the two constants, compare the
4789 return new FCmpInst(FCmpInst::FCMP_UNO, LHS->getOperand(0),
4790 RHS->getOperand(0));
4793 Value *Op0LHS, *Op0RHS, *Op1LHS, *Op1RHS;
4794 FCmpInst::Predicate Op0CC, Op1CC;
4795 if (match(Op0, m_FCmp(Op0CC, m_Value(Op0LHS), m_Value(Op0RHS))) &&
4796 match(Op1, m_FCmp(Op1CC, m_Value(Op1LHS), m_Value(Op1RHS)))) {
4797 if (Op0LHS == Op1RHS && Op0RHS == Op1LHS) {
4798 // Swap RHS operands to match LHS.
4799 Op1CC = FCmpInst::getSwappedPredicate(Op1CC);
4800 std::swap(Op1LHS, Op1RHS);
4802 if (Op0LHS == Op1LHS && Op0RHS == Op1RHS) {
4803 // Simplify (fcmp cc0 x, y) | (fcmp cc1 x, y).
4805 return new FCmpInst((FCmpInst::Predicate)Op0CC, Op0LHS, Op0RHS);
4806 else if (Op0CC == FCmpInst::FCMP_TRUE ||
4807 Op1CC == FCmpInst::FCMP_TRUE)
4808 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4809 else if (Op0CC == FCmpInst::FCMP_FALSE)
4810 return ReplaceInstUsesWith(I, Op1);
4811 else if (Op1CC == FCmpInst::FCMP_FALSE)
4812 return ReplaceInstUsesWith(I, Op0);
4815 unsigned Op0Pred = getFCmpCode(Op0CC, Op0Ordered);
4816 unsigned Op1Pred = getFCmpCode(Op1CC, Op1Ordered);
4817 if (Op0Ordered == Op1Ordered) {
4818 // If both are ordered or unordered, return a new fcmp with
4819 // or'ed predicates.
4820 Value *RV = getFCmpValue(Op0Ordered, Op0Pred|Op1Pred,
4822 if (Instruction *I = dyn_cast<Instruction>(RV))
4824 // Otherwise, it's a constant boolean value...
4825 return ReplaceInstUsesWith(I, RV);
4833 return Changed ? &I : 0;
4838 // XorSelf - Implements: X ^ X --> 0
4841 XorSelf(Value *rhs) : RHS(rhs) {}
4842 bool shouldApply(Value *LHS) const { return LHS == RHS; }
4843 Instruction *apply(BinaryOperator &Xor) const {
4850 Instruction *InstCombiner::visitXor(BinaryOperator &I) {
4851 bool Changed = SimplifyCommutative(I);
4852 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4854 if (isa<UndefValue>(Op1)) {
4855 if (isa<UndefValue>(Op0))
4856 // Handle undef ^ undef -> 0 special case. This is a common
4858 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
4859 return ReplaceInstUsesWith(I, Op1); // X ^ undef -> undef
4862 // xor X, X = 0, even if X is nested in a sequence of Xor's.
4863 if (Instruction *Result = AssociativeOpt(I, XorSelf(Op1))) {
4864 assert(Result == &I && "AssociativeOpt didn't work?"); Result=Result;
4865 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
4868 // See if we can simplify any instructions used by the instruction whose sole
4869 // purpose is to compute bits we don't care about.
4870 if (!isa<VectorType>(I.getType())) {
4871 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
4872 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
4873 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
4874 KnownZero, KnownOne))
4876 } else if (isa<ConstantAggregateZero>(Op1)) {
4877 return ReplaceInstUsesWith(I, Op0); // X ^ <0,0> -> X
4880 // Is this a ~ operation?
4881 if (Value *NotOp = dyn_castNotVal(&I)) {
4882 // ~(~X & Y) --> (X | ~Y) - De Morgan's Law
4883 // ~(~X | Y) === (X & ~Y) - De Morgan's Law
4884 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(NotOp)) {
4885 if (Op0I->getOpcode() == Instruction::And ||
4886 Op0I->getOpcode() == Instruction::Or) {
4887 if (dyn_castNotVal(Op0I->getOperand(1))) Op0I->swapOperands();
4888 if (Value *Op0NotVal = dyn_castNotVal(Op0I->getOperand(0))) {
4890 BinaryOperator::CreateNot(Op0I->getOperand(1),
4891 Op0I->getOperand(1)->getName()+".not");
4892 InsertNewInstBefore(NotY, I);
4893 if (Op0I->getOpcode() == Instruction::And)
4894 return BinaryOperator::CreateOr(Op0NotVal, NotY);
4896 return BinaryOperator::CreateAnd(Op0NotVal, NotY);
4903 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
4904 // xor (cmp A, B), true = not (cmp A, B) = !cmp A, B
4905 if (RHS == ConstantInt::getTrue() && Op0->hasOneUse()) {
4906 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Op0))
4907 return new ICmpInst(ICI->getInversePredicate(),
4908 ICI->getOperand(0), ICI->getOperand(1));
4910 if (FCmpInst *FCI = dyn_cast<FCmpInst>(Op0))
4911 return new FCmpInst(FCI->getInversePredicate(),
4912 FCI->getOperand(0), FCI->getOperand(1));
4915 // fold (xor(zext(cmp)), 1) and (xor(sext(cmp)), -1) to ext(!cmp).
4916 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
4917 if (CmpInst *CI = dyn_cast<CmpInst>(Op0C->getOperand(0))) {
4918 if (CI->hasOneUse() && Op0C->hasOneUse()) {
4919 Instruction::CastOps Opcode = Op0C->getOpcode();
4920 if (Opcode == Instruction::ZExt || Opcode == Instruction::SExt) {
4921 if (RHS == ConstantExpr::getCast(Opcode, ConstantInt::getTrue(),
4922 Op0C->getDestTy())) {
4923 Instruction *NewCI = InsertNewInstBefore(CmpInst::Create(
4924 CI->getOpcode(), CI->getInversePredicate(),
4925 CI->getOperand(0), CI->getOperand(1)), I);
4926 NewCI->takeName(CI);
4927 return CastInst::Create(Opcode, NewCI, Op0C->getType());
4934 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
4935 // ~(c-X) == X-c-1 == X+(-c-1)
4936 if (Op0I->getOpcode() == Instruction::Sub && RHS->isAllOnesValue())
4937 if (Constant *Op0I0C = dyn_cast<Constant>(Op0I->getOperand(0))) {
4938 Constant *NegOp0I0C = ConstantExpr::getNeg(Op0I0C);
4939 Constant *ConstantRHS = ConstantExpr::getSub(NegOp0I0C,
4940 ConstantInt::get(I.getType(), 1));
4941 return BinaryOperator::CreateAdd(Op0I->getOperand(1), ConstantRHS);
4944 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) {
4945 if (Op0I->getOpcode() == Instruction::Add) {
4946 // ~(X-c) --> (-c-1)-X
4947 if (RHS->isAllOnesValue()) {
4948 Constant *NegOp0CI = ConstantExpr::getNeg(Op0CI);
4949 return BinaryOperator::CreateSub(
4950 ConstantExpr::getSub(NegOp0CI,
4951 ConstantInt::get(I.getType(), 1)),
4952 Op0I->getOperand(0));
4953 } else if (RHS->getValue().isSignBit()) {
4954 // (X + C) ^ signbit -> (X + C + signbit)
4955 Constant *C = ConstantInt::get(RHS->getValue() + Op0CI->getValue());
4956 return BinaryOperator::CreateAdd(Op0I->getOperand(0), C);
4959 } else if (Op0I->getOpcode() == Instruction::Or) {
4960 // (X|C1)^C2 -> X^(C1|C2) iff X&~C1 == 0
4961 if (MaskedValueIsZero(Op0I->getOperand(0), Op0CI->getValue())) {
4962 Constant *NewRHS = ConstantExpr::getOr(Op0CI, RHS);
4963 // Anything in both C1 and C2 is known to be zero, remove it from
4965 Constant *CommonBits = And(Op0CI, RHS);
4966 NewRHS = ConstantExpr::getAnd(NewRHS,
4967 ConstantExpr::getNot(CommonBits));
4968 AddToWorkList(Op0I);
4969 I.setOperand(0, Op0I->getOperand(0));
4970 I.setOperand(1, NewRHS);
4977 // Try to fold constant and into select arguments.
4978 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
4979 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
4981 if (isa<PHINode>(Op0))
4982 if (Instruction *NV = FoldOpIntoPhi(I))
4986 if (Value *X = dyn_castNotVal(Op0)) // ~A ^ A == -1
4988 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4990 if (Value *X = dyn_castNotVal(Op1)) // A ^ ~A == -1
4992 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4995 BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1);
4998 if (match(Op1I, m_Or(m_Value(A), m_Value(B)))) {
4999 if (A == Op0) { // B^(B|A) == (A|B)^B
5000 Op1I->swapOperands();
5002 std::swap(Op0, Op1);
5003 } else if (B == Op0) { // B^(A|B) == (A|B)^B
5004 I.swapOperands(); // Simplified below.
5005 std::swap(Op0, Op1);
5007 } else if (match(Op1I, m_Xor(m_Specific(Op0), m_Value(B)))) {
5008 return ReplaceInstUsesWith(I, B); // A^(A^B) == B
5009 } else if (match(Op1I, m_Xor(m_Value(A), m_Specific(Op0)))) {
5010 return ReplaceInstUsesWith(I, A); // A^(B^A) == B
5011 } else if (match(Op1I, m_And(m_Value(A), m_Value(B))) && Op1I->hasOneUse()){
5012 if (A == Op0) { // A^(A&B) -> A^(B&A)
5013 Op1I->swapOperands();
5016 if (B == Op0) { // A^(B&A) -> (B&A)^A
5017 I.swapOperands(); // Simplified below.
5018 std::swap(Op0, Op1);
5023 BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0);
5026 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) && Op0I->hasOneUse()) {
5027 if (A == Op1) // (B|A)^B == (A|B)^B
5029 if (B == Op1) { // (A|B)^B == A & ~B
5031 InsertNewInstBefore(BinaryOperator::CreateNot(Op1, "tmp"), I);
5032 return BinaryOperator::CreateAnd(A, NotB);
5034 } else if (match(Op0I, m_Xor(m_Specific(Op1), m_Value(B)))) {
5035 return ReplaceInstUsesWith(I, B); // (A^B)^A == B
5036 } else if (match(Op0I, m_Xor(m_Value(A), m_Specific(Op1)))) {
5037 return ReplaceInstUsesWith(I, A); // (B^A)^A == B
5038 } else if (match(Op0I, m_And(m_Value(A), m_Value(B))) && Op0I->hasOneUse()){
5039 if (A == Op1) // (A&B)^A -> (B&A)^A
5041 if (B == Op1 && // (B&A)^A == ~B & A
5042 !isa<ConstantInt>(Op1)) { // Canonical form is (B&C)^C
5044 InsertNewInstBefore(BinaryOperator::CreateNot(A, "tmp"), I);
5045 return BinaryOperator::CreateAnd(N, Op1);
5050 // (X >> Z) ^ (Y >> Z) -> (X^Y) >> Z for all shifts.
5051 if (Op0I && Op1I && Op0I->isShift() &&
5052 Op0I->getOpcode() == Op1I->getOpcode() &&
5053 Op0I->getOperand(1) == Op1I->getOperand(1) &&
5054 (Op1I->hasOneUse() || Op1I->hasOneUse())) {
5055 Instruction *NewOp =
5056 InsertNewInstBefore(BinaryOperator::CreateXor(Op0I->getOperand(0),
5057 Op1I->getOperand(0),
5058 Op0I->getName()), I);
5059 return BinaryOperator::Create(Op1I->getOpcode(), NewOp,
5060 Op1I->getOperand(1));
5064 Value *A, *B, *C, *D;
5065 // (A & B)^(A | B) -> A ^ B
5066 if (match(Op0I, m_And(m_Value(A), m_Value(B))) &&
5067 match(Op1I, m_Or(m_Value(C), m_Value(D)))) {
5068 if ((A == C && B == D) || (A == D && B == C))
5069 return BinaryOperator::CreateXor(A, B);
5071 // (A | B)^(A & B) -> A ^ B
5072 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) &&
5073 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
5074 if ((A == C && B == D) || (A == D && B == C))
5075 return BinaryOperator::CreateXor(A, B);
5079 if ((Op0I->hasOneUse() || Op1I->hasOneUse()) &&
5080 match(Op0I, m_And(m_Value(A), m_Value(B))) &&
5081 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
5082 // (X & Y)^(X & Y) -> (Y^Z) & X
5083 Value *X = 0, *Y = 0, *Z = 0;
5085 X = A, Y = B, Z = D;
5087 X = A, Y = B, Z = C;
5089 X = B, Y = A, Z = D;
5091 X = B, Y = A, Z = C;
5094 Instruction *NewOp =
5095 InsertNewInstBefore(BinaryOperator::CreateXor(Y, Z, Op0->getName()), I);
5096 return BinaryOperator::CreateAnd(NewOp, X);
5101 // (icmp1 A, B) ^ (icmp2 A, B) --> (icmp3 A, B)
5102 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1)))
5103 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
5106 // fold (xor (cast A), (cast B)) -> (cast (xor A, B))
5107 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
5108 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
5109 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind?
5110 const Type *SrcTy = Op0C->getOperand(0)->getType();
5111 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
5112 // Only do this if the casts both really cause code to be generated.
5113 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
5115 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
5117 Instruction *NewOp = BinaryOperator::CreateXor(Op0C->getOperand(0),
5118 Op1C->getOperand(0),
5120 InsertNewInstBefore(NewOp, I);
5121 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
5126 return Changed ? &I : 0;
5129 /// AddWithOverflow - Compute Result = In1+In2, returning true if the result
5130 /// overflowed for this type.
5131 static bool AddWithOverflow(ConstantInt *&Result, ConstantInt *In1,
5132 ConstantInt *In2, bool IsSigned = false) {
5133 Result = cast<ConstantInt>(Add(In1, In2));
5136 if (In2->getValue().isNegative())
5137 return Result->getValue().sgt(In1->getValue());
5139 return Result->getValue().slt(In1->getValue());
5141 return Result->getValue().ult(In1->getValue());
5144 /// SubWithOverflow - Compute Result = In1-In2, returning true if the result
5145 /// overflowed for this type.
5146 static bool SubWithOverflow(ConstantInt *&Result, ConstantInt *In1,
5147 ConstantInt *In2, bool IsSigned = false) {
5148 Result = cast<ConstantInt>(Subtract(In1, In2));
5151 if (In2->getValue().isNegative())
5152 return Result->getValue().slt(In1->getValue());
5154 return Result->getValue().sgt(In1->getValue());
5156 return Result->getValue().ugt(In1->getValue());
5159 /// EmitGEPOffset - Given a getelementptr instruction/constantexpr, emit the
5160 /// code necessary to compute the offset from the base pointer (without adding
5161 /// in the base pointer). Return the result as a signed integer of intptr size.
5162 static Value *EmitGEPOffset(User *GEP, Instruction &I, InstCombiner &IC) {
5163 TargetData &TD = IC.getTargetData();
5164 gep_type_iterator GTI = gep_type_begin(GEP);
5165 const Type *IntPtrTy = TD.getIntPtrType();
5166 Value *Result = Constant::getNullValue(IntPtrTy);
5168 // Build a mask for high order bits.
5169 unsigned IntPtrWidth = TD.getPointerSizeInBits();
5170 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth);
5172 for (User::op_iterator i = GEP->op_begin() + 1, e = GEP->op_end(); i != e;
5175 uint64_t Size = TD.getABITypeSize(GTI.getIndexedType()) & PtrSizeMask;
5176 if (ConstantInt *OpC = dyn_cast<ConstantInt>(Op)) {
5177 if (OpC->isZero()) continue;
5179 // Handle a struct index, which adds its field offset to the pointer.
5180 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
5181 Size = TD.getStructLayout(STy)->getElementOffset(OpC->getZExtValue());
5183 if (ConstantInt *RC = dyn_cast<ConstantInt>(Result))
5184 Result = ConstantInt::get(RC->getValue() + APInt(IntPtrWidth, Size));
5186 Result = IC.InsertNewInstBefore(
5187 BinaryOperator::CreateAdd(Result,
5188 ConstantInt::get(IntPtrTy, Size),
5189 GEP->getName()+".offs"), I);
5193 Constant *Scale = ConstantInt::get(IntPtrTy, Size);
5194 Constant *OC = ConstantExpr::getIntegerCast(OpC, IntPtrTy, true /*SExt*/);
5195 Scale = ConstantExpr::getMul(OC, Scale);
5196 if (Constant *RC = dyn_cast<Constant>(Result))
5197 Result = ConstantExpr::getAdd(RC, Scale);
5199 // Emit an add instruction.
5200 Result = IC.InsertNewInstBefore(
5201 BinaryOperator::CreateAdd(Result, Scale,
5202 GEP->getName()+".offs"), I);
5206 // Convert to correct type.
5207 if (Op->getType() != IntPtrTy) {
5208 if (Constant *OpC = dyn_cast<Constant>(Op))
5209 Op = ConstantExpr::getSExt(OpC, IntPtrTy);
5211 Op = IC.InsertNewInstBefore(new SExtInst(Op, IntPtrTy,
5212 Op->getName()+".c"), I);
5215 Constant *Scale = ConstantInt::get(IntPtrTy, Size);
5216 if (Constant *OpC = dyn_cast<Constant>(Op))
5217 Op = ConstantExpr::getMul(OpC, Scale);
5218 else // We'll let instcombine(mul) convert this to a shl if possible.
5219 Op = IC.InsertNewInstBefore(BinaryOperator::CreateMul(Op, Scale,
5220 GEP->getName()+".idx"), I);
5223 // Emit an add instruction.
5224 if (isa<Constant>(Op) && isa<Constant>(Result))
5225 Result = ConstantExpr::getAdd(cast<Constant>(Op),
5226 cast<Constant>(Result));
5228 Result = IC.InsertNewInstBefore(BinaryOperator::CreateAdd(Op, Result,
5229 GEP->getName()+".offs"), I);
5235 /// EvaluateGEPOffsetExpression - Return an value that can be used to compare of
5236 /// the *offset* implied by GEP to zero. For example, if we have &A[i], we want
5237 /// to return 'i' for "icmp ne i, 0". Note that, in general, indices can be
5238 /// complex, and scales are involved. The above expression would also be legal
5239 /// to codegen as "icmp ne (i*4), 0" (assuming A is a pointer to i32). This
5240 /// later form is less amenable to optimization though, and we are allowed to
5241 /// generate the first by knowing that pointer arithmetic doesn't overflow.
5243 /// If we can't emit an optimized form for this expression, this returns null.
5245 static Value *EvaluateGEPOffsetExpression(User *GEP, Instruction &I,
5247 TargetData &TD = IC.getTargetData();
5248 gep_type_iterator GTI = gep_type_begin(GEP);
5250 // Check to see if this gep only has a single variable index. If so, and if
5251 // any constant indices are a multiple of its scale, then we can compute this
5252 // in terms of the scale of the variable index. For example, if the GEP
5253 // implies an offset of "12 + i*4", then we can codegen this as "3 + i",
5254 // because the expression will cross zero at the same point.
5255 unsigned i, e = GEP->getNumOperands();
5257 for (i = 1; i != e; ++i, ++GTI) {
5258 if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i))) {
5259 // Compute the aggregate offset of constant indices.
5260 if (CI->isZero()) continue;
5262 // Handle a struct index, which adds its field offset to the pointer.
5263 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
5264 Offset += TD.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
5266 uint64_t Size = TD.getABITypeSize(GTI.getIndexedType());
5267 Offset += Size*CI->getSExtValue();
5270 // Found our variable index.
5275 // If there are no variable indices, we must have a constant offset, just
5276 // evaluate it the general way.
5277 if (i == e) return 0;
5279 Value *VariableIdx = GEP->getOperand(i);
5280 // Determine the scale factor of the variable element. For example, this is
5281 // 4 if the variable index is into an array of i32.
5282 uint64_t VariableScale = TD.getABITypeSize(GTI.getIndexedType());
5284 // Verify that there are no other variable indices. If so, emit the hard way.
5285 for (++i, ++GTI; i != e; ++i, ++GTI) {
5286 ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i));
5289 // Compute the aggregate offset of constant indices.
5290 if (CI->isZero()) continue;
5292 // Handle a struct index, which adds its field offset to the pointer.
5293 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
5294 Offset += TD.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
5296 uint64_t Size = TD.getABITypeSize(GTI.getIndexedType());
5297 Offset += Size*CI->getSExtValue();
5301 // Okay, we know we have a single variable index, which must be a
5302 // pointer/array/vector index. If there is no offset, life is simple, return
5304 unsigned IntPtrWidth = TD.getPointerSizeInBits();
5306 // Cast to intptrty in case a truncation occurs. If an extension is needed,
5307 // we don't need to bother extending: the extension won't affect where the
5308 // computation crosses zero.
5309 if (VariableIdx->getType()->getPrimitiveSizeInBits() > IntPtrWidth)
5310 VariableIdx = new TruncInst(VariableIdx, TD.getIntPtrType(),
5311 VariableIdx->getNameStart(), &I);
5315 // Otherwise, there is an index. The computation we will do will be modulo
5316 // the pointer size, so get it.
5317 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth);
5319 Offset &= PtrSizeMask;
5320 VariableScale &= PtrSizeMask;
5322 // To do this transformation, any constant index must be a multiple of the
5323 // variable scale factor. For example, we can evaluate "12 + 4*i" as "3 + i",
5324 // but we can't evaluate "10 + 3*i" in terms of i. Check that the offset is a
5325 // multiple of the variable scale.
5326 int64_t NewOffs = Offset / (int64_t)VariableScale;
5327 if (Offset != NewOffs*(int64_t)VariableScale)
5330 // Okay, we can do this evaluation. Start by converting the index to intptr.
5331 const Type *IntPtrTy = TD.getIntPtrType();
5332 if (VariableIdx->getType() != IntPtrTy)
5333 VariableIdx = CastInst::CreateIntegerCast(VariableIdx, IntPtrTy,
5335 VariableIdx->getNameStart(), &I);
5336 Constant *OffsetVal = ConstantInt::get(IntPtrTy, NewOffs);
5337 return BinaryOperator::CreateAdd(VariableIdx, OffsetVal, "offset", &I);
5341 /// FoldGEPICmp - Fold comparisons between a GEP instruction and something
5342 /// else. At this point we know that the GEP is on the LHS of the comparison.
5343 Instruction *InstCombiner::FoldGEPICmp(User *GEPLHS, Value *RHS,
5344 ICmpInst::Predicate Cond,
5346 assert(dyn_castGetElementPtr(GEPLHS) && "LHS is not a getelementptr!");
5348 // Look through bitcasts.
5349 if (BitCastInst *BCI = dyn_cast<BitCastInst>(RHS))
5350 RHS = BCI->getOperand(0);
5352 Value *PtrBase = GEPLHS->getOperand(0);
5353 if (PtrBase == RHS) {
5354 // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
5355 // This transformation (ignoring the base and scales) is valid because we
5356 // know pointers can't overflow. See if we can output an optimized form.
5357 Value *Offset = EvaluateGEPOffsetExpression(GEPLHS, I, *this);
5359 // If not, synthesize the offset the hard way.
5361 Offset = EmitGEPOffset(GEPLHS, I, *this);
5362 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Offset,
5363 Constant::getNullValue(Offset->getType()));
5364 } else if (User *GEPRHS = dyn_castGetElementPtr(RHS)) {
5365 // If the base pointers are different, but the indices are the same, just
5366 // compare the base pointer.
5367 if (PtrBase != GEPRHS->getOperand(0)) {
5368 bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands();
5369 IndicesTheSame &= GEPLHS->getOperand(0)->getType() ==
5370 GEPRHS->getOperand(0)->getType();
5372 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
5373 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
5374 IndicesTheSame = false;
5378 // If all indices are the same, just compare the base pointers.
5380 return new ICmpInst(ICmpInst::getSignedPredicate(Cond),
5381 GEPLHS->getOperand(0), GEPRHS->getOperand(0));
5383 // Otherwise, the base pointers are different and the indices are
5384 // different, bail out.
5388 // If one of the GEPs has all zero indices, recurse.
5389 bool AllZeros = true;
5390 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
5391 if (!isa<Constant>(GEPLHS->getOperand(i)) ||
5392 !cast<Constant>(GEPLHS->getOperand(i))->isNullValue()) {
5397 return FoldGEPICmp(GEPRHS, GEPLHS->getOperand(0),
5398 ICmpInst::getSwappedPredicate(Cond), I);
5400 // If the other GEP has all zero indices, recurse.
5402 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
5403 if (!isa<Constant>(GEPRHS->getOperand(i)) ||
5404 !cast<Constant>(GEPRHS->getOperand(i))->isNullValue()) {
5409 return FoldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I);
5411 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
5412 // If the GEPs only differ by one index, compare it.
5413 unsigned NumDifferences = 0; // Keep track of # differences.
5414 unsigned DiffOperand = 0; // The operand that differs.
5415 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
5416 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
5417 if (GEPLHS->getOperand(i)->getType()->getPrimitiveSizeInBits() !=
5418 GEPRHS->getOperand(i)->getType()->getPrimitiveSizeInBits()) {
5419 // Irreconcilable differences.
5423 if (NumDifferences++) break;
5428 if (NumDifferences == 0) // SAME GEP?
5429 return ReplaceInstUsesWith(I, // No comparison is needed here.
5430 ConstantInt::get(Type::Int1Ty,
5431 ICmpInst::isTrueWhenEqual(Cond)));
5433 else if (NumDifferences == 1) {
5434 Value *LHSV = GEPLHS->getOperand(DiffOperand);
5435 Value *RHSV = GEPRHS->getOperand(DiffOperand);
5436 // Make sure we do a signed comparison here.
5437 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), LHSV, RHSV);
5441 // Only lower this if the icmp is the only user of the GEP or if we expect
5442 // the result to fold to a constant!
5443 if ((isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
5444 (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
5445 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
5446 Value *L = EmitGEPOffset(GEPLHS, I, *this);
5447 Value *R = EmitGEPOffset(GEPRHS, I, *this);
5448 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), L, R);
5454 /// FoldFCmp_IntToFP_Cst - Fold fcmp ([us]itofp x, cst) if possible.
5456 Instruction *InstCombiner::FoldFCmp_IntToFP_Cst(FCmpInst &I,
5459 if (!isa<ConstantFP>(RHSC)) return 0;
5460 const APFloat &RHS = cast<ConstantFP>(RHSC)->getValueAPF();
5462 // Get the width of the mantissa. We don't want to hack on conversions that
5463 // might lose information from the integer, e.g. "i64 -> float"
5464 int MantissaWidth = LHSI->getType()->getFPMantissaWidth();
5465 if (MantissaWidth == -1) return 0; // Unknown.
5467 // Check to see that the input is converted from an integer type that is small
5468 // enough that preserves all bits. TODO: check here for "known" sign bits.
5469 // This would allow us to handle (fptosi (x >>s 62) to float) if x is i64 f.e.
5470 unsigned InputSize = LHSI->getOperand(0)->getType()->getPrimitiveSizeInBits();
5472 // If this is a uitofp instruction, we need an extra bit to hold the sign.
5473 bool LHSUnsigned = isa<UIToFPInst>(LHSI);
5477 // If the conversion would lose info, don't hack on this.
5478 if ((int)InputSize > MantissaWidth)
5481 // Otherwise, we can potentially simplify the comparison. We know that it
5482 // will always come through as an integer value and we know the constant is
5483 // not a NAN (it would have been previously simplified).
5484 assert(!RHS.isNaN() && "NaN comparison not already folded!");
5486 ICmpInst::Predicate Pred;
5487 switch (I.getPredicate()) {
5488 default: assert(0 && "Unexpected predicate!");
5489 case FCmpInst::FCMP_UEQ:
5490 case FCmpInst::FCMP_OEQ:
5491 Pred = ICmpInst::ICMP_EQ;
5493 case FCmpInst::FCMP_UGT:
5494 case FCmpInst::FCMP_OGT:
5495 Pred = LHSUnsigned ? ICmpInst::ICMP_UGT : ICmpInst::ICMP_SGT;
5497 case FCmpInst::FCMP_UGE:
5498 case FCmpInst::FCMP_OGE:
5499 Pred = LHSUnsigned ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_SGE;
5501 case FCmpInst::FCMP_ULT:
5502 case FCmpInst::FCMP_OLT:
5503 Pred = LHSUnsigned ? ICmpInst::ICMP_ULT : ICmpInst::ICMP_SLT;
5505 case FCmpInst::FCMP_ULE:
5506 case FCmpInst::FCMP_OLE:
5507 Pred = LHSUnsigned ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_SLE;
5509 case FCmpInst::FCMP_UNE:
5510 case FCmpInst::FCMP_ONE:
5511 Pred = ICmpInst::ICMP_NE;
5513 case FCmpInst::FCMP_ORD:
5514 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5515 case FCmpInst::FCMP_UNO:
5516 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5519 const IntegerType *IntTy = cast<IntegerType>(LHSI->getOperand(0)->getType());
5521 // Now we know that the APFloat is a normal number, zero or inf.
5523 // See if the FP constant is too large for the integer. For example,
5524 // comparing an i8 to 300.0.
5525 unsigned IntWidth = IntTy->getPrimitiveSizeInBits();
5528 // If the RHS value is > SignedMax, fold the comparison. This handles +INF
5529 // and large values.
5530 APFloat SMax(RHS.getSemantics(), APFloat::fcZero, false);
5531 SMax.convertFromAPInt(APInt::getSignedMaxValue(IntWidth), true,
5532 APFloat::rmNearestTiesToEven);
5533 if (SMax.compare(RHS) == APFloat::cmpLessThan) { // smax < 13123.0
5534 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SLT ||
5535 Pred == ICmpInst::ICMP_SLE)
5536 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5537 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5540 // If the RHS value is > UnsignedMax, fold the comparison. This handles
5541 // +INF and large values.
5542 APFloat UMax(RHS.getSemantics(), APFloat::fcZero, false);
5543 UMax.convertFromAPInt(APInt::getMaxValue(IntWidth), false,
5544 APFloat::rmNearestTiesToEven);
5545 if (UMax.compare(RHS) == APFloat::cmpLessThan) { // umax < 13123.0
5546 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_ULT ||
5547 Pred == ICmpInst::ICMP_ULE)
5548 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5549 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5554 // See if the RHS value is < SignedMin.
5555 APFloat SMin(RHS.getSemantics(), APFloat::fcZero, false);
5556 SMin.convertFromAPInt(APInt::getSignedMinValue(IntWidth), true,
5557 APFloat::rmNearestTiesToEven);
5558 if (SMin.compare(RHS) == APFloat::cmpGreaterThan) { // smin > 12312.0
5559 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SGT ||
5560 Pred == ICmpInst::ICMP_SGE)
5561 return ReplaceInstUsesWith(I,ConstantInt::getTrue());
5562 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5566 // Okay, now we know that the FP constant fits in the range [SMIN, SMAX] or
5567 // [0, UMAX], but it may still be fractional. See if it is fractional by
5568 // casting the FP value to the integer value and back, checking for equality.
5569 // Don't do this for zero, because -0.0 is not fractional.
5570 Constant *RHSInt = ConstantExpr::getFPToSI(RHSC, IntTy);
5571 if (!RHS.isZero() &&
5572 ConstantExpr::getSIToFP(RHSInt, RHSC->getType()) != RHSC) {
5573 // If we had a comparison against a fractional value, we have to adjust the
5574 // compare predicate and sometimes the value. RHSC is rounded towards zero
5577 default: assert(0 && "Unexpected integer comparison!");
5578 case ICmpInst::ICMP_NE: // (float)int != 4.4 --> true
5579 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5580 case ICmpInst::ICMP_EQ: // (float)int == 4.4 --> false
5581 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5582 case ICmpInst::ICMP_ULE:
5583 // (float)int <= 4.4 --> int <= 4
5584 // (float)int <= -4.4 --> false
5585 if (RHS.isNegative())
5586 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5588 case ICmpInst::ICMP_SLE:
5589 // (float)int <= 4.4 --> int <= 4
5590 // (float)int <= -4.4 --> int < -4
5591 if (RHS.isNegative())
5592 Pred = ICmpInst::ICMP_SLT;
5594 case ICmpInst::ICMP_ULT:
5595 // (float)int < -4.4 --> false
5596 // (float)int < 4.4 --> int <= 4
5597 if (RHS.isNegative())
5598 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5599 Pred = ICmpInst::ICMP_ULE;
5601 case ICmpInst::ICMP_SLT:
5602 // (float)int < -4.4 --> int < -4
5603 // (float)int < 4.4 --> int <= 4
5604 if (!RHS.isNegative())
5605 Pred = ICmpInst::ICMP_SLE;
5607 case ICmpInst::ICMP_UGT:
5608 // (float)int > 4.4 --> int > 4
5609 // (float)int > -4.4 --> true
5610 if (RHS.isNegative())
5611 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5613 case ICmpInst::ICMP_SGT:
5614 // (float)int > 4.4 --> int > 4
5615 // (float)int > -4.4 --> int >= -4
5616 if (RHS.isNegative())
5617 Pred = ICmpInst::ICMP_SGE;
5619 case ICmpInst::ICMP_UGE:
5620 // (float)int >= -4.4 --> true
5621 // (float)int >= 4.4 --> int > 4
5622 if (!RHS.isNegative())
5623 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5624 Pred = ICmpInst::ICMP_UGT;
5626 case ICmpInst::ICMP_SGE:
5627 // (float)int >= -4.4 --> int >= -4
5628 // (float)int >= 4.4 --> int > 4
5629 if (!RHS.isNegative())
5630 Pred = ICmpInst::ICMP_SGT;
5635 // Lower this FP comparison into an appropriate integer version of the
5637 return new ICmpInst(Pred, LHSI->getOperand(0), RHSInt);
5640 Instruction *InstCombiner::visitFCmpInst(FCmpInst &I) {
5641 bool Changed = SimplifyCompare(I);
5642 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5644 // Fold trivial predicates.
5645 if (I.getPredicate() == FCmpInst::FCMP_FALSE)
5646 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5647 if (I.getPredicate() == FCmpInst::FCMP_TRUE)
5648 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5650 // Simplify 'fcmp pred X, X'
5652 switch (I.getPredicate()) {
5653 default: assert(0 && "Unknown predicate!");
5654 case FCmpInst::FCMP_UEQ: // True if unordered or equal
5655 case FCmpInst::FCMP_UGE: // True if unordered, greater than, or equal
5656 case FCmpInst::FCMP_ULE: // True if unordered, less than, or equal
5657 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5658 case FCmpInst::FCMP_OGT: // True if ordered and greater than
5659 case FCmpInst::FCMP_OLT: // True if ordered and less than
5660 case FCmpInst::FCMP_ONE: // True if ordered and operands are unequal
5661 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5663 case FCmpInst::FCMP_UNO: // True if unordered: isnan(X) | isnan(Y)
5664 case FCmpInst::FCMP_ULT: // True if unordered or less than
5665 case FCmpInst::FCMP_UGT: // True if unordered or greater than
5666 case FCmpInst::FCMP_UNE: // True if unordered or not equal
5667 // Canonicalize these to be 'fcmp uno %X, 0.0'.
5668 I.setPredicate(FCmpInst::FCMP_UNO);
5669 I.setOperand(1, Constant::getNullValue(Op0->getType()));
5672 case FCmpInst::FCMP_ORD: // True if ordered (no nans)
5673 case FCmpInst::FCMP_OEQ: // True if ordered and equal
5674 case FCmpInst::FCMP_OGE: // True if ordered and greater than or equal
5675 case FCmpInst::FCMP_OLE: // True if ordered and less than or equal
5676 // Canonicalize these to be 'fcmp ord %X, 0.0'.
5677 I.setPredicate(FCmpInst::FCMP_ORD);
5678 I.setOperand(1, Constant::getNullValue(Op0->getType()));
5683 if (isa<UndefValue>(Op1)) // fcmp pred X, undef -> undef
5684 return ReplaceInstUsesWith(I, UndefValue::get(Type::Int1Ty));
5686 // Handle fcmp with constant RHS
5687 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
5688 // If the constant is a nan, see if we can fold the comparison based on it.
5689 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHSC)) {
5690 if (CFP->getValueAPF().isNaN()) {
5691 if (FCmpInst::isOrdered(I.getPredicate())) // True if ordered and...
5692 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5693 assert(FCmpInst::isUnordered(I.getPredicate()) &&
5694 "Comparison must be either ordered or unordered!");
5695 // True if unordered.
5696 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5700 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
5701 switch (LHSI->getOpcode()) {
5702 case Instruction::PHI:
5703 // Only fold fcmp into the PHI if the phi and fcmp are in the same
5704 // block. If in the same block, we're encouraging jump threading. If
5705 // not, we are just pessimizing the code by making an i1 phi.
5706 if (LHSI->getParent() == I.getParent())
5707 if (Instruction *NV = FoldOpIntoPhi(I))
5710 case Instruction::SIToFP:
5711 case Instruction::UIToFP:
5712 if (Instruction *NV = FoldFCmp_IntToFP_Cst(I, LHSI, RHSC))
5715 case Instruction::Select:
5716 // If either operand of the select is a constant, we can fold the
5717 // comparison into the select arms, which will cause one to be
5718 // constant folded and the select turned into a bitwise or.
5719 Value *Op1 = 0, *Op2 = 0;
5720 if (LHSI->hasOneUse()) {
5721 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
5722 // Fold the known value into the constant operand.
5723 Op1 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
5724 // Insert a new FCmp of the other select operand.
5725 Op2 = InsertNewInstBefore(new FCmpInst(I.getPredicate(),
5726 LHSI->getOperand(2), RHSC,
5728 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
5729 // Fold the known value into the constant operand.
5730 Op2 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
5731 // Insert a new FCmp of the other select operand.
5732 Op1 = InsertNewInstBefore(new FCmpInst(I.getPredicate(),
5733 LHSI->getOperand(1), RHSC,
5739 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
5744 return Changed ? &I : 0;
5747 Instruction *InstCombiner::visitICmpInst(ICmpInst &I) {
5748 bool Changed = SimplifyCompare(I);
5749 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5750 const Type *Ty = Op0->getType();
5754 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
5755 I.isTrueWhenEqual()));
5757 if (isa<UndefValue>(Op1)) // X icmp undef -> undef
5758 return ReplaceInstUsesWith(I, UndefValue::get(Type::Int1Ty));
5760 // icmp <global/alloca*/null>, <global/alloca*/null> - Global/Stack value
5761 // addresses never equal each other! We already know that Op0 != Op1.
5762 if ((isa<GlobalValue>(Op0) || isa<AllocaInst>(Op0) ||
5763 isa<ConstantPointerNull>(Op0)) &&
5764 (isa<GlobalValue>(Op1) || isa<AllocaInst>(Op1) ||
5765 isa<ConstantPointerNull>(Op1)))
5766 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
5767 !I.isTrueWhenEqual()));
5769 // icmp's with boolean values can always be turned into bitwise operations
5770 if (Ty == Type::Int1Ty) {
5771 switch (I.getPredicate()) {
5772 default: assert(0 && "Invalid icmp instruction!");
5773 case ICmpInst::ICMP_EQ: { // icmp eq i1 A, B -> ~(A^B)
5774 Instruction *Xor = BinaryOperator::CreateXor(Op0, Op1, I.getName()+"tmp");
5775 InsertNewInstBefore(Xor, I);
5776 return BinaryOperator::CreateNot(Xor);
5778 case ICmpInst::ICMP_NE: // icmp eq i1 A, B -> A^B
5779 return BinaryOperator::CreateXor(Op0, Op1);
5781 case ICmpInst::ICMP_UGT:
5782 std::swap(Op0, Op1); // Change icmp ugt -> icmp ult
5784 case ICmpInst::ICMP_ULT:{ // icmp ult i1 A, B -> ~A & B
5785 Instruction *Not = BinaryOperator::CreateNot(Op0, I.getName()+"tmp");
5786 InsertNewInstBefore(Not, I);
5787 return BinaryOperator::CreateAnd(Not, Op1);
5789 case ICmpInst::ICMP_SGT:
5790 std::swap(Op0, Op1); // Change icmp sgt -> icmp slt
5792 case ICmpInst::ICMP_SLT: { // icmp slt i1 A, B -> A & ~B
5793 Instruction *Not = BinaryOperator::CreateNot(Op1, I.getName()+"tmp");
5794 InsertNewInstBefore(Not, I);
5795 return BinaryOperator::CreateAnd(Not, Op0);
5797 case ICmpInst::ICMP_UGE:
5798 std::swap(Op0, Op1); // Change icmp uge -> icmp ule
5800 case ICmpInst::ICMP_ULE: { // icmp ule i1 A, B -> ~A | B
5801 Instruction *Not = BinaryOperator::CreateNot(Op0, I.getName()+"tmp");
5802 InsertNewInstBefore(Not, I);
5803 return BinaryOperator::CreateOr(Not, Op1);
5805 case ICmpInst::ICMP_SGE:
5806 std::swap(Op0, Op1); // Change icmp sge -> icmp sle
5808 case ICmpInst::ICMP_SLE: { // icmp sle i1 A, B -> A | ~B
5809 Instruction *Not = BinaryOperator::CreateNot(Op1, I.getName()+"tmp");
5810 InsertNewInstBefore(Not, I);
5811 return BinaryOperator::CreateOr(Not, Op0);
5816 // See if we are doing a comparison with a constant.
5817 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
5820 // (icmp ne/eq (sub A B) 0) -> (icmp ne/eq A, B)
5821 if (I.isEquality() && CI->isNullValue() &&
5822 match(Op0, m_Sub(m_Value(A), m_Value(B)))) {
5823 // (icmp cond A B) if cond is equality
5824 return new ICmpInst(I.getPredicate(), A, B);
5827 // If we have an icmp le or icmp ge instruction, turn it into the
5828 // appropriate icmp lt or icmp gt instruction. This allows us to rely on
5829 // them being folded in the code below.
5830 switch (I.getPredicate()) {
5832 case ICmpInst::ICMP_ULE:
5833 if (CI->isMaxValue(false)) // A <=u MAX -> TRUE
5834 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5835 return new ICmpInst(ICmpInst::ICMP_ULT, Op0, AddOne(CI));
5836 case ICmpInst::ICMP_SLE:
5837 if (CI->isMaxValue(true)) // A <=s MAX -> TRUE
5838 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5839 return new ICmpInst(ICmpInst::ICMP_SLT, Op0, AddOne(CI));
5840 case ICmpInst::ICMP_UGE:
5841 if (CI->isMinValue(false)) // A >=u MIN -> TRUE
5842 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5843 return new ICmpInst( ICmpInst::ICMP_UGT, Op0, SubOne(CI));
5844 case ICmpInst::ICMP_SGE:
5845 if (CI->isMinValue(true)) // A >=s MIN -> TRUE
5846 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5847 return new ICmpInst(ICmpInst::ICMP_SGT, Op0, SubOne(CI));
5850 // See if we can fold the comparison based on range information we can get
5851 // by checking whether bits are known to be zero or one in the input.
5852 uint32_t BitWidth = cast<IntegerType>(Ty)->getBitWidth();
5853 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
5855 // If this comparison is a normal comparison, it demands all
5856 // bits, if it is a sign bit comparison, it only demands the sign bit.
5858 bool isSignBit = isSignBitCheck(I.getPredicate(), CI, UnusedBit);
5860 if (SimplifyDemandedBits(Op0,
5861 isSignBit ? APInt::getSignBit(BitWidth)
5862 : APInt::getAllOnesValue(BitWidth),
5863 KnownZero, KnownOne, 0))
5866 // Given the known and unknown bits, compute a range that the LHS could be
5867 // in. Compute the Min, Max and RHS values based on the known bits. For the
5868 // EQ and NE we use unsigned values.
5869 APInt Min(BitWidth, 0), Max(BitWidth, 0);
5870 if (ICmpInst::isSignedPredicate(I.getPredicate()))
5871 ComputeSignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne, Min, Max);
5873 ComputeUnsignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne,Min,Max);
5875 // If Min and Max are known to be the same, then SimplifyDemandedBits
5876 // figured out that the LHS is a constant. Just constant fold this now so
5877 // that code below can assume that Min != Max.
5879 return ReplaceInstUsesWith(I, ConstantExpr::getICmp(I.getPredicate(),
5880 ConstantInt::get(Min),
5883 // Based on the range information we know about the LHS, see if we can
5884 // simplify this comparison. For example, (x&4) < 8 is always true.
5885 const APInt &RHSVal = CI->getValue();
5886 switch (I.getPredicate()) { // LE/GE have been folded already.
5887 default: assert(0 && "Unknown icmp opcode!");
5888 case ICmpInst::ICMP_EQ:
5889 if (Max.ult(RHSVal) || Min.ugt(RHSVal))
5890 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5892 case ICmpInst::ICMP_NE:
5893 if (Max.ult(RHSVal) || Min.ugt(RHSVal))
5894 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5896 case ICmpInst::ICMP_ULT:
5897 if (Max.ult(RHSVal)) // A <u C -> true iff max(A) < C
5898 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5899 if (Min.uge(RHSVal)) // A <u C -> false iff min(A) >= C
5900 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5901 if (RHSVal == Max) // A <u MAX -> A != MAX
5902 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5903 if (RHSVal == Min+1) // A <u MIN+1 -> A == MIN
5904 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, SubOne(CI));
5906 // (x <u 2147483648) -> (x >s -1) -> true if sign bit clear
5907 if (CI->isMinValue(true))
5908 return new ICmpInst(ICmpInst::ICMP_SGT, Op0,
5909 ConstantInt::getAllOnesValue(Op0->getType()));
5911 case ICmpInst::ICMP_UGT:
5912 if (Min.ugt(RHSVal)) // A >u C -> true iff min(A) > C
5913 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5914 if (Max.ule(RHSVal)) // A >u C -> false iff max(A) <= C
5915 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5917 if (RHSVal == Min) // A >u MIN -> A != MIN
5918 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5919 if (RHSVal == Max-1) // A >u MAX-1 -> A == MAX
5920 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, AddOne(CI));
5922 // (x >u 2147483647) -> (x <s 0) -> true if sign bit set
5923 if (CI->isMaxValue(true))
5924 return new ICmpInst(ICmpInst::ICMP_SLT, Op0,
5925 ConstantInt::getNullValue(Op0->getType()));
5927 case ICmpInst::ICMP_SLT:
5928 if (Max.slt(RHSVal)) // A <s C -> true iff max(A) < C
5929 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5930 if (Min.sge(RHSVal)) // A <s C -> false iff min(A) >= C
5931 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5932 if (RHSVal == Max) // A <s MAX -> A != MAX
5933 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5934 if (RHSVal == Min+1) // A <s MIN+1 -> A == MIN
5935 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, SubOne(CI));
5937 case ICmpInst::ICMP_SGT:
5938 if (Min.sgt(RHSVal)) // A >s C -> true iff min(A) > C
5939 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5940 if (Max.sle(RHSVal)) // A >s C -> false iff max(A) <= C
5941 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5943 if (RHSVal == Min) // A >s MIN -> A != MIN
5944 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5945 if (RHSVal == Max-1) // A >s MAX-1 -> A == MAX
5946 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, AddOne(CI));
5951 // Test if the ICmpInst instruction is used exclusively by a select as
5952 // part of a minimum or maximum operation. If so, refrain from doing
5953 // any other folding. This helps out other analyses which understand
5954 // non-obfuscated minimum and maximum idioms, such as ScalarEvolution
5955 // and CodeGen. And in this case, at least one of the comparison
5956 // operands has at least one user besides the compare (the select),
5957 // which would often largely negate the benefit of folding anyway.
5959 if (SelectInst *SI = dyn_cast<SelectInst>(*I.use_begin()))
5960 if ((SI->getOperand(1) == Op0 && SI->getOperand(2) == Op1) ||
5961 (SI->getOperand(2) == Op0 && SI->getOperand(1) == Op1))
5964 // See if we are doing a comparison between a constant and an instruction that
5965 // can be folded into the comparison.
5966 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
5967 // Since the RHS is a ConstantInt (CI), if the left hand side is an
5968 // instruction, see if that instruction also has constants so that the
5969 // instruction can be folded into the icmp
5970 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
5971 if (Instruction *Res = visitICmpInstWithInstAndIntCst(I, LHSI, CI))
5975 // Handle icmp with constant (but not simple integer constant) RHS
5976 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
5977 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
5978 switch (LHSI->getOpcode()) {
5979 case Instruction::GetElementPtr:
5980 if (RHSC->isNullValue()) {
5981 // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
5982 bool isAllZeros = true;
5983 for (unsigned i = 1, e = LHSI->getNumOperands(); i != e; ++i)
5984 if (!isa<Constant>(LHSI->getOperand(i)) ||
5985 !cast<Constant>(LHSI->getOperand(i))->isNullValue()) {
5990 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
5991 Constant::getNullValue(LHSI->getOperand(0)->getType()));
5995 case Instruction::PHI:
5996 // Only fold icmp into the PHI if the phi and fcmp are in the same
5997 // block. If in the same block, we're encouraging jump threading. If
5998 // not, we are just pessimizing the code by making an i1 phi.
5999 if (LHSI->getParent() == I.getParent())
6000 if (Instruction *NV = FoldOpIntoPhi(I))
6003 case Instruction::Select: {
6004 // If either operand of the select is a constant, we can fold the
6005 // comparison into the select arms, which will cause one to be
6006 // constant folded and the select turned into a bitwise or.
6007 Value *Op1 = 0, *Op2 = 0;
6008 if (LHSI->hasOneUse()) {
6009 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
6010 // Fold the known value into the constant operand.
6011 Op1 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
6012 // Insert a new ICmp of the other select operand.
6013 Op2 = InsertNewInstBefore(new ICmpInst(I.getPredicate(),
6014 LHSI->getOperand(2), RHSC,
6016 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
6017 // Fold the known value into the constant operand.
6018 Op2 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
6019 // Insert a new ICmp of the other select operand.
6020 Op1 = InsertNewInstBefore(new ICmpInst(I.getPredicate(),
6021 LHSI->getOperand(1), RHSC,
6027 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
6030 case Instruction::Malloc:
6031 // If we have (malloc != null), and if the malloc has a single use, we
6032 // can assume it is successful and remove the malloc.
6033 if (LHSI->hasOneUse() && isa<ConstantPointerNull>(RHSC)) {
6034 AddToWorkList(LHSI);
6035 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
6036 !I.isTrueWhenEqual()));
6042 // If we can optimize a 'icmp GEP, P' or 'icmp P, GEP', do so now.
6043 if (User *GEP = dyn_castGetElementPtr(Op0))
6044 if (Instruction *NI = FoldGEPICmp(GEP, Op1, I.getPredicate(), I))
6046 if (User *GEP = dyn_castGetElementPtr(Op1))
6047 if (Instruction *NI = FoldGEPICmp(GEP, Op0,
6048 ICmpInst::getSwappedPredicate(I.getPredicate()), I))
6051 // Test to see if the operands of the icmp are casted versions of other
6052 // values. If the ptr->ptr cast can be stripped off both arguments, we do so
6054 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op0)) {
6055 if (isa<PointerType>(Op0->getType()) &&
6056 (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
6057 // We keep moving the cast from the left operand over to the right
6058 // operand, where it can often be eliminated completely.
6059 Op0 = CI->getOperand(0);
6061 // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast
6062 // so eliminate it as well.
6063 if (BitCastInst *CI2 = dyn_cast<BitCastInst>(Op1))
6064 Op1 = CI2->getOperand(0);
6066 // If Op1 is a constant, we can fold the cast into the constant.
6067 if (Op0->getType() != Op1->getType()) {
6068 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
6069 Op1 = ConstantExpr::getBitCast(Op1C, Op0->getType());
6071 // Otherwise, cast the RHS right before the icmp
6072 Op1 = InsertBitCastBefore(Op1, Op0->getType(), I);
6075 return new ICmpInst(I.getPredicate(), Op0, Op1);
6079 if (isa<CastInst>(Op0)) {
6080 // Handle the special case of: icmp (cast bool to X), <cst>
6081 // This comes up when you have code like
6084 // For generality, we handle any zero-extension of any operand comparison
6085 // with a constant or another cast from the same type.
6086 if (isa<ConstantInt>(Op1) || isa<CastInst>(Op1))
6087 if (Instruction *R = visitICmpInstWithCastAndCast(I))
6091 // See if it's the same type of instruction on the left and right.
6092 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
6093 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
6094 if (Op0I->getOpcode() == Op1I->getOpcode() && Op0I->hasOneUse() &&
6095 Op1I->hasOneUse() && Op0I->getOperand(1) == Op1I->getOperand(1) &&
6097 switch (Op0I->getOpcode()) {
6099 case Instruction::Add:
6100 case Instruction::Sub:
6101 case Instruction::Xor:
6102 // a+x icmp eq/ne b+x --> a icmp b
6103 return new ICmpInst(I.getPredicate(), Op0I->getOperand(0),
6104 Op1I->getOperand(0));
6106 case Instruction::Mul:
6107 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) {
6108 // a * Cst icmp eq/ne b * Cst --> a & Mask icmp b & Mask
6109 // Mask = -1 >> count-trailing-zeros(Cst).
6110 if (!CI->isZero() && !CI->isOne()) {
6111 const APInt &AP = CI->getValue();
6112 ConstantInt *Mask = ConstantInt::get(
6113 APInt::getLowBitsSet(AP.getBitWidth(),
6115 AP.countTrailingZeros()));
6116 Instruction *And1 = BinaryOperator::CreateAnd(Op0I->getOperand(0),
6118 Instruction *And2 = BinaryOperator::CreateAnd(Op1I->getOperand(0),
6120 InsertNewInstBefore(And1, I);
6121 InsertNewInstBefore(And2, I);
6122 return new ICmpInst(I.getPredicate(), And1, And2);
6131 // ~x < ~y --> y < x
6133 if (match(Op0, m_Not(m_Value(A))) &&
6134 match(Op1, m_Not(m_Value(B))))
6135 return new ICmpInst(I.getPredicate(), B, A);
6138 if (I.isEquality()) {
6139 Value *A, *B, *C, *D;
6141 // -x == -y --> x == y
6142 if (match(Op0, m_Neg(m_Value(A))) &&
6143 match(Op1, m_Neg(m_Value(B))))
6144 return new ICmpInst(I.getPredicate(), A, B);
6146 if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
6147 if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0
6148 Value *OtherVal = A == Op1 ? B : A;
6149 return new ICmpInst(I.getPredicate(), OtherVal,
6150 Constant::getNullValue(A->getType()));
6153 if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) {
6154 // A^c1 == C^c2 --> A == C^(c1^c2)
6155 ConstantInt *C1, *C2;
6156 if (match(B, m_ConstantInt(C1)) &&
6157 match(D, m_ConstantInt(C2)) && Op1->hasOneUse()) {
6158 Constant *NC = ConstantInt::get(C1->getValue() ^ C2->getValue());
6159 Instruction *Xor = BinaryOperator::CreateXor(C, NC, "tmp");
6160 return new ICmpInst(I.getPredicate(), A,
6161 InsertNewInstBefore(Xor, I));
6164 // A^B == A^D -> B == D
6165 if (A == C) return new ICmpInst(I.getPredicate(), B, D);
6166 if (A == D) return new ICmpInst(I.getPredicate(), B, C);
6167 if (B == C) return new ICmpInst(I.getPredicate(), A, D);
6168 if (B == D) return new ICmpInst(I.getPredicate(), A, C);
6172 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
6173 (A == Op0 || B == Op0)) {
6174 // A == (A^B) -> B == 0
6175 Value *OtherVal = A == Op0 ? B : A;
6176 return new ICmpInst(I.getPredicate(), OtherVal,
6177 Constant::getNullValue(A->getType()));
6180 // (A-B) == A -> B == 0
6181 if (match(Op0, m_Sub(m_Specific(Op1), m_Value(B))))
6182 return new ICmpInst(I.getPredicate(), B,
6183 Constant::getNullValue(B->getType()));
6185 // A == (A-B) -> B == 0
6186 if (match(Op1, m_Sub(m_Specific(Op0), m_Value(B))))
6187 return new ICmpInst(I.getPredicate(), B,
6188 Constant::getNullValue(B->getType()));
6190 // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
6191 if (Op0->hasOneUse() && Op1->hasOneUse() &&
6192 match(Op0, m_And(m_Value(A), m_Value(B))) &&
6193 match(Op1, m_And(m_Value(C), m_Value(D)))) {
6194 Value *X = 0, *Y = 0, *Z = 0;
6197 X = B; Y = D; Z = A;
6198 } else if (A == D) {
6199 X = B; Y = C; Z = A;
6200 } else if (B == C) {
6201 X = A; Y = D; Z = B;
6202 } else if (B == D) {
6203 X = A; Y = C; Z = B;
6206 if (X) { // Build (X^Y) & Z
6207 Op1 = InsertNewInstBefore(BinaryOperator::CreateXor(X, Y, "tmp"), I);
6208 Op1 = InsertNewInstBefore(BinaryOperator::CreateAnd(Op1, Z, "tmp"), I);
6209 I.setOperand(0, Op1);
6210 I.setOperand(1, Constant::getNullValue(Op1->getType()));
6215 return Changed ? &I : 0;
6219 /// FoldICmpDivCst - Fold "icmp pred, ([su]div X, DivRHS), CmpRHS" where DivRHS
6220 /// and CmpRHS are both known to be integer constants.
6221 Instruction *InstCombiner::FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
6222 ConstantInt *DivRHS) {
6223 ConstantInt *CmpRHS = cast<ConstantInt>(ICI.getOperand(1));
6224 const APInt &CmpRHSV = CmpRHS->getValue();
6226 // FIXME: If the operand types don't match the type of the divide
6227 // then don't attempt this transform. The code below doesn't have the
6228 // logic to deal with a signed divide and an unsigned compare (and
6229 // vice versa). This is because (x /s C1) <s C2 produces different
6230 // results than (x /s C1) <u C2 or (x /u C1) <s C2 or even
6231 // (x /u C1) <u C2. Simply casting the operands and result won't
6232 // work. :( The if statement below tests that condition and bails
6234 bool DivIsSigned = DivI->getOpcode() == Instruction::SDiv;
6235 if (!ICI.isEquality() && DivIsSigned != ICI.isSignedPredicate())
6237 if (DivRHS->isZero())
6238 return 0; // The ProdOV computation fails on divide by zero.
6239 if (DivIsSigned && DivRHS->isAllOnesValue())
6240 return 0; // The overflow computation also screws up here
6241 if (DivRHS->isOne())
6242 return 0; // Not worth bothering, and eliminates some funny cases
6245 // Compute Prod = CI * DivRHS. We are essentially solving an equation
6246 // of form X/C1=C2. We solve for X by multiplying C1 (DivRHS) and
6247 // C2 (CI). By solving for X we can turn this into a range check
6248 // instead of computing a divide.
6249 ConstantInt *Prod = Multiply(CmpRHS, DivRHS);
6251 // Determine if the product overflows by seeing if the product is
6252 // not equal to the divide. Make sure we do the same kind of divide
6253 // as in the LHS instruction that we're folding.
6254 bool ProdOV = (DivIsSigned ? ConstantExpr::getSDiv(Prod, DivRHS) :
6255 ConstantExpr::getUDiv(Prod, DivRHS)) != CmpRHS;
6257 // Get the ICmp opcode
6258 ICmpInst::Predicate Pred = ICI.getPredicate();
6260 // Figure out the interval that is being checked. For example, a comparison
6261 // like "X /u 5 == 0" is really checking that X is in the interval [0, 5).
6262 // Compute this interval based on the constants involved and the signedness of
6263 // the compare/divide. This computes a half-open interval, keeping track of
6264 // whether either value in the interval overflows. After analysis each
6265 // overflow variable is set to 0 if it's corresponding bound variable is valid
6266 // -1 if overflowed off the bottom end, or +1 if overflowed off the top end.
6267 int LoOverflow = 0, HiOverflow = 0;
6268 ConstantInt *LoBound = 0, *HiBound = 0;
6270 if (!DivIsSigned) { // udiv
6271 // e.g. X/5 op 3 --> [15, 20)
6273 HiOverflow = LoOverflow = ProdOV;
6275 HiOverflow = AddWithOverflow(HiBound, LoBound, DivRHS, false);
6276 } else if (DivRHS->getValue().isStrictlyPositive()) { // Divisor is > 0.
6277 if (CmpRHSV == 0) { // (X / pos) op 0
6278 // Can't overflow. e.g. X/2 op 0 --> [-1, 2)
6279 LoBound = cast<ConstantInt>(ConstantExpr::getNeg(SubOne(DivRHS)));
6281 } else if (CmpRHSV.isStrictlyPositive()) { // (X / pos) op pos
6282 LoBound = Prod; // e.g. X/5 op 3 --> [15, 20)
6283 HiOverflow = LoOverflow = ProdOV;
6285 HiOverflow = AddWithOverflow(HiBound, Prod, DivRHS, true);
6286 } else { // (X / pos) op neg
6287 // e.g. X/5 op -3 --> [-15-4, -15+1) --> [-19, -14)
6288 HiBound = AddOne(Prod);
6289 LoOverflow = HiOverflow = ProdOV ? -1 : 0;
6291 ConstantInt* DivNeg = cast<ConstantInt>(ConstantExpr::getNeg(DivRHS));
6292 LoOverflow = AddWithOverflow(LoBound, HiBound, DivNeg,
6296 } else if (DivRHS->getValue().isNegative()) { // Divisor is < 0.
6297 if (CmpRHSV == 0) { // (X / neg) op 0
6298 // e.g. X/-5 op 0 --> [-4, 5)
6299 LoBound = AddOne(DivRHS);
6300 HiBound = cast<ConstantInt>(ConstantExpr::getNeg(DivRHS));
6301 if (HiBound == DivRHS) { // -INTMIN = INTMIN
6302 HiOverflow = 1; // [INTMIN+1, overflow)
6303 HiBound = 0; // e.g. X/INTMIN = 0 --> X > INTMIN
6305 } else if (CmpRHSV.isStrictlyPositive()) { // (X / neg) op pos
6306 // e.g. X/-5 op 3 --> [-19, -14)
6307 HiBound = AddOne(Prod);
6308 HiOverflow = LoOverflow = ProdOV ? -1 : 0;
6310 LoOverflow = AddWithOverflow(LoBound, HiBound, DivRHS, true) ? -1 : 0;
6311 } else { // (X / neg) op neg
6312 LoBound = Prod; // e.g. X/-5 op -3 --> [15, 20)
6313 LoOverflow = HiOverflow = ProdOV;
6315 HiOverflow = SubWithOverflow(HiBound, Prod, DivRHS, true);
6318 // Dividing by a negative swaps the condition. LT <-> GT
6319 Pred = ICmpInst::getSwappedPredicate(Pred);
6322 Value *X = DivI->getOperand(0);
6324 default: assert(0 && "Unhandled icmp opcode!");
6325 case ICmpInst::ICMP_EQ:
6326 if (LoOverflow && HiOverflow)
6327 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
6328 else if (HiOverflow)
6329 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
6330 ICmpInst::ICMP_UGE, X, LoBound);
6331 else if (LoOverflow)
6332 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
6333 ICmpInst::ICMP_ULT, X, HiBound);
6335 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned, true, ICI);
6336 case ICmpInst::ICMP_NE:
6337 if (LoOverflow && HiOverflow)
6338 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
6339 else if (HiOverflow)
6340 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
6341 ICmpInst::ICMP_ULT, X, LoBound);
6342 else if (LoOverflow)
6343 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
6344 ICmpInst::ICMP_UGE, X, HiBound);
6346 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned, false, ICI);
6347 case ICmpInst::ICMP_ULT:
6348 case ICmpInst::ICMP_SLT:
6349 if (LoOverflow == +1) // Low bound is greater than input range.
6350 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
6351 if (LoOverflow == -1) // Low bound is less than input range.
6352 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
6353 return new ICmpInst(Pred, X, LoBound);
6354 case ICmpInst::ICMP_UGT:
6355 case ICmpInst::ICMP_SGT:
6356 if (HiOverflow == +1) // High bound greater than input range.
6357 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
6358 else if (HiOverflow == -1) // High bound less than input range.
6359 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
6360 if (Pred == ICmpInst::ICMP_UGT)
6361 return new ICmpInst(ICmpInst::ICMP_UGE, X, HiBound);
6363 return new ICmpInst(ICmpInst::ICMP_SGE, X, HiBound);
6368 /// visitICmpInstWithInstAndIntCst - Handle "icmp (instr, intcst)".
6370 Instruction *InstCombiner::visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
6373 const APInt &RHSV = RHS->getValue();
6375 switch (LHSI->getOpcode()) {
6376 case Instruction::Xor: // (icmp pred (xor X, XorCST), CI)
6377 if (ConstantInt *XorCST = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
6378 // If this is a comparison that tests the signbit (X < 0) or (x > -1),
6380 if ((ICI.getPredicate() == ICmpInst::ICMP_SLT && RHSV == 0) ||
6381 (ICI.getPredicate() == ICmpInst::ICMP_SGT && RHSV.isAllOnesValue())) {
6382 Value *CompareVal = LHSI->getOperand(0);
6384 // If the sign bit of the XorCST is not set, there is no change to
6385 // the operation, just stop using the Xor.
6386 if (!XorCST->getValue().isNegative()) {
6387 ICI.setOperand(0, CompareVal);
6388 AddToWorkList(LHSI);
6392 // Was the old condition true if the operand is positive?
6393 bool isTrueIfPositive = ICI.getPredicate() == ICmpInst::ICMP_SGT;
6395 // If so, the new one isn't.
6396 isTrueIfPositive ^= true;
6398 if (isTrueIfPositive)
6399 return new ICmpInst(ICmpInst::ICMP_SGT, CompareVal, SubOne(RHS));
6401 return new ICmpInst(ICmpInst::ICMP_SLT, CompareVal, AddOne(RHS));
6405 case Instruction::And: // (icmp pred (and X, AndCST), RHS)
6406 if (LHSI->hasOneUse() && isa<ConstantInt>(LHSI->getOperand(1)) &&
6407 LHSI->getOperand(0)->hasOneUse()) {
6408 ConstantInt *AndCST = cast<ConstantInt>(LHSI->getOperand(1));
6410 // If the LHS is an AND of a truncating cast, we can widen the
6411 // and/compare to be the input width without changing the value
6412 // produced, eliminating a cast.
6413 if (TruncInst *Cast = dyn_cast<TruncInst>(LHSI->getOperand(0))) {
6414 // We can do this transformation if either the AND constant does not
6415 // have its sign bit set or if it is an equality comparison.
6416 // Extending a relational comparison when we're checking the sign
6417 // bit would not work.
6418 if (Cast->hasOneUse() &&
6419 (ICI.isEquality() ||
6420 (AndCST->getValue().isNonNegative() && RHSV.isNonNegative()))) {
6422 cast<IntegerType>(Cast->getOperand(0)->getType())->getBitWidth();
6423 APInt NewCST = AndCST->getValue();
6424 NewCST.zext(BitWidth);
6426 NewCI.zext(BitWidth);
6427 Instruction *NewAnd =
6428 BinaryOperator::CreateAnd(Cast->getOperand(0),
6429 ConstantInt::get(NewCST),LHSI->getName());
6430 InsertNewInstBefore(NewAnd, ICI);
6431 return new ICmpInst(ICI.getPredicate(), NewAnd,
6432 ConstantInt::get(NewCI));
6436 // If this is: (X >> C1) & C2 != C3 (where any shift and any compare
6437 // could exist), turn it into (X & (C2 << C1)) != (C3 << C1). This
6438 // happens a LOT in code produced by the C front-end, for bitfield
6440 BinaryOperator *Shift = dyn_cast<BinaryOperator>(LHSI->getOperand(0));
6441 if (Shift && !Shift->isShift())
6445 ShAmt = Shift ? dyn_cast<ConstantInt>(Shift->getOperand(1)) : 0;
6446 const Type *Ty = Shift ? Shift->getType() : 0; // Type of the shift.
6447 const Type *AndTy = AndCST->getType(); // Type of the and.
6449 // We can fold this as long as we can't shift unknown bits
6450 // into the mask. This can only happen with signed shift
6451 // rights, as they sign-extend.
6453 bool CanFold = Shift->isLogicalShift();
6455 // To test for the bad case of the signed shr, see if any
6456 // of the bits shifted in could be tested after the mask.
6457 uint32_t TyBits = Ty->getPrimitiveSizeInBits();
6458 int ShAmtVal = TyBits - ShAmt->getLimitedValue(TyBits);
6460 uint32_t BitWidth = AndTy->getPrimitiveSizeInBits();
6461 if ((APInt::getHighBitsSet(BitWidth, BitWidth-ShAmtVal) &
6462 AndCST->getValue()) == 0)
6468 if (Shift->getOpcode() == Instruction::Shl)
6469 NewCst = ConstantExpr::getLShr(RHS, ShAmt);
6471 NewCst = ConstantExpr::getShl(RHS, ShAmt);
6473 // Check to see if we are shifting out any of the bits being
6475 if (ConstantExpr::get(Shift->getOpcode(), NewCst, ShAmt) != RHS) {
6476 // If we shifted bits out, the fold is not going to work out.
6477 // As a special case, check to see if this means that the
6478 // result is always true or false now.
6479 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
6480 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
6481 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
6482 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
6484 ICI.setOperand(1, NewCst);
6485 Constant *NewAndCST;
6486 if (Shift->getOpcode() == Instruction::Shl)
6487 NewAndCST = ConstantExpr::getLShr(AndCST, ShAmt);
6489 NewAndCST = ConstantExpr::getShl(AndCST, ShAmt);
6490 LHSI->setOperand(1, NewAndCST);
6491 LHSI->setOperand(0, Shift->getOperand(0));
6492 AddToWorkList(Shift); // Shift is dead.
6493 AddUsesToWorkList(ICI);
6499 // Turn ((X >> Y) & C) == 0 into (X & (C << Y)) == 0. The later is
6500 // preferable because it allows the C<<Y expression to be hoisted out
6501 // of a loop if Y is invariant and X is not.
6502 if (Shift && Shift->hasOneUse() && RHSV == 0 &&
6503 ICI.isEquality() && !Shift->isArithmeticShift() &&
6504 isa<Instruction>(Shift->getOperand(0))) {
6507 if (Shift->getOpcode() == Instruction::LShr) {
6508 NS = BinaryOperator::CreateShl(AndCST,
6509 Shift->getOperand(1), "tmp");
6511 // Insert a logical shift.
6512 NS = BinaryOperator::CreateLShr(AndCST,
6513 Shift->getOperand(1), "tmp");
6515 InsertNewInstBefore(cast<Instruction>(NS), ICI);
6517 // Compute X & (C << Y).
6518 Instruction *NewAnd =
6519 BinaryOperator::CreateAnd(Shift->getOperand(0), NS, LHSI->getName());
6520 InsertNewInstBefore(NewAnd, ICI);
6522 ICI.setOperand(0, NewAnd);
6528 case Instruction::Shl: { // (icmp pred (shl X, ShAmt), CI)
6529 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
6532 uint32_t TypeBits = RHSV.getBitWidth();
6534 // Check that the shift amount is in range. If not, don't perform
6535 // undefined shifts. When the shift is visited it will be
6537 if (ShAmt->uge(TypeBits))
6540 if (ICI.isEquality()) {
6541 // If we are comparing against bits always shifted out, the
6542 // comparison cannot succeed.
6544 ConstantExpr::getShl(ConstantExpr::getLShr(RHS, ShAmt), ShAmt);
6545 if (Comp != RHS) {// Comparing against a bit that we know is zero.
6546 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
6547 Constant *Cst = ConstantInt::get(Type::Int1Ty, IsICMP_NE);
6548 return ReplaceInstUsesWith(ICI, Cst);
6551 if (LHSI->hasOneUse()) {
6552 // Otherwise strength reduce the shift into an and.
6553 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
6555 ConstantInt::get(APInt::getLowBitsSet(TypeBits, TypeBits-ShAmtVal));
6558 BinaryOperator::CreateAnd(LHSI->getOperand(0),
6559 Mask, LHSI->getName()+".mask");
6560 Value *And = InsertNewInstBefore(AndI, ICI);
6561 return new ICmpInst(ICI.getPredicate(), And,
6562 ConstantInt::get(RHSV.lshr(ShAmtVal)));
6566 // Otherwise, if this is a comparison of the sign bit, simplify to and/test.
6567 bool TrueIfSigned = false;
6568 if (LHSI->hasOneUse() &&
6569 isSignBitCheck(ICI.getPredicate(), RHS, TrueIfSigned)) {
6570 // (X << 31) <s 0 --> (X&1) != 0
6571 Constant *Mask = ConstantInt::get(APInt(TypeBits, 1) <<
6572 (TypeBits-ShAmt->getZExtValue()-1));
6574 BinaryOperator::CreateAnd(LHSI->getOperand(0),
6575 Mask, LHSI->getName()+".mask");
6576 Value *And = InsertNewInstBefore(AndI, ICI);
6578 return new ICmpInst(TrueIfSigned ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ,
6579 And, Constant::getNullValue(And->getType()));
6584 case Instruction::LShr: // (icmp pred (shr X, ShAmt), CI)
6585 case Instruction::AShr: {
6586 // Only handle equality comparisons of shift-by-constant.
6587 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
6588 if (!ShAmt || !ICI.isEquality()) break;
6590 // Check that the shift amount is in range. If not, don't perform
6591 // undefined shifts. When the shift is visited it will be
6593 uint32_t TypeBits = RHSV.getBitWidth();
6594 if (ShAmt->uge(TypeBits))
6597 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
6599 // If we are comparing against bits always shifted out, the
6600 // comparison cannot succeed.
6601 APInt Comp = RHSV << ShAmtVal;
6602 if (LHSI->getOpcode() == Instruction::LShr)
6603 Comp = Comp.lshr(ShAmtVal);
6605 Comp = Comp.ashr(ShAmtVal);
6607 if (Comp != RHSV) { // Comparing against a bit that we know is zero.
6608 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
6609 Constant *Cst = ConstantInt::get(Type::Int1Ty, IsICMP_NE);
6610 return ReplaceInstUsesWith(ICI, Cst);
6613 // Otherwise, check to see if the bits shifted out are known to be zero.
6614 // If so, we can compare against the unshifted value:
6615 // (X & 4) >> 1 == 2 --> (X & 4) == 4.
6616 if (LHSI->hasOneUse() &&
6617 MaskedValueIsZero(LHSI->getOperand(0),
6618 APInt::getLowBitsSet(Comp.getBitWidth(), ShAmtVal))) {
6619 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
6620 ConstantExpr::getShl(RHS, ShAmt));
6623 if (LHSI->hasOneUse()) {
6624 // Otherwise strength reduce the shift into an and.
6625 APInt Val(APInt::getHighBitsSet(TypeBits, TypeBits - ShAmtVal));
6626 Constant *Mask = ConstantInt::get(Val);
6629 BinaryOperator::CreateAnd(LHSI->getOperand(0),
6630 Mask, LHSI->getName()+".mask");
6631 Value *And = InsertNewInstBefore(AndI, ICI);
6632 return new ICmpInst(ICI.getPredicate(), And,
6633 ConstantExpr::getShl(RHS, ShAmt));
6638 case Instruction::SDiv:
6639 case Instruction::UDiv:
6640 // Fold: icmp pred ([us]div X, C1), C2 -> range test
6641 // Fold this div into the comparison, producing a range check.
6642 // Determine, based on the divide type, what the range is being
6643 // checked. If there is an overflow on the low or high side, remember
6644 // it, otherwise compute the range [low, hi) bounding the new value.
6645 // See: InsertRangeTest above for the kinds of replacements possible.
6646 if (ConstantInt *DivRHS = dyn_cast<ConstantInt>(LHSI->getOperand(1)))
6647 if (Instruction *R = FoldICmpDivCst(ICI, cast<BinaryOperator>(LHSI),
6652 case Instruction::Add:
6653 // Fold: icmp pred (add, X, C1), C2
6655 if (!ICI.isEquality()) {
6656 ConstantInt *LHSC = dyn_cast<ConstantInt>(LHSI->getOperand(1));
6658 const APInt &LHSV = LHSC->getValue();
6660 ConstantRange CR = ICI.makeConstantRange(ICI.getPredicate(), RHSV)
6663 if (ICI.isSignedPredicate()) {
6664 if (CR.getLower().isSignBit()) {
6665 return new ICmpInst(ICmpInst::ICMP_SLT, LHSI->getOperand(0),
6666 ConstantInt::get(CR.getUpper()));
6667 } else if (CR.getUpper().isSignBit()) {
6668 return new ICmpInst(ICmpInst::ICMP_SGE, LHSI->getOperand(0),
6669 ConstantInt::get(CR.getLower()));
6672 if (CR.getLower().isMinValue()) {
6673 return new ICmpInst(ICmpInst::ICMP_ULT, LHSI->getOperand(0),
6674 ConstantInt::get(CR.getUpper()));
6675 } else if (CR.getUpper().isMinValue()) {
6676 return new ICmpInst(ICmpInst::ICMP_UGE, LHSI->getOperand(0),
6677 ConstantInt::get(CR.getLower()));
6684 // Simplify icmp_eq and icmp_ne instructions with integer constant RHS.
6685 if (ICI.isEquality()) {
6686 bool isICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
6688 // If the first operand is (add|sub|and|or|xor|rem) with a constant, and
6689 // the second operand is a constant, simplify a bit.
6690 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(LHSI)) {
6691 switch (BO->getOpcode()) {
6692 case Instruction::SRem:
6693 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
6694 if (RHSV == 0 && isa<ConstantInt>(BO->getOperand(1)) &&BO->hasOneUse()){
6695 const APInt &V = cast<ConstantInt>(BO->getOperand(1))->getValue();
6696 if (V.sgt(APInt(V.getBitWidth(), 1)) && V.isPowerOf2()) {
6697 Instruction *NewRem =
6698 BinaryOperator::CreateURem(BO->getOperand(0), BO->getOperand(1),
6700 InsertNewInstBefore(NewRem, ICI);
6701 return new ICmpInst(ICI.getPredicate(), NewRem,
6702 Constant::getNullValue(BO->getType()));
6706 case Instruction::Add:
6707 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
6708 if (ConstantInt *BOp1C = dyn_cast<ConstantInt>(BO->getOperand(1))) {
6709 if (BO->hasOneUse())
6710 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
6711 Subtract(RHS, BOp1C));
6712 } else if (RHSV == 0) {
6713 // Replace ((add A, B) != 0) with (A != -B) if A or B is
6714 // efficiently invertible, or if the add has just this one use.
6715 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
6717 if (Value *NegVal = dyn_castNegVal(BOp1))
6718 return new ICmpInst(ICI.getPredicate(), BOp0, NegVal);
6719 else if (Value *NegVal = dyn_castNegVal(BOp0))
6720 return new ICmpInst(ICI.getPredicate(), NegVal, BOp1);
6721 else if (BO->hasOneUse()) {
6722 Instruction *Neg = BinaryOperator::CreateNeg(BOp1);
6723 InsertNewInstBefore(Neg, ICI);
6725 return new ICmpInst(ICI.getPredicate(), BOp0, Neg);
6729 case Instruction::Xor:
6730 // For the xor case, we can xor two constants together, eliminating
6731 // the explicit xor.
6732 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1)))
6733 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
6734 ConstantExpr::getXor(RHS, BOC));
6737 case Instruction::Sub:
6738 // Replace (([sub|xor] A, B) != 0) with (A != B)
6740 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
6744 case Instruction::Or:
6745 // If bits are being or'd in that are not present in the constant we
6746 // are comparing against, then the comparison could never succeed!
6747 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1))) {
6748 Constant *NotCI = ConstantExpr::getNot(RHS);
6749 if (!ConstantExpr::getAnd(BOC, NotCI)->isNullValue())
6750 return ReplaceInstUsesWith(ICI, ConstantInt::get(Type::Int1Ty,
6755 case Instruction::And:
6756 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
6757 // If bits are being compared against that are and'd out, then the
6758 // comparison can never succeed!
6759 if ((RHSV & ~BOC->getValue()) != 0)
6760 return ReplaceInstUsesWith(ICI, ConstantInt::get(Type::Int1Ty,
6763 // If we have ((X & C) == C), turn it into ((X & C) != 0).
6764 if (RHS == BOC && RHSV.isPowerOf2())
6765 return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ :
6766 ICmpInst::ICMP_NE, LHSI,
6767 Constant::getNullValue(RHS->getType()));
6769 // Replace (and X, (1 << size(X)-1) != 0) with x s< 0
6770 if (BOC->getValue().isSignBit()) {
6771 Value *X = BO->getOperand(0);
6772 Constant *Zero = Constant::getNullValue(X->getType());
6773 ICmpInst::Predicate pred = isICMP_NE ?
6774 ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
6775 return new ICmpInst(pred, X, Zero);
6778 // ((X & ~7) == 0) --> X < 8
6779 if (RHSV == 0 && isHighOnes(BOC)) {
6780 Value *X = BO->getOperand(0);
6781 Constant *NegX = ConstantExpr::getNeg(BOC);
6782 ICmpInst::Predicate pred = isICMP_NE ?
6783 ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
6784 return new ICmpInst(pred, X, NegX);
6789 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(LHSI)) {
6790 // Handle icmp {eq|ne} <intrinsic>, intcst.
6791 if (II->getIntrinsicID() == Intrinsic::bswap) {
6793 ICI.setOperand(0, II->getOperand(1));
6794 ICI.setOperand(1, ConstantInt::get(RHSV.byteSwap()));
6798 } else { // Not a ICMP_EQ/ICMP_NE
6799 // If the LHS is a cast from an integral value of the same size,
6800 // then since we know the RHS is a constant, try to simlify.
6801 if (CastInst *Cast = dyn_cast<CastInst>(LHSI)) {
6802 Value *CastOp = Cast->getOperand(0);
6803 const Type *SrcTy = CastOp->getType();
6804 uint32_t SrcTySize = SrcTy->getPrimitiveSizeInBits();
6805 if (SrcTy->isInteger() &&
6806 SrcTySize == Cast->getType()->getPrimitiveSizeInBits()) {
6807 // If this is an unsigned comparison, try to make the comparison use
6808 // smaller constant values.
6809 if (ICI.getPredicate() == ICmpInst::ICMP_ULT && RHSV.isSignBit()) {
6810 // X u< 128 => X s> -1
6811 return new ICmpInst(ICmpInst::ICMP_SGT, CastOp,
6812 ConstantInt::get(APInt::getAllOnesValue(SrcTySize)));
6813 } else if (ICI.getPredicate() == ICmpInst::ICMP_UGT &&
6814 RHSV == APInt::getSignedMaxValue(SrcTySize)) {
6815 // X u> 127 => X s< 0
6816 return new ICmpInst(ICmpInst::ICMP_SLT, CastOp,
6817 Constant::getNullValue(SrcTy));
6825 /// visitICmpInstWithCastAndCast - Handle icmp (cast x to y), (cast/cst).
6826 /// We only handle extending casts so far.
6828 Instruction *InstCombiner::visitICmpInstWithCastAndCast(ICmpInst &ICI) {
6829 const CastInst *LHSCI = cast<CastInst>(ICI.getOperand(0));
6830 Value *LHSCIOp = LHSCI->getOperand(0);
6831 const Type *SrcTy = LHSCIOp->getType();
6832 const Type *DestTy = LHSCI->getType();
6835 // Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the
6836 // integer type is the same size as the pointer type.
6837 if (LHSCI->getOpcode() == Instruction::PtrToInt &&
6838 getTargetData().getPointerSizeInBits() ==
6839 cast<IntegerType>(DestTy)->getBitWidth()) {
6841 if (Constant *RHSC = dyn_cast<Constant>(ICI.getOperand(1))) {
6842 RHSOp = ConstantExpr::getIntToPtr(RHSC, SrcTy);
6843 } else if (PtrToIntInst *RHSC = dyn_cast<PtrToIntInst>(ICI.getOperand(1))) {
6844 RHSOp = RHSC->getOperand(0);
6845 // If the pointer types don't match, insert a bitcast.
6846 if (LHSCIOp->getType() != RHSOp->getType())
6847 RHSOp = InsertBitCastBefore(RHSOp, LHSCIOp->getType(), ICI);
6851 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSOp);
6854 // The code below only handles extension cast instructions, so far.
6856 if (LHSCI->getOpcode() != Instruction::ZExt &&
6857 LHSCI->getOpcode() != Instruction::SExt)
6860 bool isSignedExt = LHSCI->getOpcode() == Instruction::SExt;
6861 bool isSignedCmp = ICI.isSignedPredicate();
6863 if (CastInst *CI = dyn_cast<CastInst>(ICI.getOperand(1))) {
6864 // Not an extension from the same type?
6865 RHSCIOp = CI->getOperand(0);
6866 if (RHSCIOp->getType() != LHSCIOp->getType())
6869 // If the signedness of the two casts doesn't agree (i.e. one is a sext
6870 // and the other is a zext), then we can't handle this.
6871 if (CI->getOpcode() != LHSCI->getOpcode())
6874 // Deal with equality cases early.
6875 if (ICI.isEquality())
6876 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
6878 // A signed comparison of sign extended values simplifies into a
6879 // signed comparison.
6880 if (isSignedCmp && isSignedExt)
6881 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
6883 // The other three cases all fold into an unsigned comparison.
6884 return new ICmpInst(ICI.getUnsignedPredicate(), LHSCIOp, RHSCIOp);
6887 // If we aren't dealing with a constant on the RHS, exit early
6888 ConstantInt *CI = dyn_cast<ConstantInt>(ICI.getOperand(1));
6892 // Compute the constant that would happen if we truncated to SrcTy then
6893 // reextended to DestTy.
6894 Constant *Res1 = ConstantExpr::getTrunc(CI, SrcTy);
6895 Constant *Res2 = ConstantExpr::getCast(LHSCI->getOpcode(), Res1, DestTy);
6897 // If the re-extended constant didn't change...
6899 // Make sure that sign of the Cmp and the sign of the Cast are the same.
6900 // For example, we might have:
6901 // %A = sext short %X to uint
6902 // %B = icmp ugt uint %A, 1330
6903 // It is incorrect to transform this into
6904 // %B = icmp ugt short %X, 1330
6905 // because %A may have negative value.
6907 // However, we allow this when the compare is EQ/NE, because they are
6909 if (isSignedExt == isSignedCmp || ICI.isEquality())
6910 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
6914 // The re-extended constant changed so the constant cannot be represented
6915 // in the shorter type. Consequently, we cannot emit a simple comparison.
6917 // First, handle some easy cases. We know the result cannot be equal at this
6918 // point so handle the ICI.isEquality() cases
6919 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
6920 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
6921 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
6922 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
6924 // Evaluate the comparison for LT (we invert for GT below). LE and GE cases
6925 // should have been folded away previously and not enter in here.
6928 // We're performing a signed comparison.
6929 if (cast<ConstantInt>(CI)->getValue().isNegative())
6930 Result = ConstantInt::getFalse(); // X < (small) --> false
6932 Result = ConstantInt::getTrue(); // X < (large) --> true
6934 // We're performing an unsigned comparison.
6936 // We're performing an unsigned comp with a sign extended value.
6937 // This is true if the input is >= 0. [aka >s -1]
6938 Constant *NegOne = ConstantInt::getAllOnesValue(SrcTy);
6939 Result = InsertNewInstBefore(new ICmpInst(ICmpInst::ICMP_SGT, LHSCIOp,
6940 NegOne, ICI.getName()), ICI);
6942 // Unsigned extend & unsigned compare -> always true.
6943 Result = ConstantInt::getTrue();
6947 // Finally, return the value computed.
6948 if (ICI.getPredicate() == ICmpInst::ICMP_ULT ||
6949 ICI.getPredicate() == ICmpInst::ICMP_SLT)
6950 return ReplaceInstUsesWith(ICI, Result);
6952 assert((ICI.getPredicate()==ICmpInst::ICMP_UGT ||
6953 ICI.getPredicate()==ICmpInst::ICMP_SGT) &&
6954 "ICmp should be folded!");
6955 if (Constant *CI = dyn_cast<Constant>(Result))
6956 return ReplaceInstUsesWith(ICI, ConstantExpr::getNot(CI));
6957 return BinaryOperator::CreateNot(Result);
6960 Instruction *InstCombiner::visitShl(BinaryOperator &I) {
6961 return commonShiftTransforms(I);
6964 Instruction *InstCombiner::visitLShr(BinaryOperator &I) {
6965 return commonShiftTransforms(I);
6968 Instruction *InstCombiner::visitAShr(BinaryOperator &I) {
6969 if (Instruction *R = commonShiftTransforms(I))
6972 Value *Op0 = I.getOperand(0);
6974 // ashr int -1, X = -1 (for any arithmetic shift rights of ~0)
6975 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
6976 if (CSI->isAllOnesValue())
6977 return ReplaceInstUsesWith(I, CSI);
6979 // See if we can turn a signed shr into an unsigned shr.
6980 if (!isa<VectorType>(I.getType()) &&
6981 MaskedValueIsZero(Op0,
6982 APInt::getSignBit(I.getType()->getPrimitiveSizeInBits())))
6983 return BinaryOperator::CreateLShr(Op0, I.getOperand(1));
6988 Instruction *InstCombiner::commonShiftTransforms(BinaryOperator &I) {
6989 assert(I.getOperand(1)->getType() == I.getOperand(0)->getType());
6990 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
6992 // shl X, 0 == X and shr X, 0 == X
6993 // shl 0, X == 0 and shr 0, X == 0
6994 if (Op1 == Constant::getNullValue(Op1->getType()) ||
6995 Op0 == Constant::getNullValue(Op0->getType()))
6996 return ReplaceInstUsesWith(I, Op0);
6998 if (isa<UndefValue>(Op0)) {
6999 if (I.getOpcode() == Instruction::AShr) // undef >>s X -> undef
7000 return ReplaceInstUsesWith(I, Op0);
7001 else // undef << X -> 0, undef >>u X -> 0
7002 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
7004 if (isa<UndefValue>(Op1)) {
7005 if (I.getOpcode() == Instruction::AShr) // X >>s undef -> X
7006 return ReplaceInstUsesWith(I, Op0);
7007 else // X << undef, X >>u undef -> 0
7008 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
7011 // Try to fold constant and into select arguments.
7012 if (isa<Constant>(Op0))
7013 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
7014 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
7017 if (ConstantInt *CUI = dyn_cast<ConstantInt>(Op1))
7018 if (Instruction *Res = FoldShiftByConstant(Op0, CUI, I))
7023 Instruction *InstCombiner::FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
7024 BinaryOperator &I) {
7025 bool isLeftShift = I.getOpcode() == Instruction::Shl;
7027 // See if we can simplify any instructions used by the instruction whose sole
7028 // purpose is to compute bits we don't care about.
7029 uint32_t TypeBits = Op0->getType()->getPrimitiveSizeInBits();
7030 APInt KnownZero(TypeBits, 0), KnownOne(TypeBits, 0);
7031 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(TypeBits),
7032 KnownZero, KnownOne))
7035 // shl uint X, 32 = 0 and shr ubyte Y, 9 = 0, ... just don't eliminate shr
7036 // of a signed value.
7038 if (Op1->uge(TypeBits)) {
7039 if (I.getOpcode() != Instruction::AShr)
7040 return ReplaceInstUsesWith(I, Constant::getNullValue(Op0->getType()));
7042 I.setOperand(1, ConstantInt::get(I.getType(), TypeBits-1));
7047 // ((X*C1) << C2) == (X * (C1 << C2))
7048 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0))
7049 if (BO->getOpcode() == Instruction::Mul && isLeftShift)
7050 if (Constant *BOOp = dyn_cast<Constant>(BO->getOperand(1)))
7051 return BinaryOperator::CreateMul(BO->getOperand(0),
7052 ConstantExpr::getShl(BOOp, Op1));
7054 // Try to fold constant and into select arguments.
7055 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
7056 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
7058 if (isa<PHINode>(Op0))
7059 if (Instruction *NV = FoldOpIntoPhi(I))
7062 // Fold shift2(trunc(shift1(x,c1)), c2) -> trunc(shift2(shift1(x,c1),c2))
7063 if (TruncInst *TI = dyn_cast<TruncInst>(Op0)) {
7064 Instruction *TrOp = dyn_cast<Instruction>(TI->getOperand(0));
7065 // If 'shift2' is an ashr, we would have to get the sign bit into a funny
7066 // place. Don't try to do this transformation in this case. Also, we
7067 // require that the input operand is a shift-by-constant so that we have
7068 // confidence that the shifts will get folded together. We could do this
7069 // xform in more cases, but it is unlikely to be profitable.
7070 if (TrOp && I.isLogicalShift() && TrOp->isShift() &&
7071 isa<ConstantInt>(TrOp->getOperand(1))) {
7072 // Okay, we'll do this xform. Make the shift of shift.
7073 Constant *ShAmt = ConstantExpr::getZExt(Op1, TrOp->getType());
7074 Instruction *NSh = BinaryOperator::Create(I.getOpcode(), TrOp, ShAmt,
7076 InsertNewInstBefore(NSh, I); // (shift2 (shift1 & 0x00FF), c2)
7078 // For logical shifts, the truncation has the effect of making the high
7079 // part of the register be zeros. Emulate this by inserting an AND to
7080 // clear the top bits as needed. This 'and' will usually be zapped by
7081 // other xforms later if dead.
7082 unsigned SrcSize = TrOp->getType()->getPrimitiveSizeInBits();
7083 unsigned DstSize = TI->getType()->getPrimitiveSizeInBits();
7084 APInt MaskV(APInt::getLowBitsSet(SrcSize, DstSize));
7086 // The mask we constructed says what the trunc would do if occurring
7087 // between the shifts. We want to know the effect *after* the second
7088 // shift. We know that it is a logical shift by a constant, so adjust the
7089 // mask as appropriate.
7090 if (I.getOpcode() == Instruction::Shl)
7091 MaskV <<= Op1->getZExtValue();
7093 assert(I.getOpcode() == Instruction::LShr && "Unknown logical shift");
7094 MaskV = MaskV.lshr(Op1->getZExtValue());
7097 Instruction *And = BinaryOperator::CreateAnd(NSh, ConstantInt::get(MaskV),
7099 InsertNewInstBefore(And, I); // shift1 & 0x00FF
7101 // Return the value truncated to the interesting size.
7102 return new TruncInst(And, I.getType());
7106 if (Op0->hasOneUse()) {
7107 if (BinaryOperator *Op0BO = dyn_cast<BinaryOperator>(Op0)) {
7108 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
7111 switch (Op0BO->getOpcode()) {
7113 case Instruction::Add:
7114 case Instruction::And:
7115 case Instruction::Or:
7116 case Instruction::Xor: {
7117 // These operators commute.
7118 // Turn (Y + (X >> C)) << C -> (X + (Y << C)) & (~0 << C)
7119 if (isLeftShift && Op0BO->getOperand(1)->hasOneUse() &&
7120 match(Op0BO->getOperand(1), m_Shr(m_Value(V1), m_Specific(Op1)))){
7121 Instruction *YS = BinaryOperator::CreateShl(
7122 Op0BO->getOperand(0), Op1,
7124 InsertNewInstBefore(YS, I); // (Y << C)
7126 BinaryOperator::Create(Op0BO->getOpcode(), YS, V1,
7127 Op0BO->getOperand(1)->getName());
7128 InsertNewInstBefore(X, I); // (X + (Y << C))
7129 uint32_t Op1Val = Op1->getLimitedValue(TypeBits);
7130 return BinaryOperator::CreateAnd(X, ConstantInt::get(
7131 APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
7134 // Turn (Y + ((X >> C) & CC)) << C -> ((X & (CC << C)) + (Y << C))
7135 Value *Op0BOOp1 = Op0BO->getOperand(1);
7136 if (isLeftShift && Op0BOOp1->hasOneUse() &&
7138 m_And(m_Shr(m_Value(V1), m_Specific(Op1)),
7139 m_ConstantInt(CC))) &&
7140 cast<BinaryOperator>(Op0BOOp1)->getOperand(0)->hasOneUse()) {
7141 Instruction *YS = BinaryOperator::CreateShl(
7142 Op0BO->getOperand(0), Op1,
7144 InsertNewInstBefore(YS, I); // (Y << C)
7146 BinaryOperator::CreateAnd(V1, ConstantExpr::getShl(CC, Op1),
7147 V1->getName()+".mask");
7148 InsertNewInstBefore(XM, I); // X & (CC << C)
7150 return BinaryOperator::Create(Op0BO->getOpcode(), YS, XM);
7155 case Instruction::Sub: {
7156 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
7157 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
7158 match(Op0BO->getOperand(0), m_Shr(m_Value(V1), m_Specific(Op1)))){
7159 Instruction *YS = BinaryOperator::CreateShl(
7160 Op0BO->getOperand(1), Op1,
7162 InsertNewInstBefore(YS, I); // (Y << C)
7164 BinaryOperator::Create(Op0BO->getOpcode(), V1, YS,
7165 Op0BO->getOperand(0)->getName());
7166 InsertNewInstBefore(X, I); // (X + (Y << C))
7167 uint32_t Op1Val = Op1->getLimitedValue(TypeBits);
7168 return BinaryOperator::CreateAnd(X, ConstantInt::get(
7169 APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
7172 // Turn (((X >> C)&CC) + Y) << C -> (X + (Y << C)) & (CC << C)
7173 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
7174 match(Op0BO->getOperand(0),
7175 m_And(m_Shr(m_Value(V1), m_Value(V2)),
7176 m_ConstantInt(CC))) && V2 == Op1 &&
7177 cast<BinaryOperator>(Op0BO->getOperand(0))
7178 ->getOperand(0)->hasOneUse()) {
7179 Instruction *YS = BinaryOperator::CreateShl(
7180 Op0BO->getOperand(1), Op1,
7182 InsertNewInstBefore(YS, I); // (Y << C)
7184 BinaryOperator::CreateAnd(V1, ConstantExpr::getShl(CC, Op1),
7185 V1->getName()+".mask");
7186 InsertNewInstBefore(XM, I); // X & (CC << C)
7188 return BinaryOperator::Create(Op0BO->getOpcode(), XM, YS);
7196 // If the operand is an bitwise operator with a constant RHS, and the
7197 // shift is the only use, we can pull it out of the shift.
7198 if (ConstantInt *Op0C = dyn_cast<ConstantInt>(Op0BO->getOperand(1))) {
7199 bool isValid = true; // Valid only for And, Or, Xor
7200 bool highBitSet = false; // Transform if high bit of constant set?
7202 switch (Op0BO->getOpcode()) {
7203 default: isValid = false; break; // Do not perform transform!
7204 case Instruction::Add:
7205 isValid = isLeftShift;
7207 case Instruction::Or:
7208 case Instruction::Xor:
7211 case Instruction::And:
7216 // If this is a signed shift right, and the high bit is modified
7217 // by the logical operation, do not perform the transformation.
7218 // The highBitSet boolean indicates the value of the high bit of
7219 // the constant which would cause it to be modified for this
7222 if (isValid && I.getOpcode() == Instruction::AShr)
7223 isValid = Op0C->getValue()[TypeBits-1] == highBitSet;
7226 Constant *NewRHS = ConstantExpr::get(I.getOpcode(), Op0C, Op1);
7228 Instruction *NewShift =
7229 BinaryOperator::Create(I.getOpcode(), Op0BO->getOperand(0), Op1);
7230 InsertNewInstBefore(NewShift, I);
7231 NewShift->takeName(Op0BO);
7233 return BinaryOperator::Create(Op0BO->getOpcode(), NewShift,
7240 // Find out if this is a shift of a shift by a constant.
7241 BinaryOperator *ShiftOp = dyn_cast<BinaryOperator>(Op0);
7242 if (ShiftOp && !ShiftOp->isShift())
7245 if (ShiftOp && isa<ConstantInt>(ShiftOp->getOperand(1))) {
7246 ConstantInt *ShiftAmt1C = cast<ConstantInt>(ShiftOp->getOperand(1));
7247 uint32_t ShiftAmt1 = ShiftAmt1C->getLimitedValue(TypeBits);
7248 uint32_t ShiftAmt2 = Op1->getLimitedValue(TypeBits);
7249 assert(ShiftAmt2 != 0 && "Should have been simplified earlier");
7250 if (ShiftAmt1 == 0) return 0; // Will be simplified in the future.
7251 Value *X = ShiftOp->getOperand(0);
7253 uint32_t AmtSum = ShiftAmt1+ShiftAmt2; // Fold into one big shift.
7254 if (AmtSum > TypeBits)
7257 const IntegerType *Ty = cast<IntegerType>(I.getType());
7259 // Check for (X << c1) << c2 and (X >> c1) >> c2
7260 if (I.getOpcode() == ShiftOp->getOpcode()) {
7261 return BinaryOperator::Create(I.getOpcode(), X,
7262 ConstantInt::get(Ty, AmtSum));
7263 } else if (ShiftOp->getOpcode() == Instruction::LShr &&
7264 I.getOpcode() == Instruction::AShr) {
7265 // ((X >>u C1) >>s C2) -> (X >>u (C1+C2)) since C1 != 0.
7266 return BinaryOperator::CreateLShr(X, ConstantInt::get(Ty, AmtSum));
7267 } else if (ShiftOp->getOpcode() == Instruction::AShr &&
7268 I.getOpcode() == Instruction::LShr) {
7269 // ((X >>s C1) >>u C2) -> ((X >>s (C1+C2)) & mask) since C1 != 0.
7270 Instruction *Shift =
7271 BinaryOperator::CreateAShr(X, ConstantInt::get(Ty, AmtSum));
7272 InsertNewInstBefore(Shift, I);
7274 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
7275 return BinaryOperator::CreateAnd(Shift, ConstantInt::get(Mask));
7278 // Okay, if we get here, one shift must be left, and the other shift must be
7279 // right. See if the amounts are equal.
7280 if (ShiftAmt1 == ShiftAmt2) {
7281 // If we have ((X >>? C) << C), turn this into X & (-1 << C).
7282 if (I.getOpcode() == Instruction::Shl) {
7283 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt1));
7284 return BinaryOperator::CreateAnd(X, ConstantInt::get(Mask));
7286 // If we have ((X << C) >>u C), turn this into X & (-1 >>u C).
7287 if (I.getOpcode() == Instruction::LShr) {
7288 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt1));
7289 return BinaryOperator::CreateAnd(X, ConstantInt::get(Mask));
7291 // We can simplify ((X << C) >>s C) into a trunc + sext.
7292 // NOTE: we could do this for any C, but that would make 'unusual' integer
7293 // types. For now, just stick to ones well-supported by the code
7295 const Type *SExtType = 0;
7296 switch (Ty->getBitWidth() - ShiftAmt1) {
7303 SExtType = IntegerType::get(Ty->getBitWidth() - ShiftAmt1);
7308 Instruction *NewTrunc = new TruncInst(X, SExtType, "sext");
7309 InsertNewInstBefore(NewTrunc, I);
7310 return new SExtInst(NewTrunc, Ty);
7312 // Otherwise, we can't handle it yet.
7313 } else if (ShiftAmt1 < ShiftAmt2) {
7314 uint32_t ShiftDiff = ShiftAmt2-ShiftAmt1;
7316 // (X >>? C1) << C2 --> X << (C2-C1) & (-1 << C2)
7317 if (I.getOpcode() == Instruction::Shl) {
7318 assert(ShiftOp->getOpcode() == Instruction::LShr ||
7319 ShiftOp->getOpcode() == Instruction::AShr);
7320 Instruction *Shift =
7321 BinaryOperator::CreateShl(X, ConstantInt::get(Ty, ShiftDiff));
7322 InsertNewInstBefore(Shift, I);
7324 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
7325 return BinaryOperator::CreateAnd(Shift, ConstantInt::get(Mask));
7328 // (X << C1) >>u C2 --> X >>u (C2-C1) & (-1 >> C2)
7329 if (I.getOpcode() == Instruction::LShr) {
7330 assert(ShiftOp->getOpcode() == Instruction::Shl);
7331 Instruction *Shift =
7332 BinaryOperator::CreateLShr(X, ConstantInt::get(Ty, ShiftDiff));
7333 InsertNewInstBefore(Shift, I);
7335 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
7336 return BinaryOperator::CreateAnd(Shift, ConstantInt::get(Mask));
7339 // We can't handle (X << C1) >>s C2, it shifts arbitrary bits in.
7341 assert(ShiftAmt2 < ShiftAmt1);
7342 uint32_t ShiftDiff = ShiftAmt1-ShiftAmt2;
7344 // (X >>? C1) << C2 --> X >>? (C1-C2) & (-1 << C2)
7345 if (I.getOpcode() == Instruction::Shl) {
7346 assert(ShiftOp->getOpcode() == Instruction::LShr ||
7347 ShiftOp->getOpcode() == Instruction::AShr);
7348 Instruction *Shift =
7349 BinaryOperator::Create(ShiftOp->getOpcode(), X,
7350 ConstantInt::get(Ty, ShiftDiff));
7351 InsertNewInstBefore(Shift, I);
7353 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
7354 return BinaryOperator::CreateAnd(Shift, ConstantInt::get(Mask));
7357 // (X << C1) >>u C2 --> X << (C1-C2) & (-1 >> C2)
7358 if (I.getOpcode() == Instruction::LShr) {
7359 assert(ShiftOp->getOpcode() == Instruction::Shl);
7360 Instruction *Shift =
7361 BinaryOperator::CreateShl(X, ConstantInt::get(Ty, ShiftDiff));
7362 InsertNewInstBefore(Shift, I);
7364 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
7365 return BinaryOperator::CreateAnd(Shift, ConstantInt::get(Mask));
7368 // We can't handle (X << C1) >>a C2, it shifts arbitrary bits in.
7375 /// DecomposeSimpleLinearExpr - Analyze 'Val', seeing if it is a simple linear
7376 /// expression. If so, decompose it, returning some value X, such that Val is
7379 static Value *DecomposeSimpleLinearExpr(Value *Val, unsigned &Scale,
7381 assert(Val->getType() == Type::Int32Ty && "Unexpected allocation size type!");
7382 if (ConstantInt *CI = dyn_cast<ConstantInt>(Val)) {
7383 Offset = CI->getZExtValue();
7385 return ConstantInt::get(Type::Int32Ty, 0);
7386 } else if (BinaryOperator *I = dyn_cast<BinaryOperator>(Val)) {
7387 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
7388 if (I->getOpcode() == Instruction::Shl) {
7389 // This is a value scaled by '1 << the shift amt'.
7390 Scale = 1U << RHS->getZExtValue();
7392 return I->getOperand(0);
7393 } else if (I->getOpcode() == Instruction::Mul) {
7394 // This value is scaled by 'RHS'.
7395 Scale = RHS->getZExtValue();
7397 return I->getOperand(0);
7398 } else if (I->getOpcode() == Instruction::Add) {
7399 // We have X+C. Check to see if we really have (X*C2)+C1,
7400 // where C1 is divisible by C2.
7403 DecomposeSimpleLinearExpr(I->getOperand(0), SubScale, Offset);
7404 Offset += RHS->getZExtValue();
7411 // Otherwise, we can't look past this.
7418 /// PromoteCastOfAllocation - If we find a cast of an allocation instruction,
7419 /// try to eliminate the cast by moving the type information into the alloc.
7420 Instruction *InstCombiner::PromoteCastOfAllocation(BitCastInst &CI,
7421 AllocationInst &AI) {
7422 const PointerType *PTy = cast<PointerType>(CI.getType());
7424 // Remove any uses of AI that are dead.
7425 assert(!CI.use_empty() && "Dead instructions should be removed earlier!");
7427 for (Value::use_iterator UI = AI.use_begin(), E = AI.use_end(); UI != E; ) {
7428 Instruction *User = cast<Instruction>(*UI++);
7429 if (isInstructionTriviallyDead(User)) {
7430 while (UI != E && *UI == User)
7431 ++UI; // If this instruction uses AI more than once, don't break UI.
7434 DOUT << "IC: DCE: " << *User;
7435 EraseInstFromFunction(*User);
7439 // Get the type really allocated and the type casted to.
7440 const Type *AllocElTy = AI.getAllocatedType();
7441 const Type *CastElTy = PTy->getElementType();
7442 if (!AllocElTy->isSized() || !CastElTy->isSized()) return 0;
7444 unsigned AllocElTyAlign = TD->getABITypeAlignment(AllocElTy);
7445 unsigned CastElTyAlign = TD->getABITypeAlignment(CastElTy);
7446 if (CastElTyAlign < AllocElTyAlign) return 0;
7448 // If the allocation has multiple uses, only promote it if we are strictly
7449 // increasing the alignment of the resultant allocation. If we keep it the
7450 // same, we open the door to infinite loops of various kinds.
7451 if (!AI.hasOneUse() && CastElTyAlign == AllocElTyAlign) return 0;
7453 uint64_t AllocElTySize = TD->getABITypeSize(AllocElTy);
7454 uint64_t CastElTySize = TD->getABITypeSize(CastElTy);
7455 if (CastElTySize == 0 || AllocElTySize == 0) return 0;
7457 // See if we can satisfy the modulus by pulling a scale out of the array
7459 unsigned ArraySizeScale;
7461 Value *NumElements = // See if the array size is a decomposable linear expr.
7462 DecomposeSimpleLinearExpr(AI.getOperand(0), ArraySizeScale, ArrayOffset);
7464 // If we can now satisfy the modulus, by using a non-1 scale, we really can
7466 if ((AllocElTySize*ArraySizeScale) % CastElTySize != 0 ||
7467 (AllocElTySize*ArrayOffset ) % CastElTySize != 0) return 0;
7469 unsigned Scale = (AllocElTySize*ArraySizeScale)/CastElTySize;
7474 // If the allocation size is constant, form a constant mul expression
7475 Amt = ConstantInt::get(Type::Int32Ty, Scale);
7476 if (isa<ConstantInt>(NumElements))
7477 Amt = Multiply(cast<ConstantInt>(NumElements), cast<ConstantInt>(Amt));
7478 // otherwise multiply the amount and the number of elements
7479 else if (Scale != 1) {
7480 Instruction *Tmp = BinaryOperator::CreateMul(Amt, NumElements, "tmp");
7481 Amt = InsertNewInstBefore(Tmp, AI);
7485 if (int Offset = (AllocElTySize*ArrayOffset)/CastElTySize) {
7486 Value *Off = ConstantInt::get(Type::Int32Ty, Offset, true);
7487 Instruction *Tmp = BinaryOperator::CreateAdd(Amt, Off, "tmp");
7488 Amt = InsertNewInstBefore(Tmp, AI);
7491 AllocationInst *New;
7492 if (isa<MallocInst>(AI))
7493 New = new MallocInst(CastElTy, Amt, AI.getAlignment());
7495 New = new AllocaInst(CastElTy, Amt, AI.getAlignment());
7496 InsertNewInstBefore(New, AI);
7499 // If the allocation has multiple uses, insert a cast and change all things
7500 // that used it to use the new cast. This will also hack on CI, but it will
7502 if (!AI.hasOneUse()) {
7503 AddUsesToWorkList(AI);
7504 // New is the allocation instruction, pointer typed. AI is the original
7505 // allocation instruction, also pointer typed. Thus, cast to use is BitCast.
7506 CastInst *NewCast = new BitCastInst(New, AI.getType(), "tmpcast");
7507 InsertNewInstBefore(NewCast, AI);
7508 AI.replaceAllUsesWith(NewCast);
7510 return ReplaceInstUsesWith(CI, New);
7513 /// CanEvaluateInDifferentType - Return true if we can take the specified value
7514 /// and return it as type Ty without inserting any new casts and without
7515 /// changing the computed value. This is used by code that tries to decide
7516 /// whether promoting or shrinking integer operations to wider or smaller types
7517 /// will allow us to eliminate a truncate or extend.
7519 /// This is a truncation operation if Ty is smaller than V->getType(), or an
7520 /// extension operation if Ty is larger.
7522 /// If CastOpc is a truncation, then Ty will be a type smaller than V. We
7523 /// should return true if trunc(V) can be computed by computing V in the smaller
7524 /// type. If V is an instruction, then trunc(inst(x,y)) can be computed as
7525 /// inst(trunc(x),trunc(y)), which only makes sense if x and y can be
7526 /// efficiently truncated.
7528 /// If CastOpc is a sext or zext, we are asking if the low bits of the value can
7529 /// bit computed in a larger type, which is then and'd or sext_in_reg'd to get
7530 /// the final result.
7531 bool InstCombiner::CanEvaluateInDifferentType(Value *V, const IntegerType *Ty,
7533 int &NumCastsRemoved) {
7534 // We can always evaluate constants in another type.
7535 if (isa<ConstantInt>(V))
7538 Instruction *I = dyn_cast<Instruction>(V);
7539 if (!I) return false;
7541 const IntegerType *OrigTy = cast<IntegerType>(V->getType());
7543 // If this is an extension or truncate, we can often eliminate it.
7544 if (isa<TruncInst>(I) || isa<ZExtInst>(I) || isa<SExtInst>(I)) {
7545 // If this is a cast from the destination type, we can trivially eliminate
7546 // it, and this will remove a cast overall.
7547 if (I->getOperand(0)->getType() == Ty) {
7548 // If the first operand is itself a cast, and is eliminable, do not count
7549 // this as an eliminable cast. We would prefer to eliminate those two
7551 if (!isa<CastInst>(I->getOperand(0)) && I->hasOneUse())
7557 // We can't extend or shrink something that has multiple uses: doing so would
7558 // require duplicating the instruction in general, which isn't profitable.
7559 if (!I->hasOneUse()) return false;
7561 switch (I->getOpcode()) {
7562 case Instruction::Add:
7563 case Instruction::Sub:
7564 case Instruction::Mul:
7565 case Instruction::And:
7566 case Instruction::Or:
7567 case Instruction::Xor:
7568 // These operators can all arbitrarily be extended or truncated.
7569 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
7571 CanEvaluateInDifferentType(I->getOperand(1), Ty, CastOpc,
7574 case Instruction::Shl:
7575 // If we are truncating the result of this SHL, and if it's a shift of a
7576 // constant amount, we can always perform a SHL in a smaller type.
7577 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
7578 uint32_t BitWidth = Ty->getBitWidth();
7579 if (BitWidth < OrigTy->getBitWidth() &&
7580 CI->getLimitedValue(BitWidth) < BitWidth)
7581 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
7585 case Instruction::LShr:
7586 // If this is a truncate of a logical shr, we can truncate it to a smaller
7587 // lshr iff we know that the bits we would otherwise be shifting in are
7589 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
7590 uint32_t OrigBitWidth = OrigTy->getBitWidth();
7591 uint32_t BitWidth = Ty->getBitWidth();
7592 if (BitWidth < OrigBitWidth &&
7593 MaskedValueIsZero(I->getOperand(0),
7594 APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth)) &&
7595 CI->getLimitedValue(BitWidth) < BitWidth) {
7596 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
7601 case Instruction::ZExt:
7602 case Instruction::SExt:
7603 case Instruction::Trunc:
7604 // If this is the same kind of case as our original (e.g. zext+zext), we
7605 // can safely replace it. Note that replacing it does not reduce the number
7606 // of casts in the input.
7607 if (I->getOpcode() == CastOpc)
7610 case Instruction::Select: {
7611 SelectInst *SI = cast<SelectInst>(I);
7612 return CanEvaluateInDifferentType(SI->getTrueValue(), Ty, CastOpc,
7614 CanEvaluateInDifferentType(SI->getFalseValue(), Ty, CastOpc,
7617 case Instruction::PHI: {
7618 // We can change a phi if we can change all operands.
7619 PHINode *PN = cast<PHINode>(I);
7620 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
7621 if (!CanEvaluateInDifferentType(PN->getIncomingValue(i), Ty, CastOpc,
7627 // TODO: Can handle more cases here.
7634 /// EvaluateInDifferentType - Given an expression that
7635 /// CanEvaluateInDifferentType returns true for, actually insert the code to
7636 /// evaluate the expression.
7637 Value *InstCombiner::EvaluateInDifferentType(Value *V, const Type *Ty,
7639 if (Constant *C = dyn_cast<Constant>(V))
7640 return ConstantExpr::getIntegerCast(C, Ty, isSigned /*Sext or ZExt*/);
7642 // Otherwise, it must be an instruction.
7643 Instruction *I = cast<Instruction>(V);
7644 Instruction *Res = 0;
7645 switch (I->getOpcode()) {
7646 case Instruction::Add:
7647 case Instruction::Sub:
7648 case Instruction::Mul:
7649 case Instruction::And:
7650 case Instruction::Or:
7651 case Instruction::Xor:
7652 case Instruction::AShr:
7653 case Instruction::LShr:
7654 case Instruction::Shl: {
7655 Value *LHS = EvaluateInDifferentType(I->getOperand(0), Ty, isSigned);
7656 Value *RHS = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
7657 Res = BinaryOperator::Create((Instruction::BinaryOps)I->getOpcode(),
7661 case Instruction::Trunc:
7662 case Instruction::ZExt:
7663 case Instruction::SExt:
7664 // If the source type of the cast is the type we're trying for then we can
7665 // just return the source. There's no need to insert it because it is not
7667 if (I->getOperand(0)->getType() == Ty)
7668 return I->getOperand(0);
7670 // Otherwise, must be the same type of cast, so just reinsert a new one.
7671 Res = CastInst::Create(cast<CastInst>(I)->getOpcode(), I->getOperand(0),
7674 case Instruction::Select: {
7675 Value *True = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
7676 Value *False = EvaluateInDifferentType(I->getOperand(2), Ty, isSigned);
7677 Res = SelectInst::Create(I->getOperand(0), True, False);
7680 case Instruction::PHI: {
7681 PHINode *OPN = cast<PHINode>(I);
7682 PHINode *NPN = PHINode::Create(Ty);
7683 for (unsigned i = 0, e = OPN->getNumIncomingValues(); i != e; ++i) {
7684 Value *V =EvaluateInDifferentType(OPN->getIncomingValue(i), Ty, isSigned);
7685 NPN->addIncoming(V, OPN->getIncomingBlock(i));
7691 // TODO: Can handle more cases here.
7692 assert(0 && "Unreachable!");
7697 return InsertNewInstBefore(Res, *I);
7700 /// @brief Implement the transforms common to all CastInst visitors.
7701 Instruction *InstCombiner::commonCastTransforms(CastInst &CI) {
7702 Value *Src = CI.getOperand(0);
7704 // Many cases of "cast of a cast" are eliminable. If it's eliminable we just
7705 // eliminate it now.
7706 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
7707 if (Instruction::CastOps opc =
7708 isEliminableCastPair(CSrc, CI.getOpcode(), CI.getType(), TD)) {
7709 // The first cast (CSrc) is eliminable so we need to fix up or replace
7710 // the second cast (CI). CSrc will then have a good chance of being dead.
7711 return CastInst::Create(opc, CSrc->getOperand(0), CI.getType());
7715 // If we are casting a select then fold the cast into the select
7716 if (SelectInst *SI = dyn_cast<SelectInst>(Src))
7717 if (Instruction *NV = FoldOpIntoSelect(CI, SI, this))
7720 // If we are casting a PHI then fold the cast into the PHI
7721 if (isa<PHINode>(Src))
7722 if (Instruction *NV = FoldOpIntoPhi(CI))
7728 /// @brief Implement the transforms for cast of pointer (bitcast/ptrtoint)
7729 Instruction *InstCombiner::commonPointerCastTransforms(CastInst &CI) {
7730 Value *Src = CI.getOperand(0);
7732 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Src)) {
7733 // If casting the result of a getelementptr instruction with no offset, turn
7734 // this into a cast of the original pointer!
7735 if (GEP->hasAllZeroIndices()) {
7736 // Changing the cast operand is usually not a good idea but it is safe
7737 // here because the pointer operand is being replaced with another
7738 // pointer operand so the opcode doesn't need to change.
7740 CI.setOperand(0, GEP->getOperand(0));
7744 // If the GEP has a single use, and the base pointer is a bitcast, and the
7745 // GEP computes a constant offset, see if we can convert these three
7746 // instructions into fewer. This typically happens with unions and other
7747 // non-type-safe code.
7748 if (GEP->hasOneUse() && isa<BitCastInst>(GEP->getOperand(0))) {
7749 if (GEP->hasAllConstantIndices()) {
7750 // We are guaranteed to get a constant from EmitGEPOffset.
7751 ConstantInt *OffsetV = cast<ConstantInt>(EmitGEPOffset(GEP, CI, *this));
7752 int64_t Offset = OffsetV->getSExtValue();
7754 // Get the base pointer input of the bitcast, and the type it points to.
7755 Value *OrigBase = cast<BitCastInst>(GEP->getOperand(0))->getOperand(0);
7756 const Type *GEPIdxTy =
7757 cast<PointerType>(OrigBase->getType())->getElementType();
7758 if (GEPIdxTy->isSized()) {
7759 SmallVector<Value*, 8> NewIndices;
7761 // Start with the index over the outer type. Note that the type size
7762 // might be zero (even if the offset isn't zero) if the indexed type
7763 // is something like [0 x {int, int}]
7764 const Type *IntPtrTy = TD->getIntPtrType();
7765 int64_t FirstIdx = 0;
7766 if (int64_t TySize = TD->getABITypeSize(GEPIdxTy)) {
7767 FirstIdx = Offset/TySize;
7770 // Handle silly modulus not returning values values [0..TySize).
7774 assert(Offset >= 0);
7776 assert((uint64_t)Offset < (uint64_t)TySize &&"Out of range offset");
7779 NewIndices.push_back(ConstantInt::get(IntPtrTy, FirstIdx));
7781 // Index into the types. If we fail, set OrigBase to null.
7783 if (const StructType *STy = dyn_cast<StructType>(GEPIdxTy)) {
7784 const StructLayout *SL = TD->getStructLayout(STy);
7785 if (Offset < (int64_t)SL->getSizeInBytes()) {
7786 unsigned Elt = SL->getElementContainingOffset(Offset);
7787 NewIndices.push_back(ConstantInt::get(Type::Int32Ty, Elt));
7789 Offset -= SL->getElementOffset(Elt);
7790 GEPIdxTy = STy->getElementType(Elt);
7792 // Otherwise, we can't index into this, bail out.
7796 } else if (isa<ArrayType>(GEPIdxTy) || isa<VectorType>(GEPIdxTy)) {
7797 const SequentialType *STy = cast<SequentialType>(GEPIdxTy);
7798 if (uint64_t EltSize = TD->getABITypeSize(STy->getElementType())){
7799 NewIndices.push_back(ConstantInt::get(IntPtrTy,Offset/EltSize));
7802 NewIndices.push_back(ConstantInt::get(IntPtrTy, 0));
7804 GEPIdxTy = STy->getElementType();
7806 // Otherwise, we can't index into this, bail out.
7812 // If we were able to index down into an element, create the GEP
7813 // and bitcast the result. This eliminates one bitcast, potentially
7815 Instruction *NGEP = GetElementPtrInst::Create(OrigBase,
7817 NewIndices.end(), "");
7818 InsertNewInstBefore(NGEP, CI);
7819 NGEP->takeName(GEP);
7821 if (isa<BitCastInst>(CI))
7822 return new BitCastInst(NGEP, CI.getType());
7823 assert(isa<PtrToIntInst>(CI));
7824 return new PtrToIntInst(NGEP, CI.getType());
7831 return commonCastTransforms(CI);
7836 /// Only the TRUNC, ZEXT, SEXT, and BITCAST can both operand and result as
7837 /// integer types. This function implements the common transforms for all those
7839 /// @brief Implement the transforms common to CastInst with integer operands
7840 Instruction *InstCombiner::commonIntCastTransforms(CastInst &CI) {
7841 if (Instruction *Result = commonCastTransforms(CI))
7844 Value *Src = CI.getOperand(0);
7845 const Type *SrcTy = Src->getType();
7846 const Type *DestTy = CI.getType();
7847 uint32_t SrcBitSize = SrcTy->getPrimitiveSizeInBits();
7848 uint32_t DestBitSize = DestTy->getPrimitiveSizeInBits();
7850 // See if we can simplify any instructions used by the LHS whose sole
7851 // purpose is to compute bits we don't care about.
7852 APInt KnownZero(DestBitSize, 0), KnownOne(DestBitSize, 0);
7853 if (SimplifyDemandedBits(&CI, APInt::getAllOnesValue(DestBitSize),
7854 KnownZero, KnownOne))
7857 // If the source isn't an instruction or has more than one use then we
7858 // can't do anything more.
7859 Instruction *SrcI = dyn_cast<Instruction>(Src);
7860 if (!SrcI || !Src->hasOneUse())
7863 // Attempt to propagate the cast into the instruction for int->int casts.
7864 int NumCastsRemoved = 0;
7865 if (!isa<BitCastInst>(CI) &&
7866 CanEvaluateInDifferentType(SrcI, cast<IntegerType>(DestTy),
7867 CI.getOpcode(), NumCastsRemoved)) {
7868 // If this cast is a truncate, evaluting in a different type always
7869 // eliminates the cast, so it is always a win. If this is a zero-extension,
7870 // we need to do an AND to maintain the clear top-part of the computation,
7871 // so we require that the input have eliminated at least one cast. If this
7872 // is a sign extension, we insert two new casts (to do the extension) so we
7873 // require that two casts have been eliminated.
7875 switch (CI.getOpcode()) {
7877 // All the others use floating point so we shouldn't actually
7878 // get here because of the check above.
7879 assert(0 && "Unknown cast type");
7880 case Instruction::Trunc:
7883 case Instruction::ZExt:
7884 DoXForm = NumCastsRemoved >= 1;
7886 case Instruction::SExt:
7887 DoXForm = NumCastsRemoved >= 2;
7892 Value *Res = EvaluateInDifferentType(SrcI, DestTy,
7893 CI.getOpcode() == Instruction::SExt);
7894 assert(Res->getType() == DestTy);
7895 switch (CI.getOpcode()) {
7896 default: assert(0 && "Unknown cast type!");
7897 case Instruction::Trunc:
7898 case Instruction::BitCast:
7899 // Just replace this cast with the result.
7900 return ReplaceInstUsesWith(CI, Res);
7901 case Instruction::ZExt: {
7902 // We need to emit an AND to clear the high bits.
7903 assert(SrcBitSize < DestBitSize && "Not a zext?");
7904 Constant *C = ConstantInt::get(APInt::getLowBitsSet(DestBitSize,
7906 return BinaryOperator::CreateAnd(Res, C);
7908 case Instruction::SExt:
7909 // We need to emit a cast to truncate, then a cast to sext.
7910 return CastInst::Create(Instruction::SExt,
7911 InsertCastBefore(Instruction::Trunc, Res, Src->getType(),
7917 Value *Op0 = SrcI->getNumOperands() > 0 ? SrcI->getOperand(0) : 0;
7918 Value *Op1 = SrcI->getNumOperands() > 1 ? SrcI->getOperand(1) : 0;
7920 switch (SrcI->getOpcode()) {
7921 case Instruction::Add:
7922 case Instruction::Mul:
7923 case Instruction::And:
7924 case Instruction::Or:
7925 case Instruction::Xor:
7926 // If we are discarding information, rewrite.
7927 if (DestBitSize <= SrcBitSize && DestBitSize != 1) {
7928 // Don't insert two casts if they cannot be eliminated. We allow
7929 // two casts to be inserted if the sizes are the same. This could
7930 // only be converting signedness, which is a noop.
7931 if (DestBitSize == SrcBitSize ||
7932 !ValueRequiresCast(CI.getOpcode(), Op1, DestTy,TD) ||
7933 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
7934 Instruction::CastOps opcode = CI.getOpcode();
7935 Value *Op0c = InsertCastBefore(opcode, Op0, DestTy, *SrcI);
7936 Value *Op1c = InsertCastBefore(opcode, Op1, DestTy, *SrcI);
7937 return BinaryOperator::Create(
7938 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
7942 // cast (xor bool X, true) to int --> xor (cast bool X to int), 1
7943 if (isa<ZExtInst>(CI) && SrcBitSize == 1 &&
7944 SrcI->getOpcode() == Instruction::Xor &&
7945 Op1 == ConstantInt::getTrue() &&
7946 (!Op0->hasOneUse() || !isa<CmpInst>(Op0))) {
7947 Value *New = InsertCastBefore(Instruction::ZExt, Op0, DestTy, CI);
7948 return BinaryOperator::CreateXor(New, ConstantInt::get(CI.getType(), 1));
7951 case Instruction::SDiv:
7952 case Instruction::UDiv:
7953 case Instruction::SRem:
7954 case Instruction::URem:
7955 // If we are just changing the sign, rewrite.
7956 if (DestBitSize == SrcBitSize) {
7957 // Don't insert two casts if they cannot be eliminated. We allow
7958 // two casts to be inserted if the sizes are the same. This could
7959 // only be converting signedness, which is a noop.
7960 if (!ValueRequiresCast(CI.getOpcode(), Op1, DestTy, TD) ||
7961 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
7962 Value *Op0c = InsertCastBefore(Instruction::BitCast,
7963 Op0, DestTy, *SrcI);
7964 Value *Op1c = InsertCastBefore(Instruction::BitCast,
7965 Op1, DestTy, *SrcI);
7966 return BinaryOperator::Create(
7967 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
7972 case Instruction::Shl:
7973 // Allow changing the sign of the source operand. Do not allow
7974 // changing the size of the shift, UNLESS the shift amount is a
7975 // constant. We must not change variable sized shifts to a smaller
7976 // size, because it is undefined to shift more bits out than exist
7978 if (DestBitSize == SrcBitSize ||
7979 (DestBitSize < SrcBitSize && isa<Constant>(Op1))) {
7980 Instruction::CastOps opcode = (DestBitSize == SrcBitSize ?
7981 Instruction::BitCast : Instruction::Trunc);
7982 Value *Op0c = InsertCastBefore(opcode, Op0, DestTy, *SrcI);
7983 Value *Op1c = InsertCastBefore(opcode, Op1, DestTy, *SrcI);
7984 return BinaryOperator::CreateShl(Op0c, Op1c);
7987 case Instruction::AShr:
7988 // If this is a signed shr, and if all bits shifted in are about to be
7989 // truncated off, turn it into an unsigned shr to allow greater
7991 if (DestBitSize < SrcBitSize &&
7992 isa<ConstantInt>(Op1)) {
7993 uint32_t ShiftAmt = cast<ConstantInt>(Op1)->getLimitedValue(SrcBitSize);
7994 if (SrcBitSize > ShiftAmt && SrcBitSize-ShiftAmt >= DestBitSize) {
7995 // Insert the new logical shift right.
7996 return BinaryOperator::CreateLShr(Op0, Op1);
8004 Instruction *InstCombiner::visitTrunc(TruncInst &CI) {
8005 if (Instruction *Result = commonIntCastTransforms(CI))
8008 Value *Src = CI.getOperand(0);
8009 const Type *Ty = CI.getType();
8010 uint32_t DestBitWidth = Ty->getPrimitiveSizeInBits();
8011 uint32_t SrcBitWidth = cast<IntegerType>(Src->getType())->getBitWidth();
8013 if (Instruction *SrcI = dyn_cast<Instruction>(Src)) {
8014 switch (SrcI->getOpcode()) {
8016 case Instruction::LShr:
8017 // We can shrink lshr to something smaller if we know the bits shifted in
8018 // are already zeros.
8019 if (ConstantInt *ShAmtV = dyn_cast<ConstantInt>(SrcI->getOperand(1))) {
8020 uint32_t ShAmt = ShAmtV->getLimitedValue(SrcBitWidth);
8022 // Get a mask for the bits shifting in.
8023 APInt Mask(APInt::getLowBitsSet(SrcBitWidth, ShAmt).shl(DestBitWidth));
8024 Value* SrcIOp0 = SrcI->getOperand(0);
8025 if (SrcI->hasOneUse() && MaskedValueIsZero(SrcIOp0, Mask)) {
8026 if (ShAmt >= DestBitWidth) // All zeros.
8027 return ReplaceInstUsesWith(CI, Constant::getNullValue(Ty));
8029 // Okay, we can shrink this. Truncate the input, then return a new
8031 Value *V1 = InsertCastBefore(Instruction::Trunc, SrcIOp0, Ty, CI);
8032 Value *V2 = InsertCastBefore(Instruction::Trunc, SrcI->getOperand(1),
8034 return BinaryOperator::CreateLShr(V1, V2);
8036 } else { // This is a variable shr.
8038 // Turn 'trunc (lshr X, Y) to bool' into '(X & (1 << Y)) != 0'. This is
8039 // more LLVM instructions, but allows '1 << Y' to be hoisted if
8040 // loop-invariant and CSE'd.
8041 if (CI.getType() == Type::Int1Ty && SrcI->hasOneUse()) {
8042 Value *One = ConstantInt::get(SrcI->getType(), 1);
8044 Value *V = InsertNewInstBefore(
8045 BinaryOperator::CreateShl(One, SrcI->getOperand(1),
8047 V = InsertNewInstBefore(BinaryOperator::CreateAnd(V,
8048 SrcI->getOperand(0),
8050 Value *Zero = Constant::getNullValue(V->getType());
8051 return new ICmpInst(ICmpInst::ICMP_NE, V, Zero);
8061 /// transformZExtICmp - Transform (zext icmp) to bitwise / integer operations
8062 /// in order to eliminate the icmp.
8063 Instruction *InstCombiner::transformZExtICmp(ICmpInst *ICI, Instruction &CI,
8065 // If we are just checking for a icmp eq of a single bit and zext'ing it
8066 // to an integer, then shift the bit to the appropriate place and then
8067 // cast to integer to avoid the comparison.
8068 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(ICI->getOperand(1))) {
8069 const APInt &Op1CV = Op1C->getValue();
8071 // zext (x <s 0) to i32 --> x>>u31 true if signbit set.
8072 // zext (x >s -1) to i32 --> (x>>u31)^1 true if signbit clear.
8073 if ((ICI->getPredicate() == ICmpInst::ICMP_SLT && Op1CV == 0) ||
8074 (ICI->getPredicate() == ICmpInst::ICMP_SGT &&Op1CV.isAllOnesValue())) {
8075 if (!DoXform) return ICI;
8077 Value *In = ICI->getOperand(0);
8078 Value *Sh = ConstantInt::get(In->getType(),
8079 In->getType()->getPrimitiveSizeInBits()-1);
8080 In = InsertNewInstBefore(BinaryOperator::CreateLShr(In, Sh,
8081 In->getName()+".lobit"),
8083 if (In->getType() != CI.getType())
8084 In = CastInst::CreateIntegerCast(In, CI.getType(),
8085 false/*ZExt*/, "tmp", &CI);
8087 if (ICI->getPredicate() == ICmpInst::ICMP_SGT) {
8088 Constant *One = ConstantInt::get(In->getType(), 1);
8089 In = InsertNewInstBefore(BinaryOperator::CreateXor(In, One,
8090 In->getName()+".not"),
8094 return ReplaceInstUsesWith(CI, In);
8099 // zext (X == 0) to i32 --> X^1 iff X has only the low bit set.
8100 // zext (X == 0) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
8101 // zext (X == 1) to i32 --> X iff X has only the low bit set.
8102 // zext (X == 2) to i32 --> X>>1 iff X has only the 2nd bit set.
8103 // zext (X != 0) to i32 --> X iff X has only the low bit set.
8104 // zext (X != 0) to i32 --> X>>1 iff X has only the 2nd bit set.
8105 // zext (X != 1) to i32 --> X^1 iff X has only the low bit set.
8106 // zext (X != 2) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
8107 if ((Op1CV == 0 || Op1CV.isPowerOf2()) &&
8108 // This only works for EQ and NE
8109 ICI->isEquality()) {
8110 // If Op1C some other power of two, convert:
8111 uint32_t BitWidth = Op1C->getType()->getBitWidth();
8112 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
8113 APInt TypeMask(APInt::getAllOnesValue(BitWidth));
8114 ComputeMaskedBits(ICI->getOperand(0), TypeMask, KnownZero, KnownOne);
8116 APInt KnownZeroMask(~KnownZero);
8117 if (KnownZeroMask.isPowerOf2()) { // Exactly 1 possible 1?
8118 if (!DoXform) return ICI;
8120 bool isNE = ICI->getPredicate() == ICmpInst::ICMP_NE;
8121 if (Op1CV != 0 && (Op1CV != KnownZeroMask)) {
8122 // (X&4) == 2 --> false
8123 // (X&4) != 2 --> true
8124 Constant *Res = ConstantInt::get(Type::Int1Ty, isNE);
8125 Res = ConstantExpr::getZExt(Res, CI.getType());
8126 return ReplaceInstUsesWith(CI, Res);
8129 uint32_t ShiftAmt = KnownZeroMask.logBase2();
8130 Value *In = ICI->getOperand(0);
8132 // Perform a logical shr by shiftamt.
8133 // Insert the shift to put the result in the low bit.
8134 In = InsertNewInstBefore(BinaryOperator::CreateLShr(In,
8135 ConstantInt::get(In->getType(), ShiftAmt),
8136 In->getName()+".lobit"), CI);
8139 if ((Op1CV != 0) == isNE) { // Toggle the low bit.
8140 Constant *One = ConstantInt::get(In->getType(), 1);
8141 In = BinaryOperator::CreateXor(In, One, "tmp");
8142 InsertNewInstBefore(cast<Instruction>(In), CI);
8145 if (CI.getType() == In->getType())
8146 return ReplaceInstUsesWith(CI, In);
8148 return CastInst::CreateIntegerCast(In, CI.getType(), false/*ZExt*/);
8156 Instruction *InstCombiner::visitZExt(ZExtInst &CI) {
8157 // If one of the common conversion will work ..
8158 if (Instruction *Result = commonIntCastTransforms(CI))
8161 Value *Src = CI.getOperand(0);
8163 // If this is a cast of a cast
8164 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
8165 // If this is a TRUNC followed by a ZEXT then we are dealing with integral
8166 // types and if the sizes are just right we can convert this into a logical
8167 // 'and' which will be much cheaper than the pair of casts.
8168 if (isa<TruncInst>(CSrc)) {
8169 // Get the sizes of the types involved
8170 Value *A = CSrc->getOperand(0);
8171 uint32_t SrcSize = A->getType()->getPrimitiveSizeInBits();
8172 uint32_t MidSize = CSrc->getType()->getPrimitiveSizeInBits();
8173 uint32_t DstSize = CI.getType()->getPrimitiveSizeInBits();
8174 // If we're actually extending zero bits and the trunc is a no-op
8175 if (MidSize < DstSize && SrcSize == DstSize) {
8176 // Replace both of the casts with an And of the type mask.
8177 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
8178 Constant *AndConst = ConstantInt::get(AndValue);
8180 BinaryOperator::CreateAnd(CSrc->getOperand(0), AndConst);
8181 // Unfortunately, if the type changed, we need to cast it back.
8182 if (And->getType() != CI.getType()) {
8183 And->setName(CSrc->getName()+".mask");
8184 InsertNewInstBefore(And, CI);
8185 And = CastInst::CreateIntegerCast(And, CI.getType(), false/*ZExt*/);
8192 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src))
8193 return transformZExtICmp(ICI, CI);
8195 BinaryOperator *SrcI = dyn_cast<BinaryOperator>(Src);
8196 if (SrcI && SrcI->getOpcode() == Instruction::Or) {
8197 // zext (or icmp, icmp) --> or (zext icmp), (zext icmp) if at least one
8198 // of the (zext icmp) will be transformed.
8199 ICmpInst *LHS = dyn_cast<ICmpInst>(SrcI->getOperand(0));
8200 ICmpInst *RHS = dyn_cast<ICmpInst>(SrcI->getOperand(1));
8201 if (LHS && RHS && LHS->hasOneUse() && RHS->hasOneUse() &&
8202 (transformZExtICmp(LHS, CI, false) ||
8203 transformZExtICmp(RHS, CI, false))) {
8204 Value *LCast = InsertCastBefore(Instruction::ZExt, LHS, CI.getType(), CI);
8205 Value *RCast = InsertCastBefore(Instruction::ZExt, RHS, CI.getType(), CI);
8206 return BinaryOperator::Create(Instruction::Or, LCast, RCast);
8213 Instruction *InstCombiner::visitSExt(SExtInst &CI) {
8214 if (Instruction *I = commonIntCastTransforms(CI))
8217 Value *Src = CI.getOperand(0);
8219 // Canonicalize sign-extend from i1 to a select.
8220 if (Src->getType() == Type::Int1Ty)
8221 return SelectInst::Create(Src,
8222 ConstantInt::getAllOnesValue(CI.getType()),
8223 Constant::getNullValue(CI.getType()));
8225 // See if the value being truncated is already sign extended. If so, just
8226 // eliminate the trunc/sext pair.
8227 if (getOpcode(Src) == Instruction::Trunc) {
8228 Value *Op = cast<User>(Src)->getOperand(0);
8229 unsigned OpBits = cast<IntegerType>(Op->getType())->getBitWidth();
8230 unsigned MidBits = cast<IntegerType>(Src->getType())->getBitWidth();
8231 unsigned DestBits = cast<IntegerType>(CI.getType())->getBitWidth();
8232 unsigned NumSignBits = ComputeNumSignBits(Op);
8234 if (OpBits == DestBits) {
8235 // Op is i32, Mid is i8, and Dest is i32. If Op has more than 24 sign
8236 // bits, it is already ready.
8237 if (NumSignBits > DestBits-MidBits)
8238 return ReplaceInstUsesWith(CI, Op);
8239 } else if (OpBits < DestBits) {
8240 // Op is i32, Mid is i8, and Dest is i64. If Op has more than 24 sign
8241 // bits, just sext from i32.
8242 if (NumSignBits > OpBits-MidBits)
8243 return new SExtInst(Op, CI.getType(), "tmp");
8245 // Op is i64, Mid is i8, and Dest is i32. If Op has more than 56 sign
8246 // bits, just truncate to i32.
8247 if (NumSignBits > OpBits-MidBits)
8248 return new TruncInst(Op, CI.getType(), "tmp");
8252 // If the input is a shl/ashr pair of a same constant, then this is a sign
8253 // extension from a smaller value. If we could trust arbitrary bitwidth
8254 // integers, we could turn this into a truncate to the smaller bit and then
8255 // use a sext for the whole extension. Since we don't, look deeper and check
8256 // for a truncate. If the source and dest are the same type, eliminate the
8257 // trunc and extend and just do shifts. For example, turn:
8258 // %a = trunc i32 %i to i8
8259 // %b = shl i8 %a, 6
8260 // %c = ashr i8 %b, 6
8261 // %d = sext i8 %c to i32
8263 // %a = shl i32 %i, 30
8264 // %d = ashr i32 %a, 30
8266 ConstantInt *BA = 0, *CA = 0;
8267 if (match(Src, m_AShr(m_Shl(m_Value(A), m_ConstantInt(BA)),
8268 m_ConstantInt(CA))) &&
8269 BA == CA && isa<TruncInst>(A)) {
8270 Value *I = cast<TruncInst>(A)->getOperand(0);
8271 if (I->getType() == CI.getType()) {
8272 unsigned MidSize = Src->getType()->getPrimitiveSizeInBits();
8273 unsigned SrcDstSize = CI.getType()->getPrimitiveSizeInBits();
8274 unsigned ShAmt = CA->getZExtValue()+SrcDstSize-MidSize;
8275 Constant *ShAmtV = ConstantInt::get(CI.getType(), ShAmt);
8276 I = InsertNewInstBefore(BinaryOperator::CreateShl(I, ShAmtV,
8278 return BinaryOperator::CreateAShr(I, ShAmtV);
8285 /// FitsInFPType - Return a Constant* for the specified FP constant if it fits
8286 /// in the specified FP type without changing its value.
8287 static Constant *FitsInFPType(ConstantFP *CFP, const fltSemantics &Sem) {
8289 APFloat F = CFP->getValueAPF();
8290 (void)F.convert(Sem, APFloat::rmNearestTiesToEven, &losesInfo);
8292 return ConstantFP::get(F);
8296 /// LookThroughFPExtensions - If this is an fp extension instruction, look
8297 /// through it until we get the source value.
8298 static Value *LookThroughFPExtensions(Value *V) {
8299 if (Instruction *I = dyn_cast<Instruction>(V))
8300 if (I->getOpcode() == Instruction::FPExt)
8301 return LookThroughFPExtensions(I->getOperand(0));
8303 // If this value is a constant, return the constant in the smallest FP type
8304 // that can accurately represent it. This allows us to turn
8305 // (float)((double)X+2.0) into x+2.0f.
8306 if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
8307 if (CFP->getType() == Type::PPC_FP128Ty)
8308 return V; // No constant folding of this.
8309 // See if the value can be truncated to float and then reextended.
8310 if (Value *V = FitsInFPType(CFP, APFloat::IEEEsingle))
8312 if (CFP->getType() == Type::DoubleTy)
8313 return V; // Won't shrink.
8314 if (Value *V = FitsInFPType(CFP, APFloat::IEEEdouble))
8316 // Don't try to shrink to various long double types.
8322 Instruction *InstCombiner::visitFPTrunc(FPTruncInst &CI) {
8323 if (Instruction *I = commonCastTransforms(CI))
8326 // If we have fptrunc(add (fpextend x), (fpextend y)), where x and y are
8327 // smaller than the destination type, we can eliminate the truncate by doing
8328 // the add as the smaller type. This applies to add/sub/mul/div as well as
8329 // many builtins (sqrt, etc).
8330 BinaryOperator *OpI = dyn_cast<BinaryOperator>(CI.getOperand(0));
8331 if (OpI && OpI->hasOneUse()) {
8332 switch (OpI->getOpcode()) {
8334 case Instruction::Add:
8335 case Instruction::Sub:
8336 case Instruction::Mul:
8337 case Instruction::FDiv:
8338 case Instruction::FRem:
8339 const Type *SrcTy = OpI->getType();
8340 Value *LHSTrunc = LookThroughFPExtensions(OpI->getOperand(0));
8341 Value *RHSTrunc = LookThroughFPExtensions(OpI->getOperand(1));
8342 if (LHSTrunc->getType() != SrcTy &&
8343 RHSTrunc->getType() != SrcTy) {
8344 unsigned DstSize = CI.getType()->getPrimitiveSizeInBits();
8345 // If the source types were both smaller than the destination type of
8346 // the cast, do this xform.
8347 if (LHSTrunc->getType()->getPrimitiveSizeInBits() <= DstSize &&
8348 RHSTrunc->getType()->getPrimitiveSizeInBits() <= DstSize) {
8349 LHSTrunc = InsertCastBefore(Instruction::FPExt, LHSTrunc,
8351 RHSTrunc = InsertCastBefore(Instruction::FPExt, RHSTrunc,
8353 return BinaryOperator::Create(OpI->getOpcode(), LHSTrunc, RHSTrunc);
8362 Instruction *InstCombiner::visitFPExt(CastInst &CI) {
8363 return commonCastTransforms(CI);
8366 Instruction *InstCombiner::visitFPToUI(FPToUIInst &FI) {
8367 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0));
8369 return commonCastTransforms(FI);
8371 // fptoui(uitofp(X)) --> X
8372 // fptoui(sitofp(X)) --> X
8373 // This is safe if the intermediate type has enough bits in its mantissa to
8374 // accurately represent all values of X. For example, do not do this with
8375 // i64->float->i64. This is also safe for sitofp case, because any negative
8376 // 'X' value would cause an undefined result for the fptoui.
8377 if ((isa<UIToFPInst>(OpI) || isa<SIToFPInst>(OpI)) &&
8378 OpI->getOperand(0)->getType() == FI.getType() &&
8379 (int)FI.getType()->getPrimitiveSizeInBits() < /*extra bit for sign */
8380 OpI->getType()->getFPMantissaWidth())
8381 return ReplaceInstUsesWith(FI, OpI->getOperand(0));
8383 return commonCastTransforms(FI);
8386 Instruction *InstCombiner::visitFPToSI(FPToSIInst &FI) {
8387 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0));
8389 return commonCastTransforms(FI);
8391 // fptosi(sitofp(X)) --> X
8392 // fptosi(uitofp(X)) --> X
8393 // This is safe if the intermediate type has enough bits in its mantissa to
8394 // accurately represent all values of X. For example, do not do this with
8395 // i64->float->i64. This is also safe for sitofp case, because any negative
8396 // 'X' value would cause an undefined result for the fptoui.
8397 if ((isa<UIToFPInst>(OpI) || isa<SIToFPInst>(OpI)) &&
8398 OpI->getOperand(0)->getType() == FI.getType() &&
8399 (int)FI.getType()->getPrimitiveSizeInBits() <=
8400 OpI->getType()->getFPMantissaWidth())
8401 return ReplaceInstUsesWith(FI, OpI->getOperand(0));
8403 return commonCastTransforms(FI);
8406 Instruction *InstCombiner::visitUIToFP(CastInst &CI) {
8407 return commonCastTransforms(CI);
8410 Instruction *InstCombiner::visitSIToFP(CastInst &CI) {
8411 return commonCastTransforms(CI);
8414 Instruction *InstCombiner::visitPtrToInt(CastInst &CI) {
8415 return commonPointerCastTransforms(CI);
8418 Instruction *InstCombiner::visitIntToPtr(IntToPtrInst &CI) {
8419 if (Instruction *I = commonCastTransforms(CI))
8422 const Type *DestPointee = cast<PointerType>(CI.getType())->getElementType();
8423 if (!DestPointee->isSized()) return 0;
8425 // If this is inttoptr(add (ptrtoint x), cst), try to turn this into a GEP.
8428 if (match(CI.getOperand(0), m_Add(m_Cast<PtrToIntInst>(m_Value(X)),
8429 m_ConstantInt(Cst)))) {
8430 // If the source and destination operands have the same type, see if this
8431 // is a single-index GEP.
8432 if (X->getType() == CI.getType()) {
8433 // Get the size of the pointee type.
8434 uint64_t Size = TD->getABITypeSize(DestPointee);
8436 // Convert the constant to intptr type.
8437 APInt Offset = Cst->getValue();
8438 Offset.sextOrTrunc(TD->getPointerSizeInBits());
8440 // If Offset is evenly divisible by Size, we can do this xform.
8441 if (Size && !APIntOps::srem(Offset, APInt(Offset.getBitWidth(), Size))){
8442 Offset = APIntOps::sdiv(Offset, APInt(Offset.getBitWidth(), Size));
8443 return GetElementPtrInst::Create(X, ConstantInt::get(Offset));
8446 // TODO: Could handle other cases, e.g. where add is indexing into field of
8448 } else if (CI.getOperand(0)->hasOneUse() &&
8449 match(CI.getOperand(0), m_Add(m_Value(X), m_ConstantInt(Cst)))) {
8450 // Otherwise, if this is inttoptr(add x, cst), try to turn this into an
8451 // "inttoptr+GEP" instead of "add+intptr".
8453 // Get the size of the pointee type.
8454 uint64_t Size = TD->getABITypeSize(DestPointee);
8456 // Convert the constant to intptr type.
8457 APInt Offset = Cst->getValue();
8458 Offset.sextOrTrunc(TD->getPointerSizeInBits());
8460 // If Offset is evenly divisible by Size, we can do this xform.
8461 if (Size && !APIntOps::srem(Offset, APInt(Offset.getBitWidth(), Size))){
8462 Offset = APIntOps::sdiv(Offset, APInt(Offset.getBitWidth(), Size));
8464 Instruction *P = InsertNewInstBefore(new IntToPtrInst(X, CI.getType(),
8466 return GetElementPtrInst::Create(P, ConstantInt::get(Offset), "tmp");
8472 Instruction *InstCombiner::visitBitCast(BitCastInst &CI) {
8473 // If the operands are integer typed then apply the integer transforms,
8474 // otherwise just apply the common ones.
8475 Value *Src = CI.getOperand(0);
8476 const Type *SrcTy = Src->getType();
8477 const Type *DestTy = CI.getType();
8479 if (SrcTy->isInteger() && DestTy->isInteger()) {
8480 if (Instruction *Result = commonIntCastTransforms(CI))
8482 } else if (isa<PointerType>(SrcTy)) {
8483 if (Instruction *I = commonPointerCastTransforms(CI))
8486 if (Instruction *Result = commonCastTransforms(CI))
8491 // Get rid of casts from one type to the same type. These are useless and can
8492 // be replaced by the operand.
8493 if (DestTy == Src->getType())
8494 return ReplaceInstUsesWith(CI, Src);
8496 if (const PointerType *DstPTy = dyn_cast<PointerType>(DestTy)) {
8497 const PointerType *SrcPTy = cast<PointerType>(SrcTy);
8498 const Type *DstElTy = DstPTy->getElementType();
8499 const Type *SrcElTy = SrcPTy->getElementType();
8501 // If the address spaces don't match, don't eliminate the bitcast, which is
8502 // required for changing types.
8503 if (SrcPTy->getAddressSpace() != DstPTy->getAddressSpace())
8506 // If we are casting a malloc or alloca to a pointer to a type of the same
8507 // size, rewrite the allocation instruction to allocate the "right" type.
8508 if (AllocationInst *AI = dyn_cast<AllocationInst>(Src))
8509 if (Instruction *V = PromoteCastOfAllocation(CI, *AI))
8512 // If the source and destination are pointers, and this cast is equivalent
8513 // to a getelementptr X, 0, 0, 0... turn it into the appropriate gep.
8514 // This can enhance SROA and other transforms that want type-safe pointers.
8515 Constant *ZeroUInt = Constant::getNullValue(Type::Int32Ty);
8516 unsigned NumZeros = 0;
8517 while (SrcElTy != DstElTy &&
8518 isa<CompositeType>(SrcElTy) && !isa<PointerType>(SrcElTy) &&
8519 SrcElTy->getNumContainedTypes() /* not "{}" */) {
8520 SrcElTy = cast<CompositeType>(SrcElTy)->getTypeAtIndex(ZeroUInt);
8524 // If we found a path from the src to dest, create the getelementptr now.
8525 if (SrcElTy == DstElTy) {
8526 SmallVector<Value*, 8> Idxs(NumZeros+1, ZeroUInt);
8527 return GetElementPtrInst::Create(Src, Idxs.begin(), Idxs.end(), "",
8528 ((Instruction*) NULL));
8532 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(Src)) {
8533 if (SVI->hasOneUse()) {
8534 // Okay, we have (bitconvert (shuffle ..)). Check to see if this is
8535 // a bitconvert to a vector with the same # elts.
8536 if (isa<VectorType>(DestTy) &&
8537 cast<VectorType>(DestTy)->getNumElements() ==
8538 SVI->getType()->getNumElements() &&
8539 SVI->getType()->getNumElements() ==
8540 cast<VectorType>(SVI->getOperand(0)->getType())->getNumElements()) {
8542 // If either of the operands is a cast from CI.getType(), then
8543 // evaluating the shuffle in the casted destination's type will allow
8544 // us to eliminate at least one cast.
8545 if (((Tmp = dyn_cast<CastInst>(SVI->getOperand(0))) &&
8546 Tmp->getOperand(0)->getType() == DestTy) ||
8547 ((Tmp = dyn_cast<CastInst>(SVI->getOperand(1))) &&
8548 Tmp->getOperand(0)->getType() == DestTy)) {
8549 Value *LHS = InsertCastBefore(Instruction::BitCast,
8550 SVI->getOperand(0), DestTy, CI);
8551 Value *RHS = InsertCastBefore(Instruction::BitCast,
8552 SVI->getOperand(1), DestTy, CI);
8553 // Return a new shuffle vector. Use the same element ID's, as we
8554 // know the vector types match #elts.
8555 return new ShuffleVectorInst(LHS, RHS, SVI->getOperand(2));
8563 /// GetSelectFoldableOperands - We want to turn code that looks like this:
8565 /// %D = select %cond, %C, %A
8567 /// %C = select %cond, %B, 0
8570 /// Assuming that the specified instruction is an operand to the select, return
8571 /// a bitmask indicating which operands of this instruction are foldable if they
8572 /// equal the other incoming value of the select.
8574 static unsigned GetSelectFoldableOperands(Instruction *I) {
8575 switch (I->getOpcode()) {
8576 case Instruction::Add:
8577 case Instruction::Mul:
8578 case Instruction::And:
8579 case Instruction::Or:
8580 case Instruction::Xor:
8581 return 3; // Can fold through either operand.
8582 case Instruction::Sub: // Can only fold on the amount subtracted.
8583 case Instruction::Shl: // Can only fold on the shift amount.
8584 case Instruction::LShr:
8585 case Instruction::AShr:
8588 return 0; // Cannot fold
8592 /// GetSelectFoldableConstant - For the same transformation as the previous
8593 /// function, return the identity constant that goes into the select.
8594 static Constant *GetSelectFoldableConstant(Instruction *I) {
8595 switch (I->getOpcode()) {
8596 default: assert(0 && "This cannot happen!"); abort();
8597 case Instruction::Add:
8598 case Instruction::Sub:
8599 case Instruction::Or:
8600 case Instruction::Xor:
8601 case Instruction::Shl:
8602 case Instruction::LShr:
8603 case Instruction::AShr:
8604 return Constant::getNullValue(I->getType());
8605 case Instruction::And:
8606 return Constant::getAllOnesValue(I->getType());
8607 case Instruction::Mul:
8608 return ConstantInt::get(I->getType(), 1);
8612 /// FoldSelectOpOp - Here we have (select c, TI, FI), and we know that TI and FI
8613 /// have the same opcode and only one use each. Try to simplify this.
8614 Instruction *InstCombiner::FoldSelectOpOp(SelectInst &SI, Instruction *TI,
8616 if (TI->getNumOperands() == 1) {
8617 // If this is a non-volatile load or a cast from the same type,
8620 if (TI->getOperand(0)->getType() != FI->getOperand(0)->getType())
8623 return 0; // unknown unary op.
8626 // Fold this by inserting a select from the input values.
8627 SelectInst *NewSI = SelectInst::Create(SI.getCondition(), TI->getOperand(0),
8628 FI->getOperand(0), SI.getName()+".v");
8629 InsertNewInstBefore(NewSI, SI);
8630 return CastInst::Create(Instruction::CastOps(TI->getOpcode()), NewSI,
8634 // Only handle binary operators here.
8635 if (!isa<BinaryOperator>(TI))
8638 // Figure out if the operations have any operands in common.
8639 Value *MatchOp, *OtherOpT, *OtherOpF;
8641 if (TI->getOperand(0) == FI->getOperand(0)) {
8642 MatchOp = TI->getOperand(0);
8643 OtherOpT = TI->getOperand(1);
8644 OtherOpF = FI->getOperand(1);
8645 MatchIsOpZero = true;
8646 } else if (TI->getOperand(1) == FI->getOperand(1)) {
8647 MatchOp = TI->getOperand(1);
8648 OtherOpT = TI->getOperand(0);
8649 OtherOpF = FI->getOperand(0);
8650 MatchIsOpZero = false;
8651 } else if (!TI->isCommutative()) {
8653 } else if (TI->getOperand(0) == FI->getOperand(1)) {
8654 MatchOp = TI->getOperand(0);
8655 OtherOpT = TI->getOperand(1);
8656 OtherOpF = FI->getOperand(0);
8657 MatchIsOpZero = true;
8658 } else if (TI->getOperand(1) == FI->getOperand(0)) {
8659 MatchOp = TI->getOperand(1);
8660 OtherOpT = TI->getOperand(0);
8661 OtherOpF = FI->getOperand(1);
8662 MatchIsOpZero = true;
8667 // If we reach here, they do have operations in common.
8668 SelectInst *NewSI = SelectInst::Create(SI.getCondition(), OtherOpT,
8669 OtherOpF, SI.getName()+".v");
8670 InsertNewInstBefore(NewSI, SI);
8672 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TI)) {
8674 return BinaryOperator::Create(BO->getOpcode(), MatchOp, NewSI);
8676 return BinaryOperator::Create(BO->getOpcode(), NewSI, MatchOp);
8678 assert(0 && "Shouldn't get here");
8682 /// visitSelectInstWithICmp - Visit a SelectInst that has an
8683 /// ICmpInst as its first operand.
8685 Instruction *InstCombiner::visitSelectInstWithICmp(SelectInst &SI,
8687 bool Changed = false;
8688 ICmpInst::Predicate Pred = ICI->getPredicate();
8689 Value *CmpLHS = ICI->getOperand(0);
8690 Value *CmpRHS = ICI->getOperand(1);
8691 Value *TrueVal = SI.getTrueValue();
8692 Value *FalseVal = SI.getFalseValue();
8694 // Check cases where the comparison is with a constant that
8695 // can be adjusted to fit the min/max idiom. We may edit ICI in
8696 // place here, so make sure the select is the only user.
8697 if (ICI->hasOneUse())
8698 if (ConstantInt *CI = dyn_cast<ConstantInt>(CmpRHS)) {
8701 case ICmpInst::ICMP_ULT:
8702 case ICmpInst::ICMP_SLT: {
8703 // X < MIN ? T : F --> F
8704 if (CI->isMinValue(Pred == ICmpInst::ICMP_SLT))
8705 return ReplaceInstUsesWith(SI, FalseVal);
8706 // X < C ? X : C-1 --> X > C-1 ? C-1 : X
8707 Constant *AdjustedRHS = SubOne(CI);
8708 if ((CmpLHS == TrueVal && AdjustedRHS == FalseVal) ||
8709 (CmpLHS == FalseVal && AdjustedRHS == TrueVal)) {
8710 Pred = ICmpInst::getSwappedPredicate(Pred);
8711 CmpRHS = AdjustedRHS;
8712 std::swap(FalseVal, TrueVal);
8713 ICI->setPredicate(Pred);
8714 ICI->setOperand(1, CmpRHS);
8715 SI.setOperand(1, TrueVal);
8716 SI.setOperand(2, FalseVal);
8721 case ICmpInst::ICMP_UGT:
8722 case ICmpInst::ICMP_SGT: {
8723 // X > MAX ? T : F --> F
8724 if (CI->isMaxValue(Pred == ICmpInst::ICMP_SGT))
8725 return ReplaceInstUsesWith(SI, FalseVal);
8726 // X > C ? X : C+1 --> X < C+1 ? C+1 : X
8727 Constant *AdjustedRHS = AddOne(CI);
8728 if ((CmpLHS == TrueVal && AdjustedRHS == FalseVal) ||
8729 (CmpLHS == FalseVal && AdjustedRHS == TrueVal)) {
8730 Pred = ICmpInst::getSwappedPredicate(Pred);
8731 CmpRHS = AdjustedRHS;
8732 std::swap(FalseVal, TrueVal);
8733 ICI->setPredicate(Pred);
8734 ICI->setOperand(1, CmpRHS);
8735 SI.setOperand(1, TrueVal);
8736 SI.setOperand(2, FalseVal);
8743 // (x <s 0) ? -1 : 0 -> ashr x, 31 -> all ones if signed
8744 // (x >s -1) ? -1 : 0 -> ashr x, 31 -> all ones if not signed
8745 CmpInst::Predicate Pred = CmpInst::BAD_ICMP_PREDICATE;
8746 if (match(TrueVal, m_ConstantInt(-1)) &&
8747 match(FalseVal, m_ConstantInt(0)))
8748 Pred = ICI->getPredicate();
8749 else if (match(TrueVal, m_ConstantInt(0)) &&
8750 match(FalseVal, m_ConstantInt(-1)))
8751 Pred = CmpInst::getInversePredicate(ICI->getPredicate());
8753 if (Pred != CmpInst::BAD_ICMP_PREDICATE) {
8754 // If we are just checking for a icmp eq of a single bit and zext'ing it
8755 // to an integer, then shift the bit to the appropriate place and then
8756 // cast to integer to avoid the comparison.
8757 const APInt &Op1CV = CI->getValue();
8759 // sext (x <s 0) to i32 --> x>>s31 true if signbit set.
8760 // sext (x >s -1) to i32 --> (x>>s31)^-1 true if signbit clear.
8761 if ((Pred == ICmpInst::ICMP_SLT && Op1CV == 0) ||
8762 (Pred == ICmpInst::ICMP_SGT && Op1CV.isAllOnesValue())) {
8763 Value *In = ICI->getOperand(0);
8764 Value *Sh = ConstantInt::get(In->getType(),
8765 In->getType()->getPrimitiveSizeInBits()-1);
8766 In = InsertNewInstBefore(BinaryOperator::CreateAShr(In, Sh,
8767 In->getName()+".lobit"),
8769 if (In->getType() != SI.getType())
8770 In = CastInst::CreateIntegerCast(In, SI.getType(),
8771 true/*SExt*/, "tmp", ICI);
8773 if (Pred == ICmpInst::ICMP_SGT)
8774 In = InsertNewInstBefore(BinaryOperator::CreateNot(In,
8775 In->getName()+".not"), *ICI);
8777 return ReplaceInstUsesWith(SI, In);
8782 if (CmpLHS == TrueVal && CmpRHS == FalseVal) {
8783 // Transform (X == Y) ? X : Y -> Y
8784 if (Pred == ICmpInst::ICMP_EQ)
8785 return ReplaceInstUsesWith(SI, FalseVal);
8786 // Transform (X != Y) ? X : Y -> X
8787 if (Pred == ICmpInst::ICMP_NE)
8788 return ReplaceInstUsesWith(SI, TrueVal);
8789 /// NOTE: if we wanted to, this is where to detect integer MIN/MAX
8791 } else if (CmpLHS == FalseVal && CmpRHS == TrueVal) {
8792 // Transform (X == Y) ? Y : X -> X
8793 if (Pred == ICmpInst::ICMP_EQ)
8794 return ReplaceInstUsesWith(SI, FalseVal);
8795 // Transform (X != Y) ? Y : X -> Y
8796 if (Pred == ICmpInst::ICMP_NE)
8797 return ReplaceInstUsesWith(SI, TrueVal);
8798 /// NOTE: if we wanted to, this is where to detect integer MIN/MAX
8801 /// NOTE: if we wanted to, this is where to detect integer ABS
8803 return Changed ? &SI : 0;
8806 Instruction *InstCombiner::visitSelectInst(SelectInst &SI) {
8807 Value *CondVal = SI.getCondition();
8808 Value *TrueVal = SI.getTrueValue();
8809 Value *FalseVal = SI.getFalseValue();
8811 // select true, X, Y -> X
8812 // select false, X, Y -> Y
8813 if (ConstantInt *C = dyn_cast<ConstantInt>(CondVal))
8814 return ReplaceInstUsesWith(SI, C->getZExtValue() ? TrueVal : FalseVal);
8816 // select C, X, X -> X
8817 if (TrueVal == FalseVal)
8818 return ReplaceInstUsesWith(SI, TrueVal);
8820 if (isa<UndefValue>(TrueVal)) // select C, undef, X -> X
8821 return ReplaceInstUsesWith(SI, FalseVal);
8822 if (isa<UndefValue>(FalseVal)) // select C, X, undef -> X
8823 return ReplaceInstUsesWith(SI, TrueVal);
8824 if (isa<UndefValue>(CondVal)) { // select undef, X, Y -> X or Y
8825 if (isa<Constant>(TrueVal))
8826 return ReplaceInstUsesWith(SI, TrueVal);
8828 return ReplaceInstUsesWith(SI, FalseVal);
8831 if (SI.getType() == Type::Int1Ty) {
8832 if (ConstantInt *C = dyn_cast<ConstantInt>(TrueVal)) {
8833 if (C->getZExtValue()) {
8834 // Change: A = select B, true, C --> A = or B, C
8835 return BinaryOperator::CreateOr(CondVal, FalseVal);
8837 // Change: A = select B, false, C --> A = and !B, C
8839 InsertNewInstBefore(BinaryOperator::CreateNot(CondVal,
8840 "not."+CondVal->getName()), SI);
8841 return BinaryOperator::CreateAnd(NotCond, FalseVal);
8843 } else if (ConstantInt *C = dyn_cast<ConstantInt>(FalseVal)) {
8844 if (C->getZExtValue() == false) {
8845 // Change: A = select B, C, false --> A = and B, C
8846 return BinaryOperator::CreateAnd(CondVal, TrueVal);
8848 // Change: A = select B, C, true --> A = or !B, C
8850 InsertNewInstBefore(BinaryOperator::CreateNot(CondVal,
8851 "not."+CondVal->getName()), SI);
8852 return BinaryOperator::CreateOr(NotCond, TrueVal);
8856 // select a, b, a -> a&b
8857 // select a, a, b -> a|b
8858 if (CondVal == TrueVal)
8859 return BinaryOperator::CreateOr(CondVal, FalseVal);
8860 else if (CondVal == FalseVal)
8861 return BinaryOperator::CreateAnd(CondVal, TrueVal);
8864 // Selecting between two integer constants?
8865 if (ConstantInt *TrueValC = dyn_cast<ConstantInt>(TrueVal))
8866 if (ConstantInt *FalseValC = dyn_cast<ConstantInt>(FalseVal)) {
8867 // select C, 1, 0 -> zext C to int
8868 if (FalseValC->isZero() && TrueValC->getValue() == 1) {
8869 return CastInst::Create(Instruction::ZExt, CondVal, SI.getType());
8870 } else if (TrueValC->isZero() && FalseValC->getValue() == 1) {
8871 // select C, 0, 1 -> zext !C to int
8873 InsertNewInstBefore(BinaryOperator::CreateNot(CondVal,
8874 "not."+CondVal->getName()), SI);
8875 return CastInst::Create(Instruction::ZExt, NotCond, SI.getType());
8878 if (ICmpInst *IC = dyn_cast<ICmpInst>(SI.getCondition())) {
8880 // (x <s 0) ? -1 : 0 -> ashr x, 31
8881 if (TrueValC->isAllOnesValue() && FalseValC->isZero())
8882 if (ConstantInt *CmpCst = dyn_cast<ConstantInt>(IC->getOperand(1))) {
8883 if (IC->getPredicate() == ICmpInst::ICMP_SLT && CmpCst->isZero()) {
8884 // The comparison constant and the result are not neccessarily the
8885 // same width. Make an all-ones value by inserting a AShr.
8886 Value *X = IC->getOperand(0);
8887 uint32_t Bits = X->getType()->getPrimitiveSizeInBits();
8888 Constant *ShAmt = ConstantInt::get(X->getType(), Bits-1);
8889 Instruction *SRA = BinaryOperator::Create(Instruction::AShr, X,
8891 InsertNewInstBefore(SRA, SI);
8893 // Then cast to the appropriate width.
8894 return CastInst::CreateIntegerCast(SRA, SI.getType(), true);
8899 // If one of the constants is zero (we know they can't both be) and we
8900 // have an icmp instruction with zero, and we have an 'and' with the
8901 // non-constant value, eliminate this whole mess. This corresponds to
8902 // cases like this: ((X & 27) ? 27 : 0)
8903 if (TrueValC->isZero() || FalseValC->isZero())
8904 if (IC->isEquality() && isa<ConstantInt>(IC->getOperand(1)) &&
8905 cast<Constant>(IC->getOperand(1))->isNullValue())
8906 if (Instruction *ICA = dyn_cast<Instruction>(IC->getOperand(0)))
8907 if (ICA->getOpcode() == Instruction::And &&
8908 isa<ConstantInt>(ICA->getOperand(1)) &&
8909 (ICA->getOperand(1) == TrueValC ||
8910 ICA->getOperand(1) == FalseValC) &&
8911 isOneBitSet(cast<ConstantInt>(ICA->getOperand(1)))) {
8912 // Okay, now we know that everything is set up, we just don't
8913 // know whether we have a icmp_ne or icmp_eq and whether the
8914 // true or false val is the zero.
8915 bool ShouldNotVal = !TrueValC->isZero();
8916 ShouldNotVal ^= IC->getPredicate() == ICmpInst::ICMP_NE;
8919 V = InsertNewInstBefore(BinaryOperator::Create(
8920 Instruction::Xor, V, ICA->getOperand(1)), SI);
8921 return ReplaceInstUsesWith(SI, V);
8926 // See if we are selecting two values based on a comparison of the two values.
8927 if (FCmpInst *FCI = dyn_cast<FCmpInst>(CondVal)) {
8928 if (FCI->getOperand(0) == TrueVal && FCI->getOperand(1) == FalseVal) {
8929 // Transform (X == Y) ? X : Y -> Y
8930 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ) {
8931 // This is not safe in general for floating point:
8932 // consider X== -0, Y== +0.
8933 // It becomes safe if either operand is a nonzero constant.
8934 ConstantFP *CFPt, *CFPf;
8935 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) &&
8936 !CFPt->getValueAPF().isZero()) ||
8937 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) &&
8938 !CFPf->getValueAPF().isZero()))
8939 return ReplaceInstUsesWith(SI, FalseVal);
8941 // Transform (X != Y) ? X : Y -> X
8942 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
8943 return ReplaceInstUsesWith(SI, TrueVal);
8944 // NOTE: if we wanted to, this is where to detect MIN/MAX
8946 } else if (FCI->getOperand(0) == FalseVal && FCI->getOperand(1) == TrueVal){
8947 // Transform (X == Y) ? Y : X -> X
8948 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ) {
8949 // This is not safe in general for floating point:
8950 // consider X== -0, Y== +0.
8951 // It becomes safe if either operand is a nonzero constant.
8952 ConstantFP *CFPt, *CFPf;
8953 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) &&
8954 !CFPt->getValueAPF().isZero()) ||
8955 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) &&
8956 !CFPf->getValueAPF().isZero()))
8957 return ReplaceInstUsesWith(SI, FalseVal);
8959 // Transform (X != Y) ? Y : X -> Y
8960 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
8961 return ReplaceInstUsesWith(SI, TrueVal);
8962 // NOTE: if we wanted to, this is where to detect MIN/MAX
8964 // NOTE: if we wanted to, this is where to detect ABS
8967 // See if we are selecting two values based on a comparison of the two values.
8968 if (ICmpInst *ICI = dyn_cast<ICmpInst>(CondVal))
8969 if (Instruction *Result = visitSelectInstWithICmp(SI, ICI))
8972 if (Instruction *TI = dyn_cast<Instruction>(TrueVal))
8973 if (Instruction *FI = dyn_cast<Instruction>(FalseVal))
8974 if (TI->hasOneUse() && FI->hasOneUse()) {
8975 Instruction *AddOp = 0, *SubOp = 0;
8977 // Turn (select C, (op X, Y), (op X, Z)) -> (op X, (select C, Y, Z))
8978 if (TI->getOpcode() == FI->getOpcode())
8979 if (Instruction *IV = FoldSelectOpOp(SI, TI, FI))
8982 // Turn select C, (X+Y), (X-Y) --> (X+(select C, Y, (-Y))). This is
8983 // even legal for FP.
8984 if (TI->getOpcode() == Instruction::Sub &&
8985 FI->getOpcode() == Instruction::Add) {
8986 AddOp = FI; SubOp = TI;
8987 } else if (FI->getOpcode() == Instruction::Sub &&
8988 TI->getOpcode() == Instruction::Add) {
8989 AddOp = TI; SubOp = FI;
8993 Value *OtherAddOp = 0;
8994 if (SubOp->getOperand(0) == AddOp->getOperand(0)) {
8995 OtherAddOp = AddOp->getOperand(1);
8996 } else if (SubOp->getOperand(0) == AddOp->getOperand(1)) {
8997 OtherAddOp = AddOp->getOperand(0);
9001 // So at this point we know we have (Y -> OtherAddOp):
9002 // select C, (add X, Y), (sub X, Z)
9003 Value *NegVal; // Compute -Z
9004 if (Constant *C = dyn_cast<Constant>(SubOp->getOperand(1))) {
9005 NegVal = ConstantExpr::getNeg(C);
9007 NegVal = InsertNewInstBefore(
9008 BinaryOperator::CreateNeg(SubOp->getOperand(1), "tmp"), SI);
9011 Value *NewTrueOp = OtherAddOp;
9012 Value *NewFalseOp = NegVal;
9014 std::swap(NewTrueOp, NewFalseOp);
9015 Instruction *NewSel =
9016 SelectInst::Create(CondVal, NewTrueOp,
9017 NewFalseOp, SI.getName() + ".p");
9019 NewSel = InsertNewInstBefore(NewSel, SI);
9020 return BinaryOperator::CreateAdd(SubOp->getOperand(0), NewSel);
9025 // See if we can fold the select into one of our operands.
9026 if (SI.getType()->isInteger()) {
9027 // See the comment above GetSelectFoldableOperands for a description of the
9028 // transformation we are doing here.
9029 if (Instruction *TVI = dyn_cast<Instruction>(TrueVal))
9030 if (TVI->hasOneUse() && TVI->getNumOperands() == 2 &&
9031 !isa<Constant>(FalseVal))
9032 if (unsigned SFO = GetSelectFoldableOperands(TVI)) {
9033 unsigned OpToFold = 0;
9034 if ((SFO & 1) && FalseVal == TVI->getOperand(0)) {
9036 } else if ((SFO & 2) && FalseVal == TVI->getOperand(1)) {
9041 Constant *C = GetSelectFoldableConstant(TVI);
9042 Instruction *NewSel =
9043 SelectInst::Create(SI.getCondition(),
9044 TVI->getOperand(2-OpToFold), C);
9045 InsertNewInstBefore(NewSel, SI);
9046 NewSel->takeName(TVI);
9047 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TVI))
9048 return BinaryOperator::Create(BO->getOpcode(), FalseVal, NewSel);
9050 assert(0 && "Unknown instruction!!");
9055 if (Instruction *FVI = dyn_cast<Instruction>(FalseVal))
9056 if (FVI->hasOneUse() && FVI->getNumOperands() == 2 &&
9057 !isa<Constant>(TrueVal))
9058 if (unsigned SFO = GetSelectFoldableOperands(FVI)) {
9059 unsigned OpToFold = 0;
9060 if ((SFO & 1) && TrueVal == FVI->getOperand(0)) {
9062 } else if ((SFO & 2) && TrueVal == FVI->getOperand(1)) {
9067 Constant *C = GetSelectFoldableConstant(FVI);
9068 Instruction *NewSel =
9069 SelectInst::Create(SI.getCondition(), C,
9070 FVI->getOperand(2-OpToFold));
9071 InsertNewInstBefore(NewSel, SI);
9072 NewSel->takeName(FVI);
9073 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(FVI))
9074 return BinaryOperator::Create(BO->getOpcode(), TrueVal, NewSel);
9076 assert(0 && "Unknown instruction!!");
9081 if (BinaryOperator::isNot(CondVal)) {
9082 SI.setOperand(0, BinaryOperator::getNotArgument(CondVal));
9083 SI.setOperand(1, FalseVal);
9084 SI.setOperand(2, TrueVal);
9091 /// EnforceKnownAlignment - If the specified pointer points to an object that
9092 /// we control, modify the object's alignment to PrefAlign. This isn't
9093 /// often possible though. If alignment is important, a more reliable approach
9094 /// is to simply align all global variables and allocation instructions to
9095 /// their preferred alignment from the beginning.
9097 static unsigned EnforceKnownAlignment(Value *V,
9098 unsigned Align, unsigned PrefAlign) {
9100 User *U = dyn_cast<User>(V);
9101 if (!U) return Align;
9103 switch (getOpcode(U)) {
9105 case Instruction::BitCast:
9106 return EnforceKnownAlignment(U->getOperand(0), Align, PrefAlign);
9107 case Instruction::GetElementPtr: {
9108 // If all indexes are zero, it is just the alignment of the base pointer.
9109 bool AllZeroOperands = true;
9110 for (User::op_iterator i = U->op_begin() + 1, e = U->op_end(); i != e; ++i)
9111 if (!isa<Constant>(*i) ||
9112 !cast<Constant>(*i)->isNullValue()) {
9113 AllZeroOperands = false;
9117 if (AllZeroOperands) {
9118 // Treat this like a bitcast.
9119 return EnforceKnownAlignment(U->getOperand(0), Align, PrefAlign);
9125 if (GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
9126 // If there is a large requested alignment and we can, bump up the alignment
9128 if (!GV->isDeclaration()) {
9129 GV->setAlignment(PrefAlign);
9132 } else if (AllocationInst *AI = dyn_cast<AllocationInst>(V)) {
9133 // If there is a requested alignment and if this is an alloca, round up. We
9134 // don't do this for malloc, because some systems can't respect the request.
9135 if (isa<AllocaInst>(AI)) {
9136 AI->setAlignment(PrefAlign);
9144 /// GetOrEnforceKnownAlignment - If the specified pointer has an alignment that
9145 /// we can determine, return it, otherwise return 0. If PrefAlign is specified,
9146 /// and it is more than the alignment of the ultimate object, see if we can
9147 /// increase the alignment of the ultimate object, making this check succeed.
9148 unsigned InstCombiner::GetOrEnforceKnownAlignment(Value *V,
9149 unsigned PrefAlign) {
9150 unsigned BitWidth = TD ? TD->getTypeSizeInBits(V->getType()) :
9151 sizeof(PrefAlign) * CHAR_BIT;
9152 APInt Mask = APInt::getAllOnesValue(BitWidth);
9153 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
9154 ComputeMaskedBits(V, Mask, KnownZero, KnownOne);
9155 unsigned TrailZ = KnownZero.countTrailingOnes();
9156 unsigned Align = 1u << std::min(BitWidth - 1, TrailZ);
9158 if (PrefAlign > Align)
9159 Align = EnforceKnownAlignment(V, Align, PrefAlign);
9161 // We don't need to make any adjustment.
9165 Instruction *InstCombiner::SimplifyMemTransfer(MemIntrinsic *MI) {
9166 unsigned DstAlign = GetOrEnforceKnownAlignment(MI->getOperand(1));
9167 unsigned SrcAlign = GetOrEnforceKnownAlignment(MI->getOperand(2));
9168 unsigned MinAlign = std::min(DstAlign, SrcAlign);
9169 unsigned CopyAlign = MI->getAlignment()->getZExtValue();
9171 if (CopyAlign < MinAlign) {
9172 MI->setAlignment(ConstantInt::get(Type::Int32Ty, MinAlign));
9176 // If MemCpyInst length is 1/2/4/8 bytes then replace memcpy with
9178 ConstantInt *MemOpLength = dyn_cast<ConstantInt>(MI->getOperand(3));
9179 if (MemOpLength == 0) return 0;
9181 // Source and destination pointer types are always "i8*" for intrinsic. See
9182 // if the size is something we can handle with a single primitive load/store.
9183 // A single load+store correctly handles overlapping memory in the memmove
9185 unsigned Size = MemOpLength->getZExtValue();
9186 if (Size == 0) return MI; // Delete this mem transfer.
9188 if (Size > 8 || (Size&(Size-1)))
9189 return 0; // If not 1/2/4/8 bytes, exit.
9191 // Use an integer load+store unless we can find something better.
9192 Type *NewPtrTy = PointerType::getUnqual(IntegerType::get(Size<<3));
9194 // Memcpy forces the use of i8* for the source and destination. That means
9195 // that if you're using memcpy to move one double around, you'll get a cast
9196 // from double* to i8*. We'd much rather use a double load+store rather than
9197 // an i64 load+store, here because this improves the odds that the source or
9198 // dest address will be promotable. See if we can find a better type than the
9199 // integer datatype.
9200 if (Value *Op = getBitCastOperand(MI->getOperand(1))) {
9201 const Type *SrcETy = cast<PointerType>(Op->getType())->getElementType();
9202 if (SrcETy->isSized() && TD->getTypeStoreSize(SrcETy) == Size) {
9203 // The SrcETy might be something like {{{double}}} or [1 x double]. Rip
9204 // down through these levels if so.
9205 while (!SrcETy->isSingleValueType()) {
9206 if (const StructType *STy = dyn_cast<StructType>(SrcETy)) {
9207 if (STy->getNumElements() == 1)
9208 SrcETy = STy->getElementType(0);
9211 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(SrcETy)) {
9212 if (ATy->getNumElements() == 1)
9213 SrcETy = ATy->getElementType();
9220 if (SrcETy->isSingleValueType())
9221 NewPtrTy = PointerType::getUnqual(SrcETy);
9226 // If the memcpy/memmove provides better alignment info than we can
9228 SrcAlign = std::max(SrcAlign, CopyAlign);
9229 DstAlign = std::max(DstAlign, CopyAlign);
9231 Value *Src = InsertBitCastBefore(MI->getOperand(2), NewPtrTy, *MI);
9232 Value *Dest = InsertBitCastBefore(MI->getOperand(1), NewPtrTy, *MI);
9233 Instruction *L = new LoadInst(Src, "tmp", false, SrcAlign);
9234 InsertNewInstBefore(L, *MI);
9235 InsertNewInstBefore(new StoreInst(L, Dest, false, DstAlign), *MI);
9237 // Set the size of the copy to 0, it will be deleted on the next iteration.
9238 MI->setOperand(3, Constant::getNullValue(MemOpLength->getType()));
9242 Instruction *InstCombiner::SimplifyMemSet(MemSetInst *MI) {
9243 unsigned Alignment = GetOrEnforceKnownAlignment(MI->getDest());
9244 if (MI->getAlignment()->getZExtValue() < Alignment) {
9245 MI->setAlignment(ConstantInt::get(Type::Int32Ty, Alignment));
9249 // Extract the length and alignment and fill if they are constant.
9250 ConstantInt *LenC = dyn_cast<ConstantInt>(MI->getLength());
9251 ConstantInt *FillC = dyn_cast<ConstantInt>(MI->getValue());
9252 if (!LenC || !FillC || FillC->getType() != Type::Int8Ty)
9254 uint64_t Len = LenC->getZExtValue();
9255 Alignment = MI->getAlignment()->getZExtValue();
9257 // If the length is zero, this is a no-op
9258 if (Len == 0) return MI; // memset(d,c,0,a) -> noop
9260 // memset(s,c,n) -> store s, c (for n=1,2,4,8)
9261 if (Len <= 8 && isPowerOf2_32((uint32_t)Len)) {
9262 const Type *ITy = IntegerType::get(Len*8); // n=1 -> i8.
9264 Value *Dest = MI->getDest();
9265 Dest = InsertBitCastBefore(Dest, PointerType::getUnqual(ITy), *MI);
9267 // Alignment 0 is identity for alignment 1 for memset, but not store.
9268 if (Alignment == 0) Alignment = 1;
9270 // Extract the fill value and store.
9271 uint64_t Fill = FillC->getZExtValue()*0x0101010101010101ULL;
9272 InsertNewInstBefore(new StoreInst(ConstantInt::get(ITy, Fill), Dest, false,
9275 // Set the size of the copy to 0, it will be deleted on the next iteration.
9276 MI->setLength(Constant::getNullValue(LenC->getType()));
9284 /// visitCallInst - CallInst simplification. This mostly only handles folding
9285 /// of intrinsic instructions. For normal calls, it allows visitCallSite to do
9286 /// the heavy lifting.
9288 Instruction *InstCombiner::visitCallInst(CallInst &CI) {
9289 IntrinsicInst *II = dyn_cast<IntrinsicInst>(&CI);
9290 if (!II) return visitCallSite(&CI);
9292 // Intrinsics cannot occur in an invoke, so handle them here instead of in
9294 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(II)) {
9295 bool Changed = false;
9297 // memmove/cpy/set of zero bytes is a noop.
9298 if (Constant *NumBytes = dyn_cast<Constant>(MI->getLength())) {
9299 if (NumBytes->isNullValue()) return EraseInstFromFunction(CI);
9301 if (ConstantInt *CI = dyn_cast<ConstantInt>(NumBytes))
9302 if (CI->getZExtValue() == 1) {
9303 // Replace the instruction with just byte operations. We would
9304 // transform other cases to loads/stores, but we don't know if
9305 // alignment is sufficient.
9309 // If we have a memmove and the source operation is a constant global,
9310 // then the source and dest pointers can't alias, so we can change this
9311 // into a call to memcpy.
9312 if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(MI)) {
9313 if (GlobalVariable *GVSrc = dyn_cast<GlobalVariable>(MMI->getSource()))
9314 if (GVSrc->isConstant()) {
9315 Module *M = CI.getParent()->getParent()->getParent();
9316 Intrinsic::ID MemCpyID = Intrinsic::memcpy;
9318 Tys[0] = CI.getOperand(3)->getType();
9320 Intrinsic::getDeclaration(M, MemCpyID, Tys, 1));
9324 // memmove(x,x,size) -> noop.
9325 if (MMI->getSource() == MMI->getDest())
9326 return EraseInstFromFunction(CI);
9329 // If we can determine a pointer alignment that is bigger than currently
9330 // set, update the alignment.
9331 if (isa<MemCpyInst>(MI) || isa<MemMoveInst>(MI)) {
9332 if (Instruction *I = SimplifyMemTransfer(MI))
9334 } else if (MemSetInst *MSI = dyn_cast<MemSetInst>(MI)) {
9335 if (Instruction *I = SimplifyMemSet(MSI))
9339 if (Changed) return II;
9342 switch (II->getIntrinsicID()) {
9344 case Intrinsic::bswap:
9345 // bswap(bswap(x)) -> x
9346 if (IntrinsicInst *Operand = dyn_cast<IntrinsicInst>(II->getOperand(1)))
9347 if (Operand->getIntrinsicID() == Intrinsic::bswap)
9348 return ReplaceInstUsesWith(CI, Operand->getOperand(1));
9350 case Intrinsic::ppc_altivec_lvx:
9351 case Intrinsic::ppc_altivec_lvxl:
9352 case Intrinsic::x86_sse_loadu_ps:
9353 case Intrinsic::x86_sse2_loadu_pd:
9354 case Intrinsic::x86_sse2_loadu_dq:
9355 // Turn PPC lvx -> load if the pointer is known aligned.
9356 // Turn X86 loadups -> load if the pointer is known aligned.
9357 if (GetOrEnforceKnownAlignment(II->getOperand(1), 16) >= 16) {
9358 Value *Ptr = InsertBitCastBefore(II->getOperand(1),
9359 PointerType::getUnqual(II->getType()),
9361 return new LoadInst(Ptr);
9364 case Intrinsic::ppc_altivec_stvx:
9365 case Intrinsic::ppc_altivec_stvxl:
9366 // Turn stvx -> store if the pointer is known aligned.
9367 if (GetOrEnforceKnownAlignment(II->getOperand(2), 16) >= 16) {
9368 const Type *OpPtrTy =
9369 PointerType::getUnqual(II->getOperand(1)->getType());
9370 Value *Ptr = InsertBitCastBefore(II->getOperand(2), OpPtrTy, CI);
9371 return new StoreInst(II->getOperand(1), Ptr);
9374 case Intrinsic::x86_sse_storeu_ps:
9375 case Intrinsic::x86_sse2_storeu_pd:
9376 case Intrinsic::x86_sse2_storeu_dq:
9377 // Turn X86 storeu -> store if the pointer is known aligned.
9378 if (GetOrEnforceKnownAlignment(II->getOperand(1), 16) >= 16) {
9379 const Type *OpPtrTy =
9380 PointerType::getUnqual(II->getOperand(2)->getType());
9381 Value *Ptr = InsertBitCastBefore(II->getOperand(1), OpPtrTy, CI);
9382 return new StoreInst(II->getOperand(2), Ptr);
9386 case Intrinsic::x86_sse_cvttss2si: {
9387 // These intrinsics only demands the 0th element of its input vector. If
9388 // we can simplify the input based on that, do so now.
9390 if (Value *V = SimplifyDemandedVectorElts(II->getOperand(1), 1,
9392 II->setOperand(1, V);
9398 case Intrinsic::ppc_altivec_vperm:
9399 // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant.
9400 if (ConstantVector *Mask = dyn_cast<ConstantVector>(II->getOperand(3))) {
9401 assert(Mask->getNumOperands() == 16 && "Bad type for intrinsic!");
9403 // Check that all of the elements are integer constants or undefs.
9404 bool AllEltsOk = true;
9405 for (unsigned i = 0; i != 16; ++i) {
9406 if (!isa<ConstantInt>(Mask->getOperand(i)) &&
9407 !isa<UndefValue>(Mask->getOperand(i))) {
9414 // Cast the input vectors to byte vectors.
9415 Value *Op0 =InsertBitCastBefore(II->getOperand(1),Mask->getType(),CI);
9416 Value *Op1 =InsertBitCastBefore(II->getOperand(2),Mask->getType(),CI);
9417 Value *Result = UndefValue::get(Op0->getType());
9419 // Only extract each element once.
9420 Value *ExtractedElts[32];
9421 memset(ExtractedElts, 0, sizeof(ExtractedElts));
9423 for (unsigned i = 0; i != 16; ++i) {
9424 if (isa<UndefValue>(Mask->getOperand(i)))
9426 unsigned Idx=cast<ConstantInt>(Mask->getOperand(i))->getZExtValue();
9427 Idx &= 31; // Match the hardware behavior.
9429 if (ExtractedElts[Idx] == 0) {
9431 new ExtractElementInst(Idx < 16 ? Op0 : Op1, Idx&15, "tmp");
9432 InsertNewInstBefore(Elt, CI);
9433 ExtractedElts[Idx] = Elt;
9436 // Insert this value into the result vector.
9437 Result = InsertElementInst::Create(Result, ExtractedElts[Idx],
9439 InsertNewInstBefore(cast<Instruction>(Result), CI);
9441 return CastInst::Create(Instruction::BitCast, Result, CI.getType());
9446 case Intrinsic::stackrestore: {
9447 // If the save is right next to the restore, remove the restore. This can
9448 // happen when variable allocas are DCE'd.
9449 if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getOperand(1))) {
9450 if (SS->getIntrinsicID() == Intrinsic::stacksave) {
9451 BasicBlock::iterator BI = SS;
9453 return EraseInstFromFunction(CI);
9457 // Scan down this block to see if there is another stack restore in the
9458 // same block without an intervening call/alloca.
9459 BasicBlock::iterator BI = II;
9460 TerminatorInst *TI = II->getParent()->getTerminator();
9461 bool CannotRemove = false;
9462 for (++BI; &*BI != TI; ++BI) {
9463 if (isa<AllocaInst>(BI)) {
9464 CannotRemove = true;
9467 if (CallInst *BCI = dyn_cast<CallInst>(BI)) {
9468 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(BCI)) {
9469 // If there is a stackrestore below this one, remove this one.
9470 if (II->getIntrinsicID() == Intrinsic::stackrestore)
9471 return EraseInstFromFunction(CI);
9472 // Otherwise, ignore the intrinsic.
9474 // If we found a non-intrinsic call, we can't remove the stack
9476 CannotRemove = true;
9482 // If the stack restore is in a return/unwind block and if there are no
9483 // allocas or calls between the restore and the return, nuke the restore.
9484 if (!CannotRemove && (isa<ReturnInst>(TI) || isa<UnwindInst>(TI)))
9485 return EraseInstFromFunction(CI);
9490 return visitCallSite(II);
9493 // InvokeInst simplification
9495 Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) {
9496 return visitCallSite(&II);
9499 /// isSafeToEliminateVarargsCast - If this cast does not affect the value
9500 /// passed through the varargs area, we can eliminate the use of the cast.
9501 static bool isSafeToEliminateVarargsCast(const CallSite CS,
9502 const CastInst * const CI,
9503 const TargetData * const TD,
9505 if (!CI->isLosslessCast())
9508 // The size of ByVal arguments is derived from the type, so we
9509 // can't change to a type with a different size. If the size were
9510 // passed explicitly we could avoid this check.
9511 if (!CS.paramHasAttr(ix, Attribute::ByVal))
9515 cast<PointerType>(CI->getOperand(0)->getType())->getElementType();
9516 const Type* DstTy = cast<PointerType>(CI->getType())->getElementType();
9517 if (!SrcTy->isSized() || !DstTy->isSized())
9519 if (TD->getABITypeSize(SrcTy) != TD->getABITypeSize(DstTy))
9524 // visitCallSite - Improvements for call and invoke instructions.
9526 Instruction *InstCombiner::visitCallSite(CallSite CS) {
9527 bool Changed = false;
9529 // If the callee is a constexpr cast of a function, attempt to move the cast
9530 // to the arguments of the call/invoke.
9531 if (transformConstExprCastCall(CS)) return 0;
9533 Value *Callee = CS.getCalledValue();
9535 if (Function *CalleeF = dyn_cast<Function>(Callee))
9536 if (CalleeF->getCallingConv() != CS.getCallingConv()) {
9537 Instruction *OldCall = CS.getInstruction();
9538 // If the call and callee calling conventions don't match, this call must
9539 // be unreachable, as the call is undefined.
9540 new StoreInst(ConstantInt::getTrue(),
9541 UndefValue::get(PointerType::getUnqual(Type::Int1Ty)),
9543 if (!OldCall->use_empty())
9544 OldCall->replaceAllUsesWith(UndefValue::get(OldCall->getType()));
9545 if (isa<CallInst>(OldCall)) // Not worth removing an invoke here.
9546 return EraseInstFromFunction(*OldCall);
9550 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) {
9551 // This instruction is not reachable, just remove it. We insert a store to
9552 // undef so that we know that this code is not reachable, despite the fact
9553 // that we can't modify the CFG here.
9554 new StoreInst(ConstantInt::getTrue(),
9555 UndefValue::get(PointerType::getUnqual(Type::Int1Ty)),
9556 CS.getInstruction());
9558 if (!CS.getInstruction()->use_empty())
9559 CS.getInstruction()->
9560 replaceAllUsesWith(UndefValue::get(CS.getInstruction()->getType()));
9562 if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) {
9563 // Don't break the CFG, insert a dummy cond branch.
9564 BranchInst::Create(II->getNormalDest(), II->getUnwindDest(),
9565 ConstantInt::getTrue(), II);
9567 return EraseInstFromFunction(*CS.getInstruction());
9570 if (BitCastInst *BC = dyn_cast<BitCastInst>(Callee))
9571 if (IntrinsicInst *In = dyn_cast<IntrinsicInst>(BC->getOperand(0)))
9572 if (In->getIntrinsicID() == Intrinsic::init_trampoline)
9573 return transformCallThroughTrampoline(CS);
9575 const PointerType *PTy = cast<PointerType>(Callee->getType());
9576 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
9577 if (FTy->isVarArg()) {
9578 int ix = FTy->getNumParams() + (isa<InvokeInst>(Callee) ? 3 : 1);
9579 // See if we can optimize any arguments passed through the varargs area of
9581 for (CallSite::arg_iterator I = CS.arg_begin()+FTy->getNumParams(),
9582 E = CS.arg_end(); I != E; ++I, ++ix) {
9583 CastInst *CI = dyn_cast<CastInst>(*I);
9584 if (CI && isSafeToEliminateVarargsCast(CS, CI, TD, ix)) {
9585 *I = CI->getOperand(0);
9591 if (isa<InlineAsm>(Callee) && !CS.doesNotThrow()) {
9592 // Inline asm calls cannot throw - mark them 'nounwind'.
9593 CS.setDoesNotThrow();
9597 return Changed ? CS.getInstruction() : 0;
9600 // transformConstExprCastCall - If the callee is a constexpr cast of a function,
9601 // attempt to move the cast to the arguments of the call/invoke.
9603 bool InstCombiner::transformConstExprCastCall(CallSite CS) {
9604 if (!isa<ConstantExpr>(CS.getCalledValue())) return false;
9605 ConstantExpr *CE = cast<ConstantExpr>(CS.getCalledValue());
9606 if (CE->getOpcode() != Instruction::BitCast ||
9607 !isa<Function>(CE->getOperand(0)))
9609 Function *Callee = cast<Function>(CE->getOperand(0));
9610 Instruction *Caller = CS.getInstruction();
9611 const AttrListPtr &CallerPAL = CS.getAttributes();
9613 // Okay, this is a cast from a function to a different type. Unless doing so
9614 // would cause a type conversion of one of our arguments, change this call to
9615 // be a direct call with arguments casted to the appropriate types.
9617 const FunctionType *FT = Callee->getFunctionType();
9618 const Type *OldRetTy = Caller->getType();
9619 const Type *NewRetTy = FT->getReturnType();
9621 if (isa<StructType>(NewRetTy))
9622 return false; // TODO: Handle multiple return values.
9624 // Check to see if we are changing the return type...
9625 if (OldRetTy != NewRetTy) {
9626 if (Callee->isDeclaration() &&
9627 // Conversion is ok if changing from one pointer type to another or from
9628 // a pointer to an integer of the same size.
9629 !((isa<PointerType>(OldRetTy) || OldRetTy == TD->getIntPtrType()) &&
9630 (isa<PointerType>(NewRetTy) || NewRetTy == TD->getIntPtrType())))
9631 return false; // Cannot transform this return value.
9633 if (!Caller->use_empty() &&
9634 // void -> non-void is handled specially
9635 NewRetTy != Type::VoidTy && !CastInst::isCastable(NewRetTy, OldRetTy))
9636 return false; // Cannot transform this return value.
9638 if (!CallerPAL.isEmpty() && !Caller->use_empty()) {
9639 Attributes RAttrs = CallerPAL.getRetAttributes();
9640 if (RAttrs & Attribute::typeIncompatible(NewRetTy))
9641 return false; // Attribute not compatible with transformed value.
9644 // If the callsite is an invoke instruction, and the return value is used by
9645 // a PHI node in a successor, we cannot change the return type of the call
9646 // because there is no place to put the cast instruction (without breaking
9647 // the critical edge). Bail out in this case.
9648 if (!Caller->use_empty())
9649 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller))
9650 for (Value::use_iterator UI = II->use_begin(), E = II->use_end();
9652 if (PHINode *PN = dyn_cast<PHINode>(*UI))
9653 if (PN->getParent() == II->getNormalDest() ||
9654 PN->getParent() == II->getUnwindDest())
9658 unsigned NumActualArgs = unsigned(CS.arg_end()-CS.arg_begin());
9659 unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs);
9661 CallSite::arg_iterator AI = CS.arg_begin();
9662 for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) {
9663 const Type *ParamTy = FT->getParamType(i);
9664 const Type *ActTy = (*AI)->getType();
9666 if (!CastInst::isCastable(ActTy, ParamTy))
9667 return false; // Cannot transform this parameter value.
9669 if (CallerPAL.getParamAttributes(i + 1)
9670 & Attribute::typeIncompatible(ParamTy))
9671 return false; // Attribute not compatible with transformed value.
9673 // Converting from one pointer type to another or between a pointer and an
9674 // integer of the same size is safe even if we do not have a body.
9675 bool isConvertible = ActTy == ParamTy ||
9676 ((isa<PointerType>(ParamTy) || ParamTy == TD->getIntPtrType()) &&
9677 (isa<PointerType>(ActTy) || ActTy == TD->getIntPtrType()));
9678 if (Callee->isDeclaration() && !isConvertible) return false;
9681 if (FT->getNumParams() < NumActualArgs && !FT->isVarArg() &&
9682 Callee->isDeclaration())
9683 return false; // Do not delete arguments unless we have a function body.
9685 if (FT->getNumParams() < NumActualArgs && FT->isVarArg() &&
9686 !CallerPAL.isEmpty())
9687 // In this case we have more arguments than the new function type, but we
9688 // won't be dropping them. Check that these extra arguments have attributes
9689 // that are compatible with being a vararg call argument.
9690 for (unsigned i = CallerPAL.getNumSlots(); i; --i) {
9691 if (CallerPAL.getSlot(i - 1).Index <= FT->getNumParams())
9693 Attributes PAttrs = CallerPAL.getSlot(i - 1).Attrs;
9694 if (PAttrs & Attribute::VarArgsIncompatible)
9698 // Okay, we decided that this is a safe thing to do: go ahead and start
9699 // inserting cast instructions as necessary...
9700 std::vector<Value*> Args;
9701 Args.reserve(NumActualArgs);
9702 SmallVector<AttributeWithIndex, 8> attrVec;
9703 attrVec.reserve(NumCommonArgs);
9705 // Get any return attributes.
9706 Attributes RAttrs = CallerPAL.getRetAttributes();
9708 // If the return value is not being used, the type may not be compatible
9709 // with the existing attributes. Wipe out any problematic attributes.
9710 RAttrs &= ~Attribute::typeIncompatible(NewRetTy);
9712 // Add the new return attributes.
9714 attrVec.push_back(AttributeWithIndex::get(0, RAttrs));
9716 AI = CS.arg_begin();
9717 for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) {
9718 const Type *ParamTy = FT->getParamType(i);
9719 if ((*AI)->getType() == ParamTy) {
9720 Args.push_back(*AI);
9722 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI,
9723 false, ParamTy, false);
9724 CastInst *NewCast = CastInst::Create(opcode, *AI, ParamTy, "tmp");
9725 Args.push_back(InsertNewInstBefore(NewCast, *Caller));
9728 // Add any parameter attributes.
9729 if (Attributes PAttrs = CallerPAL.getParamAttributes(i + 1))
9730 attrVec.push_back(AttributeWithIndex::get(i + 1, PAttrs));
9733 // If the function takes more arguments than the call was taking, add them
9735 for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i)
9736 Args.push_back(Constant::getNullValue(FT->getParamType(i)));
9738 // If we are removing arguments to the function, emit an obnoxious warning...
9739 if (FT->getNumParams() < NumActualArgs) {
9740 if (!FT->isVarArg()) {
9741 cerr << "WARNING: While resolving call to function '"
9742 << Callee->getName() << "' arguments were dropped!\n";
9744 // Add all of the arguments in their promoted form to the arg list...
9745 for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) {
9746 const Type *PTy = getPromotedType((*AI)->getType());
9747 if (PTy != (*AI)->getType()) {
9748 // Must promote to pass through va_arg area!
9749 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI, false,
9751 Instruction *Cast = CastInst::Create(opcode, *AI, PTy, "tmp");
9752 InsertNewInstBefore(Cast, *Caller);
9753 Args.push_back(Cast);
9755 Args.push_back(*AI);
9758 // Add any parameter attributes.
9759 if (Attributes PAttrs = CallerPAL.getParamAttributes(i + 1))
9760 attrVec.push_back(AttributeWithIndex::get(i + 1, PAttrs));
9765 if (Attributes FnAttrs = CallerPAL.getFnAttributes())
9766 attrVec.push_back(AttributeWithIndex::get(~0, FnAttrs));
9768 if (NewRetTy == Type::VoidTy)
9769 Caller->setName(""); // Void type should not have a name.
9771 const AttrListPtr &NewCallerPAL = AttrListPtr::get(attrVec.begin(),attrVec.end());
9774 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
9775 NC = InvokeInst::Create(Callee, II->getNormalDest(), II->getUnwindDest(),
9776 Args.begin(), Args.end(),
9777 Caller->getName(), Caller);
9778 cast<InvokeInst>(NC)->setCallingConv(II->getCallingConv());
9779 cast<InvokeInst>(NC)->setAttributes(NewCallerPAL);
9781 NC = CallInst::Create(Callee, Args.begin(), Args.end(),
9782 Caller->getName(), Caller);
9783 CallInst *CI = cast<CallInst>(Caller);
9784 if (CI->isTailCall())
9785 cast<CallInst>(NC)->setTailCall();
9786 cast<CallInst>(NC)->setCallingConv(CI->getCallingConv());
9787 cast<CallInst>(NC)->setAttributes(NewCallerPAL);
9790 // Insert a cast of the return type as necessary.
9792 if (OldRetTy != NV->getType() && !Caller->use_empty()) {
9793 if (NV->getType() != Type::VoidTy) {
9794 Instruction::CastOps opcode = CastInst::getCastOpcode(NC, false,
9796 NV = NC = CastInst::Create(opcode, NC, OldRetTy, "tmp");
9798 // If this is an invoke instruction, we should insert it after the first
9799 // non-phi, instruction in the normal successor block.
9800 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
9801 BasicBlock::iterator I = II->getNormalDest()->getFirstNonPHI();
9802 InsertNewInstBefore(NC, *I);
9804 // Otherwise, it's a call, just insert cast right after the call instr
9805 InsertNewInstBefore(NC, *Caller);
9807 AddUsersToWorkList(*Caller);
9809 NV = UndefValue::get(Caller->getType());
9813 if (Caller->getType() != Type::VoidTy && !Caller->use_empty())
9814 Caller->replaceAllUsesWith(NV);
9815 Caller->eraseFromParent();
9816 RemoveFromWorkList(Caller);
9820 // transformCallThroughTrampoline - Turn a call to a function created by the
9821 // init_trampoline intrinsic into a direct call to the underlying function.
9823 Instruction *InstCombiner::transformCallThroughTrampoline(CallSite CS) {
9824 Value *Callee = CS.getCalledValue();
9825 const PointerType *PTy = cast<PointerType>(Callee->getType());
9826 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
9827 const AttrListPtr &Attrs = CS.getAttributes();
9829 // If the call already has the 'nest' attribute somewhere then give up -
9830 // otherwise 'nest' would occur twice after splicing in the chain.
9831 if (Attrs.hasAttrSomewhere(Attribute::Nest))
9834 IntrinsicInst *Tramp =
9835 cast<IntrinsicInst>(cast<BitCastInst>(Callee)->getOperand(0));
9837 Function *NestF = cast<Function>(Tramp->getOperand(2)->stripPointerCasts());
9838 const PointerType *NestFPTy = cast<PointerType>(NestF->getType());
9839 const FunctionType *NestFTy = cast<FunctionType>(NestFPTy->getElementType());
9841 const AttrListPtr &NestAttrs = NestF->getAttributes();
9842 if (!NestAttrs.isEmpty()) {
9843 unsigned NestIdx = 1;
9844 const Type *NestTy = 0;
9845 Attributes NestAttr = Attribute::None;
9847 // Look for a parameter marked with the 'nest' attribute.
9848 for (FunctionType::param_iterator I = NestFTy->param_begin(),
9849 E = NestFTy->param_end(); I != E; ++NestIdx, ++I)
9850 if (NestAttrs.paramHasAttr(NestIdx, Attribute::Nest)) {
9851 // Record the parameter type and any other attributes.
9853 NestAttr = NestAttrs.getParamAttributes(NestIdx);
9858 Instruction *Caller = CS.getInstruction();
9859 std::vector<Value*> NewArgs;
9860 NewArgs.reserve(unsigned(CS.arg_end()-CS.arg_begin())+1);
9862 SmallVector<AttributeWithIndex, 8> NewAttrs;
9863 NewAttrs.reserve(Attrs.getNumSlots() + 1);
9865 // Insert the nest argument into the call argument list, which may
9866 // mean appending it. Likewise for attributes.
9868 // Add any result attributes.
9869 if (Attributes Attr = Attrs.getRetAttributes())
9870 NewAttrs.push_back(AttributeWithIndex::get(0, Attr));
9874 CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
9876 if (Idx == NestIdx) {
9877 // Add the chain argument and attributes.
9878 Value *NestVal = Tramp->getOperand(3);
9879 if (NestVal->getType() != NestTy)
9880 NestVal = new BitCastInst(NestVal, NestTy, "nest", Caller);
9881 NewArgs.push_back(NestVal);
9882 NewAttrs.push_back(AttributeWithIndex::get(NestIdx, NestAttr));
9888 // Add the original argument and attributes.
9889 NewArgs.push_back(*I);
9890 if (Attributes Attr = Attrs.getParamAttributes(Idx))
9892 (AttributeWithIndex::get(Idx + (Idx >= NestIdx), Attr));
9898 // Add any function attributes.
9899 if (Attributes Attr = Attrs.getFnAttributes())
9900 NewAttrs.push_back(AttributeWithIndex::get(~0, Attr));
9902 // The trampoline may have been bitcast to a bogus type (FTy).
9903 // Handle this by synthesizing a new function type, equal to FTy
9904 // with the chain parameter inserted.
9906 std::vector<const Type*> NewTypes;
9907 NewTypes.reserve(FTy->getNumParams()+1);
9909 // Insert the chain's type into the list of parameter types, which may
9910 // mean appending it.
9913 FunctionType::param_iterator I = FTy->param_begin(),
9914 E = FTy->param_end();
9918 // Add the chain's type.
9919 NewTypes.push_back(NestTy);
9924 // Add the original type.
9925 NewTypes.push_back(*I);
9931 // Replace the trampoline call with a direct call. Let the generic
9932 // code sort out any function type mismatches.
9933 FunctionType *NewFTy =
9934 FunctionType::get(FTy->getReturnType(), NewTypes, FTy->isVarArg());
9935 Constant *NewCallee = NestF->getType() == PointerType::getUnqual(NewFTy) ?
9936 NestF : ConstantExpr::getBitCast(NestF, PointerType::getUnqual(NewFTy));
9937 const AttrListPtr &NewPAL = AttrListPtr::get(NewAttrs.begin(),NewAttrs.end());
9939 Instruction *NewCaller;
9940 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
9941 NewCaller = InvokeInst::Create(NewCallee,
9942 II->getNormalDest(), II->getUnwindDest(),
9943 NewArgs.begin(), NewArgs.end(),
9944 Caller->getName(), Caller);
9945 cast<InvokeInst>(NewCaller)->setCallingConv(II->getCallingConv());
9946 cast<InvokeInst>(NewCaller)->setAttributes(NewPAL);
9948 NewCaller = CallInst::Create(NewCallee, NewArgs.begin(), NewArgs.end(),
9949 Caller->getName(), Caller);
9950 if (cast<CallInst>(Caller)->isTailCall())
9951 cast<CallInst>(NewCaller)->setTailCall();
9952 cast<CallInst>(NewCaller)->
9953 setCallingConv(cast<CallInst>(Caller)->getCallingConv());
9954 cast<CallInst>(NewCaller)->setAttributes(NewPAL);
9956 if (Caller->getType() != Type::VoidTy && !Caller->use_empty())
9957 Caller->replaceAllUsesWith(NewCaller);
9958 Caller->eraseFromParent();
9959 RemoveFromWorkList(Caller);
9964 // Replace the trampoline call with a direct call. Since there is no 'nest'
9965 // parameter, there is no need to adjust the argument list. Let the generic
9966 // code sort out any function type mismatches.
9967 Constant *NewCallee =
9968 NestF->getType() == PTy ? NestF : ConstantExpr::getBitCast(NestF, PTy);
9969 CS.setCalledFunction(NewCallee);
9970 return CS.getInstruction();
9973 /// FoldPHIArgBinOpIntoPHI - If we have something like phi [add (a,b), add(c,d)]
9974 /// and if a/b/c/d and the add's all have a single use, turn this into two phi's
9975 /// and a single binop.
9976 Instruction *InstCombiner::FoldPHIArgBinOpIntoPHI(PHINode &PN) {
9977 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
9978 assert(isa<BinaryOperator>(FirstInst) || isa<CmpInst>(FirstInst));
9979 unsigned Opc = FirstInst->getOpcode();
9980 Value *LHSVal = FirstInst->getOperand(0);
9981 Value *RHSVal = FirstInst->getOperand(1);
9983 const Type *LHSType = LHSVal->getType();
9984 const Type *RHSType = RHSVal->getType();
9986 // Scan to see if all operands are the same opcode, all have one use, and all
9987 // kill their operands (i.e. the operands have one use).
9988 for (unsigned i = 1; i != PN.getNumIncomingValues(); ++i) {
9989 Instruction *I = dyn_cast<Instruction>(PN.getIncomingValue(i));
9990 if (!I || I->getOpcode() != Opc || !I->hasOneUse() ||
9991 // Verify type of the LHS matches so we don't fold cmp's of different
9992 // types or GEP's with different index types.
9993 I->getOperand(0)->getType() != LHSType ||
9994 I->getOperand(1)->getType() != RHSType)
9997 // If they are CmpInst instructions, check their predicates
9998 if (Opc == Instruction::ICmp || Opc == Instruction::FCmp)
9999 if (cast<CmpInst>(I)->getPredicate() !=
10000 cast<CmpInst>(FirstInst)->getPredicate())
10003 // Keep track of which operand needs a phi node.
10004 if (I->getOperand(0) != LHSVal) LHSVal = 0;
10005 if (I->getOperand(1) != RHSVal) RHSVal = 0;
10008 // Otherwise, this is safe to transform!
10010 Value *InLHS = FirstInst->getOperand(0);
10011 Value *InRHS = FirstInst->getOperand(1);
10012 PHINode *NewLHS = 0, *NewRHS = 0;
10014 NewLHS = PHINode::Create(LHSType,
10015 FirstInst->getOperand(0)->getName() + ".pn");
10016 NewLHS->reserveOperandSpace(PN.getNumOperands()/2);
10017 NewLHS->addIncoming(InLHS, PN.getIncomingBlock(0));
10018 InsertNewInstBefore(NewLHS, PN);
10023 NewRHS = PHINode::Create(RHSType,
10024 FirstInst->getOperand(1)->getName() + ".pn");
10025 NewRHS->reserveOperandSpace(PN.getNumOperands()/2);
10026 NewRHS->addIncoming(InRHS, PN.getIncomingBlock(0));
10027 InsertNewInstBefore(NewRHS, PN);
10031 // Add all operands to the new PHIs.
10032 if (NewLHS || NewRHS) {
10033 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
10034 Instruction *InInst = cast<Instruction>(PN.getIncomingValue(i));
10036 Value *NewInLHS = InInst->getOperand(0);
10037 NewLHS->addIncoming(NewInLHS, PN.getIncomingBlock(i));
10040 Value *NewInRHS = InInst->getOperand(1);
10041 NewRHS->addIncoming(NewInRHS, PN.getIncomingBlock(i));
10046 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
10047 return BinaryOperator::Create(BinOp->getOpcode(), LHSVal, RHSVal);
10048 CmpInst *CIOp = cast<CmpInst>(FirstInst);
10049 return CmpInst::Create(CIOp->getOpcode(), CIOp->getPredicate(), LHSVal,
10053 Instruction *InstCombiner::FoldPHIArgGEPIntoPHI(PHINode &PN) {
10054 GetElementPtrInst *FirstInst =cast<GetElementPtrInst>(PN.getIncomingValue(0));
10056 SmallVector<Value*, 16> FixedOperands(FirstInst->op_begin(),
10057 FirstInst->op_end());
10059 // Scan to see if all operands are the same opcode, all have one use, and all
10060 // kill their operands (i.e. the operands have one use).
10061 for (unsigned i = 1; i != PN.getNumIncomingValues(); ++i) {
10062 GetElementPtrInst *GEP= dyn_cast<GetElementPtrInst>(PN.getIncomingValue(i));
10063 if (!GEP || !GEP->hasOneUse() || GEP->getType() != FirstInst->getType() ||
10064 GEP->getNumOperands() != FirstInst->getNumOperands())
10067 // Compare the operand lists.
10068 for (unsigned op = 0, e = FirstInst->getNumOperands(); op != e; ++op) {
10069 if (FirstInst->getOperand(op) == GEP->getOperand(op))
10072 // Don't merge two GEPs when two operands differ (introducing phi nodes)
10073 // if one of the PHIs has a constant for the index. The index may be
10074 // substantially cheaper to compute for the constants, so making it a
10075 // variable index could pessimize the path. This also handles the case
10076 // for struct indices, which must always be constant.
10077 if (isa<ConstantInt>(FirstInst->getOperand(op)) ||
10078 isa<ConstantInt>(GEP->getOperand(op)))
10081 if (FirstInst->getOperand(op)->getType() !=GEP->getOperand(op)->getType())
10083 FixedOperands[op] = 0; // Needs a PHI.
10087 // Otherwise, this is safe to transform. Insert PHI nodes for each operand
10088 // that is variable.
10089 SmallVector<PHINode*, 16> OperandPhis(FixedOperands.size());
10091 bool HasAnyPHIs = false;
10092 for (unsigned i = 0, e = FixedOperands.size(); i != e; ++i) {
10093 if (FixedOperands[i]) continue; // operand doesn't need a phi.
10094 Value *FirstOp = FirstInst->getOperand(i);
10095 PHINode *NewPN = PHINode::Create(FirstOp->getType(),
10096 FirstOp->getName()+".pn");
10097 InsertNewInstBefore(NewPN, PN);
10099 NewPN->reserveOperandSpace(e);
10100 NewPN->addIncoming(FirstOp, PN.getIncomingBlock(0));
10101 OperandPhis[i] = NewPN;
10102 FixedOperands[i] = NewPN;
10107 // Add all operands to the new PHIs.
10109 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
10110 GetElementPtrInst *InGEP =cast<GetElementPtrInst>(PN.getIncomingValue(i));
10111 BasicBlock *InBB = PN.getIncomingBlock(i);
10113 for (unsigned op = 0, e = OperandPhis.size(); op != e; ++op)
10114 if (PHINode *OpPhi = OperandPhis[op])
10115 OpPhi->addIncoming(InGEP->getOperand(op), InBB);
10119 Value *Base = FixedOperands[0];
10120 return GetElementPtrInst::Create(Base, FixedOperands.begin()+1,
10121 FixedOperands.end());
10125 /// isSafeToSinkLoad - Return true if we know that it is safe sink the load out
10126 /// of the block that defines it. This means that it must be obvious the value
10127 /// of the load is not changed from the point of the load to the end of the
10128 /// block it is in.
10130 /// Finally, it is safe, but not profitable, to sink a load targetting a
10131 /// non-address-taken alloca. Doing so will cause us to not promote the alloca
10133 static bool isSafeToSinkLoad(LoadInst *L) {
10134 BasicBlock::iterator BBI = L, E = L->getParent()->end();
10136 for (++BBI; BBI != E; ++BBI)
10137 if (BBI->mayWriteToMemory())
10140 // Check for non-address taken alloca. If not address-taken already, it isn't
10141 // profitable to do this xform.
10142 if (AllocaInst *AI = dyn_cast<AllocaInst>(L->getOperand(0))) {
10143 bool isAddressTaken = false;
10144 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
10146 if (isa<LoadInst>(UI)) continue;
10147 if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
10148 // If storing TO the alloca, then the address isn't taken.
10149 if (SI->getOperand(1) == AI) continue;
10151 isAddressTaken = true;
10155 if (!isAddressTaken)
10163 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
10164 // operator and they all are only used by the PHI, PHI together their
10165 // inputs, and do the operation once, to the result of the PHI.
10166 Instruction *InstCombiner::FoldPHIArgOpIntoPHI(PHINode &PN) {
10167 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
10169 // Scan the instruction, looking for input operations that can be folded away.
10170 // If all input operands to the phi are the same instruction (e.g. a cast from
10171 // the same type or "+42") we can pull the operation through the PHI, reducing
10172 // code size and simplifying code.
10173 Constant *ConstantOp = 0;
10174 const Type *CastSrcTy = 0;
10175 bool isVolatile = false;
10176 if (isa<CastInst>(FirstInst)) {
10177 CastSrcTy = FirstInst->getOperand(0)->getType();
10178 } else if (isa<BinaryOperator>(FirstInst) || isa<CmpInst>(FirstInst)) {
10179 // Can fold binop, compare or shift here if the RHS is a constant,
10180 // otherwise call FoldPHIArgBinOpIntoPHI.
10181 ConstantOp = dyn_cast<Constant>(FirstInst->getOperand(1));
10182 if (ConstantOp == 0)
10183 return FoldPHIArgBinOpIntoPHI(PN);
10184 } else if (LoadInst *LI = dyn_cast<LoadInst>(FirstInst)) {
10185 isVolatile = LI->isVolatile();
10186 // We can't sink the load if the loaded value could be modified between the
10187 // load and the PHI.
10188 if (LI->getParent() != PN.getIncomingBlock(0) ||
10189 !isSafeToSinkLoad(LI))
10192 // If the PHI is of volatile loads and the load block has multiple
10193 // successors, sinking it would remove a load of the volatile value from
10194 // the path through the other successor.
10196 LI->getParent()->getTerminator()->getNumSuccessors() != 1)
10199 } else if (isa<GetElementPtrInst>(FirstInst)) {
10200 return FoldPHIArgGEPIntoPHI(PN);
10202 return 0; // Cannot fold this operation.
10205 // Check to see if all arguments are the same operation.
10206 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
10207 if (!isa<Instruction>(PN.getIncomingValue(i))) return 0;
10208 Instruction *I = cast<Instruction>(PN.getIncomingValue(i));
10209 if (!I->hasOneUse() || !I->isSameOperationAs(FirstInst))
10212 if (I->getOperand(0)->getType() != CastSrcTy)
10213 return 0; // Cast operation must match.
10214 } else if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
10215 // We can't sink the load if the loaded value could be modified between
10216 // the load and the PHI.
10217 if (LI->isVolatile() != isVolatile ||
10218 LI->getParent() != PN.getIncomingBlock(i) ||
10219 !isSafeToSinkLoad(LI))
10222 // If the PHI is of volatile loads and the load block has multiple
10223 // successors, sinking it would remove a load of the volatile value from
10224 // the path through the other successor.
10226 LI->getParent()->getTerminator()->getNumSuccessors() != 1)
10230 } else if (I->getOperand(1) != ConstantOp) {
10235 // Okay, they are all the same operation. Create a new PHI node of the
10236 // correct type, and PHI together all of the LHS's of the instructions.
10237 PHINode *NewPN = PHINode::Create(FirstInst->getOperand(0)->getType(),
10238 PN.getName()+".in");
10239 NewPN->reserveOperandSpace(PN.getNumOperands()/2);
10241 Value *InVal = FirstInst->getOperand(0);
10242 NewPN->addIncoming(InVal, PN.getIncomingBlock(0));
10244 // Add all operands to the new PHI.
10245 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
10246 Value *NewInVal = cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
10247 if (NewInVal != InVal)
10249 NewPN->addIncoming(NewInVal, PN.getIncomingBlock(i));
10254 // The new PHI unions all of the same values together. This is really
10255 // common, so we handle it intelligently here for compile-time speed.
10259 InsertNewInstBefore(NewPN, PN);
10263 // Insert and return the new operation.
10264 if (CastInst* FirstCI = dyn_cast<CastInst>(FirstInst))
10265 return CastInst::Create(FirstCI->getOpcode(), PhiVal, PN.getType());
10266 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
10267 return BinaryOperator::Create(BinOp->getOpcode(), PhiVal, ConstantOp);
10268 if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst))
10269 return CmpInst::Create(CIOp->getOpcode(), CIOp->getPredicate(),
10270 PhiVal, ConstantOp);
10271 assert(isa<LoadInst>(FirstInst) && "Unknown operation");
10273 // If this was a volatile load that we are merging, make sure to loop through
10274 // and mark all the input loads as non-volatile. If we don't do this, we will
10275 // insert a new volatile load and the old ones will not be deletable.
10277 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i)
10278 cast<LoadInst>(PN.getIncomingValue(i))->setVolatile(false);
10280 return new LoadInst(PhiVal, "", isVolatile);
10283 /// DeadPHICycle - Return true if this PHI node is only used by a PHI node cycle
10285 static bool DeadPHICycle(PHINode *PN,
10286 SmallPtrSet<PHINode*, 16> &PotentiallyDeadPHIs) {
10287 if (PN->use_empty()) return true;
10288 if (!PN->hasOneUse()) return false;
10290 // Remember this node, and if we find the cycle, return.
10291 if (!PotentiallyDeadPHIs.insert(PN))
10294 // Don't scan crazily complex things.
10295 if (PotentiallyDeadPHIs.size() == 16)
10298 if (PHINode *PU = dyn_cast<PHINode>(PN->use_back()))
10299 return DeadPHICycle(PU, PotentiallyDeadPHIs);
10304 /// PHIsEqualValue - Return true if this phi node is always equal to
10305 /// NonPhiInVal. This happens with mutually cyclic phi nodes like:
10306 /// z = some value; x = phi (y, z); y = phi (x, z)
10307 static bool PHIsEqualValue(PHINode *PN, Value *NonPhiInVal,
10308 SmallPtrSet<PHINode*, 16> &ValueEqualPHIs) {
10309 // See if we already saw this PHI node.
10310 if (!ValueEqualPHIs.insert(PN))
10313 // Don't scan crazily complex things.
10314 if (ValueEqualPHIs.size() == 16)
10317 // Scan the operands to see if they are either phi nodes or are equal to
10319 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
10320 Value *Op = PN->getIncomingValue(i);
10321 if (PHINode *OpPN = dyn_cast<PHINode>(Op)) {
10322 if (!PHIsEqualValue(OpPN, NonPhiInVal, ValueEqualPHIs))
10324 } else if (Op != NonPhiInVal)
10332 // PHINode simplification
10334 Instruction *InstCombiner::visitPHINode(PHINode &PN) {
10335 // If LCSSA is around, don't mess with Phi nodes
10336 if (MustPreserveLCSSA) return 0;
10338 if (Value *V = PN.hasConstantValue())
10339 return ReplaceInstUsesWith(PN, V);
10341 // If all PHI operands are the same operation, pull them through the PHI,
10342 // reducing code size.
10343 if (isa<Instruction>(PN.getIncomingValue(0)) &&
10344 isa<Instruction>(PN.getIncomingValue(1)) &&
10345 cast<Instruction>(PN.getIncomingValue(0))->getOpcode() ==
10346 cast<Instruction>(PN.getIncomingValue(1))->getOpcode() &&
10347 // FIXME: The hasOneUse check will fail for PHIs that use the value more
10348 // than themselves more than once.
10349 PN.getIncomingValue(0)->hasOneUse())
10350 if (Instruction *Result = FoldPHIArgOpIntoPHI(PN))
10353 // If this is a trivial cycle in the PHI node graph, remove it. Basically, if
10354 // this PHI only has a single use (a PHI), and if that PHI only has one use (a
10355 // PHI)... break the cycle.
10356 if (PN.hasOneUse()) {
10357 Instruction *PHIUser = cast<Instruction>(PN.use_back());
10358 if (PHINode *PU = dyn_cast<PHINode>(PHIUser)) {
10359 SmallPtrSet<PHINode*, 16> PotentiallyDeadPHIs;
10360 PotentiallyDeadPHIs.insert(&PN);
10361 if (DeadPHICycle(PU, PotentiallyDeadPHIs))
10362 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
10365 // If this phi has a single use, and if that use just computes a value for
10366 // the next iteration of a loop, delete the phi. This occurs with unused
10367 // induction variables, e.g. "for (int j = 0; ; ++j);". Detecting this
10368 // common case here is good because the only other things that catch this
10369 // are induction variable analysis (sometimes) and ADCE, which is only run
10371 if (PHIUser->hasOneUse() &&
10372 (isa<BinaryOperator>(PHIUser) || isa<GetElementPtrInst>(PHIUser)) &&
10373 PHIUser->use_back() == &PN) {
10374 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
10378 // We sometimes end up with phi cycles that non-obviously end up being the
10379 // same value, for example:
10380 // z = some value; x = phi (y, z); y = phi (x, z)
10381 // where the phi nodes don't necessarily need to be in the same block. Do a
10382 // quick check to see if the PHI node only contains a single non-phi value, if
10383 // so, scan to see if the phi cycle is actually equal to that value.
10385 unsigned InValNo = 0, NumOperandVals = PN.getNumIncomingValues();
10386 // Scan for the first non-phi operand.
10387 while (InValNo != NumOperandVals &&
10388 isa<PHINode>(PN.getIncomingValue(InValNo)))
10391 if (InValNo != NumOperandVals) {
10392 Value *NonPhiInVal = PN.getOperand(InValNo);
10394 // Scan the rest of the operands to see if there are any conflicts, if so
10395 // there is no need to recursively scan other phis.
10396 for (++InValNo; InValNo != NumOperandVals; ++InValNo) {
10397 Value *OpVal = PN.getIncomingValue(InValNo);
10398 if (OpVal != NonPhiInVal && !isa<PHINode>(OpVal))
10402 // If we scanned over all operands, then we have one unique value plus
10403 // phi values. Scan PHI nodes to see if they all merge in each other or
10405 if (InValNo == NumOperandVals) {
10406 SmallPtrSet<PHINode*, 16> ValueEqualPHIs;
10407 if (PHIsEqualValue(&PN, NonPhiInVal, ValueEqualPHIs))
10408 return ReplaceInstUsesWith(PN, NonPhiInVal);
10415 static Value *InsertCastToIntPtrTy(Value *V, const Type *DTy,
10416 Instruction *InsertPoint,
10417 InstCombiner *IC) {
10418 unsigned PtrSize = DTy->getPrimitiveSizeInBits();
10419 unsigned VTySize = V->getType()->getPrimitiveSizeInBits();
10420 // We must cast correctly to the pointer type. Ensure that we
10421 // sign extend the integer value if it is smaller as this is
10422 // used for address computation.
10423 Instruction::CastOps opcode =
10424 (VTySize < PtrSize ? Instruction::SExt :
10425 (VTySize == PtrSize ? Instruction::BitCast : Instruction::Trunc));
10426 return IC->InsertCastBefore(opcode, V, DTy, *InsertPoint);
10430 Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) {
10431 Value *PtrOp = GEP.getOperand(0);
10432 // Is it 'getelementptr %P, i32 0' or 'getelementptr %P'
10433 // If so, eliminate the noop.
10434 if (GEP.getNumOperands() == 1)
10435 return ReplaceInstUsesWith(GEP, PtrOp);
10437 if (isa<UndefValue>(GEP.getOperand(0)))
10438 return ReplaceInstUsesWith(GEP, UndefValue::get(GEP.getType()));
10440 bool HasZeroPointerIndex = false;
10441 if (Constant *C = dyn_cast<Constant>(GEP.getOperand(1)))
10442 HasZeroPointerIndex = C->isNullValue();
10444 if (GEP.getNumOperands() == 2 && HasZeroPointerIndex)
10445 return ReplaceInstUsesWith(GEP, PtrOp);
10447 // Eliminate unneeded casts for indices.
10448 bool MadeChange = false;
10450 gep_type_iterator GTI = gep_type_begin(GEP);
10451 for (User::op_iterator i = GEP.op_begin() + 1, e = GEP.op_end();
10452 i != e; ++i, ++GTI) {
10453 if (isa<SequentialType>(*GTI)) {
10454 if (CastInst *CI = dyn_cast<CastInst>(*i)) {
10455 if (CI->getOpcode() == Instruction::ZExt ||
10456 CI->getOpcode() == Instruction::SExt) {
10457 const Type *SrcTy = CI->getOperand(0)->getType();
10458 // We can eliminate a cast from i32 to i64 iff the target
10459 // is a 32-bit pointer target.
10460 if (SrcTy->getPrimitiveSizeInBits() >= TD->getPointerSizeInBits()) {
10462 *i = CI->getOperand(0);
10466 // If we are using a wider index than needed for this platform, shrink it
10467 // to what we need. If narrower, sign-extend it to what we need.
10468 // If the incoming value needs a cast instruction,
10469 // insert it. This explicit cast can make subsequent optimizations more
10472 if (TD->getTypeSizeInBits(Op->getType()) > TD->getPointerSizeInBits()) {
10473 if (Constant *C = dyn_cast<Constant>(Op)) {
10474 *i = ConstantExpr::getTrunc(C, TD->getIntPtrType());
10477 Op = InsertCastBefore(Instruction::Trunc, Op, TD->getIntPtrType(),
10482 } else if (TD->getTypeSizeInBits(Op->getType()) < TD->getPointerSizeInBits()) {
10483 if (Constant *C = dyn_cast<Constant>(Op)) {
10484 *i = ConstantExpr::getSExt(C, TD->getIntPtrType());
10487 Op = InsertCastBefore(Instruction::SExt, Op, TD->getIntPtrType(),
10495 if (MadeChange) return &GEP;
10497 // If this GEP instruction doesn't move the pointer, and if the input operand
10498 // is a bitcast of another pointer, just replace the GEP with a bitcast of the
10499 // real input to the dest type.
10500 if (GEP.hasAllZeroIndices()) {
10501 if (BitCastInst *BCI = dyn_cast<BitCastInst>(GEP.getOperand(0))) {
10502 // If the bitcast is of an allocation, and the allocation will be
10503 // converted to match the type of the cast, don't touch this.
10504 if (isa<AllocationInst>(BCI->getOperand(0))) {
10505 // See if the bitcast simplifies, if so, don't nuke this GEP yet.
10506 if (Instruction *I = visitBitCast(*BCI)) {
10509 BCI->getParent()->getInstList().insert(BCI, I);
10510 ReplaceInstUsesWith(*BCI, I);
10515 return new BitCastInst(BCI->getOperand(0), GEP.getType());
10519 // Combine Indices - If the source pointer to this getelementptr instruction
10520 // is a getelementptr instruction, combine the indices of the two
10521 // getelementptr instructions into a single instruction.
10523 SmallVector<Value*, 8> SrcGEPOperands;
10524 if (User *Src = dyn_castGetElementPtr(PtrOp))
10525 SrcGEPOperands.append(Src->op_begin(), Src->op_end());
10527 if (!SrcGEPOperands.empty()) {
10528 // Note that if our source is a gep chain itself that we wait for that
10529 // chain to be resolved before we perform this transformation. This
10530 // avoids us creating a TON of code in some cases.
10532 if (isa<GetElementPtrInst>(SrcGEPOperands[0]) &&
10533 cast<Instruction>(SrcGEPOperands[0])->getNumOperands() == 2)
10534 return 0; // Wait until our source is folded to completion.
10536 SmallVector<Value*, 8> Indices;
10538 // Find out whether the last index in the source GEP is a sequential idx.
10539 bool EndsWithSequential = false;
10540 for (gep_type_iterator I = gep_type_begin(*cast<User>(PtrOp)),
10541 E = gep_type_end(*cast<User>(PtrOp)); I != E; ++I)
10542 EndsWithSequential = !isa<StructType>(*I);
10544 // Can we combine the two pointer arithmetics offsets?
10545 if (EndsWithSequential) {
10546 // Replace: gep (gep %P, long B), long A, ...
10547 // With: T = long A+B; gep %P, T, ...
10549 Value *Sum, *SO1 = SrcGEPOperands.back(), *GO1 = GEP.getOperand(1);
10550 if (SO1 == Constant::getNullValue(SO1->getType())) {
10552 } else if (GO1 == Constant::getNullValue(GO1->getType())) {
10555 // If they aren't the same type, convert both to an integer of the
10556 // target's pointer size.
10557 if (SO1->getType() != GO1->getType()) {
10558 if (Constant *SO1C = dyn_cast<Constant>(SO1)) {
10559 SO1 = ConstantExpr::getIntegerCast(SO1C, GO1->getType(), true);
10560 } else if (Constant *GO1C = dyn_cast<Constant>(GO1)) {
10561 GO1 = ConstantExpr::getIntegerCast(GO1C, SO1->getType(), true);
10563 unsigned PS = TD->getPointerSizeInBits();
10564 if (TD->getTypeSizeInBits(SO1->getType()) == PS) {
10565 // Convert GO1 to SO1's type.
10566 GO1 = InsertCastToIntPtrTy(GO1, SO1->getType(), &GEP, this);
10568 } else if (TD->getTypeSizeInBits(GO1->getType()) == PS) {
10569 // Convert SO1 to GO1's type.
10570 SO1 = InsertCastToIntPtrTy(SO1, GO1->getType(), &GEP, this);
10572 const Type *PT = TD->getIntPtrType();
10573 SO1 = InsertCastToIntPtrTy(SO1, PT, &GEP, this);
10574 GO1 = InsertCastToIntPtrTy(GO1, PT, &GEP, this);
10578 if (isa<Constant>(SO1) && isa<Constant>(GO1))
10579 Sum = ConstantExpr::getAdd(cast<Constant>(SO1), cast<Constant>(GO1));
10581 Sum = BinaryOperator::CreateAdd(SO1, GO1, PtrOp->getName()+".sum");
10582 InsertNewInstBefore(cast<Instruction>(Sum), GEP);
10586 // Recycle the GEP we already have if possible.
10587 if (SrcGEPOperands.size() == 2) {
10588 GEP.setOperand(0, SrcGEPOperands[0]);
10589 GEP.setOperand(1, Sum);
10592 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
10593 SrcGEPOperands.end()-1);
10594 Indices.push_back(Sum);
10595 Indices.insert(Indices.end(), GEP.op_begin()+2, GEP.op_end());
10597 } else if (isa<Constant>(*GEP.idx_begin()) &&
10598 cast<Constant>(*GEP.idx_begin())->isNullValue() &&
10599 SrcGEPOperands.size() != 1) {
10600 // Otherwise we can do the fold if the first index of the GEP is a zero
10601 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
10602 SrcGEPOperands.end());
10603 Indices.insert(Indices.end(), GEP.idx_begin()+1, GEP.idx_end());
10606 if (!Indices.empty())
10607 return GetElementPtrInst::Create(SrcGEPOperands[0], Indices.begin(),
10608 Indices.end(), GEP.getName());
10610 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(PtrOp)) {
10611 // GEP of global variable. If all of the indices for this GEP are
10612 // constants, we can promote this to a constexpr instead of an instruction.
10614 // Scan for nonconstants...
10615 SmallVector<Constant*, 8> Indices;
10616 User::op_iterator I = GEP.idx_begin(), E = GEP.idx_end();
10617 for (; I != E && isa<Constant>(*I); ++I)
10618 Indices.push_back(cast<Constant>(*I));
10620 if (I == E) { // If they are all constants...
10621 Constant *CE = ConstantExpr::getGetElementPtr(GV,
10622 &Indices[0],Indices.size());
10624 // Replace all uses of the GEP with the new constexpr...
10625 return ReplaceInstUsesWith(GEP, CE);
10627 } else if (Value *X = getBitCastOperand(PtrOp)) { // Is the operand a cast?
10628 if (!isa<PointerType>(X->getType())) {
10629 // Not interesting. Source pointer must be a cast from pointer.
10630 } else if (HasZeroPointerIndex) {
10631 // transform: GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ...
10632 // into : GEP [10 x i8]* X, i32 0, ...
10634 // This occurs when the program declares an array extern like "int X[];"
10636 const PointerType *CPTy = cast<PointerType>(PtrOp->getType());
10637 const PointerType *XTy = cast<PointerType>(X->getType());
10638 if (const ArrayType *XATy =
10639 dyn_cast<ArrayType>(XTy->getElementType()))
10640 if (const ArrayType *CATy =
10641 dyn_cast<ArrayType>(CPTy->getElementType()))
10642 if (CATy->getElementType() == XATy->getElementType()) {
10643 // At this point, we know that the cast source type is a pointer
10644 // to an array of the same type as the destination pointer
10645 // array. Because the array type is never stepped over (there
10646 // is a leading zero) we can fold the cast into this GEP.
10647 GEP.setOperand(0, X);
10650 } else if (GEP.getNumOperands() == 2) {
10651 // Transform things like:
10652 // %t = getelementptr i32* bitcast ([2 x i32]* %str to i32*), i32 %V
10653 // into: %t1 = getelementptr [2 x i32]* %str, i32 0, i32 %V; bitcast
10654 const Type *SrcElTy = cast<PointerType>(X->getType())->getElementType();
10655 const Type *ResElTy=cast<PointerType>(PtrOp->getType())->getElementType();
10656 if (isa<ArrayType>(SrcElTy) &&
10657 TD->getABITypeSize(cast<ArrayType>(SrcElTy)->getElementType()) ==
10658 TD->getABITypeSize(ResElTy)) {
10660 Idx[0] = Constant::getNullValue(Type::Int32Ty);
10661 Idx[1] = GEP.getOperand(1);
10662 Value *V = InsertNewInstBefore(
10663 GetElementPtrInst::Create(X, Idx, Idx + 2, GEP.getName()), GEP);
10664 // V and GEP are both pointer types --> BitCast
10665 return new BitCastInst(V, GEP.getType());
10668 // Transform things like:
10669 // getelementptr i8* bitcast ([100 x double]* X to i8*), i32 %tmp
10670 // (where tmp = 8*tmp2) into:
10671 // getelementptr [100 x double]* %arr, i32 0, i32 %tmp2; bitcast
10673 if (isa<ArrayType>(SrcElTy) && ResElTy == Type::Int8Ty) {
10674 uint64_t ArrayEltSize =
10675 TD->getABITypeSize(cast<ArrayType>(SrcElTy)->getElementType());
10677 // Check to see if "tmp" is a scale by a multiple of ArrayEltSize. We
10678 // allow either a mul, shift, or constant here.
10680 ConstantInt *Scale = 0;
10681 if (ArrayEltSize == 1) {
10682 NewIdx = GEP.getOperand(1);
10683 Scale = ConstantInt::get(NewIdx->getType(), 1);
10684 } else if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP.getOperand(1))) {
10685 NewIdx = ConstantInt::get(CI->getType(), 1);
10687 } else if (Instruction *Inst =dyn_cast<Instruction>(GEP.getOperand(1))){
10688 if (Inst->getOpcode() == Instruction::Shl &&
10689 isa<ConstantInt>(Inst->getOperand(1))) {
10690 ConstantInt *ShAmt = cast<ConstantInt>(Inst->getOperand(1));
10691 uint32_t ShAmtVal = ShAmt->getLimitedValue(64);
10692 Scale = ConstantInt::get(Inst->getType(), 1ULL << ShAmtVal);
10693 NewIdx = Inst->getOperand(0);
10694 } else if (Inst->getOpcode() == Instruction::Mul &&
10695 isa<ConstantInt>(Inst->getOperand(1))) {
10696 Scale = cast<ConstantInt>(Inst->getOperand(1));
10697 NewIdx = Inst->getOperand(0);
10701 // If the index will be to exactly the right offset with the scale taken
10702 // out, perform the transformation. Note, we don't know whether Scale is
10703 // signed or not. We'll use unsigned version of division/modulo
10704 // operation after making sure Scale doesn't have the sign bit set.
10705 if (Scale && Scale->getSExtValue() >= 0LL &&
10706 Scale->getZExtValue() % ArrayEltSize == 0) {
10707 Scale = ConstantInt::get(Scale->getType(),
10708 Scale->getZExtValue() / ArrayEltSize);
10709 if (Scale->getZExtValue() != 1) {
10710 Constant *C = ConstantExpr::getIntegerCast(Scale, NewIdx->getType(),
10712 Instruction *Sc = BinaryOperator::CreateMul(NewIdx, C, "idxscale");
10713 NewIdx = InsertNewInstBefore(Sc, GEP);
10716 // Insert the new GEP instruction.
10718 Idx[0] = Constant::getNullValue(Type::Int32Ty);
10720 Instruction *NewGEP =
10721 GetElementPtrInst::Create(X, Idx, Idx + 2, GEP.getName());
10722 NewGEP = InsertNewInstBefore(NewGEP, GEP);
10723 // The NewGEP must be pointer typed, so must the old one -> BitCast
10724 return new BitCastInst(NewGEP, GEP.getType());
10733 Instruction *InstCombiner::visitAllocationInst(AllocationInst &AI) {
10734 // Convert: malloc Ty, C - where C is a constant != 1 into: malloc [C x Ty], 1
10735 if (AI.isArrayAllocation()) { // Check C != 1
10736 if (const ConstantInt *C = dyn_cast<ConstantInt>(AI.getArraySize())) {
10737 const Type *NewTy =
10738 ArrayType::get(AI.getAllocatedType(), C->getZExtValue());
10739 AllocationInst *New = 0;
10741 // Create and insert the replacement instruction...
10742 if (isa<MallocInst>(AI))
10743 New = new MallocInst(NewTy, 0, AI.getAlignment(), AI.getName());
10745 assert(isa<AllocaInst>(AI) && "Unknown type of allocation inst!");
10746 New = new AllocaInst(NewTy, 0, AI.getAlignment(), AI.getName());
10749 InsertNewInstBefore(New, AI);
10751 // Scan to the end of the allocation instructions, to skip over a block of
10752 // allocas if possible...
10754 BasicBlock::iterator It = New;
10755 while (isa<AllocationInst>(*It)) ++It;
10757 // Now that I is pointing to the first non-allocation-inst in the block,
10758 // insert our getelementptr instruction...
10760 Value *NullIdx = Constant::getNullValue(Type::Int32Ty);
10764 Value *V = GetElementPtrInst::Create(New, Idx, Idx + 2,
10765 New->getName()+".sub", It);
10767 // Now make everything use the getelementptr instead of the original
10769 return ReplaceInstUsesWith(AI, V);
10770 } else if (isa<UndefValue>(AI.getArraySize())) {
10771 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
10775 // If alloca'ing a zero byte object, replace the alloca with a null pointer.
10776 // Note that we only do this for alloca's, because malloc should allocate and
10777 // return a unique pointer, even for a zero byte allocation.
10778 if (isa<AllocaInst>(AI) && AI.getAllocatedType()->isSized() &&
10779 TD->getABITypeSize(AI.getAllocatedType()) == 0)
10780 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
10785 Instruction *InstCombiner::visitFreeInst(FreeInst &FI) {
10786 Value *Op = FI.getOperand(0);
10788 // free undef -> unreachable.
10789 if (isa<UndefValue>(Op)) {
10790 // Insert a new store to null because we cannot modify the CFG here.
10791 new StoreInst(ConstantInt::getTrue(),
10792 UndefValue::get(PointerType::getUnqual(Type::Int1Ty)), &FI);
10793 return EraseInstFromFunction(FI);
10796 // If we have 'free null' delete the instruction. This can happen in stl code
10797 // when lots of inlining happens.
10798 if (isa<ConstantPointerNull>(Op))
10799 return EraseInstFromFunction(FI);
10801 // Change free <ty>* (cast <ty2>* X to <ty>*) into free <ty2>* X
10802 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op)) {
10803 FI.setOperand(0, CI->getOperand(0));
10807 // Change free (gep X, 0,0,0,0) into free(X)
10808 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op)) {
10809 if (GEPI->hasAllZeroIndices()) {
10810 AddToWorkList(GEPI);
10811 FI.setOperand(0, GEPI->getOperand(0));
10816 // Change free(malloc) into nothing, if the malloc has a single use.
10817 if (MallocInst *MI = dyn_cast<MallocInst>(Op))
10818 if (MI->hasOneUse()) {
10819 EraseInstFromFunction(FI);
10820 return EraseInstFromFunction(*MI);
10827 /// InstCombineLoadCast - Fold 'load (cast P)' -> cast (load P)' when possible.
10828 static Instruction *InstCombineLoadCast(InstCombiner &IC, LoadInst &LI,
10829 const TargetData *TD) {
10830 User *CI = cast<User>(LI.getOperand(0));
10831 Value *CastOp = CI->getOperand(0);
10833 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(CI)) {
10834 // Instead of loading constant c string, use corresponding integer value
10835 // directly if string length is small enough.
10837 if (GetConstantStringInfo(CE->getOperand(0), Str) && !Str.empty()) {
10838 unsigned len = Str.length();
10839 const Type *Ty = cast<PointerType>(CE->getType())->getElementType();
10840 unsigned numBits = Ty->getPrimitiveSizeInBits();
10841 // Replace LI with immediate integer store.
10842 if ((numBits >> 3) == len + 1) {
10843 APInt StrVal(numBits, 0);
10844 APInt SingleChar(numBits, 0);
10845 if (TD->isLittleEndian()) {
10846 for (signed i = len-1; i >= 0; i--) {
10847 SingleChar = (uint64_t) Str[i];
10848 StrVal = (StrVal << 8) | SingleChar;
10851 for (unsigned i = 0; i < len; i++) {
10852 SingleChar = (uint64_t) Str[i];
10853 StrVal = (StrVal << 8) | SingleChar;
10855 // Append NULL at the end.
10857 StrVal = (StrVal << 8) | SingleChar;
10859 Value *NL = ConstantInt::get(StrVal);
10860 return IC.ReplaceInstUsesWith(LI, NL);
10865 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
10866 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
10867 const Type *SrcPTy = SrcTy->getElementType();
10869 if (DestPTy->isInteger() || isa<PointerType>(DestPTy) ||
10870 isa<VectorType>(DestPTy)) {
10871 // If the source is an array, the code below will not succeed. Check to
10872 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
10874 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
10875 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
10876 if (ASrcTy->getNumElements() != 0) {
10878 Idxs[0] = Idxs[1] = Constant::getNullValue(Type::Int32Ty);
10879 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs, 2);
10880 SrcTy = cast<PointerType>(CastOp->getType());
10881 SrcPTy = SrcTy->getElementType();
10884 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy) ||
10885 isa<VectorType>(SrcPTy)) &&
10886 // Do not allow turning this into a load of an integer, which is then
10887 // casted to a pointer, this pessimizes pointer analysis a lot.
10888 (isa<PointerType>(SrcPTy) == isa<PointerType>(LI.getType())) &&
10889 IC.getTargetData().getTypeSizeInBits(SrcPTy) ==
10890 IC.getTargetData().getTypeSizeInBits(DestPTy)) {
10892 // Okay, we are casting from one integer or pointer type to another of
10893 // the same size. Instead of casting the pointer before the load, cast
10894 // the result of the loaded value.
10895 Value *NewLoad = IC.InsertNewInstBefore(new LoadInst(CastOp,
10897 LI.isVolatile()),LI);
10898 // Now cast the result of the load.
10899 return new BitCastInst(NewLoad, LI.getType());
10906 /// isSafeToLoadUnconditionally - Return true if we know that executing a load
10907 /// from this value cannot trap. If it is not obviously safe to load from the
10908 /// specified pointer, we do a quick local scan of the basic block containing
10909 /// ScanFrom, to determine if the address is already accessed.
10910 static bool isSafeToLoadUnconditionally(Value *V, Instruction *ScanFrom) {
10911 // If it is an alloca it is always safe to load from.
10912 if (isa<AllocaInst>(V)) return true;
10914 // If it is a global variable it is mostly safe to load from.
10915 if (const GlobalValue *GV = dyn_cast<GlobalVariable>(V))
10916 // Don't try to evaluate aliases. External weak GV can be null.
10917 return !isa<GlobalAlias>(GV) && !GV->hasExternalWeakLinkage();
10919 // Otherwise, be a little bit agressive by scanning the local block where we
10920 // want to check to see if the pointer is already being loaded or stored
10921 // from/to. If so, the previous load or store would have already trapped,
10922 // so there is no harm doing an extra load (also, CSE will later eliminate
10923 // the load entirely).
10924 BasicBlock::iterator BBI = ScanFrom, E = ScanFrom->getParent()->begin();
10929 // If we see a free or a call (which might do a free) the pointer could be
10931 if (isa<FreeInst>(BBI) || isa<CallInst>(BBI))
10934 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
10935 if (LI->getOperand(0) == V) return true;
10936 } else if (StoreInst *SI = dyn_cast<StoreInst>(BBI)) {
10937 if (SI->getOperand(1) == V) return true;
10944 Instruction *InstCombiner::visitLoadInst(LoadInst &LI) {
10945 Value *Op = LI.getOperand(0);
10947 // Attempt to improve the alignment.
10948 unsigned KnownAlign = GetOrEnforceKnownAlignment(Op);
10950 (LI.getAlignment() == 0 ? TD->getABITypeAlignment(LI.getType()) :
10951 LI.getAlignment()))
10952 LI.setAlignment(KnownAlign);
10954 // load (cast X) --> cast (load X) iff safe
10955 if (isa<CastInst>(Op))
10956 if (Instruction *Res = InstCombineLoadCast(*this, LI, TD))
10959 // None of the following transforms are legal for volatile loads.
10960 if (LI.isVolatile()) return 0;
10962 // Do really simple store-to-load forwarding and load CSE, to catch cases
10963 // where there are several consequtive memory accesses to the same location,
10964 // separated by a few arithmetic operations.
10965 BasicBlock::iterator BBI = &LI;
10966 if (Value *AvailableVal = FindAvailableLoadedValue(Op, LI.getParent(), BBI,6))
10967 return ReplaceInstUsesWith(LI, AvailableVal);
10969 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op)) {
10970 const Value *GEPI0 = GEPI->getOperand(0);
10971 // TODO: Consider a target hook for valid address spaces for this xform.
10972 if (isa<ConstantPointerNull>(GEPI0) &&
10973 cast<PointerType>(GEPI0->getType())->getAddressSpace() == 0) {
10974 // Insert a new store to null instruction before the load to indicate
10975 // that this code is not reachable. We do this instead of inserting
10976 // an unreachable instruction directly because we cannot modify the
10978 new StoreInst(UndefValue::get(LI.getType()),
10979 Constant::getNullValue(Op->getType()), &LI);
10980 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
10984 if (Constant *C = dyn_cast<Constant>(Op)) {
10985 // load null/undef -> undef
10986 // TODO: Consider a target hook for valid address spaces for this xform.
10987 if (isa<UndefValue>(C) || (C->isNullValue() &&
10988 cast<PointerType>(Op->getType())->getAddressSpace() == 0)) {
10989 // Insert a new store to null instruction before the load to indicate that
10990 // this code is not reachable. We do this instead of inserting an
10991 // unreachable instruction directly because we cannot modify the CFG.
10992 new StoreInst(UndefValue::get(LI.getType()),
10993 Constant::getNullValue(Op->getType()), &LI);
10994 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
10997 // Instcombine load (constant global) into the value loaded.
10998 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Op))
10999 if (GV->isConstant() && !GV->isDeclaration())
11000 return ReplaceInstUsesWith(LI, GV->getInitializer());
11002 // Instcombine load (constantexpr_GEP global, 0, ...) into the value loaded.
11003 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Op)) {
11004 if (CE->getOpcode() == Instruction::GetElementPtr) {
11005 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(CE->getOperand(0)))
11006 if (GV->isConstant() && !GV->isDeclaration())
11008 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE))
11009 return ReplaceInstUsesWith(LI, V);
11010 if (CE->getOperand(0)->isNullValue()) {
11011 // Insert a new store to null instruction before the load to indicate
11012 // that this code is not reachable. We do this instead of inserting
11013 // an unreachable instruction directly because we cannot modify the
11015 new StoreInst(UndefValue::get(LI.getType()),
11016 Constant::getNullValue(Op->getType()), &LI);
11017 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
11020 } else if (CE->isCast()) {
11021 if (Instruction *Res = InstCombineLoadCast(*this, LI, TD))
11027 // If this load comes from anywhere in a constant global, and if the global
11028 // is all undef or zero, we know what it loads.
11029 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Op->getUnderlyingObject())){
11030 if (GV->isConstant() && GV->hasInitializer()) {
11031 if (GV->getInitializer()->isNullValue())
11032 return ReplaceInstUsesWith(LI, Constant::getNullValue(LI.getType()));
11033 else if (isa<UndefValue>(GV->getInitializer()))
11034 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
11038 if (Op->hasOneUse()) {
11039 // Change select and PHI nodes to select values instead of addresses: this
11040 // helps alias analysis out a lot, allows many others simplifications, and
11041 // exposes redundancy in the code.
11043 // Note that we cannot do the transformation unless we know that the
11044 // introduced loads cannot trap! Something like this is valid as long as
11045 // the condition is always false: load (select bool %C, int* null, int* %G),
11046 // but it would not be valid if we transformed it to load from null
11047 // unconditionally.
11049 if (SelectInst *SI = dyn_cast<SelectInst>(Op)) {
11050 // load (select (Cond, &V1, &V2)) --> select(Cond, load &V1, load &V2).
11051 if (isSafeToLoadUnconditionally(SI->getOperand(1), SI) &&
11052 isSafeToLoadUnconditionally(SI->getOperand(2), SI)) {
11053 Value *V1 = InsertNewInstBefore(new LoadInst(SI->getOperand(1),
11054 SI->getOperand(1)->getName()+".val"), LI);
11055 Value *V2 = InsertNewInstBefore(new LoadInst(SI->getOperand(2),
11056 SI->getOperand(2)->getName()+".val"), LI);
11057 return SelectInst::Create(SI->getCondition(), V1, V2);
11060 // load (select (cond, null, P)) -> load P
11061 if (Constant *C = dyn_cast<Constant>(SI->getOperand(1)))
11062 if (C->isNullValue()) {
11063 LI.setOperand(0, SI->getOperand(2));
11067 // load (select (cond, P, null)) -> load P
11068 if (Constant *C = dyn_cast<Constant>(SI->getOperand(2)))
11069 if (C->isNullValue()) {
11070 LI.setOperand(0, SI->getOperand(1));
11078 /// InstCombineStoreToCast - Fold store V, (cast P) -> store (cast V), P
11080 static Instruction *InstCombineStoreToCast(InstCombiner &IC, StoreInst &SI) {
11081 User *CI = cast<User>(SI.getOperand(1));
11082 Value *CastOp = CI->getOperand(0);
11084 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
11085 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
11086 const Type *SrcPTy = SrcTy->getElementType();
11088 if (DestPTy->isInteger() || isa<PointerType>(DestPTy)) {
11089 // If the source is an array, the code below will not succeed. Check to
11090 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
11092 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
11093 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
11094 if (ASrcTy->getNumElements() != 0) {
11096 Idxs[0] = Idxs[1] = Constant::getNullValue(Type::Int32Ty);
11097 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs, 2);
11098 SrcTy = cast<PointerType>(CastOp->getType());
11099 SrcPTy = SrcTy->getElementType();
11102 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy)) &&
11103 IC.getTargetData().getTypeSizeInBits(SrcPTy) ==
11104 IC.getTargetData().getTypeSizeInBits(DestPTy)) {
11106 // Okay, we are casting from one integer or pointer type to another of
11107 // the same size. Instead of casting the pointer before
11108 // the store, cast the value to be stored.
11110 Value *SIOp0 = SI.getOperand(0);
11111 Instruction::CastOps opcode = Instruction::BitCast;
11112 const Type* CastSrcTy = SIOp0->getType();
11113 const Type* CastDstTy = SrcPTy;
11114 if (isa<PointerType>(CastDstTy)) {
11115 if (CastSrcTy->isInteger())
11116 opcode = Instruction::IntToPtr;
11117 } else if (isa<IntegerType>(CastDstTy)) {
11118 if (isa<PointerType>(SIOp0->getType()))
11119 opcode = Instruction::PtrToInt;
11121 if (Constant *C = dyn_cast<Constant>(SIOp0))
11122 NewCast = ConstantExpr::getCast(opcode, C, CastDstTy);
11124 NewCast = IC.InsertNewInstBefore(
11125 CastInst::Create(opcode, SIOp0, CastDstTy, SIOp0->getName()+".c"),
11127 return new StoreInst(NewCast, CastOp);
11134 /// equivalentAddressValues - Test if A and B will obviously have the same
11135 /// value. This includes recognizing that %t0 and %t1 will have the same
11136 /// value in code like this:
11137 /// %t0 = getelementptr @a, 0, 3
11138 /// store i32 0, i32* %t0
11139 /// %t1 = getelementptr @a, 0, 3
11140 /// %t2 = load i32* %t1
11142 static bool equivalentAddressValues(Value *A, Value *B) {
11143 // Test if the values are trivially equivalent.
11144 if (A == B) return true;
11146 // Test if the values come form identical arithmetic instructions.
11147 if (isa<BinaryOperator>(A) ||
11148 isa<CastInst>(A) ||
11150 isa<GetElementPtrInst>(A))
11151 if (Instruction *BI = dyn_cast<Instruction>(B))
11152 if (cast<Instruction>(A)->isIdenticalTo(BI))
11155 // Otherwise they may not be equivalent.
11159 Instruction *InstCombiner::visitStoreInst(StoreInst &SI) {
11160 Value *Val = SI.getOperand(0);
11161 Value *Ptr = SI.getOperand(1);
11163 if (isa<UndefValue>(Ptr)) { // store X, undef -> noop (even if volatile)
11164 EraseInstFromFunction(SI);
11169 // If the RHS is an alloca with a single use, zapify the store, making the
11171 if (Ptr->hasOneUse() && !SI.isVolatile()) {
11172 if (isa<AllocaInst>(Ptr)) {
11173 EraseInstFromFunction(SI);
11178 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr))
11179 if (isa<AllocaInst>(GEP->getOperand(0)) &&
11180 GEP->getOperand(0)->hasOneUse()) {
11181 EraseInstFromFunction(SI);
11187 // Attempt to improve the alignment.
11188 unsigned KnownAlign = GetOrEnforceKnownAlignment(Ptr);
11190 (SI.getAlignment() == 0 ? TD->getABITypeAlignment(Val->getType()) :
11191 SI.getAlignment()))
11192 SI.setAlignment(KnownAlign);
11194 // Do really simple DSE, to catch cases where there are several consequtive
11195 // stores to the same location, separated by a few arithmetic operations. This
11196 // situation often occurs with bitfield accesses.
11197 BasicBlock::iterator BBI = &SI;
11198 for (unsigned ScanInsts = 6; BBI != SI.getParent()->begin() && ScanInsts;
11202 if (StoreInst *PrevSI = dyn_cast<StoreInst>(BBI)) {
11203 // Prev store isn't volatile, and stores to the same location?
11204 if (!PrevSI->isVolatile() &&equivalentAddressValues(PrevSI->getOperand(1),
11205 SI.getOperand(1))) {
11208 EraseInstFromFunction(*PrevSI);
11214 // If this is a load, we have to stop. However, if the loaded value is from
11215 // the pointer we're loading and is producing the pointer we're storing,
11216 // then *this* store is dead (X = load P; store X -> P).
11217 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
11218 if (LI == Val && equivalentAddressValues(LI->getOperand(0), Ptr) &&
11219 !SI.isVolatile()) {
11220 EraseInstFromFunction(SI);
11224 // Otherwise, this is a load from some other location. Stores before it
11225 // may not be dead.
11229 // Don't skip over loads or things that can modify memory.
11230 if (BBI->mayWriteToMemory() || BBI->mayReadFromMemory())
11235 if (SI.isVolatile()) return 0; // Don't hack volatile stores.
11237 // store X, null -> turns into 'unreachable' in SimplifyCFG
11238 if (isa<ConstantPointerNull>(Ptr)) {
11239 if (!isa<UndefValue>(Val)) {
11240 SI.setOperand(0, UndefValue::get(Val->getType()));
11241 if (Instruction *U = dyn_cast<Instruction>(Val))
11242 AddToWorkList(U); // Dropped a use.
11245 return 0; // Do not modify these!
11248 // store undef, Ptr -> noop
11249 if (isa<UndefValue>(Val)) {
11250 EraseInstFromFunction(SI);
11255 // If the pointer destination is a cast, see if we can fold the cast into the
11257 if (isa<CastInst>(Ptr))
11258 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
11260 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
11262 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
11266 // If this store is the last instruction in the basic block, and if the block
11267 // ends with an unconditional branch, try to move it to the successor block.
11269 if (BranchInst *BI = dyn_cast<BranchInst>(BBI))
11270 if (BI->isUnconditional())
11271 if (SimplifyStoreAtEndOfBlock(SI))
11272 return 0; // xform done!
11277 /// SimplifyStoreAtEndOfBlock - Turn things like:
11278 /// if () { *P = v1; } else { *P = v2 }
11279 /// into a phi node with a store in the successor.
11281 /// Simplify things like:
11282 /// *P = v1; if () { *P = v2; }
11283 /// into a phi node with a store in the successor.
11285 bool InstCombiner::SimplifyStoreAtEndOfBlock(StoreInst &SI) {
11286 BasicBlock *StoreBB = SI.getParent();
11288 // Check to see if the successor block has exactly two incoming edges. If
11289 // so, see if the other predecessor contains a store to the same location.
11290 // if so, insert a PHI node (if needed) and move the stores down.
11291 BasicBlock *DestBB = StoreBB->getTerminator()->getSuccessor(0);
11293 // Determine whether Dest has exactly two predecessors and, if so, compute
11294 // the other predecessor.
11295 pred_iterator PI = pred_begin(DestBB);
11296 BasicBlock *OtherBB = 0;
11297 if (*PI != StoreBB)
11300 if (PI == pred_end(DestBB))
11303 if (*PI != StoreBB) {
11308 if (++PI != pred_end(DestBB))
11311 // Bail out if all the relevant blocks aren't distinct (this can happen,
11312 // for example, if SI is in an infinite loop)
11313 if (StoreBB == DestBB || OtherBB == DestBB)
11316 // Verify that the other block ends in a branch and is not otherwise empty.
11317 BasicBlock::iterator BBI = OtherBB->getTerminator();
11318 BranchInst *OtherBr = dyn_cast<BranchInst>(BBI);
11319 if (!OtherBr || BBI == OtherBB->begin())
11322 // If the other block ends in an unconditional branch, check for the 'if then
11323 // else' case. there is an instruction before the branch.
11324 StoreInst *OtherStore = 0;
11325 if (OtherBr->isUnconditional()) {
11326 // If this isn't a store, or isn't a store to the same location, bail out.
11328 OtherStore = dyn_cast<StoreInst>(BBI);
11329 if (!OtherStore || OtherStore->getOperand(1) != SI.getOperand(1))
11332 // Otherwise, the other block ended with a conditional branch. If one of the
11333 // destinations is StoreBB, then we have the if/then case.
11334 if (OtherBr->getSuccessor(0) != StoreBB &&
11335 OtherBr->getSuccessor(1) != StoreBB)
11338 // Okay, we know that OtherBr now goes to Dest and StoreBB, so this is an
11339 // if/then triangle. See if there is a store to the same ptr as SI that
11340 // lives in OtherBB.
11342 // Check to see if we find the matching store.
11343 if ((OtherStore = dyn_cast<StoreInst>(BBI))) {
11344 if (OtherStore->getOperand(1) != SI.getOperand(1))
11348 // If we find something that may be using or overwriting the stored
11349 // value, or if we run out of instructions, we can't do the xform.
11350 if (BBI->mayReadFromMemory() || BBI->mayWriteToMemory() ||
11351 BBI == OtherBB->begin())
11355 // In order to eliminate the store in OtherBr, we have to
11356 // make sure nothing reads or overwrites the stored value in
11358 for (BasicBlock::iterator I = StoreBB->begin(); &*I != &SI; ++I) {
11359 // FIXME: This should really be AA driven.
11360 if (I->mayReadFromMemory() || I->mayWriteToMemory())
11365 // Insert a PHI node now if we need it.
11366 Value *MergedVal = OtherStore->getOperand(0);
11367 if (MergedVal != SI.getOperand(0)) {
11368 PHINode *PN = PHINode::Create(MergedVal->getType(), "storemerge");
11369 PN->reserveOperandSpace(2);
11370 PN->addIncoming(SI.getOperand(0), SI.getParent());
11371 PN->addIncoming(OtherStore->getOperand(0), OtherBB);
11372 MergedVal = InsertNewInstBefore(PN, DestBB->front());
11375 // Advance to a place where it is safe to insert the new store and
11377 BBI = DestBB->getFirstNonPHI();
11378 InsertNewInstBefore(new StoreInst(MergedVal, SI.getOperand(1),
11379 OtherStore->isVolatile()), *BBI);
11381 // Nuke the old stores.
11382 EraseInstFromFunction(SI);
11383 EraseInstFromFunction(*OtherStore);
11389 Instruction *InstCombiner::visitBranchInst(BranchInst &BI) {
11390 // Change br (not X), label True, label False to: br X, label False, True
11392 BasicBlock *TrueDest;
11393 BasicBlock *FalseDest;
11394 if (match(&BI, m_Br(m_Not(m_Value(X)), TrueDest, FalseDest)) &&
11395 !isa<Constant>(X)) {
11396 // Swap Destinations and condition...
11397 BI.setCondition(X);
11398 BI.setSuccessor(0, FalseDest);
11399 BI.setSuccessor(1, TrueDest);
11403 // Cannonicalize fcmp_one -> fcmp_oeq
11404 FCmpInst::Predicate FPred; Value *Y;
11405 if (match(&BI, m_Br(m_FCmp(FPred, m_Value(X), m_Value(Y)),
11406 TrueDest, FalseDest)))
11407 if ((FPred == FCmpInst::FCMP_ONE || FPred == FCmpInst::FCMP_OLE ||
11408 FPred == FCmpInst::FCMP_OGE) && BI.getCondition()->hasOneUse()) {
11409 FCmpInst *I = cast<FCmpInst>(BI.getCondition());
11410 FCmpInst::Predicate NewPred = FCmpInst::getInversePredicate(FPred);
11411 Instruction *NewSCC = new FCmpInst(NewPred, X, Y, "", I);
11412 NewSCC->takeName(I);
11413 // Swap Destinations and condition...
11414 BI.setCondition(NewSCC);
11415 BI.setSuccessor(0, FalseDest);
11416 BI.setSuccessor(1, TrueDest);
11417 RemoveFromWorkList(I);
11418 I->eraseFromParent();
11419 AddToWorkList(NewSCC);
11423 // Cannonicalize icmp_ne -> icmp_eq
11424 ICmpInst::Predicate IPred;
11425 if (match(&BI, m_Br(m_ICmp(IPred, m_Value(X), m_Value(Y)),
11426 TrueDest, FalseDest)))
11427 if ((IPred == ICmpInst::ICMP_NE || IPred == ICmpInst::ICMP_ULE ||
11428 IPred == ICmpInst::ICMP_SLE || IPred == ICmpInst::ICMP_UGE ||
11429 IPred == ICmpInst::ICMP_SGE) && BI.getCondition()->hasOneUse()) {
11430 ICmpInst *I = cast<ICmpInst>(BI.getCondition());
11431 ICmpInst::Predicate NewPred = ICmpInst::getInversePredicate(IPred);
11432 Instruction *NewSCC = new ICmpInst(NewPred, X, Y, "", I);
11433 NewSCC->takeName(I);
11434 // Swap Destinations and condition...
11435 BI.setCondition(NewSCC);
11436 BI.setSuccessor(0, FalseDest);
11437 BI.setSuccessor(1, TrueDest);
11438 RemoveFromWorkList(I);
11439 I->eraseFromParent();;
11440 AddToWorkList(NewSCC);
11447 Instruction *InstCombiner::visitSwitchInst(SwitchInst &SI) {
11448 Value *Cond = SI.getCondition();
11449 if (Instruction *I = dyn_cast<Instruction>(Cond)) {
11450 if (I->getOpcode() == Instruction::Add)
11451 if (ConstantInt *AddRHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
11452 // change 'switch (X+4) case 1:' into 'switch (X) case -3'
11453 for (unsigned i = 2, e = SI.getNumOperands(); i != e; i += 2)
11454 SI.setOperand(i,ConstantExpr::getSub(cast<Constant>(SI.getOperand(i)),
11456 SI.setOperand(0, I->getOperand(0));
11464 Instruction *InstCombiner::visitExtractValueInst(ExtractValueInst &EV) {
11465 Value *Agg = EV.getAggregateOperand();
11467 if (!EV.hasIndices())
11468 return ReplaceInstUsesWith(EV, Agg);
11470 if (Constant *C = dyn_cast<Constant>(Agg)) {
11471 if (isa<UndefValue>(C))
11472 return ReplaceInstUsesWith(EV, UndefValue::get(EV.getType()));
11474 if (isa<ConstantAggregateZero>(C))
11475 return ReplaceInstUsesWith(EV, Constant::getNullValue(EV.getType()));
11477 if (isa<ConstantArray>(C) || isa<ConstantStruct>(C)) {
11478 // Extract the element indexed by the first index out of the constant
11479 Value *V = C->getOperand(*EV.idx_begin());
11480 if (EV.getNumIndices() > 1)
11481 // Extract the remaining indices out of the constant indexed by the
11483 return ExtractValueInst::Create(V, EV.idx_begin() + 1, EV.idx_end());
11485 return ReplaceInstUsesWith(EV, V);
11487 return 0; // Can't handle other constants
11489 if (InsertValueInst *IV = dyn_cast<InsertValueInst>(Agg)) {
11490 // We're extracting from an insertvalue instruction, compare the indices
11491 const unsigned *exti, *exte, *insi, *inse;
11492 for (exti = EV.idx_begin(), insi = IV->idx_begin(),
11493 exte = EV.idx_end(), inse = IV->idx_end();
11494 exti != exte && insi != inse;
11496 if (*insi != *exti)
11497 // The insert and extract both reference distinctly different elements.
11498 // This means the extract is not influenced by the insert, and we can
11499 // replace the aggregate operand of the extract with the aggregate
11500 // operand of the insert. i.e., replace
11501 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1
11502 // %E = extractvalue { i32, { i32 } } %I, 0
11504 // %E = extractvalue { i32, { i32 } } %A, 0
11505 return ExtractValueInst::Create(IV->getAggregateOperand(),
11506 EV.idx_begin(), EV.idx_end());
11508 if (exti == exte && insi == inse)
11509 // Both iterators are at the end: Index lists are identical. Replace
11510 // %B = insertvalue { i32, { i32 } } %A, i32 42, 1, 0
11511 // %C = extractvalue { i32, { i32 } } %B, 1, 0
11513 return ReplaceInstUsesWith(EV, IV->getInsertedValueOperand());
11514 if (exti == exte) {
11515 // The extract list is a prefix of the insert list. i.e. replace
11516 // %I = insertvalue { i32, { i32 } } %A, i32 42, 1, 0
11517 // %E = extractvalue { i32, { i32 } } %I, 1
11519 // %X = extractvalue { i32, { i32 } } %A, 1
11520 // %E = insertvalue { i32 } %X, i32 42, 0
11521 // by switching the order of the insert and extract (though the
11522 // insertvalue should be left in, since it may have other uses).
11523 Value *NewEV = InsertNewInstBefore(
11524 ExtractValueInst::Create(IV->getAggregateOperand(),
11525 EV.idx_begin(), EV.idx_end()),
11527 return InsertValueInst::Create(NewEV, IV->getInsertedValueOperand(),
11531 // The insert list is a prefix of the extract list
11532 // We can simply remove the common indices from the extract and make it
11533 // operate on the inserted value instead of the insertvalue result.
11535 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1
11536 // %E = extractvalue { i32, { i32 } } %I, 1, 0
11538 // %E extractvalue { i32 } { i32 42 }, 0
11539 return ExtractValueInst::Create(IV->getInsertedValueOperand(),
11542 // Can't simplify extracts from other values. Note that nested extracts are
11543 // already simplified implicitely by the above (extract ( extract (insert) )
11544 // will be translated into extract ( insert ( extract ) ) first and then just
11545 // the value inserted, if appropriate).
11549 /// CheapToScalarize - Return true if the value is cheaper to scalarize than it
11550 /// is to leave as a vector operation.
11551 static bool CheapToScalarize(Value *V, bool isConstant) {
11552 if (isa<ConstantAggregateZero>(V))
11554 if (ConstantVector *C = dyn_cast<ConstantVector>(V)) {
11555 if (isConstant) return true;
11556 // If all elts are the same, we can extract.
11557 Constant *Op0 = C->getOperand(0);
11558 for (unsigned i = 1; i < C->getNumOperands(); ++i)
11559 if (C->getOperand(i) != Op0)
11563 Instruction *I = dyn_cast<Instruction>(V);
11564 if (!I) return false;
11566 // Insert element gets simplified to the inserted element or is deleted if
11567 // this is constant idx extract element and its a constant idx insertelt.
11568 if (I->getOpcode() == Instruction::InsertElement && isConstant &&
11569 isa<ConstantInt>(I->getOperand(2)))
11571 if (I->getOpcode() == Instruction::Load && I->hasOneUse())
11573 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I))
11574 if (BO->hasOneUse() &&
11575 (CheapToScalarize(BO->getOperand(0), isConstant) ||
11576 CheapToScalarize(BO->getOperand(1), isConstant)))
11578 if (CmpInst *CI = dyn_cast<CmpInst>(I))
11579 if (CI->hasOneUse() &&
11580 (CheapToScalarize(CI->getOperand(0), isConstant) ||
11581 CheapToScalarize(CI->getOperand(1), isConstant)))
11587 /// Read and decode a shufflevector mask.
11589 /// It turns undef elements into values that are larger than the number of
11590 /// elements in the input.
11591 static std::vector<unsigned> getShuffleMask(const ShuffleVectorInst *SVI) {
11592 unsigned NElts = SVI->getType()->getNumElements();
11593 if (isa<ConstantAggregateZero>(SVI->getOperand(2)))
11594 return std::vector<unsigned>(NElts, 0);
11595 if (isa<UndefValue>(SVI->getOperand(2)))
11596 return std::vector<unsigned>(NElts, 2*NElts);
11598 std::vector<unsigned> Result;
11599 const ConstantVector *CP = cast<ConstantVector>(SVI->getOperand(2));
11600 for (User::const_op_iterator i = CP->op_begin(), e = CP->op_end(); i!=e; ++i)
11601 if (isa<UndefValue>(*i))
11602 Result.push_back(NElts*2); // undef -> 8
11604 Result.push_back(cast<ConstantInt>(*i)->getZExtValue());
11608 /// FindScalarElement - Given a vector and an element number, see if the scalar
11609 /// value is already around as a register, for example if it were inserted then
11610 /// extracted from the vector.
11611 static Value *FindScalarElement(Value *V, unsigned EltNo) {
11612 assert(isa<VectorType>(V->getType()) && "Not looking at a vector?");
11613 const VectorType *PTy = cast<VectorType>(V->getType());
11614 unsigned Width = PTy->getNumElements();
11615 if (EltNo >= Width) // Out of range access.
11616 return UndefValue::get(PTy->getElementType());
11618 if (isa<UndefValue>(V))
11619 return UndefValue::get(PTy->getElementType());
11620 else if (isa<ConstantAggregateZero>(V))
11621 return Constant::getNullValue(PTy->getElementType());
11622 else if (ConstantVector *CP = dyn_cast<ConstantVector>(V))
11623 return CP->getOperand(EltNo);
11624 else if (InsertElementInst *III = dyn_cast<InsertElementInst>(V)) {
11625 // If this is an insert to a variable element, we don't know what it is.
11626 if (!isa<ConstantInt>(III->getOperand(2)))
11628 unsigned IIElt = cast<ConstantInt>(III->getOperand(2))->getZExtValue();
11630 // If this is an insert to the element we are looking for, return the
11632 if (EltNo == IIElt)
11633 return III->getOperand(1);
11635 // Otherwise, the insertelement doesn't modify the value, recurse on its
11637 return FindScalarElement(III->getOperand(0), EltNo);
11638 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(V)) {
11639 unsigned LHSWidth =
11640 cast<VectorType>(SVI->getOperand(0)->getType())->getNumElements();
11641 unsigned InEl = getShuffleMask(SVI)[EltNo];
11642 if (InEl < LHSWidth)
11643 return FindScalarElement(SVI->getOperand(0), InEl);
11644 else if (InEl < LHSWidth*2)
11645 return FindScalarElement(SVI->getOperand(1), InEl - LHSWidth);
11647 return UndefValue::get(PTy->getElementType());
11650 // Otherwise, we don't know.
11654 Instruction *InstCombiner::visitExtractElementInst(ExtractElementInst &EI) {
11655 // If vector val is undef, replace extract with scalar undef.
11656 if (isa<UndefValue>(EI.getOperand(0)))
11657 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
11659 // If vector val is constant 0, replace extract with scalar 0.
11660 if (isa<ConstantAggregateZero>(EI.getOperand(0)))
11661 return ReplaceInstUsesWith(EI, Constant::getNullValue(EI.getType()));
11663 if (ConstantVector *C = dyn_cast<ConstantVector>(EI.getOperand(0))) {
11664 // If vector val is constant with all elements the same, replace EI with
11665 // that element. When the elements are not identical, we cannot replace yet
11666 // (we do that below, but only when the index is constant).
11667 Constant *op0 = C->getOperand(0);
11668 for (unsigned i = 1; i < C->getNumOperands(); ++i)
11669 if (C->getOperand(i) != op0) {
11674 return ReplaceInstUsesWith(EI, op0);
11677 // If extracting a specified index from the vector, see if we can recursively
11678 // find a previously computed scalar that was inserted into the vector.
11679 if (ConstantInt *IdxC = dyn_cast<ConstantInt>(EI.getOperand(1))) {
11680 unsigned IndexVal = IdxC->getZExtValue();
11681 unsigned VectorWidth =
11682 cast<VectorType>(EI.getOperand(0)->getType())->getNumElements();
11684 // If this is extracting an invalid index, turn this into undef, to avoid
11685 // crashing the code below.
11686 if (IndexVal >= VectorWidth)
11687 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
11689 // This instruction only demands the single element from the input vector.
11690 // If the input vector has a single use, simplify it based on this use
11692 if (EI.getOperand(0)->hasOneUse() && VectorWidth != 1) {
11693 uint64_t UndefElts;
11694 if (Value *V = SimplifyDemandedVectorElts(EI.getOperand(0),
11697 EI.setOperand(0, V);
11702 if (Value *Elt = FindScalarElement(EI.getOperand(0), IndexVal))
11703 return ReplaceInstUsesWith(EI, Elt);
11705 // If the this extractelement is directly using a bitcast from a vector of
11706 // the same number of elements, see if we can find the source element from
11707 // it. In this case, we will end up needing to bitcast the scalars.
11708 if (BitCastInst *BCI = dyn_cast<BitCastInst>(EI.getOperand(0))) {
11709 if (const VectorType *VT =
11710 dyn_cast<VectorType>(BCI->getOperand(0)->getType()))
11711 if (VT->getNumElements() == VectorWidth)
11712 if (Value *Elt = FindScalarElement(BCI->getOperand(0), IndexVal))
11713 return new BitCastInst(Elt, EI.getType());
11717 if (Instruction *I = dyn_cast<Instruction>(EI.getOperand(0))) {
11718 if (I->hasOneUse()) {
11719 // Push extractelement into predecessor operation if legal and
11720 // profitable to do so
11721 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
11722 bool isConstantElt = isa<ConstantInt>(EI.getOperand(1));
11723 if (CheapToScalarize(BO, isConstantElt)) {
11724 ExtractElementInst *newEI0 =
11725 new ExtractElementInst(BO->getOperand(0), EI.getOperand(1),
11726 EI.getName()+".lhs");
11727 ExtractElementInst *newEI1 =
11728 new ExtractElementInst(BO->getOperand(1), EI.getOperand(1),
11729 EI.getName()+".rhs");
11730 InsertNewInstBefore(newEI0, EI);
11731 InsertNewInstBefore(newEI1, EI);
11732 return BinaryOperator::Create(BO->getOpcode(), newEI0, newEI1);
11734 } else if (isa<LoadInst>(I)) {
11736 cast<PointerType>(I->getOperand(0)->getType())->getAddressSpace();
11737 Value *Ptr = InsertBitCastBefore(I->getOperand(0),
11738 PointerType::get(EI.getType(), AS),EI);
11739 GetElementPtrInst *GEP =
11740 GetElementPtrInst::Create(Ptr, EI.getOperand(1), I->getName()+".gep");
11741 InsertNewInstBefore(GEP, EI);
11742 return new LoadInst(GEP);
11745 if (InsertElementInst *IE = dyn_cast<InsertElementInst>(I)) {
11746 // Extracting the inserted element?
11747 if (IE->getOperand(2) == EI.getOperand(1))
11748 return ReplaceInstUsesWith(EI, IE->getOperand(1));
11749 // If the inserted and extracted elements are constants, they must not
11750 // be the same value, extract from the pre-inserted value instead.
11751 if (isa<Constant>(IE->getOperand(2)) &&
11752 isa<Constant>(EI.getOperand(1))) {
11753 AddUsesToWorkList(EI);
11754 EI.setOperand(0, IE->getOperand(0));
11757 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I)) {
11758 // If this is extracting an element from a shufflevector, figure out where
11759 // it came from and extract from the appropriate input element instead.
11760 if (ConstantInt *Elt = dyn_cast<ConstantInt>(EI.getOperand(1))) {
11761 unsigned SrcIdx = getShuffleMask(SVI)[Elt->getZExtValue()];
11763 unsigned LHSWidth =
11764 cast<VectorType>(SVI->getOperand(0)->getType())->getNumElements();
11766 if (SrcIdx < LHSWidth)
11767 Src = SVI->getOperand(0);
11768 else if (SrcIdx < LHSWidth*2) {
11769 SrcIdx -= LHSWidth;
11770 Src = SVI->getOperand(1);
11772 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
11774 return new ExtractElementInst(Src, SrcIdx);
11781 /// CollectSingleShuffleElements - If V is a shuffle of values that ONLY returns
11782 /// elements from either LHS or RHS, return the shuffle mask and true.
11783 /// Otherwise, return false.
11784 static bool CollectSingleShuffleElements(Value *V, Value *LHS, Value *RHS,
11785 std::vector<Constant*> &Mask) {
11786 assert(V->getType() == LHS->getType() && V->getType() == RHS->getType() &&
11787 "Invalid CollectSingleShuffleElements");
11788 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
11790 if (isa<UndefValue>(V)) {
11791 Mask.assign(NumElts, UndefValue::get(Type::Int32Ty));
11793 } else if (V == LHS) {
11794 for (unsigned i = 0; i != NumElts; ++i)
11795 Mask.push_back(ConstantInt::get(Type::Int32Ty, i));
11797 } else if (V == RHS) {
11798 for (unsigned i = 0; i != NumElts; ++i)
11799 Mask.push_back(ConstantInt::get(Type::Int32Ty, i+NumElts));
11801 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
11802 // If this is an insert of an extract from some other vector, include it.
11803 Value *VecOp = IEI->getOperand(0);
11804 Value *ScalarOp = IEI->getOperand(1);
11805 Value *IdxOp = IEI->getOperand(2);
11807 if (!isa<ConstantInt>(IdxOp))
11809 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
11811 if (isa<UndefValue>(ScalarOp)) { // inserting undef into vector.
11812 // Okay, we can handle this if the vector we are insertinting into is
11813 // transitively ok.
11814 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
11815 // If so, update the mask to reflect the inserted undef.
11816 Mask[InsertedIdx] = UndefValue::get(Type::Int32Ty);
11819 } else if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)){
11820 if (isa<ConstantInt>(EI->getOperand(1)) &&
11821 EI->getOperand(0)->getType() == V->getType()) {
11822 unsigned ExtractedIdx =
11823 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
11825 // This must be extracting from either LHS or RHS.
11826 if (EI->getOperand(0) == LHS || EI->getOperand(0) == RHS) {
11827 // Okay, we can handle this if the vector we are insertinting into is
11828 // transitively ok.
11829 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
11830 // If so, update the mask to reflect the inserted value.
11831 if (EI->getOperand(0) == LHS) {
11832 Mask[InsertedIdx % NumElts] =
11833 ConstantInt::get(Type::Int32Ty, ExtractedIdx);
11835 assert(EI->getOperand(0) == RHS);
11836 Mask[InsertedIdx % NumElts] =
11837 ConstantInt::get(Type::Int32Ty, ExtractedIdx+NumElts);
11846 // TODO: Handle shufflevector here!
11851 /// CollectShuffleElements - We are building a shuffle of V, using RHS as the
11852 /// RHS of the shuffle instruction, if it is not null. Return a shuffle mask
11853 /// that computes V and the LHS value of the shuffle.
11854 static Value *CollectShuffleElements(Value *V, std::vector<Constant*> &Mask,
11856 assert(isa<VectorType>(V->getType()) &&
11857 (RHS == 0 || V->getType() == RHS->getType()) &&
11858 "Invalid shuffle!");
11859 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
11861 if (isa<UndefValue>(V)) {
11862 Mask.assign(NumElts, UndefValue::get(Type::Int32Ty));
11864 } else if (isa<ConstantAggregateZero>(V)) {
11865 Mask.assign(NumElts, ConstantInt::get(Type::Int32Ty, 0));
11867 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
11868 // If this is an insert of an extract from some other vector, include it.
11869 Value *VecOp = IEI->getOperand(0);
11870 Value *ScalarOp = IEI->getOperand(1);
11871 Value *IdxOp = IEI->getOperand(2);
11873 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
11874 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
11875 EI->getOperand(0)->getType() == V->getType()) {
11876 unsigned ExtractedIdx =
11877 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
11878 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
11880 // Either the extracted from or inserted into vector must be RHSVec,
11881 // otherwise we'd end up with a shuffle of three inputs.
11882 if (EI->getOperand(0) == RHS || RHS == 0) {
11883 RHS = EI->getOperand(0);
11884 Value *V = CollectShuffleElements(VecOp, Mask, RHS);
11885 Mask[InsertedIdx % NumElts] =
11886 ConstantInt::get(Type::Int32Ty, NumElts+ExtractedIdx);
11890 if (VecOp == RHS) {
11891 Value *V = CollectShuffleElements(EI->getOperand(0), Mask, RHS);
11892 // Everything but the extracted element is replaced with the RHS.
11893 for (unsigned i = 0; i != NumElts; ++i) {
11894 if (i != InsertedIdx)
11895 Mask[i] = ConstantInt::get(Type::Int32Ty, NumElts+i);
11900 // If this insertelement is a chain that comes from exactly these two
11901 // vectors, return the vector and the effective shuffle.
11902 if (CollectSingleShuffleElements(IEI, EI->getOperand(0), RHS, Mask))
11903 return EI->getOperand(0);
11908 // TODO: Handle shufflevector here!
11910 // Otherwise, can't do anything fancy. Return an identity vector.
11911 for (unsigned i = 0; i != NumElts; ++i)
11912 Mask.push_back(ConstantInt::get(Type::Int32Ty, i));
11916 Instruction *InstCombiner::visitInsertElementInst(InsertElementInst &IE) {
11917 Value *VecOp = IE.getOperand(0);
11918 Value *ScalarOp = IE.getOperand(1);
11919 Value *IdxOp = IE.getOperand(2);
11921 // Inserting an undef or into an undefined place, remove this.
11922 if (isa<UndefValue>(ScalarOp) || isa<UndefValue>(IdxOp))
11923 ReplaceInstUsesWith(IE, VecOp);
11925 // If the inserted element was extracted from some other vector, and if the
11926 // indexes are constant, try to turn this into a shufflevector operation.
11927 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
11928 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
11929 EI->getOperand(0)->getType() == IE.getType()) {
11930 unsigned NumVectorElts = IE.getType()->getNumElements();
11931 unsigned ExtractedIdx =
11932 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
11933 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
11935 if (ExtractedIdx >= NumVectorElts) // Out of range extract.
11936 return ReplaceInstUsesWith(IE, VecOp);
11938 if (InsertedIdx >= NumVectorElts) // Out of range insert.
11939 return ReplaceInstUsesWith(IE, UndefValue::get(IE.getType()));
11941 // If we are extracting a value from a vector, then inserting it right
11942 // back into the same place, just use the input vector.
11943 if (EI->getOperand(0) == VecOp && ExtractedIdx == InsertedIdx)
11944 return ReplaceInstUsesWith(IE, VecOp);
11946 // We could theoretically do this for ANY input. However, doing so could
11947 // turn chains of insertelement instructions into a chain of shufflevector
11948 // instructions, and right now we do not merge shufflevectors. As such,
11949 // only do this in a situation where it is clear that there is benefit.
11950 if (isa<UndefValue>(VecOp) || isa<ConstantAggregateZero>(VecOp)) {
11951 // Turn this into shuffle(EIOp0, VecOp, Mask). The result has all of
11952 // the values of VecOp, except then one read from EIOp0.
11953 // Build a new shuffle mask.
11954 std::vector<Constant*> Mask;
11955 if (isa<UndefValue>(VecOp))
11956 Mask.assign(NumVectorElts, UndefValue::get(Type::Int32Ty));
11958 assert(isa<ConstantAggregateZero>(VecOp) && "Unknown thing");
11959 Mask.assign(NumVectorElts, ConstantInt::get(Type::Int32Ty,
11962 Mask[InsertedIdx] = ConstantInt::get(Type::Int32Ty, ExtractedIdx);
11963 return new ShuffleVectorInst(EI->getOperand(0), VecOp,
11964 ConstantVector::get(Mask));
11967 // If this insertelement isn't used by some other insertelement, turn it
11968 // (and any insertelements it points to), into one big shuffle.
11969 if (!IE.hasOneUse() || !isa<InsertElementInst>(IE.use_back())) {
11970 std::vector<Constant*> Mask;
11972 Value *LHS = CollectShuffleElements(&IE, Mask, RHS);
11973 if (RHS == 0) RHS = UndefValue::get(LHS->getType());
11974 // We now have a shuffle of LHS, RHS, Mask.
11975 return new ShuffleVectorInst(LHS, RHS, ConstantVector::get(Mask));
11984 Instruction *InstCombiner::visitShuffleVectorInst(ShuffleVectorInst &SVI) {
11985 Value *LHS = SVI.getOperand(0);
11986 Value *RHS = SVI.getOperand(1);
11987 std::vector<unsigned> Mask = getShuffleMask(&SVI);
11989 bool MadeChange = false;
11991 // Undefined shuffle mask -> undefined value.
11992 if (isa<UndefValue>(SVI.getOperand(2)))
11993 return ReplaceInstUsesWith(SVI, UndefValue::get(SVI.getType()));
11995 uint64_t UndefElts;
11996 unsigned VWidth = cast<VectorType>(SVI.getType())->getNumElements();
11998 if (VWidth != cast<VectorType>(LHS->getType())->getNumElements())
12001 uint64_t AllOnesEltMask = ~0ULL >> (64-VWidth);
12002 if (VWidth <= 64 &&
12003 SimplifyDemandedVectorElts(&SVI, AllOnesEltMask, UndefElts)) {
12004 LHS = SVI.getOperand(0);
12005 RHS = SVI.getOperand(1);
12009 // Canonicalize shuffle(x ,x,mask) -> shuffle(x, undef,mask')
12010 // Canonicalize shuffle(undef,x,mask) -> shuffle(x, undef,mask').
12011 if (LHS == RHS || isa<UndefValue>(LHS)) {
12012 if (isa<UndefValue>(LHS) && LHS == RHS) {
12013 // shuffle(undef,undef,mask) -> undef.
12014 return ReplaceInstUsesWith(SVI, LHS);
12017 // Remap any references to RHS to use LHS.
12018 std::vector<Constant*> Elts;
12019 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
12020 if (Mask[i] >= 2*e)
12021 Elts.push_back(UndefValue::get(Type::Int32Ty));
12023 if ((Mask[i] >= e && isa<UndefValue>(RHS)) ||
12024 (Mask[i] < e && isa<UndefValue>(LHS))) {
12025 Mask[i] = 2*e; // Turn into undef.
12026 Elts.push_back(UndefValue::get(Type::Int32Ty));
12028 Mask[i] = Mask[i] % e; // Force to LHS.
12029 Elts.push_back(ConstantInt::get(Type::Int32Ty, Mask[i]));
12033 SVI.setOperand(0, SVI.getOperand(1));
12034 SVI.setOperand(1, UndefValue::get(RHS->getType()));
12035 SVI.setOperand(2, ConstantVector::get(Elts));
12036 LHS = SVI.getOperand(0);
12037 RHS = SVI.getOperand(1);
12041 // Analyze the shuffle, are the LHS or RHS and identity shuffles?
12042 bool isLHSID = true, isRHSID = true;
12044 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
12045 if (Mask[i] >= e*2) continue; // Ignore undef values.
12046 // Is this an identity shuffle of the LHS value?
12047 isLHSID &= (Mask[i] == i);
12049 // Is this an identity shuffle of the RHS value?
12050 isRHSID &= (Mask[i]-e == i);
12053 // Eliminate identity shuffles.
12054 if (isLHSID) return ReplaceInstUsesWith(SVI, LHS);
12055 if (isRHSID) return ReplaceInstUsesWith(SVI, RHS);
12057 // If the LHS is a shufflevector itself, see if we can combine it with this
12058 // one without producing an unusual shuffle. Here we are really conservative:
12059 // we are absolutely afraid of producing a shuffle mask not in the input
12060 // program, because the code gen may not be smart enough to turn a merged
12061 // shuffle into two specific shuffles: it may produce worse code. As such,
12062 // we only merge two shuffles if the result is one of the two input shuffle
12063 // masks. In this case, merging the shuffles just removes one instruction,
12064 // which we know is safe. This is good for things like turning:
12065 // (splat(splat)) -> splat.
12066 if (ShuffleVectorInst *LHSSVI = dyn_cast<ShuffleVectorInst>(LHS)) {
12067 if (isa<UndefValue>(RHS)) {
12068 std::vector<unsigned> LHSMask = getShuffleMask(LHSSVI);
12070 std::vector<unsigned> NewMask;
12071 for (unsigned i = 0, e = Mask.size(); i != e; ++i)
12072 if (Mask[i] >= 2*e)
12073 NewMask.push_back(2*e);
12075 NewMask.push_back(LHSMask[Mask[i]]);
12077 // If the result mask is equal to the src shuffle or this shuffle mask, do
12078 // the replacement.
12079 if (NewMask == LHSMask || NewMask == Mask) {
12080 std::vector<Constant*> Elts;
12081 for (unsigned i = 0, e = NewMask.size(); i != e; ++i) {
12082 if (NewMask[i] >= e*2) {
12083 Elts.push_back(UndefValue::get(Type::Int32Ty));
12085 Elts.push_back(ConstantInt::get(Type::Int32Ty, NewMask[i]));
12088 return new ShuffleVectorInst(LHSSVI->getOperand(0),
12089 LHSSVI->getOperand(1),
12090 ConstantVector::get(Elts));
12095 return MadeChange ? &SVI : 0;
12101 /// TryToSinkInstruction - Try to move the specified instruction from its
12102 /// current block into the beginning of DestBlock, which can only happen if it's
12103 /// safe to move the instruction past all of the instructions between it and the
12104 /// end of its block.
12105 static bool TryToSinkInstruction(Instruction *I, BasicBlock *DestBlock) {
12106 assert(I->hasOneUse() && "Invariants didn't hold!");
12108 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
12109 if (isa<PHINode>(I) || I->mayWriteToMemory() || isa<TerminatorInst>(I))
12112 // Do not sink alloca instructions out of the entry block.
12113 if (isa<AllocaInst>(I) && I->getParent() ==
12114 &DestBlock->getParent()->getEntryBlock())
12117 // We can only sink load instructions if there is nothing between the load and
12118 // the end of block that could change the value.
12119 if (I->mayReadFromMemory()) {
12120 for (BasicBlock::iterator Scan = I, E = I->getParent()->end();
12122 if (Scan->mayWriteToMemory())
12126 BasicBlock::iterator InsertPos = DestBlock->getFirstNonPHI();
12128 I->moveBefore(InsertPos);
12134 /// AddReachableCodeToWorklist - Walk the function in depth-first order, adding
12135 /// all reachable code to the worklist.
12137 /// This has a couple of tricks to make the code faster and more powerful. In
12138 /// particular, we constant fold and DCE instructions as we go, to avoid adding
12139 /// them to the worklist (this significantly speeds up instcombine on code where
12140 /// many instructions are dead or constant). Additionally, if we find a branch
12141 /// whose condition is a known constant, we only visit the reachable successors.
12143 static void AddReachableCodeToWorklist(BasicBlock *BB,
12144 SmallPtrSet<BasicBlock*, 64> &Visited,
12146 const TargetData *TD) {
12147 SmallVector<BasicBlock*, 256> Worklist;
12148 Worklist.push_back(BB);
12150 while (!Worklist.empty()) {
12151 BB = Worklist.back();
12152 Worklist.pop_back();
12154 // We have now visited this block! If we've already been here, ignore it.
12155 if (!Visited.insert(BB)) continue;
12157 DbgInfoIntrinsic *DBI_Prev = NULL;
12158 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) {
12159 Instruction *Inst = BBI++;
12161 // DCE instruction if trivially dead.
12162 if (isInstructionTriviallyDead(Inst)) {
12164 DOUT << "IC: DCE: " << *Inst;
12165 Inst->eraseFromParent();
12169 // ConstantProp instruction if trivially constant.
12170 if (Constant *C = ConstantFoldInstruction(Inst, TD)) {
12171 DOUT << "IC: ConstFold to: " << *C << " from: " << *Inst;
12172 Inst->replaceAllUsesWith(C);
12174 Inst->eraseFromParent();
12178 // If there are two consecutive llvm.dbg.stoppoint calls then
12179 // it is likely that the optimizer deleted code in between these
12181 DbgInfoIntrinsic *DBI_Next = dyn_cast<DbgInfoIntrinsic>(Inst);
12184 && DBI_Prev->getIntrinsicID() == llvm::Intrinsic::dbg_stoppoint
12185 && DBI_Next->getIntrinsicID() == llvm::Intrinsic::dbg_stoppoint) {
12186 IC.RemoveFromWorkList(DBI_Prev);
12187 DBI_Prev->eraseFromParent();
12189 DBI_Prev = DBI_Next;
12192 IC.AddToWorkList(Inst);
12195 // Recursively visit successors. If this is a branch or switch on a
12196 // constant, only visit the reachable successor.
12197 TerminatorInst *TI = BB->getTerminator();
12198 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
12199 if (BI->isConditional() && isa<ConstantInt>(BI->getCondition())) {
12200 bool CondVal = cast<ConstantInt>(BI->getCondition())->getZExtValue();
12201 BasicBlock *ReachableBB = BI->getSuccessor(!CondVal);
12202 Worklist.push_back(ReachableBB);
12205 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
12206 if (ConstantInt *Cond = dyn_cast<ConstantInt>(SI->getCondition())) {
12207 // See if this is an explicit destination.
12208 for (unsigned i = 1, e = SI->getNumSuccessors(); i != e; ++i)
12209 if (SI->getCaseValue(i) == Cond) {
12210 BasicBlock *ReachableBB = SI->getSuccessor(i);
12211 Worklist.push_back(ReachableBB);
12215 // Otherwise it is the default destination.
12216 Worklist.push_back(SI->getSuccessor(0));
12221 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
12222 Worklist.push_back(TI->getSuccessor(i));
12226 bool InstCombiner::DoOneIteration(Function &F, unsigned Iteration) {
12227 bool Changed = false;
12228 TD = &getAnalysis<TargetData>();
12230 DEBUG(DOUT << "\n\nINSTCOMBINE ITERATION #" << Iteration << " on "
12231 << F.getNameStr() << "\n");
12234 // Do a depth-first traversal of the function, populate the worklist with
12235 // the reachable instructions. Ignore blocks that are not reachable. Keep
12236 // track of which blocks we visit.
12237 SmallPtrSet<BasicBlock*, 64> Visited;
12238 AddReachableCodeToWorklist(F.begin(), Visited, *this, TD);
12240 // Do a quick scan over the function. If we find any blocks that are
12241 // unreachable, remove any instructions inside of them. This prevents
12242 // the instcombine code from having to deal with some bad special cases.
12243 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
12244 if (!Visited.count(BB)) {
12245 Instruction *Term = BB->getTerminator();
12246 while (Term != BB->begin()) { // Remove instrs bottom-up
12247 BasicBlock::iterator I = Term; --I;
12249 DOUT << "IC: DCE: " << *I;
12252 if (!I->use_empty())
12253 I->replaceAllUsesWith(UndefValue::get(I->getType()));
12254 I->eraseFromParent();
12259 while (!Worklist.empty()) {
12260 Instruction *I = RemoveOneFromWorkList();
12261 if (I == 0) continue; // skip null values.
12263 // Check to see if we can DCE the instruction.
12264 if (isInstructionTriviallyDead(I)) {
12265 // Add operands to the worklist.
12266 if (I->getNumOperands() < 4)
12267 AddUsesToWorkList(*I);
12270 DOUT << "IC: DCE: " << *I;
12272 I->eraseFromParent();
12273 RemoveFromWorkList(I);
12277 // Instruction isn't dead, see if we can constant propagate it.
12278 if (Constant *C = ConstantFoldInstruction(I, TD)) {
12279 DOUT << "IC: ConstFold to: " << *C << " from: " << *I;
12281 // Add operands to the worklist.
12282 AddUsesToWorkList(*I);
12283 ReplaceInstUsesWith(*I, C);
12286 I->eraseFromParent();
12287 RemoveFromWorkList(I);
12291 if (TD && I->getType()->getTypeID() == Type::VoidTyID) {
12292 // See if we can constant fold its operands.
12293 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i) {
12294 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(i)) {
12295 if (Constant *NewC = ConstantFoldConstantExpression(CE, TD))
12301 // See if we can trivially sink this instruction to a successor basic block.
12302 if (I->hasOneUse()) {
12303 BasicBlock *BB = I->getParent();
12304 BasicBlock *UserParent = cast<Instruction>(I->use_back())->getParent();
12305 if (UserParent != BB) {
12306 bool UserIsSuccessor = false;
12307 // See if the user is one of our successors.
12308 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI)
12309 if (*SI == UserParent) {
12310 UserIsSuccessor = true;
12314 // If the user is one of our immediate successors, and if that successor
12315 // only has us as a predecessors (we'd have to split the critical edge
12316 // otherwise), we can keep going.
12317 if (UserIsSuccessor && !isa<PHINode>(I->use_back()) &&
12318 next(pred_begin(UserParent)) == pred_end(UserParent))
12319 // Okay, the CFG is simple enough, try to sink this instruction.
12320 Changed |= TryToSinkInstruction(I, UserParent);
12324 // Now that we have an instruction, try combining it to simplify it...
12328 DEBUG(std::ostringstream SS; I->print(SS); OrigI = SS.str(););
12329 if (Instruction *Result = visit(*I)) {
12331 // Should we replace the old instruction with a new one?
12333 DOUT << "IC: Old = " << *I
12334 << " New = " << *Result;
12336 // Everything uses the new instruction now.
12337 I->replaceAllUsesWith(Result);
12339 // Push the new instruction and any users onto the worklist.
12340 AddToWorkList(Result);
12341 AddUsersToWorkList(*Result);
12343 // Move the name to the new instruction first.
12344 Result->takeName(I);
12346 // Insert the new instruction into the basic block...
12347 BasicBlock *InstParent = I->getParent();
12348 BasicBlock::iterator InsertPos = I;
12350 if (!isa<PHINode>(Result)) // If combining a PHI, don't insert
12351 while (isa<PHINode>(InsertPos)) // middle of a block of PHIs.
12354 InstParent->getInstList().insert(InsertPos, Result);
12356 // Make sure that we reprocess all operands now that we reduced their
12358 AddUsesToWorkList(*I);
12360 // Instructions can end up on the worklist more than once. Make sure
12361 // we do not process an instruction that has been deleted.
12362 RemoveFromWorkList(I);
12364 // Erase the old instruction.
12365 InstParent->getInstList().erase(I);
12368 DOUT << "IC: Mod = " << OrigI
12369 << " New = " << *I;
12372 // If the instruction was modified, it's possible that it is now dead.
12373 // if so, remove it.
12374 if (isInstructionTriviallyDead(I)) {
12375 // Make sure we process all operands now that we are reducing their
12377 AddUsesToWorkList(*I);
12379 // Instructions may end up in the worklist more than once. Erase all
12380 // occurrences of this instruction.
12381 RemoveFromWorkList(I);
12382 I->eraseFromParent();
12385 AddUsersToWorkList(*I);
12392 assert(WorklistMap.empty() && "Worklist empty, but map not?");
12394 // Do an explicit clear, this shrinks the map if needed.
12395 WorklistMap.clear();
12400 bool InstCombiner::runOnFunction(Function &F) {
12401 MustPreserveLCSSA = mustPreserveAnalysisID(LCSSAID);
12403 bool EverMadeChange = false;
12405 // Iterate while there is work to do.
12406 unsigned Iteration = 0;
12407 while (DoOneIteration(F, Iteration++))
12408 EverMadeChange = true;
12409 return EverMadeChange;
12412 FunctionPass *llvm::createInstructionCombiningPass() {
12413 return new InstCombiner();