1 //===- InstructionCombining.cpp - Combine multiple instructions -----------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // InstructionCombining - Combine instructions to form fewer, simple
11 // instructions. This pass does not modify the CFG. This pass is where
12 // algebraic simplification happens.
14 // This pass combines things like:
20 // This is a simple worklist driven algorithm.
22 // This pass guarantees that the following canonicalizations are performed on
24 // 1. If a binary operator has a constant operand, it is moved to the RHS
25 // 2. Bitwise operators with constant operands are always grouped so that
26 // shifts are performed first, then or's, then and's, then xor's.
27 // 3. Compare instructions are converted from <,>,<=,>= to ==,!= if possible
28 // 4. All cmp instructions on boolean values are replaced with logical ops
29 // 5. add X, X is represented as (X*2) => (X << 1)
30 // 6. Multiplies with a power-of-two constant argument are transformed into
34 //===----------------------------------------------------------------------===//
36 #define DEBUG_TYPE "instcombine"
37 #include "llvm/Transforms/Scalar.h"
38 #include "llvm/IntrinsicInst.h"
39 #include "llvm/Pass.h"
40 #include "llvm/DerivedTypes.h"
41 #include "llvm/GlobalVariable.h"
42 #include "llvm/Analysis/ConstantFolding.h"
43 #include "llvm/Analysis/ValueTracking.h"
44 #include "llvm/Target/TargetData.h"
45 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
46 #include "llvm/Transforms/Utils/Local.h"
47 #include "llvm/Support/CallSite.h"
48 #include "llvm/Support/ConstantRange.h"
49 #include "llvm/Support/Debug.h"
50 #include "llvm/Support/GetElementPtrTypeIterator.h"
51 #include "llvm/Support/InstVisitor.h"
52 #include "llvm/Support/MathExtras.h"
53 #include "llvm/Support/PatternMatch.h"
54 #include "llvm/Support/Compiler.h"
55 #include "llvm/ADT/DenseMap.h"
56 #include "llvm/ADT/SmallVector.h"
57 #include "llvm/ADT/SmallPtrSet.h"
58 #include "llvm/ADT/Statistic.h"
59 #include "llvm/ADT/STLExtras.h"
64 using namespace llvm::PatternMatch;
66 STATISTIC(NumCombined , "Number of insts combined");
67 STATISTIC(NumConstProp, "Number of constant folds");
68 STATISTIC(NumDeadInst , "Number of dead inst eliminated");
69 STATISTIC(NumDeadStore, "Number of dead stores eliminated");
70 STATISTIC(NumSunkInst , "Number of instructions sunk");
73 class VISIBILITY_HIDDEN InstCombiner
74 : public FunctionPass,
75 public InstVisitor<InstCombiner, Instruction*> {
76 // Worklist of all of the instructions that need to be simplified.
77 SmallVector<Instruction*, 256> Worklist;
78 DenseMap<Instruction*, unsigned> WorklistMap;
80 bool MustPreserveLCSSA;
82 static char ID; // Pass identification, replacement for typeid
83 InstCombiner() : FunctionPass(&ID) {}
85 /// AddToWorkList - Add the specified instruction to the worklist if it
86 /// isn't already in it.
87 void AddToWorkList(Instruction *I) {
88 if (WorklistMap.insert(std::make_pair(I, Worklist.size())).second)
89 Worklist.push_back(I);
92 // RemoveFromWorkList - remove I from the worklist if it exists.
93 void RemoveFromWorkList(Instruction *I) {
94 DenseMap<Instruction*, unsigned>::iterator It = WorklistMap.find(I);
95 if (It == WorklistMap.end()) return; // Not in worklist.
97 // Don't bother moving everything down, just null out the slot.
98 Worklist[It->second] = 0;
100 WorklistMap.erase(It);
103 Instruction *RemoveOneFromWorkList() {
104 Instruction *I = Worklist.back();
106 WorklistMap.erase(I);
111 /// AddUsersToWorkList - When an instruction is simplified, add all users of
112 /// the instruction to the work lists because they might get more simplified
115 void AddUsersToWorkList(Value &I) {
116 for (Value::use_iterator UI = I.use_begin(), UE = I.use_end();
118 AddToWorkList(cast<Instruction>(*UI));
121 /// AddUsesToWorkList - When an instruction is simplified, add operands to
122 /// the work lists because they might get more simplified now.
124 void AddUsesToWorkList(Instruction &I) {
125 for (User::op_iterator i = I.op_begin(), e = I.op_end(); i != e; ++i)
126 if (Instruction *Op = dyn_cast<Instruction>(*i))
130 /// AddSoonDeadInstToWorklist - The specified instruction is about to become
131 /// dead. Add all of its operands to the worklist, turning them into
132 /// undef's to reduce the number of uses of those instructions.
134 /// Return the specified operand before it is turned into an undef.
136 Value *AddSoonDeadInstToWorklist(Instruction &I, unsigned op) {
137 Value *R = I.getOperand(op);
139 for (User::op_iterator i = I.op_begin(), e = I.op_end(); i != e; ++i)
140 if (Instruction *Op = dyn_cast<Instruction>(*i)) {
142 // Set the operand to undef to drop the use.
143 *i = UndefValue::get(Op->getType());
150 virtual bool runOnFunction(Function &F);
152 bool DoOneIteration(Function &F, unsigned ItNum);
154 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
155 AU.addRequired<TargetData>();
156 AU.addPreservedID(LCSSAID);
157 AU.setPreservesCFG();
160 TargetData &getTargetData() const { return *TD; }
162 // Visitation implementation - Implement instruction combining for different
163 // instruction types. The semantics are as follows:
165 // null - No change was made
166 // I - Change was made, I is still valid, I may be dead though
167 // otherwise - Change was made, replace I with returned instruction
169 Instruction *visitAdd(BinaryOperator &I);
170 Instruction *visitSub(BinaryOperator &I);
171 Instruction *visitMul(BinaryOperator &I);
172 Instruction *visitURem(BinaryOperator &I);
173 Instruction *visitSRem(BinaryOperator &I);
174 Instruction *visitFRem(BinaryOperator &I);
175 bool SimplifyDivRemOfSelect(BinaryOperator &I);
176 Instruction *commonRemTransforms(BinaryOperator &I);
177 Instruction *commonIRemTransforms(BinaryOperator &I);
178 Instruction *commonDivTransforms(BinaryOperator &I);
179 Instruction *commonIDivTransforms(BinaryOperator &I);
180 Instruction *visitUDiv(BinaryOperator &I);
181 Instruction *visitSDiv(BinaryOperator &I);
182 Instruction *visitFDiv(BinaryOperator &I);
183 Instruction *FoldAndOfICmps(Instruction &I, ICmpInst *LHS, ICmpInst *RHS);
184 Instruction *visitAnd(BinaryOperator &I);
185 Instruction *FoldOrOfICmps(Instruction &I, ICmpInst *LHS, ICmpInst *RHS);
186 Instruction *FoldOrWithConstants(BinaryOperator &I, Value *Op,
187 Value *A, Value *B, Value *C);
188 Instruction *visitOr (BinaryOperator &I);
189 Instruction *visitXor(BinaryOperator &I);
190 Instruction *visitShl(BinaryOperator &I);
191 Instruction *visitAShr(BinaryOperator &I);
192 Instruction *visitLShr(BinaryOperator &I);
193 Instruction *commonShiftTransforms(BinaryOperator &I);
194 Instruction *FoldFCmp_IntToFP_Cst(FCmpInst &I, Instruction *LHSI,
196 Instruction *visitFCmpInst(FCmpInst &I);
197 Instruction *visitICmpInst(ICmpInst &I);
198 Instruction *visitICmpInstWithCastAndCast(ICmpInst &ICI);
199 Instruction *visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
202 Instruction *FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
203 ConstantInt *DivRHS);
205 Instruction *FoldGEPICmp(User *GEPLHS, Value *RHS,
206 ICmpInst::Predicate Cond, Instruction &I);
207 Instruction *FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
209 Instruction *commonCastTransforms(CastInst &CI);
210 Instruction *commonIntCastTransforms(CastInst &CI);
211 Instruction *commonPointerCastTransforms(CastInst &CI);
212 Instruction *visitTrunc(TruncInst &CI);
213 Instruction *visitZExt(ZExtInst &CI);
214 Instruction *visitSExt(SExtInst &CI);
215 Instruction *visitFPTrunc(FPTruncInst &CI);
216 Instruction *visitFPExt(CastInst &CI);
217 Instruction *visitFPToUI(FPToUIInst &FI);
218 Instruction *visitFPToSI(FPToSIInst &FI);
219 Instruction *visitUIToFP(CastInst &CI);
220 Instruction *visitSIToFP(CastInst &CI);
221 Instruction *visitPtrToInt(CastInst &CI);
222 Instruction *visitIntToPtr(IntToPtrInst &CI);
223 Instruction *visitBitCast(BitCastInst &CI);
224 Instruction *FoldSelectOpOp(SelectInst &SI, Instruction *TI,
226 Instruction *visitSelectInst(SelectInst &SI);
227 Instruction *visitSelectInstWithICmp(SelectInst &SI, ICmpInst *ICI);
228 Instruction *visitCallInst(CallInst &CI);
229 Instruction *visitInvokeInst(InvokeInst &II);
230 Instruction *visitPHINode(PHINode &PN);
231 Instruction *visitGetElementPtrInst(GetElementPtrInst &GEP);
232 Instruction *visitAllocationInst(AllocationInst &AI);
233 Instruction *visitFreeInst(FreeInst &FI);
234 Instruction *visitLoadInst(LoadInst &LI);
235 Instruction *visitStoreInst(StoreInst &SI);
236 Instruction *visitBranchInst(BranchInst &BI);
237 Instruction *visitSwitchInst(SwitchInst &SI);
238 Instruction *visitInsertElementInst(InsertElementInst &IE);
239 Instruction *visitExtractElementInst(ExtractElementInst &EI);
240 Instruction *visitShuffleVectorInst(ShuffleVectorInst &SVI);
241 Instruction *visitExtractValueInst(ExtractValueInst &EV);
243 // visitInstruction - Specify what to return for unhandled instructions...
244 Instruction *visitInstruction(Instruction &I) { return 0; }
247 Instruction *visitCallSite(CallSite CS);
248 bool transformConstExprCastCall(CallSite CS);
249 Instruction *transformCallThroughTrampoline(CallSite CS);
250 Instruction *transformZExtICmp(ICmpInst *ICI, Instruction &CI,
251 bool DoXform = true);
252 bool WillNotOverflowSignedAdd(Value *LHS, Value *RHS);
255 // InsertNewInstBefore - insert an instruction New before instruction Old
256 // in the program. Add the new instruction to the worklist.
258 Instruction *InsertNewInstBefore(Instruction *New, Instruction &Old) {
259 assert(New && New->getParent() == 0 &&
260 "New instruction already inserted into a basic block!");
261 BasicBlock *BB = Old.getParent();
262 BB->getInstList().insert(&Old, New); // Insert inst
267 /// InsertCastBefore - Insert a cast of V to TY before the instruction POS.
268 /// This also adds the cast to the worklist. Finally, this returns the
270 Value *InsertCastBefore(Instruction::CastOps opc, Value *V, const Type *Ty,
272 if (V->getType() == Ty) return V;
274 if (Constant *CV = dyn_cast<Constant>(V))
275 return ConstantExpr::getCast(opc, CV, Ty);
277 Instruction *C = CastInst::Create(opc, V, Ty, V->getName(), &Pos);
282 Value *InsertBitCastBefore(Value *V, const Type *Ty, Instruction &Pos) {
283 return InsertCastBefore(Instruction::BitCast, V, Ty, Pos);
287 // ReplaceInstUsesWith - This method is to be used when an instruction is
288 // found to be dead, replacable with another preexisting expression. Here
289 // we add all uses of I to the worklist, replace all uses of I with the new
290 // value, then return I, so that the inst combiner will know that I was
293 Instruction *ReplaceInstUsesWith(Instruction &I, Value *V) {
294 AddUsersToWorkList(I); // Add all modified instrs to worklist
296 I.replaceAllUsesWith(V);
299 // If we are replacing the instruction with itself, this must be in a
300 // segment of unreachable code, so just clobber the instruction.
301 I.replaceAllUsesWith(UndefValue::get(I.getType()));
306 // UpdateValueUsesWith - This method is to be used when an value is
307 // found to be replacable with another preexisting expression or was
308 // updated. Here we add all uses of I to the worklist, replace all uses of
309 // I with the new value (unless the instruction was just updated), then
310 // return true, so that the inst combiner will know that I was modified.
312 bool UpdateValueUsesWith(Value *Old, Value *New) {
313 AddUsersToWorkList(*Old); // Add all modified instrs to worklist
315 Old->replaceAllUsesWith(New);
316 if (Instruction *I = dyn_cast<Instruction>(Old))
318 if (Instruction *I = dyn_cast<Instruction>(New))
323 // EraseInstFromFunction - When dealing with an instruction that has side
324 // effects or produces a void value, we can't rely on DCE to delete the
325 // instruction. Instead, visit methods should return the value returned by
327 Instruction *EraseInstFromFunction(Instruction &I) {
328 assert(I.use_empty() && "Cannot erase instruction that is used!");
329 AddUsesToWorkList(I);
330 RemoveFromWorkList(&I);
332 return 0; // Don't do anything with FI
335 void ComputeMaskedBits(Value *V, const APInt &Mask, APInt &KnownZero,
336 APInt &KnownOne, unsigned Depth = 0) const {
337 return llvm::ComputeMaskedBits(V, Mask, KnownZero, KnownOne, TD, Depth);
340 bool MaskedValueIsZero(Value *V, const APInt &Mask,
341 unsigned Depth = 0) const {
342 return llvm::MaskedValueIsZero(V, Mask, TD, Depth);
344 unsigned ComputeNumSignBits(Value *Op, unsigned Depth = 0) const {
345 return llvm::ComputeNumSignBits(Op, TD, Depth);
350 /// SimplifyCommutative - This performs a few simplifications for
351 /// commutative operators.
352 bool SimplifyCommutative(BinaryOperator &I);
354 /// SimplifyCompare - This reorders the operands of a CmpInst to get them in
355 /// most-complex to least-complex order.
356 bool SimplifyCompare(CmpInst &I);
358 /// SimplifyDemandedBits - Attempts to replace V with a simpler value based
359 /// on the demanded bits.
360 bool SimplifyDemandedBits(Value *V, APInt DemandedMask,
361 APInt& KnownZero, APInt& KnownOne,
364 Value *SimplifyDemandedVectorElts(Value *V, uint64_t DemandedElts,
365 uint64_t &UndefElts, unsigned Depth = 0);
367 // FoldOpIntoPhi - Given a binary operator or cast instruction which has a
368 // PHI node as operand #0, see if we can fold the instruction into the PHI
369 // (which is only possible if all operands to the PHI are constants).
370 Instruction *FoldOpIntoPhi(Instruction &I);
372 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
373 // operator and they all are only used by the PHI, PHI together their
374 // inputs, and do the operation once, to the result of the PHI.
375 Instruction *FoldPHIArgOpIntoPHI(PHINode &PN);
376 Instruction *FoldPHIArgBinOpIntoPHI(PHINode &PN);
377 Instruction *FoldPHIArgGEPIntoPHI(PHINode &PN);
380 Instruction *OptAndOp(Instruction *Op, ConstantInt *OpRHS,
381 ConstantInt *AndRHS, BinaryOperator &TheAnd);
383 Value *FoldLogicalPlusAnd(Value *LHS, Value *RHS, ConstantInt *Mask,
384 bool isSub, Instruction &I);
385 Instruction *InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
386 bool isSigned, bool Inside, Instruction &IB);
387 Instruction *PromoteCastOfAllocation(BitCastInst &CI, AllocationInst &AI);
388 Instruction *MatchBSwap(BinaryOperator &I);
389 bool SimplifyStoreAtEndOfBlock(StoreInst &SI);
390 Instruction *SimplifyMemTransfer(MemIntrinsic *MI);
391 Instruction *SimplifyMemSet(MemSetInst *MI);
394 Value *EvaluateInDifferentType(Value *V, const Type *Ty, bool isSigned);
396 bool CanEvaluateInDifferentType(Value *V, const IntegerType *Ty,
398 int &NumCastsRemoved);
399 unsigned GetOrEnforceKnownAlignment(Value *V,
400 unsigned PrefAlign = 0);
405 char InstCombiner::ID = 0;
406 static RegisterPass<InstCombiner>
407 X("instcombine", "Combine redundant instructions");
409 // getComplexity: Assign a complexity or rank value to LLVM Values...
410 // 0 -> undef, 1 -> Const, 2 -> Other, 3 -> Arg, 3 -> Unary, 4 -> OtherInst
411 static unsigned getComplexity(Value *V) {
412 if (isa<Instruction>(V)) {
413 if (BinaryOperator::isNeg(V) || BinaryOperator::isNot(V))
417 if (isa<Argument>(V)) return 3;
418 return isa<Constant>(V) ? (isa<UndefValue>(V) ? 0 : 1) : 2;
421 // isOnlyUse - Return true if this instruction will be deleted if we stop using
423 static bool isOnlyUse(Value *V) {
424 return V->hasOneUse() || isa<Constant>(V);
427 // getPromotedType - Return the specified type promoted as it would be to pass
428 // though a va_arg area...
429 static const Type *getPromotedType(const Type *Ty) {
430 if (const IntegerType* ITy = dyn_cast<IntegerType>(Ty)) {
431 if (ITy->getBitWidth() < 32)
432 return Type::Int32Ty;
437 /// getBitCastOperand - If the specified operand is a CastInst, a constant
438 /// expression bitcast, or a GetElementPtrInst with all zero indices, return the
439 /// operand value, otherwise return null.
440 static Value *getBitCastOperand(Value *V) {
441 if (BitCastInst *I = dyn_cast<BitCastInst>(V))
443 return I->getOperand(0);
444 else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(V)) {
445 // GetElementPtrInst?
446 if (GEP->hasAllZeroIndices())
447 return GEP->getOperand(0);
448 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) {
449 if (CE->getOpcode() == Instruction::BitCast)
450 // BitCast ConstantExp?
451 return CE->getOperand(0);
452 else if (CE->getOpcode() == Instruction::GetElementPtr) {
453 // GetElementPtr ConstantExp?
454 for (User::op_iterator I = CE->op_begin() + 1, E = CE->op_end();
456 ConstantInt *CI = dyn_cast<ConstantInt>(I);
457 if (!CI || !CI->isZero())
458 // Any non-zero indices? Not cast-like.
461 // All-zero indices? This is just like casting.
462 return CE->getOperand(0);
468 /// This function is a wrapper around CastInst::isEliminableCastPair. It
469 /// simply extracts arguments and returns what that function returns.
470 static Instruction::CastOps
471 isEliminableCastPair(
472 const CastInst *CI, ///< The first cast instruction
473 unsigned opcode, ///< The opcode of the second cast instruction
474 const Type *DstTy, ///< The target type for the second cast instruction
475 TargetData *TD ///< The target data for pointer size
478 const Type *SrcTy = CI->getOperand(0)->getType(); // A from above
479 const Type *MidTy = CI->getType(); // B from above
481 // Get the opcodes of the two Cast instructions
482 Instruction::CastOps firstOp = Instruction::CastOps(CI->getOpcode());
483 Instruction::CastOps secondOp = Instruction::CastOps(opcode);
485 return Instruction::CastOps(
486 CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy,
487 DstTy, TD->getIntPtrType()));
490 /// ValueRequiresCast - Return true if the cast from "V to Ty" actually results
491 /// in any code being generated. It does not require codegen if V is simple
492 /// enough or if the cast can be folded into other casts.
493 static bool ValueRequiresCast(Instruction::CastOps opcode, const Value *V,
494 const Type *Ty, TargetData *TD) {
495 if (V->getType() == Ty || isa<Constant>(V)) return false;
497 // If this is another cast that can be eliminated, it isn't codegen either.
498 if (const CastInst *CI = dyn_cast<CastInst>(V))
499 if (isEliminableCastPair(CI, opcode, Ty, TD))
504 // SimplifyCommutative - This performs a few simplifications for commutative
507 // 1. Order operands such that they are listed from right (least complex) to
508 // left (most complex). This puts constants before unary operators before
511 // 2. Transform: (op (op V, C1), C2) ==> (op V, (op C1, C2))
512 // 3. Transform: (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
514 bool InstCombiner::SimplifyCommutative(BinaryOperator &I) {
515 bool Changed = false;
516 if (getComplexity(I.getOperand(0)) < getComplexity(I.getOperand(1)))
517 Changed = !I.swapOperands();
519 if (!I.isAssociative()) return Changed;
520 Instruction::BinaryOps Opcode = I.getOpcode();
521 if (BinaryOperator *Op = dyn_cast<BinaryOperator>(I.getOperand(0)))
522 if (Op->getOpcode() == Opcode && isa<Constant>(Op->getOperand(1))) {
523 if (isa<Constant>(I.getOperand(1))) {
524 Constant *Folded = ConstantExpr::get(I.getOpcode(),
525 cast<Constant>(I.getOperand(1)),
526 cast<Constant>(Op->getOperand(1)));
527 I.setOperand(0, Op->getOperand(0));
528 I.setOperand(1, Folded);
530 } else if (BinaryOperator *Op1=dyn_cast<BinaryOperator>(I.getOperand(1)))
531 if (Op1->getOpcode() == Opcode && isa<Constant>(Op1->getOperand(1)) &&
532 isOnlyUse(Op) && isOnlyUse(Op1)) {
533 Constant *C1 = cast<Constant>(Op->getOperand(1));
534 Constant *C2 = cast<Constant>(Op1->getOperand(1));
536 // Fold (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
537 Constant *Folded = ConstantExpr::get(I.getOpcode(), C1, C2);
538 Instruction *New = BinaryOperator::Create(Opcode, Op->getOperand(0),
542 I.setOperand(0, New);
543 I.setOperand(1, Folded);
550 /// SimplifyCompare - For a CmpInst this function just orders the operands
551 /// so that theyare listed from right (least complex) to left (most complex).
552 /// This puts constants before unary operators before binary operators.
553 bool InstCombiner::SimplifyCompare(CmpInst &I) {
554 if (getComplexity(I.getOperand(0)) >= getComplexity(I.getOperand(1)))
557 // Compare instructions are not associative so there's nothing else we can do.
561 // dyn_castNegVal - Given a 'sub' instruction, return the RHS of the instruction
562 // if the LHS is a constant zero (which is the 'negate' form).
564 static inline Value *dyn_castNegVal(Value *V) {
565 if (BinaryOperator::isNeg(V))
566 return BinaryOperator::getNegArgument(V);
568 // Constants can be considered to be negated values if they can be folded.
569 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
570 return ConstantExpr::getNeg(C);
572 if (ConstantVector *C = dyn_cast<ConstantVector>(V))
573 if (C->getType()->getElementType()->isInteger())
574 return ConstantExpr::getNeg(C);
579 static inline Value *dyn_castNotVal(Value *V) {
580 if (BinaryOperator::isNot(V))
581 return BinaryOperator::getNotArgument(V);
583 // Constants can be considered to be not'ed values...
584 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
585 return ConstantInt::get(~C->getValue());
589 // dyn_castFoldableMul - If this value is a multiply that can be folded into
590 // other computations (because it has a constant operand), return the
591 // non-constant operand of the multiply, and set CST to point to the multiplier.
592 // Otherwise, return null.
594 static inline Value *dyn_castFoldableMul(Value *V, ConstantInt *&CST) {
595 if (V->hasOneUse() && V->getType()->isInteger())
596 if (Instruction *I = dyn_cast<Instruction>(V)) {
597 if (I->getOpcode() == Instruction::Mul)
598 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1))))
599 return I->getOperand(0);
600 if (I->getOpcode() == Instruction::Shl)
601 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1)))) {
602 // The multiplier is really 1 << CST.
603 uint32_t BitWidth = cast<IntegerType>(V->getType())->getBitWidth();
604 uint32_t CSTVal = CST->getLimitedValue(BitWidth);
605 CST = ConstantInt::get(APInt(BitWidth, 1).shl(CSTVal));
606 return I->getOperand(0);
612 /// dyn_castGetElementPtr - If this is a getelementptr instruction or constant
613 /// expression, return it.
614 static User *dyn_castGetElementPtr(Value *V) {
615 if (isa<GetElementPtrInst>(V)) return cast<User>(V);
616 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
617 if (CE->getOpcode() == Instruction::GetElementPtr)
618 return cast<User>(V);
622 /// getOpcode - If this is an Instruction or a ConstantExpr, return the
623 /// opcode value. Otherwise return UserOp1.
624 static unsigned getOpcode(const Value *V) {
625 if (const Instruction *I = dyn_cast<Instruction>(V))
626 return I->getOpcode();
627 if (const ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
628 return CE->getOpcode();
629 // Use UserOp1 to mean there's no opcode.
630 return Instruction::UserOp1;
633 /// AddOne - Add one to a ConstantInt
634 static ConstantInt *AddOne(ConstantInt *C) {
635 APInt Val(C->getValue());
636 return ConstantInt::get(++Val);
638 /// SubOne - Subtract one from a ConstantInt
639 static ConstantInt *SubOne(ConstantInt *C) {
640 APInt Val(C->getValue());
641 return ConstantInt::get(--Val);
643 /// Add - Add two ConstantInts together
644 static ConstantInt *Add(ConstantInt *C1, ConstantInt *C2) {
645 return ConstantInt::get(C1->getValue() + C2->getValue());
647 /// And - Bitwise AND two ConstantInts together
648 static ConstantInt *And(ConstantInt *C1, ConstantInt *C2) {
649 return ConstantInt::get(C1->getValue() & C2->getValue());
651 /// Subtract - Subtract one ConstantInt from another
652 static ConstantInt *Subtract(ConstantInt *C1, ConstantInt *C2) {
653 return ConstantInt::get(C1->getValue() - C2->getValue());
655 /// Multiply - Multiply two ConstantInts together
656 static ConstantInt *Multiply(ConstantInt *C1, ConstantInt *C2) {
657 return ConstantInt::get(C1->getValue() * C2->getValue());
659 /// MultiplyOverflows - True if the multiply can not be expressed in an int
661 static bool MultiplyOverflows(ConstantInt *C1, ConstantInt *C2, bool sign) {
662 uint32_t W = C1->getBitWidth();
663 APInt LHSExt = C1->getValue(), RHSExt = C2->getValue();
672 APInt MulExt = LHSExt * RHSExt;
675 APInt Min = APInt::getSignedMinValue(W).sext(W * 2);
676 APInt Max = APInt::getSignedMaxValue(W).sext(W * 2);
677 return MulExt.slt(Min) || MulExt.sgt(Max);
679 return MulExt.ugt(APInt::getLowBitsSet(W * 2, W));
683 /// ShrinkDemandedConstant - Check to see if the specified operand of the
684 /// specified instruction is a constant integer. If so, check to see if there
685 /// are any bits set in the constant that are not demanded. If so, shrink the
686 /// constant and return true.
687 static bool ShrinkDemandedConstant(Instruction *I, unsigned OpNo,
689 assert(I && "No instruction?");
690 assert(OpNo < I->getNumOperands() && "Operand index too large");
692 // If the operand is not a constant integer, nothing to do.
693 ConstantInt *OpC = dyn_cast<ConstantInt>(I->getOperand(OpNo));
694 if (!OpC) return false;
696 // If there are no bits set that aren't demanded, nothing to do.
697 Demanded.zextOrTrunc(OpC->getValue().getBitWidth());
698 if ((~Demanded & OpC->getValue()) == 0)
701 // This instruction is producing bits that are not demanded. Shrink the RHS.
702 Demanded &= OpC->getValue();
703 I->setOperand(OpNo, ConstantInt::get(Demanded));
707 // ComputeSignedMinMaxValuesFromKnownBits - Given a signed integer type and a
708 // set of known zero and one bits, compute the maximum and minimum values that
709 // could have the specified known zero and known one bits, returning them in
711 static void ComputeSignedMinMaxValuesFromKnownBits(const Type *Ty,
712 const APInt& KnownZero,
713 const APInt& KnownOne,
714 APInt& Min, APInt& Max) {
715 uint32_t BitWidth = cast<IntegerType>(Ty)->getBitWidth();
716 assert(KnownZero.getBitWidth() == BitWidth &&
717 KnownOne.getBitWidth() == BitWidth &&
718 Min.getBitWidth() == BitWidth && Max.getBitWidth() == BitWidth &&
719 "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth.");
720 APInt UnknownBits = ~(KnownZero|KnownOne);
722 // The minimum value is when all unknown bits are zeros, EXCEPT for the sign
723 // bit if it is unknown.
725 Max = KnownOne|UnknownBits;
727 if (UnknownBits[BitWidth-1]) { // Sign bit is unknown
729 Max.clear(BitWidth-1);
733 // ComputeUnsignedMinMaxValuesFromKnownBits - Given an unsigned integer type and
734 // a set of known zero and one bits, compute the maximum and minimum values that
735 // could have the specified known zero and known one bits, returning them in
737 static void ComputeUnsignedMinMaxValuesFromKnownBits(const Type *Ty,
738 const APInt &KnownZero,
739 const APInt &KnownOne,
740 APInt &Min, APInt &Max) {
741 uint32_t BitWidth = cast<IntegerType>(Ty)->getBitWidth(); BitWidth = BitWidth;
742 assert(KnownZero.getBitWidth() == BitWidth &&
743 KnownOne.getBitWidth() == BitWidth &&
744 Min.getBitWidth() == BitWidth && Max.getBitWidth() &&
745 "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth.");
746 APInt UnknownBits = ~(KnownZero|KnownOne);
748 // The minimum value is when the unknown bits are all zeros.
750 // The maximum value is when the unknown bits are all ones.
751 Max = KnownOne|UnknownBits;
754 /// SimplifyDemandedBits - This function attempts to replace V with a simpler
755 /// value based on the demanded bits. When this function is called, it is known
756 /// that only the bits set in DemandedMask of the result of V are ever used
757 /// downstream. Consequently, depending on the mask and V, it may be possible
758 /// to replace V with a constant or one of its operands. In such cases, this
759 /// function does the replacement and returns true. In all other cases, it
760 /// returns false after analyzing the expression and setting KnownOne and known
761 /// to be one in the expression. KnownZero contains all the bits that are known
762 /// to be zero in the expression. These are provided to potentially allow the
763 /// caller (which might recursively be SimplifyDemandedBits itself) to simplify
764 /// the expression. KnownOne and KnownZero always follow the invariant that
765 /// KnownOne & KnownZero == 0. That is, a bit can't be both 1 and 0. Note that
766 /// the bits in KnownOne and KnownZero may only be accurate for those bits set
767 /// in DemandedMask. Note also that the bitwidth of V, DemandedMask, KnownZero
768 /// and KnownOne must all be the same.
769 bool InstCombiner::SimplifyDemandedBits(Value *V, APInt DemandedMask,
770 APInt& KnownZero, APInt& KnownOne,
772 assert(V != 0 && "Null pointer of Value???");
773 assert(Depth <= 6 && "Limit Search Depth");
774 uint32_t BitWidth = DemandedMask.getBitWidth();
775 const IntegerType *VTy = cast<IntegerType>(V->getType());
776 assert(VTy->getBitWidth() == BitWidth &&
777 KnownZero.getBitWidth() == BitWidth &&
778 KnownOne.getBitWidth() == BitWidth &&
779 "Value *V, DemandedMask, KnownZero and KnownOne \
780 must have same BitWidth");
781 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
782 // We know all of the bits for a constant!
783 KnownOne = CI->getValue() & DemandedMask;
784 KnownZero = ~KnownOne & DemandedMask;
790 if (!V->hasOneUse()) { // Other users may use these bits.
791 if (Depth != 0) { // Not at the root.
792 // Just compute the KnownZero/KnownOne bits to simplify things downstream.
793 ComputeMaskedBits(V, DemandedMask, KnownZero, KnownOne, Depth);
796 // If this is the root being simplified, allow it to have multiple uses,
797 // just set the DemandedMask to all bits.
798 DemandedMask = APInt::getAllOnesValue(BitWidth);
799 } else if (DemandedMask == 0) { // Not demanding any bits from V.
800 if (V != UndefValue::get(VTy))
801 return UpdateValueUsesWith(V, UndefValue::get(VTy));
803 } else if (Depth == 6) { // Limit search depth.
807 Instruction *I = dyn_cast<Instruction>(V);
808 if (!I) return false; // Only analyze instructions.
810 APInt LHSKnownZero(BitWidth, 0), LHSKnownOne(BitWidth, 0);
811 APInt &RHSKnownZero = KnownZero, &RHSKnownOne = KnownOne;
812 switch (I->getOpcode()) {
814 ComputeMaskedBits(V, DemandedMask, RHSKnownZero, RHSKnownOne, Depth);
816 case Instruction::And:
817 // If either the LHS or the RHS are Zero, the result is zero.
818 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
819 RHSKnownZero, RHSKnownOne, Depth+1))
821 assert((RHSKnownZero & RHSKnownOne) == 0 &&
822 "Bits known to be one AND zero?");
824 // If something is known zero on the RHS, the bits aren't demanded on the
826 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask & ~RHSKnownZero,
827 LHSKnownZero, LHSKnownOne, Depth+1))
829 assert((LHSKnownZero & LHSKnownOne) == 0 &&
830 "Bits known to be one AND zero?");
832 // If all of the demanded bits are known 1 on one side, return the other.
833 // These bits cannot contribute to the result of the 'and'.
834 if ((DemandedMask & ~LHSKnownZero & RHSKnownOne) ==
835 (DemandedMask & ~LHSKnownZero))
836 return UpdateValueUsesWith(I, I->getOperand(0));
837 if ((DemandedMask & ~RHSKnownZero & LHSKnownOne) ==
838 (DemandedMask & ~RHSKnownZero))
839 return UpdateValueUsesWith(I, I->getOperand(1));
841 // If all of the demanded bits in the inputs are known zeros, return zero.
842 if ((DemandedMask & (RHSKnownZero|LHSKnownZero)) == DemandedMask)
843 return UpdateValueUsesWith(I, Constant::getNullValue(VTy));
845 // If the RHS is a constant, see if we can simplify it.
846 if (ShrinkDemandedConstant(I, 1, DemandedMask & ~LHSKnownZero))
847 return UpdateValueUsesWith(I, I);
849 // Output known-1 bits are only known if set in both the LHS & RHS.
850 RHSKnownOne &= LHSKnownOne;
851 // Output known-0 are known to be clear if zero in either the LHS | RHS.
852 RHSKnownZero |= LHSKnownZero;
854 case Instruction::Or:
855 // If either the LHS or the RHS are One, the result is One.
856 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
857 RHSKnownZero, RHSKnownOne, Depth+1))
859 assert((RHSKnownZero & RHSKnownOne) == 0 &&
860 "Bits known to be one AND zero?");
861 // If something is known one on the RHS, the bits aren't demanded on the
863 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask & ~RHSKnownOne,
864 LHSKnownZero, LHSKnownOne, Depth+1))
866 assert((LHSKnownZero & LHSKnownOne) == 0 &&
867 "Bits known to be one AND zero?");
869 // If all of the demanded bits are known zero on one side, return the other.
870 // These bits cannot contribute to the result of the 'or'.
871 if ((DemandedMask & ~LHSKnownOne & RHSKnownZero) ==
872 (DemandedMask & ~LHSKnownOne))
873 return UpdateValueUsesWith(I, I->getOperand(0));
874 if ((DemandedMask & ~RHSKnownOne & LHSKnownZero) ==
875 (DemandedMask & ~RHSKnownOne))
876 return UpdateValueUsesWith(I, I->getOperand(1));
878 // If all of the potentially set bits on one side are known to be set on
879 // the other side, just use the 'other' side.
880 if ((DemandedMask & (~RHSKnownZero) & LHSKnownOne) ==
881 (DemandedMask & (~RHSKnownZero)))
882 return UpdateValueUsesWith(I, I->getOperand(0));
883 if ((DemandedMask & (~LHSKnownZero) & RHSKnownOne) ==
884 (DemandedMask & (~LHSKnownZero)))
885 return UpdateValueUsesWith(I, I->getOperand(1));
887 // If the RHS is a constant, see if we can simplify it.
888 if (ShrinkDemandedConstant(I, 1, DemandedMask))
889 return UpdateValueUsesWith(I, I);
891 // Output known-0 bits are only known if clear in both the LHS & RHS.
892 RHSKnownZero &= LHSKnownZero;
893 // Output known-1 are known to be set if set in either the LHS | RHS.
894 RHSKnownOne |= LHSKnownOne;
896 case Instruction::Xor: {
897 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
898 RHSKnownZero, RHSKnownOne, Depth+1))
900 assert((RHSKnownZero & RHSKnownOne) == 0 &&
901 "Bits known to be one AND zero?");
902 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
903 LHSKnownZero, LHSKnownOne, Depth+1))
905 assert((LHSKnownZero & LHSKnownOne) == 0 &&
906 "Bits known to be one AND zero?");
908 // If all of the demanded bits are known zero on one side, return the other.
909 // These bits cannot contribute to the result of the 'xor'.
910 if ((DemandedMask & RHSKnownZero) == DemandedMask)
911 return UpdateValueUsesWith(I, I->getOperand(0));
912 if ((DemandedMask & LHSKnownZero) == DemandedMask)
913 return UpdateValueUsesWith(I, I->getOperand(1));
915 // Output known-0 bits are known if clear or set in both the LHS & RHS.
916 APInt KnownZeroOut = (RHSKnownZero & LHSKnownZero) |
917 (RHSKnownOne & LHSKnownOne);
918 // Output known-1 are known to be set if set in only one of the LHS, RHS.
919 APInt KnownOneOut = (RHSKnownZero & LHSKnownOne) |
920 (RHSKnownOne & LHSKnownZero);
922 // If all of the demanded bits are known to be zero on one side or the
923 // other, turn this into an *inclusive* or.
924 // e.g. (A & C1)^(B & C2) -> (A & C1)|(B & C2) iff C1&C2 == 0
925 if ((DemandedMask & ~RHSKnownZero & ~LHSKnownZero) == 0) {
927 BinaryOperator::CreateOr(I->getOperand(0), I->getOperand(1),
929 InsertNewInstBefore(Or, *I);
930 return UpdateValueUsesWith(I, Or);
933 // If all of the demanded bits on one side are known, and all of the set
934 // bits on that side are also known to be set on the other side, turn this
935 // into an AND, as we know the bits will be cleared.
936 // e.g. (X | C1) ^ C2 --> (X | C1) & ~C2 iff (C1&C2) == C2
937 if ((DemandedMask & (RHSKnownZero|RHSKnownOne)) == DemandedMask) {
939 if ((RHSKnownOne & LHSKnownOne) == RHSKnownOne) {
940 Constant *AndC = ConstantInt::get(~RHSKnownOne & DemandedMask);
942 BinaryOperator::CreateAnd(I->getOperand(0), AndC, "tmp");
943 InsertNewInstBefore(And, *I);
944 return UpdateValueUsesWith(I, And);
948 // If the RHS is a constant, see if we can simplify it.
949 // FIXME: for XOR, we prefer to force bits to 1 if they will make a -1.
950 if (ShrinkDemandedConstant(I, 1, DemandedMask))
951 return UpdateValueUsesWith(I, I);
953 RHSKnownZero = KnownZeroOut;
954 RHSKnownOne = KnownOneOut;
957 case Instruction::Select:
958 if (SimplifyDemandedBits(I->getOperand(2), DemandedMask,
959 RHSKnownZero, RHSKnownOne, Depth+1))
961 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
962 LHSKnownZero, LHSKnownOne, Depth+1))
964 assert((RHSKnownZero & RHSKnownOne) == 0 &&
965 "Bits known to be one AND zero?");
966 assert((LHSKnownZero & LHSKnownOne) == 0 &&
967 "Bits known to be one AND zero?");
969 // If the operands are constants, see if we can simplify them.
970 if (ShrinkDemandedConstant(I, 1, DemandedMask))
971 return UpdateValueUsesWith(I, I);
972 if (ShrinkDemandedConstant(I, 2, DemandedMask))
973 return UpdateValueUsesWith(I, I);
975 // Only known if known in both the LHS and RHS.
976 RHSKnownOne &= LHSKnownOne;
977 RHSKnownZero &= LHSKnownZero;
979 case Instruction::Trunc: {
981 cast<IntegerType>(I->getOperand(0)->getType())->getBitWidth();
982 DemandedMask.zext(truncBf);
983 RHSKnownZero.zext(truncBf);
984 RHSKnownOne.zext(truncBf);
985 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
986 RHSKnownZero, RHSKnownOne, Depth+1))
988 DemandedMask.trunc(BitWidth);
989 RHSKnownZero.trunc(BitWidth);
990 RHSKnownOne.trunc(BitWidth);
991 assert((RHSKnownZero & RHSKnownOne) == 0 &&
992 "Bits known to be one AND zero?");
995 case Instruction::BitCast:
996 if (!I->getOperand(0)->getType()->isInteger())
999 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1000 RHSKnownZero, RHSKnownOne, Depth+1))
1002 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1003 "Bits known to be one AND zero?");
1005 case Instruction::ZExt: {
1006 // Compute the bits in the result that are not present in the input.
1007 const IntegerType *SrcTy = cast<IntegerType>(I->getOperand(0)->getType());
1008 uint32_t SrcBitWidth = SrcTy->getBitWidth();
1010 DemandedMask.trunc(SrcBitWidth);
1011 RHSKnownZero.trunc(SrcBitWidth);
1012 RHSKnownOne.trunc(SrcBitWidth);
1013 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1014 RHSKnownZero, RHSKnownOne, Depth+1))
1016 DemandedMask.zext(BitWidth);
1017 RHSKnownZero.zext(BitWidth);
1018 RHSKnownOne.zext(BitWidth);
1019 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1020 "Bits known to be one AND zero?");
1021 // The top bits are known to be zero.
1022 RHSKnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
1025 case Instruction::SExt: {
1026 // Compute the bits in the result that are not present in the input.
1027 const IntegerType *SrcTy = cast<IntegerType>(I->getOperand(0)->getType());
1028 uint32_t SrcBitWidth = SrcTy->getBitWidth();
1030 APInt InputDemandedBits = DemandedMask &
1031 APInt::getLowBitsSet(BitWidth, SrcBitWidth);
1033 APInt NewBits(APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth));
1034 // If any of the sign extended bits are demanded, we know that the sign
1036 if ((NewBits & DemandedMask) != 0)
1037 InputDemandedBits.set(SrcBitWidth-1);
1039 InputDemandedBits.trunc(SrcBitWidth);
1040 RHSKnownZero.trunc(SrcBitWidth);
1041 RHSKnownOne.trunc(SrcBitWidth);
1042 if (SimplifyDemandedBits(I->getOperand(0), InputDemandedBits,
1043 RHSKnownZero, RHSKnownOne, Depth+1))
1045 InputDemandedBits.zext(BitWidth);
1046 RHSKnownZero.zext(BitWidth);
1047 RHSKnownOne.zext(BitWidth);
1048 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1049 "Bits known to be one AND zero?");
1051 // If the sign bit of the input is known set or clear, then we know the
1052 // top bits of the result.
1054 // If the input sign bit is known zero, or if the NewBits are not demanded
1055 // convert this into a zero extension.
1056 if (RHSKnownZero[SrcBitWidth-1] || (NewBits & ~DemandedMask) == NewBits)
1058 // Convert to ZExt cast
1059 CastInst *NewCast = new ZExtInst(I->getOperand(0), VTy, I->getName(), I);
1060 return UpdateValueUsesWith(I, NewCast);
1061 } else if (RHSKnownOne[SrcBitWidth-1]) { // Input sign bit known set
1062 RHSKnownOne |= NewBits;
1066 case Instruction::Add: {
1067 // Figure out what the input bits are. If the top bits of the and result
1068 // are not demanded, then the add doesn't demand them from its input
1070 uint32_t NLZ = DemandedMask.countLeadingZeros();
1072 // If there is a constant on the RHS, there are a variety of xformations
1074 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
1075 // If null, this should be simplified elsewhere. Some of the xforms here
1076 // won't work if the RHS is zero.
1080 // If the top bit of the output is demanded, demand everything from the
1081 // input. Otherwise, we demand all the input bits except NLZ top bits.
1082 APInt InDemandedBits(APInt::getLowBitsSet(BitWidth, BitWidth - NLZ));
1084 // Find information about known zero/one bits in the input.
1085 if (SimplifyDemandedBits(I->getOperand(0), InDemandedBits,
1086 LHSKnownZero, LHSKnownOne, Depth+1))
1089 // If the RHS of the add has bits set that can't affect the input, reduce
1091 if (ShrinkDemandedConstant(I, 1, InDemandedBits))
1092 return UpdateValueUsesWith(I, I);
1094 // Avoid excess work.
1095 if (LHSKnownZero == 0 && LHSKnownOne == 0)
1098 // Turn it into OR if input bits are zero.
1099 if ((LHSKnownZero & RHS->getValue()) == RHS->getValue()) {
1101 BinaryOperator::CreateOr(I->getOperand(0), I->getOperand(1),
1103 InsertNewInstBefore(Or, *I);
1104 return UpdateValueUsesWith(I, Or);
1107 // We can say something about the output known-zero and known-one bits,
1108 // depending on potential carries from the input constant and the
1109 // unknowns. For example if the LHS is known to have at most the 0x0F0F0
1110 // bits set and the RHS constant is 0x01001, then we know we have a known
1111 // one mask of 0x00001 and a known zero mask of 0xE0F0E.
1113 // To compute this, we first compute the potential carry bits. These are
1114 // the bits which may be modified. I'm not aware of a better way to do
1116 const APInt& RHSVal = RHS->getValue();
1117 APInt CarryBits((~LHSKnownZero + RHSVal) ^ (~LHSKnownZero ^ RHSVal));
1119 // Now that we know which bits have carries, compute the known-1/0 sets.
1121 // Bits are known one if they are known zero in one operand and one in the
1122 // other, and there is no input carry.
1123 RHSKnownOne = ((LHSKnownZero & RHSVal) |
1124 (LHSKnownOne & ~RHSVal)) & ~CarryBits;
1126 // Bits are known zero if they are known zero in both operands and there
1127 // is no input carry.
1128 RHSKnownZero = LHSKnownZero & ~RHSVal & ~CarryBits;
1130 // If the high-bits of this ADD are not demanded, then it does not demand
1131 // the high bits of its LHS or RHS.
1132 if (DemandedMask[BitWidth-1] == 0) {
1133 // Right fill the mask of bits for this ADD to demand the most
1134 // significant bit and all those below it.
1135 APInt DemandedFromOps(APInt::getLowBitsSet(BitWidth, BitWidth-NLZ));
1136 if (SimplifyDemandedBits(I->getOperand(0), DemandedFromOps,
1137 LHSKnownZero, LHSKnownOne, Depth+1))
1139 if (SimplifyDemandedBits(I->getOperand(1), DemandedFromOps,
1140 LHSKnownZero, LHSKnownOne, Depth+1))
1146 case Instruction::Sub:
1147 // If the high-bits of this SUB are not demanded, then it does not demand
1148 // the high bits of its LHS or RHS.
1149 if (DemandedMask[BitWidth-1] == 0) {
1150 // Right fill the mask of bits for this SUB to demand the most
1151 // significant bit and all those below it.
1152 uint32_t NLZ = DemandedMask.countLeadingZeros();
1153 APInt DemandedFromOps(APInt::getLowBitsSet(BitWidth, BitWidth-NLZ));
1154 if (SimplifyDemandedBits(I->getOperand(0), DemandedFromOps,
1155 LHSKnownZero, LHSKnownOne, Depth+1))
1157 if (SimplifyDemandedBits(I->getOperand(1), DemandedFromOps,
1158 LHSKnownZero, LHSKnownOne, Depth+1))
1161 // Otherwise just hand the sub off to ComputeMaskedBits to fill in
1162 // the known zeros and ones.
1163 ComputeMaskedBits(V, DemandedMask, RHSKnownZero, RHSKnownOne, Depth);
1165 case Instruction::Shl:
1166 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1167 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
1168 APInt DemandedMaskIn(DemandedMask.lshr(ShiftAmt));
1169 if (SimplifyDemandedBits(I->getOperand(0), DemandedMaskIn,
1170 RHSKnownZero, RHSKnownOne, Depth+1))
1172 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1173 "Bits known to be one AND zero?");
1174 RHSKnownZero <<= ShiftAmt;
1175 RHSKnownOne <<= ShiftAmt;
1176 // low bits known zero.
1178 RHSKnownZero |= APInt::getLowBitsSet(BitWidth, ShiftAmt);
1181 case Instruction::LShr:
1182 // For a logical shift right
1183 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1184 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
1186 // Unsigned shift right.
1187 APInt DemandedMaskIn(DemandedMask.shl(ShiftAmt));
1188 if (SimplifyDemandedBits(I->getOperand(0), DemandedMaskIn,
1189 RHSKnownZero, RHSKnownOne, Depth+1))
1191 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1192 "Bits known to be one AND zero?");
1193 RHSKnownZero = APIntOps::lshr(RHSKnownZero, ShiftAmt);
1194 RHSKnownOne = APIntOps::lshr(RHSKnownOne, ShiftAmt);
1196 // Compute the new bits that are at the top now.
1197 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
1198 RHSKnownZero |= HighBits; // high bits known zero.
1202 case Instruction::AShr:
1203 // If this is an arithmetic shift right and only the low-bit is set, we can
1204 // always convert this into a logical shr, even if the shift amount is
1205 // variable. The low bit of the shift cannot be an input sign bit unless
1206 // the shift amount is >= the size of the datatype, which is undefined.
1207 if (DemandedMask == 1) {
1208 // Perform the logical shift right.
1209 Value *NewVal = BinaryOperator::CreateLShr(
1210 I->getOperand(0), I->getOperand(1), I->getName());
1211 InsertNewInstBefore(cast<Instruction>(NewVal), *I);
1212 return UpdateValueUsesWith(I, NewVal);
1215 // If the sign bit is the only bit demanded by this ashr, then there is no
1216 // need to do it, the shift doesn't change the high bit.
1217 if (DemandedMask.isSignBit())
1218 return UpdateValueUsesWith(I, I->getOperand(0));
1220 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1221 uint32_t ShiftAmt = SA->getLimitedValue(BitWidth);
1223 // Signed shift right.
1224 APInt DemandedMaskIn(DemandedMask.shl(ShiftAmt));
1225 // If any of the "high bits" are demanded, we should set the sign bit as
1227 if (DemandedMask.countLeadingZeros() <= ShiftAmt)
1228 DemandedMaskIn.set(BitWidth-1);
1229 if (SimplifyDemandedBits(I->getOperand(0),
1231 RHSKnownZero, RHSKnownOne, Depth+1))
1233 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1234 "Bits known to be one AND zero?");
1235 // Compute the new bits that are at the top now.
1236 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
1237 RHSKnownZero = APIntOps::lshr(RHSKnownZero, ShiftAmt);
1238 RHSKnownOne = APIntOps::lshr(RHSKnownOne, ShiftAmt);
1240 // Handle the sign bits.
1241 APInt SignBit(APInt::getSignBit(BitWidth));
1242 // Adjust to where it is now in the mask.
1243 SignBit = APIntOps::lshr(SignBit, ShiftAmt);
1245 // If the input sign bit is known to be zero, or if none of the top bits
1246 // are demanded, turn this into an unsigned shift right.
1247 if (BitWidth <= ShiftAmt || RHSKnownZero[BitWidth-ShiftAmt-1] ||
1248 (HighBits & ~DemandedMask) == HighBits) {
1249 // Perform the logical shift right.
1250 Value *NewVal = BinaryOperator::CreateLShr(
1251 I->getOperand(0), SA, I->getName());
1252 InsertNewInstBefore(cast<Instruction>(NewVal), *I);
1253 return UpdateValueUsesWith(I, NewVal);
1254 } else if ((RHSKnownOne & SignBit) != 0) { // New bits are known one.
1255 RHSKnownOne |= HighBits;
1259 case Instruction::SRem:
1260 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
1261 APInt RA = Rem->getValue().abs();
1262 if (RA.isPowerOf2()) {
1263 if (DemandedMask.ule(RA)) // srem won't affect demanded bits
1264 return UpdateValueUsesWith(I, I->getOperand(0));
1266 APInt LowBits = RA - 1;
1267 APInt Mask2 = LowBits | APInt::getSignBit(BitWidth);
1268 if (SimplifyDemandedBits(I->getOperand(0), Mask2,
1269 LHSKnownZero, LHSKnownOne, Depth+1))
1272 if (LHSKnownZero[BitWidth-1] || ((LHSKnownZero & LowBits) == LowBits))
1273 LHSKnownZero |= ~LowBits;
1275 KnownZero |= LHSKnownZero & DemandedMask;
1277 assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?");
1281 case Instruction::URem: {
1282 APInt KnownZero2(BitWidth, 0), KnownOne2(BitWidth, 0);
1283 APInt AllOnes = APInt::getAllOnesValue(BitWidth);
1284 if (SimplifyDemandedBits(I->getOperand(0), AllOnes,
1285 KnownZero2, KnownOne2, Depth+1))
1288 uint32_t Leaders = KnownZero2.countLeadingOnes();
1289 if (SimplifyDemandedBits(I->getOperand(1), AllOnes,
1290 KnownZero2, KnownOne2, Depth+1))
1293 Leaders = std::max(Leaders,
1294 KnownZero2.countLeadingOnes());
1295 KnownZero = APInt::getHighBitsSet(BitWidth, Leaders) & DemandedMask;
1298 case Instruction::Call:
1299 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
1300 switch (II->getIntrinsicID()) {
1302 case Intrinsic::bswap: {
1303 // If the only bits demanded come from one byte of the bswap result,
1304 // just shift the input byte into position to eliminate the bswap.
1305 unsigned NLZ = DemandedMask.countLeadingZeros();
1306 unsigned NTZ = DemandedMask.countTrailingZeros();
1308 // Round NTZ down to the next byte. If we have 11 trailing zeros, then
1309 // we need all the bits down to bit 8. Likewise, round NLZ. If we
1310 // have 14 leading zeros, round to 8.
1313 // If we need exactly one byte, we can do this transformation.
1314 if (BitWidth-NLZ-NTZ == 8) {
1315 unsigned ResultBit = NTZ;
1316 unsigned InputBit = BitWidth-NTZ-8;
1318 // Replace this with either a left or right shift to get the byte into
1320 Instruction *NewVal;
1321 if (InputBit > ResultBit)
1322 NewVal = BinaryOperator::CreateLShr(I->getOperand(1),
1323 ConstantInt::get(I->getType(), InputBit-ResultBit));
1325 NewVal = BinaryOperator::CreateShl(I->getOperand(1),
1326 ConstantInt::get(I->getType(), ResultBit-InputBit));
1327 NewVal->takeName(I);
1328 InsertNewInstBefore(NewVal, *I);
1329 return UpdateValueUsesWith(I, NewVal);
1332 // TODO: Could compute known zero/one bits based on the input.
1337 ComputeMaskedBits(V, DemandedMask, RHSKnownZero, RHSKnownOne, Depth);
1341 // If the client is only demanding bits that we know, return the known
1343 if ((DemandedMask & (RHSKnownZero|RHSKnownOne)) == DemandedMask)
1344 return UpdateValueUsesWith(I, ConstantInt::get(RHSKnownOne));
1349 /// SimplifyDemandedVectorElts - The specified value produces a vector with
1350 /// 64 or fewer elements. DemandedElts contains the set of elements that are
1351 /// actually used by the caller. This method analyzes which elements of the
1352 /// operand are undef and returns that information in UndefElts.
1354 /// If the information about demanded elements can be used to simplify the
1355 /// operation, the operation is simplified, then the resultant value is
1356 /// returned. This returns null if no change was made.
1357 Value *InstCombiner::SimplifyDemandedVectorElts(Value *V, uint64_t DemandedElts,
1358 uint64_t &UndefElts,
1360 unsigned VWidth = cast<VectorType>(V->getType())->getNumElements();
1361 assert(VWidth <= 64 && "Vector too wide to analyze!");
1362 uint64_t EltMask = ~0ULL >> (64-VWidth);
1363 assert((DemandedElts & ~EltMask) == 0 && "Invalid DemandedElts!");
1365 if (isa<UndefValue>(V)) {
1366 // If the entire vector is undefined, just return this info.
1367 UndefElts = EltMask;
1369 } else if (DemandedElts == 0) { // If nothing is demanded, provide undef.
1370 UndefElts = EltMask;
1371 return UndefValue::get(V->getType());
1375 if (ConstantVector *CP = dyn_cast<ConstantVector>(V)) {
1376 const Type *EltTy = cast<VectorType>(V->getType())->getElementType();
1377 Constant *Undef = UndefValue::get(EltTy);
1379 std::vector<Constant*> Elts;
1380 for (unsigned i = 0; i != VWidth; ++i)
1381 if (!(DemandedElts & (1ULL << i))) { // If not demanded, set to undef.
1382 Elts.push_back(Undef);
1383 UndefElts |= (1ULL << i);
1384 } else if (isa<UndefValue>(CP->getOperand(i))) { // Already undef.
1385 Elts.push_back(Undef);
1386 UndefElts |= (1ULL << i);
1387 } else { // Otherwise, defined.
1388 Elts.push_back(CP->getOperand(i));
1391 // If we changed the constant, return it.
1392 Constant *NewCP = ConstantVector::get(Elts);
1393 return NewCP != CP ? NewCP : 0;
1394 } else if (isa<ConstantAggregateZero>(V)) {
1395 // Simplify the CAZ to a ConstantVector where the non-demanded elements are
1398 // Check if this is identity. If so, return 0 since we are not simplifying
1400 if (DemandedElts == ((1ULL << VWidth) -1))
1403 const Type *EltTy = cast<VectorType>(V->getType())->getElementType();
1404 Constant *Zero = Constant::getNullValue(EltTy);
1405 Constant *Undef = UndefValue::get(EltTy);
1406 std::vector<Constant*> Elts;
1407 for (unsigned i = 0; i != VWidth; ++i)
1408 Elts.push_back((DemandedElts & (1ULL << i)) ? Zero : Undef);
1409 UndefElts = DemandedElts ^ EltMask;
1410 return ConstantVector::get(Elts);
1413 // Limit search depth.
1417 // If multiple users are using the root value, procede with
1418 // simplification conservatively assuming that all elements
1420 if (!V->hasOneUse()) {
1421 // Quit if we find multiple users of a non-root value though.
1422 // They'll be handled when it's their turn to be visited by
1423 // the main instcombine process.
1425 // TODO: Just compute the UndefElts information recursively.
1428 // Conservatively assume that all elements are needed.
1429 DemandedElts = EltMask;
1432 Instruction *I = dyn_cast<Instruction>(V);
1433 if (!I) return false; // Only analyze instructions.
1435 bool MadeChange = false;
1436 uint64_t UndefElts2;
1438 switch (I->getOpcode()) {
1441 case Instruction::InsertElement: {
1442 // If this is a variable index, we don't know which element it overwrites.
1443 // demand exactly the same input as we produce.
1444 ConstantInt *Idx = dyn_cast<ConstantInt>(I->getOperand(2));
1446 // Note that we can't propagate undef elt info, because we don't know
1447 // which elt is getting updated.
1448 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
1449 UndefElts2, Depth+1);
1450 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1454 // If this is inserting an element that isn't demanded, remove this
1456 unsigned IdxNo = Idx->getZExtValue();
1457 if (IdxNo >= VWidth || (DemandedElts & (1ULL << IdxNo)) == 0)
1458 return AddSoonDeadInstToWorklist(*I, 0);
1460 // Otherwise, the element inserted overwrites whatever was there, so the
1461 // input demanded set is simpler than the output set.
1462 TmpV = SimplifyDemandedVectorElts(I->getOperand(0),
1463 DemandedElts & ~(1ULL << IdxNo),
1464 UndefElts, Depth+1);
1465 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1467 // The inserted element is defined.
1468 UndefElts &= ~(1ULL << IdxNo);
1471 case Instruction::ShuffleVector: {
1472 ShuffleVectorInst *Shuffle = cast<ShuffleVectorInst>(I);
1473 uint64_t LHSVWidth =
1474 cast<VectorType>(Shuffle->getOperand(0)->getType())->getNumElements();
1475 uint64_t LeftDemanded = 0, RightDemanded = 0;
1476 for (unsigned i = 0; i < VWidth; i++) {
1477 if (DemandedElts & (1ULL << i)) {
1478 unsigned MaskVal = Shuffle->getMaskValue(i);
1479 if (MaskVal != -1u) {
1480 assert(MaskVal < LHSVWidth * 2 &&
1481 "shufflevector mask index out of range!");
1482 if (MaskVal < LHSVWidth)
1483 LeftDemanded |= 1ULL << MaskVal;
1485 RightDemanded |= 1ULL << (MaskVal - LHSVWidth);
1490 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), LeftDemanded,
1491 UndefElts2, Depth+1);
1492 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1494 uint64_t UndefElts3;
1495 TmpV = SimplifyDemandedVectorElts(I->getOperand(1), RightDemanded,
1496 UndefElts3, Depth+1);
1497 if (TmpV) { I->setOperand(1, TmpV); MadeChange = true; }
1499 bool NewUndefElts = false;
1500 for (unsigned i = 0; i < VWidth; i++) {
1501 unsigned MaskVal = Shuffle->getMaskValue(i);
1502 if (MaskVal == -1u) {
1503 uint64_t NewBit = 1ULL << i;
1504 UndefElts |= NewBit;
1505 } else if (MaskVal < LHSVWidth) {
1506 uint64_t NewBit = ((UndefElts2 >> MaskVal) & 1) << i;
1507 NewUndefElts |= NewBit;
1508 UndefElts |= NewBit;
1510 uint64_t NewBit = ((UndefElts3 >> (MaskVal - LHSVWidth)) & 1) << i;
1511 NewUndefElts |= NewBit;
1512 UndefElts |= NewBit;
1517 // Add additional discovered undefs.
1518 std::vector<Constant*> Elts;
1519 for (unsigned i = 0; i < VWidth; ++i) {
1520 if (UndefElts & (1ULL << i))
1521 Elts.push_back(UndefValue::get(Type::Int32Ty));
1523 Elts.push_back(ConstantInt::get(Type::Int32Ty,
1524 Shuffle->getMaskValue(i)));
1526 I->setOperand(2, ConstantVector::get(Elts));
1531 case Instruction::BitCast: {
1532 // Vector->vector casts only.
1533 const VectorType *VTy = dyn_cast<VectorType>(I->getOperand(0)->getType());
1535 unsigned InVWidth = VTy->getNumElements();
1536 uint64_t InputDemandedElts = 0;
1539 if (VWidth == InVWidth) {
1540 // If we are converting from <4 x i32> -> <4 x f32>, we demand the same
1541 // elements as are demanded of us.
1543 InputDemandedElts = DemandedElts;
1544 } else if (VWidth > InVWidth) {
1548 // If there are more elements in the result than there are in the source,
1549 // then an input element is live if any of the corresponding output
1550 // elements are live.
1551 Ratio = VWidth/InVWidth;
1552 for (unsigned OutIdx = 0; OutIdx != VWidth; ++OutIdx) {
1553 if (DemandedElts & (1ULL << OutIdx))
1554 InputDemandedElts |= 1ULL << (OutIdx/Ratio);
1560 // If there are more elements in the source than there are in the result,
1561 // then an input element is live if the corresponding output element is
1563 Ratio = InVWidth/VWidth;
1564 for (unsigned InIdx = 0; InIdx != InVWidth; ++InIdx)
1565 if (DemandedElts & (1ULL << InIdx/Ratio))
1566 InputDemandedElts |= 1ULL << InIdx;
1569 // div/rem demand all inputs, because they don't want divide by zero.
1570 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), InputDemandedElts,
1571 UndefElts2, Depth+1);
1573 I->setOperand(0, TmpV);
1577 UndefElts = UndefElts2;
1578 if (VWidth > InVWidth) {
1579 assert(0 && "Unimp");
1580 // If there are more elements in the result than there are in the source,
1581 // then an output element is undef if the corresponding input element is
1583 for (unsigned OutIdx = 0; OutIdx != VWidth; ++OutIdx)
1584 if (UndefElts2 & (1ULL << (OutIdx/Ratio)))
1585 UndefElts |= 1ULL << OutIdx;
1586 } else if (VWidth < InVWidth) {
1587 assert(0 && "Unimp");
1588 // If there are more elements in the source than there are in the result,
1589 // then a result element is undef if all of the corresponding input
1590 // elements are undef.
1591 UndefElts = ~0ULL >> (64-VWidth); // Start out all undef.
1592 for (unsigned InIdx = 0; InIdx != InVWidth; ++InIdx)
1593 if ((UndefElts2 & (1ULL << InIdx)) == 0) // Not undef?
1594 UndefElts &= ~(1ULL << (InIdx/Ratio)); // Clear undef bit.
1598 case Instruction::And:
1599 case Instruction::Or:
1600 case Instruction::Xor:
1601 case Instruction::Add:
1602 case Instruction::Sub:
1603 case Instruction::Mul:
1604 // div/rem demand all inputs, because they don't want divide by zero.
1605 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
1606 UndefElts, Depth+1);
1607 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1608 TmpV = SimplifyDemandedVectorElts(I->getOperand(1), DemandedElts,
1609 UndefElts2, Depth+1);
1610 if (TmpV) { I->setOperand(1, TmpV); MadeChange = true; }
1612 // Output elements are undefined if both are undefined. Consider things
1613 // like undef&0. The result is known zero, not undef.
1614 UndefElts &= UndefElts2;
1617 case Instruction::Call: {
1618 IntrinsicInst *II = dyn_cast<IntrinsicInst>(I);
1620 switch (II->getIntrinsicID()) {
1623 // Binary vector operations that work column-wise. A dest element is a
1624 // function of the corresponding input elements from the two inputs.
1625 case Intrinsic::x86_sse_sub_ss:
1626 case Intrinsic::x86_sse_mul_ss:
1627 case Intrinsic::x86_sse_min_ss:
1628 case Intrinsic::x86_sse_max_ss:
1629 case Intrinsic::x86_sse2_sub_sd:
1630 case Intrinsic::x86_sse2_mul_sd:
1631 case Intrinsic::x86_sse2_min_sd:
1632 case Intrinsic::x86_sse2_max_sd:
1633 TmpV = SimplifyDemandedVectorElts(II->getOperand(1), DemandedElts,
1634 UndefElts, Depth+1);
1635 if (TmpV) { II->setOperand(1, TmpV); MadeChange = true; }
1636 TmpV = SimplifyDemandedVectorElts(II->getOperand(2), DemandedElts,
1637 UndefElts2, Depth+1);
1638 if (TmpV) { II->setOperand(2, TmpV); MadeChange = true; }
1640 // If only the low elt is demanded and this is a scalarizable intrinsic,
1641 // scalarize it now.
1642 if (DemandedElts == 1) {
1643 switch (II->getIntrinsicID()) {
1645 case Intrinsic::x86_sse_sub_ss:
1646 case Intrinsic::x86_sse_mul_ss:
1647 case Intrinsic::x86_sse2_sub_sd:
1648 case Intrinsic::x86_sse2_mul_sd:
1649 // TODO: Lower MIN/MAX/ABS/etc
1650 Value *LHS = II->getOperand(1);
1651 Value *RHS = II->getOperand(2);
1652 // Extract the element as scalars.
1653 LHS = InsertNewInstBefore(new ExtractElementInst(LHS, 0U,"tmp"), *II);
1654 RHS = InsertNewInstBefore(new ExtractElementInst(RHS, 0U,"tmp"), *II);
1656 switch (II->getIntrinsicID()) {
1657 default: assert(0 && "Case stmts out of sync!");
1658 case Intrinsic::x86_sse_sub_ss:
1659 case Intrinsic::x86_sse2_sub_sd:
1660 TmpV = InsertNewInstBefore(BinaryOperator::CreateSub(LHS, RHS,
1661 II->getName()), *II);
1663 case Intrinsic::x86_sse_mul_ss:
1664 case Intrinsic::x86_sse2_mul_sd:
1665 TmpV = InsertNewInstBefore(BinaryOperator::CreateMul(LHS, RHS,
1666 II->getName()), *II);
1671 InsertElementInst::Create(UndefValue::get(II->getType()), TmpV, 0U,
1673 InsertNewInstBefore(New, *II);
1674 AddSoonDeadInstToWorklist(*II, 0);
1679 // Output elements are undefined if both are undefined. Consider things
1680 // like undef&0. The result is known zero, not undef.
1681 UndefElts &= UndefElts2;
1687 return MadeChange ? I : 0;
1691 /// AssociativeOpt - Perform an optimization on an associative operator. This
1692 /// function is designed to check a chain of associative operators for a
1693 /// potential to apply a certain optimization. Since the optimization may be
1694 /// applicable if the expression was reassociated, this checks the chain, then
1695 /// reassociates the expression as necessary to expose the optimization
1696 /// opportunity. This makes use of a special Functor, which must define
1697 /// 'shouldApply' and 'apply' methods.
1699 template<typename Functor>
1700 static Instruction *AssociativeOpt(BinaryOperator &Root, const Functor &F) {
1701 unsigned Opcode = Root.getOpcode();
1702 Value *LHS = Root.getOperand(0);
1704 // Quick check, see if the immediate LHS matches...
1705 if (F.shouldApply(LHS))
1706 return F.apply(Root);
1708 // Otherwise, if the LHS is not of the same opcode as the root, return.
1709 Instruction *LHSI = dyn_cast<Instruction>(LHS);
1710 while (LHSI && LHSI->getOpcode() == Opcode && LHSI->hasOneUse()) {
1711 // Should we apply this transform to the RHS?
1712 bool ShouldApply = F.shouldApply(LHSI->getOperand(1));
1714 // If not to the RHS, check to see if we should apply to the LHS...
1715 if (!ShouldApply && F.shouldApply(LHSI->getOperand(0))) {
1716 cast<BinaryOperator>(LHSI)->swapOperands(); // Make the LHS the RHS
1720 // If the functor wants to apply the optimization to the RHS of LHSI,
1721 // reassociate the expression from ((? op A) op B) to (? op (A op B))
1723 // Now all of the instructions are in the current basic block, go ahead
1724 // and perform the reassociation.
1725 Instruction *TmpLHSI = cast<Instruction>(Root.getOperand(0));
1727 // First move the selected RHS to the LHS of the root...
1728 Root.setOperand(0, LHSI->getOperand(1));
1730 // Make what used to be the LHS of the root be the user of the root...
1731 Value *ExtraOperand = TmpLHSI->getOperand(1);
1732 if (&Root == TmpLHSI) {
1733 Root.replaceAllUsesWith(Constant::getNullValue(TmpLHSI->getType()));
1736 Root.replaceAllUsesWith(TmpLHSI); // Users now use TmpLHSI
1737 TmpLHSI->setOperand(1, &Root); // TmpLHSI now uses the root
1738 BasicBlock::iterator ARI = &Root; ++ARI;
1739 TmpLHSI->moveBefore(ARI); // Move TmpLHSI to after Root
1742 // Now propagate the ExtraOperand down the chain of instructions until we
1744 while (TmpLHSI != LHSI) {
1745 Instruction *NextLHSI = cast<Instruction>(TmpLHSI->getOperand(0));
1746 // Move the instruction to immediately before the chain we are
1747 // constructing to avoid breaking dominance properties.
1748 NextLHSI->moveBefore(ARI);
1751 Value *NextOp = NextLHSI->getOperand(1);
1752 NextLHSI->setOperand(1, ExtraOperand);
1754 ExtraOperand = NextOp;
1757 // Now that the instructions are reassociated, have the functor perform
1758 // the transformation...
1759 return F.apply(Root);
1762 LHSI = dyn_cast<Instruction>(LHSI->getOperand(0));
1769 // AddRHS - Implements: X + X --> X << 1
1772 AddRHS(Value *rhs) : RHS(rhs) {}
1773 bool shouldApply(Value *LHS) const { return LHS == RHS; }
1774 Instruction *apply(BinaryOperator &Add) const {
1775 return BinaryOperator::CreateShl(Add.getOperand(0),
1776 ConstantInt::get(Add.getType(), 1));
1780 // AddMaskingAnd - Implements (A & C1)+(B & C2) --> (A & C1)|(B & C2)
1782 struct AddMaskingAnd {
1784 AddMaskingAnd(Constant *c) : C2(c) {}
1785 bool shouldApply(Value *LHS) const {
1787 return match(LHS, m_And(m_Value(), m_ConstantInt(C1))) &&
1788 ConstantExpr::getAnd(C1, C2)->isNullValue();
1790 Instruction *apply(BinaryOperator &Add) const {
1791 return BinaryOperator::CreateOr(Add.getOperand(0), Add.getOperand(1));
1797 static Value *FoldOperationIntoSelectOperand(Instruction &I, Value *SO,
1799 if (CastInst *CI = dyn_cast<CastInst>(&I)) {
1800 return IC->InsertCastBefore(CI->getOpcode(), SO, I.getType(), I);
1803 // Figure out if the constant is the left or the right argument.
1804 bool ConstIsRHS = isa<Constant>(I.getOperand(1));
1805 Constant *ConstOperand = cast<Constant>(I.getOperand(ConstIsRHS));
1807 if (Constant *SOC = dyn_cast<Constant>(SO)) {
1809 return ConstantExpr::get(I.getOpcode(), SOC, ConstOperand);
1810 return ConstantExpr::get(I.getOpcode(), ConstOperand, SOC);
1813 Value *Op0 = SO, *Op1 = ConstOperand;
1815 std::swap(Op0, Op1);
1817 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
1818 New = BinaryOperator::Create(BO->getOpcode(), Op0, Op1,SO->getName()+".op");
1819 else if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1820 New = CmpInst::Create(CI->getOpcode(), CI->getPredicate(), Op0, Op1,
1821 SO->getName()+".cmp");
1823 assert(0 && "Unknown binary instruction type!");
1826 return IC->InsertNewInstBefore(New, I);
1829 // FoldOpIntoSelect - Given an instruction with a select as one operand and a
1830 // constant as the other operand, try to fold the binary operator into the
1831 // select arguments. This also works for Cast instructions, which obviously do
1832 // not have a second operand.
1833 static Instruction *FoldOpIntoSelect(Instruction &Op, SelectInst *SI,
1835 // Don't modify shared select instructions
1836 if (!SI->hasOneUse()) return 0;
1837 Value *TV = SI->getOperand(1);
1838 Value *FV = SI->getOperand(2);
1840 if (isa<Constant>(TV) || isa<Constant>(FV)) {
1841 // Bool selects with constant operands can be folded to logical ops.
1842 if (SI->getType() == Type::Int1Ty) return 0;
1844 Value *SelectTrueVal = FoldOperationIntoSelectOperand(Op, TV, IC);
1845 Value *SelectFalseVal = FoldOperationIntoSelectOperand(Op, FV, IC);
1847 return SelectInst::Create(SI->getCondition(), SelectTrueVal,
1854 /// FoldOpIntoPhi - Given a binary operator or cast instruction which has a PHI
1855 /// node as operand #0, see if we can fold the instruction into the PHI (which
1856 /// is only possible if all operands to the PHI are constants).
1857 Instruction *InstCombiner::FoldOpIntoPhi(Instruction &I) {
1858 PHINode *PN = cast<PHINode>(I.getOperand(0));
1859 unsigned NumPHIValues = PN->getNumIncomingValues();
1860 if (!PN->hasOneUse() || NumPHIValues == 0) return 0;
1862 // Check to see if all of the operands of the PHI are constants. If there is
1863 // one non-constant value, remember the BB it is. If there is more than one
1864 // or if *it* is a PHI, bail out.
1865 BasicBlock *NonConstBB = 0;
1866 for (unsigned i = 0; i != NumPHIValues; ++i)
1867 if (!isa<Constant>(PN->getIncomingValue(i))) {
1868 if (NonConstBB) return 0; // More than one non-const value.
1869 if (isa<PHINode>(PN->getIncomingValue(i))) return 0; // Itself a phi.
1870 NonConstBB = PN->getIncomingBlock(i);
1872 // If the incoming non-constant value is in I's block, we have an infinite
1874 if (NonConstBB == I.getParent())
1878 // If there is exactly one non-constant value, we can insert a copy of the
1879 // operation in that block. However, if this is a critical edge, we would be
1880 // inserting the computation one some other paths (e.g. inside a loop). Only
1881 // do this if the pred block is unconditionally branching into the phi block.
1883 BranchInst *BI = dyn_cast<BranchInst>(NonConstBB->getTerminator());
1884 if (!BI || !BI->isUnconditional()) return 0;
1887 // Okay, we can do the transformation: create the new PHI node.
1888 PHINode *NewPN = PHINode::Create(I.getType(), "");
1889 NewPN->reserveOperandSpace(PN->getNumOperands()/2);
1890 InsertNewInstBefore(NewPN, *PN);
1891 NewPN->takeName(PN);
1893 // Next, add all of the operands to the PHI.
1894 if (I.getNumOperands() == 2) {
1895 Constant *C = cast<Constant>(I.getOperand(1));
1896 for (unsigned i = 0; i != NumPHIValues; ++i) {
1898 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
1899 if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1900 InV = ConstantExpr::getCompare(CI->getPredicate(), InC, C);
1902 InV = ConstantExpr::get(I.getOpcode(), InC, C);
1904 assert(PN->getIncomingBlock(i) == NonConstBB);
1905 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
1906 InV = BinaryOperator::Create(BO->getOpcode(),
1907 PN->getIncomingValue(i), C, "phitmp",
1908 NonConstBB->getTerminator());
1909 else if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1910 InV = CmpInst::Create(CI->getOpcode(),
1912 PN->getIncomingValue(i), C, "phitmp",
1913 NonConstBB->getTerminator());
1915 assert(0 && "Unknown binop!");
1917 AddToWorkList(cast<Instruction>(InV));
1919 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
1922 CastInst *CI = cast<CastInst>(&I);
1923 const Type *RetTy = CI->getType();
1924 for (unsigned i = 0; i != NumPHIValues; ++i) {
1926 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
1927 InV = ConstantExpr::getCast(CI->getOpcode(), InC, RetTy);
1929 assert(PN->getIncomingBlock(i) == NonConstBB);
1930 InV = CastInst::Create(CI->getOpcode(), PN->getIncomingValue(i),
1931 I.getType(), "phitmp",
1932 NonConstBB->getTerminator());
1933 AddToWorkList(cast<Instruction>(InV));
1935 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
1938 return ReplaceInstUsesWith(I, NewPN);
1942 /// WillNotOverflowSignedAdd - Return true if we can prove that:
1943 /// (sext (add LHS, RHS)) === (add (sext LHS), (sext RHS))
1944 /// This basically requires proving that the add in the original type would not
1945 /// overflow to change the sign bit or have a carry out.
1946 bool InstCombiner::WillNotOverflowSignedAdd(Value *LHS, Value *RHS) {
1947 // There are different heuristics we can use for this. Here are some simple
1950 // Add has the property that adding any two 2's complement numbers can only
1951 // have one carry bit which can change a sign. As such, if LHS and RHS each
1952 // have at least two sign bits, we know that the addition of the two values will
1953 // sign extend fine.
1954 if (ComputeNumSignBits(LHS) > 1 && ComputeNumSignBits(RHS) > 1)
1958 // If one of the operands only has one non-zero bit, and if the other operand
1959 // has a known-zero bit in a more significant place than it (not including the
1960 // sign bit) the ripple may go up to and fill the zero, but won't change the
1961 // sign. For example, (X & ~4) + 1.
1969 Instruction *InstCombiner::visitAdd(BinaryOperator &I) {
1970 bool Changed = SimplifyCommutative(I);
1971 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
1973 if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
1974 // X + undef -> undef
1975 if (isa<UndefValue>(RHS))
1976 return ReplaceInstUsesWith(I, RHS);
1979 if (!I.getType()->isFPOrFPVector()) { // NOTE: -0 + +0 = +0.
1980 if (RHSC->isNullValue())
1981 return ReplaceInstUsesWith(I, LHS);
1982 } else if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHSC)) {
1983 if (CFP->isExactlyValue(ConstantFP::getNegativeZero
1984 (I.getType())->getValueAPF()))
1985 return ReplaceInstUsesWith(I, LHS);
1988 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHSC)) {
1989 // X + (signbit) --> X ^ signbit
1990 const APInt& Val = CI->getValue();
1991 uint32_t BitWidth = Val.getBitWidth();
1992 if (Val == APInt::getSignBit(BitWidth))
1993 return BinaryOperator::CreateXor(LHS, RHS);
1995 // See if SimplifyDemandedBits can simplify this. This handles stuff like
1996 // (X & 254)+1 -> (X&254)|1
1997 if (!isa<VectorType>(I.getType())) {
1998 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
1999 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
2000 KnownZero, KnownOne))
2004 // zext(i1) - 1 -> select i1, 0, -1
2005 if (ZExtInst *ZI = dyn_cast<ZExtInst>(LHS))
2006 if (CI->isAllOnesValue() &&
2007 ZI->getOperand(0)->getType() == Type::Int1Ty)
2008 return SelectInst::Create(ZI->getOperand(0),
2009 Constant::getNullValue(I.getType()),
2010 ConstantInt::getAllOnesValue(I.getType()));
2013 if (isa<PHINode>(LHS))
2014 if (Instruction *NV = FoldOpIntoPhi(I))
2017 ConstantInt *XorRHS = 0;
2019 if (isa<ConstantInt>(RHSC) &&
2020 match(LHS, m_Xor(m_Value(XorLHS), m_ConstantInt(XorRHS)))) {
2021 uint32_t TySizeBits = I.getType()->getPrimitiveSizeInBits();
2022 const APInt& RHSVal = cast<ConstantInt>(RHSC)->getValue();
2024 uint32_t Size = TySizeBits / 2;
2025 APInt C0080Val(APInt(TySizeBits, 1ULL).shl(Size - 1));
2026 APInt CFF80Val(-C0080Val);
2028 if (TySizeBits > Size) {
2029 // If we have ADD(XOR(AND(X, 0xFF), 0x80), 0xF..F80), it's a sext.
2030 // If we have ADD(XOR(AND(X, 0xFF), 0xF..F80), 0x80), it's a sext.
2031 if ((RHSVal == CFF80Val && XorRHS->getValue() == C0080Val) ||
2032 (RHSVal == C0080Val && XorRHS->getValue() == CFF80Val)) {
2033 // This is a sign extend if the top bits are known zero.
2034 if (!MaskedValueIsZero(XorLHS,
2035 APInt::getHighBitsSet(TySizeBits, TySizeBits - Size)))
2036 Size = 0; // Not a sign ext, but can't be any others either.
2041 C0080Val = APIntOps::lshr(C0080Val, Size);
2042 CFF80Val = APIntOps::ashr(CFF80Val, Size);
2043 } while (Size >= 1);
2045 // FIXME: This shouldn't be necessary. When the backends can handle types
2046 // with funny bit widths then this switch statement should be removed. It
2047 // is just here to get the size of the "middle" type back up to something
2048 // that the back ends can handle.
2049 const Type *MiddleType = 0;
2052 case 32: MiddleType = Type::Int32Ty; break;
2053 case 16: MiddleType = Type::Int16Ty; break;
2054 case 8: MiddleType = Type::Int8Ty; break;
2057 Instruction *NewTrunc = new TruncInst(XorLHS, MiddleType, "sext");
2058 InsertNewInstBefore(NewTrunc, I);
2059 return new SExtInst(NewTrunc, I.getType(), I.getName());
2064 if (I.getType() == Type::Int1Ty)
2065 return BinaryOperator::CreateXor(LHS, RHS);
2068 if (I.getType()->isInteger()) {
2069 if (Instruction *Result = AssociativeOpt(I, AddRHS(RHS))) return Result;
2071 if (Instruction *RHSI = dyn_cast<Instruction>(RHS)) {
2072 if (RHSI->getOpcode() == Instruction::Sub)
2073 if (LHS == RHSI->getOperand(1)) // A + (B - A) --> B
2074 return ReplaceInstUsesWith(I, RHSI->getOperand(0));
2076 if (Instruction *LHSI = dyn_cast<Instruction>(LHS)) {
2077 if (LHSI->getOpcode() == Instruction::Sub)
2078 if (RHS == LHSI->getOperand(1)) // (B - A) + A --> B
2079 return ReplaceInstUsesWith(I, LHSI->getOperand(0));
2084 // -A + -B --> -(A + B)
2085 if (Value *LHSV = dyn_castNegVal(LHS)) {
2086 if (LHS->getType()->isIntOrIntVector()) {
2087 if (Value *RHSV = dyn_castNegVal(RHS)) {
2088 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSV, RHSV, "sum");
2089 InsertNewInstBefore(NewAdd, I);
2090 return BinaryOperator::CreateNeg(NewAdd);
2094 return BinaryOperator::CreateSub(RHS, LHSV);
2098 if (!isa<Constant>(RHS))
2099 if (Value *V = dyn_castNegVal(RHS))
2100 return BinaryOperator::CreateSub(LHS, V);
2104 if (Value *X = dyn_castFoldableMul(LHS, C2)) {
2105 if (X == RHS) // X*C + X --> X * (C+1)
2106 return BinaryOperator::CreateMul(RHS, AddOne(C2));
2108 // X*C1 + X*C2 --> X * (C1+C2)
2110 if (X == dyn_castFoldableMul(RHS, C1))
2111 return BinaryOperator::CreateMul(X, Add(C1, C2));
2114 // X + X*C --> X * (C+1)
2115 if (dyn_castFoldableMul(RHS, C2) == LHS)
2116 return BinaryOperator::CreateMul(LHS, AddOne(C2));
2118 // X + ~X --> -1 since ~X = -X-1
2119 if (dyn_castNotVal(LHS) == RHS || dyn_castNotVal(RHS) == LHS)
2120 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
2123 // (A & C1)+(B & C2) --> (A & C1)|(B & C2) iff C1&C2 == 0
2124 if (match(RHS, m_And(m_Value(), m_ConstantInt(C2))))
2125 if (Instruction *R = AssociativeOpt(I, AddMaskingAnd(C2)))
2128 // A+B --> A|B iff A and B have no bits set in common.
2129 if (const IntegerType *IT = dyn_cast<IntegerType>(I.getType())) {
2130 APInt Mask = APInt::getAllOnesValue(IT->getBitWidth());
2131 APInt LHSKnownOne(IT->getBitWidth(), 0);
2132 APInt LHSKnownZero(IT->getBitWidth(), 0);
2133 ComputeMaskedBits(LHS, Mask, LHSKnownZero, LHSKnownOne);
2134 if (LHSKnownZero != 0) {
2135 APInt RHSKnownOne(IT->getBitWidth(), 0);
2136 APInt RHSKnownZero(IT->getBitWidth(), 0);
2137 ComputeMaskedBits(RHS, Mask, RHSKnownZero, RHSKnownOne);
2139 // No bits in common -> bitwise or.
2140 if ((LHSKnownZero|RHSKnownZero).isAllOnesValue())
2141 return BinaryOperator::CreateOr(LHS, RHS);
2145 // W*X + Y*Z --> W * (X+Z) iff W == Y
2146 if (I.getType()->isIntOrIntVector()) {
2147 Value *W, *X, *Y, *Z;
2148 if (match(LHS, m_Mul(m_Value(W), m_Value(X))) &&
2149 match(RHS, m_Mul(m_Value(Y), m_Value(Z)))) {
2153 } else if (Y == X) {
2155 } else if (X == Z) {
2162 Value *NewAdd = InsertNewInstBefore(BinaryOperator::CreateAdd(X, Z,
2163 LHS->getName()), I);
2164 return BinaryOperator::CreateMul(W, NewAdd);
2169 if (ConstantInt *CRHS = dyn_cast<ConstantInt>(RHS)) {
2171 if (match(LHS, m_Not(m_Value(X)))) // ~X + C --> (C-1) - X
2172 return BinaryOperator::CreateSub(SubOne(CRHS), X);
2174 // (X & FF00) + xx00 -> (X+xx00) & FF00
2175 if (LHS->hasOneUse() && match(LHS, m_And(m_Value(X), m_ConstantInt(C2)))) {
2176 Constant *Anded = And(CRHS, C2);
2177 if (Anded == CRHS) {
2178 // See if all bits from the first bit set in the Add RHS up are included
2179 // in the mask. First, get the rightmost bit.
2180 const APInt& AddRHSV = CRHS->getValue();
2182 // Form a mask of all bits from the lowest bit added through the top.
2183 APInt AddRHSHighBits(~((AddRHSV & -AddRHSV)-1));
2185 // See if the and mask includes all of these bits.
2186 APInt AddRHSHighBitsAnd(AddRHSHighBits & C2->getValue());
2188 if (AddRHSHighBits == AddRHSHighBitsAnd) {
2189 // Okay, the xform is safe. Insert the new add pronto.
2190 Value *NewAdd = InsertNewInstBefore(BinaryOperator::CreateAdd(X, CRHS,
2191 LHS->getName()), I);
2192 return BinaryOperator::CreateAnd(NewAdd, C2);
2197 // Try to fold constant add into select arguments.
2198 if (SelectInst *SI = dyn_cast<SelectInst>(LHS))
2199 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2203 // add (cast *A to intptrtype) B ->
2204 // cast (GEP (cast *A to sbyte*) B) --> intptrtype
2206 CastInst *CI = dyn_cast<CastInst>(LHS);
2209 CI = dyn_cast<CastInst>(RHS);
2212 if (CI && CI->getType()->isSized() &&
2213 (CI->getType()->getPrimitiveSizeInBits() ==
2214 TD->getIntPtrType()->getPrimitiveSizeInBits())
2215 && isa<PointerType>(CI->getOperand(0)->getType())) {
2217 cast<PointerType>(CI->getOperand(0)->getType())->getAddressSpace();
2218 Value *I2 = InsertBitCastBefore(CI->getOperand(0),
2219 PointerType::get(Type::Int8Ty, AS), I);
2220 I2 = InsertNewInstBefore(GetElementPtrInst::Create(I2, Other, "ctg2"), I);
2221 return new PtrToIntInst(I2, CI->getType());
2225 // add (select X 0 (sub n A)) A --> select X A n
2227 SelectInst *SI = dyn_cast<SelectInst>(LHS);
2230 SI = dyn_cast<SelectInst>(RHS);
2233 if (SI && SI->hasOneUse()) {
2234 Value *TV = SI->getTrueValue();
2235 Value *FV = SI->getFalseValue();
2238 // Can we fold the add into the argument of the select?
2239 // We check both true and false select arguments for a matching subtract.
2240 if (match(FV, m_Zero()) && match(TV, m_Sub(m_Value(N), m_Specific(A))))
2241 // Fold the add into the true select value.
2242 return SelectInst::Create(SI->getCondition(), N, A);
2243 if (match(TV, m_Zero()) && match(FV, m_Sub(m_Value(N), m_Specific(A))))
2244 // Fold the add into the false select value.
2245 return SelectInst::Create(SI->getCondition(), A, N);
2249 // Check for X+0.0. Simplify it to X if we know X is not -0.0.
2250 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHS))
2251 if (CFP->getValueAPF().isPosZero() && CannotBeNegativeZero(LHS))
2252 return ReplaceInstUsesWith(I, LHS);
2254 // Check for (add (sext x), y), see if we can merge this into an
2255 // integer add followed by a sext.
2256 if (SExtInst *LHSConv = dyn_cast<SExtInst>(LHS)) {
2257 // (add (sext x), cst) --> (sext (add x, cst'))
2258 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS)) {
2260 ConstantExpr::getTrunc(RHSC, LHSConv->getOperand(0)->getType());
2261 if (LHSConv->hasOneUse() &&
2262 ConstantExpr::getSExt(CI, I.getType()) == RHSC &&
2263 WillNotOverflowSignedAdd(LHSConv->getOperand(0), CI)) {
2264 // Insert the new, smaller add.
2265 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSConv->getOperand(0),
2267 InsertNewInstBefore(NewAdd, I);
2268 return new SExtInst(NewAdd, I.getType());
2272 // (add (sext x), (sext y)) --> (sext (add int x, y))
2273 if (SExtInst *RHSConv = dyn_cast<SExtInst>(RHS)) {
2274 // Only do this if x/y have the same type, if at last one of them has a
2275 // single use (so we don't increase the number of sexts), and if the
2276 // integer add will not overflow.
2277 if (LHSConv->getOperand(0)->getType()==RHSConv->getOperand(0)->getType()&&
2278 (LHSConv->hasOneUse() || RHSConv->hasOneUse()) &&
2279 WillNotOverflowSignedAdd(LHSConv->getOperand(0),
2280 RHSConv->getOperand(0))) {
2281 // Insert the new integer add.
2282 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSConv->getOperand(0),
2283 RHSConv->getOperand(0),
2285 InsertNewInstBefore(NewAdd, I);
2286 return new SExtInst(NewAdd, I.getType());
2291 // Check for (add double (sitofp x), y), see if we can merge this into an
2292 // integer add followed by a promotion.
2293 if (SIToFPInst *LHSConv = dyn_cast<SIToFPInst>(LHS)) {
2294 // (add double (sitofp x), fpcst) --> (sitofp (add int x, intcst))
2295 // ... if the constant fits in the integer value. This is useful for things
2296 // like (double)(x & 1234) + 4.0 -> (double)((X & 1234)+4) which no longer
2297 // requires a constant pool load, and generally allows the add to be better
2299 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHS)) {
2301 ConstantExpr::getFPToSI(CFP, LHSConv->getOperand(0)->getType());
2302 if (LHSConv->hasOneUse() &&
2303 ConstantExpr::getSIToFP(CI, I.getType()) == CFP &&
2304 WillNotOverflowSignedAdd(LHSConv->getOperand(0), CI)) {
2305 // Insert the new integer add.
2306 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSConv->getOperand(0),
2308 InsertNewInstBefore(NewAdd, I);
2309 return new SIToFPInst(NewAdd, I.getType());
2313 // (add double (sitofp x), (sitofp y)) --> (sitofp (add int x, y))
2314 if (SIToFPInst *RHSConv = dyn_cast<SIToFPInst>(RHS)) {
2315 // Only do this if x/y have the same type, if at last one of them has a
2316 // single use (so we don't increase the number of int->fp conversions),
2317 // and if the integer add will not overflow.
2318 if (LHSConv->getOperand(0)->getType()==RHSConv->getOperand(0)->getType()&&
2319 (LHSConv->hasOneUse() || RHSConv->hasOneUse()) &&
2320 WillNotOverflowSignedAdd(LHSConv->getOperand(0),
2321 RHSConv->getOperand(0))) {
2322 // Insert the new integer add.
2323 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSConv->getOperand(0),
2324 RHSConv->getOperand(0),
2326 InsertNewInstBefore(NewAdd, I);
2327 return new SIToFPInst(NewAdd, I.getType());
2332 return Changed ? &I : 0;
2335 Instruction *InstCombiner::visitSub(BinaryOperator &I) {
2336 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2338 if (Op0 == Op1 && // sub X, X -> 0
2339 !I.getType()->isFPOrFPVector())
2340 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2342 // If this is a 'B = x-(-A)', change to B = x+A...
2343 if (Value *V = dyn_castNegVal(Op1))
2344 return BinaryOperator::CreateAdd(Op0, V);
2346 if (isa<UndefValue>(Op0))
2347 return ReplaceInstUsesWith(I, Op0); // undef - X -> undef
2348 if (isa<UndefValue>(Op1))
2349 return ReplaceInstUsesWith(I, Op1); // X - undef -> undef
2351 if (ConstantInt *C = dyn_cast<ConstantInt>(Op0)) {
2352 // Replace (-1 - A) with (~A)...
2353 if (C->isAllOnesValue())
2354 return BinaryOperator::CreateNot(Op1);
2356 // C - ~X == X + (1+C)
2358 if (match(Op1, m_Not(m_Value(X))))
2359 return BinaryOperator::CreateAdd(X, AddOne(C));
2361 // -(X >>u 31) -> (X >>s 31)
2362 // -(X >>s 31) -> (X >>u 31)
2364 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op1)) {
2365 if (SI->getOpcode() == Instruction::LShr) {
2366 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
2367 // Check to see if we are shifting out everything but the sign bit.
2368 if (CU->getLimitedValue(SI->getType()->getPrimitiveSizeInBits()) ==
2369 SI->getType()->getPrimitiveSizeInBits()-1) {
2370 // Ok, the transformation is safe. Insert AShr.
2371 return BinaryOperator::Create(Instruction::AShr,
2372 SI->getOperand(0), CU, SI->getName());
2376 else if (SI->getOpcode() == Instruction::AShr) {
2377 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
2378 // Check to see if we are shifting out everything but the sign bit.
2379 if (CU->getLimitedValue(SI->getType()->getPrimitiveSizeInBits()) ==
2380 SI->getType()->getPrimitiveSizeInBits()-1) {
2381 // Ok, the transformation is safe. Insert LShr.
2382 return BinaryOperator::CreateLShr(
2383 SI->getOperand(0), CU, SI->getName());
2390 // Try to fold constant sub into select arguments.
2391 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
2392 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2396 if (I.getType() == Type::Int1Ty)
2397 return BinaryOperator::CreateXor(Op0, Op1);
2399 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
2400 if (Op1I->getOpcode() == Instruction::Add &&
2401 !Op0->getType()->isFPOrFPVector()) {
2402 if (Op1I->getOperand(0) == Op0) // X-(X+Y) == -Y
2403 return BinaryOperator::CreateNeg(Op1I->getOperand(1), I.getName());
2404 else if (Op1I->getOperand(1) == Op0) // X-(Y+X) == -Y
2405 return BinaryOperator::CreateNeg(Op1I->getOperand(0), I.getName());
2406 else if (ConstantInt *CI1 = dyn_cast<ConstantInt>(I.getOperand(0))) {
2407 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(Op1I->getOperand(1)))
2408 // C1-(X+C2) --> (C1-C2)-X
2409 return BinaryOperator::CreateSub(Subtract(CI1, CI2),
2410 Op1I->getOperand(0));
2414 if (Op1I->hasOneUse()) {
2415 // Replace (x - (y - z)) with (x + (z - y)) if the (y - z) subexpression
2416 // is not used by anyone else...
2418 if (Op1I->getOpcode() == Instruction::Sub &&
2419 !Op1I->getType()->isFPOrFPVector()) {
2420 // Swap the two operands of the subexpr...
2421 Value *IIOp0 = Op1I->getOperand(0), *IIOp1 = Op1I->getOperand(1);
2422 Op1I->setOperand(0, IIOp1);
2423 Op1I->setOperand(1, IIOp0);
2425 // Create the new top level add instruction...
2426 return BinaryOperator::CreateAdd(Op0, Op1);
2429 // Replace (A - (A & B)) with (A & ~B) if this is the only use of (A&B)...
2431 if (Op1I->getOpcode() == Instruction::And &&
2432 (Op1I->getOperand(0) == Op0 || Op1I->getOperand(1) == Op0)) {
2433 Value *OtherOp = Op1I->getOperand(Op1I->getOperand(0) == Op0);
2436 InsertNewInstBefore(BinaryOperator::CreateNot(OtherOp, "B.not"), I);
2437 return BinaryOperator::CreateAnd(Op0, NewNot);
2440 // 0 - (X sdiv C) -> (X sdiv -C)
2441 if (Op1I->getOpcode() == Instruction::SDiv)
2442 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
2444 if (Constant *DivRHS = dyn_cast<Constant>(Op1I->getOperand(1)))
2445 return BinaryOperator::CreateSDiv(Op1I->getOperand(0),
2446 ConstantExpr::getNeg(DivRHS));
2448 // X - X*C --> X * (1-C)
2449 ConstantInt *C2 = 0;
2450 if (dyn_castFoldableMul(Op1I, C2) == Op0) {
2451 Constant *CP1 = Subtract(ConstantInt::get(I.getType(), 1), C2);
2452 return BinaryOperator::CreateMul(Op0, CP1);
2457 if (!Op0->getType()->isFPOrFPVector())
2458 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
2459 if (Op0I->getOpcode() == Instruction::Add) {
2460 if (Op0I->getOperand(0) == Op1) // (Y+X)-Y == X
2461 return ReplaceInstUsesWith(I, Op0I->getOperand(1));
2462 else if (Op0I->getOperand(1) == Op1) // (X+Y)-Y == X
2463 return ReplaceInstUsesWith(I, Op0I->getOperand(0));
2464 } else if (Op0I->getOpcode() == Instruction::Sub) {
2465 if (Op0I->getOperand(0) == Op1) // (X-Y)-X == -Y
2466 return BinaryOperator::CreateNeg(Op0I->getOperand(1), I.getName());
2471 if (Value *X = dyn_castFoldableMul(Op0, C1)) {
2472 if (X == Op1) // X*C - X --> X * (C-1)
2473 return BinaryOperator::CreateMul(Op1, SubOne(C1));
2475 ConstantInt *C2; // X*C1 - X*C2 -> X * (C1-C2)
2476 if (X == dyn_castFoldableMul(Op1, C2))
2477 return BinaryOperator::CreateMul(X, Subtract(C1, C2));
2482 /// isSignBitCheck - Given an exploded icmp instruction, return true if the
2483 /// comparison only checks the sign bit. If it only checks the sign bit, set
2484 /// TrueIfSigned if the result of the comparison is true when the input value is
2486 static bool isSignBitCheck(ICmpInst::Predicate pred, ConstantInt *RHS,
2487 bool &TrueIfSigned) {
2489 case ICmpInst::ICMP_SLT: // True if LHS s< 0
2490 TrueIfSigned = true;
2491 return RHS->isZero();
2492 case ICmpInst::ICMP_SLE: // True if LHS s<= RHS and RHS == -1
2493 TrueIfSigned = true;
2494 return RHS->isAllOnesValue();
2495 case ICmpInst::ICMP_SGT: // True if LHS s> -1
2496 TrueIfSigned = false;
2497 return RHS->isAllOnesValue();
2498 case ICmpInst::ICMP_UGT:
2499 // True if LHS u> RHS and RHS == high-bit-mask - 1
2500 TrueIfSigned = true;
2501 return RHS->getValue() ==
2502 APInt::getSignedMaxValue(RHS->getType()->getPrimitiveSizeInBits());
2503 case ICmpInst::ICMP_UGE:
2504 // True if LHS u>= RHS and RHS == high-bit-mask (2^7, 2^15, 2^31, etc)
2505 TrueIfSigned = true;
2506 return RHS->getValue().isSignBit();
2512 Instruction *InstCombiner::visitMul(BinaryOperator &I) {
2513 bool Changed = SimplifyCommutative(I);
2514 Value *Op0 = I.getOperand(0);
2516 if (isa<UndefValue>(I.getOperand(1))) // undef * X -> 0
2517 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2519 // Simplify mul instructions with a constant RHS...
2520 if (Constant *Op1 = dyn_cast<Constant>(I.getOperand(1))) {
2521 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2523 // ((X << C1)*C2) == (X * (C2 << C1))
2524 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op0))
2525 if (SI->getOpcode() == Instruction::Shl)
2526 if (Constant *ShOp = dyn_cast<Constant>(SI->getOperand(1)))
2527 return BinaryOperator::CreateMul(SI->getOperand(0),
2528 ConstantExpr::getShl(CI, ShOp));
2531 return ReplaceInstUsesWith(I, Op1); // X * 0 == 0
2532 if (CI->equalsInt(1)) // X * 1 == X
2533 return ReplaceInstUsesWith(I, Op0);
2534 if (CI->isAllOnesValue()) // X * -1 == 0 - X
2535 return BinaryOperator::CreateNeg(Op0, I.getName());
2537 const APInt& Val = cast<ConstantInt>(CI)->getValue();
2538 if (Val.isPowerOf2()) { // Replace X*(2^C) with X << C
2539 return BinaryOperator::CreateShl(Op0,
2540 ConstantInt::get(Op0->getType(), Val.logBase2()));
2542 } else if (ConstantFP *Op1F = dyn_cast<ConstantFP>(Op1)) {
2543 if (Op1F->isNullValue())
2544 return ReplaceInstUsesWith(I, Op1);
2546 // "In IEEE floating point, x*1 is not equivalent to x for nans. However,
2547 // ANSI says we can drop signals, so we can do this anyway." (from GCC)
2548 if (Op1F->isExactlyValue(1.0))
2549 return ReplaceInstUsesWith(I, Op0); // Eliminate 'mul double %X, 1.0'
2550 } else if (isa<VectorType>(Op1->getType())) {
2551 if (isa<ConstantAggregateZero>(Op1))
2552 return ReplaceInstUsesWith(I, Op1);
2554 if (ConstantVector *Op1V = dyn_cast<ConstantVector>(Op1)) {
2555 if (Op1V->isAllOnesValue()) // X * -1 == 0 - X
2556 return BinaryOperator::CreateNeg(Op0, I.getName());
2558 // As above, vector X*splat(1.0) -> X in all defined cases.
2559 if (Constant *Splat = Op1V->getSplatValue()) {
2560 if (ConstantFP *F = dyn_cast<ConstantFP>(Splat))
2561 if (F->isExactlyValue(1.0))
2562 return ReplaceInstUsesWith(I, Op0);
2563 if (ConstantInt *CI = dyn_cast<ConstantInt>(Splat))
2564 if (CI->equalsInt(1))
2565 return ReplaceInstUsesWith(I, Op0);
2570 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
2571 if (Op0I->getOpcode() == Instruction::Add && Op0I->hasOneUse() &&
2572 isa<ConstantInt>(Op0I->getOperand(1)) && isa<ConstantInt>(Op1)) {
2573 // Canonicalize (X+C1)*C2 -> X*C2+C1*C2.
2574 Instruction *Add = BinaryOperator::CreateMul(Op0I->getOperand(0),
2576 InsertNewInstBefore(Add, I);
2577 Value *C1C2 = ConstantExpr::getMul(Op1,
2578 cast<Constant>(Op0I->getOperand(1)));
2579 return BinaryOperator::CreateAdd(Add, C1C2);
2583 // Try to fold constant mul into select arguments.
2584 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2585 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2588 if (isa<PHINode>(Op0))
2589 if (Instruction *NV = FoldOpIntoPhi(I))
2593 if (Value *Op0v = dyn_castNegVal(Op0)) // -X * -Y = X*Y
2594 if (Value *Op1v = dyn_castNegVal(I.getOperand(1)))
2595 return BinaryOperator::CreateMul(Op0v, Op1v);
2597 // (X / Y) * Y = X - (X % Y)
2598 // (X / Y) * -Y = (X % Y) - X
2600 Value *Op1 = I.getOperand(1);
2601 BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0);
2603 (BO->getOpcode() != Instruction::UDiv &&
2604 BO->getOpcode() != Instruction::SDiv)) {
2606 BO = dyn_cast<BinaryOperator>(I.getOperand(1));
2608 Value *Neg = dyn_castNegVal(Op1);
2609 if (BO && BO->hasOneUse() &&
2610 (BO->getOperand(1) == Op1 || BO->getOperand(1) == Neg) &&
2611 (BO->getOpcode() == Instruction::UDiv ||
2612 BO->getOpcode() == Instruction::SDiv)) {
2613 Value *Op0BO = BO->getOperand(0), *Op1BO = BO->getOperand(1);
2616 if (BO->getOpcode() == Instruction::UDiv)
2617 Rem = BinaryOperator::CreateURem(Op0BO, Op1BO);
2619 Rem = BinaryOperator::CreateSRem(Op0BO, Op1BO);
2621 InsertNewInstBefore(Rem, I);
2625 return BinaryOperator::CreateSub(Op0BO, Rem);
2627 return BinaryOperator::CreateSub(Rem, Op0BO);
2631 if (I.getType() == Type::Int1Ty)
2632 return BinaryOperator::CreateAnd(Op0, I.getOperand(1));
2634 // If one of the operands of the multiply is a cast from a boolean value, then
2635 // we know the bool is either zero or one, so this is a 'masking' multiply.
2636 // See if we can simplify things based on how the boolean was originally
2638 CastInst *BoolCast = 0;
2639 if (ZExtInst *CI = dyn_cast<ZExtInst>(Op0))
2640 if (CI->getOperand(0)->getType() == Type::Int1Ty)
2643 if (ZExtInst *CI = dyn_cast<ZExtInst>(I.getOperand(1)))
2644 if (CI->getOperand(0)->getType() == Type::Int1Ty)
2647 if (ICmpInst *SCI = dyn_cast<ICmpInst>(BoolCast->getOperand(0))) {
2648 Value *SCIOp0 = SCI->getOperand(0), *SCIOp1 = SCI->getOperand(1);
2649 const Type *SCOpTy = SCIOp0->getType();
2652 // If the icmp is true iff the sign bit of X is set, then convert this
2653 // multiply into a shift/and combination.
2654 if (isa<ConstantInt>(SCIOp1) &&
2655 isSignBitCheck(SCI->getPredicate(), cast<ConstantInt>(SCIOp1), TIS) &&
2657 // Shift the X value right to turn it into "all signbits".
2658 Constant *Amt = ConstantInt::get(SCIOp0->getType(),
2659 SCOpTy->getPrimitiveSizeInBits()-1);
2661 InsertNewInstBefore(
2662 BinaryOperator::Create(Instruction::AShr, SCIOp0, Amt,
2663 BoolCast->getOperand(0)->getName()+
2666 // If the multiply type is not the same as the source type, sign extend
2667 // or truncate to the multiply type.
2668 if (I.getType() != V->getType()) {
2669 uint32_t SrcBits = V->getType()->getPrimitiveSizeInBits();
2670 uint32_t DstBits = I.getType()->getPrimitiveSizeInBits();
2671 Instruction::CastOps opcode =
2672 (SrcBits == DstBits ? Instruction::BitCast :
2673 (SrcBits < DstBits ? Instruction::SExt : Instruction::Trunc));
2674 V = InsertCastBefore(opcode, V, I.getType(), I);
2677 Value *OtherOp = Op0 == BoolCast ? I.getOperand(1) : Op0;
2678 return BinaryOperator::CreateAnd(V, OtherOp);
2683 return Changed ? &I : 0;
2686 /// SimplifyDivRemOfSelect - Try to fold a divide or remainder of a select
2688 bool InstCombiner::SimplifyDivRemOfSelect(BinaryOperator &I) {
2689 SelectInst *SI = cast<SelectInst>(I.getOperand(1));
2691 // div/rem X, (Cond ? 0 : Y) -> div/rem X, Y
2692 int NonNullOperand = -1;
2693 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(1)))
2694 if (ST->isNullValue())
2696 // div/rem X, (Cond ? Y : 0) -> div/rem X, Y
2697 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(2)))
2698 if (ST->isNullValue())
2701 if (NonNullOperand == -1)
2704 Value *SelectCond = SI->getOperand(0);
2706 // Change the div/rem to use 'Y' instead of the select.
2707 I.setOperand(1, SI->getOperand(NonNullOperand));
2709 // Okay, we know we replace the operand of the div/rem with 'Y' with no
2710 // problem. However, the select, or the condition of the select may have
2711 // multiple uses. Based on our knowledge that the operand must be non-zero,
2712 // propagate the known value for the select into other uses of it, and
2713 // propagate a known value of the condition into its other users.
2715 // If the select and condition only have a single use, don't bother with this,
2717 if (SI->use_empty() && SelectCond->hasOneUse())
2720 // Scan the current block backward, looking for other uses of SI.
2721 BasicBlock::iterator BBI = &I, BBFront = I.getParent()->begin();
2723 while (BBI != BBFront) {
2725 // If we found a call to a function, we can't assume it will return, so
2726 // information from below it cannot be propagated above it.
2727 if (isa<CallInst>(BBI) && !isa<IntrinsicInst>(BBI))
2730 // Replace uses of the select or its condition with the known values.
2731 for (Instruction::op_iterator I = BBI->op_begin(), E = BBI->op_end();
2734 *I = SI->getOperand(NonNullOperand);
2736 } else if (*I == SelectCond) {
2737 *I = NonNullOperand == 1 ? ConstantInt::getTrue() :
2738 ConstantInt::getFalse();
2743 // If we past the instruction, quit looking for it.
2746 if (&*BBI == SelectCond)
2749 // If we ran out of things to eliminate, break out of the loop.
2750 if (SelectCond == 0 && SI == 0)
2758 /// This function implements the transforms on div instructions that work
2759 /// regardless of the kind of div instruction it is (udiv, sdiv, or fdiv). It is
2760 /// used by the visitors to those instructions.
2761 /// @brief Transforms common to all three div instructions
2762 Instruction *InstCombiner::commonDivTransforms(BinaryOperator &I) {
2763 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2765 // undef / X -> 0 for integer.
2766 // undef / X -> undef for FP (the undef could be a snan).
2767 if (isa<UndefValue>(Op0)) {
2768 if (Op0->getType()->isFPOrFPVector())
2769 return ReplaceInstUsesWith(I, Op0);
2770 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2773 // X / undef -> undef
2774 if (isa<UndefValue>(Op1))
2775 return ReplaceInstUsesWith(I, Op1);
2780 /// This function implements the transforms common to both integer division
2781 /// instructions (udiv and sdiv). It is called by the visitors to those integer
2782 /// division instructions.
2783 /// @brief Common integer divide transforms
2784 Instruction *InstCombiner::commonIDivTransforms(BinaryOperator &I) {
2785 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2787 // (sdiv X, X) --> 1 (udiv X, X) --> 1
2789 if (const VectorType *Ty = dyn_cast<VectorType>(I.getType())) {
2790 ConstantInt *CI = ConstantInt::get(Ty->getElementType(), 1);
2791 std::vector<Constant*> Elts(Ty->getNumElements(), CI);
2792 return ReplaceInstUsesWith(I, ConstantVector::get(Elts));
2795 ConstantInt *CI = ConstantInt::get(I.getType(), 1);
2796 return ReplaceInstUsesWith(I, CI);
2799 if (Instruction *Common = commonDivTransforms(I))
2802 // Handle cases involving: [su]div X, (select Cond, Y, Z)
2803 // This does not apply for fdiv.
2804 if (isa<SelectInst>(Op1) && SimplifyDivRemOfSelect(I))
2807 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2809 if (RHS->equalsInt(1))
2810 return ReplaceInstUsesWith(I, Op0);
2812 // (X / C1) / C2 -> X / (C1*C2)
2813 if (Instruction *LHS = dyn_cast<Instruction>(Op0))
2814 if (Instruction::BinaryOps(LHS->getOpcode()) == I.getOpcode())
2815 if (ConstantInt *LHSRHS = dyn_cast<ConstantInt>(LHS->getOperand(1))) {
2816 if (MultiplyOverflows(RHS, LHSRHS, I.getOpcode()==Instruction::SDiv))
2817 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2819 return BinaryOperator::Create(I.getOpcode(), LHS->getOperand(0),
2820 Multiply(RHS, LHSRHS));
2823 if (!RHS->isZero()) { // avoid X udiv 0
2824 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2825 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2827 if (isa<PHINode>(Op0))
2828 if (Instruction *NV = FoldOpIntoPhi(I))
2833 // 0 / X == 0, we don't need to preserve faults!
2834 if (ConstantInt *LHS = dyn_cast<ConstantInt>(Op0))
2835 if (LHS->equalsInt(0))
2836 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2838 // It can't be division by zero, hence it must be division by one.
2839 if (I.getType() == Type::Int1Ty)
2840 return ReplaceInstUsesWith(I, Op0);
2842 if (ConstantVector *Op1V = dyn_cast<ConstantVector>(Op1)) {
2843 if (ConstantInt *X = cast_or_null<ConstantInt>(Op1V->getSplatValue()))
2846 return ReplaceInstUsesWith(I, Op0);
2852 Instruction *InstCombiner::visitUDiv(BinaryOperator &I) {
2853 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2855 // Handle the integer div common cases
2856 if (Instruction *Common = commonIDivTransforms(I))
2859 if (ConstantInt *C = dyn_cast<ConstantInt>(Op1)) {
2860 // X udiv C^2 -> X >> C
2861 // Check to see if this is an unsigned division with an exact power of 2,
2862 // if so, convert to a right shift.
2863 if (C->getValue().isPowerOf2()) // 0 not included in isPowerOf2
2864 return BinaryOperator::CreateLShr(Op0,
2865 ConstantInt::get(Op0->getType(), C->getValue().logBase2()));
2867 // X udiv C, where C >= signbit
2868 if (C->getValue().isNegative()) {
2869 Value *IC = InsertNewInstBefore(new ICmpInst(ICmpInst::ICMP_ULT, Op0, C),
2871 return SelectInst::Create(IC, Constant::getNullValue(I.getType()),
2872 ConstantInt::get(I.getType(), 1));
2876 // X udiv (C1 << N), where C1 is "1<<C2" --> X >> (N+C2)
2877 if (BinaryOperator *RHSI = dyn_cast<BinaryOperator>(I.getOperand(1))) {
2878 if (RHSI->getOpcode() == Instruction::Shl &&
2879 isa<ConstantInt>(RHSI->getOperand(0))) {
2880 const APInt& C1 = cast<ConstantInt>(RHSI->getOperand(0))->getValue();
2881 if (C1.isPowerOf2()) {
2882 Value *N = RHSI->getOperand(1);
2883 const Type *NTy = N->getType();
2884 if (uint32_t C2 = C1.logBase2()) {
2885 Constant *C2V = ConstantInt::get(NTy, C2);
2886 N = InsertNewInstBefore(BinaryOperator::CreateAdd(N, C2V, "tmp"), I);
2888 return BinaryOperator::CreateLShr(Op0, N);
2893 // udiv X, (Select Cond, C1, C2) --> Select Cond, (shr X, C1), (shr X, C2)
2894 // where C1&C2 are powers of two.
2895 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
2896 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
2897 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
2898 const APInt &TVA = STO->getValue(), &FVA = SFO->getValue();
2899 if (TVA.isPowerOf2() && FVA.isPowerOf2()) {
2900 // Compute the shift amounts
2901 uint32_t TSA = TVA.logBase2(), FSA = FVA.logBase2();
2902 // Construct the "on true" case of the select
2903 Constant *TC = ConstantInt::get(Op0->getType(), TSA);
2904 Instruction *TSI = BinaryOperator::CreateLShr(
2905 Op0, TC, SI->getName()+".t");
2906 TSI = InsertNewInstBefore(TSI, I);
2908 // Construct the "on false" case of the select
2909 Constant *FC = ConstantInt::get(Op0->getType(), FSA);
2910 Instruction *FSI = BinaryOperator::CreateLShr(
2911 Op0, FC, SI->getName()+".f");
2912 FSI = InsertNewInstBefore(FSI, I);
2914 // construct the select instruction and return it.
2915 return SelectInst::Create(SI->getOperand(0), TSI, FSI, SI->getName());
2921 Instruction *InstCombiner::visitSDiv(BinaryOperator &I) {
2922 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2924 // Handle the integer div common cases
2925 if (Instruction *Common = commonIDivTransforms(I))
2928 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2930 if (RHS->isAllOnesValue())
2931 return BinaryOperator::CreateNeg(Op0);
2934 // If the sign bits of both operands are zero (i.e. we can prove they are
2935 // unsigned inputs), turn this into a udiv.
2936 if (I.getType()->isInteger()) {
2937 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()));
2938 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
2939 // X sdiv Y -> X udiv Y, iff X and Y don't have sign bit set
2940 return BinaryOperator::CreateUDiv(Op0, Op1, I.getName());
2947 Instruction *InstCombiner::visitFDiv(BinaryOperator &I) {
2948 return commonDivTransforms(I);
2951 /// This function implements the transforms on rem instructions that work
2952 /// regardless of the kind of rem instruction it is (urem, srem, or frem). It
2953 /// is used by the visitors to those instructions.
2954 /// @brief Transforms common to all three rem instructions
2955 Instruction *InstCombiner::commonRemTransforms(BinaryOperator &I) {
2956 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2958 // 0 % X == 0 for integer, we don't need to preserve faults!
2959 if (Constant *LHS = dyn_cast<Constant>(Op0))
2960 if (LHS->isNullValue())
2961 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2963 if (isa<UndefValue>(Op0)) { // undef % X -> 0
2964 if (I.getType()->isFPOrFPVector())
2965 return ReplaceInstUsesWith(I, Op0); // X % undef -> undef (could be SNaN)
2966 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2968 if (isa<UndefValue>(Op1))
2969 return ReplaceInstUsesWith(I, Op1); // X % undef -> undef
2971 // Handle cases involving: rem X, (select Cond, Y, Z)
2972 if (isa<SelectInst>(Op1) && SimplifyDivRemOfSelect(I))
2978 /// This function implements the transforms common to both integer remainder
2979 /// instructions (urem and srem). It is called by the visitors to those integer
2980 /// remainder instructions.
2981 /// @brief Common integer remainder transforms
2982 Instruction *InstCombiner::commonIRemTransforms(BinaryOperator &I) {
2983 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2985 if (Instruction *common = commonRemTransforms(I))
2988 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2989 // X % 0 == undef, we don't need to preserve faults!
2990 if (RHS->equalsInt(0))
2991 return ReplaceInstUsesWith(I, UndefValue::get(I.getType()));
2993 if (RHS->equalsInt(1)) // X % 1 == 0
2994 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2996 if (Instruction *Op0I = dyn_cast<Instruction>(Op0)) {
2997 if (SelectInst *SI = dyn_cast<SelectInst>(Op0I)) {
2998 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3000 } else if (isa<PHINode>(Op0I)) {
3001 if (Instruction *NV = FoldOpIntoPhi(I))
3005 // See if we can fold away this rem instruction.
3006 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
3007 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
3008 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
3009 KnownZero, KnownOne))
3017 Instruction *InstCombiner::visitURem(BinaryOperator &I) {
3018 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3020 if (Instruction *common = commonIRemTransforms(I))
3023 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3024 // X urem C^2 -> X and C
3025 // Check to see if this is an unsigned remainder with an exact power of 2,
3026 // if so, convert to a bitwise and.
3027 if (ConstantInt *C = dyn_cast<ConstantInt>(RHS))
3028 if (C->getValue().isPowerOf2())
3029 return BinaryOperator::CreateAnd(Op0, SubOne(C));
3032 if (Instruction *RHSI = dyn_cast<Instruction>(I.getOperand(1))) {
3033 // Turn A % (C << N), where C is 2^k, into A & ((C << N)-1)
3034 if (RHSI->getOpcode() == Instruction::Shl &&
3035 isa<ConstantInt>(RHSI->getOperand(0))) {
3036 if (cast<ConstantInt>(RHSI->getOperand(0))->getValue().isPowerOf2()) {
3037 Constant *N1 = ConstantInt::getAllOnesValue(I.getType());
3038 Value *Add = InsertNewInstBefore(BinaryOperator::CreateAdd(RHSI, N1,
3040 return BinaryOperator::CreateAnd(Op0, Add);
3045 // urem X, (select Cond, 2^C1, 2^C2) --> select Cond, (and X, C1), (and X, C2)
3046 // where C1&C2 are powers of two.
3047 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
3048 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
3049 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
3050 // STO == 0 and SFO == 0 handled above.
3051 if ((STO->getValue().isPowerOf2()) &&
3052 (SFO->getValue().isPowerOf2())) {
3053 Value *TrueAnd = InsertNewInstBefore(
3054 BinaryOperator::CreateAnd(Op0, SubOne(STO), SI->getName()+".t"), I);
3055 Value *FalseAnd = InsertNewInstBefore(
3056 BinaryOperator::CreateAnd(Op0, SubOne(SFO), SI->getName()+".f"), I);
3057 return SelectInst::Create(SI->getOperand(0), TrueAnd, FalseAnd);
3065 Instruction *InstCombiner::visitSRem(BinaryOperator &I) {
3066 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3068 // Handle the integer rem common cases
3069 if (Instruction *common = commonIRemTransforms(I))
3072 if (Value *RHSNeg = dyn_castNegVal(Op1))
3073 if (!isa<Constant>(RHSNeg) ||
3074 (isa<ConstantInt>(RHSNeg) &&
3075 cast<ConstantInt>(RHSNeg)->getValue().isStrictlyPositive())) {
3077 AddUsesToWorkList(I);
3078 I.setOperand(1, RHSNeg);
3082 // If the sign bits of both operands are zero (i.e. we can prove they are
3083 // unsigned inputs), turn this into a urem.
3084 if (I.getType()->isInteger()) {
3085 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()));
3086 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
3087 // X srem Y -> X urem Y, iff X and Y don't have sign bit set
3088 return BinaryOperator::CreateURem(Op0, Op1, I.getName());
3092 // If it's a constant vector, flip any negative values positive.
3093 if (ConstantVector *RHSV = dyn_cast<ConstantVector>(Op1)) {
3094 unsigned VWidth = RHSV->getNumOperands();
3096 bool hasNegative = false;
3097 for (unsigned i = 0; !hasNegative && i != VWidth; ++i)
3098 if (ConstantInt *RHS = dyn_cast<ConstantInt>(RHSV->getOperand(i)))
3099 if (RHS->getValue().isNegative())
3103 std::vector<Constant *> Elts(VWidth);
3104 for (unsigned i = 0; i != VWidth; ++i) {
3105 if (ConstantInt *RHS = dyn_cast<ConstantInt>(RHSV->getOperand(i))) {
3106 if (RHS->getValue().isNegative())
3107 Elts[i] = cast<ConstantInt>(ConstantExpr::getNeg(RHS));
3113 Constant *NewRHSV = ConstantVector::get(Elts);
3114 if (NewRHSV != RHSV) {
3115 AddUsesToWorkList(I);
3116 I.setOperand(1, NewRHSV);
3125 Instruction *InstCombiner::visitFRem(BinaryOperator &I) {
3126 return commonRemTransforms(I);
3129 // isOneBitSet - Return true if there is exactly one bit set in the specified
3131 static bool isOneBitSet(const ConstantInt *CI) {
3132 return CI->getValue().isPowerOf2();
3135 // isHighOnes - Return true if the constant is of the form 1+0+.
3136 // This is the same as lowones(~X).
3137 static bool isHighOnes(const ConstantInt *CI) {
3138 return (~CI->getValue() + 1).isPowerOf2();
3141 /// getICmpCode - Encode a icmp predicate into a three bit mask. These bits
3142 /// are carefully arranged to allow folding of expressions such as:
3144 /// (A < B) | (A > B) --> (A != B)
3146 /// Note that this is only valid if the first and second predicates have the
3147 /// same sign. Is illegal to do: (A u< B) | (A s> B)
3149 /// Three bits are used to represent the condition, as follows:
3154 /// <=> Value Definition
3155 /// 000 0 Always false
3162 /// 111 7 Always true
3164 static unsigned getICmpCode(const ICmpInst *ICI) {
3165 switch (ICI->getPredicate()) {
3167 case ICmpInst::ICMP_UGT: return 1; // 001
3168 case ICmpInst::ICMP_SGT: return 1; // 001
3169 case ICmpInst::ICMP_EQ: return 2; // 010
3170 case ICmpInst::ICMP_UGE: return 3; // 011
3171 case ICmpInst::ICMP_SGE: return 3; // 011
3172 case ICmpInst::ICMP_ULT: return 4; // 100
3173 case ICmpInst::ICMP_SLT: return 4; // 100
3174 case ICmpInst::ICMP_NE: return 5; // 101
3175 case ICmpInst::ICMP_ULE: return 6; // 110
3176 case ICmpInst::ICMP_SLE: return 6; // 110
3179 assert(0 && "Invalid ICmp predicate!");
3184 /// getFCmpCode - Similar to getICmpCode but for FCmpInst. This encodes a fcmp
3185 /// predicate into a three bit mask. It also returns whether it is an ordered
3186 /// predicate by reference.
3187 static unsigned getFCmpCode(FCmpInst::Predicate CC, bool &isOrdered) {
3190 case FCmpInst::FCMP_ORD: isOrdered = true; return 0; // 000
3191 case FCmpInst::FCMP_UNO: return 0; // 000
3192 case FCmpInst::FCMP_OGT: isOrdered = true; return 1; // 001
3193 case FCmpInst::FCMP_UGT: return 1; // 001
3194 case FCmpInst::FCMP_OEQ: isOrdered = true; return 2; // 010
3195 case FCmpInst::FCMP_UEQ: return 2; // 010
3196 case FCmpInst::FCMP_OGE: isOrdered = true; return 3; // 011
3197 case FCmpInst::FCMP_UGE: return 3; // 011
3198 case FCmpInst::FCMP_OLT: isOrdered = true; return 4; // 100
3199 case FCmpInst::FCMP_ULT: return 4; // 100
3200 case FCmpInst::FCMP_ONE: isOrdered = true; return 5; // 101
3201 case FCmpInst::FCMP_UNE: return 5; // 101
3202 case FCmpInst::FCMP_OLE: isOrdered = true; return 6; // 110
3203 case FCmpInst::FCMP_ULE: return 6; // 110
3206 // Not expecting FCMP_FALSE and FCMP_TRUE;
3207 assert(0 && "Unexpected FCmp predicate!");
3212 /// getICmpValue - This is the complement of getICmpCode, which turns an
3213 /// opcode and two operands into either a constant true or false, or a brand
3214 /// new ICmp instruction. The sign is passed in to determine which kind
3215 /// of predicate to use in the new icmp instruction.
3216 static Value *getICmpValue(bool sign, unsigned code, Value *LHS, Value *RHS) {
3218 default: assert(0 && "Illegal ICmp code!");
3219 case 0: return ConstantInt::getFalse();
3222 return new ICmpInst(ICmpInst::ICMP_SGT, LHS, RHS);
3224 return new ICmpInst(ICmpInst::ICMP_UGT, LHS, RHS);
3225 case 2: return new ICmpInst(ICmpInst::ICMP_EQ, LHS, RHS);
3228 return new ICmpInst(ICmpInst::ICMP_SGE, LHS, RHS);
3230 return new ICmpInst(ICmpInst::ICMP_UGE, LHS, RHS);
3233 return new ICmpInst(ICmpInst::ICMP_SLT, LHS, RHS);
3235 return new ICmpInst(ICmpInst::ICMP_ULT, LHS, RHS);
3236 case 5: return new ICmpInst(ICmpInst::ICMP_NE, LHS, RHS);
3239 return new ICmpInst(ICmpInst::ICMP_SLE, LHS, RHS);
3241 return new ICmpInst(ICmpInst::ICMP_ULE, LHS, RHS);
3242 case 7: return ConstantInt::getTrue();
3246 /// getFCmpValue - This is the complement of getFCmpCode, which turns an
3247 /// opcode and two operands into either a FCmp instruction. isordered is passed
3248 /// in to determine which kind of predicate to use in the new fcmp instruction.
3249 static Value *getFCmpValue(bool isordered, unsigned code,
3250 Value *LHS, Value *RHS) {
3252 default: assert(0 && "Illegal FCmp code!");
3255 return new FCmpInst(FCmpInst::FCMP_ORD, LHS, RHS);
3257 return new FCmpInst(FCmpInst::FCMP_UNO, LHS, RHS);
3260 return new FCmpInst(FCmpInst::FCMP_OGT, LHS, RHS);
3262 return new FCmpInst(FCmpInst::FCMP_UGT, LHS, RHS);
3265 return new FCmpInst(FCmpInst::FCMP_OEQ, LHS, RHS);
3267 return new FCmpInst(FCmpInst::FCMP_UEQ, LHS, RHS);
3270 return new FCmpInst(FCmpInst::FCMP_OGE, LHS, RHS);
3272 return new FCmpInst(FCmpInst::FCMP_UGE, LHS, RHS);
3275 return new FCmpInst(FCmpInst::FCMP_OLT, LHS, RHS);
3277 return new FCmpInst(FCmpInst::FCMP_ULT, LHS, RHS);
3280 return new FCmpInst(FCmpInst::FCMP_ONE, LHS, RHS);
3282 return new FCmpInst(FCmpInst::FCMP_UNE, LHS, RHS);
3285 return new FCmpInst(FCmpInst::FCMP_OLE, LHS, RHS);
3287 return new FCmpInst(FCmpInst::FCMP_ULE, LHS, RHS);
3288 case 7: return ConstantInt::getTrue();
3292 /// PredicatesFoldable - Return true if both predicates match sign or if at
3293 /// least one of them is an equality comparison (which is signless).
3294 static bool PredicatesFoldable(ICmpInst::Predicate p1, ICmpInst::Predicate p2) {
3295 return (ICmpInst::isSignedPredicate(p1) == ICmpInst::isSignedPredicate(p2)) ||
3296 (ICmpInst::isSignedPredicate(p1) && ICmpInst::isEquality(p2)) ||
3297 (ICmpInst::isSignedPredicate(p2) && ICmpInst::isEquality(p1));
3301 // FoldICmpLogical - Implements (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
3302 struct FoldICmpLogical {
3305 ICmpInst::Predicate pred;
3306 FoldICmpLogical(InstCombiner &ic, ICmpInst *ICI)
3307 : IC(ic), LHS(ICI->getOperand(0)), RHS(ICI->getOperand(1)),
3308 pred(ICI->getPredicate()) {}
3309 bool shouldApply(Value *V) const {
3310 if (ICmpInst *ICI = dyn_cast<ICmpInst>(V))
3311 if (PredicatesFoldable(pred, ICI->getPredicate()))
3312 return ((ICI->getOperand(0) == LHS && ICI->getOperand(1) == RHS) ||
3313 (ICI->getOperand(0) == RHS && ICI->getOperand(1) == LHS));
3316 Instruction *apply(Instruction &Log) const {
3317 ICmpInst *ICI = cast<ICmpInst>(Log.getOperand(0));
3318 if (ICI->getOperand(0) != LHS) {
3319 assert(ICI->getOperand(1) == LHS);
3320 ICI->swapOperands(); // Swap the LHS and RHS of the ICmp
3323 ICmpInst *RHSICI = cast<ICmpInst>(Log.getOperand(1));
3324 unsigned LHSCode = getICmpCode(ICI);
3325 unsigned RHSCode = getICmpCode(RHSICI);
3327 switch (Log.getOpcode()) {
3328 case Instruction::And: Code = LHSCode & RHSCode; break;
3329 case Instruction::Or: Code = LHSCode | RHSCode; break;
3330 case Instruction::Xor: Code = LHSCode ^ RHSCode; break;
3331 default: assert(0 && "Illegal logical opcode!"); return 0;
3334 bool isSigned = ICmpInst::isSignedPredicate(RHSICI->getPredicate()) ||
3335 ICmpInst::isSignedPredicate(ICI->getPredicate());
3337 Value *RV = getICmpValue(isSigned, Code, LHS, RHS);
3338 if (Instruction *I = dyn_cast<Instruction>(RV))
3340 // Otherwise, it's a constant boolean value...
3341 return IC.ReplaceInstUsesWith(Log, RV);
3344 } // end anonymous namespace
3346 // OptAndOp - This handles expressions of the form ((val OP C1) & C2). Where
3347 // the Op parameter is 'OP', OpRHS is 'C1', and AndRHS is 'C2'. Op is
3348 // guaranteed to be a binary operator.
3349 Instruction *InstCombiner::OptAndOp(Instruction *Op,
3351 ConstantInt *AndRHS,
3352 BinaryOperator &TheAnd) {
3353 Value *X = Op->getOperand(0);
3354 Constant *Together = 0;
3356 Together = And(AndRHS, OpRHS);
3358 switch (Op->getOpcode()) {
3359 case Instruction::Xor:
3360 if (Op->hasOneUse()) {
3361 // (X ^ C1) & C2 --> (X & C2) ^ (C1&C2)
3362 Instruction *And = BinaryOperator::CreateAnd(X, AndRHS);
3363 InsertNewInstBefore(And, TheAnd);
3365 return BinaryOperator::CreateXor(And, Together);
3368 case Instruction::Or:
3369 if (Together == AndRHS) // (X | C) & C --> C
3370 return ReplaceInstUsesWith(TheAnd, AndRHS);
3372 if (Op->hasOneUse() && Together != OpRHS) {
3373 // (X | C1) & C2 --> (X | (C1&C2)) & C2
3374 Instruction *Or = BinaryOperator::CreateOr(X, Together);
3375 InsertNewInstBefore(Or, TheAnd);
3377 return BinaryOperator::CreateAnd(Or, AndRHS);
3380 case Instruction::Add:
3381 if (Op->hasOneUse()) {
3382 // Adding a one to a single bit bit-field should be turned into an XOR
3383 // of the bit. First thing to check is to see if this AND is with a
3384 // single bit constant.
3385 const APInt& AndRHSV = cast<ConstantInt>(AndRHS)->getValue();
3387 // If there is only one bit set...
3388 if (isOneBitSet(cast<ConstantInt>(AndRHS))) {
3389 // Ok, at this point, we know that we are masking the result of the
3390 // ADD down to exactly one bit. If the constant we are adding has
3391 // no bits set below this bit, then we can eliminate the ADD.
3392 const APInt& AddRHS = cast<ConstantInt>(OpRHS)->getValue();
3394 // Check to see if any bits below the one bit set in AndRHSV are set.
3395 if ((AddRHS & (AndRHSV-1)) == 0) {
3396 // If not, the only thing that can effect the output of the AND is
3397 // the bit specified by AndRHSV. If that bit is set, the effect of
3398 // the XOR is to toggle the bit. If it is clear, then the ADD has
3400 if ((AddRHS & AndRHSV) == 0) { // Bit is not set, noop
3401 TheAnd.setOperand(0, X);
3404 // Pull the XOR out of the AND.
3405 Instruction *NewAnd = BinaryOperator::CreateAnd(X, AndRHS);
3406 InsertNewInstBefore(NewAnd, TheAnd);
3407 NewAnd->takeName(Op);
3408 return BinaryOperator::CreateXor(NewAnd, AndRHS);
3415 case Instruction::Shl: {
3416 // We know that the AND will not produce any of the bits shifted in, so if
3417 // the anded constant includes them, clear them now!
3419 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3420 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3421 APInt ShlMask(APInt::getHighBitsSet(BitWidth, BitWidth-OpRHSVal));
3422 ConstantInt *CI = ConstantInt::get(AndRHS->getValue() & ShlMask);
3424 if (CI->getValue() == ShlMask) {
3425 // Masking out bits that the shift already masks
3426 return ReplaceInstUsesWith(TheAnd, Op); // No need for the and.
3427 } else if (CI != AndRHS) { // Reducing bits set in and.
3428 TheAnd.setOperand(1, CI);
3433 case Instruction::LShr:
3435 // We know that the AND will not produce any of the bits shifted in, so if
3436 // the anded constant includes them, clear them now! This only applies to
3437 // unsigned shifts, because a signed shr may bring in set bits!
3439 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3440 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3441 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
3442 ConstantInt *CI = ConstantInt::get(AndRHS->getValue() & ShrMask);
3444 if (CI->getValue() == ShrMask) {
3445 // Masking out bits that the shift already masks.
3446 return ReplaceInstUsesWith(TheAnd, Op);
3447 } else if (CI != AndRHS) {
3448 TheAnd.setOperand(1, CI); // Reduce bits set in and cst.
3453 case Instruction::AShr:
3455 // See if this is shifting in some sign extension, then masking it out
3457 if (Op->hasOneUse()) {
3458 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3459 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3460 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
3461 Constant *C = ConstantInt::get(AndRHS->getValue() & ShrMask);
3462 if (C == AndRHS) { // Masking out bits shifted in.
3463 // (Val ashr C1) & C2 -> (Val lshr C1) & C2
3464 // Make the argument unsigned.
3465 Value *ShVal = Op->getOperand(0);
3466 ShVal = InsertNewInstBefore(
3467 BinaryOperator::CreateLShr(ShVal, OpRHS,
3468 Op->getName()), TheAnd);
3469 return BinaryOperator::CreateAnd(ShVal, AndRHS, TheAnd.getName());
3478 /// InsertRangeTest - Emit a computation of: (V >= Lo && V < Hi) if Inside is
3479 /// true, otherwise (V < Lo || V >= Hi). In pratice, we emit the more efficient
3480 /// (V-Lo) <u Hi-Lo. This method expects that Lo <= Hi. isSigned indicates
3481 /// whether to treat the V, Lo and HI as signed or not. IB is the location to
3482 /// insert new instructions.
3483 Instruction *InstCombiner::InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
3484 bool isSigned, bool Inside,
3486 assert(cast<ConstantInt>(ConstantExpr::getICmp((isSigned ?
3487 ICmpInst::ICMP_SLE:ICmpInst::ICMP_ULE), Lo, Hi))->getZExtValue() &&
3488 "Lo is not <= Hi in range emission code!");
3491 if (Lo == Hi) // Trivially false.
3492 return new ICmpInst(ICmpInst::ICMP_NE, V, V);
3494 // V >= Min && V < Hi --> V < Hi
3495 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
3496 ICmpInst::Predicate pred = (isSigned ?
3497 ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT);
3498 return new ICmpInst(pred, V, Hi);
3501 // Emit V-Lo <u Hi-Lo
3502 Constant *NegLo = ConstantExpr::getNeg(Lo);
3503 Instruction *Add = BinaryOperator::CreateAdd(V, NegLo, V->getName()+".off");
3504 InsertNewInstBefore(Add, IB);
3505 Constant *UpperBound = ConstantExpr::getAdd(NegLo, Hi);
3506 return new ICmpInst(ICmpInst::ICMP_ULT, Add, UpperBound);
3509 if (Lo == Hi) // Trivially true.
3510 return new ICmpInst(ICmpInst::ICMP_EQ, V, V);
3512 // V < Min || V >= Hi -> V > Hi-1
3513 Hi = SubOne(cast<ConstantInt>(Hi));
3514 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
3515 ICmpInst::Predicate pred = (isSigned ?
3516 ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT);
3517 return new ICmpInst(pred, V, Hi);
3520 // Emit V-Lo >u Hi-1-Lo
3521 // Note that Hi has already had one subtracted from it, above.
3522 ConstantInt *NegLo = cast<ConstantInt>(ConstantExpr::getNeg(Lo));
3523 Instruction *Add = BinaryOperator::CreateAdd(V, NegLo, V->getName()+".off");
3524 InsertNewInstBefore(Add, IB);
3525 Constant *LowerBound = ConstantExpr::getAdd(NegLo, Hi);
3526 return new ICmpInst(ICmpInst::ICMP_UGT, Add, LowerBound);
3529 // isRunOfOnes - Returns true iff Val consists of one contiguous run of 1s with
3530 // any number of 0s on either side. The 1s are allowed to wrap from LSB to
3531 // MSB, so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs. 0x0F0F0000 is
3532 // not, since all 1s are not contiguous.
3533 static bool isRunOfOnes(ConstantInt *Val, uint32_t &MB, uint32_t &ME) {
3534 const APInt& V = Val->getValue();
3535 uint32_t BitWidth = Val->getType()->getBitWidth();
3536 if (!APIntOps::isShiftedMask(BitWidth, V)) return false;
3538 // look for the first zero bit after the run of ones
3539 MB = BitWidth - ((V - 1) ^ V).countLeadingZeros();
3540 // look for the first non-zero bit
3541 ME = V.getActiveBits();
3545 /// FoldLogicalPlusAnd - This is part of an expression (LHS +/- RHS) & Mask,
3546 /// where isSub determines whether the operator is a sub. If we can fold one of
3547 /// the following xforms:
3549 /// ((A & N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == Mask
3550 /// ((A | N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3551 /// ((A ^ N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3553 /// return (A +/- B).
3555 Value *InstCombiner::FoldLogicalPlusAnd(Value *LHS, Value *RHS,
3556 ConstantInt *Mask, bool isSub,
3558 Instruction *LHSI = dyn_cast<Instruction>(LHS);
3559 if (!LHSI || LHSI->getNumOperands() != 2 ||
3560 !isa<ConstantInt>(LHSI->getOperand(1))) return 0;
3562 ConstantInt *N = cast<ConstantInt>(LHSI->getOperand(1));
3564 switch (LHSI->getOpcode()) {
3566 case Instruction::And:
3567 if (And(N, Mask) == Mask) {
3568 // If the AndRHS is a power of two minus one (0+1+), this is simple.
3569 if ((Mask->getValue().countLeadingZeros() +
3570 Mask->getValue().countPopulation()) ==
3571 Mask->getValue().getBitWidth())
3574 // Otherwise, if Mask is 0+1+0+, and if B is known to have the low 0+
3575 // part, we don't need any explicit masks to take them out of A. If that
3576 // is all N is, ignore it.
3577 uint32_t MB = 0, ME = 0;
3578 if (isRunOfOnes(Mask, MB, ME)) { // begin/end bit of run, inclusive
3579 uint32_t BitWidth = cast<IntegerType>(RHS->getType())->getBitWidth();
3580 APInt Mask(APInt::getLowBitsSet(BitWidth, MB-1));
3581 if (MaskedValueIsZero(RHS, Mask))
3586 case Instruction::Or:
3587 case Instruction::Xor:
3588 // If the AndRHS is a power of two minus one (0+1+), and N&Mask == 0
3589 if ((Mask->getValue().countLeadingZeros() +
3590 Mask->getValue().countPopulation()) == Mask->getValue().getBitWidth()
3591 && And(N, Mask)->isZero())
3598 New = BinaryOperator::CreateSub(LHSI->getOperand(0), RHS, "fold");
3600 New = BinaryOperator::CreateAdd(LHSI->getOperand(0), RHS, "fold");
3601 return InsertNewInstBefore(New, I);
3604 /// FoldAndOfICmps - Fold (icmp)&(icmp) if possible.
3605 Instruction *InstCombiner::FoldAndOfICmps(Instruction &I,
3606 ICmpInst *LHS, ICmpInst *RHS) {
3608 ConstantInt *LHSCst, *RHSCst;
3609 ICmpInst::Predicate LHSCC, RHSCC;
3611 // This only handles icmp of constants: (icmp1 A, C1) & (icmp2 B, C2).
3612 if (!match(LHS, m_ICmp(LHSCC, m_Value(Val), m_ConstantInt(LHSCst))) ||
3613 !match(RHS, m_ICmp(RHSCC, m_Value(Val2), m_ConstantInt(RHSCst))))
3616 // (icmp ult A, C) & (icmp ult B, C) --> (icmp ult (A|B), C)
3617 // where C is a power of 2
3618 if (LHSCst == RHSCst && LHSCC == RHSCC && LHSCC == ICmpInst::ICMP_ULT &&
3619 LHSCst->getValue().isPowerOf2()) {
3620 Instruction *NewOr = BinaryOperator::CreateOr(Val, Val2);
3621 InsertNewInstBefore(NewOr, I);
3622 return new ICmpInst(LHSCC, NewOr, LHSCst);
3625 // From here on, we only handle:
3626 // (icmp1 A, C1) & (icmp2 A, C2) --> something simpler.
3627 if (Val != Val2) return 0;
3629 // ICMP_[US][GL]E X, CST is folded to ICMP_[US][GL]T elsewhere.
3630 if (LHSCC == ICmpInst::ICMP_UGE || LHSCC == ICmpInst::ICMP_ULE ||
3631 RHSCC == ICmpInst::ICMP_UGE || RHSCC == ICmpInst::ICMP_ULE ||
3632 LHSCC == ICmpInst::ICMP_SGE || LHSCC == ICmpInst::ICMP_SLE ||
3633 RHSCC == ICmpInst::ICMP_SGE || RHSCC == ICmpInst::ICMP_SLE)
3636 // We can't fold (ugt x, C) & (sgt x, C2).
3637 if (!PredicatesFoldable(LHSCC, RHSCC))
3640 // Ensure that the larger constant is on the RHS.
3642 if (ICmpInst::isSignedPredicate(LHSCC) ||
3643 (ICmpInst::isEquality(LHSCC) &&
3644 ICmpInst::isSignedPredicate(RHSCC)))
3645 ShouldSwap = LHSCst->getValue().sgt(RHSCst->getValue());
3647 ShouldSwap = LHSCst->getValue().ugt(RHSCst->getValue());
3650 std::swap(LHS, RHS);
3651 std::swap(LHSCst, RHSCst);
3652 std::swap(LHSCC, RHSCC);
3655 // At this point, we know we have have two icmp instructions
3656 // comparing a value against two constants and and'ing the result
3657 // together. Because of the above check, we know that we only have
3658 // icmp eq, icmp ne, icmp [su]lt, and icmp [SU]gt here. We also know
3659 // (from the FoldICmpLogical check above), that the two constants
3660 // are not equal and that the larger constant is on the RHS
3661 assert(LHSCst != RHSCst && "Compares not folded above?");
3664 default: assert(0 && "Unknown integer condition code!");
3665 case ICmpInst::ICMP_EQ:
3667 default: assert(0 && "Unknown integer condition code!");
3668 case ICmpInst::ICMP_EQ: // (X == 13 & X == 15) -> false
3669 case ICmpInst::ICMP_UGT: // (X == 13 & X > 15) -> false
3670 case ICmpInst::ICMP_SGT: // (X == 13 & X > 15) -> false
3671 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3672 case ICmpInst::ICMP_NE: // (X == 13 & X != 15) -> X == 13
3673 case ICmpInst::ICMP_ULT: // (X == 13 & X < 15) -> X == 13
3674 case ICmpInst::ICMP_SLT: // (X == 13 & X < 15) -> X == 13
3675 return ReplaceInstUsesWith(I, LHS);
3677 case ICmpInst::ICMP_NE:
3679 default: assert(0 && "Unknown integer condition code!");
3680 case ICmpInst::ICMP_ULT:
3681 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X u< 14) -> X < 13
3682 return new ICmpInst(ICmpInst::ICMP_ULT, Val, LHSCst);
3683 break; // (X != 13 & X u< 15) -> no change
3684 case ICmpInst::ICMP_SLT:
3685 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X s< 14) -> X < 13
3686 return new ICmpInst(ICmpInst::ICMP_SLT, Val, LHSCst);
3687 break; // (X != 13 & X s< 15) -> no change
3688 case ICmpInst::ICMP_EQ: // (X != 13 & X == 15) -> X == 15
3689 case ICmpInst::ICMP_UGT: // (X != 13 & X u> 15) -> X u> 15
3690 case ICmpInst::ICMP_SGT: // (X != 13 & X s> 15) -> X s> 15
3691 return ReplaceInstUsesWith(I, RHS);
3692 case ICmpInst::ICMP_NE:
3693 if (LHSCst == SubOne(RHSCst)){// (X != 13 & X != 14) -> X-13 >u 1
3694 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
3695 Instruction *Add = BinaryOperator::CreateAdd(Val, AddCST,
3696 Val->getName()+".off");
3697 InsertNewInstBefore(Add, I);
3698 return new ICmpInst(ICmpInst::ICMP_UGT, Add,
3699 ConstantInt::get(Add->getType(), 1));
3701 break; // (X != 13 & X != 15) -> no change
3704 case ICmpInst::ICMP_ULT:
3706 default: assert(0 && "Unknown integer condition code!");
3707 case ICmpInst::ICMP_EQ: // (X u< 13 & X == 15) -> false
3708 case ICmpInst::ICMP_UGT: // (X u< 13 & X u> 15) -> false
3709 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3710 case ICmpInst::ICMP_SGT: // (X u< 13 & X s> 15) -> no change
3712 case ICmpInst::ICMP_NE: // (X u< 13 & X != 15) -> X u< 13
3713 case ICmpInst::ICMP_ULT: // (X u< 13 & X u< 15) -> X u< 13
3714 return ReplaceInstUsesWith(I, LHS);
3715 case ICmpInst::ICMP_SLT: // (X u< 13 & X s< 15) -> no change
3719 case ICmpInst::ICMP_SLT:
3721 default: assert(0 && "Unknown integer condition code!");
3722 case ICmpInst::ICMP_EQ: // (X s< 13 & X == 15) -> false
3723 case ICmpInst::ICMP_SGT: // (X s< 13 & X s> 15) -> false
3724 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3725 case ICmpInst::ICMP_UGT: // (X s< 13 & X u> 15) -> no change
3727 case ICmpInst::ICMP_NE: // (X s< 13 & X != 15) -> X < 13
3728 case ICmpInst::ICMP_SLT: // (X s< 13 & X s< 15) -> X < 13
3729 return ReplaceInstUsesWith(I, LHS);
3730 case ICmpInst::ICMP_ULT: // (X s< 13 & X u< 15) -> no change
3734 case ICmpInst::ICMP_UGT:
3736 default: assert(0 && "Unknown integer condition code!");
3737 case ICmpInst::ICMP_EQ: // (X u> 13 & X == 15) -> X == 15
3738 case ICmpInst::ICMP_UGT: // (X u> 13 & X u> 15) -> X u> 15
3739 return ReplaceInstUsesWith(I, RHS);
3740 case ICmpInst::ICMP_SGT: // (X u> 13 & X s> 15) -> no change
3742 case ICmpInst::ICMP_NE:
3743 if (RHSCst == AddOne(LHSCst)) // (X u> 13 & X != 14) -> X u> 14
3744 return new ICmpInst(LHSCC, Val, RHSCst);
3745 break; // (X u> 13 & X != 15) -> no change
3746 case ICmpInst::ICMP_ULT: // (X u> 13 & X u< 15) -> (X-14) <u 1
3747 return InsertRangeTest(Val, AddOne(LHSCst), RHSCst, false, true, I);
3748 case ICmpInst::ICMP_SLT: // (X u> 13 & X s< 15) -> no change
3752 case ICmpInst::ICMP_SGT:
3754 default: assert(0 && "Unknown integer condition code!");
3755 case ICmpInst::ICMP_EQ: // (X s> 13 & X == 15) -> X == 15
3756 case ICmpInst::ICMP_SGT: // (X s> 13 & X s> 15) -> X s> 15
3757 return ReplaceInstUsesWith(I, RHS);
3758 case ICmpInst::ICMP_UGT: // (X s> 13 & X u> 15) -> no change
3760 case ICmpInst::ICMP_NE:
3761 if (RHSCst == AddOne(LHSCst)) // (X s> 13 & X != 14) -> X s> 14
3762 return new ICmpInst(LHSCC, Val, RHSCst);
3763 break; // (X s> 13 & X != 15) -> no change
3764 case ICmpInst::ICMP_SLT: // (X s> 13 & X s< 15) -> (X-14) s< 1
3765 return InsertRangeTest(Val, AddOne(LHSCst), RHSCst, true, true, I);
3766 case ICmpInst::ICMP_ULT: // (X s> 13 & X u< 15) -> no change
3776 Instruction *InstCombiner::visitAnd(BinaryOperator &I) {
3777 bool Changed = SimplifyCommutative(I);
3778 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3780 if (isa<UndefValue>(Op1)) // X & undef -> 0
3781 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3785 return ReplaceInstUsesWith(I, Op1);
3787 // See if we can simplify any instructions used by the instruction whose sole
3788 // purpose is to compute bits we don't care about.
3789 if (!isa<VectorType>(I.getType())) {
3790 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
3791 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
3792 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
3793 KnownZero, KnownOne))
3796 if (ConstantVector *CP = dyn_cast<ConstantVector>(Op1)) {
3797 if (CP->isAllOnesValue()) // X & <-1,-1> -> X
3798 return ReplaceInstUsesWith(I, I.getOperand(0));
3799 } else if (isa<ConstantAggregateZero>(Op1)) {
3800 return ReplaceInstUsesWith(I, Op1); // X & <0,0> -> <0,0>
3804 if (ConstantInt *AndRHS = dyn_cast<ConstantInt>(Op1)) {
3805 const APInt& AndRHSMask = AndRHS->getValue();
3806 APInt NotAndRHS(~AndRHSMask);
3808 // Optimize a variety of ((val OP C1) & C2) combinations...
3809 if (isa<BinaryOperator>(Op0)) {
3810 Instruction *Op0I = cast<Instruction>(Op0);
3811 Value *Op0LHS = Op0I->getOperand(0);
3812 Value *Op0RHS = Op0I->getOperand(1);
3813 switch (Op0I->getOpcode()) {
3814 case Instruction::Xor:
3815 case Instruction::Or:
3816 // If the mask is only needed on one incoming arm, push it up.
3817 if (Op0I->hasOneUse()) {
3818 if (MaskedValueIsZero(Op0LHS, NotAndRHS)) {
3819 // Not masking anything out for the LHS, move to RHS.
3820 Instruction *NewRHS = BinaryOperator::CreateAnd(Op0RHS, AndRHS,
3821 Op0RHS->getName()+".masked");
3822 InsertNewInstBefore(NewRHS, I);
3823 return BinaryOperator::Create(
3824 cast<BinaryOperator>(Op0I)->getOpcode(), Op0LHS, NewRHS);
3826 if (!isa<Constant>(Op0RHS) &&
3827 MaskedValueIsZero(Op0RHS, NotAndRHS)) {
3828 // Not masking anything out for the RHS, move to LHS.
3829 Instruction *NewLHS = BinaryOperator::CreateAnd(Op0LHS, AndRHS,
3830 Op0LHS->getName()+".masked");
3831 InsertNewInstBefore(NewLHS, I);
3832 return BinaryOperator::Create(
3833 cast<BinaryOperator>(Op0I)->getOpcode(), NewLHS, Op0RHS);
3838 case Instruction::Add:
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, false, I))
3843 return BinaryOperator::CreateAnd(V, AndRHS);
3844 if (Value *V = FoldLogicalPlusAnd(Op0RHS, Op0LHS, AndRHS, false, I))
3845 return BinaryOperator::CreateAnd(V, AndRHS); // Add commutes
3848 case Instruction::Sub:
3849 // ((A & N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == AndRHS.
3850 // ((A | N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
3851 // ((A ^ N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
3852 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, true, I))
3853 return BinaryOperator::CreateAnd(V, AndRHS);
3855 // (A - N) & AndRHS -> -N & AndRHS iff A&AndRHS==0 and AndRHS
3856 // has 1's for all bits that the subtraction with A might affect.
3857 if (Op0I->hasOneUse()) {
3858 uint32_t BitWidth = AndRHSMask.getBitWidth();
3859 uint32_t Zeros = AndRHSMask.countLeadingZeros();
3860 APInt Mask = APInt::getLowBitsSet(BitWidth, BitWidth - Zeros);
3862 ConstantInt *A = dyn_cast<ConstantInt>(Op0LHS);
3863 if (!(A && A->isZero()) && // avoid infinite recursion.
3864 MaskedValueIsZero(Op0LHS, Mask)) {
3865 Instruction *NewNeg = BinaryOperator::CreateNeg(Op0RHS);
3866 InsertNewInstBefore(NewNeg, I);
3867 return BinaryOperator::CreateAnd(NewNeg, AndRHS);
3872 case Instruction::Shl:
3873 case Instruction::LShr:
3874 // (1 << x) & 1 --> zext(x == 0)
3875 // (1 >> x) & 1 --> zext(x == 0)
3876 if (AndRHSMask == 1 && Op0LHS == AndRHS) {
3877 Instruction *NewICmp = new ICmpInst(ICmpInst::ICMP_EQ, Op0RHS,
3878 Constant::getNullValue(I.getType()));
3879 InsertNewInstBefore(NewICmp, I);
3880 return new ZExtInst(NewICmp, I.getType());
3885 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
3886 if (Instruction *Res = OptAndOp(Op0I, Op0CI, AndRHS, I))
3888 } else if (CastInst *CI = dyn_cast<CastInst>(Op0)) {
3889 // If this is an integer truncation or change from signed-to-unsigned, and
3890 // if the source is an and/or with immediate, transform it. This
3891 // frequently occurs for bitfield accesses.
3892 if (Instruction *CastOp = dyn_cast<Instruction>(CI->getOperand(0))) {
3893 if ((isa<TruncInst>(CI) || isa<BitCastInst>(CI)) &&
3894 CastOp->getNumOperands() == 2)
3895 if (ConstantInt *AndCI = dyn_cast<ConstantInt>(CastOp->getOperand(1))) {
3896 if (CastOp->getOpcode() == Instruction::And) {
3897 // Change: and (cast (and X, C1) to T), C2
3898 // into : and (cast X to T), trunc_or_bitcast(C1)&C2
3899 // This will fold the two constants together, which may allow
3900 // other simplifications.
3901 Instruction *NewCast = CastInst::CreateTruncOrBitCast(
3902 CastOp->getOperand(0), I.getType(),
3903 CastOp->getName()+".shrunk");
3904 NewCast = InsertNewInstBefore(NewCast, I);
3905 // trunc_or_bitcast(C1)&C2
3906 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
3907 C3 = ConstantExpr::getAnd(C3, AndRHS);
3908 return BinaryOperator::CreateAnd(NewCast, C3);
3909 } else if (CastOp->getOpcode() == Instruction::Or) {
3910 // Change: and (cast (or X, C1) to T), C2
3911 // into : trunc(C1)&C2 iff trunc(C1)&C2 == C2
3912 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
3913 if (ConstantExpr::getAnd(C3, AndRHS) == AndRHS) // trunc(C1)&C2
3914 return ReplaceInstUsesWith(I, AndRHS);
3920 // Try to fold constant and into select arguments.
3921 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3922 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3924 if (isa<PHINode>(Op0))
3925 if (Instruction *NV = FoldOpIntoPhi(I))
3929 Value *Op0NotVal = dyn_castNotVal(Op0);
3930 Value *Op1NotVal = dyn_castNotVal(Op1);
3932 if (Op0NotVal == Op1 || Op1NotVal == Op0) // A & ~A == ~A & A == 0
3933 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3935 // (~A & ~B) == (~(A | B)) - De Morgan's Law
3936 if (Op0NotVal && Op1NotVal && isOnlyUse(Op0) && isOnlyUse(Op1)) {
3937 Instruction *Or = BinaryOperator::CreateOr(Op0NotVal, Op1NotVal,
3938 I.getName()+".demorgan");
3939 InsertNewInstBefore(Or, I);
3940 return BinaryOperator::CreateNot(Or);
3944 Value *A = 0, *B = 0, *C = 0, *D = 0;
3945 if (match(Op0, m_Or(m_Value(A), m_Value(B)))) {
3946 if (A == Op1 || B == Op1) // (A | ?) & A --> A
3947 return ReplaceInstUsesWith(I, Op1);
3949 // (A|B) & ~(A&B) -> A^B
3950 if (match(Op1, m_Not(m_And(m_Value(C), m_Value(D))))) {
3951 if ((A == C && B == D) || (A == D && B == C))
3952 return BinaryOperator::CreateXor(A, B);
3956 if (match(Op1, m_Or(m_Value(A), m_Value(B)))) {
3957 if (A == Op0 || B == Op0) // A & (A | ?) --> A
3958 return ReplaceInstUsesWith(I, Op0);
3960 // ~(A&B) & (A|B) -> A^B
3961 if (match(Op0, m_Not(m_And(m_Value(C), m_Value(D))))) {
3962 if ((A == C && B == D) || (A == D && B == C))
3963 return BinaryOperator::CreateXor(A, B);
3967 if (Op0->hasOneUse() &&
3968 match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
3969 if (A == Op1) { // (A^B)&A -> A&(A^B)
3970 I.swapOperands(); // Simplify below
3971 std::swap(Op0, Op1);
3972 } else if (B == Op1) { // (A^B)&B -> B&(B^A)
3973 cast<BinaryOperator>(Op0)->swapOperands();
3974 I.swapOperands(); // Simplify below
3975 std::swap(Op0, Op1);
3979 if (Op1->hasOneUse() &&
3980 match(Op1, m_Xor(m_Value(A), m_Value(B)))) {
3981 if (B == Op0) { // B&(A^B) -> B&(B^A)
3982 cast<BinaryOperator>(Op1)->swapOperands();
3985 if (A == Op0) { // A&(A^B) -> A & ~B
3986 Instruction *NotB = BinaryOperator::CreateNot(B, "tmp");
3987 InsertNewInstBefore(NotB, I);
3988 return BinaryOperator::CreateAnd(A, NotB);
3992 // (A&((~A)|B)) -> A&B
3993 if (match(Op0, m_Or(m_Not(m_Specific(Op1)), m_Value(A))) ||
3994 match(Op0, m_Or(m_Value(A), m_Not(m_Specific(Op1)))))
3995 return BinaryOperator::CreateAnd(A, Op1);
3996 if (match(Op1, m_Or(m_Not(m_Specific(Op0)), m_Value(A))) ||
3997 match(Op1, m_Or(m_Value(A), m_Not(m_Specific(Op0)))))
3998 return BinaryOperator::CreateAnd(A, Op0);
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<-1, 0>(m_Value(Cond))))
4282 // ((cond?-1:0)&C) | (B&(cond?0:-1)) -> cond ? C : B.
4283 if (match(D, m_SelectCst<0, -1>(m_Specific(Cond))))
4284 return SelectInst::Create(Cond, C, B);
4285 if (match(D, m_Not(m_SelectCst<-1, 0>(m_Specific(Cond)))))
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<0, -1>(m_Specific(Cond))))
4289 return SelectInst::Create(Cond, C, D);
4290 if (match(B, m_Not(m_SelectCst<-1, 0>(m_Specific(Cond)))))
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 /// FoldOrWithConstants - This helper function folds:
4458 /// ((A | B) & C1) | (B & C2)
4464 /// when the XOR of the two constants is "all ones" (-1).
4465 Instruction *InstCombiner::FoldOrWithConstants(BinaryOperator &I, Value *Op,
4466 Value *A, Value *B, Value *C) {
4467 ConstantInt *CI1 = dyn_cast<ConstantInt>(C);
4471 ConstantInt *CI2 = 0;
4472 if (!match(Op, m_And(m_Value(V1), m_ConstantInt(CI2)))) return 0;
4474 APInt Xor = CI1->getValue() ^ CI2->getValue();
4475 if (!Xor.isAllOnesValue()) return 0;
4477 if (V1 == A || V1 == B) {
4478 Instruction *NewOp =
4479 InsertNewInstBefore(BinaryOperator::CreateAnd((V1 == A) ? B : A, CI1), I);
4480 return BinaryOperator::CreateOr(NewOp, V1);
4486 Instruction *InstCombiner::visitOr(BinaryOperator &I) {
4487 bool Changed = SimplifyCommutative(I);
4488 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4490 if (isa<UndefValue>(Op1)) // X | undef -> -1
4491 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4495 return ReplaceInstUsesWith(I, Op0);
4497 // See if we can simplify any instructions used by the instruction whose sole
4498 // purpose is to compute bits we don't care about.
4499 if (!isa<VectorType>(I.getType())) {
4500 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
4501 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
4502 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
4503 KnownZero, KnownOne))
4505 } else if (isa<ConstantAggregateZero>(Op1)) {
4506 return ReplaceInstUsesWith(I, Op0); // X | <0,0> -> X
4507 } else if (ConstantVector *CP = dyn_cast<ConstantVector>(Op1)) {
4508 if (CP->isAllOnesValue()) // X | <-1,-1> -> <-1,-1>
4509 return ReplaceInstUsesWith(I, I.getOperand(1));
4515 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
4516 ConstantInt *C1 = 0; Value *X = 0;
4517 // (X & C1) | C2 --> (X | C2) & (C1|C2)
4518 if (match(Op0, m_And(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
4519 Instruction *Or = BinaryOperator::CreateOr(X, RHS);
4520 InsertNewInstBefore(Or, I);
4522 return BinaryOperator::CreateAnd(Or,
4523 ConstantInt::get(RHS->getValue() | C1->getValue()));
4526 // (X ^ C1) | C2 --> (X | C2) ^ (C1&~C2)
4527 if (match(Op0, m_Xor(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
4528 Instruction *Or = BinaryOperator::CreateOr(X, RHS);
4529 InsertNewInstBefore(Or, I);
4531 return BinaryOperator::CreateXor(Or,
4532 ConstantInt::get(C1->getValue() & ~RHS->getValue()));
4535 // Try to fold constant and into select arguments.
4536 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
4537 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
4539 if (isa<PHINode>(Op0))
4540 if (Instruction *NV = FoldOpIntoPhi(I))
4544 Value *A = 0, *B = 0;
4545 ConstantInt *C1 = 0, *C2 = 0;
4547 if (match(Op0, m_And(m_Value(A), m_Value(B))))
4548 if (A == Op1 || B == Op1) // (A & ?) | A --> A
4549 return ReplaceInstUsesWith(I, Op1);
4550 if (match(Op1, m_And(m_Value(A), m_Value(B))))
4551 if (A == Op0 || B == Op0) // A | (A & ?) --> A
4552 return ReplaceInstUsesWith(I, Op0);
4554 // (A | B) | C and A | (B | C) -> bswap if possible.
4555 // (A >> B) | (C << D) and (A << B) | (B >> C) -> bswap if possible.
4556 if (match(Op0, m_Or(m_Value(), m_Value())) ||
4557 match(Op1, m_Or(m_Value(), m_Value())) ||
4558 (match(Op0, m_Shift(m_Value(), m_Value())) &&
4559 match(Op1, m_Shift(m_Value(), m_Value())))) {
4560 if (Instruction *BSwap = MatchBSwap(I))
4564 // (X^C)|Y -> (X|Y)^C iff Y&C == 0
4565 if (Op0->hasOneUse() && match(Op0, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
4566 MaskedValueIsZero(Op1, C1->getValue())) {
4567 Instruction *NOr = BinaryOperator::CreateOr(A, Op1);
4568 InsertNewInstBefore(NOr, I);
4570 return BinaryOperator::CreateXor(NOr, C1);
4573 // Y|(X^C) -> (X|Y)^C iff Y&C == 0
4574 if (Op1->hasOneUse() && match(Op1, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
4575 MaskedValueIsZero(Op0, C1->getValue())) {
4576 Instruction *NOr = BinaryOperator::CreateOr(A, Op0);
4577 InsertNewInstBefore(NOr, I);
4579 return BinaryOperator::CreateXor(NOr, C1);
4583 Value *C = 0, *D = 0;
4584 if (match(Op0, m_And(m_Value(A), m_Value(C))) &&
4585 match(Op1, m_And(m_Value(B), m_Value(D)))) {
4586 Value *V1 = 0, *V2 = 0, *V3 = 0;
4587 C1 = dyn_cast<ConstantInt>(C);
4588 C2 = dyn_cast<ConstantInt>(D);
4589 if (C1 && C2) { // (A & C1)|(B & C2)
4590 // If we have: ((V + N) & C1) | (V & C2)
4591 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
4592 // replace with V+N.
4593 if (C1->getValue() == ~C2->getValue()) {
4594 if ((C2->getValue() & (C2->getValue()+1)) == 0 && // C2 == 0+1+
4595 match(A, m_Add(m_Value(V1), m_Value(V2)))) {
4596 // Add commutes, try both ways.
4597 if (V1 == B && MaskedValueIsZero(V2, C2->getValue()))
4598 return ReplaceInstUsesWith(I, A);
4599 if (V2 == B && MaskedValueIsZero(V1, C2->getValue()))
4600 return ReplaceInstUsesWith(I, A);
4602 // Or commutes, try both ways.
4603 if ((C1->getValue() & (C1->getValue()+1)) == 0 &&
4604 match(B, m_Add(m_Value(V1), m_Value(V2)))) {
4605 // Add commutes, try both ways.
4606 if (V1 == A && MaskedValueIsZero(V2, C1->getValue()))
4607 return ReplaceInstUsesWith(I, B);
4608 if (V2 == A && MaskedValueIsZero(V1, C1->getValue()))
4609 return ReplaceInstUsesWith(I, B);
4612 V1 = 0; V2 = 0; V3 = 0;
4615 // Check to see if we have any common things being and'ed. If so, find the
4616 // terms for V1 & (V2|V3).
4617 if (isOnlyUse(Op0) || isOnlyUse(Op1)) {
4618 if (A == B) // (A & C)|(A & D) == A & (C|D)
4619 V1 = A, V2 = C, V3 = D;
4620 else if (A == D) // (A & C)|(B & A) == A & (B|C)
4621 V1 = A, V2 = B, V3 = C;
4622 else if (C == B) // (A & C)|(C & D) == C & (A|D)
4623 V1 = C, V2 = A, V3 = D;
4624 else if (C == D) // (A & C)|(B & C) == C & (A|B)
4625 V1 = C, V2 = A, V3 = B;
4629 InsertNewInstBefore(BinaryOperator::CreateOr(V2, V3, "tmp"), I);
4630 return BinaryOperator::CreateAnd(V1, Or);
4634 // (A & (C0?-1:0)) | (B & ~(C0?-1:0)) -> C0 ? A : B, and commuted variants
4635 if (Instruction *Match = MatchSelectFromAndOr(A, B, C, D))
4637 if (Instruction *Match = MatchSelectFromAndOr(B, A, D, C))
4639 if (Instruction *Match = MatchSelectFromAndOr(C, B, A, D))
4641 if (Instruction *Match = MatchSelectFromAndOr(D, A, B, C))
4644 // ((A&~B)|(~A&B)) -> A^B
4645 if ((match(C, m_Not(m_Specific(D))) &&
4646 match(B, m_Not(m_Specific(A)))))
4647 return BinaryOperator::CreateXor(A, D);
4648 // ((~B&A)|(~A&B)) -> A^B
4649 if ((match(A, m_Not(m_Specific(D))) &&
4650 match(B, m_Not(m_Specific(C)))))
4651 return BinaryOperator::CreateXor(C, D);
4652 // ((A&~B)|(B&~A)) -> A^B
4653 if ((match(C, m_Not(m_Specific(B))) &&
4654 match(D, m_Not(m_Specific(A)))))
4655 return BinaryOperator::CreateXor(A, B);
4656 // ((~B&A)|(B&~A)) -> A^B
4657 if ((match(A, m_Not(m_Specific(B))) &&
4658 match(D, m_Not(m_Specific(C)))))
4659 return BinaryOperator::CreateXor(C, B);
4662 // (X >> Z) | (Y >> Z) -> (X|Y) >> Z for all shifts.
4663 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
4664 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
4665 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
4666 SI0->getOperand(1) == SI1->getOperand(1) &&
4667 (SI0->hasOneUse() || SI1->hasOneUse())) {
4668 Instruction *NewOp =
4669 InsertNewInstBefore(BinaryOperator::CreateOr(SI0->getOperand(0),
4671 SI0->getName()), I);
4672 return BinaryOperator::Create(SI1->getOpcode(), NewOp,
4673 SI1->getOperand(1));
4677 // ((A|B)&1)|(B&-2) -> (A&1) | B
4678 if (match(Op0, m_And(m_Or(m_Value(A), m_Value(B)), m_Value(C))) ||
4679 match(Op0, m_And(m_Value(C), m_Or(m_Value(A), m_Value(B))))) {
4680 Instruction *Ret = FoldOrWithConstants(I, Op1, A, B, C);
4681 if (Ret) return Ret;
4683 // (B&-2)|((A|B)&1) -> (A&1) | B
4684 if (match(Op1, m_And(m_Or(m_Value(A), m_Value(B)), m_Value(C))) ||
4685 match(Op1, m_And(m_Value(C), m_Or(m_Value(A), m_Value(B))))) {
4686 Instruction *Ret = FoldOrWithConstants(I, Op0, A, B, C);
4687 if (Ret) return Ret;
4690 if (match(Op0, m_Not(m_Value(A)))) { // ~A | Op1
4691 if (A == Op1) // ~A | A == -1
4692 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4696 // Note, A is still live here!
4697 if (match(Op1, m_Not(m_Value(B)))) { // Op0 | ~B
4699 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4701 // (~A | ~B) == (~(A & B)) - De Morgan's Law
4702 if (A && isOnlyUse(Op0) && isOnlyUse(Op1)) {
4703 Value *And = InsertNewInstBefore(BinaryOperator::CreateAnd(A, B,
4704 I.getName()+".demorgan"), I);
4705 return BinaryOperator::CreateNot(And);
4709 // (icmp1 A, B) | (icmp2 A, B) --> (icmp3 A, B)
4710 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1))) {
4711 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
4714 if (ICmpInst *LHS = dyn_cast<ICmpInst>(I.getOperand(0)))
4715 if (Instruction *Res = FoldOrOfICmps(I, LHS, RHS))
4719 // fold (or (cast A), (cast B)) -> (cast (or A, B))
4720 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
4721 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4722 if (Op0C->getOpcode() == Op1C->getOpcode()) {// same cast kind ?
4723 if (!isa<ICmpInst>(Op0C->getOperand(0)) ||
4724 !isa<ICmpInst>(Op1C->getOperand(0))) {
4725 const Type *SrcTy = Op0C->getOperand(0)->getType();
4726 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
4727 // Only do this if the casts both really cause code to be
4729 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4731 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4733 Instruction *NewOp = BinaryOperator::CreateOr(Op0C->getOperand(0),
4734 Op1C->getOperand(0),
4736 InsertNewInstBefore(NewOp, I);
4737 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
4744 // (fcmp uno x, c) | (fcmp uno y, c) -> (fcmp uno x, y)
4745 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0))) {
4746 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1))) {
4747 if (LHS->getPredicate() == FCmpInst::FCMP_UNO &&
4748 RHS->getPredicate() == FCmpInst::FCMP_UNO &&
4749 LHS->getOperand(0)->getType() == RHS->getOperand(0)->getType()) {
4750 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
4751 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
4752 // If either of the constants are nans, then the whole thing returns
4754 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
4755 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4757 // Otherwise, no need to compare the two constants, compare the
4759 return new FCmpInst(FCmpInst::FCMP_UNO, LHS->getOperand(0),
4760 RHS->getOperand(0));
4763 Value *Op0LHS, *Op0RHS, *Op1LHS, *Op1RHS;
4764 FCmpInst::Predicate Op0CC, Op1CC;
4765 if (match(Op0, m_FCmp(Op0CC, m_Value(Op0LHS), m_Value(Op0RHS))) &&
4766 match(Op1, m_FCmp(Op1CC, m_Value(Op1LHS), m_Value(Op1RHS)))) {
4767 if (Op0LHS == Op1RHS && Op0RHS == Op1LHS) {
4768 // Swap RHS operands to match LHS.
4769 Op1CC = FCmpInst::getSwappedPredicate(Op1CC);
4770 std::swap(Op1LHS, Op1RHS);
4772 if (Op0LHS == Op1LHS && Op0RHS == Op1RHS) {
4773 // Simplify (fcmp cc0 x, y) | (fcmp cc1 x, y).
4775 return new FCmpInst((FCmpInst::Predicate)Op0CC, Op0LHS, Op0RHS);
4776 else if (Op0CC == FCmpInst::FCMP_TRUE ||
4777 Op1CC == FCmpInst::FCMP_TRUE)
4778 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4779 else if (Op0CC == FCmpInst::FCMP_FALSE)
4780 return ReplaceInstUsesWith(I, Op1);
4781 else if (Op1CC == FCmpInst::FCMP_FALSE)
4782 return ReplaceInstUsesWith(I, Op0);
4785 unsigned Op0Pred = getFCmpCode(Op0CC, Op0Ordered);
4786 unsigned Op1Pred = getFCmpCode(Op1CC, Op1Ordered);
4787 if (Op0Ordered == Op1Ordered) {
4788 // If both are ordered or unordered, return a new fcmp with
4789 // or'ed predicates.
4790 Value *RV = getFCmpValue(Op0Ordered, Op0Pred|Op1Pred,
4792 if (Instruction *I = dyn_cast<Instruction>(RV))
4794 // Otherwise, it's a constant boolean value...
4795 return ReplaceInstUsesWith(I, RV);
4803 return Changed ? &I : 0;
4808 // XorSelf - Implements: X ^ X --> 0
4811 XorSelf(Value *rhs) : RHS(rhs) {}
4812 bool shouldApply(Value *LHS) const { return LHS == RHS; }
4813 Instruction *apply(BinaryOperator &Xor) const {
4820 Instruction *InstCombiner::visitXor(BinaryOperator &I) {
4821 bool Changed = SimplifyCommutative(I);
4822 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4824 if (isa<UndefValue>(Op1)) {
4825 if (isa<UndefValue>(Op0))
4826 // Handle undef ^ undef -> 0 special case. This is a common
4828 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
4829 return ReplaceInstUsesWith(I, Op1); // X ^ undef -> undef
4832 // xor X, X = 0, even if X is nested in a sequence of Xor's.
4833 if (Instruction *Result = AssociativeOpt(I, XorSelf(Op1))) {
4834 assert(Result == &I && "AssociativeOpt didn't work?"); Result=Result;
4835 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
4838 // See if we can simplify any instructions used by the instruction whose sole
4839 // purpose is to compute bits we don't care about.
4840 if (!isa<VectorType>(I.getType())) {
4841 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
4842 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
4843 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
4844 KnownZero, KnownOne))
4846 } else if (isa<ConstantAggregateZero>(Op1)) {
4847 return ReplaceInstUsesWith(I, Op0); // X ^ <0,0> -> X
4850 // Is this a ~ operation?
4851 if (Value *NotOp = dyn_castNotVal(&I)) {
4852 // ~(~X & Y) --> (X | ~Y) - De Morgan's Law
4853 // ~(~X | Y) === (X & ~Y) - De Morgan's Law
4854 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(NotOp)) {
4855 if (Op0I->getOpcode() == Instruction::And ||
4856 Op0I->getOpcode() == Instruction::Or) {
4857 if (dyn_castNotVal(Op0I->getOperand(1))) Op0I->swapOperands();
4858 if (Value *Op0NotVal = dyn_castNotVal(Op0I->getOperand(0))) {
4860 BinaryOperator::CreateNot(Op0I->getOperand(1),
4861 Op0I->getOperand(1)->getName()+".not");
4862 InsertNewInstBefore(NotY, I);
4863 if (Op0I->getOpcode() == Instruction::And)
4864 return BinaryOperator::CreateOr(Op0NotVal, NotY);
4866 return BinaryOperator::CreateAnd(Op0NotVal, NotY);
4873 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
4874 if (RHS == ConstantInt::getTrue() && Op0->hasOneUse()) {
4875 // xor (cmp A, B), true = not (cmp A, B) = !cmp A, B
4876 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Op0))
4877 return new ICmpInst(ICI->getInversePredicate(),
4878 ICI->getOperand(0), ICI->getOperand(1));
4880 if (FCmpInst *FCI = dyn_cast<FCmpInst>(Op0))
4881 return new FCmpInst(FCI->getInversePredicate(),
4882 FCI->getOperand(0), FCI->getOperand(1));
4885 // fold (xor(zext(cmp)), 1) and (xor(sext(cmp)), -1) to ext(!cmp).
4886 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
4887 if (CmpInst *CI = dyn_cast<CmpInst>(Op0C->getOperand(0))) {
4888 if (CI->hasOneUse() && Op0C->hasOneUse()) {
4889 Instruction::CastOps Opcode = Op0C->getOpcode();
4890 if (Opcode == Instruction::ZExt || Opcode == Instruction::SExt) {
4891 if (RHS == ConstantExpr::getCast(Opcode, ConstantInt::getTrue(),
4892 Op0C->getDestTy())) {
4893 Instruction *NewCI = InsertNewInstBefore(CmpInst::Create(
4894 CI->getOpcode(), CI->getInversePredicate(),
4895 CI->getOperand(0), CI->getOperand(1)), I);
4896 NewCI->takeName(CI);
4897 return CastInst::Create(Opcode, NewCI, Op0C->getType());
4904 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
4905 // ~(c-X) == X-c-1 == X+(-c-1)
4906 if (Op0I->getOpcode() == Instruction::Sub && RHS->isAllOnesValue())
4907 if (Constant *Op0I0C = dyn_cast<Constant>(Op0I->getOperand(0))) {
4908 Constant *NegOp0I0C = ConstantExpr::getNeg(Op0I0C);
4909 Constant *ConstantRHS = ConstantExpr::getSub(NegOp0I0C,
4910 ConstantInt::get(I.getType(), 1));
4911 return BinaryOperator::CreateAdd(Op0I->getOperand(1), ConstantRHS);
4914 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) {
4915 if (Op0I->getOpcode() == Instruction::Add) {
4916 // ~(X-c) --> (-c-1)-X
4917 if (RHS->isAllOnesValue()) {
4918 Constant *NegOp0CI = ConstantExpr::getNeg(Op0CI);
4919 return BinaryOperator::CreateSub(
4920 ConstantExpr::getSub(NegOp0CI,
4921 ConstantInt::get(I.getType(), 1)),
4922 Op0I->getOperand(0));
4923 } else if (RHS->getValue().isSignBit()) {
4924 // (X + C) ^ signbit -> (X + C + signbit)
4925 Constant *C = ConstantInt::get(RHS->getValue() + Op0CI->getValue());
4926 return BinaryOperator::CreateAdd(Op0I->getOperand(0), C);
4929 } else if (Op0I->getOpcode() == Instruction::Or) {
4930 // (X|C1)^C2 -> X^(C1|C2) iff X&~C1 == 0
4931 if (MaskedValueIsZero(Op0I->getOperand(0), Op0CI->getValue())) {
4932 Constant *NewRHS = ConstantExpr::getOr(Op0CI, RHS);
4933 // Anything in both C1 and C2 is known to be zero, remove it from
4935 Constant *CommonBits = And(Op0CI, RHS);
4936 NewRHS = ConstantExpr::getAnd(NewRHS,
4937 ConstantExpr::getNot(CommonBits));
4938 AddToWorkList(Op0I);
4939 I.setOperand(0, Op0I->getOperand(0));
4940 I.setOperand(1, NewRHS);
4947 // Try to fold constant and into select arguments.
4948 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
4949 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
4951 if (isa<PHINode>(Op0))
4952 if (Instruction *NV = FoldOpIntoPhi(I))
4956 if (Value *X = dyn_castNotVal(Op0)) // ~A ^ A == -1
4958 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4960 if (Value *X = dyn_castNotVal(Op1)) // A ^ ~A == -1
4962 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4965 BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1);
4968 if (match(Op1I, m_Or(m_Value(A), m_Value(B)))) {
4969 if (A == Op0) { // B^(B|A) == (A|B)^B
4970 Op1I->swapOperands();
4972 std::swap(Op0, Op1);
4973 } else if (B == Op0) { // B^(A|B) == (A|B)^B
4974 I.swapOperands(); // Simplified below.
4975 std::swap(Op0, Op1);
4977 } else if (match(Op1I, m_Xor(m_Specific(Op0), m_Value(B)))) {
4978 return ReplaceInstUsesWith(I, B); // A^(A^B) == B
4979 } else if (match(Op1I, m_Xor(m_Value(A), m_Specific(Op0)))) {
4980 return ReplaceInstUsesWith(I, A); // A^(B^A) == B
4981 } else if (match(Op1I, m_And(m_Value(A), m_Value(B))) && Op1I->hasOneUse()){
4982 if (A == Op0) { // A^(A&B) -> A^(B&A)
4983 Op1I->swapOperands();
4986 if (B == Op0) { // A^(B&A) -> (B&A)^A
4987 I.swapOperands(); // Simplified below.
4988 std::swap(Op0, Op1);
4993 BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0);
4996 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) && Op0I->hasOneUse()) {
4997 if (A == Op1) // (B|A)^B == (A|B)^B
4999 if (B == Op1) { // (A|B)^B == A & ~B
5001 InsertNewInstBefore(BinaryOperator::CreateNot(Op1, "tmp"), I);
5002 return BinaryOperator::CreateAnd(A, NotB);
5004 } else if (match(Op0I, m_Xor(m_Specific(Op1), m_Value(B)))) {
5005 return ReplaceInstUsesWith(I, B); // (A^B)^A == B
5006 } else if (match(Op0I, m_Xor(m_Value(A), m_Specific(Op1)))) {
5007 return ReplaceInstUsesWith(I, A); // (B^A)^A == B
5008 } else if (match(Op0I, m_And(m_Value(A), m_Value(B))) && Op0I->hasOneUse()){
5009 if (A == Op1) // (A&B)^A -> (B&A)^A
5011 if (B == Op1 && // (B&A)^A == ~B & A
5012 !isa<ConstantInt>(Op1)) { // Canonical form is (B&C)^C
5014 InsertNewInstBefore(BinaryOperator::CreateNot(A, "tmp"), I);
5015 return BinaryOperator::CreateAnd(N, Op1);
5020 // (X >> Z) ^ (Y >> Z) -> (X^Y) >> Z for all shifts.
5021 if (Op0I && Op1I && Op0I->isShift() &&
5022 Op0I->getOpcode() == Op1I->getOpcode() &&
5023 Op0I->getOperand(1) == Op1I->getOperand(1) &&
5024 (Op1I->hasOneUse() || Op1I->hasOneUse())) {
5025 Instruction *NewOp =
5026 InsertNewInstBefore(BinaryOperator::CreateXor(Op0I->getOperand(0),
5027 Op1I->getOperand(0),
5028 Op0I->getName()), I);
5029 return BinaryOperator::Create(Op1I->getOpcode(), NewOp,
5030 Op1I->getOperand(1));
5034 Value *A, *B, *C, *D;
5035 // (A & B)^(A | B) -> A ^ B
5036 if (match(Op0I, m_And(m_Value(A), m_Value(B))) &&
5037 match(Op1I, m_Or(m_Value(C), m_Value(D)))) {
5038 if ((A == C && B == D) || (A == D && B == C))
5039 return BinaryOperator::CreateXor(A, B);
5041 // (A | B)^(A & B) -> A ^ B
5042 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) &&
5043 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
5044 if ((A == C && B == D) || (A == D && B == C))
5045 return BinaryOperator::CreateXor(A, B);
5049 if ((Op0I->hasOneUse() || Op1I->hasOneUse()) &&
5050 match(Op0I, m_And(m_Value(A), m_Value(B))) &&
5051 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
5052 // (X & Y)^(X & Y) -> (Y^Z) & X
5053 Value *X = 0, *Y = 0, *Z = 0;
5055 X = A, Y = B, Z = D;
5057 X = A, Y = B, Z = C;
5059 X = B, Y = A, Z = D;
5061 X = B, Y = A, Z = C;
5064 Instruction *NewOp =
5065 InsertNewInstBefore(BinaryOperator::CreateXor(Y, Z, Op0->getName()), I);
5066 return BinaryOperator::CreateAnd(NewOp, X);
5071 // (icmp1 A, B) ^ (icmp2 A, B) --> (icmp3 A, B)
5072 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1)))
5073 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
5076 // fold (xor (cast A), (cast B)) -> (cast (xor A, B))
5077 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
5078 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
5079 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind?
5080 const Type *SrcTy = Op0C->getOperand(0)->getType();
5081 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
5082 // Only do this if the casts both really cause code to be generated.
5083 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
5085 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
5087 Instruction *NewOp = BinaryOperator::CreateXor(Op0C->getOperand(0),
5088 Op1C->getOperand(0),
5090 InsertNewInstBefore(NewOp, I);
5091 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
5096 return Changed ? &I : 0;
5099 /// AddWithOverflow - Compute Result = In1+In2, returning true if the result
5100 /// overflowed for this type.
5101 static bool AddWithOverflow(ConstantInt *&Result, ConstantInt *In1,
5102 ConstantInt *In2, bool IsSigned = false) {
5103 Result = cast<ConstantInt>(Add(In1, In2));
5106 if (In2->getValue().isNegative())
5107 return Result->getValue().sgt(In1->getValue());
5109 return Result->getValue().slt(In1->getValue());
5111 return Result->getValue().ult(In1->getValue());
5114 /// SubWithOverflow - Compute Result = In1-In2, returning true if the result
5115 /// overflowed for this type.
5116 static bool SubWithOverflow(ConstantInt *&Result, ConstantInt *In1,
5117 ConstantInt *In2, bool IsSigned = false) {
5118 Result = cast<ConstantInt>(Subtract(In1, In2));
5121 if (In2->getValue().isNegative())
5122 return Result->getValue().slt(In1->getValue());
5124 return Result->getValue().sgt(In1->getValue());
5126 return Result->getValue().ugt(In1->getValue());
5129 /// EmitGEPOffset - Given a getelementptr instruction/constantexpr, emit the
5130 /// code necessary to compute the offset from the base pointer (without adding
5131 /// in the base pointer). Return the result as a signed integer of intptr size.
5132 static Value *EmitGEPOffset(User *GEP, Instruction &I, InstCombiner &IC) {
5133 TargetData &TD = IC.getTargetData();
5134 gep_type_iterator GTI = gep_type_begin(GEP);
5135 const Type *IntPtrTy = TD.getIntPtrType();
5136 Value *Result = Constant::getNullValue(IntPtrTy);
5138 // Build a mask for high order bits.
5139 unsigned IntPtrWidth = TD.getPointerSizeInBits();
5140 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth);
5142 for (User::op_iterator i = GEP->op_begin() + 1, e = GEP->op_end(); i != e;
5145 uint64_t Size = TD.getABITypeSize(GTI.getIndexedType()) & PtrSizeMask;
5146 if (ConstantInt *OpC = dyn_cast<ConstantInt>(Op)) {
5147 if (OpC->isZero()) continue;
5149 // Handle a struct index, which adds its field offset to the pointer.
5150 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
5151 Size = TD.getStructLayout(STy)->getElementOffset(OpC->getZExtValue());
5153 if (ConstantInt *RC = dyn_cast<ConstantInt>(Result))
5154 Result = ConstantInt::get(RC->getValue() + APInt(IntPtrWidth, Size));
5156 Result = IC.InsertNewInstBefore(
5157 BinaryOperator::CreateAdd(Result,
5158 ConstantInt::get(IntPtrTy, Size),
5159 GEP->getName()+".offs"), I);
5163 Constant *Scale = ConstantInt::get(IntPtrTy, Size);
5164 Constant *OC = ConstantExpr::getIntegerCast(OpC, IntPtrTy, true /*SExt*/);
5165 Scale = ConstantExpr::getMul(OC, Scale);
5166 if (Constant *RC = dyn_cast<Constant>(Result))
5167 Result = ConstantExpr::getAdd(RC, Scale);
5169 // Emit an add instruction.
5170 Result = IC.InsertNewInstBefore(
5171 BinaryOperator::CreateAdd(Result, Scale,
5172 GEP->getName()+".offs"), I);
5176 // Convert to correct type.
5177 if (Op->getType() != IntPtrTy) {
5178 if (Constant *OpC = dyn_cast<Constant>(Op))
5179 Op = ConstantExpr::getSExt(OpC, IntPtrTy);
5181 Op = IC.InsertNewInstBefore(new SExtInst(Op, IntPtrTy,
5182 Op->getName()+".c"), I);
5185 Constant *Scale = ConstantInt::get(IntPtrTy, Size);
5186 if (Constant *OpC = dyn_cast<Constant>(Op))
5187 Op = ConstantExpr::getMul(OpC, Scale);
5188 else // We'll let instcombine(mul) convert this to a shl if possible.
5189 Op = IC.InsertNewInstBefore(BinaryOperator::CreateMul(Op, Scale,
5190 GEP->getName()+".idx"), I);
5193 // Emit an add instruction.
5194 if (isa<Constant>(Op) && isa<Constant>(Result))
5195 Result = ConstantExpr::getAdd(cast<Constant>(Op),
5196 cast<Constant>(Result));
5198 Result = IC.InsertNewInstBefore(BinaryOperator::CreateAdd(Op, Result,
5199 GEP->getName()+".offs"), I);
5205 /// EvaluateGEPOffsetExpression - Return an value that can be used to compare of
5206 /// the *offset* implied by GEP to zero. For example, if we have &A[i], we want
5207 /// to return 'i' for "icmp ne i, 0". Note that, in general, indices can be
5208 /// complex, and scales are involved. The above expression would also be legal
5209 /// to codegen as "icmp ne (i*4), 0" (assuming A is a pointer to i32). This
5210 /// later form is less amenable to optimization though, and we are allowed to
5211 /// generate the first by knowing that pointer arithmetic doesn't overflow.
5213 /// If we can't emit an optimized form for this expression, this returns null.
5215 static Value *EvaluateGEPOffsetExpression(User *GEP, Instruction &I,
5217 TargetData &TD = IC.getTargetData();
5218 gep_type_iterator GTI = gep_type_begin(GEP);
5220 // Check to see if this gep only has a single variable index. If so, and if
5221 // any constant indices are a multiple of its scale, then we can compute this
5222 // in terms of the scale of the variable index. For example, if the GEP
5223 // implies an offset of "12 + i*4", then we can codegen this as "3 + i",
5224 // because the expression will cross zero at the same point.
5225 unsigned i, e = GEP->getNumOperands();
5227 for (i = 1; i != e; ++i, ++GTI) {
5228 if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i))) {
5229 // Compute the aggregate offset of constant indices.
5230 if (CI->isZero()) continue;
5232 // Handle a struct index, which adds its field offset to the pointer.
5233 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
5234 Offset += TD.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
5236 uint64_t Size = TD.getABITypeSize(GTI.getIndexedType());
5237 Offset += Size*CI->getSExtValue();
5240 // Found our variable index.
5245 // If there are no variable indices, we must have a constant offset, just
5246 // evaluate it the general way.
5247 if (i == e) return 0;
5249 Value *VariableIdx = GEP->getOperand(i);
5250 // Determine the scale factor of the variable element. For example, this is
5251 // 4 if the variable index is into an array of i32.
5252 uint64_t VariableScale = TD.getABITypeSize(GTI.getIndexedType());
5254 // Verify that there are no other variable indices. If so, emit the hard way.
5255 for (++i, ++GTI; i != e; ++i, ++GTI) {
5256 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();
5271 // Okay, we know we have a single variable index, which must be a
5272 // pointer/array/vector index. If there is no offset, life is simple, return
5274 unsigned IntPtrWidth = TD.getPointerSizeInBits();
5276 // Cast to intptrty in case a truncation occurs. If an extension is needed,
5277 // we don't need to bother extending: the extension won't affect where the
5278 // computation crosses zero.
5279 if (VariableIdx->getType()->getPrimitiveSizeInBits() > IntPtrWidth)
5280 VariableIdx = new TruncInst(VariableIdx, TD.getIntPtrType(),
5281 VariableIdx->getNameStart(), &I);
5285 // Otherwise, there is an index. The computation we will do will be modulo
5286 // the pointer size, so get it.
5287 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth);
5289 Offset &= PtrSizeMask;
5290 VariableScale &= PtrSizeMask;
5292 // To do this transformation, any constant index must be a multiple of the
5293 // variable scale factor. For example, we can evaluate "12 + 4*i" as "3 + i",
5294 // but we can't evaluate "10 + 3*i" in terms of i. Check that the offset is a
5295 // multiple of the variable scale.
5296 int64_t NewOffs = Offset / (int64_t)VariableScale;
5297 if (Offset != NewOffs*(int64_t)VariableScale)
5300 // Okay, we can do this evaluation. Start by converting the index to intptr.
5301 const Type *IntPtrTy = TD.getIntPtrType();
5302 if (VariableIdx->getType() != IntPtrTy)
5303 VariableIdx = CastInst::CreateIntegerCast(VariableIdx, IntPtrTy,
5305 VariableIdx->getNameStart(), &I);
5306 Constant *OffsetVal = ConstantInt::get(IntPtrTy, NewOffs);
5307 return BinaryOperator::CreateAdd(VariableIdx, OffsetVal, "offset", &I);
5311 /// FoldGEPICmp - Fold comparisons between a GEP instruction and something
5312 /// else. At this point we know that the GEP is on the LHS of the comparison.
5313 Instruction *InstCombiner::FoldGEPICmp(User *GEPLHS, Value *RHS,
5314 ICmpInst::Predicate Cond,
5316 assert(dyn_castGetElementPtr(GEPLHS) && "LHS is not a getelementptr!");
5318 // Look through bitcasts.
5319 if (BitCastInst *BCI = dyn_cast<BitCastInst>(RHS))
5320 RHS = BCI->getOperand(0);
5322 Value *PtrBase = GEPLHS->getOperand(0);
5323 if (PtrBase == RHS) {
5324 // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
5325 // This transformation (ignoring the base and scales) is valid because we
5326 // know pointers can't overflow. See if we can output an optimized form.
5327 Value *Offset = EvaluateGEPOffsetExpression(GEPLHS, I, *this);
5329 // If not, synthesize the offset the hard way.
5331 Offset = EmitGEPOffset(GEPLHS, I, *this);
5332 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Offset,
5333 Constant::getNullValue(Offset->getType()));
5334 } else if (User *GEPRHS = dyn_castGetElementPtr(RHS)) {
5335 // If the base pointers are different, but the indices are the same, just
5336 // compare the base pointer.
5337 if (PtrBase != GEPRHS->getOperand(0)) {
5338 bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands();
5339 IndicesTheSame &= GEPLHS->getOperand(0)->getType() ==
5340 GEPRHS->getOperand(0)->getType();
5342 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
5343 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
5344 IndicesTheSame = false;
5348 // If all indices are the same, just compare the base pointers.
5350 return new ICmpInst(ICmpInst::getSignedPredicate(Cond),
5351 GEPLHS->getOperand(0), GEPRHS->getOperand(0));
5353 // Otherwise, the base pointers are different and the indices are
5354 // different, bail out.
5358 // If one of the GEPs has all zero indices, recurse.
5359 bool AllZeros = true;
5360 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
5361 if (!isa<Constant>(GEPLHS->getOperand(i)) ||
5362 !cast<Constant>(GEPLHS->getOperand(i))->isNullValue()) {
5367 return FoldGEPICmp(GEPRHS, GEPLHS->getOperand(0),
5368 ICmpInst::getSwappedPredicate(Cond), I);
5370 // If the other GEP has all zero indices, recurse.
5372 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
5373 if (!isa<Constant>(GEPRHS->getOperand(i)) ||
5374 !cast<Constant>(GEPRHS->getOperand(i))->isNullValue()) {
5379 return FoldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I);
5381 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
5382 // If the GEPs only differ by one index, compare it.
5383 unsigned NumDifferences = 0; // Keep track of # differences.
5384 unsigned DiffOperand = 0; // The operand that differs.
5385 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
5386 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
5387 if (GEPLHS->getOperand(i)->getType()->getPrimitiveSizeInBits() !=
5388 GEPRHS->getOperand(i)->getType()->getPrimitiveSizeInBits()) {
5389 // Irreconcilable differences.
5393 if (NumDifferences++) break;
5398 if (NumDifferences == 0) // SAME GEP?
5399 return ReplaceInstUsesWith(I, // No comparison is needed here.
5400 ConstantInt::get(Type::Int1Ty,
5401 ICmpInst::isTrueWhenEqual(Cond)));
5403 else if (NumDifferences == 1) {
5404 Value *LHSV = GEPLHS->getOperand(DiffOperand);
5405 Value *RHSV = GEPRHS->getOperand(DiffOperand);
5406 // Make sure we do a signed comparison here.
5407 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), LHSV, RHSV);
5411 // Only lower this if the icmp is the only user of the GEP or if we expect
5412 // the result to fold to a constant!
5413 if ((isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
5414 (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
5415 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
5416 Value *L = EmitGEPOffset(GEPLHS, I, *this);
5417 Value *R = EmitGEPOffset(GEPRHS, I, *this);
5418 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), L, R);
5424 /// FoldFCmp_IntToFP_Cst - Fold fcmp ([us]itofp x, cst) if possible.
5426 Instruction *InstCombiner::FoldFCmp_IntToFP_Cst(FCmpInst &I,
5429 if (!isa<ConstantFP>(RHSC)) return 0;
5430 const APFloat &RHS = cast<ConstantFP>(RHSC)->getValueAPF();
5432 // Get the width of the mantissa. We don't want to hack on conversions that
5433 // might lose information from the integer, e.g. "i64 -> float"
5434 int MantissaWidth = LHSI->getType()->getFPMantissaWidth();
5435 if (MantissaWidth == -1) return 0; // Unknown.
5437 // Check to see that the input is converted from an integer type that is small
5438 // enough that preserves all bits. TODO: check here for "known" sign bits.
5439 // This would allow us to handle (fptosi (x >>s 62) to float) if x is i64 f.e.
5440 unsigned InputSize = LHSI->getOperand(0)->getType()->getPrimitiveSizeInBits();
5442 // If this is a uitofp instruction, we need an extra bit to hold the sign.
5443 bool LHSUnsigned = isa<UIToFPInst>(LHSI);
5447 // If the conversion would lose info, don't hack on this.
5448 if ((int)InputSize > MantissaWidth)
5451 // Otherwise, we can potentially simplify the comparison. We know that it
5452 // will always come through as an integer value and we know the constant is
5453 // not a NAN (it would have been previously simplified).
5454 assert(!RHS.isNaN() && "NaN comparison not already folded!");
5456 ICmpInst::Predicate Pred;
5457 switch (I.getPredicate()) {
5458 default: assert(0 && "Unexpected predicate!");
5459 case FCmpInst::FCMP_UEQ:
5460 case FCmpInst::FCMP_OEQ:
5461 Pred = ICmpInst::ICMP_EQ;
5463 case FCmpInst::FCMP_UGT:
5464 case FCmpInst::FCMP_OGT:
5465 Pred = LHSUnsigned ? ICmpInst::ICMP_UGT : ICmpInst::ICMP_SGT;
5467 case FCmpInst::FCMP_UGE:
5468 case FCmpInst::FCMP_OGE:
5469 Pred = LHSUnsigned ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_SGE;
5471 case FCmpInst::FCMP_ULT:
5472 case FCmpInst::FCMP_OLT:
5473 Pred = LHSUnsigned ? ICmpInst::ICMP_ULT : ICmpInst::ICMP_SLT;
5475 case FCmpInst::FCMP_ULE:
5476 case FCmpInst::FCMP_OLE:
5477 Pred = LHSUnsigned ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_SLE;
5479 case FCmpInst::FCMP_UNE:
5480 case FCmpInst::FCMP_ONE:
5481 Pred = ICmpInst::ICMP_NE;
5483 case FCmpInst::FCMP_ORD:
5484 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5485 case FCmpInst::FCMP_UNO:
5486 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5489 const IntegerType *IntTy = cast<IntegerType>(LHSI->getOperand(0)->getType());
5491 // Now we know that the APFloat is a normal number, zero or inf.
5493 // See if the FP constant is too large for the integer. For example,
5494 // comparing an i8 to 300.0.
5495 unsigned IntWidth = IntTy->getPrimitiveSizeInBits();
5498 // If the RHS value is > SignedMax, fold the comparison. This handles +INF
5499 // and large values.
5500 APFloat SMax(RHS.getSemantics(), APFloat::fcZero, false);
5501 SMax.convertFromAPInt(APInt::getSignedMaxValue(IntWidth), true,
5502 APFloat::rmNearestTiesToEven);
5503 if (SMax.compare(RHS) == APFloat::cmpLessThan) { // smax < 13123.0
5504 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SLT ||
5505 Pred == ICmpInst::ICMP_SLE)
5506 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5507 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5510 // If the RHS value is > UnsignedMax, fold the comparison. This handles
5511 // +INF and large values.
5512 APFloat UMax(RHS.getSemantics(), APFloat::fcZero, false);
5513 UMax.convertFromAPInt(APInt::getMaxValue(IntWidth), false,
5514 APFloat::rmNearestTiesToEven);
5515 if (UMax.compare(RHS) == APFloat::cmpLessThan) { // umax < 13123.0
5516 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_ULT ||
5517 Pred == ICmpInst::ICMP_ULE)
5518 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5519 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5524 // See if the RHS value is < SignedMin.
5525 APFloat SMin(RHS.getSemantics(), APFloat::fcZero, false);
5526 SMin.convertFromAPInt(APInt::getSignedMinValue(IntWidth), true,
5527 APFloat::rmNearestTiesToEven);
5528 if (SMin.compare(RHS) == APFloat::cmpGreaterThan) { // smin > 12312.0
5529 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SGT ||
5530 Pred == ICmpInst::ICMP_SGE)
5531 return ReplaceInstUsesWith(I,ConstantInt::getTrue());
5532 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5536 // Okay, now we know that the FP constant fits in the range [SMIN, SMAX] or
5537 // [0, UMAX], but it may still be fractional. See if it is fractional by
5538 // casting the FP value to the integer value and back, checking for equality.
5539 // Don't do this for zero, because -0.0 is not fractional.
5540 Constant *RHSInt = ConstantExpr::getFPToSI(RHSC, IntTy);
5541 if (!RHS.isZero() &&
5542 ConstantExpr::getSIToFP(RHSInt, RHSC->getType()) != RHSC) {
5543 // If we had a comparison against a fractional value, we have to adjust the
5544 // compare predicate and sometimes the value. RHSC is rounded towards zero
5547 default: assert(0 && "Unexpected integer comparison!");
5548 case ICmpInst::ICMP_NE: // (float)int != 4.4 --> true
5549 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5550 case ICmpInst::ICMP_EQ: // (float)int == 4.4 --> false
5551 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5552 case ICmpInst::ICMP_ULE:
5553 // (float)int <= 4.4 --> int <= 4
5554 // (float)int <= -4.4 --> false
5555 if (RHS.isNegative())
5556 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5558 case ICmpInst::ICMP_SLE:
5559 // (float)int <= 4.4 --> int <= 4
5560 // (float)int <= -4.4 --> int < -4
5561 if (RHS.isNegative())
5562 Pred = ICmpInst::ICMP_SLT;
5564 case ICmpInst::ICMP_ULT:
5565 // (float)int < -4.4 --> false
5566 // (float)int < 4.4 --> int <= 4
5567 if (RHS.isNegative())
5568 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5569 Pred = ICmpInst::ICMP_ULE;
5571 case ICmpInst::ICMP_SLT:
5572 // (float)int < -4.4 --> int < -4
5573 // (float)int < 4.4 --> int <= 4
5574 if (!RHS.isNegative())
5575 Pred = ICmpInst::ICMP_SLE;
5577 case ICmpInst::ICMP_UGT:
5578 // (float)int > 4.4 --> int > 4
5579 // (float)int > -4.4 --> true
5580 if (RHS.isNegative())
5581 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5583 case ICmpInst::ICMP_SGT:
5584 // (float)int > 4.4 --> int > 4
5585 // (float)int > -4.4 --> int >= -4
5586 if (RHS.isNegative())
5587 Pred = ICmpInst::ICMP_SGE;
5589 case ICmpInst::ICMP_UGE:
5590 // (float)int >= -4.4 --> true
5591 // (float)int >= 4.4 --> int > 4
5592 if (!RHS.isNegative())
5593 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5594 Pred = ICmpInst::ICMP_UGT;
5596 case ICmpInst::ICMP_SGE:
5597 // (float)int >= -4.4 --> int >= -4
5598 // (float)int >= 4.4 --> int > 4
5599 if (!RHS.isNegative())
5600 Pred = ICmpInst::ICMP_SGT;
5605 // Lower this FP comparison into an appropriate integer version of the
5607 return new ICmpInst(Pred, LHSI->getOperand(0), RHSInt);
5610 Instruction *InstCombiner::visitFCmpInst(FCmpInst &I) {
5611 bool Changed = SimplifyCompare(I);
5612 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5614 // Fold trivial predicates.
5615 if (I.getPredicate() == FCmpInst::FCMP_FALSE)
5616 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5617 if (I.getPredicate() == FCmpInst::FCMP_TRUE)
5618 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5620 // Simplify 'fcmp pred X, X'
5622 switch (I.getPredicate()) {
5623 default: assert(0 && "Unknown predicate!");
5624 case FCmpInst::FCMP_UEQ: // True if unordered or equal
5625 case FCmpInst::FCMP_UGE: // True if unordered, greater than, or equal
5626 case FCmpInst::FCMP_ULE: // True if unordered, less than, or equal
5627 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5628 case FCmpInst::FCMP_OGT: // True if ordered and greater than
5629 case FCmpInst::FCMP_OLT: // True if ordered and less than
5630 case FCmpInst::FCMP_ONE: // True if ordered and operands are unequal
5631 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5633 case FCmpInst::FCMP_UNO: // True if unordered: isnan(X) | isnan(Y)
5634 case FCmpInst::FCMP_ULT: // True if unordered or less than
5635 case FCmpInst::FCMP_UGT: // True if unordered or greater than
5636 case FCmpInst::FCMP_UNE: // True if unordered or not equal
5637 // Canonicalize these to be 'fcmp uno %X, 0.0'.
5638 I.setPredicate(FCmpInst::FCMP_UNO);
5639 I.setOperand(1, Constant::getNullValue(Op0->getType()));
5642 case FCmpInst::FCMP_ORD: // True if ordered (no nans)
5643 case FCmpInst::FCMP_OEQ: // True if ordered and equal
5644 case FCmpInst::FCMP_OGE: // True if ordered and greater than or equal
5645 case FCmpInst::FCMP_OLE: // True if ordered and less than or equal
5646 // Canonicalize these to be 'fcmp ord %X, 0.0'.
5647 I.setPredicate(FCmpInst::FCMP_ORD);
5648 I.setOperand(1, Constant::getNullValue(Op0->getType()));
5653 if (isa<UndefValue>(Op1)) // fcmp pred X, undef -> undef
5654 return ReplaceInstUsesWith(I, UndefValue::get(Type::Int1Ty));
5656 // Handle fcmp with constant RHS
5657 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
5658 // If the constant is a nan, see if we can fold the comparison based on it.
5659 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHSC)) {
5660 if (CFP->getValueAPF().isNaN()) {
5661 if (FCmpInst::isOrdered(I.getPredicate())) // True if ordered and...
5662 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5663 assert(FCmpInst::isUnordered(I.getPredicate()) &&
5664 "Comparison must be either ordered or unordered!");
5665 // True if unordered.
5666 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5670 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
5671 switch (LHSI->getOpcode()) {
5672 case Instruction::PHI:
5673 // Only fold fcmp into the PHI if the phi and fcmp are in the same
5674 // block. If in the same block, we're encouraging jump threading. If
5675 // not, we are just pessimizing the code by making an i1 phi.
5676 if (LHSI->getParent() == I.getParent())
5677 if (Instruction *NV = FoldOpIntoPhi(I))
5680 case Instruction::SIToFP:
5681 case Instruction::UIToFP:
5682 if (Instruction *NV = FoldFCmp_IntToFP_Cst(I, LHSI, RHSC))
5685 case Instruction::Select:
5686 // If either operand of the select is a constant, we can fold the
5687 // comparison into the select arms, which will cause one to be
5688 // constant folded and the select turned into a bitwise or.
5689 Value *Op1 = 0, *Op2 = 0;
5690 if (LHSI->hasOneUse()) {
5691 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
5692 // Fold the known value into the constant operand.
5693 Op1 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
5694 // Insert a new FCmp of the other select operand.
5695 Op2 = InsertNewInstBefore(new FCmpInst(I.getPredicate(),
5696 LHSI->getOperand(2), RHSC,
5698 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
5699 // Fold the known value into the constant operand.
5700 Op2 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
5701 // Insert a new FCmp of the other select operand.
5702 Op1 = InsertNewInstBefore(new FCmpInst(I.getPredicate(),
5703 LHSI->getOperand(1), RHSC,
5709 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
5714 return Changed ? &I : 0;
5717 Instruction *InstCombiner::visitICmpInst(ICmpInst &I) {
5718 bool Changed = SimplifyCompare(I);
5719 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5720 const Type *Ty = Op0->getType();
5724 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
5725 I.isTrueWhenEqual()));
5727 if (isa<UndefValue>(Op1)) // X icmp undef -> undef
5728 return ReplaceInstUsesWith(I, UndefValue::get(Type::Int1Ty));
5730 // icmp <global/alloca*/null>, <global/alloca*/null> - Global/Stack value
5731 // addresses never equal each other! We already know that Op0 != Op1.
5732 if ((isa<GlobalValue>(Op0) || isa<AllocaInst>(Op0) ||
5733 isa<ConstantPointerNull>(Op0)) &&
5734 (isa<GlobalValue>(Op1) || isa<AllocaInst>(Op1) ||
5735 isa<ConstantPointerNull>(Op1)))
5736 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
5737 !I.isTrueWhenEqual()));
5739 // icmp's with boolean values can always be turned into bitwise operations
5740 if (Ty == Type::Int1Ty) {
5741 switch (I.getPredicate()) {
5742 default: assert(0 && "Invalid icmp instruction!");
5743 case ICmpInst::ICMP_EQ: { // icmp eq i1 A, B -> ~(A^B)
5744 Instruction *Xor = BinaryOperator::CreateXor(Op0, Op1, I.getName()+"tmp");
5745 InsertNewInstBefore(Xor, I);
5746 return BinaryOperator::CreateNot(Xor);
5748 case ICmpInst::ICMP_NE: // icmp eq i1 A, B -> A^B
5749 return BinaryOperator::CreateXor(Op0, Op1);
5751 case ICmpInst::ICMP_UGT:
5752 std::swap(Op0, Op1); // Change icmp ugt -> icmp ult
5754 case ICmpInst::ICMP_ULT:{ // icmp ult i1 A, B -> ~A & B
5755 Instruction *Not = BinaryOperator::CreateNot(Op0, I.getName()+"tmp");
5756 InsertNewInstBefore(Not, I);
5757 return BinaryOperator::CreateAnd(Not, Op1);
5759 case ICmpInst::ICMP_SGT:
5760 std::swap(Op0, Op1); // Change icmp sgt -> icmp slt
5762 case ICmpInst::ICMP_SLT: { // icmp slt i1 A, B -> A & ~B
5763 Instruction *Not = BinaryOperator::CreateNot(Op1, I.getName()+"tmp");
5764 InsertNewInstBefore(Not, I);
5765 return BinaryOperator::CreateAnd(Not, Op0);
5767 case ICmpInst::ICMP_UGE:
5768 std::swap(Op0, Op1); // Change icmp uge -> icmp ule
5770 case ICmpInst::ICMP_ULE: { // icmp ule i1 A, B -> ~A | B
5771 Instruction *Not = BinaryOperator::CreateNot(Op0, I.getName()+"tmp");
5772 InsertNewInstBefore(Not, I);
5773 return BinaryOperator::CreateOr(Not, Op1);
5775 case ICmpInst::ICMP_SGE:
5776 std::swap(Op0, Op1); // Change icmp sge -> icmp sle
5778 case ICmpInst::ICMP_SLE: { // icmp sle i1 A, B -> A | ~B
5779 Instruction *Not = BinaryOperator::CreateNot(Op1, I.getName()+"tmp");
5780 InsertNewInstBefore(Not, I);
5781 return BinaryOperator::CreateOr(Not, Op0);
5786 // See if we are doing a comparison with a constant.
5787 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
5790 // (icmp ne/eq (sub A B) 0) -> (icmp ne/eq A, B)
5791 if (I.isEquality() && CI->isNullValue() &&
5792 match(Op0, m_Sub(m_Value(A), m_Value(B)))) {
5793 // (icmp cond A B) if cond is equality
5794 return new ICmpInst(I.getPredicate(), A, B);
5797 // If we have an icmp le or icmp ge instruction, turn it into the
5798 // appropriate icmp lt or icmp gt instruction. This allows us to rely on
5799 // them being folded in the code below.
5800 switch (I.getPredicate()) {
5802 case ICmpInst::ICMP_ULE:
5803 if (CI->isMaxValue(false)) // A <=u MAX -> TRUE
5804 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5805 return new ICmpInst(ICmpInst::ICMP_ULT, Op0, AddOne(CI));
5806 case ICmpInst::ICMP_SLE:
5807 if (CI->isMaxValue(true)) // A <=s MAX -> TRUE
5808 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5809 return new ICmpInst(ICmpInst::ICMP_SLT, Op0, AddOne(CI));
5810 case ICmpInst::ICMP_UGE:
5811 if (CI->isMinValue(false)) // A >=u MIN -> TRUE
5812 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5813 return new ICmpInst( ICmpInst::ICMP_UGT, Op0, SubOne(CI));
5814 case ICmpInst::ICMP_SGE:
5815 if (CI->isMinValue(true)) // A >=s MIN -> TRUE
5816 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5817 return new ICmpInst(ICmpInst::ICMP_SGT, Op0, SubOne(CI));
5820 // See if we can fold the comparison based on range information we can get
5821 // by checking whether bits are known to be zero or one in the input.
5822 uint32_t BitWidth = cast<IntegerType>(Ty)->getBitWidth();
5823 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
5825 // If this comparison is a normal comparison, it demands all
5826 // bits, if it is a sign bit comparison, it only demands the sign bit.
5828 bool isSignBit = isSignBitCheck(I.getPredicate(), CI, UnusedBit);
5830 if (SimplifyDemandedBits(Op0,
5831 isSignBit ? APInt::getSignBit(BitWidth)
5832 : APInt::getAllOnesValue(BitWidth),
5833 KnownZero, KnownOne, 0))
5836 // Given the known and unknown bits, compute a range that the LHS could be
5837 // in. Compute the Min, Max and RHS values based on the known bits. For the
5838 // EQ and NE we use unsigned values.
5839 APInt Min(BitWidth, 0), Max(BitWidth, 0);
5840 if (ICmpInst::isSignedPredicate(I.getPredicate()))
5841 ComputeSignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne, Min, Max);
5843 ComputeUnsignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne,Min,Max);
5845 // If Min and Max are known to be the same, then SimplifyDemandedBits
5846 // figured out that the LHS is a constant. Just constant fold this now so
5847 // that code below can assume that Min != Max.
5849 return ReplaceInstUsesWith(I, ConstantExpr::getICmp(I.getPredicate(),
5850 ConstantInt::get(Min),
5853 // Based on the range information we know about the LHS, see if we can
5854 // simplify this comparison. For example, (x&4) < 8 is always true.
5855 const APInt &RHSVal = CI->getValue();
5856 switch (I.getPredicate()) { // LE/GE have been folded already.
5857 default: assert(0 && "Unknown icmp opcode!");
5858 case ICmpInst::ICMP_EQ:
5859 if (Max.ult(RHSVal) || Min.ugt(RHSVal))
5860 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5862 case ICmpInst::ICMP_NE:
5863 if (Max.ult(RHSVal) || Min.ugt(RHSVal))
5864 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5866 case ICmpInst::ICMP_ULT:
5867 if (Max.ult(RHSVal)) // A <u C -> true iff max(A) < C
5868 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5869 if (Min.uge(RHSVal)) // A <u C -> false iff min(A) >= C
5870 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5871 if (RHSVal == Max) // A <u MAX -> A != MAX
5872 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5873 if (RHSVal == Min+1) // A <u MIN+1 -> A == MIN
5874 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, SubOne(CI));
5876 // (x <u 2147483648) -> (x >s -1) -> true if sign bit clear
5877 if (CI->isMinValue(true))
5878 return new ICmpInst(ICmpInst::ICMP_SGT, Op0,
5879 ConstantInt::getAllOnesValue(Op0->getType()));
5881 case ICmpInst::ICMP_UGT:
5882 if (Min.ugt(RHSVal)) // A >u C -> true iff min(A) > C
5883 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5884 if (Max.ule(RHSVal)) // A >u C -> false iff max(A) <= C
5885 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5887 if (RHSVal == Min) // A >u MIN -> A != MIN
5888 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5889 if (RHSVal == Max-1) // A >u MAX-1 -> A == MAX
5890 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, AddOne(CI));
5892 // (x >u 2147483647) -> (x <s 0) -> true if sign bit set
5893 if (CI->isMaxValue(true))
5894 return new ICmpInst(ICmpInst::ICMP_SLT, Op0,
5895 ConstantInt::getNullValue(Op0->getType()));
5897 case ICmpInst::ICMP_SLT:
5898 if (Max.slt(RHSVal)) // A <s C -> true iff max(A) < C
5899 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5900 if (Min.sge(RHSVal)) // A <s C -> false iff min(A) >= C
5901 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5902 if (RHSVal == Max) // A <s MAX -> A != MAX
5903 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5904 if (RHSVal == Min+1) // A <s MIN+1 -> A == MIN
5905 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, SubOne(CI));
5907 case ICmpInst::ICMP_SGT:
5908 if (Min.sgt(RHSVal)) // A >s C -> true iff min(A) > C
5909 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5910 if (Max.sle(RHSVal)) // A >s C -> false iff max(A) <= C
5911 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5913 if (RHSVal == Min) // A >s MIN -> A != MIN
5914 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5915 if (RHSVal == Max-1) // A >s MAX-1 -> A == MAX
5916 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, AddOne(CI));
5921 // Test if the ICmpInst instruction is used exclusively by a select as
5922 // part of a minimum or maximum operation. If so, refrain from doing
5923 // any other folding. This helps out other analyses which understand
5924 // non-obfuscated minimum and maximum idioms, such as ScalarEvolution
5925 // and CodeGen. And in this case, at least one of the comparison
5926 // operands has at least one user besides the compare (the select),
5927 // which would often largely negate the benefit of folding anyway.
5929 if (SelectInst *SI = dyn_cast<SelectInst>(*I.use_begin()))
5930 if ((SI->getOperand(1) == Op0 && SI->getOperand(2) == Op1) ||
5931 (SI->getOperand(2) == Op0 && SI->getOperand(1) == Op1))
5934 // See if we are doing a comparison between a constant and an instruction that
5935 // can be folded into the comparison.
5936 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
5937 // Since the RHS is a ConstantInt (CI), if the left hand side is an
5938 // instruction, see if that instruction also has constants so that the
5939 // instruction can be folded into the icmp
5940 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
5941 if (Instruction *Res = visitICmpInstWithInstAndIntCst(I, LHSI, CI))
5945 // Handle icmp with constant (but not simple integer constant) RHS
5946 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
5947 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
5948 switch (LHSI->getOpcode()) {
5949 case Instruction::GetElementPtr:
5950 if (RHSC->isNullValue()) {
5951 // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
5952 bool isAllZeros = true;
5953 for (unsigned i = 1, e = LHSI->getNumOperands(); i != e; ++i)
5954 if (!isa<Constant>(LHSI->getOperand(i)) ||
5955 !cast<Constant>(LHSI->getOperand(i))->isNullValue()) {
5960 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
5961 Constant::getNullValue(LHSI->getOperand(0)->getType()));
5965 case Instruction::PHI:
5966 // Only fold icmp into the PHI if the phi and fcmp are in the same
5967 // block. If in the same block, we're encouraging jump threading. If
5968 // not, we are just pessimizing the code by making an i1 phi.
5969 if (LHSI->getParent() == I.getParent())
5970 if (Instruction *NV = FoldOpIntoPhi(I))
5973 case Instruction::Select: {
5974 // If either operand of the select is a constant, we can fold the
5975 // comparison into the select arms, which will cause one to be
5976 // constant folded and the select turned into a bitwise or.
5977 Value *Op1 = 0, *Op2 = 0;
5978 if (LHSI->hasOneUse()) {
5979 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
5980 // Fold the known value into the constant operand.
5981 Op1 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
5982 // Insert a new ICmp of the other select operand.
5983 Op2 = InsertNewInstBefore(new ICmpInst(I.getPredicate(),
5984 LHSI->getOperand(2), RHSC,
5986 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
5987 // Fold the known value into the constant operand.
5988 Op2 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
5989 // Insert a new ICmp of the other select operand.
5990 Op1 = InsertNewInstBefore(new ICmpInst(I.getPredicate(),
5991 LHSI->getOperand(1), RHSC,
5997 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
6000 case Instruction::Malloc:
6001 // If we have (malloc != null), and if the malloc has a single use, we
6002 // can assume it is successful and remove the malloc.
6003 if (LHSI->hasOneUse() && isa<ConstantPointerNull>(RHSC)) {
6004 AddToWorkList(LHSI);
6005 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
6006 !I.isTrueWhenEqual()));
6012 // If we can optimize a 'icmp GEP, P' or 'icmp P, GEP', do so now.
6013 if (User *GEP = dyn_castGetElementPtr(Op0))
6014 if (Instruction *NI = FoldGEPICmp(GEP, Op1, I.getPredicate(), I))
6016 if (User *GEP = dyn_castGetElementPtr(Op1))
6017 if (Instruction *NI = FoldGEPICmp(GEP, Op0,
6018 ICmpInst::getSwappedPredicate(I.getPredicate()), I))
6021 // Test to see if the operands of the icmp are casted versions of other
6022 // values. If the ptr->ptr cast can be stripped off both arguments, we do so
6024 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op0)) {
6025 if (isa<PointerType>(Op0->getType()) &&
6026 (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
6027 // We keep moving the cast from the left operand over to the right
6028 // operand, where it can often be eliminated completely.
6029 Op0 = CI->getOperand(0);
6031 // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast
6032 // so eliminate it as well.
6033 if (BitCastInst *CI2 = dyn_cast<BitCastInst>(Op1))
6034 Op1 = CI2->getOperand(0);
6036 // If Op1 is a constant, we can fold the cast into the constant.
6037 if (Op0->getType() != Op1->getType()) {
6038 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
6039 Op1 = ConstantExpr::getBitCast(Op1C, Op0->getType());
6041 // Otherwise, cast the RHS right before the icmp
6042 Op1 = InsertBitCastBefore(Op1, Op0->getType(), I);
6045 return new ICmpInst(I.getPredicate(), Op0, Op1);
6049 if (isa<CastInst>(Op0)) {
6050 // Handle the special case of: icmp (cast bool to X), <cst>
6051 // This comes up when you have code like
6054 // For generality, we handle any zero-extension of any operand comparison
6055 // with a constant or another cast from the same type.
6056 if (isa<ConstantInt>(Op1) || isa<CastInst>(Op1))
6057 if (Instruction *R = visitICmpInstWithCastAndCast(I))
6061 // See if it's the same type of instruction on the left and right.
6062 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
6063 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
6064 if (Op0I->getOpcode() == Op1I->getOpcode() && Op0I->hasOneUse() &&
6065 Op1I->hasOneUse() && Op0I->getOperand(1) == Op1I->getOperand(1) &&
6067 switch (Op0I->getOpcode()) {
6069 case Instruction::Add:
6070 case Instruction::Sub:
6071 case Instruction::Xor:
6072 // a+x icmp eq/ne b+x --> a icmp b
6073 return new ICmpInst(I.getPredicate(), Op0I->getOperand(0),
6074 Op1I->getOperand(0));
6076 case Instruction::Mul:
6077 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) {
6078 // a * Cst icmp eq/ne b * Cst --> a & Mask icmp b & Mask
6079 // Mask = -1 >> count-trailing-zeros(Cst).
6080 if (!CI->isZero() && !CI->isOne()) {
6081 const APInt &AP = CI->getValue();
6082 ConstantInt *Mask = ConstantInt::get(
6083 APInt::getLowBitsSet(AP.getBitWidth(),
6085 AP.countTrailingZeros()));
6086 Instruction *And1 = BinaryOperator::CreateAnd(Op0I->getOperand(0),
6088 Instruction *And2 = BinaryOperator::CreateAnd(Op1I->getOperand(0),
6090 InsertNewInstBefore(And1, I);
6091 InsertNewInstBefore(And2, I);
6092 return new ICmpInst(I.getPredicate(), And1, And2);
6101 // ~x < ~y --> y < x
6103 if (match(Op0, m_Not(m_Value(A))) &&
6104 match(Op1, m_Not(m_Value(B))))
6105 return new ICmpInst(I.getPredicate(), B, A);
6108 if (I.isEquality()) {
6109 Value *A, *B, *C, *D;
6111 // -x == -y --> x == y
6112 if (match(Op0, m_Neg(m_Value(A))) &&
6113 match(Op1, m_Neg(m_Value(B))))
6114 return new ICmpInst(I.getPredicate(), A, B);
6116 if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
6117 if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0
6118 Value *OtherVal = A == Op1 ? B : A;
6119 return new ICmpInst(I.getPredicate(), OtherVal,
6120 Constant::getNullValue(A->getType()));
6123 if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) {
6124 // A^c1 == C^c2 --> A == C^(c1^c2)
6125 ConstantInt *C1, *C2;
6126 if (match(B, m_ConstantInt(C1)) &&
6127 match(D, m_ConstantInt(C2)) && Op1->hasOneUse()) {
6128 Constant *NC = ConstantInt::get(C1->getValue() ^ C2->getValue());
6129 Instruction *Xor = BinaryOperator::CreateXor(C, NC, "tmp");
6130 return new ICmpInst(I.getPredicate(), A,
6131 InsertNewInstBefore(Xor, I));
6134 // A^B == A^D -> B == D
6135 if (A == C) return new ICmpInst(I.getPredicate(), B, D);
6136 if (A == D) return new ICmpInst(I.getPredicate(), B, C);
6137 if (B == C) return new ICmpInst(I.getPredicate(), A, D);
6138 if (B == D) return new ICmpInst(I.getPredicate(), A, C);
6142 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
6143 (A == Op0 || B == Op0)) {
6144 // A == (A^B) -> B == 0
6145 Value *OtherVal = A == Op0 ? B : A;
6146 return new ICmpInst(I.getPredicate(), OtherVal,
6147 Constant::getNullValue(A->getType()));
6150 // (A-B) == A -> B == 0
6151 if (match(Op0, m_Sub(m_Specific(Op1), m_Value(B))))
6152 return new ICmpInst(I.getPredicate(), B,
6153 Constant::getNullValue(B->getType()));
6155 // A == (A-B) -> B == 0
6156 if (match(Op1, m_Sub(m_Specific(Op0), m_Value(B))))
6157 return new ICmpInst(I.getPredicate(), B,
6158 Constant::getNullValue(B->getType()));
6160 // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
6161 if (Op0->hasOneUse() && Op1->hasOneUse() &&
6162 match(Op0, m_And(m_Value(A), m_Value(B))) &&
6163 match(Op1, m_And(m_Value(C), m_Value(D)))) {
6164 Value *X = 0, *Y = 0, *Z = 0;
6167 X = B; Y = D; Z = A;
6168 } else if (A == D) {
6169 X = B; Y = C; Z = A;
6170 } else if (B == C) {
6171 X = A; Y = D; Z = B;
6172 } else if (B == D) {
6173 X = A; Y = C; Z = B;
6176 if (X) { // Build (X^Y) & Z
6177 Op1 = InsertNewInstBefore(BinaryOperator::CreateXor(X, Y, "tmp"), I);
6178 Op1 = InsertNewInstBefore(BinaryOperator::CreateAnd(Op1, Z, "tmp"), I);
6179 I.setOperand(0, Op1);
6180 I.setOperand(1, Constant::getNullValue(Op1->getType()));
6185 return Changed ? &I : 0;
6189 /// FoldICmpDivCst - Fold "icmp pred, ([su]div X, DivRHS), CmpRHS" where DivRHS
6190 /// and CmpRHS are both known to be integer constants.
6191 Instruction *InstCombiner::FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
6192 ConstantInt *DivRHS) {
6193 ConstantInt *CmpRHS = cast<ConstantInt>(ICI.getOperand(1));
6194 const APInt &CmpRHSV = CmpRHS->getValue();
6196 // FIXME: If the operand types don't match the type of the divide
6197 // then don't attempt this transform. The code below doesn't have the
6198 // logic to deal with a signed divide and an unsigned compare (and
6199 // vice versa). This is because (x /s C1) <s C2 produces different
6200 // results than (x /s C1) <u C2 or (x /u C1) <s C2 or even
6201 // (x /u C1) <u C2. Simply casting the operands and result won't
6202 // work. :( The if statement below tests that condition and bails
6204 bool DivIsSigned = DivI->getOpcode() == Instruction::SDiv;
6205 if (!ICI.isEquality() && DivIsSigned != ICI.isSignedPredicate())
6207 if (DivRHS->isZero())
6208 return 0; // The ProdOV computation fails on divide by zero.
6209 if (DivIsSigned && DivRHS->isAllOnesValue())
6210 return 0; // The overflow computation also screws up here
6211 if (DivRHS->isOne())
6212 return 0; // Not worth bothering, and eliminates some funny cases
6215 // Compute Prod = CI * DivRHS. We are essentially solving an equation
6216 // of form X/C1=C2. We solve for X by multiplying C1 (DivRHS) and
6217 // C2 (CI). By solving for X we can turn this into a range check
6218 // instead of computing a divide.
6219 ConstantInt *Prod = Multiply(CmpRHS, DivRHS);
6221 // Determine if the product overflows by seeing if the product is
6222 // not equal to the divide. Make sure we do the same kind of divide
6223 // as in the LHS instruction that we're folding.
6224 bool ProdOV = (DivIsSigned ? ConstantExpr::getSDiv(Prod, DivRHS) :
6225 ConstantExpr::getUDiv(Prod, DivRHS)) != CmpRHS;
6227 // Get the ICmp opcode
6228 ICmpInst::Predicate Pred = ICI.getPredicate();
6230 // Figure out the interval that is being checked. For example, a comparison
6231 // like "X /u 5 == 0" is really checking that X is in the interval [0, 5).
6232 // Compute this interval based on the constants involved and the signedness of
6233 // the compare/divide. This computes a half-open interval, keeping track of
6234 // whether either value in the interval overflows. After analysis each
6235 // overflow variable is set to 0 if it's corresponding bound variable is valid
6236 // -1 if overflowed off the bottom end, or +1 if overflowed off the top end.
6237 int LoOverflow = 0, HiOverflow = 0;
6238 ConstantInt *LoBound = 0, *HiBound = 0;
6240 if (!DivIsSigned) { // udiv
6241 // e.g. X/5 op 3 --> [15, 20)
6243 HiOverflow = LoOverflow = ProdOV;
6245 HiOverflow = AddWithOverflow(HiBound, LoBound, DivRHS, false);
6246 } else if (DivRHS->getValue().isStrictlyPositive()) { // Divisor is > 0.
6247 if (CmpRHSV == 0) { // (X / pos) op 0
6248 // Can't overflow. e.g. X/2 op 0 --> [-1, 2)
6249 LoBound = cast<ConstantInt>(ConstantExpr::getNeg(SubOne(DivRHS)));
6251 } else if (CmpRHSV.isStrictlyPositive()) { // (X / pos) op pos
6252 LoBound = Prod; // e.g. X/5 op 3 --> [15, 20)
6253 HiOverflow = LoOverflow = ProdOV;
6255 HiOverflow = AddWithOverflow(HiBound, Prod, DivRHS, true);
6256 } else { // (X / pos) op neg
6257 // e.g. X/5 op -3 --> [-15-4, -15+1) --> [-19, -14)
6258 HiBound = AddOne(Prod);
6259 LoOverflow = HiOverflow = ProdOV ? -1 : 0;
6261 ConstantInt* DivNeg = cast<ConstantInt>(ConstantExpr::getNeg(DivRHS));
6262 LoOverflow = AddWithOverflow(LoBound, HiBound, DivNeg,
6266 } else if (DivRHS->getValue().isNegative()) { // Divisor is < 0.
6267 if (CmpRHSV == 0) { // (X / neg) op 0
6268 // e.g. X/-5 op 0 --> [-4, 5)
6269 LoBound = AddOne(DivRHS);
6270 HiBound = cast<ConstantInt>(ConstantExpr::getNeg(DivRHS));
6271 if (HiBound == DivRHS) { // -INTMIN = INTMIN
6272 HiOverflow = 1; // [INTMIN+1, overflow)
6273 HiBound = 0; // e.g. X/INTMIN = 0 --> X > INTMIN
6275 } else if (CmpRHSV.isStrictlyPositive()) { // (X / neg) op pos
6276 // e.g. X/-5 op 3 --> [-19, -14)
6277 HiBound = AddOne(Prod);
6278 HiOverflow = LoOverflow = ProdOV ? -1 : 0;
6280 LoOverflow = AddWithOverflow(LoBound, HiBound, DivRHS, true) ? -1 : 0;
6281 } else { // (X / neg) op neg
6282 LoBound = Prod; // e.g. X/-5 op -3 --> [15, 20)
6283 LoOverflow = HiOverflow = ProdOV;
6285 HiOverflow = SubWithOverflow(HiBound, Prod, DivRHS, true);
6288 // Dividing by a negative swaps the condition. LT <-> GT
6289 Pred = ICmpInst::getSwappedPredicate(Pred);
6292 Value *X = DivI->getOperand(0);
6294 default: assert(0 && "Unhandled icmp opcode!");
6295 case ICmpInst::ICMP_EQ:
6296 if (LoOverflow && HiOverflow)
6297 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
6298 else if (HiOverflow)
6299 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
6300 ICmpInst::ICMP_UGE, X, LoBound);
6301 else if (LoOverflow)
6302 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
6303 ICmpInst::ICMP_ULT, X, HiBound);
6305 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned, true, ICI);
6306 case ICmpInst::ICMP_NE:
6307 if (LoOverflow && HiOverflow)
6308 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
6309 else if (HiOverflow)
6310 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
6311 ICmpInst::ICMP_ULT, X, LoBound);
6312 else if (LoOverflow)
6313 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
6314 ICmpInst::ICMP_UGE, X, HiBound);
6316 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned, false, ICI);
6317 case ICmpInst::ICMP_ULT:
6318 case ICmpInst::ICMP_SLT:
6319 if (LoOverflow == +1) // Low bound is greater than input range.
6320 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
6321 if (LoOverflow == -1) // Low bound is less than input range.
6322 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
6323 return new ICmpInst(Pred, X, LoBound);
6324 case ICmpInst::ICMP_UGT:
6325 case ICmpInst::ICMP_SGT:
6326 if (HiOverflow == +1) // High bound greater than input range.
6327 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
6328 else if (HiOverflow == -1) // High bound less than input range.
6329 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
6330 if (Pred == ICmpInst::ICMP_UGT)
6331 return new ICmpInst(ICmpInst::ICMP_UGE, X, HiBound);
6333 return new ICmpInst(ICmpInst::ICMP_SGE, X, HiBound);
6338 /// visitICmpInstWithInstAndIntCst - Handle "icmp (instr, intcst)".
6340 Instruction *InstCombiner::visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
6343 const APInt &RHSV = RHS->getValue();
6345 switch (LHSI->getOpcode()) {
6346 case Instruction::Xor: // (icmp pred (xor X, XorCST), CI)
6347 if (ConstantInt *XorCST = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
6348 // If this is a comparison that tests the signbit (X < 0) or (x > -1),
6350 if ((ICI.getPredicate() == ICmpInst::ICMP_SLT && RHSV == 0) ||
6351 (ICI.getPredicate() == ICmpInst::ICMP_SGT && RHSV.isAllOnesValue())) {
6352 Value *CompareVal = LHSI->getOperand(0);
6354 // If the sign bit of the XorCST is not set, there is no change to
6355 // the operation, just stop using the Xor.
6356 if (!XorCST->getValue().isNegative()) {
6357 ICI.setOperand(0, CompareVal);
6358 AddToWorkList(LHSI);
6362 // Was the old condition true if the operand is positive?
6363 bool isTrueIfPositive = ICI.getPredicate() == ICmpInst::ICMP_SGT;
6365 // If so, the new one isn't.
6366 isTrueIfPositive ^= true;
6368 if (isTrueIfPositive)
6369 return new ICmpInst(ICmpInst::ICMP_SGT, CompareVal, SubOne(RHS));
6371 return new ICmpInst(ICmpInst::ICMP_SLT, CompareVal, AddOne(RHS));
6375 case Instruction::And: // (icmp pred (and X, AndCST), RHS)
6376 if (LHSI->hasOneUse() && isa<ConstantInt>(LHSI->getOperand(1)) &&
6377 LHSI->getOperand(0)->hasOneUse()) {
6378 ConstantInt *AndCST = cast<ConstantInt>(LHSI->getOperand(1));
6380 // If the LHS is an AND of a truncating cast, we can widen the
6381 // and/compare to be the input width without changing the value
6382 // produced, eliminating a cast.
6383 if (TruncInst *Cast = dyn_cast<TruncInst>(LHSI->getOperand(0))) {
6384 // We can do this transformation if either the AND constant does not
6385 // have its sign bit set or if it is an equality comparison.
6386 // Extending a relational comparison when we're checking the sign
6387 // bit would not work.
6388 if (Cast->hasOneUse() &&
6389 (ICI.isEquality() ||
6390 (AndCST->getValue().isNonNegative() && RHSV.isNonNegative()))) {
6392 cast<IntegerType>(Cast->getOperand(0)->getType())->getBitWidth();
6393 APInt NewCST = AndCST->getValue();
6394 NewCST.zext(BitWidth);
6396 NewCI.zext(BitWidth);
6397 Instruction *NewAnd =
6398 BinaryOperator::CreateAnd(Cast->getOperand(0),
6399 ConstantInt::get(NewCST),LHSI->getName());
6400 InsertNewInstBefore(NewAnd, ICI);
6401 return new ICmpInst(ICI.getPredicate(), NewAnd,
6402 ConstantInt::get(NewCI));
6406 // If this is: (X >> C1) & C2 != C3 (where any shift and any compare
6407 // could exist), turn it into (X & (C2 << C1)) != (C3 << C1). This
6408 // happens a LOT in code produced by the C front-end, for bitfield
6410 BinaryOperator *Shift = dyn_cast<BinaryOperator>(LHSI->getOperand(0));
6411 if (Shift && !Shift->isShift())
6415 ShAmt = Shift ? dyn_cast<ConstantInt>(Shift->getOperand(1)) : 0;
6416 const Type *Ty = Shift ? Shift->getType() : 0; // Type of the shift.
6417 const Type *AndTy = AndCST->getType(); // Type of the and.
6419 // We can fold this as long as we can't shift unknown bits
6420 // into the mask. This can only happen with signed shift
6421 // rights, as they sign-extend.
6423 bool CanFold = Shift->isLogicalShift();
6425 // To test for the bad case of the signed shr, see if any
6426 // of the bits shifted in could be tested after the mask.
6427 uint32_t TyBits = Ty->getPrimitiveSizeInBits();
6428 int ShAmtVal = TyBits - ShAmt->getLimitedValue(TyBits);
6430 uint32_t BitWidth = AndTy->getPrimitiveSizeInBits();
6431 if ((APInt::getHighBitsSet(BitWidth, BitWidth-ShAmtVal) &
6432 AndCST->getValue()) == 0)
6438 if (Shift->getOpcode() == Instruction::Shl)
6439 NewCst = ConstantExpr::getLShr(RHS, ShAmt);
6441 NewCst = ConstantExpr::getShl(RHS, ShAmt);
6443 // Check to see if we are shifting out any of the bits being
6445 if (ConstantExpr::get(Shift->getOpcode(), NewCst, ShAmt) != RHS) {
6446 // If we shifted bits out, the fold is not going to work out.
6447 // As a special case, check to see if this means that the
6448 // result is always true or false now.
6449 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
6450 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
6451 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
6452 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
6454 ICI.setOperand(1, NewCst);
6455 Constant *NewAndCST;
6456 if (Shift->getOpcode() == Instruction::Shl)
6457 NewAndCST = ConstantExpr::getLShr(AndCST, ShAmt);
6459 NewAndCST = ConstantExpr::getShl(AndCST, ShAmt);
6460 LHSI->setOperand(1, NewAndCST);
6461 LHSI->setOperand(0, Shift->getOperand(0));
6462 AddToWorkList(Shift); // Shift is dead.
6463 AddUsesToWorkList(ICI);
6469 // Turn ((X >> Y) & C) == 0 into (X & (C << Y)) == 0. The later is
6470 // preferable because it allows the C<<Y expression to be hoisted out
6471 // of a loop if Y is invariant and X is not.
6472 if (Shift && Shift->hasOneUse() && RHSV == 0 &&
6473 ICI.isEquality() && !Shift->isArithmeticShift() &&
6474 isa<Instruction>(Shift->getOperand(0))) {
6477 if (Shift->getOpcode() == Instruction::LShr) {
6478 NS = BinaryOperator::CreateShl(AndCST,
6479 Shift->getOperand(1), "tmp");
6481 // Insert a logical shift.
6482 NS = BinaryOperator::CreateLShr(AndCST,
6483 Shift->getOperand(1), "tmp");
6485 InsertNewInstBefore(cast<Instruction>(NS), ICI);
6487 // Compute X & (C << Y).
6488 Instruction *NewAnd =
6489 BinaryOperator::CreateAnd(Shift->getOperand(0), NS, LHSI->getName());
6490 InsertNewInstBefore(NewAnd, ICI);
6492 ICI.setOperand(0, NewAnd);
6498 case Instruction::Shl: { // (icmp pred (shl X, ShAmt), CI)
6499 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
6502 uint32_t TypeBits = RHSV.getBitWidth();
6504 // Check that the shift amount is in range. If not, don't perform
6505 // undefined shifts. When the shift is visited it will be
6507 if (ShAmt->uge(TypeBits))
6510 if (ICI.isEquality()) {
6511 // If we are comparing against bits always shifted out, the
6512 // comparison cannot succeed.
6514 ConstantExpr::getShl(ConstantExpr::getLShr(RHS, ShAmt), ShAmt);
6515 if (Comp != RHS) {// Comparing against a bit that we know is zero.
6516 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
6517 Constant *Cst = ConstantInt::get(Type::Int1Ty, IsICMP_NE);
6518 return ReplaceInstUsesWith(ICI, Cst);
6521 if (LHSI->hasOneUse()) {
6522 // Otherwise strength reduce the shift into an and.
6523 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
6525 ConstantInt::get(APInt::getLowBitsSet(TypeBits, TypeBits-ShAmtVal));
6528 BinaryOperator::CreateAnd(LHSI->getOperand(0),
6529 Mask, LHSI->getName()+".mask");
6530 Value *And = InsertNewInstBefore(AndI, ICI);
6531 return new ICmpInst(ICI.getPredicate(), And,
6532 ConstantInt::get(RHSV.lshr(ShAmtVal)));
6536 // Otherwise, if this is a comparison of the sign bit, simplify to and/test.
6537 bool TrueIfSigned = false;
6538 if (LHSI->hasOneUse() &&
6539 isSignBitCheck(ICI.getPredicate(), RHS, TrueIfSigned)) {
6540 // (X << 31) <s 0 --> (X&1) != 0
6541 Constant *Mask = ConstantInt::get(APInt(TypeBits, 1) <<
6542 (TypeBits-ShAmt->getZExtValue()-1));
6544 BinaryOperator::CreateAnd(LHSI->getOperand(0),
6545 Mask, LHSI->getName()+".mask");
6546 Value *And = InsertNewInstBefore(AndI, ICI);
6548 return new ICmpInst(TrueIfSigned ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ,
6549 And, Constant::getNullValue(And->getType()));
6554 case Instruction::LShr: // (icmp pred (shr X, ShAmt), CI)
6555 case Instruction::AShr: {
6556 // Only handle equality comparisons of shift-by-constant.
6557 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
6558 if (!ShAmt || !ICI.isEquality()) break;
6560 // Check that the shift amount is in range. If not, don't perform
6561 // undefined shifts. When the shift is visited it will be
6563 uint32_t TypeBits = RHSV.getBitWidth();
6564 if (ShAmt->uge(TypeBits))
6567 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
6569 // If we are comparing against bits always shifted out, the
6570 // comparison cannot succeed.
6571 APInt Comp = RHSV << ShAmtVal;
6572 if (LHSI->getOpcode() == Instruction::LShr)
6573 Comp = Comp.lshr(ShAmtVal);
6575 Comp = Comp.ashr(ShAmtVal);
6577 if (Comp != RHSV) { // Comparing against a bit that we know is zero.
6578 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
6579 Constant *Cst = ConstantInt::get(Type::Int1Ty, IsICMP_NE);
6580 return ReplaceInstUsesWith(ICI, Cst);
6583 // Otherwise, check to see if the bits shifted out are known to be zero.
6584 // If so, we can compare against the unshifted value:
6585 // (X & 4) >> 1 == 2 --> (X & 4) == 4.
6586 if (LHSI->hasOneUse() &&
6587 MaskedValueIsZero(LHSI->getOperand(0),
6588 APInt::getLowBitsSet(Comp.getBitWidth(), ShAmtVal))) {
6589 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
6590 ConstantExpr::getShl(RHS, ShAmt));
6593 if (LHSI->hasOneUse()) {
6594 // Otherwise strength reduce the shift into an and.
6595 APInt Val(APInt::getHighBitsSet(TypeBits, TypeBits - ShAmtVal));
6596 Constant *Mask = ConstantInt::get(Val);
6599 BinaryOperator::CreateAnd(LHSI->getOperand(0),
6600 Mask, LHSI->getName()+".mask");
6601 Value *And = InsertNewInstBefore(AndI, ICI);
6602 return new ICmpInst(ICI.getPredicate(), And,
6603 ConstantExpr::getShl(RHS, ShAmt));
6608 case Instruction::SDiv:
6609 case Instruction::UDiv:
6610 // Fold: icmp pred ([us]div X, C1), C2 -> range test
6611 // Fold this div into the comparison, producing a range check.
6612 // Determine, based on the divide type, what the range is being
6613 // checked. If there is an overflow on the low or high side, remember
6614 // it, otherwise compute the range [low, hi) bounding the new value.
6615 // See: InsertRangeTest above for the kinds of replacements possible.
6616 if (ConstantInt *DivRHS = dyn_cast<ConstantInt>(LHSI->getOperand(1)))
6617 if (Instruction *R = FoldICmpDivCst(ICI, cast<BinaryOperator>(LHSI),
6622 case Instruction::Add:
6623 // Fold: icmp pred (add, X, C1), C2
6625 if (!ICI.isEquality()) {
6626 ConstantInt *LHSC = dyn_cast<ConstantInt>(LHSI->getOperand(1));
6628 const APInt &LHSV = LHSC->getValue();
6630 ConstantRange CR = ICI.makeConstantRange(ICI.getPredicate(), RHSV)
6633 if (ICI.isSignedPredicate()) {
6634 if (CR.getLower().isSignBit()) {
6635 return new ICmpInst(ICmpInst::ICMP_SLT, LHSI->getOperand(0),
6636 ConstantInt::get(CR.getUpper()));
6637 } else if (CR.getUpper().isSignBit()) {
6638 return new ICmpInst(ICmpInst::ICMP_SGE, LHSI->getOperand(0),
6639 ConstantInt::get(CR.getLower()));
6642 if (CR.getLower().isMinValue()) {
6643 return new ICmpInst(ICmpInst::ICMP_ULT, LHSI->getOperand(0),
6644 ConstantInt::get(CR.getUpper()));
6645 } else if (CR.getUpper().isMinValue()) {
6646 return new ICmpInst(ICmpInst::ICMP_UGE, LHSI->getOperand(0),
6647 ConstantInt::get(CR.getLower()));
6654 // Simplify icmp_eq and icmp_ne instructions with integer constant RHS.
6655 if (ICI.isEquality()) {
6656 bool isICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
6658 // If the first operand is (add|sub|and|or|xor|rem) with a constant, and
6659 // the second operand is a constant, simplify a bit.
6660 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(LHSI)) {
6661 switch (BO->getOpcode()) {
6662 case Instruction::SRem:
6663 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
6664 if (RHSV == 0 && isa<ConstantInt>(BO->getOperand(1)) &&BO->hasOneUse()){
6665 const APInt &V = cast<ConstantInt>(BO->getOperand(1))->getValue();
6666 if (V.sgt(APInt(V.getBitWidth(), 1)) && V.isPowerOf2()) {
6667 Instruction *NewRem =
6668 BinaryOperator::CreateURem(BO->getOperand(0), BO->getOperand(1),
6670 InsertNewInstBefore(NewRem, ICI);
6671 return new ICmpInst(ICI.getPredicate(), NewRem,
6672 Constant::getNullValue(BO->getType()));
6676 case Instruction::Add:
6677 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
6678 if (ConstantInt *BOp1C = dyn_cast<ConstantInt>(BO->getOperand(1))) {
6679 if (BO->hasOneUse())
6680 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
6681 Subtract(RHS, BOp1C));
6682 } else if (RHSV == 0) {
6683 // Replace ((add A, B) != 0) with (A != -B) if A or B is
6684 // efficiently invertible, or if the add has just this one use.
6685 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
6687 if (Value *NegVal = dyn_castNegVal(BOp1))
6688 return new ICmpInst(ICI.getPredicate(), BOp0, NegVal);
6689 else if (Value *NegVal = dyn_castNegVal(BOp0))
6690 return new ICmpInst(ICI.getPredicate(), NegVal, BOp1);
6691 else if (BO->hasOneUse()) {
6692 Instruction *Neg = BinaryOperator::CreateNeg(BOp1);
6693 InsertNewInstBefore(Neg, ICI);
6695 return new ICmpInst(ICI.getPredicate(), BOp0, Neg);
6699 case Instruction::Xor:
6700 // For the xor case, we can xor two constants together, eliminating
6701 // the explicit xor.
6702 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1)))
6703 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
6704 ConstantExpr::getXor(RHS, BOC));
6707 case Instruction::Sub:
6708 // Replace (([sub|xor] A, B) != 0) with (A != B)
6710 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
6714 case Instruction::Or:
6715 // If bits are being or'd in that are not present in the constant we
6716 // are comparing against, then the comparison could never succeed!
6717 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1))) {
6718 Constant *NotCI = ConstantExpr::getNot(RHS);
6719 if (!ConstantExpr::getAnd(BOC, NotCI)->isNullValue())
6720 return ReplaceInstUsesWith(ICI, ConstantInt::get(Type::Int1Ty,
6725 case Instruction::And:
6726 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
6727 // If bits are being compared against that are and'd out, then the
6728 // comparison can never succeed!
6729 if ((RHSV & ~BOC->getValue()) != 0)
6730 return ReplaceInstUsesWith(ICI, ConstantInt::get(Type::Int1Ty,
6733 // If we have ((X & C) == C), turn it into ((X & C) != 0).
6734 if (RHS == BOC && RHSV.isPowerOf2())
6735 return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ :
6736 ICmpInst::ICMP_NE, LHSI,
6737 Constant::getNullValue(RHS->getType()));
6739 // Replace (and X, (1 << size(X)-1) != 0) with x s< 0
6740 if (BOC->getValue().isSignBit()) {
6741 Value *X = BO->getOperand(0);
6742 Constant *Zero = Constant::getNullValue(X->getType());
6743 ICmpInst::Predicate pred = isICMP_NE ?
6744 ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
6745 return new ICmpInst(pred, X, Zero);
6748 // ((X & ~7) == 0) --> X < 8
6749 if (RHSV == 0 && isHighOnes(BOC)) {
6750 Value *X = BO->getOperand(0);
6751 Constant *NegX = ConstantExpr::getNeg(BOC);
6752 ICmpInst::Predicate pred = isICMP_NE ?
6753 ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
6754 return new ICmpInst(pred, X, NegX);
6759 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(LHSI)) {
6760 // Handle icmp {eq|ne} <intrinsic>, intcst.
6761 if (II->getIntrinsicID() == Intrinsic::bswap) {
6763 ICI.setOperand(0, II->getOperand(1));
6764 ICI.setOperand(1, ConstantInt::get(RHSV.byteSwap()));
6772 /// visitICmpInstWithCastAndCast - Handle icmp (cast x to y), (cast/cst).
6773 /// We only handle extending casts so far.
6775 Instruction *InstCombiner::visitICmpInstWithCastAndCast(ICmpInst &ICI) {
6776 const CastInst *LHSCI = cast<CastInst>(ICI.getOperand(0));
6777 Value *LHSCIOp = LHSCI->getOperand(0);
6778 const Type *SrcTy = LHSCIOp->getType();
6779 const Type *DestTy = LHSCI->getType();
6782 // Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the
6783 // integer type is the same size as the pointer type.
6784 if (LHSCI->getOpcode() == Instruction::PtrToInt &&
6785 getTargetData().getPointerSizeInBits() ==
6786 cast<IntegerType>(DestTy)->getBitWidth()) {
6788 if (Constant *RHSC = dyn_cast<Constant>(ICI.getOperand(1))) {
6789 RHSOp = ConstantExpr::getIntToPtr(RHSC, SrcTy);
6790 } else if (PtrToIntInst *RHSC = dyn_cast<PtrToIntInst>(ICI.getOperand(1))) {
6791 RHSOp = RHSC->getOperand(0);
6792 // If the pointer types don't match, insert a bitcast.
6793 if (LHSCIOp->getType() != RHSOp->getType())
6794 RHSOp = InsertBitCastBefore(RHSOp, LHSCIOp->getType(), ICI);
6798 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSOp);
6801 // The code below only handles extension cast instructions, so far.
6803 if (LHSCI->getOpcode() != Instruction::ZExt &&
6804 LHSCI->getOpcode() != Instruction::SExt)
6807 bool isSignedExt = LHSCI->getOpcode() == Instruction::SExt;
6808 bool isSignedCmp = ICI.isSignedPredicate();
6810 if (CastInst *CI = dyn_cast<CastInst>(ICI.getOperand(1))) {
6811 // Not an extension from the same type?
6812 RHSCIOp = CI->getOperand(0);
6813 if (RHSCIOp->getType() != LHSCIOp->getType())
6816 // If the signedness of the two casts doesn't agree (i.e. one is a sext
6817 // and the other is a zext), then we can't handle this.
6818 if (CI->getOpcode() != LHSCI->getOpcode())
6821 // Deal with equality cases early.
6822 if (ICI.isEquality())
6823 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
6825 // A signed comparison of sign extended values simplifies into a
6826 // signed comparison.
6827 if (isSignedCmp && isSignedExt)
6828 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
6830 // The other three cases all fold into an unsigned comparison.
6831 return new ICmpInst(ICI.getUnsignedPredicate(), LHSCIOp, RHSCIOp);
6834 // If we aren't dealing with a constant on the RHS, exit early
6835 ConstantInt *CI = dyn_cast<ConstantInt>(ICI.getOperand(1));
6839 // Compute the constant that would happen if we truncated to SrcTy then
6840 // reextended to DestTy.
6841 Constant *Res1 = ConstantExpr::getTrunc(CI, SrcTy);
6842 Constant *Res2 = ConstantExpr::getCast(LHSCI->getOpcode(), Res1, DestTy);
6844 // If the re-extended constant didn't change...
6846 // Make sure that sign of the Cmp and the sign of the Cast are the same.
6847 // For example, we might have:
6848 // %A = sext short %X to uint
6849 // %B = icmp ugt uint %A, 1330
6850 // It is incorrect to transform this into
6851 // %B = icmp ugt short %X, 1330
6852 // because %A may have negative value.
6854 // However, we allow this when the compare is EQ/NE, because they are
6856 if (isSignedExt == isSignedCmp || ICI.isEquality())
6857 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
6861 // The re-extended constant changed so the constant cannot be represented
6862 // in the shorter type. Consequently, we cannot emit a simple comparison.
6864 // First, handle some easy cases. We know the result cannot be equal at this
6865 // point so handle the ICI.isEquality() cases
6866 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
6867 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
6868 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
6869 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
6871 // Evaluate the comparison for LT (we invert for GT below). LE and GE cases
6872 // should have been folded away previously and not enter in here.
6875 // We're performing a signed comparison.
6876 if (cast<ConstantInt>(CI)->getValue().isNegative())
6877 Result = ConstantInt::getFalse(); // X < (small) --> false
6879 Result = ConstantInt::getTrue(); // X < (large) --> true
6881 // We're performing an unsigned comparison.
6883 // We're performing an unsigned comp with a sign extended value.
6884 // This is true if the input is >= 0. [aka >s -1]
6885 Constant *NegOne = ConstantInt::getAllOnesValue(SrcTy);
6886 Result = InsertNewInstBefore(new ICmpInst(ICmpInst::ICMP_SGT, LHSCIOp,
6887 NegOne, ICI.getName()), ICI);
6889 // Unsigned extend & unsigned compare -> always true.
6890 Result = ConstantInt::getTrue();
6894 // Finally, return the value computed.
6895 if (ICI.getPredicate() == ICmpInst::ICMP_ULT ||
6896 ICI.getPredicate() == ICmpInst::ICMP_SLT)
6897 return ReplaceInstUsesWith(ICI, Result);
6899 assert((ICI.getPredicate()==ICmpInst::ICMP_UGT ||
6900 ICI.getPredicate()==ICmpInst::ICMP_SGT) &&
6901 "ICmp should be folded!");
6902 if (Constant *CI = dyn_cast<Constant>(Result))
6903 return ReplaceInstUsesWith(ICI, ConstantExpr::getNot(CI));
6904 return BinaryOperator::CreateNot(Result);
6907 Instruction *InstCombiner::visitShl(BinaryOperator &I) {
6908 return commonShiftTransforms(I);
6911 Instruction *InstCombiner::visitLShr(BinaryOperator &I) {
6912 return commonShiftTransforms(I);
6915 Instruction *InstCombiner::visitAShr(BinaryOperator &I) {
6916 if (Instruction *R = commonShiftTransforms(I))
6919 Value *Op0 = I.getOperand(0);
6921 // ashr int -1, X = -1 (for any arithmetic shift rights of ~0)
6922 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
6923 if (CSI->isAllOnesValue())
6924 return ReplaceInstUsesWith(I, CSI);
6926 // See if we can turn a signed shr into an unsigned shr.
6927 if (!isa<VectorType>(I.getType()) &&
6928 MaskedValueIsZero(Op0,
6929 APInt::getSignBit(I.getType()->getPrimitiveSizeInBits())))
6930 return BinaryOperator::CreateLShr(Op0, I.getOperand(1));
6935 Instruction *InstCombiner::commonShiftTransforms(BinaryOperator &I) {
6936 assert(I.getOperand(1)->getType() == I.getOperand(0)->getType());
6937 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
6939 // shl X, 0 == X and shr X, 0 == X
6940 // shl 0, X == 0 and shr 0, X == 0
6941 if (Op1 == Constant::getNullValue(Op1->getType()) ||
6942 Op0 == Constant::getNullValue(Op0->getType()))
6943 return ReplaceInstUsesWith(I, Op0);
6945 if (isa<UndefValue>(Op0)) {
6946 if (I.getOpcode() == Instruction::AShr) // undef >>s X -> undef
6947 return ReplaceInstUsesWith(I, Op0);
6948 else // undef << X -> 0, undef >>u X -> 0
6949 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
6951 if (isa<UndefValue>(Op1)) {
6952 if (I.getOpcode() == Instruction::AShr) // X >>s undef -> X
6953 return ReplaceInstUsesWith(I, Op0);
6954 else // X << undef, X >>u undef -> 0
6955 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
6958 // Try to fold constant and into select arguments.
6959 if (isa<Constant>(Op0))
6960 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
6961 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
6964 if (ConstantInt *CUI = dyn_cast<ConstantInt>(Op1))
6965 if (Instruction *Res = FoldShiftByConstant(Op0, CUI, I))
6970 Instruction *InstCombiner::FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
6971 BinaryOperator &I) {
6972 bool isLeftShift = I.getOpcode() == Instruction::Shl;
6974 // See if we can simplify any instructions used by the instruction whose sole
6975 // purpose is to compute bits we don't care about.
6976 uint32_t TypeBits = Op0->getType()->getPrimitiveSizeInBits();
6977 APInt KnownZero(TypeBits, 0), KnownOne(TypeBits, 0);
6978 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(TypeBits),
6979 KnownZero, KnownOne))
6982 // shl uint X, 32 = 0 and shr ubyte Y, 9 = 0, ... just don't eliminate shr
6983 // of a signed value.
6985 if (Op1->uge(TypeBits)) {
6986 if (I.getOpcode() != Instruction::AShr)
6987 return ReplaceInstUsesWith(I, Constant::getNullValue(Op0->getType()));
6989 I.setOperand(1, ConstantInt::get(I.getType(), TypeBits-1));
6994 // ((X*C1) << C2) == (X * (C1 << C2))
6995 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0))
6996 if (BO->getOpcode() == Instruction::Mul && isLeftShift)
6997 if (Constant *BOOp = dyn_cast<Constant>(BO->getOperand(1)))
6998 return BinaryOperator::CreateMul(BO->getOperand(0),
6999 ConstantExpr::getShl(BOOp, Op1));
7001 // Try to fold constant and into select arguments.
7002 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
7003 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
7005 if (isa<PHINode>(Op0))
7006 if (Instruction *NV = FoldOpIntoPhi(I))
7009 // Fold shift2(trunc(shift1(x,c1)), c2) -> trunc(shift2(shift1(x,c1),c2))
7010 if (TruncInst *TI = dyn_cast<TruncInst>(Op0)) {
7011 Instruction *TrOp = dyn_cast<Instruction>(TI->getOperand(0));
7012 // If 'shift2' is an ashr, we would have to get the sign bit into a funny
7013 // place. Don't try to do this transformation in this case. Also, we
7014 // require that the input operand is a shift-by-constant so that we have
7015 // confidence that the shifts will get folded together. We could do this
7016 // xform in more cases, but it is unlikely to be profitable.
7017 if (TrOp && I.isLogicalShift() && TrOp->isShift() &&
7018 isa<ConstantInt>(TrOp->getOperand(1))) {
7019 // Okay, we'll do this xform. Make the shift of shift.
7020 Constant *ShAmt = ConstantExpr::getZExt(Op1, TrOp->getType());
7021 Instruction *NSh = BinaryOperator::Create(I.getOpcode(), TrOp, ShAmt,
7023 InsertNewInstBefore(NSh, I); // (shift2 (shift1 & 0x00FF), c2)
7025 // For logical shifts, the truncation has the effect of making the high
7026 // part of the register be zeros. Emulate this by inserting an AND to
7027 // clear the top bits as needed. This 'and' will usually be zapped by
7028 // other xforms later if dead.
7029 unsigned SrcSize = TrOp->getType()->getPrimitiveSizeInBits();
7030 unsigned DstSize = TI->getType()->getPrimitiveSizeInBits();
7031 APInt MaskV(APInt::getLowBitsSet(SrcSize, DstSize));
7033 // The mask we constructed says what the trunc would do if occurring
7034 // between the shifts. We want to know the effect *after* the second
7035 // shift. We know that it is a logical shift by a constant, so adjust the
7036 // mask as appropriate.
7037 if (I.getOpcode() == Instruction::Shl)
7038 MaskV <<= Op1->getZExtValue();
7040 assert(I.getOpcode() == Instruction::LShr && "Unknown logical shift");
7041 MaskV = MaskV.lshr(Op1->getZExtValue());
7044 Instruction *And = BinaryOperator::CreateAnd(NSh, ConstantInt::get(MaskV),
7046 InsertNewInstBefore(And, I); // shift1 & 0x00FF
7048 // Return the value truncated to the interesting size.
7049 return new TruncInst(And, I.getType());
7053 if (Op0->hasOneUse()) {
7054 if (BinaryOperator *Op0BO = dyn_cast<BinaryOperator>(Op0)) {
7055 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
7058 switch (Op0BO->getOpcode()) {
7060 case Instruction::Add:
7061 case Instruction::And:
7062 case Instruction::Or:
7063 case Instruction::Xor: {
7064 // These operators commute.
7065 // Turn (Y + (X >> C)) << C -> (X + (Y << C)) & (~0 << C)
7066 if (isLeftShift && Op0BO->getOperand(1)->hasOneUse() &&
7067 match(Op0BO->getOperand(1), m_Shr(m_Value(V1), m_Specific(Op1)))){
7068 Instruction *YS = BinaryOperator::CreateShl(
7069 Op0BO->getOperand(0), Op1,
7071 InsertNewInstBefore(YS, I); // (Y << C)
7073 BinaryOperator::Create(Op0BO->getOpcode(), YS, V1,
7074 Op0BO->getOperand(1)->getName());
7075 InsertNewInstBefore(X, I); // (X + (Y << C))
7076 uint32_t Op1Val = Op1->getLimitedValue(TypeBits);
7077 return BinaryOperator::CreateAnd(X, ConstantInt::get(
7078 APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
7081 // Turn (Y + ((X >> C) & CC)) << C -> ((X & (CC << C)) + (Y << C))
7082 Value *Op0BOOp1 = Op0BO->getOperand(1);
7083 if (isLeftShift && Op0BOOp1->hasOneUse() &&
7085 m_And(m_Shr(m_Value(V1), m_Specific(Op1)),
7086 m_ConstantInt(CC))) &&
7087 cast<BinaryOperator>(Op0BOOp1)->getOperand(0)->hasOneUse()) {
7088 Instruction *YS = BinaryOperator::CreateShl(
7089 Op0BO->getOperand(0), Op1,
7091 InsertNewInstBefore(YS, I); // (Y << C)
7093 BinaryOperator::CreateAnd(V1, ConstantExpr::getShl(CC, Op1),
7094 V1->getName()+".mask");
7095 InsertNewInstBefore(XM, I); // X & (CC << C)
7097 return BinaryOperator::Create(Op0BO->getOpcode(), YS, XM);
7102 case Instruction::Sub: {
7103 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
7104 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
7105 match(Op0BO->getOperand(0), m_Shr(m_Value(V1), m_Specific(Op1)))){
7106 Instruction *YS = BinaryOperator::CreateShl(
7107 Op0BO->getOperand(1), Op1,
7109 InsertNewInstBefore(YS, I); // (Y << C)
7111 BinaryOperator::Create(Op0BO->getOpcode(), V1, YS,
7112 Op0BO->getOperand(0)->getName());
7113 InsertNewInstBefore(X, I); // (X + (Y << C))
7114 uint32_t Op1Val = Op1->getLimitedValue(TypeBits);
7115 return BinaryOperator::CreateAnd(X, ConstantInt::get(
7116 APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
7119 // Turn (((X >> C)&CC) + Y) << C -> (X + (Y << C)) & (CC << C)
7120 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
7121 match(Op0BO->getOperand(0),
7122 m_And(m_Shr(m_Value(V1), m_Value(V2)),
7123 m_ConstantInt(CC))) && V2 == Op1 &&
7124 cast<BinaryOperator>(Op0BO->getOperand(0))
7125 ->getOperand(0)->hasOneUse()) {
7126 Instruction *YS = BinaryOperator::CreateShl(
7127 Op0BO->getOperand(1), Op1,
7129 InsertNewInstBefore(YS, I); // (Y << C)
7131 BinaryOperator::CreateAnd(V1, ConstantExpr::getShl(CC, Op1),
7132 V1->getName()+".mask");
7133 InsertNewInstBefore(XM, I); // X & (CC << C)
7135 return BinaryOperator::Create(Op0BO->getOpcode(), XM, YS);
7143 // If the operand is an bitwise operator with a constant RHS, and the
7144 // shift is the only use, we can pull it out of the shift.
7145 if (ConstantInt *Op0C = dyn_cast<ConstantInt>(Op0BO->getOperand(1))) {
7146 bool isValid = true; // Valid only for And, Or, Xor
7147 bool highBitSet = false; // Transform if high bit of constant set?
7149 switch (Op0BO->getOpcode()) {
7150 default: isValid = false; break; // Do not perform transform!
7151 case Instruction::Add:
7152 isValid = isLeftShift;
7154 case Instruction::Or:
7155 case Instruction::Xor:
7158 case Instruction::And:
7163 // If this is a signed shift right, and the high bit is modified
7164 // by the logical operation, do not perform the transformation.
7165 // The highBitSet boolean indicates the value of the high bit of
7166 // the constant which would cause it to be modified for this
7169 if (isValid && I.getOpcode() == Instruction::AShr)
7170 isValid = Op0C->getValue()[TypeBits-1] == highBitSet;
7173 Constant *NewRHS = ConstantExpr::get(I.getOpcode(), Op0C, Op1);
7175 Instruction *NewShift =
7176 BinaryOperator::Create(I.getOpcode(), Op0BO->getOperand(0), Op1);
7177 InsertNewInstBefore(NewShift, I);
7178 NewShift->takeName(Op0BO);
7180 return BinaryOperator::Create(Op0BO->getOpcode(), NewShift,
7187 // Find out if this is a shift of a shift by a constant.
7188 BinaryOperator *ShiftOp = dyn_cast<BinaryOperator>(Op0);
7189 if (ShiftOp && !ShiftOp->isShift())
7192 if (ShiftOp && isa<ConstantInt>(ShiftOp->getOperand(1))) {
7193 ConstantInt *ShiftAmt1C = cast<ConstantInt>(ShiftOp->getOperand(1));
7194 uint32_t ShiftAmt1 = ShiftAmt1C->getLimitedValue(TypeBits);
7195 uint32_t ShiftAmt2 = Op1->getLimitedValue(TypeBits);
7196 assert(ShiftAmt2 != 0 && "Should have been simplified earlier");
7197 if (ShiftAmt1 == 0) return 0; // Will be simplified in the future.
7198 Value *X = ShiftOp->getOperand(0);
7200 uint32_t AmtSum = ShiftAmt1+ShiftAmt2; // Fold into one big shift.
7201 if (AmtSum > TypeBits)
7204 const IntegerType *Ty = cast<IntegerType>(I.getType());
7206 // Check for (X << c1) << c2 and (X >> c1) >> c2
7207 if (I.getOpcode() == ShiftOp->getOpcode()) {
7208 return BinaryOperator::Create(I.getOpcode(), X,
7209 ConstantInt::get(Ty, AmtSum));
7210 } else if (ShiftOp->getOpcode() == Instruction::LShr &&
7211 I.getOpcode() == Instruction::AShr) {
7212 // ((X >>u C1) >>s C2) -> (X >>u (C1+C2)) since C1 != 0.
7213 return BinaryOperator::CreateLShr(X, ConstantInt::get(Ty, AmtSum));
7214 } else if (ShiftOp->getOpcode() == Instruction::AShr &&
7215 I.getOpcode() == Instruction::LShr) {
7216 // ((X >>s C1) >>u C2) -> ((X >>s (C1+C2)) & mask) since C1 != 0.
7217 Instruction *Shift =
7218 BinaryOperator::CreateAShr(X, ConstantInt::get(Ty, AmtSum));
7219 InsertNewInstBefore(Shift, I);
7221 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
7222 return BinaryOperator::CreateAnd(Shift, ConstantInt::get(Mask));
7225 // Okay, if we get here, one shift must be left, and the other shift must be
7226 // right. See if the amounts are equal.
7227 if (ShiftAmt1 == ShiftAmt2) {
7228 // If we have ((X >>? C) << C), turn this into X & (-1 << C).
7229 if (I.getOpcode() == Instruction::Shl) {
7230 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt1));
7231 return BinaryOperator::CreateAnd(X, ConstantInt::get(Mask));
7233 // If we have ((X << C) >>u C), turn this into X & (-1 >>u C).
7234 if (I.getOpcode() == Instruction::LShr) {
7235 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt1));
7236 return BinaryOperator::CreateAnd(X, ConstantInt::get(Mask));
7238 // We can simplify ((X << C) >>s C) into a trunc + sext.
7239 // NOTE: we could do this for any C, but that would make 'unusual' integer
7240 // types. For now, just stick to ones well-supported by the code
7242 const Type *SExtType = 0;
7243 switch (Ty->getBitWidth() - ShiftAmt1) {
7250 SExtType = IntegerType::get(Ty->getBitWidth() - ShiftAmt1);
7255 Instruction *NewTrunc = new TruncInst(X, SExtType, "sext");
7256 InsertNewInstBefore(NewTrunc, I);
7257 return new SExtInst(NewTrunc, Ty);
7259 // Otherwise, we can't handle it yet.
7260 } else if (ShiftAmt1 < ShiftAmt2) {
7261 uint32_t ShiftDiff = ShiftAmt2-ShiftAmt1;
7263 // (X >>? C1) << C2 --> X << (C2-C1) & (-1 << C2)
7264 if (I.getOpcode() == Instruction::Shl) {
7265 assert(ShiftOp->getOpcode() == Instruction::LShr ||
7266 ShiftOp->getOpcode() == Instruction::AShr);
7267 Instruction *Shift =
7268 BinaryOperator::CreateShl(X, ConstantInt::get(Ty, ShiftDiff));
7269 InsertNewInstBefore(Shift, I);
7271 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
7272 return BinaryOperator::CreateAnd(Shift, ConstantInt::get(Mask));
7275 // (X << C1) >>u C2 --> X >>u (C2-C1) & (-1 >> C2)
7276 if (I.getOpcode() == Instruction::LShr) {
7277 assert(ShiftOp->getOpcode() == Instruction::Shl);
7278 Instruction *Shift =
7279 BinaryOperator::CreateLShr(X, ConstantInt::get(Ty, ShiftDiff));
7280 InsertNewInstBefore(Shift, I);
7282 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
7283 return BinaryOperator::CreateAnd(Shift, ConstantInt::get(Mask));
7286 // We can't handle (X << C1) >>s C2, it shifts arbitrary bits in.
7288 assert(ShiftAmt2 < ShiftAmt1);
7289 uint32_t ShiftDiff = ShiftAmt1-ShiftAmt2;
7291 // (X >>? C1) << C2 --> X >>? (C1-C2) & (-1 << C2)
7292 if (I.getOpcode() == Instruction::Shl) {
7293 assert(ShiftOp->getOpcode() == Instruction::LShr ||
7294 ShiftOp->getOpcode() == Instruction::AShr);
7295 Instruction *Shift =
7296 BinaryOperator::Create(ShiftOp->getOpcode(), X,
7297 ConstantInt::get(Ty, ShiftDiff));
7298 InsertNewInstBefore(Shift, I);
7300 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
7301 return BinaryOperator::CreateAnd(Shift, ConstantInt::get(Mask));
7304 // (X << C1) >>u C2 --> X << (C1-C2) & (-1 >> C2)
7305 if (I.getOpcode() == Instruction::LShr) {
7306 assert(ShiftOp->getOpcode() == Instruction::Shl);
7307 Instruction *Shift =
7308 BinaryOperator::CreateShl(X, ConstantInt::get(Ty, ShiftDiff));
7309 InsertNewInstBefore(Shift, I);
7311 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
7312 return BinaryOperator::CreateAnd(Shift, ConstantInt::get(Mask));
7315 // We can't handle (X << C1) >>a C2, it shifts arbitrary bits in.
7322 /// DecomposeSimpleLinearExpr - Analyze 'Val', seeing if it is a simple linear
7323 /// expression. If so, decompose it, returning some value X, such that Val is
7326 static Value *DecomposeSimpleLinearExpr(Value *Val, unsigned &Scale,
7328 assert(Val->getType() == Type::Int32Ty && "Unexpected allocation size type!");
7329 if (ConstantInt *CI = dyn_cast<ConstantInt>(Val)) {
7330 Offset = CI->getZExtValue();
7332 return ConstantInt::get(Type::Int32Ty, 0);
7333 } else if (BinaryOperator *I = dyn_cast<BinaryOperator>(Val)) {
7334 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
7335 if (I->getOpcode() == Instruction::Shl) {
7336 // This is a value scaled by '1 << the shift amt'.
7337 Scale = 1U << RHS->getZExtValue();
7339 return I->getOperand(0);
7340 } else if (I->getOpcode() == Instruction::Mul) {
7341 // This value is scaled by 'RHS'.
7342 Scale = RHS->getZExtValue();
7344 return I->getOperand(0);
7345 } else if (I->getOpcode() == Instruction::Add) {
7346 // We have X+C. Check to see if we really have (X*C2)+C1,
7347 // where C1 is divisible by C2.
7350 DecomposeSimpleLinearExpr(I->getOperand(0), SubScale, Offset);
7351 Offset += RHS->getZExtValue();
7358 // Otherwise, we can't look past this.
7365 /// PromoteCastOfAllocation - If we find a cast of an allocation instruction,
7366 /// try to eliminate the cast by moving the type information into the alloc.
7367 Instruction *InstCombiner::PromoteCastOfAllocation(BitCastInst &CI,
7368 AllocationInst &AI) {
7369 const PointerType *PTy = cast<PointerType>(CI.getType());
7371 // Remove any uses of AI that are dead.
7372 assert(!CI.use_empty() && "Dead instructions should be removed earlier!");
7374 for (Value::use_iterator UI = AI.use_begin(), E = AI.use_end(); UI != E; ) {
7375 Instruction *User = cast<Instruction>(*UI++);
7376 if (isInstructionTriviallyDead(User)) {
7377 while (UI != E && *UI == User)
7378 ++UI; // If this instruction uses AI more than once, don't break UI.
7381 DOUT << "IC: DCE: " << *User;
7382 EraseInstFromFunction(*User);
7386 // Get the type really allocated and the type casted to.
7387 const Type *AllocElTy = AI.getAllocatedType();
7388 const Type *CastElTy = PTy->getElementType();
7389 if (!AllocElTy->isSized() || !CastElTy->isSized()) return 0;
7391 unsigned AllocElTyAlign = TD->getABITypeAlignment(AllocElTy);
7392 unsigned CastElTyAlign = TD->getABITypeAlignment(CastElTy);
7393 if (CastElTyAlign < AllocElTyAlign) return 0;
7395 // If the allocation has multiple uses, only promote it if we are strictly
7396 // increasing the alignment of the resultant allocation. If we keep it the
7397 // same, we open the door to infinite loops of various kinds.
7398 if (!AI.hasOneUse() && CastElTyAlign == AllocElTyAlign) return 0;
7400 uint64_t AllocElTySize = TD->getABITypeSize(AllocElTy);
7401 uint64_t CastElTySize = TD->getABITypeSize(CastElTy);
7402 if (CastElTySize == 0 || AllocElTySize == 0) return 0;
7404 // See if we can satisfy the modulus by pulling a scale out of the array
7406 unsigned ArraySizeScale;
7408 Value *NumElements = // See if the array size is a decomposable linear expr.
7409 DecomposeSimpleLinearExpr(AI.getOperand(0), ArraySizeScale, ArrayOffset);
7411 // If we can now satisfy the modulus, by using a non-1 scale, we really can
7413 if ((AllocElTySize*ArraySizeScale) % CastElTySize != 0 ||
7414 (AllocElTySize*ArrayOffset ) % CastElTySize != 0) return 0;
7416 unsigned Scale = (AllocElTySize*ArraySizeScale)/CastElTySize;
7421 // If the allocation size is constant, form a constant mul expression
7422 Amt = ConstantInt::get(Type::Int32Ty, Scale);
7423 if (isa<ConstantInt>(NumElements))
7424 Amt = Multiply(cast<ConstantInt>(NumElements), cast<ConstantInt>(Amt));
7425 // otherwise multiply the amount and the number of elements
7426 else if (Scale != 1) {
7427 Instruction *Tmp = BinaryOperator::CreateMul(Amt, NumElements, "tmp");
7428 Amt = InsertNewInstBefore(Tmp, AI);
7432 if (int Offset = (AllocElTySize*ArrayOffset)/CastElTySize) {
7433 Value *Off = ConstantInt::get(Type::Int32Ty, Offset, true);
7434 Instruction *Tmp = BinaryOperator::CreateAdd(Amt, Off, "tmp");
7435 Amt = InsertNewInstBefore(Tmp, AI);
7438 AllocationInst *New;
7439 if (isa<MallocInst>(AI))
7440 New = new MallocInst(CastElTy, Amt, AI.getAlignment());
7442 New = new AllocaInst(CastElTy, Amt, AI.getAlignment());
7443 InsertNewInstBefore(New, AI);
7446 // If the allocation has multiple uses, insert a cast and change all things
7447 // that used it to use the new cast. This will also hack on CI, but it will
7449 if (!AI.hasOneUse()) {
7450 AddUsesToWorkList(AI);
7451 // New is the allocation instruction, pointer typed. AI is the original
7452 // allocation instruction, also pointer typed. Thus, cast to use is BitCast.
7453 CastInst *NewCast = new BitCastInst(New, AI.getType(), "tmpcast");
7454 InsertNewInstBefore(NewCast, AI);
7455 AI.replaceAllUsesWith(NewCast);
7457 return ReplaceInstUsesWith(CI, New);
7460 /// CanEvaluateInDifferentType - Return true if we can take the specified value
7461 /// and return it as type Ty without inserting any new casts and without
7462 /// changing the computed value. This is used by code that tries to decide
7463 /// whether promoting or shrinking integer operations to wider or smaller types
7464 /// will allow us to eliminate a truncate or extend.
7466 /// This is a truncation operation if Ty is smaller than V->getType(), or an
7467 /// extension operation if Ty is larger.
7469 /// If CastOpc is a truncation, then Ty will be a type smaller than V. We
7470 /// should return true if trunc(V) can be computed by computing V in the smaller
7471 /// type. If V is an instruction, then trunc(inst(x,y)) can be computed as
7472 /// inst(trunc(x),trunc(y)), which only makes sense if x and y can be
7473 /// efficiently truncated.
7475 /// If CastOpc is a sext or zext, we are asking if the low bits of the value can
7476 /// bit computed in a larger type, which is then and'd or sext_in_reg'd to get
7477 /// the final result.
7478 bool InstCombiner::CanEvaluateInDifferentType(Value *V, const IntegerType *Ty,
7480 int &NumCastsRemoved) {
7481 // We can always evaluate constants in another type.
7482 if (isa<ConstantInt>(V))
7485 Instruction *I = dyn_cast<Instruction>(V);
7486 if (!I) return false;
7488 const IntegerType *OrigTy = cast<IntegerType>(V->getType());
7490 // If this is an extension or truncate, we can often eliminate it.
7491 if (isa<TruncInst>(I) || isa<ZExtInst>(I) || isa<SExtInst>(I)) {
7492 // If this is a cast from the destination type, we can trivially eliminate
7493 // it, and this will remove a cast overall.
7494 if (I->getOperand(0)->getType() == Ty) {
7495 // If the first operand is itself a cast, and is eliminable, do not count
7496 // this as an eliminable cast. We would prefer to eliminate those two
7498 if (!isa<CastInst>(I->getOperand(0)) && I->hasOneUse())
7504 // We can't extend or shrink something that has multiple uses: doing so would
7505 // require duplicating the instruction in general, which isn't profitable.
7506 if (!I->hasOneUse()) return false;
7508 switch (I->getOpcode()) {
7509 case Instruction::Add:
7510 case Instruction::Sub:
7511 case Instruction::Mul:
7512 case Instruction::And:
7513 case Instruction::Or:
7514 case Instruction::Xor:
7515 // These operators can all arbitrarily be extended or truncated.
7516 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
7518 CanEvaluateInDifferentType(I->getOperand(1), Ty, CastOpc,
7521 case Instruction::Shl:
7522 // If we are truncating the result of this SHL, and if it's a shift of a
7523 // constant amount, we can always perform a SHL in a smaller type.
7524 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
7525 uint32_t BitWidth = Ty->getBitWidth();
7526 if (BitWidth < OrigTy->getBitWidth() &&
7527 CI->getLimitedValue(BitWidth) < BitWidth)
7528 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
7532 case Instruction::LShr:
7533 // If this is a truncate of a logical shr, we can truncate it to a smaller
7534 // lshr iff we know that the bits we would otherwise be shifting in are
7536 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
7537 uint32_t OrigBitWidth = OrigTy->getBitWidth();
7538 uint32_t BitWidth = Ty->getBitWidth();
7539 if (BitWidth < OrigBitWidth &&
7540 MaskedValueIsZero(I->getOperand(0),
7541 APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth)) &&
7542 CI->getLimitedValue(BitWidth) < BitWidth) {
7543 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
7548 case Instruction::ZExt:
7549 case Instruction::SExt:
7550 case Instruction::Trunc:
7551 // If this is the same kind of case as our original (e.g. zext+zext), we
7552 // can safely replace it. Note that replacing it does not reduce the number
7553 // of casts in the input.
7554 if (I->getOpcode() == CastOpc)
7557 case Instruction::Select: {
7558 SelectInst *SI = cast<SelectInst>(I);
7559 return CanEvaluateInDifferentType(SI->getTrueValue(), Ty, CastOpc,
7561 CanEvaluateInDifferentType(SI->getFalseValue(), Ty, CastOpc,
7564 case Instruction::PHI: {
7565 // We can change a phi if we can change all operands.
7566 PHINode *PN = cast<PHINode>(I);
7567 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
7568 if (!CanEvaluateInDifferentType(PN->getIncomingValue(i), Ty, CastOpc,
7574 // TODO: Can handle more cases here.
7581 /// EvaluateInDifferentType - Given an expression that
7582 /// CanEvaluateInDifferentType returns true for, actually insert the code to
7583 /// evaluate the expression.
7584 Value *InstCombiner::EvaluateInDifferentType(Value *V, const Type *Ty,
7586 if (Constant *C = dyn_cast<Constant>(V))
7587 return ConstantExpr::getIntegerCast(C, Ty, isSigned /*Sext or ZExt*/);
7589 // Otherwise, it must be an instruction.
7590 Instruction *I = cast<Instruction>(V);
7591 Instruction *Res = 0;
7592 switch (I->getOpcode()) {
7593 case Instruction::Add:
7594 case Instruction::Sub:
7595 case Instruction::Mul:
7596 case Instruction::And:
7597 case Instruction::Or:
7598 case Instruction::Xor:
7599 case Instruction::AShr:
7600 case Instruction::LShr:
7601 case Instruction::Shl: {
7602 Value *LHS = EvaluateInDifferentType(I->getOperand(0), Ty, isSigned);
7603 Value *RHS = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
7604 Res = BinaryOperator::Create((Instruction::BinaryOps)I->getOpcode(),
7608 case Instruction::Trunc:
7609 case Instruction::ZExt:
7610 case Instruction::SExt:
7611 // If the source type of the cast is the type we're trying for then we can
7612 // just return the source. There's no need to insert it because it is not
7614 if (I->getOperand(0)->getType() == Ty)
7615 return I->getOperand(0);
7617 // Otherwise, must be the same type of cast, so just reinsert a new one.
7618 Res = CastInst::Create(cast<CastInst>(I)->getOpcode(), I->getOperand(0),
7621 case Instruction::Select: {
7622 Value *True = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
7623 Value *False = EvaluateInDifferentType(I->getOperand(2), Ty, isSigned);
7624 Res = SelectInst::Create(I->getOperand(0), True, False);
7627 case Instruction::PHI: {
7628 PHINode *OPN = cast<PHINode>(I);
7629 PHINode *NPN = PHINode::Create(Ty);
7630 for (unsigned i = 0, e = OPN->getNumIncomingValues(); i != e; ++i) {
7631 Value *V =EvaluateInDifferentType(OPN->getIncomingValue(i), Ty, isSigned);
7632 NPN->addIncoming(V, OPN->getIncomingBlock(i));
7638 // TODO: Can handle more cases here.
7639 assert(0 && "Unreachable!");
7644 return InsertNewInstBefore(Res, *I);
7647 /// @brief Implement the transforms common to all CastInst visitors.
7648 Instruction *InstCombiner::commonCastTransforms(CastInst &CI) {
7649 Value *Src = CI.getOperand(0);
7651 // Many cases of "cast of a cast" are eliminable. If it's eliminable we just
7652 // eliminate it now.
7653 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
7654 if (Instruction::CastOps opc =
7655 isEliminableCastPair(CSrc, CI.getOpcode(), CI.getType(), TD)) {
7656 // The first cast (CSrc) is eliminable so we need to fix up or replace
7657 // the second cast (CI). CSrc will then have a good chance of being dead.
7658 return CastInst::Create(opc, CSrc->getOperand(0), CI.getType());
7662 // If we are casting a select then fold the cast into the select
7663 if (SelectInst *SI = dyn_cast<SelectInst>(Src))
7664 if (Instruction *NV = FoldOpIntoSelect(CI, SI, this))
7667 // If we are casting a PHI then fold the cast into the PHI
7668 if (isa<PHINode>(Src))
7669 if (Instruction *NV = FoldOpIntoPhi(CI))
7675 /// FindElementAtOffset - Given a type and a constant offset, determine whether
7676 /// or not there is a sequence of GEP indices into the type that will land us at
7677 /// the specified offset. If so, fill them into NewIndices and return true,
7678 /// otherwise return false.
7679 static bool FindElementAtOffset(const Type *Ty, int64_t Offset,
7680 SmallVectorImpl<Value*> &NewIndices,
7681 const TargetData *TD) {
7682 if (!Ty->isSized()) return false;
7684 // Start with the index over the outer type. Note that the type size
7685 // might be zero (even if the offset isn't zero) if the indexed type
7686 // is something like [0 x {int, int}]
7687 const Type *IntPtrTy = TD->getIntPtrType();
7688 int64_t FirstIdx = 0;
7689 if (int64_t TySize = TD->getABITypeSize(Ty)) {
7690 FirstIdx = Offset/TySize;
7693 // Handle silly modulus not returning values values [0..TySize).
7697 assert(Offset >= 0);
7699 assert((uint64_t)Offset < (uint64_t)TySize && "Out of range offset");
7702 NewIndices.push_back(ConstantInt::get(IntPtrTy, FirstIdx));
7704 // Index into the types. If we fail, set OrigBase to null.
7706 if (const StructType *STy = dyn_cast<StructType>(Ty)) {
7707 const StructLayout *SL = TD->getStructLayout(STy);
7708 if (Offset >= (int64_t)SL->getSizeInBytes()) {
7709 // We can't index into this, bail out.
7712 unsigned Elt = SL->getElementContainingOffset(Offset);
7713 NewIndices.push_back(ConstantInt::get(Type::Int32Ty, Elt));
7715 Offset -= SL->getElementOffset(Elt);
7716 Ty = STy->getElementType(Elt);
7717 } else if (isa<ArrayType>(Ty) || isa<VectorType>(Ty)) {
7718 const SequentialType *STy = cast<SequentialType>(Ty);
7719 if (uint64_t EltSize = TD->getABITypeSize(STy->getElementType())) {
7720 NewIndices.push_back(ConstantInt::get(IntPtrTy,Offset/EltSize));
7723 NewIndices.push_back(ConstantInt::get(IntPtrTy, 0));
7725 Ty = STy->getElementType();
7727 // Otherwise, we can't index into this, bail out.
7735 /// @brief Implement the transforms for cast of pointer (bitcast/ptrtoint)
7736 Instruction *InstCombiner::commonPointerCastTransforms(CastInst &CI) {
7737 Value *Src = CI.getOperand(0);
7739 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Src)) {
7740 // If casting the result of a getelementptr instruction with no offset, turn
7741 // this into a cast of the original pointer!
7742 if (GEP->hasAllZeroIndices()) {
7743 // Changing the cast operand is usually not a good idea but it is safe
7744 // here because the pointer operand is being replaced with another
7745 // pointer operand so the opcode doesn't need to change.
7747 CI.setOperand(0, GEP->getOperand(0));
7751 // If the GEP has a single use, and the base pointer is a bitcast, and the
7752 // GEP computes a constant offset, see if we can convert these three
7753 // instructions into fewer. This typically happens with unions and other
7754 // non-type-safe code.
7755 if (GEP->hasOneUse() && isa<BitCastInst>(GEP->getOperand(0))) {
7756 if (GEP->hasAllConstantIndices()) {
7757 // We are guaranteed to get a constant from EmitGEPOffset.
7758 ConstantInt *OffsetV = cast<ConstantInt>(EmitGEPOffset(GEP, CI, *this));
7759 int64_t Offset = OffsetV->getSExtValue();
7761 // Get the base pointer input of the bitcast, and the type it points to.
7762 Value *OrigBase = cast<BitCastInst>(GEP->getOperand(0))->getOperand(0);
7763 const Type *GEPIdxTy =
7764 cast<PointerType>(OrigBase->getType())->getElementType();
7765 SmallVector<Value*, 8> NewIndices;
7766 if (FindElementAtOffset(GEPIdxTy, Offset, NewIndices, TD)) {
7767 // If we were able to index down into an element, create the GEP
7768 // and bitcast the result. This eliminates one bitcast, potentially
7770 Instruction *NGEP = GetElementPtrInst::Create(OrigBase,
7772 NewIndices.end(), "");
7773 InsertNewInstBefore(NGEP, CI);
7774 NGEP->takeName(GEP);
7776 if (isa<BitCastInst>(CI))
7777 return new BitCastInst(NGEP, CI.getType());
7778 assert(isa<PtrToIntInst>(CI));
7779 return new PtrToIntInst(NGEP, CI.getType());
7785 return commonCastTransforms(CI);
7790 /// Only the TRUNC, ZEXT, SEXT, and BITCAST can both operand and result as
7791 /// integer types. This function implements the common transforms for all those
7793 /// @brief Implement the transforms common to CastInst with integer operands
7794 Instruction *InstCombiner::commonIntCastTransforms(CastInst &CI) {
7795 if (Instruction *Result = commonCastTransforms(CI))
7798 Value *Src = CI.getOperand(0);
7799 const Type *SrcTy = Src->getType();
7800 const Type *DestTy = CI.getType();
7801 uint32_t SrcBitSize = SrcTy->getPrimitiveSizeInBits();
7802 uint32_t DestBitSize = DestTy->getPrimitiveSizeInBits();
7804 // See if we can simplify any instructions used by the LHS whose sole
7805 // purpose is to compute bits we don't care about.
7806 APInt KnownZero(DestBitSize, 0), KnownOne(DestBitSize, 0);
7807 if (SimplifyDemandedBits(&CI, APInt::getAllOnesValue(DestBitSize),
7808 KnownZero, KnownOne))
7811 // If the source isn't an instruction or has more than one use then we
7812 // can't do anything more.
7813 Instruction *SrcI = dyn_cast<Instruction>(Src);
7814 if (!SrcI || !Src->hasOneUse())
7817 // Attempt to propagate the cast into the instruction for int->int casts.
7818 int NumCastsRemoved = 0;
7819 if (!isa<BitCastInst>(CI) &&
7820 CanEvaluateInDifferentType(SrcI, cast<IntegerType>(DestTy),
7821 CI.getOpcode(), NumCastsRemoved)) {
7822 // If this cast is a truncate, evaluting in a different type always
7823 // eliminates the cast, so it is always a win. If this is a zero-extension,
7824 // we need to do an AND to maintain the clear top-part of the computation,
7825 // so we require that the input have eliminated at least one cast. If this
7826 // is a sign extension, we insert two new casts (to do the extension) so we
7827 // require that two casts have been eliminated.
7829 switch (CI.getOpcode()) {
7831 // All the others use floating point so we shouldn't actually
7832 // get here because of the check above.
7833 assert(0 && "Unknown cast type");
7834 case Instruction::Trunc:
7837 case Instruction::ZExt:
7838 DoXForm = NumCastsRemoved >= 1;
7840 case Instruction::SExt:
7841 DoXForm = NumCastsRemoved >= 2;
7846 Value *Res = EvaluateInDifferentType(SrcI, DestTy,
7847 CI.getOpcode() == Instruction::SExt);
7848 assert(Res->getType() == DestTy);
7849 switch (CI.getOpcode()) {
7850 default: assert(0 && "Unknown cast type!");
7851 case Instruction::Trunc:
7852 case Instruction::BitCast:
7853 // Just replace this cast with the result.
7854 return ReplaceInstUsesWith(CI, Res);
7855 case Instruction::ZExt: {
7856 // We need to emit an AND to clear the high bits.
7857 assert(SrcBitSize < DestBitSize && "Not a zext?");
7858 Constant *C = ConstantInt::get(APInt::getLowBitsSet(DestBitSize,
7860 return BinaryOperator::CreateAnd(Res, C);
7862 case Instruction::SExt:
7863 // We need to emit a cast to truncate, then a cast to sext.
7864 return CastInst::Create(Instruction::SExt,
7865 InsertCastBefore(Instruction::Trunc, Res, Src->getType(),
7871 Value *Op0 = SrcI->getNumOperands() > 0 ? SrcI->getOperand(0) : 0;
7872 Value *Op1 = SrcI->getNumOperands() > 1 ? SrcI->getOperand(1) : 0;
7874 switch (SrcI->getOpcode()) {
7875 case Instruction::Add:
7876 case Instruction::Mul:
7877 case Instruction::And:
7878 case Instruction::Or:
7879 case Instruction::Xor:
7880 // If we are discarding information, rewrite.
7881 if (DestBitSize <= SrcBitSize && DestBitSize != 1) {
7882 // Don't insert two casts if they cannot be eliminated. We allow
7883 // two casts to be inserted if the sizes are the same. This could
7884 // only be converting signedness, which is a noop.
7885 if (DestBitSize == SrcBitSize ||
7886 !ValueRequiresCast(CI.getOpcode(), Op1, DestTy,TD) ||
7887 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
7888 Instruction::CastOps opcode = CI.getOpcode();
7889 Value *Op0c = InsertCastBefore(opcode, Op0, DestTy, *SrcI);
7890 Value *Op1c = InsertCastBefore(opcode, Op1, DestTy, *SrcI);
7891 return BinaryOperator::Create(
7892 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
7896 // cast (xor bool X, true) to int --> xor (cast bool X to int), 1
7897 if (isa<ZExtInst>(CI) && SrcBitSize == 1 &&
7898 SrcI->getOpcode() == Instruction::Xor &&
7899 Op1 == ConstantInt::getTrue() &&
7900 (!Op0->hasOneUse() || !isa<CmpInst>(Op0))) {
7901 Value *New = InsertCastBefore(Instruction::ZExt, Op0, DestTy, CI);
7902 return BinaryOperator::CreateXor(New, ConstantInt::get(CI.getType(), 1));
7905 case Instruction::SDiv:
7906 case Instruction::UDiv:
7907 case Instruction::SRem:
7908 case Instruction::URem:
7909 // If we are just changing the sign, rewrite.
7910 if (DestBitSize == SrcBitSize) {
7911 // Don't insert two casts if they cannot be eliminated. We allow
7912 // two casts to be inserted if the sizes are the same. This could
7913 // only be converting signedness, which is a noop.
7914 if (!ValueRequiresCast(CI.getOpcode(), Op1, DestTy, TD) ||
7915 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
7916 Value *Op0c = InsertCastBefore(Instruction::BitCast,
7917 Op0, DestTy, *SrcI);
7918 Value *Op1c = InsertCastBefore(Instruction::BitCast,
7919 Op1, DestTy, *SrcI);
7920 return BinaryOperator::Create(
7921 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
7926 case Instruction::Shl:
7927 // Allow changing the sign of the source operand. Do not allow
7928 // changing the size of the shift, UNLESS the shift amount is a
7929 // constant. We must not change variable sized shifts to a smaller
7930 // size, because it is undefined to shift more bits out than exist
7932 if (DestBitSize == SrcBitSize ||
7933 (DestBitSize < SrcBitSize && isa<Constant>(Op1))) {
7934 Instruction::CastOps opcode = (DestBitSize == SrcBitSize ?
7935 Instruction::BitCast : Instruction::Trunc);
7936 Value *Op0c = InsertCastBefore(opcode, Op0, DestTy, *SrcI);
7937 Value *Op1c = InsertCastBefore(opcode, Op1, DestTy, *SrcI);
7938 return BinaryOperator::CreateShl(Op0c, Op1c);
7941 case Instruction::AShr:
7942 // If this is a signed shr, and if all bits shifted in are about to be
7943 // truncated off, turn it into an unsigned shr to allow greater
7945 if (DestBitSize < SrcBitSize &&
7946 isa<ConstantInt>(Op1)) {
7947 uint32_t ShiftAmt = cast<ConstantInt>(Op1)->getLimitedValue(SrcBitSize);
7948 if (SrcBitSize > ShiftAmt && SrcBitSize-ShiftAmt >= DestBitSize) {
7949 // Insert the new logical shift right.
7950 return BinaryOperator::CreateLShr(Op0, Op1);
7958 Instruction *InstCombiner::visitTrunc(TruncInst &CI) {
7959 if (Instruction *Result = commonIntCastTransforms(CI))
7962 Value *Src = CI.getOperand(0);
7963 const Type *Ty = CI.getType();
7964 uint32_t DestBitWidth = Ty->getPrimitiveSizeInBits();
7965 uint32_t SrcBitWidth = cast<IntegerType>(Src->getType())->getBitWidth();
7967 if (Instruction *SrcI = dyn_cast<Instruction>(Src)) {
7968 switch (SrcI->getOpcode()) {
7970 case Instruction::LShr:
7971 // We can shrink lshr to something smaller if we know the bits shifted in
7972 // are already zeros.
7973 if (ConstantInt *ShAmtV = dyn_cast<ConstantInt>(SrcI->getOperand(1))) {
7974 uint32_t ShAmt = ShAmtV->getLimitedValue(SrcBitWidth);
7976 // Get a mask for the bits shifting in.
7977 APInt Mask(APInt::getLowBitsSet(SrcBitWidth, ShAmt).shl(DestBitWidth));
7978 Value* SrcIOp0 = SrcI->getOperand(0);
7979 if (SrcI->hasOneUse() && MaskedValueIsZero(SrcIOp0, Mask)) {
7980 if (ShAmt >= DestBitWidth) // All zeros.
7981 return ReplaceInstUsesWith(CI, Constant::getNullValue(Ty));
7983 // Okay, we can shrink this. Truncate the input, then return a new
7985 Value *V1 = InsertCastBefore(Instruction::Trunc, SrcIOp0, Ty, CI);
7986 Value *V2 = InsertCastBefore(Instruction::Trunc, SrcI->getOperand(1),
7988 return BinaryOperator::CreateLShr(V1, V2);
7990 } else { // This is a variable shr.
7992 // Turn 'trunc (lshr X, Y) to bool' into '(X & (1 << Y)) != 0'. This is
7993 // more LLVM instructions, but allows '1 << Y' to be hoisted if
7994 // loop-invariant and CSE'd.
7995 if (CI.getType() == Type::Int1Ty && SrcI->hasOneUse()) {
7996 Value *One = ConstantInt::get(SrcI->getType(), 1);
7998 Value *V = InsertNewInstBefore(
7999 BinaryOperator::CreateShl(One, SrcI->getOperand(1),
8001 V = InsertNewInstBefore(BinaryOperator::CreateAnd(V,
8002 SrcI->getOperand(0),
8004 Value *Zero = Constant::getNullValue(V->getType());
8005 return new ICmpInst(ICmpInst::ICMP_NE, V, Zero);
8015 /// transformZExtICmp - Transform (zext icmp) to bitwise / integer operations
8016 /// in order to eliminate the icmp.
8017 Instruction *InstCombiner::transformZExtICmp(ICmpInst *ICI, Instruction &CI,
8019 // If we are just checking for a icmp eq of a single bit and zext'ing it
8020 // to an integer, then shift the bit to the appropriate place and then
8021 // cast to integer to avoid the comparison.
8022 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(ICI->getOperand(1))) {
8023 const APInt &Op1CV = Op1C->getValue();
8025 // zext (x <s 0) to i32 --> x>>u31 true if signbit set.
8026 // zext (x >s -1) to i32 --> (x>>u31)^1 true if signbit clear.
8027 if ((ICI->getPredicate() == ICmpInst::ICMP_SLT && Op1CV == 0) ||
8028 (ICI->getPredicate() == ICmpInst::ICMP_SGT &&Op1CV.isAllOnesValue())) {
8029 if (!DoXform) return ICI;
8031 Value *In = ICI->getOperand(0);
8032 Value *Sh = ConstantInt::get(In->getType(),
8033 In->getType()->getPrimitiveSizeInBits()-1);
8034 In = InsertNewInstBefore(BinaryOperator::CreateLShr(In, Sh,
8035 In->getName()+".lobit"),
8037 if (In->getType() != CI.getType())
8038 In = CastInst::CreateIntegerCast(In, CI.getType(),
8039 false/*ZExt*/, "tmp", &CI);
8041 if (ICI->getPredicate() == ICmpInst::ICMP_SGT) {
8042 Constant *One = ConstantInt::get(In->getType(), 1);
8043 In = InsertNewInstBefore(BinaryOperator::CreateXor(In, One,
8044 In->getName()+".not"),
8048 return ReplaceInstUsesWith(CI, In);
8053 // zext (X == 0) to i32 --> X^1 iff X has only the low bit set.
8054 // zext (X == 0) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
8055 // zext (X == 1) to i32 --> X iff X has only the low bit set.
8056 // zext (X == 2) to i32 --> X>>1 iff X has only the 2nd bit set.
8057 // zext (X != 0) to i32 --> X iff X has only the low bit set.
8058 // zext (X != 0) to i32 --> X>>1 iff X has only the 2nd bit set.
8059 // zext (X != 1) to i32 --> X^1 iff X has only the low bit set.
8060 // zext (X != 2) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
8061 if ((Op1CV == 0 || Op1CV.isPowerOf2()) &&
8062 // This only works for EQ and NE
8063 ICI->isEquality()) {
8064 // If Op1C some other power of two, convert:
8065 uint32_t BitWidth = Op1C->getType()->getBitWidth();
8066 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
8067 APInt TypeMask(APInt::getAllOnesValue(BitWidth));
8068 ComputeMaskedBits(ICI->getOperand(0), TypeMask, KnownZero, KnownOne);
8070 APInt KnownZeroMask(~KnownZero);
8071 if (KnownZeroMask.isPowerOf2()) { // Exactly 1 possible 1?
8072 if (!DoXform) return ICI;
8074 bool isNE = ICI->getPredicate() == ICmpInst::ICMP_NE;
8075 if (Op1CV != 0 && (Op1CV != KnownZeroMask)) {
8076 // (X&4) == 2 --> false
8077 // (X&4) != 2 --> true
8078 Constant *Res = ConstantInt::get(Type::Int1Ty, isNE);
8079 Res = ConstantExpr::getZExt(Res, CI.getType());
8080 return ReplaceInstUsesWith(CI, Res);
8083 uint32_t ShiftAmt = KnownZeroMask.logBase2();
8084 Value *In = ICI->getOperand(0);
8086 // Perform a logical shr by shiftamt.
8087 // Insert the shift to put the result in the low bit.
8088 In = InsertNewInstBefore(BinaryOperator::CreateLShr(In,
8089 ConstantInt::get(In->getType(), ShiftAmt),
8090 In->getName()+".lobit"), CI);
8093 if ((Op1CV != 0) == isNE) { // Toggle the low bit.
8094 Constant *One = ConstantInt::get(In->getType(), 1);
8095 In = BinaryOperator::CreateXor(In, One, "tmp");
8096 InsertNewInstBefore(cast<Instruction>(In), CI);
8099 if (CI.getType() == In->getType())
8100 return ReplaceInstUsesWith(CI, In);
8102 return CastInst::CreateIntegerCast(In, CI.getType(), false/*ZExt*/);
8110 Instruction *InstCombiner::visitZExt(ZExtInst &CI) {
8111 // If one of the common conversion will work ..
8112 if (Instruction *Result = commonIntCastTransforms(CI))
8115 Value *Src = CI.getOperand(0);
8117 // If this is a cast of a cast
8118 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
8119 // If this is a TRUNC followed by a ZEXT then we are dealing with integral
8120 // types and if the sizes are just right we can convert this into a logical
8121 // 'and' which will be much cheaper than the pair of casts.
8122 if (isa<TruncInst>(CSrc)) {
8123 // Get the sizes of the types involved
8124 Value *A = CSrc->getOperand(0);
8125 uint32_t SrcSize = A->getType()->getPrimitiveSizeInBits();
8126 uint32_t MidSize = CSrc->getType()->getPrimitiveSizeInBits();
8127 uint32_t DstSize = CI.getType()->getPrimitiveSizeInBits();
8128 // If we're actually extending zero bits and the trunc is a no-op
8129 if (MidSize < DstSize && SrcSize == DstSize) {
8130 // Replace both of the casts with an And of the type mask.
8131 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
8132 Constant *AndConst = ConstantInt::get(AndValue);
8134 BinaryOperator::CreateAnd(CSrc->getOperand(0), AndConst);
8135 // Unfortunately, if the type changed, we need to cast it back.
8136 if (And->getType() != CI.getType()) {
8137 And->setName(CSrc->getName()+".mask");
8138 InsertNewInstBefore(And, CI);
8139 And = CastInst::CreateIntegerCast(And, CI.getType(), false/*ZExt*/);
8146 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src))
8147 return transformZExtICmp(ICI, CI);
8149 BinaryOperator *SrcI = dyn_cast<BinaryOperator>(Src);
8150 if (SrcI && SrcI->getOpcode() == Instruction::Or) {
8151 // zext (or icmp, icmp) --> or (zext icmp), (zext icmp) if at least one
8152 // of the (zext icmp) will be transformed.
8153 ICmpInst *LHS = dyn_cast<ICmpInst>(SrcI->getOperand(0));
8154 ICmpInst *RHS = dyn_cast<ICmpInst>(SrcI->getOperand(1));
8155 if (LHS && RHS && LHS->hasOneUse() && RHS->hasOneUse() &&
8156 (transformZExtICmp(LHS, CI, false) ||
8157 transformZExtICmp(RHS, CI, false))) {
8158 Value *LCast = InsertCastBefore(Instruction::ZExt, LHS, CI.getType(), CI);
8159 Value *RCast = InsertCastBefore(Instruction::ZExt, RHS, CI.getType(), CI);
8160 return BinaryOperator::Create(Instruction::Or, LCast, RCast);
8167 Instruction *InstCombiner::visitSExt(SExtInst &CI) {
8168 if (Instruction *I = commonIntCastTransforms(CI))
8171 Value *Src = CI.getOperand(0);
8173 // Canonicalize sign-extend from i1 to a select.
8174 if (Src->getType() == Type::Int1Ty)
8175 return SelectInst::Create(Src,
8176 ConstantInt::getAllOnesValue(CI.getType()),
8177 Constant::getNullValue(CI.getType()));
8179 // See if the value being truncated is already sign extended. If so, just
8180 // eliminate the trunc/sext pair.
8181 if (getOpcode(Src) == Instruction::Trunc) {
8182 Value *Op = cast<User>(Src)->getOperand(0);
8183 unsigned OpBits = cast<IntegerType>(Op->getType())->getBitWidth();
8184 unsigned MidBits = cast<IntegerType>(Src->getType())->getBitWidth();
8185 unsigned DestBits = cast<IntegerType>(CI.getType())->getBitWidth();
8186 unsigned NumSignBits = ComputeNumSignBits(Op);
8188 if (OpBits == DestBits) {
8189 // Op is i32, Mid is i8, and Dest is i32. If Op has more than 24 sign
8190 // bits, it is already ready.
8191 if (NumSignBits > DestBits-MidBits)
8192 return ReplaceInstUsesWith(CI, Op);
8193 } else if (OpBits < DestBits) {
8194 // Op is i32, Mid is i8, and Dest is i64. If Op has more than 24 sign
8195 // bits, just sext from i32.
8196 if (NumSignBits > OpBits-MidBits)
8197 return new SExtInst(Op, CI.getType(), "tmp");
8199 // Op is i64, Mid is i8, and Dest is i32. If Op has more than 56 sign
8200 // bits, just truncate to i32.
8201 if (NumSignBits > OpBits-MidBits)
8202 return new TruncInst(Op, CI.getType(), "tmp");
8206 // If the input is a shl/ashr pair of a same constant, then this is a sign
8207 // extension from a smaller value. If we could trust arbitrary bitwidth
8208 // integers, we could turn this into a truncate to the smaller bit and then
8209 // use a sext for the whole extension. Since we don't, look deeper and check
8210 // for a truncate. If the source and dest are the same type, eliminate the
8211 // trunc and extend and just do shifts. For example, turn:
8212 // %a = trunc i32 %i to i8
8213 // %b = shl i8 %a, 6
8214 // %c = ashr i8 %b, 6
8215 // %d = sext i8 %c to i32
8217 // %a = shl i32 %i, 30
8218 // %d = ashr i32 %a, 30
8220 ConstantInt *BA = 0, *CA = 0;
8221 if (match(Src, m_AShr(m_Shl(m_Value(A), m_ConstantInt(BA)),
8222 m_ConstantInt(CA))) &&
8223 BA == CA && isa<TruncInst>(A)) {
8224 Value *I = cast<TruncInst>(A)->getOperand(0);
8225 if (I->getType() == CI.getType()) {
8226 unsigned MidSize = Src->getType()->getPrimitiveSizeInBits();
8227 unsigned SrcDstSize = CI.getType()->getPrimitiveSizeInBits();
8228 unsigned ShAmt = CA->getZExtValue()+SrcDstSize-MidSize;
8229 Constant *ShAmtV = ConstantInt::get(CI.getType(), ShAmt);
8230 I = InsertNewInstBefore(BinaryOperator::CreateShl(I, ShAmtV,
8232 return BinaryOperator::CreateAShr(I, ShAmtV);
8239 /// FitsInFPType - Return a Constant* for the specified FP constant if it fits
8240 /// in the specified FP type without changing its value.
8241 static Constant *FitsInFPType(ConstantFP *CFP, const fltSemantics &Sem) {
8243 APFloat F = CFP->getValueAPF();
8244 (void)F.convert(Sem, APFloat::rmNearestTiesToEven, &losesInfo);
8246 return ConstantFP::get(F);
8250 /// LookThroughFPExtensions - If this is an fp extension instruction, look
8251 /// through it until we get the source value.
8252 static Value *LookThroughFPExtensions(Value *V) {
8253 if (Instruction *I = dyn_cast<Instruction>(V))
8254 if (I->getOpcode() == Instruction::FPExt)
8255 return LookThroughFPExtensions(I->getOperand(0));
8257 // If this value is a constant, return the constant in the smallest FP type
8258 // that can accurately represent it. This allows us to turn
8259 // (float)((double)X+2.0) into x+2.0f.
8260 if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
8261 if (CFP->getType() == Type::PPC_FP128Ty)
8262 return V; // No constant folding of this.
8263 // See if the value can be truncated to float and then reextended.
8264 if (Value *V = FitsInFPType(CFP, APFloat::IEEEsingle))
8266 if (CFP->getType() == Type::DoubleTy)
8267 return V; // Won't shrink.
8268 if (Value *V = FitsInFPType(CFP, APFloat::IEEEdouble))
8270 // Don't try to shrink to various long double types.
8276 Instruction *InstCombiner::visitFPTrunc(FPTruncInst &CI) {
8277 if (Instruction *I = commonCastTransforms(CI))
8280 // If we have fptrunc(add (fpextend x), (fpextend y)), where x and y are
8281 // smaller than the destination type, we can eliminate the truncate by doing
8282 // the add as the smaller type. This applies to add/sub/mul/div as well as
8283 // many builtins (sqrt, etc).
8284 BinaryOperator *OpI = dyn_cast<BinaryOperator>(CI.getOperand(0));
8285 if (OpI && OpI->hasOneUse()) {
8286 switch (OpI->getOpcode()) {
8288 case Instruction::Add:
8289 case Instruction::Sub:
8290 case Instruction::Mul:
8291 case Instruction::FDiv:
8292 case Instruction::FRem:
8293 const Type *SrcTy = OpI->getType();
8294 Value *LHSTrunc = LookThroughFPExtensions(OpI->getOperand(0));
8295 Value *RHSTrunc = LookThroughFPExtensions(OpI->getOperand(1));
8296 if (LHSTrunc->getType() != SrcTy &&
8297 RHSTrunc->getType() != SrcTy) {
8298 unsigned DstSize = CI.getType()->getPrimitiveSizeInBits();
8299 // If the source types were both smaller than the destination type of
8300 // the cast, do this xform.
8301 if (LHSTrunc->getType()->getPrimitiveSizeInBits() <= DstSize &&
8302 RHSTrunc->getType()->getPrimitiveSizeInBits() <= DstSize) {
8303 LHSTrunc = InsertCastBefore(Instruction::FPExt, LHSTrunc,
8305 RHSTrunc = InsertCastBefore(Instruction::FPExt, RHSTrunc,
8307 return BinaryOperator::Create(OpI->getOpcode(), LHSTrunc, RHSTrunc);
8316 Instruction *InstCombiner::visitFPExt(CastInst &CI) {
8317 return commonCastTransforms(CI);
8320 Instruction *InstCombiner::visitFPToUI(FPToUIInst &FI) {
8321 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0));
8323 return commonCastTransforms(FI);
8325 // fptoui(uitofp(X)) --> X
8326 // fptoui(sitofp(X)) --> X
8327 // This is safe if the intermediate type has enough bits in its mantissa to
8328 // accurately represent all values of X. For example, do not do this with
8329 // i64->float->i64. This is also safe for sitofp case, because any negative
8330 // 'X' value would cause an undefined result for the fptoui.
8331 if ((isa<UIToFPInst>(OpI) || isa<SIToFPInst>(OpI)) &&
8332 OpI->getOperand(0)->getType() == FI.getType() &&
8333 (int)FI.getType()->getPrimitiveSizeInBits() < /*extra bit for sign */
8334 OpI->getType()->getFPMantissaWidth())
8335 return ReplaceInstUsesWith(FI, OpI->getOperand(0));
8337 return commonCastTransforms(FI);
8340 Instruction *InstCombiner::visitFPToSI(FPToSIInst &FI) {
8341 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0));
8343 return commonCastTransforms(FI);
8345 // fptosi(sitofp(X)) --> X
8346 // fptosi(uitofp(X)) --> X
8347 // This is safe if the intermediate type has enough bits in its mantissa to
8348 // accurately represent all values of X. For example, do not do this with
8349 // i64->float->i64. This is also safe for sitofp case, because any negative
8350 // 'X' value would cause an undefined result for the fptoui.
8351 if ((isa<UIToFPInst>(OpI) || isa<SIToFPInst>(OpI)) &&
8352 OpI->getOperand(0)->getType() == FI.getType() &&
8353 (int)FI.getType()->getPrimitiveSizeInBits() <=
8354 OpI->getType()->getFPMantissaWidth())
8355 return ReplaceInstUsesWith(FI, OpI->getOperand(0));
8357 return commonCastTransforms(FI);
8360 Instruction *InstCombiner::visitUIToFP(CastInst &CI) {
8361 return commonCastTransforms(CI);
8364 Instruction *InstCombiner::visitSIToFP(CastInst &CI) {
8365 return commonCastTransforms(CI);
8368 Instruction *InstCombiner::visitPtrToInt(CastInst &CI) {
8369 return commonPointerCastTransforms(CI);
8372 Instruction *InstCombiner::visitIntToPtr(IntToPtrInst &CI) {
8373 if (Instruction *I = commonCastTransforms(CI))
8376 const Type *DestPointee = cast<PointerType>(CI.getType())->getElementType();
8377 if (!DestPointee->isSized()) return 0;
8379 // If this is inttoptr(add (ptrtoint x), cst), try to turn this into a GEP.
8382 if (match(CI.getOperand(0), m_Add(m_Cast<PtrToIntInst>(m_Value(X)),
8383 m_ConstantInt(Cst)))) {
8384 // If the source and destination operands have the same type, see if this
8385 // is a single-index GEP.
8386 if (X->getType() == CI.getType()) {
8387 // Get the size of the pointee type.
8388 uint64_t Size = TD->getABITypeSize(DestPointee);
8390 // Convert the constant to intptr type.
8391 APInt Offset = Cst->getValue();
8392 Offset.sextOrTrunc(TD->getPointerSizeInBits());
8394 // If Offset is evenly divisible by Size, we can do this xform.
8395 if (Size && !APIntOps::srem(Offset, APInt(Offset.getBitWidth(), Size))){
8396 Offset = APIntOps::sdiv(Offset, APInt(Offset.getBitWidth(), Size));
8397 return GetElementPtrInst::Create(X, ConstantInt::get(Offset));
8400 // TODO: Could handle other cases, e.g. where add is indexing into field of
8402 } else if (CI.getOperand(0)->hasOneUse() &&
8403 match(CI.getOperand(0), m_Add(m_Value(X), m_ConstantInt(Cst)))) {
8404 // Otherwise, if this is inttoptr(add x, cst), try to turn this into an
8405 // "inttoptr+GEP" instead of "add+intptr".
8407 // Get the size of the pointee type.
8408 uint64_t Size = TD->getABITypeSize(DestPointee);
8410 // Convert the constant to intptr type.
8411 APInt Offset = Cst->getValue();
8412 Offset.sextOrTrunc(TD->getPointerSizeInBits());
8414 // If Offset is evenly divisible by Size, we can do this xform.
8415 if (Size && !APIntOps::srem(Offset, APInt(Offset.getBitWidth(), Size))){
8416 Offset = APIntOps::sdiv(Offset, APInt(Offset.getBitWidth(), Size));
8418 Instruction *P = InsertNewInstBefore(new IntToPtrInst(X, CI.getType(),
8420 return GetElementPtrInst::Create(P, ConstantInt::get(Offset), "tmp");
8426 Instruction *InstCombiner::visitBitCast(BitCastInst &CI) {
8427 // If the operands are integer typed then apply the integer transforms,
8428 // otherwise just apply the common ones.
8429 Value *Src = CI.getOperand(0);
8430 const Type *SrcTy = Src->getType();
8431 const Type *DestTy = CI.getType();
8433 if (SrcTy->isInteger() && DestTy->isInteger()) {
8434 if (Instruction *Result = commonIntCastTransforms(CI))
8436 } else if (isa<PointerType>(SrcTy)) {
8437 if (Instruction *I = commonPointerCastTransforms(CI))
8440 if (Instruction *Result = commonCastTransforms(CI))
8445 // Get rid of casts from one type to the same type. These are useless and can
8446 // be replaced by the operand.
8447 if (DestTy == Src->getType())
8448 return ReplaceInstUsesWith(CI, Src);
8450 if (const PointerType *DstPTy = dyn_cast<PointerType>(DestTy)) {
8451 const PointerType *SrcPTy = cast<PointerType>(SrcTy);
8452 const Type *DstElTy = DstPTy->getElementType();
8453 const Type *SrcElTy = SrcPTy->getElementType();
8455 // If the address spaces don't match, don't eliminate the bitcast, which is
8456 // required for changing types.
8457 if (SrcPTy->getAddressSpace() != DstPTy->getAddressSpace())
8460 // If we are casting a malloc or alloca to a pointer to a type of the same
8461 // size, rewrite the allocation instruction to allocate the "right" type.
8462 if (AllocationInst *AI = dyn_cast<AllocationInst>(Src))
8463 if (Instruction *V = PromoteCastOfAllocation(CI, *AI))
8466 // If the source and destination are pointers, and this cast is equivalent
8467 // to a getelementptr X, 0, 0, 0... turn it into the appropriate gep.
8468 // This can enhance SROA and other transforms that want type-safe pointers.
8469 Constant *ZeroUInt = Constant::getNullValue(Type::Int32Ty);
8470 unsigned NumZeros = 0;
8471 while (SrcElTy != DstElTy &&
8472 isa<CompositeType>(SrcElTy) && !isa<PointerType>(SrcElTy) &&
8473 SrcElTy->getNumContainedTypes() /* not "{}" */) {
8474 SrcElTy = cast<CompositeType>(SrcElTy)->getTypeAtIndex(ZeroUInt);
8478 // If we found a path from the src to dest, create the getelementptr now.
8479 if (SrcElTy == DstElTy) {
8480 SmallVector<Value*, 8> Idxs(NumZeros+1, ZeroUInt);
8481 return GetElementPtrInst::Create(Src, Idxs.begin(), Idxs.end(), "",
8482 ((Instruction*) NULL));
8486 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(Src)) {
8487 if (SVI->hasOneUse()) {
8488 // Okay, we have (bitconvert (shuffle ..)). Check to see if this is
8489 // a bitconvert to a vector with the same # elts.
8490 if (isa<VectorType>(DestTy) &&
8491 cast<VectorType>(DestTy)->getNumElements() ==
8492 SVI->getType()->getNumElements() &&
8493 SVI->getType()->getNumElements() ==
8494 cast<VectorType>(SVI->getOperand(0)->getType())->getNumElements()) {
8496 // If either of the operands is a cast from CI.getType(), then
8497 // evaluating the shuffle in the casted destination's type will allow
8498 // us to eliminate at least one cast.
8499 if (((Tmp = dyn_cast<CastInst>(SVI->getOperand(0))) &&
8500 Tmp->getOperand(0)->getType() == DestTy) ||
8501 ((Tmp = dyn_cast<CastInst>(SVI->getOperand(1))) &&
8502 Tmp->getOperand(0)->getType() == DestTy)) {
8503 Value *LHS = InsertCastBefore(Instruction::BitCast,
8504 SVI->getOperand(0), DestTy, CI);
8505 Value *RHS = InsertCastBefore(Instruction::BitCast,
8506 SVI->getOperand(1), DestTy, CI);
8507 // Return a new shuffle vector. Use the same element ID's, as we
8508 // know the vector types match #elts.
8509 return new ShuffleVectorInst(LHS, RHS, SVI->getOperand(2));
8517 /// GetSelectFoldableOperands - We want to turn code that looks like this:
8519 /// %D = select %cond, %C, %A
8521 /// %C = select %cond, %B, 0
8524 /// Assuming that the specified instruction is an operand to the select, return
8525 /// a bitmask indicating which operands of this instruction are foldable if they
8526 /// equal the other incoming value of the select.
8528 static unsigned GetSelectFoldableOperands(Instruction *I) {
8529 switch (I->getOpcode()) {
8530 case Instruction::Add:
8531 case Instruction::Mul:
8532 case Instruction::And:
8533 case Instruction::Or:
8534 case Instruction::Xor:
8535 return 3; // Can fold through either operand.
8536 case Instruction::Sub: // Can only fold on the amount subtracted.
8537 case Instruction::Shl: // Can only fold on the shift amount.
8538 case Instruction::LShr:
8539 case Instruction::AShr:
8542 return 0; // Cannot fold
8546 /// GetSelectFoldableConstant - For the same transformation as the previous
8547 /// function, return the identity constant that goes into the select.
8548 static Constant *GetSelectFoldableConstant(Instruction *I) {
8549 switch (I->getOpcode()) {
8550 default: assert(0 && "This cannot happen!"); abort();
8551 case Instruction::Add:
8552 case Instruction::Sub:
8553 case Instruction::Or:
8554 case Instruction::Xor:
8555 case Instruction::Shl:
8556 case Instruction::LShr:
8557 case Instruction::AShr:
8558 return Constant::getNullValue(I->getType());
8559 case Instruction::And:
8560 return Constant::getAllOnesValue(I->getType());
8561 case Instruction::Mul:
8562 return ConstantInt::get(I->getType(), 1);
8566 /// FoldSelectOpOp - Here we have (select c, TI, FI), and we know that TI and FI
8567 /// have the same opcode and only one use each. Try to simplify this.
8568 Instruction *InstCombiner::FoldSelectOpOp(SelectInst &SI, Instruction *TI,
8570 if (TI->getNumOperands() == 1) {
8571 // If this is a non-volatile load or a cast from the same type,
8574 if (TI->getOperand(0)->getType() != FI->getOperand(0)->getType())
8577 return 0; // unknown unary op.
8580 // Fold this by inserting a select from the input values.
8581 SelectInst *NewSI = SelectInst::Create(SI.getCondition(), TI->getOperand(0),
8582 FI->getOperand(0), SI.getName()+".v");
8583 InsertNewInstBefore(NewSI, SI);
8584 return CastInst::Create(Instruction::CastOps(TI->getOpcode()), NewSI,
8588 // Only handle binary operators here.
8589 if (!isa<BinaryOperator>(TI))
8592 // Figure out if the operations have any operands in common.
8593 Value *MatchOp, *OtherOpT, *OtherOpF;
8595 if (TI->getOperand(0) == FI->getOperand(0)) {
8596 MatchOp = TI->getOperand(0);
8597 OtherOpT = TI->getOperand(1);
8598 OtherOpF = FI->getOperand(1);
8599 MatchIsOpZero = true;
8600 } else if (TI->getOperand(1) == FI->getOperand(1)) {
8601 MatchOp = TI->getOperand(1);
8602 OtherOpT = TI->getOperand(0);
8603 OtherOpF = FI->getOperand(0);
8604 MatchIsOpZero = false;
8605 } else if (!TI->isCommutative()) {
8607 } else if (TI->getOperand(0) == FI->getOperand(1)) {
8608 MatchOp = TI->getOperand(0);
8609 OtherOpT = TI->getOperand(1);
8610 OtherOpF = FI->getOperand(0);
8611 MatchIsOpZero = true;
8612 } else if (TI->getOperand(1) == FI->getOperand(0)) {
8613 MatchOp = TI->getOperand(1);
8614 OtherOpT = TI->getOperand(0);
8615 OtherOpF = FI->getOperand(1);
8616 MatchIsOpZero = true;
8621 // If we reach here, they do have operations in common.
8622 SelectInst *NewSI = SelectInst::Create(SI.getCondition(), OtherOpT,
8623 OtherOpF, SI.getName()+".v");
8624 InsertNewInstBefore(NewSI, SI);
8626 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TI)) {
8628 return BinaryOperator::Create(BO->getOpcode(), MatchOp, NewSI);
8630 return BinaryOperator::Create(BO->getOpcode(), NewSI, MatchOp);
8632 assert(0 && "Shouldn't get here");
8636 /// visitSelectInstWithICmp - Visit a SelectInst that has an
8637 /// ICmpInst as its first operand.
8639 Instruction *InstCombiner::visitSelectInstWithICmp(SelectInst &SI,
8641 bool Changed = false;
8642 ICmpInst::Predicate Pred = ICI->getPredicate();
8643 Value *CmpLHS = ICI->getOperand(0);
8644 Value *CmpRHS = ICI->getOperand(1);
8645 Value *TrueVal = SI.getTrueValue();
8646 Value *FalseVal = SI.getFalseValue();
8648 // Check cases where the comparison is with a constant that
8649 // can be adjusted to fit the min/max idiom. We may edit ICI in
8650 // place here, so make sure the select is the only user.
8651 if (ICI->hasOneUse())
8652 if (ConstantInt *CI = dyn_cast<ConstantInt>(CmpRHS)) {
8655 case ICmpInst::ICMP_ULT:
8656 case ICmpInst::ICMP_SLT: {
8657 // X < MIN ? T : F --> F
8658 if (CI->isMinValue(Pred == ICmpInst::ICMP_SLT))
8659 return ReplaceInstUsesWith(SI, FalseVal);
8660 // X < C ? X : C-1 --> X > C-1 ? C-1 : X
8661 Constant *AdjustedRHS = SubOne(CI);
8662 if ((CmpLHS == TrueVal && AdjustedRHS == FalseVal) ||
8663 (CmpLHS == FalseVal && AdjustedRHS == TrueVal)) {
8664 Pred = ICmpInst::getSwappedPredicate(Pred);
8665 CmpRHS = AdjustedRHS;
8666 std::swap(FalseVal, TrueVal);
8667 ICI->setPredicate(Pred);
8668 ICI->setOperand(1, CmpRHS);
8669 SI.setOperand(1, TrueVal);
8670 SI.setOperand(2, FalseVal);
8675 case ICmpInst::ICMP_UGT:
8676 case ICmpInst::ICMP_SGT: {
8677 // X > MAX ? T : F --> F
8678 if (CI->isMaxValue(Pred == ICmpInst::ICMP_SGT))
8679 return ReplaceInstUsesWith(SI, FalseVal);
8680 // X > C ? X : C+1 --> X < C+1 ? C+1 : X
8681 Constant *AdjustedRHS = AddOne(CI);
8682 if ((CmpLHS == TrueVal && AdjustedRHS == FalseVal) ||
8683 (CmpLHS == FalseVal && AdjustedRHS == TrueVal)) {
8684 Pred = ICmpInst::getSwappedPredicate(Pred);
8685 CmpRHS = AdjustedRHS;
8686 std::swap(FalseVal, TrueVal);
8687 ICI->setPredicate(Pred);
8688 ICI->setOperand(1, CmpRHS);
8689 SI.setOperand(1, TrueVal);
8690 SI.setOperand(2, FalseVal);
8697 // (x <s 0) ? -1 : 0 -> ashr x, 31 -> all ones if signed
8698 // (x >s -1) ? -1 : 0 -> ashr x, 31 -> all ones if not signed
8699 CmpInst::Predicate Pred = CmpInst::BAD_ICMP_PREDICATE;
8700 if (match(TrueVal, m_ConstantInt<-1>()) &&
8701 match(FalseVal, m_ConstantInt<0>()))
8702 Pred = ICI->getPredicate();
8703 else if (match(TrueVal, m_ConstantInt<0>()) &&
8704 match(FalseVal, m_ConstantInt<-1>()))
8705 Pred = CmpInst::getInversePredicate(ICI->getPredicate());
8707 if (Pred != CmpInst::BAD_ICMP_PREDICATE) {
8708 // If we are just checking for a icmp eq of a single bit and zext'ing it
8709 // to an integer, then shift the bit to the appropriate place and then
8710 // cast to integer to avoid the comparison.
8711 const APInt &Op1CV = CI->getValue();
8713 // sext (x <s 0) to i32 --> x>>s31 true if signbit set.
8714 // sext (x >s -1) to i32 --> (x>>s31)^-1 true if signbit clear.
8715 if ((Pred == ICmpInst::ICMP_SLT && Op1CV == 0) ||
8716 (Pred == ICmpInst::ICMP_SGT && Op1CV.isAllOnesValue())) {
8717 Value *In = ICI->getOperand(0);
8718 Value *Sh = ConstantInt::get(In->getType(),
8719 In->getType()->getPrimitiveSizeInBits()-1);
8720 In = InsertNewInstBefore(BinaryOperator::CreateAShr(In, Sh,
8721 In->getName()+".lobit"),
8723 if (In->getType() != SI.getType())
8724 In = CastInst::CreateIntegerCast(In, SI.getType(),
8725 true/*SExt*/, "tmp", ICI);
8727 if (Pred == ICmpInst::ICMP_SGT)
8728 In = InsertNewInstBefore(BinaryOperator::CreateNot(In,
8729 In->getName()+".not"), *ICI);
8731 return ReplaceInstUsesWith(SI, In);
8736 if (CmpLHS == TrueVal && CmpRHS == FalseVal) {
8737 // Transform (X == Y) ? X : Y -> Y
8738 if (Pred == ICmpInst::ICMP_EQ)
8739 return ReplaceInstUsesWith(SI, FalseVal);
8740 // Transform (X != Y) ? X : Y -> X
8741 if (Pred == ICmpInst::ICMP_NE)
8742 return ReplaceInstUsesWith(SI, TrueVal);
8743 /// NOTE: if we wanted to, this is where to detect integer MIN/MAX
8745 } else if (CmpLHS == FalseVal && CmpRHS == TrueVal) {
8746 // Transform (X == Y) ? Y : X -> X
8747 if (Pred == ICmpInst::ICMP_EQ)
8748 return ReplaceInstUsesWith(SI, FalseVal);
8749 // Transform (X != Y) ? Y : X -> Y
8750 if (Pred == ICmpInst::ICMP_NE)
8751 return ReplaceInstUsesWith(SI, TrueVal);
8752 /// NOTE: if we wanted to, this is where to detect integer MIN/MAX
8755 /// NOTE: if we wanted to, this is where to detect integer ABS
8757 return Changed ? &SI : 0;
8760 Instruction *InstCombiner::visitSelectInst(SelectInst &SI) {
8761 Value *CondVal = SI.getCondition();
8762 Value *TrueVal = SI.getTrueValue();
8763 Value *FalseVal = SI.getFalseValue();
8765 // select true, X, Y -> X
8766 // select false, X, Y -> Y
8767 if (ConstantInt *C = dyn_cast<ConstantInt>(CondVal))
8768 return ReplaceInstUsesWith(SI, C->getZExtValue() ? TrueVal : FalseVal);
8770 // select C, X, X -> X
8771 if (TrueVal == FalseVal)
8772 return ReplaceInstUsesWith(SI, TrueVal);
8774 if (isa<UndefValue>(TrueVal)) // select C, undef, X -> X
8775 return ReplaceInstUsesWith(SI, FalseVal);
8776 if (isa<UndefValue>(FalseVal)) // select C, X, undef -> X
8777 return ReplaceInstUsesWith(SI, TrueVal);
8778 if (isa<UndefValue>(CondVal)) { // select undef, X, Y -> X or Y
8779 if (isa<Constant>(TrueVal))
8780 return ReplaceInstUsesWith(SI, TrueVal);
8782 return ReplaceInstUsesWith(SI, FalseVal);
8785 if (SI.getType() == Type::Int1Ty) {
8786 if (ConstantInt *C = dyn_cast<ConstantInt>(TrueVal)) {
8787 if (C->getZExtValue()) {
8788 // Change: A = select B, true, C --> A = or B, C
8789 return BinaryOperator::CreateOr(CondVal, FalseVal);
8791 // Change: A = select B, false, C --> A = and !B, C
8793 InsertNewInstBefore(BinaryOperator::CreateNot(CondVal,
8794 "not."+CondVal->getName()), SI);
8795 return BinaryOperator::CreateAnd(NotCond, FalseVal);
8797 } else if (ConstantInt *C = dyn_cast<ConstantInt>(FalseVal)) {
8798 if (C->getZExtValue() == false) {
8799 // Change: A = select B, C, false --> A = and B, C
8800 return BinaryOperator::CreateAnd(CondVal, TrueVal);
8802 // Change: A = select B, C, true --> A = or !B, C
8804 InsertNewInstBefore(BinaryOperator::CreateNot(CondVal,
8805 "not."+CondVal->getName()), SI);
8806 return BinaryOperator::CreateOr(NotCond, TrueVal);
8810 // select a, b, a -> a&b
8811 // select a, a, b -> a|b
8812 if (CondVal == TrueVal)
8813 return BinaryOperator::CreateOr(CondVal, FalseVal);
8814 else if (CondVal == FalseVal)
8815 return BinaryOperator::CreateAnd(CondVal, TrueVal);
8818 // Selecting between two integer constants?
8819 if (ConstantInt *TrueValC = dyn_cast<ConstantInt>(TrueVal))
8820 if (ConstantInt *FalseValC = dyn_cast<ConstantInt>(FalseVal)) {
8821 // select C, 1, 0 -> zext C to int
8822 if (FalseValC->isZero() && TrueValC->getValue() == 1) {
8823 return CastInst::Create(Instruction::ZExt, CondVal, SI.getType());
8824 } else if (TrueValC->isZero() && FalseValC->getValue() == 1) {
8825 // select C, 0, 1 -> zext !C to int
8827 InsertNewInstBefore(BinaryOperator::CreateNot(CondVal,
8828 "not."+CondVal->getName()), SI);
8829 return CastInst::Create(Instruction::ZExt, NotCond, SI.getType());
8832 if (ICmpInst *IC = dyn_cast<ICmpInst>(SI.getCondition())) {
8834 // (x <s 0) ? -1 : 0 -> ashr x, 31
8835 if (TrueValC->isAllOnesValue() && FalseValC->isZero())
8836 if (ConstantInt *CmpCst = dyn_cast<ConstantInt>(IC->getOperand(1))) {
8837 if (IC->getPredicate() == ICmpInst::ICMP_SLT && CmpCst->isZero()) {
8838 // The comparison constant and the result are not neccessarily the
8839 // same width. Make an all-ones value by inserting a AShr.
8840 Value *X = IC->getOperand(0);
8841 uint32_t Bits = X->getType()->getPrimitiveSizeInBits();
8842 Constant *ShAmt = ConstantInt::get(X->getType(), Bits-1);
8843 Instruction *SRA = BinaryOperator::Create(Instruction::AShr, X,
8845 InsertNewInstBefore(SRA, SI);
8847 // Then cast to the appropriate width.
8848 return CastInst::CreateIntegerCast(SRA, SI.getType(), true);
8853 // If one of the constants is zero (we know they can't both be) and we
8854 // have an icmp instruction with zero, and we have an 'and' with the
8855 // non-constant value, eliminate this whole mess. This corresponds to
8856 // cases like this: ((X & 27) ? 27 : 0)
8857 if (TrueValC->isZero() || FalseValC->isZero())
8858 if (IC->isEquality() && isa<ConstantInt>(IC->getOperand(1)) &&
8859 cast<Constant>(IC->getOperand(1))->isNullValue())
8860 if (Instruction *ICA = dyn_cast<Instruction>(IC->getOperand(0)))
8861 if (ICA->getOpcode() == Instruction::And &&
8862 isa<ConstantInt>(ICA->getOperand(1)) &&
8863 (ICA->getOperand(1) == TrueValC ||
8864 ICA->getOperand(1) == FalseValC) &&
8865 isOneBitSet(cast<ConstantInt>(ICA->getOperand(1)))) {
8866 // Okay, now we know that everything is set up, we just don't
8867 // know whether we have a icmp_ne or icmp_eq and whether the
8868 // true or false val is the zero.
8869 bool ShouldNotVal = !TrueValC->isZero();
8870 ShouldNotVal ^= IC->getPredicate() == ICmpInst::ICMP_NE;
8873 V = InsertNewInstBefore(BinaryOperator::Create(
8874 Instruction::Xor, V, ICA->getOperand(1)), SI);
8875 return ReplaceInstUsesWith(SI, V);
8880 // See if we are selecting two values based on a comparison of the two values.
8881 if (FCmpInst *FCI = dyn_cast<FCmpInst>(CondVal)) {
8882 if (FCI->getOperand(0) == TrueVal && FCI->getOperand(1) == FalseVal) {
8883 // Transform (X == Y) ? X : Y -> Y
8884 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ) {
8885 // This is not safe in general for floating point:
8886 // consider X== -0, Y== +0.
8887 // It becomes safe if either operand is a nonzero constant.
8888 ConstantFP *CFPt, *CFPf;
8889 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) &&
8890 !CFPt->getValueAPF().isZero()) ||
8891 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) &&
8892 !CFPf->getValueAPF().isZero()))
8893 return ReplaceInstUsesWith(SI, FalseVal);
8895 // Transform (X != Y) ? X : Y -> X
8896 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
8897 return ReplaceInstUsesWith(SI, TrueVal);
8898 // NOTE: if we wanted to, this is where to detect MIN/MAX
8900 } else if (FCI->getOperand(0) == FalseVal && FCI->getOperand(1) == TrueVal){
8901 // Transform (X == Y) ? Y : X -> X
8902 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ) {
8903 // This is not safe in general for floating point:
8904 // consider X== -0, Y== +0.
8905 // It becomes safe if either operand is a nonzero constant.
8906 ConstantFP *CFPt, *CFPf;
8907 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) &&
8908 !CFPt->getValueAPF().isZero()) ||
8909 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) &&
8910 !CFPf->getValueAPF().isZero()))
8911 return ReplaceInstUsesWith(SI, FalseVal);
8913 // Transform (X != Y) ? Y : X -> Y
8914 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
8915 return ReplaceInstUsesWith(SI, TrueVal);
8916 // NOTE: if we wanted to, this is where to detect MIN/MAX
8918 // NOTE: if we wanted to, this is where to detect ABS
8921 // See if we are selecting two values based on a comparison of the two values.
8922 if (ICmpInst *ICI = dyn_cast<ICmpInst>(CondVal))
8923 if (Instruction *Result = visitSelectInstWithICmp(SI, ICI))
8926 if (Instruction *TI = dyn_cast<Instruction>(TrueVal))
8927 if (Instruction *FI = dyn_cast<Instruction>(FalseVal))
8928 if (TI->hasOneUse() && FI->hasOneUse()) {
8929 Instruction *AddOp = 0, *SubOp = 0;
8931 // Turn (select C, (op X, Y), (op X, Z)) -> (op X, (select C, Y, Z))
8932 if (TI->getOpcode() == FI->getOpcode())
8933 if (Instruction *IV = FoldSelectOpOp(SI, TI, FI))
8936 // Turn select C, (X+Y), (X-Y) --> (X+(select C, Y, (-Y))). This is
8937 // even legal for FP.
8938 if (TI->getOpcode() == Instruction::Sub &&
8939 FI->getOpcode() == Instruction::Add) {
8940 AddOp = FI; SubOp = TI;
8941 } else if (FI->getOpcode() == Instruction::Sub &&
8942 TI->getOpcode() == Instruction::Add) {
8943 AddOp = TI; SubOp = FI;
8947 Value *OtherAddOp = 0;
8948 if (SubOp->getOperand(0) == AddOp->getOperand(0)) {
8949 OtherAddOp = AddOp->getOperand(1);
8950 } else if (SubOp->getOperand(0) == AddOp->getOperand(1)) {
8951 OtherAddOp = AddOp->getOperand(0);
8955 // So at this point we know we have (Y -> OtherAddOp):
8956 // select C, (add X, Y), (sub X, Z)
8957 Value *NegVal; // Compute -Z
8958 if (Constant *C = dyn_cast<Constant>(SubOp->getOperand(1))) {
8959 NegVal = ConstantExpr::getNeg(C);
8961 NegVal = InsertNewInstBefore(
8962 BinaryOperator::CreateNeg(SubOp->getOperand(1), "tmp"), SI);
8965 Value *NewTrueOp = OtherAddOp;
8966 Value *NewFalseOp = NegVal;
8968 std::swap(NewTrueOp, NewFalseOp);
8969 Instruction *NewSel =
8970 SelectInst::Create(CondVal, NewTrueOp,
8971 NewFalseOp, SI.getName() + ".p");
8973 NewSel = InsertNewInstBefore(NewSel, SI);
8974 return BinaryOperator::CreateAdd(SubOp->getOperand(0), NewSel);
8979 // See if we can fold the select into one of our operands.
8980 if (SI.getType()->isInteger()) {
8981 // See the comment above GetSelectFoldableOperands for a description of the
8982 // transformation we are doing here.
8983 if (Instruction *TVI = dyn_cast<Instruction>(TrueVal))
8984 if (TVI->hasOneUse() && TVI->getNumOperands() == 2 &&
8985 !isa<Constant>(FalseVal))
8986 if (unsigned SFO = GetSelectFoldableOperands(TVI)) {
8987 unsigned OpToFold = 0;
8988 if ((SFO & 1) && FalseVal == TVI->getOperand(0)) {
8990 } else if ((SFO & 2) && FalseVal == TVI->getOperand(1)) {
8995 Constant *C = GetSelectFoldableConstant(TVI);
8996 Instruction *NewSel =
8997 SelectInst::Create(SI.getCondition(),
8998 TVI->getOperand(2-OpToFold), C);
8999 InsertNewInstBefore(NewSel, SI);
9000 NewSel->takeName(TVI);
9001 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TVI))
9002 return BinaryOperator::Create(BO->getOpcode(), FalseVal, NewSel);
9004 assert(0 && "Unknown instruction!!");
9009 if (Instruction *FVI = dyn_cast<Instruction>(FalseVal))
9010 if (FVI->hasOneUse() && FVI->getNumOperands() == 2 &&
9011 !isa<Constant>(TrueVal))
9012 if (unsigned SFO = GetSelectFoldableOperands(FVI)) {
9013 unsigned OpToFold = 0;
9014 if ((SFO & 1) && TrueVal == FVI->getOperand(0)) {
9016 } else if ((SFO & 2) && TrueVal == FVI->getOperand(1)) {
9021 Constant *C = GetSelectFoldableConstant(FVI);
9022 Instruction *NewSel =
9023 SelectInst::Create(SI.getCondition(), C,
9024 FVI->getOperand(2-OpToFold));
9025 InsertNewInstBefore(NewSel, SI);
9026 NewSel->takeName(FVI);
9027 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(FVI))
9028 return BinaryOperator::Create(BO->getOpcode(), TrueVal, NewSel);
9030 assert(0 && "Unknown instruction!!");
9035 if (BinaryOperator::isNot(CondVal)) {
9036 SI.setOperand(0, BinaryOperator::getNotArgument(CondVal));
9037 SI.setOperand(1, FalseVal);
9038 SI.setOperand(2, TrueVal);
9045 /// EnforceKnownAlignment - If the specified pointer points to an object that
9046 /// we control, modify the object's alignment to PrefAlign. This isn't
9047 /// often possible though. If alignment is important, a more reliable approach
9048 /// is to simply align all global variables and allocation instructions to
9049 /// their preferred alignment from the beginning.
9051 static unsigned EnforceKnownAlignment(Value *V,
9052 unsigned Align, unsigned PrefAlign) {
9054 User *U = dyn_cast<User>(V);
9055 if (!U) return Align;
9057 switch (getOpcode(U)) {
9059 case Instruction::BitCast:
9060 return EnforceKnownAlignment(U->getOperand(0), Align, PrefAlign);
9061 case Instruction::GetElementPtr: {
9062 // If all indexes are zero, it is just the alignment of the base pointer.
9063 bool AllZeroOperands = true;
9064 for (User::op_iterator i = U->op_begin() + 1, e = U->op_end(); i != e; ++i)
9065 if (!isa<Constant>(*i) ||
9066 !cast<Constant>(*i)->isNullValue()) {
9067 AllZeroOperands = false;
9071 if (AllZeroOperands) {
9072 // Treat this like a bitcast.
9073 return EnforceKnownAlignment(U->getOperand(0), Align, PrefAlign);
9079 if (GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
9080 // If there is a large requested alignment and we can, bump up the alignment
9082 if (!GV->isDeclaration()) {
9083 GV->setAlignment(PrefAlign);
9086 } else if (AllocationInst *AI = dyn_cast<AllocationInst>(V)) {
9087 // If there is a requested alignment and if this is an alloca, round up. We
9088 // don't do this for malloc, because some systems can't respect the request.
9089 if (isa<AllocaInst>(AI)) {
9090 AI->setAlignment(PrefAlign);
9098 /// GetOrEnforceKnownAlignment - If the specified pointer has an alignment that
9099 /// we can determine, return it, otherwise return 0. If PrefAlign is specified,
9100 /// and it is more than the alignment of the ultimate object, see if we can
9101 /// increase the alignment of the ultimate object, making this check succeed.
9102 unsigned InstCombiner::GetOrEnforceKnownAlignment(Value *V,
9103 unsigned PrefAlign) {
9104 unsigned BitWidth = TD ? TD->getTypeSizeInBits(V->getType()) :
9105 sizeof(PrefAlign) * CHAR_BIT;
9106 APInt Mask = APInt::getAllOnesValue(BitWidth);
9107 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
9108 ComputeMaskedBits(V, Mask, KnownZero, KnownOne);
9109 unsigned TrailZ = KnownZero.countTrailingOnes();
9110 unsigned Align = 1u << std::min(BitWidth - 1, TrailZ);
9112 if (PrefAlign > Align)
9113 Align = EnforceKnownAlignment(V, Align, PrefAlign);
9115 // We don't need to make any adjustment.
9119 Instruction *InstCombiner::SimplifyMemTransfer(MemIntrinsic *MI) {
9120 unsigned DstAlign = GetOrEnforceKnownAlignment(MI->getOperand(1));
9121 unsigned SrcAlign = GetOrEnforceKnownAlignment(MI->getOperand(2));
9122 unsigned MinAlign = std::min(DstAlign, SrcAlign);
9123 unsigned CopyAlign = MI->getAlignment()->getZExtValue();
9125 if (CopyAlign < MinAlign) {
9126 MI->setAlignment(ConstantInt::get(Type::Int32Ty, MinAlign));
9130 // If MemCpyInst length is 1/2/4/8 bytes then replace memcpy with
9132 ConstantInt *MemOpLength = dyn_cast<ConstantInt>(MI->getOperand(3));
9133 if (MemOpLength == 0) return 0;
9135 // Source and destination pointer types are always "i8*" for intrinsic. See
9136 // if the size is something we can handle with a single primitive load/store.
9137 // A single load+store correctly handles overlapping memory in the memmove
9139 unsigned Size = MemOpLength->getZExtValue();
9140 if (Size == 0) return MI; // Delete this mem transfer.
9142 if (Size > 8 || (Size&(Size-1)))
9143 return 0; // If not 1/2/4/8 bytes, exit.
9145 // Use an integer load+store unless we can find something better.
9146 Type *NewPtrTy = PointerType::getUnqual(IntegerType::get(Size<<3));
9148 // Memcpy forces the use of i8* for the source and destination. That means
9149 // that if you're using memcpy to move one double around, you'll get a cast
9150 // from double* to i8*. We'd much rather use a double load+store rather than
9151 // an i64 load+store, here because this improves the odds that the source or
9152 // dest address will be promotable. See if we can find a better type than the
9153 // integer datatype.
9154 if (Value *Op = getBitCastOperand(MI->getOperand(1))) {
9155 const Type *SrcETy = cast<PointerType>(Op->getType())->getElementType();
9156 if (SrcETy->isSized() && TD->getTypeStoreSize(SrcETy) == Size) {
9157 // The SrcETy might be something like {{{double}}} or [1 x double]. Rip
9158 // down through these levels if so.
9159 while (!SrcETy->isSingleValueType()) {
9160 if (const StructType *STy = dyn_cast<StructType>(SrcETy)) {
9161 if (STy->getNumElements() == 1)
9162 SrcETy = STy->getElementType(0);
9165 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(SrcETy)) {
9166 if (ATy->getNumElements() == 1)
9167 SrcETy = ATy->getElementType();
9174 if (SrcETy->isSingleValueType())
9175 NewPtrTy = PointerType::getUnqual(SrcETy);
9180 // If the memcpy/memmove provides better alignment info than we can
9182 SrcAlign = std::max(SrcAlign, CopyAlign);
9183 DstAlign = std::max(DstAlign, CopyAlign);
9185 Value *Src = InsertBitCastBefore(MI->getOperand(2), NewPtrTy, *MI);
9186 Value *Dest = InsertBitCastBefore(MI->getOperand(1), NewPtrTy, *MI);
9187 Instruction *L = new LoadInst(Src, "tmp", false, SrcAlign);
9188 InsertNewInstBefore(L, *MI);
9189 InsertNewInstBefore(new StoreInst(L, Dest, false, DstAlign), *MI);
9191 // Set the size of the copy to 0, it will be deleted on the next iteration.
9192 MI->setOperand(3, Constant::getNullValue(MemOpLength->getType()));
9196 Instruction *InstCombiner::SimplifyMemSet(MemSetInst *MI) {
9197 unsigned Alignment = GetOrEnforceKnownAlignment(MI->getDest());
9198 if (MI->getAlignment()->getZExtValue() < Alignment) {
9199 MI->setAlignment(ConstantInt::get(Type::Int32Ty, Alignment));
9203 // Extract the length and alignment and fill if they are constant.
9204 ConstantInt *LenC = dyn_cast<ConstantInt>(MI->getLength());
9205 ConstantInt *FillC = dyn_cast<ConstantInt>(MI->getValue());
9206 if (!LenC || !FillC || FillC->getType() != Type::Int8Ty)
9208 uint64_t Len = LenC->getZExtValue();
9209 Alignment = MI->getAlignment()->getZExtValue();
9211 // If the length is zero, this is a no-op
9212 if (Len == 0) return MI; // memset(d,c,0,a) -> noop
9214 // memset(s,c,n) -> store s, c (for n=1,2,4,8)
9215 if (Len <= 8 && isPowerOf2_32((uint32_t)Len)) {
9216 const Type *ITy = IntegerType::get(Len*8); // n=1 -> i8.
9218 Value *Dest = MI->getDest();
9219 Dest = InsertBitCastBefore(Dest, PointerType::getUnqual(ITy), *MI);
9221 // Alignment 0 is identity for alignment 1 for memset, but not store.
9222 if (Alignment == 0) Alignment = 1;
9224 // Extract the fill value and store.
9225 uint64_t Fill = FillC->getZExtValue()*0x0101010101010101ULL;
9226 InsertNewInstBefore(new StoreInst(ConstantInt::get(ITy, Fill), Dest, false,
9229 // Set the size of the copy to 0, it will be deleted on the next iteration.
9230 MI->setLength(Constant::getNullValue(LenC->getType()));
9238 /// visitCallInst - CallInst simplification. This mostly only handles folding
9239 /// of intrinsic instructions. For normal calls, it allows visitCallSite to do
9240 /// the heavy lifting.
9242 Instruction *InstCombiner::visitCallInst(CallInst &CI) {
9243 IntrinsicInst *II = dyn_cast<IntrinsicInst>(&CI);
9244 if (!II) return visitCallSite(&CI);
9246 // Intrinsics cannot occur in an invoke, so handle them here instead of in
9248 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(II)) {
9249 bool Changed = false;
9251 // memmove/cpy/set of zero bytes is a noop.
9252 if (Constant *NumBytes = dyn_cast<Constant>(MI->getLength())) {
9253 if (NumBytes->isNullValue()) return EraseInstFromFunction(CI);
9255 if (ConstantInt *CI = dyn_cast<ConstantInt>(NumBytes))
9256 if (CI->getZExtValue() == 1) {
9257 // Replace the instruction with just byte operations. We would
9258 // transform other cases to loads/stores, but we don't know if
9259 // alignment is sufficient.
9263 // If we have a memmove and the source operation is a constant global,
9264 // then the source and dest pointers can't alias, so we can change this
9265 // into a call to memcpy.
9266 if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(MI)) {
9267 if (GlobalVariable *GVSrc = dyn_cast<GlobalVariable>(MMI->getSource()))
9268 if (GVSrc->isConstant()) {
9269 Module *M = CI.getParent()->getParent()->getParent();
9270 Intrinsic::ID MemCpyID = Intrinsic::memcpy;
9272 Tys[0] = CI.getOperand(3)->getType();
9274 Intrinsic::getDeclaration(M, MemCpyID, Tys, 1));
9278 // memmove(x,x,size) -> noop.
9279 if (MMI->getSource() == MMI->getDest())
9280 return EraseInstFromFunction(CI);
9283 // If we can determine a pointer alignment that is bigger than currently
9284 // set, update the alignment.
9285 if (isa<MemCpyInst>(MI) || isa<MemMoveInst>(MI)) {
9286 if (Instruction *I = SimplifyMemTransfer(MI))
9288 } else if (MemSetInst *MSI = dyn_cast<MemSetInst>(MI)) {
9289 if (Instruction *I = SimplifyMemSet(MSI))
9293 if (Changed) return II;
9296 switch (II->getIntrinsicID()) {
9298 case Intrinsic::bswap:
9299 // bswap(bswap(x)) -> x
9300 if (IntrinsicInst *Operand = dyn_cast<IntrinsicInst>(II->getOperand(1)))
9301 if (Operand->getIntrinsicID() == Intrinsic::bswap)
9302 return ReplaceInstUsesWith(CI, Operand->getOperand(1));
9304 case Intrinsic::ppc_altivec_lvx:
9305 case Intrinsic::ppc_altivec_lvxl:
9306 case Intrinsic::x86_sse_loadu_ps:
9307 case Intrinsic::x86_sse2_loadu_pd:
9308 case Intrinsic::x86_sse2_loadu_dq:
9309 // Turn PPC lvx -> load if the pointer is known aligned.
9310 // Turn X86 loadups -> load if the pointer is known aligned.
9311 if (GetOrEnforceKnownAlignment(II->getOperand(1), 16) >= 16) {
9312 Value *Ptr = InsertBitCastBefore(II->getOperand(1),
9313 PointerType::getUnqual(II->getType()),
9315 return new LoadInst(Ptr);
9318 case Intrinsic::ppc_altivec_stvx:
9319 case Intrinsic::ppc_altivec_stvxl:
9320 // Turn stvx -> store if the pointer is known aligned.
9321 if (GetOrEnforceKnownAlignment(II->getOperand(2), 16) >= 16) {
9322 const Type *OpPtrTy =
9323 PointerType::getUnqual(II->getOperand(1)->getType());
9324 Value *Ptr = InsertBitCastBefore(II->getOperand(2), OpPtrTy, CI);
9325 return new StoreInst(II->getOperand(1), Ptr);
9328 case Intrinsic::x86_sse_storeu_ps:
9329 case Intrinsic::x86_sse2_storeu_pd:
9330 case Intrinsic::x86_sse2_storeu_dq:
9331 // Turn X86 storeu -> store if the pointer is known aligned.
9332 if (GetOrEnforceKnownAlignment(II->getOperand(1), 16) >= 16) {
9333 const Type *OpPtrTy =
9334 PointerType::getUnqual(II->getOperand(2)->getType());
9335 Value *Ptr = InsertBitCastBefore(II->getOperand(1), OpPtrTy, CI);
9336 return new StoreInst(II->getOperand(2), Ptr);
9340 case Intrinsic::x86_sse_cvttss2si: {
9341 // These intrinsics only demands the 0th element of its input vector. If
9342 // we can simplify the input based on that, do so now.
9344 if (Value *V = SimplifyDemandedVectorElts(II->getOperand(1), 1,
9346 II->setOperand(1, V);
9352 case Intrinsic::ppc_altivec_vperm:
9353 // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant.
9354 if (ConstantVector *Mask = dyn_cast<ConstantVector>(II->getOperand(3))) {
9355 assert(Mask->getNumOperands() == 16 && "Bad type for intrinsic!");
9357 // Check that all of the elements are integer constants or undefs.
9358 bool AllEltsOk = true;
9359 for (unsigned i = 0; i != 16; ++i) {
9360 if (!isa<ConstantInt>(Mask->getOperand(i)) &&
9361 !isa<UndefValue>(Mask->getOperand(i))) {
9368 // Cast the input vectors to byte vectors.
9369 Value *Op0 =InsertBitCastBefore(II->getOperand(1),Mask->getType(),CI);
9370 Value *Op1 =InsertBitCastBefore(II->getOperand(2),Mask->getType(),CI);
9371 Value *Result = UndefValue::get(Op0->getType());
9373 // Only extract each element once.
9374 Value *ExtractedElts[32];
9375 memset(ExtractedElts, 0, sizeof(ExtractedElts));
9377 for (unsigned i = 0; i != 16; ++i) {
9378 if (isa<UndefValue>(Mask->getOperand(i)))
9380 unsigned Idx=cast<ConstantInt>(Mask->getOperand(i))->getZExtValue();
9381 Idx &= 31; // Match the hardware behavior.
9383 if (ExtractedElts[Idx] == 0) {
9385 new ExtractElementInst(Idx < 16 ? Op0 : Op1, Idx&15, "tmp");
9386 InsertNewInstBefore(Elt, CI);
9387 ExtractedElts[Idx] = Elt;
9390 // Insert this value into the result vector.
9391 Result = InsertElementInst::Create(Result, ExtractedElts[Idx],
9393 InsertNewInstBefore(cast<Instruction>(Result), CI);
9395 return CastInst::Create(Instruction::BitCast, Result, CI.getType());
9400 case Intrinsic::stackrestore: {
9401 // If the save is right next to the restore, remove the restore. This can
9402 // happen when variable allocas are DCE'd.
9403 if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getOperand(1))) {
9404 if (SS->getIntrinsicID() == Intrinsic::stacksave) {
9405 BasicBlock::iterator BI = SS;
9407 return EraseInstFromFunction(CI);
9411 // Scan down this block to see if there is another stack restore in the
9412 // same block without an intervening call/alloca.
9413 BasicBlock::iterator BI = II;
9414 TerminatorInst *TI = II->getParent()->getTerminator();
9415 bool CannotRemove = false;
9416 for (++BI; &*BI != TI; ++BI) {
9417 if (isa<AllocaInst>(BI)) {
9418 CannotRemove = true;
9421 if (CallInst *BCI = dyn_cast<CallInst>(BI)) {
9422 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(BCI)) {
9423 // If there is a stackrestore below this one, remove this one.
9424 if (II->getIntrinsicID() == Intrinsic::stackrestore)
9425 return EraseInstFromFunction(CI);
9426 // Otherwise, ignore the intrinsic.
9428 // If we found a non-intrinsic call, we can't remove the stack
9430 CannotRemove = true;
9436 // If the stack restore is in a return/unwind block and if there are no
9437 // allocas or calls between the restore and the return, nuke the restore.
9438 if (!CannotRemove && (isa<ReturnInst>(TI) || isa<UnwindInst>(TI)))
9439 return EraseInstFromFunction(CI);
9444 return visitCallSite(II);
9447 // InvokeInst simplification
9449 Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) {
9450 return visitCallSite(&II);
9453 /// isSafeToEliminateVarargsCast - If this cast does not affect the value
9454 /// passed through the varargs area, we can eliminate the use of the cast.
9455 static bool isSafeToEliminateVarargsCast(const CallSite CS,
9456 const CastInst * const CI,
9457 const TargetData * const TD,
9459 if (!CI->isLosslessCast())
9462 // The size of ByVal arguments is derived from the type, so we
9463 // can't change to a type with a different size. If the size were
9464 // passed explicitly we could avoid this check.
9465 if (!CS.paramHasAttr(ix, Attribute::ByVal))
9469 cast<PointerType>(CI->getOperand(0)->getType())->getElementType();
9470 const Type* DstTy = cast<PointerType>(CI->getType())->getElementType();
9471 if (!SrcTy->isSized() || !DstTy->isSized())
9473 if (TD->getABITypeSize(SrcTy) != TD->getABITypeSize(DstTy))
9478 // visitCallSite - Improvements for call and invoke instructions.
9480 Instruction *InstCombiner::visitCallSite(CallSite CS) {
9481 bool Changed = false;
9483 // If the callee is a constexpr cast of a function, attempt to move the cast
9484 // to the arguments of the call/invoke.
9485 if (transformConstExprCastCall(CS)) return 0;
9487 Value *Callee = CS.getCalledValue();
9489 if (Function *CalleeF = dyn_cast<Function>(Callee))
9490 if (CalleeF->getCallingConv() != CS.getCallingConv()) {
9491 Instruction *OldCall = CS.getInstruction();
9492 // If the call and callee calling conventions don't match, this call must
9493 // be unreachable, as the call is undefined.
9494 new StoreInst(ConstantInt::getTrue(),
9495 UndefValue::get(PointerType::getUnqual(Type::Int1Ty)),
9497 if (!OldCall->use_empty())
9498 OldCall->replaceAllUsesWith(UndefValue::get(OldCall->getType()));
9499 if (isa<CallInst>(OldCall)) // Not worth removing an invoke here.
9500 return EraseInstFromFunction(*OldCall);
9504 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) {
9505 // This instruction is not reachable, just remove it. We insert a store to
9506 // undef so that we know that this code is not reachable, despite the fact
9507 // that we can't modify the CFG here.
9508 new StoreInst(ConstantInt::getTrue(),
9509 UndefValue::get(PointerType::getUnqual(Type::Int1Ty)),
9510 CS.getInstruction());
9512 if (!CS.getInstruction()->use_empty())
9513 CS.getInstruction()->
9514 replaceAllUsesWith(UndefValue::get(CS.getInstruction()->getType()));
9516 if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) {
9517 // Don't break the CFG, insert a dummy cond branch.
9518 BranchInst::Create(II->getNormalDest(), II->getUnwindDest(),
9519 ConstantInt::getTrue(), II);
9521 return EraseInstFromFunction(*CS.getInstruction());
9524 if (BitCastInst *BC = dyn_cast<BitCastInst>(Callee))
9525 if (IntrinsicInst *In = dyn_cast<IntrinsicInst>(BC->getOperand(0)))
9526 if (In->getIntrinsicID() == Intrinsic::init_trampoline)
9527 return transformCallThroughTrampoline(CS);
9529 const PointerType *PTy = cast<PointerType>(Callee->getType());
9530 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
9531 if (FTy->isVarArg()) {
9532 int ix = FTy->getNumParams() + (isa<InvokeInst>(Callee) ? 3 : 1);
9533 // See if we can optimize any arguments passed through the varargs area of
9535 for (CallSite::arg_iterator I = CS.arg_begin()+FTy->getNumParams(),
9536 E = CS.arg_end(); I != E; ++I, ++ix) {
9537 CastInst *CI = dyn_cast<CastInst>(*I);
9538 if (CI && isSafeToEliminateVarargsCast(CS, CI, TD, ix)) {
9539 *I = CI->getOperand(0);
9545 if (isa<InlineAsm>(Callee) && !CS.doesNotThrow()) {
9546 // Inline asm calls cannot throw - mark them 'nounwind'.
9547 CS.setDoesNotThrow();
9551 return Changed ? CS.getInstruction() : 0;
9554 // transformConstExprCastCall - If the callee is a constexpr cast of a function,
9555 // attempt to move the cast to the arguments of the call/invoke.
9557 bool InstCombiner::transformConstExprCastCall(CallSite CS) {
9558 if (!isa<ConstantExpr>(CS.getCalledValue())) return false;
9559 ConstantExpr *CE = cast<ConstantExpr>(CS.getCalledValue());
9560 if (CE->getOpcode() != Instruction::BitCast ||
9561 !isa<Function>(CE->getOperand(0)))
9563 Function *Callee = cast<Function>(CE->getOperand(0));
9564 Instruction *Caller = CS.getInstruction();
9565 const AttrListPtr &CallerPAL = CS.getAttributes();
9567 // Okay, this is a cast from a function to a different type. Unless doing so
9568 // would cause a type conversion of one of our arguments, change this call to
9569 // be a direct call with arguments casted to the appropriate types.
9571 const FunctionType *FT = Callee->getFunctionType();
9572 const Type *OldRetTy = Caller->getType();
9573 const Type *NewRetTy = FT->getReturnType();
9575 if (isa<StructType>(NewRetTy))
9576 return false; // TODO: Handle multiple return values.
9578 // Check to see if we are changing the return type...
9579 if (OldRetTy != NewRetTy) {
9580 if (Callee->isDeclaration() &&
9581 // Conversion is ok if changing from one pointer type to another or from
9582 // a pointer to an integer of the same size.
9583 !((isa<PointerType>(OldRetTy) || OldRetTy == TD->getIntPtrType()) &&
9584 (isa<PointerType>(NewRetTy) || NewRetTy == TD->getIntPtrType())))
9585 return false; // Cannot transform this return value.
9587 if (!Caller->use_empty() &&
9588 // void -> non-void is handled specially
9589 NewRetTy != Type::VoidTy && !CastInst::isCastable(NewRetTy, OldRetTy))
9590 return false; // Cannot transform this return value.
9592 if (!CallerPAL.isEmpty() && !Caller->use_empty()) {
9593 Attributes RAttrs = CallerPAL.getRetAttributes();
9594 if (RAttrs & Attribute::typeIncompatible(NewRetTy))
9595 return false; // Attribute not compatible with transformed value.
9598 // If the callsite is an invoke instruction, and the return value is used by
9599 // a PHI node in a successor, we cannot change the return type of the call
9600 // because there is no place to put the cast instruction (without breaking
9601 // the critical edge). Bail out in this case.
9602 if (!Caller->use_empty())
9603 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller))
9604 for (Value::use_iterator UI = II->use_begin(), E = II->use_end();
9606 if (PHINode *PN = dyn_cast<PHINode>(*UI))
9607 if (PN->getParent() == II->getNormalDest() ||
9608 PN->getParent() == II->getUnwindDest())
9612 unsigned NumActualArgs = unsigned(CS.arg_end()-CS.arg_begin());
9613 unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs);
9615 CallSite::arg_iterator AI = CS.arg_begin();
9616 for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) {
9617 const Type *ParamTy = FT->getParamType(i);
9618 const Type *ActTy = (*AI)->getType();
9620 if (!CastInst::isCastable(ActTy, ParamTy))
9621 return false; // Cannot transform this parameter value.
9623 if (CallerPAL.getParamAttributes(i + 1)
9624 & Attribute::typeIncompatible(ParamTy))
9625 return false; // Attribute not compatible with transformed value.
9627 // Converting from one pointer type to another or between a pointer and an
9628 // integer of the same size is safe even if we do not have a body.
9629 bool isConvertible = ActTy == ParamTy ||
9630 ((isa<PointerType>(ParamTy) || ParamTy == TD->getIntPtrType()) &&
9631 (isa<PointerType>(ActTy) || ActTy == TD->getIntPtrType()));
9632 if (Callee->isDeclaration() && !isConvertible) return false;
9635 if (FT->getNumParams() < NumActualArgs && !FT->isVarArg() &&
9636 Callee->isDeclaration())
9637 return false; // Do not delete arguments unless we have a function body.
9639 if (FT->getNumParams() < NumActualArgs && FT->isVarArg() &&
9640 !CallerPAL.isEmpty())
9641 // In this case we have more arguments than the new function type, but we
9642 // won't be dropping them. Check that these extra arguments have attributes
9643 // that are compatible with being a vararg call argument.
9644 for (unsigned i = CallerPAL.getNumSlots(); i; --i) {
9645 if (CallerPAL.getSlot(i - 1).Index <= FT->getNumParams())
9647 Attributes PAttrs = CallerPAL.getSlot(i - 1).Attrs;
9648 if (PAttrs & Attribute::VarArgsIncompatible)
9652 // Okay, we decided that this is a safe thing to do: go ahead and start
9653 // inserting cast instructions as necessary...
9654 std::vector<Value*> Args;
9655 Args.reserve(NumActualArgs);
9656 SmallVector<AttributeWithIndex, 8> attrVec;
9657 attrVec.reserve(NumCommonArgs);
9659 // Get any return attributes.
9660 Attributes RAttrs = CallerPAL.getRetAttributes();
9662 // If the return value is not being used, the type may not be compatible
9663 // with the existing attributes. Wipe out any problematic attributes.
9664 RAttrs &= ~Attribute::typeIncompatible(NewRetTy);
9666 // Add the new return attributes.
9668 attrVec.push_back(AttributeWithIndex::get(0, RAttrs));
9670 AI = CS.arg_begin();
9671 for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) {
9672 const Type *ParamTy = FT->getParamType(i);
9673 if ((*AI)->getType() == ParamTy) {
9674 Args.push_back(*AI);
9676 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI,
9677 false, ParamTy, false);
9678 CastInst *NewCast = CastInst::Create(opcode, *AI, ParamTy, "tmp");
9679 Args.push_back(InsertNewInstBefore(NewCast, *Caller));
9682 // Add any parameter attributes.
9683 if (Attributes PAttrs = CallerPAL.getParamAttributes(i + 1))
9684 attrVec.push_back(AttributeWithIndex::get(i + 1, PAttrs));
9687 // If the function takes more arguments than the call was taking, add them
9689 for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i)
9690 Args.push_back(Constant::getNullValue(FT->getParamType(i)));
9692 // If we are removing arguments to the function, emit an obnoxious warning...
9693 if (FT->getNumParams() < NumActualArgs) {
9694 if (!FT->isVarArg()) {
9695 cerr << "WARNING: While resolving call to function '"
9696 << Callee->getName() << "' arguments were dropped!\n";
9698 // Add all of the arguments in their promoted form to the arg list...
9699 for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) {
9700 const Type *PTy = getPromotedType((*AI)->getType());
9701 if (PTy != (*AI)->getType()) {
9702 // Must promote to pass through va_arg area!
9703 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI, false,
9705 Instruction *Cast = CastInst::Create(opcode, *AI, PTy, "tmp");
9706 InsertNewInstBefore(Cast, *Caller);
9707 Args.push_back(Cast);
9709 Args.push_back(*AI);
9712 // Add any parameter attributes.
9713 if (Attributes PAttrs = CallerPAL.getParamAttributes(i + 1))
9714 attrVec.push_back(AttributeWithIndex::get(i + 1, PAttrs));
9719 if (Attributes FnAttrs = CallerPAL.getFnAttributes())
9720 attrVec.push_back(AttributeWithIndex::get(~0, FnAttrs));
9722 if (NewRetTy == Type::VoidTy)
9723 Caller->setName(""); // Void type should not have a name.
9725 const AttrListPtr &NewCallerPAL = AttrListPtr::get(attrVec.begin(),attrVec.end());
9728 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
9729 NC = InvokeInst::Create(Callee, II->getNormalDest(), II->getUnwindDest(),
9730 Args.begin(), Args.end(),
9731 Caller->getName(), Caller);
9732 cast<InvokeInst>(NC)->setCallingConv(II->getCallingConv());
9733 cast<InvokeInst>(NC)->setAttributes(NewCallerPAL);
9735 NC = CallInst::Create(Callee, Args.begin(), Args.end(),
9736 Caller->getName(), Caller);
9737 CallInst *CI = cast<CallInst>(Caller);
9738 if (CI->isTailCall())
9739 cast<CallInst>(NC)->setTailCall();
9740 cast<CallInst>(NC)->setCallingConv(CI->getCallingConv());
9741 cast<CallInst>(NC)->setAttributes(NewCallerPAL);
9744 // Insert a cast of the return type as necessary.
9746 if (OldRetTy != NV->getType() && !Caller->use_empty()) {
9747 if (NV->getType() != Type::VoidTy) {
9748 Instruction::CastOps opcode = CastInst::getCastOpcode(NC, false,
9750 NV = NC = CastInst::Create(opcode, NC, OldRetTy, "tmp");
9752 // If this is an invoke instruction, we should insert it after the first
9753 // non-phi, instruction in the normal successor block.
9754 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
9755 BasicBlock::iterator I = II->getNormalDest()->getFirstNonPHI();
9756 InsertNewInstBefore(NC, *I);
9758 // Otherwise, it's a call, just insert cast right after the call instr
9759 InsertNewInstBefore(NC, *Caller);
9761 AddUsersToWorkList(*Caller);
9763 NV = UndefValue::get(Caller->getType());
9767 if (Caller->getType() != Type::VoidTy && !Caller->use_empty())
9768 Caller->replaceAllUsesWith(NV);
9769 Caller->eraseFromParent();
9770 RemoveFromWorkList(Caller);
9774 // transformCallThroughTrampoline - Turn a call to a function created by the
9775 // init_trampoline intrinsic into a direct call to the underlying function.
9777 Instruction *InstCombiner::transformCallThroughTrampoline(CallSite CS) {
9778 Value *Callee = CS.getCalledValue();
9779 const PointerType *PTy = cast<PointerType>(Callee->getType());
9780 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
9781 const AttrListPtr &Attrs = CS.getAttributes();
9783 // If the call already has the 'nest' attribute somewhere then give up -
9784 // otherwise 'nest' would occur twice after splicing in the chain.
9785 if (Attrs.hasAttrSomewhere(Attribute::Nest))
9788 IntrinsicInst *Tramp =
9789 cast<IntrinsicInst>(cast<BitCastInst>(Callee)->getOperand(0));
9791 Function *NestF = cast<Function>(Tramp->getOperand(2)->stripPointerCasts());
9792 const PointerType *NestFPTy = cast<PointerType>(NestF->getType());
9793 const FunctionType *NestFTy = cast<FunctionType>(NestFPTy->getElementType());
9795 const AttrListPtr &NestAttrs = NestF->getAttributes();
9796 if (!NestAttrs.isEmpty()) {
9797 unsigned NestIdx = 1;
9798 const Type *NestTy = 0;
9799 Attributes NestAttr = Attribute::None;
9801 // Look for a parameter marked with the 'nest' attribute.
9802 for (FunctionType::param_iterator I = NestFTy->param_begin(),
9803 E = NestFTy->param_end(); I != E; ++NestIdx, ++I)
9804 if (NestAttrs.paramHasAttr(NestIdx, Attribute::Nest)) {
9805 // Record the parameter type and any other attributes.
9807 NestAttr = NestAttrs.getParamAttributes(NestIdx);
9812 Instruction *Caller = CS.getInstruction();
9813 std::vector<Value*> NewArgs;
9814 NewArgs.reserve(unsigned(CS.arg_end()-CS.arg_begin())+1);
9816 SmallVector<AttributeWithIndex, 8> NewAttrs;
9817 NewAttrs.reserve(Attrs.getNumSlots() + 1);
9819 // Insert the nest argument into the call argument list, which may
9820 // mean appending it. Likewise for attributes.
9822 // Add any result attributes.
9823 if (Attributes Attr = Attrs.getRetAttributes())
9824 NewAttrs.push_back(AttributeWithIndex::get(0, Attr));
9828 CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
9830 if (Idx == NestIdx) {
9831 // Add the chain argument and attributes.
9832 Value *NestVal = Tramp->getOperand(3);
9833 if (NestVal->getType() != NestTy)
9834 NestVal = new BitCastInst(NestVal, NestTy, "nest", Caller);
9835 NewArgs.push_back(NestVal);
9836 NewAttrs.push_back(AttributeWithIndex::get(NestIdx, NestAttr));
9842 // Add the original argument and attributes.
9843 NewArgs.push_back(*I);
9844 if (Attributes Attr = Attrs.getParamAttributes(Idx))
9846 (AttributeWithIndex::get(Idx + (Idx >= NestIdx), Attr));
9852 // Add any function attributes.
9853 if (Attributes Attr = Attrs.getFnAttributes())
9854 NewAttrs.push_back(AttributeWithIndex::get(~0, Attr));
9856 // The trampoline may have been bitcast to a bogus type (FTy).
9857 // Handle this by synthesizing a new function type, equal to FTy
9858 // with the chain parameter inserted.
9860 std::vector<const Type*> NewTypes;
9861 NewTypes.reserve(FTy->getNumParams()+1);
9863 // Insert the chain's type into the list of parameter types, which may
9864 // mean appending it.
9867 FunctionType::param_iterator I = FTy->param_begin(),
9868 E = FTy->param_end();
9872 // Add the chain's type.
9873 NewTypes.push_back(NestTy);
9878 // Add the original type.
9879 NewTypes.push_back(*I);
9885 // Replace the trampoline call with a direct call. Let the generic
9886 // code sort out any function type mismatches.
9887 FunctionType *NewFTy =
9888 FunctionType::get(FTy->getReturnType(), NewTypes, FTy->isVarArg());
9889 Constant *NewCallee = NestF->getType() == PointerType::getUnqual(NewFTy) ?
9890 NestF : ConstantExpr::getBitCast(NestF, PointerType::getUnqual(NewFTy));
9891 const AttrListPtr &NewPAL = AttrListPtr::get(NewAttrs.begin(),NewAttrs.end());
9893 Instruction *NewCaller;
9894 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
9895 NewCaller = InvokeInst::Create(NewCallee,
9896 II->getNormalDest(), II->getUnwindDest(),
9897 NewArgs.begin(), NewArgs.end(),
9898 Caller->getName(), Caller);
9899 cast<InvokeInst>(NewCaller)->setCallingConv(II->getCallingConv());
9900 cast<InvokeInst>(NewCaller)->setAttributes(NewPAL);
9902 NewCaller = CallInst::Create(NewCallee, NewArgs.begin(), NewArgs.end(),
9903 Caller->getName(), Caller);
9904 if (cast<CallInst>(Caller)->isTailCall())
9905 cast<CallInst>(NewCaller)->setTailCall();
9906 cast<CallInst>(NewCaller)->
9907 setCallingConv(cast<CallInst>(Caller)->getCallingConv());
9908 cast<CallInst>(NewCaller)->setAttributes(NewPAL);
9910 if (Caller->getType() != Type::VoidTy && !Caller->use_empty())
9911 Caller->replaceAllUsesWith(NewCaller);
9912 Caller->eraseFromParent();
9913 RemoveFromWorkList(Caller);
9918 // Replace the trampoline call with a direct call. Since there is no 'nest'
9919 // parameter, there is no need to adjust the argument list. Let the generic
9920 // code sort out any function type mismatches.
9921 Constant *NewCallee =
9922 NestF->getType() == PTy ? NestF : ConstantExpr::getBitCast(NestF, PTy);
9923 CS.setCalledFunction(NewCallee);
9924 return CS.getInstruction();
9927 /// FoldPHIArgBinOpIntoPHI - If we have something like phi [add (a,b), add(c,d)]
9928 /// and if a/b/c/d and the add's all have a single use, turn this into two phi's
9929 /// and a single binop.
9930 Instruction *InstCombiner::FoldPHIArgBinOpIntoPHI(PHINode &PN) {
9931 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
9932 assert(isa<BinaryOperator>(FirstInst) || isa<CmpInst>(FirstInst));
9933 unsigned Opc = FirstInst->getOpcode();
9934 Value *LHSVal = FirstInst->getOperand(0);
9935 Value *RHSVal = FirstInst->getOperand(1);
9937 const Type *LHSType = LHSVal->getType();
9938 const Type *RHSType = RHSVal->getType();
9940 // Scan to see if all operands are the same opcode, all have one use, and all
9941 // kill their operands (i.e. the operands have one use).
9942 for (unsigned i = 1; i != PN.getNumIncomingValues(); ++i) {
9943 Instruction *I = dyn_cast<Instruction>(PN.getIncomingValue(i));
9944 if (!I || I->getOpcode() != Opc || !I->hasOneUse() ||
9945 // Verify type of the LHS matches so we don't fold cmp's of different
9946 // types or GEP's with different index types.
9947 I->getOperand(0)->getType() != LHSType ||
9948 I->getOperand(1)->getType() != RHSType)
9951 // If they are CmpInst instructions, check their predicates
9952 if (Opc == Instruction::ICmp || Opc == Instruction::FCmp)
9953 if (cast<CmpInst>(I)->getPredicate() !=
9954 cast<CmpInst>(FirstInst)->getPredicate())
9957 // Keep track of which operand needs a phi node.
9958 if (I->getOperand(0) != LHSVal) LHSVal = 0;
9959 if (I->getOperand(1) != RHSVal) RHSVal = 0;
9962 // Otherwise, this is safe to transform!
9964 Value *InLHS = FirstInst->getOperand(0);
9965 Value *InRHS = FirstInst->getOperand(1);
9966 PHINode *NewLHS = 0, *NewRHS = 0;
9968 NewLHS = PHINode::Create(LHSType,
9969 FirstInst->getOperand(0)->getName() + ".pn");
9970 NewLHS->reserveOperandSpace(PN.getNumOperands()/2);
9971 NewLHS->addIncoming(InLHS, PN.getIncomingBlock(0));
9972 InsertNewInstBefore(NewLHS, PN);
9977 NewRHS = PHINode::Create(RHSType,
9978 FirstInst->getOperand(1)->getName() + ".pn");
9979 NewRHS->reserveOperandSpace(PN.getNumOperands()/2);
9980 NewRHS->addIncoming(InRHS, PN.getIncomingBlock(0));
9981 InsertNewInstBefore(NewRHS, PN);
9985 // Add all operands to the new PHIs.
9986 if (NewLHS || NewRHS) {
9987 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
9988 Instruction *InInst = cast<Instruction>(PN.getIncomingValue(i));
9990 Value *NewInLHS = InInst->getOperand(0);
9991 NewLHS->addIncoming(NewInLHS, PN.getIncomingBlock(i));
9994 Value *NewInRHS = InInst->getOperand(1);
9995 NewRHS->addIncoming(NewInRHS, PN.getIncomingBlock(i));
10000 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
10001 return BinaryOperator::Create(BinOp->getOpcode(), LHSVal, RHSVal);
10002 CmpInst *CIOp = cast<CmpInst>(FirstInst);
10003 return CmpInst::Create(CIOp->getOpcode(), CIOp->getPredicate(), LHSVal,
10007 Instruction *InstCombiner::FoldPHIArgGEPIntoPHI(PHINode &PN) {
10008 GetElementPtrInst *FirstInst =cast<GetElementPtrInst>(PN.getIncomingValue(0));
10010 SmallVector<Value*, 16> FixedOperands(FirstInst->op_begin(),
10011 FirstInst->op_end());
10013 // Scan to see if all operands are the same opcode, all have one use, and all
10014 // kill their operands (i.e. the operands have one use).
10015 for (unsigned i = 1; i != PN.getNumIncomingValues(); ++i) {
10016 GetElementPtrInst *GEP= dyn_cast<GetElementPtrInst>(PN.getIncomingValue(i));
10017 if (!GEP || !GEP->hasOneUse() || GEP->getType() != FirstInst->getType() ||
10018 GEP->getNumOperands() != FirstInst->getNumOperands())
10021 // Compare the operand lists.
10022 for (unsigned op = 0, e = FirstInst->getNumOperands(); op != e; ++op) {
10023 if (FirstInst->getOperand(op) == GEP->getOperand(op))
10026 // Don't merge two GEPs when two operands differ (introducing phi nodes)
10027 // if one of the PHIs has a constant for the index. The index may be
10028 // substantially cheaper to compute for the constants, so making it a
10029 // variable index could pessimize the path. This also handles the case
10030 // for struct indices, which must always be constant.
10031 if (isa<ConstantInt>(FirstInst->getOperand(op)) ||
10032 isa<ConstantInt>(GEP->getOperand(op)))
10035 if (FirstInst->getOperand(op)->getType() !=GEP->getOperand(op)->getType())
10037 FixedOperands[op] = 0; // Needs a PHI.
10041 // Otherwise, this is safe to transform. Insert PHI nodes for each operand
10042 // that is variable.
10043 SmallVector<PHINode*, 16> OperandPhis(FixedOperands.size());
10045 bool HasAnyPHIs = false;
10046 for (unsigned i = 0, e = FixedOperands.size(); i != e; ++i) {
10047 if (FixedOperands[i]) continue; // operand doesn't need a phi.
10048 Value *FirstOp = FirstInst->getOperand(i);
10049 PHINode *NewPN = PHINode::Create(FirstOp->getType(),
10050 FirstOp->getName()+".pn");
10051 InsertNewInstBefore(NewPN, PN);
10053 NewPN->reserveOperandSpace(e);
10054 NewPN->addIncoming(FirstOp, PN.getIncomingBlock(0));
10055 OperandPhis[i] = NewPN;
10056 FixedOperands[i] = NewPN;
10061 // Add all operands to the new PHIs.
10063 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
10064 GetElementPtrInst *InGEP =cast<GetElementPtrInst>(PN.getIncomingValue(i));
10065 BasicBlock *InBB = PN.getIncomingBlock(i);
10067 for (unsigned op = 0, e = OperandPhis.size(); op != e; ++op)
10068 if (PHINode *OpPhi = OperandPhis[op])
10069 OpPhi->addIncoming(InGEP->getOperand(op), InBB);
10073 Value *Base = FixedOperands[0];
10074 return GetElementPtrInst::Create(Base, FixedOperands.begin()+1,
10075 FixedOperands.end());
10079 /// isSafeToSinkLoad - Return true if we know that it is safe sink the load out
10080 /// of the block that defines it. This means that it must be obvious the value
10081 /// of the load is not changed from the point of the load to the end of the
10082 /// block it is in.
10084 /// Finally, it is safe, but not profitable, to sink a load targetting a
10085 /// non-address-taken alloca. Doing so will cause us to not promote the alloca
10087 static bool isSafeToSinkLoad(LoadInst *L) {
10088 BasicBlock::iterator BBI = L, E = L->getParent()->end();
10090 for (++BBI; BBI != E; ++BBI)
10091 if (BBI->mayWriteToMemory())
10094 // Check for non-address taken alloca. If not address-taken already, it isn't
10095 // profitable to do this xform.
10096 if (AllocaInst *AI = dyn_cast<AllocaInst>(L->getOperand(0))) {
10097 bool isAddressTaken = false;
10098 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
10100 if (isa<LoadInst>(UI)) continue;
10101 if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
10102 // If storing TO the alloca, then the address isn't taken.
10103 if (SI->getOperand(1) == AI) continue;
10105 isAddressTaken = true;
10109 if (!isAddressTaken)
10117 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
10118 // operator and they all are only used by the PHI, PHI together their
10119 // inputs, and do the operation once, to the result of the PHI.
10120 Instruction *InstCombiner::FoldPHIArgOpIntoPHI(PHINode &PN) {
10121 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
10123 // Scan the instruction, looking for input operations that can be folded away.
10124 // If all input operands to the phi are the same instruction (e.g. a cast from
10125 // the same type or "+42") we can pull the operation through the PHI, reducing
10126 // code size and simplifying code.
10127 Constant *ConstantOp = 0;
10128 const Type *CastSrcTy = 0;
10129 bool isVolatile = false;
10130 if (isa<CastInst>(FirstInst)) {
10131 CastSrcTy = FirstInst->getOperand(0)->getType();
10132 } else if (isa<BinaryOperator>(FirstInst) || isa<CmpInst>(FirstInst)) {
10133 // Can fold binop, compare or shift here if the RHS is a constant,
10134 // otherwise call FoldPHIArgBinOpIntoPHI.
10135 ConstantOp = dyn_cast<Constant>(FirstInst->getOperand(1));
10136 if (ConstantOp == 0)
10137 return FoldPHIArgBinOpIntoPHI(PN);
10138 } else if (LoadInst *LI = dyn_cast<LoadInst>(FirstInst)) {
10139 isVolatile = LI->isVolatile();
10140 // We can't sink the load if the loaded value could be modified between the
10141 // load and the PHI.
10142 if (LI->getParent() != PN.getIncomingBlock(0) ||
10143 !isSafeToSinkLoad(LI))
10146 // If the PHI is of volatile loads and the load block has multiple
10147 // successors, sinking it would remove a load of the volatile value from
10148 // the path through the other successor.
10150 LI->getParent()->getTerminator()->getNumSuccessors() != 1)
10153 } else if (isa<GetElementPtrInst>(FirstInst)) {
10154 return FoldPHIArgGEPIntoPHI(PN);
10156 return 0; // Cannot fold this operation.
10159 // Check to see if all arguments are the same operation.
10160 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
10161 if (!isa<Instruction>(PN.getIncomingValue(i))) return 0;
10162 Instruction *I = cast<Instruction>(PN.getIncomingValue(i));
10163 if (!I->hasOneUse() || !I->isSameOperationAs(FirstInst))
10166 if (I->getOperand(0)->getType() != CastSrcTy)
10167 return 0; // Cast operation must match.
10168 } else if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
10169 // We can't sink the load if the loaded value could be modified between
10170 // the load and the PHI.
10171 if (LI->isVolatile() != isVolatile ||
10172 LI->getParent() != PN.getIncomingBlock(i) ||
10173 !isSafeToSinkLoad(LI))
10176 // If the PHI is of volatile loads and the load block has multiple
10177 // successors, sinking it would remove a load of the volatile value from
10178 // the path through the other successor.
10180 LI->getParent()->getTerminator()->getNumSuccessors() != 1)
10184 } else if (I->getOperand(1) != ConstantOp) {
10189 // Okay, they are all the same operation. Create a new PHI node of the
10190 // correct type, and PHI together all of the LHS's of the instructions.
10191 PHINode *NewPN = PHINode::Create(FirstInst->getOperand(0)->getType(),
10192 PN.getName()+".in");
10193 NewPN->reserveOperandSpace(PN.getNumOperands()/2);
10195 Value *InVal = FirstInst->getOperand(0);
10196 NewPN->addIncoming(InVal, PN.getIncomingBlock(0));
10198 // Add all operands to the new PHI.
10199 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
10200 Value *NewInVal = cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
10201 if (NewInVal != InVal)
10203 NewPN->addIncoming(NewInVal, PN.getIncomingBlock(i));
10208 // The new PHI unions all of the same values together. This is really
10209 // common, so we handle it intelligently here for compile-time speed.
10213 InsertNewInstBefore(NewPN, PN);
10217 // Insert and return the new operation.
10218 if (CastInst* FirstCI = dyn_cast<CastInst>(FirstInst))
10219 return CastInst::Create(FirstCI->getOpcode(), PhiVal, PN.getType());
10220 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
10221 return BinaryOperator::Create(BinOp->getOpcode(), PhiVal, ConstantOp);
10222 if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst))
10223 return CmpInst::Create(CIOp->getOpcode(), CIOp->getPredicate(),
10224 PhiVal, ConstantOp);
10225 assert(isa<LoadInst>(FirstInst) && "Unknown operation");
10227 // If this was a volatile load that we are merging, make sure to loop through
10228 // and mark all the input loads as non-volatile. If we don't do this, we will
10229 // insert a new volatile load and the old ones will not be deletable.
10231 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i)
10232 cast<LoadInst>(PN.getIncomingValue(i))->setVolatile(false);
10234 return new LoadInst(PhiVal, "", isVolatile);
10237 /// DeadPHICycle - Return true if this PHI node is only used by a PHI node cycle
10239 static bool DeadPHICycle(PHINode *PN,
10240 SmallPtrSet<PHINode*, 16> &PotentiallyDeadPHIs) {
10241 if (PN->use_empty()) return true;
10242 if (!PN->hasOneUse()) return false;
10244 // Remember this node, and if we find the cycle, return.
10245 if (!PotentiallyDeadPHIs.insert(PN))
10248 // Don't scan crazily complex things.
10249 if (PotentiallyDeadPHIs.size() == 16)
10252 if (PHINode *PU = dyn_cast<PHINode>(PN->use_back()))
10253 return DeadPHICycle(PU, PotentiallyDeadPHIs);
10258 /// PHIsEqualValue - Return true if this phi node is always equal to
10259 /// NonPhiInVal. This happens with mutually cyclic phi nodes like:
10260 /// z = some value; x = phi (y, z); y = phi (x, z)
10261 static bool PHIsEqualValue(PHINode *PN, Value *NonPhiInVal,
10262 SmallPtrSet<PHINode*, 16> &ValueEqualPHIs) {
10263 // See if we already saw this PHI node.
10264 if (!ValueEqualPHIs.insert(PN))
10267 // Don't scan crazily complex things.
10268 if (ValueEqualPHIs.size() == 16)
10271 // Scan the operands to see if they are either phi nodes or are equal to
10273 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
10274 Value *Op = PN->getIncomingValue(i);
10275 if (PHINode *OpPN = dyn_cast<PHINode>(Op)) {
10276 if (!PHIsEqualValue(OpPN, NonPhiInVal, ValueEqualPHIs))
10278 } else if (Op != NonPhiInVal)
10286 // PHINode simplification
10288 Instruction *InstCombiner::visitPHINode(PHINode &PN) {
10289 // If LCSSA is around, don't mess with Phi nodes
10290 if (MustPreserveLCSSA) return 0;
10292 if (Value *V = PN.hasConstantValue())
10293 return ReplaceInstUsesWith(PN, V);
10295 // If all PHI operands are the same operation, pull them through the PHI,
10296 // reducing code size.
10297 if (isa<Instruction>(PN.getIncomingValue(0)) &&
10298 isa<Instruction>(PN.getIncomingValue(1)) &&
10299 cast<Instruction>(PN.getIncomingValue(0))->getOpcode() ==
10300 cast<Instruction>(PN.getIncomingValue(1))->getOpcode() &&
10301 // FIXME: The hasOneUse check will fail for PHIs that use the value more
10302 // than themselves more than once.
10303 PN.getIncomingValue(0)->hasOneUse())
10304 if (Instruction *Result = FoldPHIArgOpIntoPHI(PN))
10307 // If this is a trivial cycle in the PHI node graph, remove it. Basically, if
10308 // this PHI only has a single use (a PHI), and if that PHI only has one use (a
10309 // PHI)... break the cycle.
10310 if (PN.hasOneUse()) {
10311 Instruction *PHIUser = cast<Instruction>(PN.use_back());
10312 if (PHINode *PU = dyn_cast<PHINode>(PHIUser)) {
10313 SmallPtrSet<PHINode*, 16> PotentiallyDeadPHIs;
10314 PotentiallyDeadPHIs.insert(&PN);
10315 if (DeadPHICycle(PU, PotentiallyDeadPHIs))
10316 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
10319 // If this phi has a single use, and if that use just computes a value for
10320 // the next iteration of a loop, delete the phi. This occurs with unused
10321 // induction variables, e.g. "for (int j = 0; ; ++j);". Detecting this
10322 // common case here is good because the only other things that catch this
10323 // are induction variable analysis (sometimes) and ADCE, which is only run
10325 if (PHIUser->hasOneUse() &&
10326 (isa<BinaryOperator>(PHIUser) || isa<GetElementPtrInst>(PHIUser)) &&
10327 PHIUser->use_back() == &PN) {
10328 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
10332 // We sometimes end up with phi cycles that non-obviously end up being the
10333 // same value, for example:
10334 // z = some value; x = phi (y, z); y = phi (x, z)
10335 // where the phi nodes don't necessarily need to be in the same block. Do a
10336 // quick check to see if the PHI node only contains a single non-phi value, if
10337 // so, scan to see if the phi cycle is actually equal to that value.
10339 unsigned InValNo = 0, NumOperandVals = PN.getNumIncomingValues();
10340 // Scan for the first non-phi operand.
10341 while (InValNo != NumOperandVals &&
10342 isa<PHINode>(PN.getIncomingValue(InValNo)))
10345 if (InValNo != NumOperandVals) {
10346 Value *NonPhiInVal = PN.getOperand(InValNo);
10348 // Scan the rest of the operands to see if there are any conflicts, if so
10349 // there is no need to recursively scan other phis.
10350 for (++InValNo; InValNo != NumOperandVals; ++InValNo) {
10351 Value *OpVal = PN.getIncomingValue(InValNo);
10352 if (OpVal != NonPhiInVal && !isa<PHINode>(OpVal))
10356 // If we scanned over all operands, then we have one unique value plus
10357 // phi values. Scan PHI nodes to see if they all merge in each other or
10359 if (InValNo == NumOperandVals) {
10360 SmallPtrSet<PHINode*, 16> ValueEqualPHIs;
10361 if (PHIsEqualValue(&PN, NonPhiInVal, ValueEqualPHIs))
10362 return ReplaceInstUsesWith(PN, NonPhiInVal);
10369 static Value *InsertCastToIntPtrTy(Value *V, const Type *DTy,
10370 Instruction *InsertPoint,
10371 InstCombiner *IC) {
10372 unsigned PtrSize = DTy->getPrimitiveSizeInBits();
10373 unsigned VTySize = V->getType()->getPrimitiveSizeInBits();
10374 // We must cast correctly to the pointer type. Ensure that we
10375 // sign extend the integer value if it is smaller as this is
10376 // used for address computation.
10377 Instruction::CastOps opcode =
10378 (VTySize < PtrSize ? Instruction::SExt :
10379 (VTySize == PtrSize ? Instruction::BitCast : Instruction::Trunc));
10380 return IC->InsertCastBefore(opcode, V, DTy, *InsertPoint);
10384 Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) {
10385 Value *PtrOp = GEP.getOperand(0);
10386 // Is it 'getelementptr %P, i32 0' or 'getelementptr %P'
10387 // If so, eliminate the noop.
10388 if (GEP.getNumOperands() == 1)
10389 return ReplaceInstUsesWith(GEP, PtrOp);
10391 if (isa<UndefValue>(GEP.getOperand(0)))
10392 return ReplaceInstUsesWith(GEP, UndefValue::get(GEP.getType()));
10394 bool HasZeroPointerIndex = false;
10395 if (Constant *C = dyn_cast<Constant>(GEP.getOperand(1)))
10396 HasZeroPointerIndex = C->isNullValue();
10398 if (GEP.getNumOperands() == 2 && HasZeroPointerIndex)
10399 return ReplaceInstUsesWith(GEP, PtrOp);
10401 // Eliminate unneeded casts for indices.
10402 bool MadeChange = false;
10404 gep_type_iterator GTI = gep_type_begin(GEP);
10405 for (User::op_iterator i = GEP.op_begin() + 1, e = GEP.op_end();
10406 i != e; ++i, ++GTI) {
10407 if (isa<SequentialType>(*GTI)) {
10408 if (CastInst *CI = dyn_cast<CastInst>(*i)) {
10409 if (CI->getOpcode() == Instruction::ZExt ||
10410 CI->getOpcode() == Instruction::SExt) {
10411 const Type *SrcTy = CI->getOperand(0)->getType();
10412 // We can eliminate a cast from i32 to i64 iff the target
10413 // is a 32-bit pointer target.
10414 if (SrcTy->getPrimitiveSizeInBits() >= TD->getPointerSizeInBits()) {
10416 *i = CI->getOperand(0);
10420 // If we are using a wider index than needed for this platform, shrink it
10421 // to what we need. If narrower, sign-extend it to what we need.
10422 // If the incoming value needs a cast instruction,
10423 // insert it. This explicit cast can make subsequent optimizations more
10426 if (TD->getTypeSizeInBits(Op->getType()) > TD->getPointerSizeInBits()) {
10427 if (Constant *C = dyn_cast<Constant>(Op)) {
10428 *i = ConstantExpr::getTrunc(C, TD->getIntPtrType());
10431 Op = InsertCastBefore(Instruction::Trunc, Op, TD->getIntPtrType(),
10436 } else if (TD->getTypeSizeInBits(Op->getType()) < TD->getPointerSizeInBits()) {
10437 if (Constant *C = dyn_cast<Constant>(Op)) {
10438 *i = ConstantExpr::getSExt(C, TD->getIntPtrType());
10441 Op = InsertCastBefore(Instruction::SExt, Op, TD->getIntPtrType(),
10449 if (MadeChange) return &GEP;
10451 // Combine Indices - If the source pointer to this getelementptr instruction
10452 // is a getelementptr instruction, combine the indices of the two
10453 // getelementptr instructions into a single instruction.
10455 SmallVector<Value*, 8> SrcGEPOperands;
10456 if (User *Src = dyn_castGetElementPtr(PtrOp))
10457 SrcGEPOperands.append(Src->op_begin(), Src->op_end());
10459 if (!SrcGEPOperands.empty()) {
10460 // Note that if our source is a gep chain itself that we wait for that
10461 // chain to be resolved before we perform this transformation. This
10462 // avoids us creating a TON of code in some cases.
10464 if (isa<GetElementPtrInst>(SrcGEPOperands[0]) &&
10465 cast<Instruction>(SrcGEPOperands[0])->getNumOperands() == 2)
10466 return 0; // Wait until our source is folded to completion.
10468 SmallVector<Value*, 8> Indices;
10470 // Find out whether the last index in the source GEP is a sequential idx.
10471 bool EndsWithSequential = false;
10472 for (gep_type_iterator I = gep_type_begin(*cast<User>(PtrOp)),
10473 E = gep_type_end(*cast<User>(PtrOp)); I != E; ++I)
10474 EndsWithSequential = !isa<StructType>(*I);
10476 // Can we combine the two pointer arithmetics offsets?
10477 if (EndsWithSequential) {
10478 // Replace: gep (gep %P, long B), long A, ...
10479 // With: T = long A+B; gep %P, T, ...
10481 Value *Sum, *SO1 = SrcGEPOperands.back(), *GO1 = GEP.getOperand(1);
10482 if (SO1 == Constant::getNullValue(SO1->getType())) {
10484 } else if (GO1 == Constant::getNullValue(GO1->getType())) {
10487 // If they aren't the same type, convert both to an integer of the
10488 // target's pointer size.
10489 if (SO1->getType() != GO1->getType()) {
10490 if (Constant *SO1C = dyn_cast<Constant>(SO1)) {
10491 SO1 = ConstantExpr::getIntegerCast(SO1C, GO1->getType(), true);
10492 } else if (Constant *GO1C = dyn_cast<Constant>(GO1)) {
10493 GO1 = ConstantExpr::getIntegerCast(GO1C, SO1->getType(), true);
10495 unsigned PS = TD->getPointerSizeInBits();
10496 if (TD->getTypeSizeInBits(SO1->getType()) == PS) {
10497 // Convert GO1 to SO1's type.
10498 GO1 = InsertCastToIntPtrTy(GO1, SO1->getType(), &GEP, this);
10500 } else if (TD->getTypeSizeInBits(GO1->getType()) == PS) {
10501 // Convert SO1 to GO1's type.
10502 SO1 = InsertCastToIntPtrTy(SO1, GO1->getType(), &GEP, this);
10504 const Type *PT = TD->getIntPtrType();
10505 SO1 = InsertCastToIntPtrTy(SO1, PT, &GEP, this);
10506 GO1 = InsertCastToIntPtrTy(GO1, PT, &GEP, this);
10510 if (isa<Constant>(SO1) && isa<Constant>(GO1))
10511 Sum = ConstantExpr::getAdd(cast<Constant>(SO1), cast<Constant>(GO1));
10513 Sum = BinaryOperator::CreateAdd(SO1, GO1, PtrOp->getName()+".sum");
10514 InsertNewInstBefore(cast<Instruction>(Sum), GEP);
10518 // Recycle the GEP we already have if possible.
10519 if (SrcGEPOperands.size() == 2) {
10520 GEP.setOperand(0, SrcGEPOperands[0]);
10521 GEP.setOperand(1, Sum);
10524 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
10525 SrcGEPOperands.end()-1);
10526 Indices.push_back(Sum);
10527 Indices.insert(Indices.end(), GEP.op_begin()+2, GEP.op_end());
10529 } else if (isa<Constant>(*GEP.idx_begin()) &&
10530 cast<Constant>(*GEP.idx_begin())->isNullValue() &&
10531 SrcGEPOperands.size() != 1) {
10532 // Otherwise we can do the fold if the first index of the GEP is a zero
10533 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
10534 SrcGEPOperands.end());
10535 Indices.insert(Indices.end(), GEP.idx_begin()+1, GEP.idx_end());
10538 if (!Indices.empty())
10539 return GetElementPtrInst::Create(SrcGEPOperands[0], Indices.begin(),
10540 Indices.end(), GEP.getName());
10542 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(PtrOp)) {
10543 // GEP of global variable. If all of the indices for this GEP are
10544 // constants, we can promote this to a constexpr instead of an instruction.
10546 // Scan for nonconstants...
10547 SmallVector<Constant*, 8> Indices;
10548 User::op_iterator I = GEP.idx_begin(), E = GEP.idx_end();
10549 for (; I != E && isa<Constant>(*I); ++I)
10550 Indices.push_back(cast<Constant>(*I));
10552 if (I == E) { // If they are all constants...
10553 Constant *CE = ConstantExpr::getGetElementPtr(GV,
10554 &Indices[0],Indices.size());
10556 // Replace all uses of the GEP with the new constexpr...
10557 return ReplaceInstUsesWith(GEP, CE);
10559 } else if (Value *X = getBitCastOperand(PtrOp)) { // Is the operand a cast?
10560 if (!isa<PointerType>(X->getType())) {
10561 // Not interesting. Source pointer must be a cast from pointer.
10562 } else if (HasZeroPointerIndex) {
10563 // transform: GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ...
10564 // into : GEP [10 x i8]* X, i32 0, ...
10566 // This occurs when the program declares an array extern like "int X[];"
10568 const PointerType *CPTy = cast<PointerType>(PtrOp->getType());
10569 const PointerType *XTy = cast<PointerType>(X->getType());
10570 if (const ArrayType *XATy =
10571 dyn_cast<ArrayType>(XTy->getElementType()))
10572 if (const ArrayType *CATy =
10573 dyn_cast<ArrayType>(CPTy->getElementType()))
10574 if (CATy->getElementType() == XATy->getElementType()) {
10575 // At this point, we know that the cast source type is a pointer
10576 // to an array of the same type as the destination pointer
10577 // array. Because the array type is never stepped over (there
10578 // is a leading zero) we can fold the cast into this GEP.
10579 GEP.setOperand(0, X);
10582 } else if (GEP.getNumOperands() == 2) {
10583 // Transform things like:
10584 // %t = getelementptr i32* bitcast ([2 x i32]* %str to i32*), i32 %V
10585 // into: %t1 = getelementptr [2 x i32]* %str, i32 0, i32 %V; bitcast
10586 const Type *SrcElTy = cast<PointerType>(X->getType())->getElementType();
10587 const Type *ResElTy=cast<PointerType>(PtrOp->getType())->getElementType();
10588 if (isa<ArrayType>(SrcElTy) &&
10589 TD->getABITypeSize(cast<ArrayType>(SrcElTy)->getElementType()) ==
10590 TD->getABITypeSize(ResElTy)) {
10592 Idx[0] = Constant::getNullValue(Type::Int32Ty);
10593 Idx[1] = GEP.getOperand(1);
10594 Value *V = InsertNewInstBefore(
10595 GetElementPtrInst::Create(X, Idx, Idx + 2, GEP.getName()), GEP);
10596 // V and GEP are both pointer types --> BitCast
10597 return new BitCastInst(V, GEP.getType());
10600 // Transform things like:
10601 // getelementptr i8* bitcast ([100 x double]* X to i8*), i32 %tmp
10602 // (where tmp = 8*tmp2) into:
10603 // getelementptr [100 x double]* %arr, i32 0, i32 %tmp2; bitcast
10605 if (isa<ArrayType>(SrcElTy) && ResElTy == Type::Int8Ty) {
10606 uint64_t ArrayEltSize =
10607 TD->getABITypeSize(cast<ArrayType>(SrcElTy)->getElementType());
10609 // Check to see if "tmp" is a scale by a multiple of ArrayEltSize. We
10610 // allow either a mul, shift, or constant here.
10612 ConstantInt *Scale = 0;
10613 if (ArrayEltSize == 1) {
10614 NewIdx = GEP.getOperand(1);
10615 Scale = ConstantInt::get(NewIdx->getType(), 1);
10616 } else if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP.getOperand(1))) {
10617 NewIdx = ConstantInt::get(CI->getType(), 1);
10619 } else if (Instruction *Inst =dyn_cast<Instruction>(GEP.getOperand(1))){
10620 if (Inst->getOpcode() == Instruction::Shl &&
10621 isa<ConstantInt>(Inst->getOperand(1))) {
10622 ConstantInt *ShAmt = cast<ConstantInt>(Inst->getOperand(1));
10623 uint32_t ShAmtVal = ShAmt->getLimitedValue(64);
10624 Scale = ConstantInt::get(Inst->getType(), 1ULL << ShAmtVal);
10625 NewIdx = Inst->getOperand(0);
10626 } else if (Inst->getOpcode() == Instruction::Mul &&
10627 isa<ConstantInt>(Inst->getOperand(1))) {
10628 Scale = cast<ConstantInt>(Inst->getOperand(1));
10629 NewIdx = Inst->getOperand(0);
10633 // If the index will be to exactly the right offset with the scale taken
10634 // out, perform the transformation. Note, we don't know whether Scale is
10635 // signed or not. We'll use unsigned version of division/modulo
10636 // operation after making sure Scale doesn't have the sign bit set.
10637 if (Scale && Scale->getSExtValue() >= 0LL &&
10638 Scale->getZExtValue() % ArrayEltSize == 0) {
10639 Scale = ConstantInt::get(Scale->getType(),
10640 Scale->getZExtValue() / ArrayEltSize);
10641 if (Scale->getZExtValue() != 1) {
10642 Constant *C = ConstantExpr::getIntegerCast(Scale, NewIdx->getType(),
10644 Instruction *Sc = BinaryOperator::CreateMul(NewIdx, C, "idxscale");
10645 NewIdx = InsertNewInstBefore(Sc, GEP);
10648 // Insert the new GEP instruction.
10650 Idx[0] = Constant::getNullValue(Type::Int32Ty);
10652 Instruction *NewGEP =
10653 GetElementPtrInst::Create(X, Idx, Idx + 2, GEP.getName());
10654 NewGEP = InsertNewInstBefore(NewGEP, GEP);
10655 // The NewGEP must be pointer typed, so must the old one -> BitCast
10656 return new BitCastInst(NewGEP, GEP.getType());
10662 /// See if we can simplify:
10663 /// X = bitcast A to B*
10664 /// Y = gep X, <...constant indices...>
10665 /// into a gep of the original struct. This is important for SROA and alias
10666 /// analysis of unions. If "A" is also a bitcast, wait for A/X to be merged.
10667 if (BitCastInst *BCI = dyn_cast<BitCastInst>(PtrOp)) {
10668 if (!isa<BitCastInst>(BCI->getOperand(0)) && GEP.hasAllConstantIndices()) {
10669 // Determine how much the GEP moves the pointer. We are guaranteed to get
10670 // a constant back from EmitGEPOffset.
10671 ConstantInt *OffsetV = cast<ConstantInt>(EmitGEPOffset(&GEP, GEP, *this));
10672 int64_t Offset = OffsetV->getSExtValue();
10674 // If this GEP instruction doesn't move the pointer, just replace the GEP
10675 // with a bitcast of the real input to the dest type.
10677 // If the bitcast is of an allocation, and the allocation will be
10678 // converted to match the type of the cast, don't touch this.
10679 if (isa<AllocationInst>(BCI->getOperand(0))) {
10680 // See if the bitcast simplifies, if so, don't nuke this GEP yet.
10681 if (Instruction *I = visitBitCast(*BCI)) {
10684 BCI->getParent()->getInstList().insert(BCI, I);
10685 ReplaceInstUsesWith(*BCI, I);
10690 return new BitCastInst(BCI->getOperand(0), GEP.getType());
10693 // Otherwise, if the offset is non-zero, we need to find out if there is a
10694 // field at Offset in 'A's type. If so, we can pull the cast through the
10696 SmallVector<Value*, 8> NewIndices;
10698 cast<PointerType>(BCI->getOperand(0)->getType())->getElementType();
10699 if (FindElementAtOffset(InTy, Offset, NewIndices, TD)) {
10700 Instruction *NGEP =
10701 GetElementPtrInst::Create(BCI->getOperand(0), NewIndices.begin(),
10703 if (NGEP->getType() == GEP.getType()) return NGEP;
10704 InsertNewInstBefore(NGEP, GEP);
10705 NGEP->takeName(&GEP);
10706 return new BitCastInst(NGEP, GEP.getType());
10714 Instruction *InstCombiner::visitAllocationInst(AllocationInst &AI) {
10715 // Convert: malloc Ty, C - where C is a constant != 1 into: malloc [C x Ty], 1
10716 if (AI.isArrayAllocation()) { // Check C != 1
10717 if (const ConstantInt *C = dyn_cast<ConstantInt>(AI.getArraySize())) {
10718 const Type *NewTy =
10719 ArrayType::get(AI.getAllocatedType(), C->getZExtValue());
10720 AllocationInst *New = 0;
10722 // Create and insert the replacement instruction...
10723 if (isa<MallocInst>(AI))
10724 New = new MallocInst(NewTy, 0, AI.getAlignment(), AI.getName());
10726 assert(isa<AllocaInst>(AI) && "Unknown type of allocation inst!");
10727 New = new AllocaInst(NewTy, 0, AI.getAlignment(), AI.getName());
10730 InsertNewInstBefore(New, AI);
10732 // Scan to the end of the allocation instructions, to skip over a block of
10733 // allocas if possible...
10735 BasicBlock::iterator It = New;
10736 while (isa<AllocationInst>(*It)) ++It;
10738 // Now that I is pointing to the first non-allocation-inst in the block,
10739 // insert our getelementptr instruction...
10741 Value *NullIdx = Constant::getNullValue(Type::Int32Ty);
10745 Value *V = GetElementPtrInst::Create(New, Idx, Idx + 2,
10746 New->getName()+".sub", It);
10748 // Now make everything use the getelementptr instead of the original
10750 return ReplaceInstUsesWith(AI, V);
10751 } else if (isa<UndefValue>(AI.getArraySize())) {
10752 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
10756 // If alloca'ing a zero byte object, replace the alloca with a null pointer.
10757 // Note that we only do this for alloca's, because malloc should allocate and
10758 // return a unique pointer, even for a zero byte allocation.
10759 if (isa<AllocaInst>(AI) && AI.getAllocatedType()->isSized() &&
10760 TD->getABITypeSize(AI.getAllocatedType()) == 0)
10761 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
10766 Instruction *InstCombiner::visitFreeInst(FreeInst &FI) {
10767 Value *Op = FI.getOperand(0);
10769 // free undef -> unreachable.
10770 if (isa<UndefValue>(Op)) {
10771 // Insert a new store to null because we cannot modify the CFG here.
10772 new StoreInst(ConstantInt::getTrue(),
10773 UndefValue::get(PointerType::getUnqual(Type::Int1Ty)), &FI);
10774 return EraseInstFromFunction(FI);
10777 // If we have 'free null' delete the instruction. This can happen in stl code
10778 // when lots of inlining happens.
10779 if (isa<ConstantPointerNull>(Op))
10780 return EraseInstFromFunction(FI);
10782 // Change free <ty>* (cast <ty2>* X to <ty>*) into free <ty2>* X
10783 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op)) {
10784 FI.setOperand(0, CI->getOperand(0));
10788 // Change free (gep X, 0,0,0,0) into free(X)
10789 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op)) {
10790 if (GEPI->hasAllZeroIndices()) {
10791 AddToWorkList(GEPI);
10792 FI.setOperand(0, GEPI->getOperand(0));
10797 // Change free(malloc) into nothing, if the malloc has a single use.
10798 if (MallocInst *MI = dyn_cast<MallocInst>(Op))
10799 if (MI->hasOneUse()) {
10800 EraseInstFromFunction(FI);
10801 return EraseInstFromFunction(*MI);
10808 /// InstCombineLoadCast - Fold 'load (cast P)' -> cast (load P)' when possible.
10809 static Instruction *InstCombineLoadCast(InstCombiner &IC, LoadInst &LI,
10810 const TargetData *TD) {
10811 User *CI = cast<User>(LI.getOperand(0));
10812 Value *CastOp = CI->getOperand(0);
10814 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(CI)) {
10815 // Instead of loading constant c string, use corresponding integer value
10816 // directly if string length is small enough.
10818 if (GetConstantStringInfo(CE->getOperand(0), Str) && !Str.empty()) {
10819 unsigned len = Str.length();
10820 const Type *Ty = cast<PointerType>(CE->getType())->getElementType();
10821 unsigned numBits = Ty->getPrimitiveSizeInBits();
10822 // Replace LI with immediate integer store.
10823 if ((numBits >> 3) == len + 1) {
10824 APInt StrVal(numBits, 0);
10825 APInt SingleChar(numBits, 0);
10826 if (TD->isLittleEndian()) {
10827 for (signed i = len-1; i >= 0; i--) {
10828 SingleChar = (uint64_t) Str[i];
10829 StrVal = (StrVal << 8) | SingleChar;
10832 for (unsigned i = 0; i < len; i++) {
10833 SingleChar = (uint64_t) Str[i];
10834 StrVal = (StrVal << 8) | SingleChar;
10836 // Append NULL at the end.
10838 StrVal = (StrVal << 8) | SingleChar;
10840 Value *NL = ConstantInt::get(StrVal);
10841 return IC.ReplaceInstUsesWith(LI, NL);
10846 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
10847 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
10848 const Type *SrcPTy = SrcTy->getElementType();
10850 if (DestPTy->isInteger() || isa<PointerType>(DestPTy) ||
10851 isa<VectorType>(DestPTy)) {
10852 // If the source is an array, the code below will not succeed. Check to
10853 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
10855 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
10856 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
10857 if (ASrcTy->getNumElements() != 0) {
10859 Idxs[0] = Idxs[1] = Constant::getNullValue(Type::Int32Ty);
10860 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs, 2);
10861 SrcTy = cast<PointerType>(CastOp->getType());
10862 SrcPTy = SrcTy->getElementType();
10865 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy) ||
10866 isa<VectorType>(SrcPTy)) &&
10867 // Do not allow turning this into a load of an integer, which is then
10868 // casted to a pointer, this pessimizes pointer analysis a lot.
10869 (isa<PointerType>(SrcPTy) == isa<PointerType>(LI.getType())) &&
10870 IC.getTargetData().getTypeSizeInBits(SrcPTy) ==
10871 IC.getTargetData().getTypeSizeInBits(DestPTy)) {
10873 // Okay, we are casting from one integer or pointer type to another of
10874 // the same size. Instead of casting the pointer before the load, cast
10875 // the result of the loaded value.
10876 Value *NewLoad = IC.InsertNewInstBefore(new LoadInst(CastOp,
10878 LI.isVolatile()),LI);
10879 // Now cast the result of the load.
10880 return new BitCastInst(NewLoad, LI.getType());
10887 /// isSafeToLoadUnconditionally - Return true if we know that executing a load
10888 /// from this value cannot trap. If it is not obviously safe to load from the
10889 /// specified pointer, we do a quick local scan of the basic block containing
10890 /// ScanFrom, to determine if the address is already accessed.
10891 static bool isSafeToLoadUnconditionally(Value *V, Instruction *ScanFrom) {
10892 // If it is an alloca it is always safe to load from.
10893 if (isa<AllocaInst>(V)) return true;
10895 // If it is a global variable it is mostly safe to load from.
10896 if (const GlobalValue *GV = dyn_cast<GlobalVariable>(V))
10897 // Don't try to evaluate aliases. External weak GV can be null.
10898 return !isa<GlobalAlias>(GV) && !GV->hasExternalWeakLinkage();
10900 // Otherwise, be a little bit agressive by scanning the local block where we
10901 // want to check to see if the pointer is already being loaded or stored
10902 // from/to. If so, the previous load or store would have already trapped,
10903 // so there is no harm doing an extra load (also, CSE will later eliminate
10904 // the load entirely).
10905 BasicBlock::iterator BBI = ScanFrom, E = ScanFrom->getParent()->begin();
10910 // If we see a free or a call (which might do a free) the pointer could be
10912 if (isa<FreeInst>(BBI) || isa<CallInst>(BBI))
10915 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
10916 if (LI->getOperand(0) == V) return true;
10917 } else if (StoreInst *SI = dyn_cast<StoreInst>(BBI)) {
10918 if (SI->getOperand(1) == V) return true;
10925 Instruction *InstCombiner::visitLoadInst(LoadInst &LI) {
10926 Value *Op = LI.getOperand(0);
10928 // Attempt to improve the alignment.
10929 unsigned KnownAlign = GetOrEnforceKnownAlignment(Op);
10931 (LI.getAlignment() == 0 ? TD->getABITypeAlignment(LI.getType()) :
10932 LI.getAlignment()))
10933 LI.setAlignment(KnownAlign);
10935 // load (cast X) --> cast (load X) iff safe
10936 if (isa<CastInst>(Op))
10937 if (Instruction *Res = InstCombineLoadCast(*this, LI, TD))
10940 // None of the following transforms are legal for volatile loads.
10941 if (LI.isVolatile()) return 0;
10943 // Do really simple store-to-load forwarding and load CSE, to catch cases
10944 // where there are several consequtive memory accesses to the same location,
10945 // separated by a few arithmetic operations.
10946 BasicBlock::iterator BBI = &LI;
10947 if (Value *AvailableVal = FindAvailableLoadedValue(Op, LI.getParent(), BBI,6))
10948 return ReplaceInstUsesWith(LI, AvailableVal);
10950 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op)) {
10951 const Value *GEPI0 = GEPI->getOperand(0);
10952 // TODO: Consider a target hook for valid address spaces for this xform.
10953 if (isa<ConstantPointerNull>(GEPI0) &&
10954 cast<PointerType>(GEPI0->getType())->getAddressSpace() == 0) {
10955 // Insert a new store to null instruction before the load to indicate
10956 // that this code is not reachable. We do this instead of inserting
10957 // an unreachable instruction directly because we cannot modify the
10959 new StoreInst(UndefValue::get(LI.getType()),
10960 Constant::getNullValue(Op->getType()), &LI);
10961 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
10965 if (Constant *C = dyn_cast<Constant>(Op)) {
10966 // load null/undef -> undef
10967 // TODO: Consider a target hook for valid address spaces for this xform.
10968 if (isa<UndefValue>(C) || (C->isNullValue() &&
10969 cast<PointerType>(Op->getType())->getAddressSpace() == 0)) {
10970 // Insert a new store to null instruction before the load to indicate that
10971 // this code is not reachable. We do this instead of inserting an
10972 // unreachable instruction directly because we cannot modify the CFG.
10973 new StoreInst(UndefValue::get(LI.getType()),
10974 Constant::getNullValue(Op->getType()), &LI);
10975 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
10978 // Instcombine load (constant global) into the value loaded.
10979 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Op))
10980 if (GV->isConstant() && !GV->isDeclaration())
10981 return ReplaceInstUsesWith(LI, GV->getInitializer());
10983 // Instcombine load (constantexpr_GEP global, 0, ...) into the value loaded.
10984 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Op)) {
10985 if (CE->getOpcode() == Instruction::GetElementPtr) {
10986 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(CE->getOperand(0)))
10987 if (GV->isConstant() && !GV->isDeclaration())
10989 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE))
10990 return ReplaceInstUsesWith(LI, V);
10991 if (CE->getOperand(0)->isNullValue()) {
10992 // Insert a new store to null instruction before the load to indicate
10993 // that this code is not reachable. We do this instead of inserting
10994 // an unreachable instruction directly because we cannot modify the
10996 new StoreInst(UndefValue::get(LI.getType()),
10997 Constant::getNullValue(Op->getType()), &LI);
10998 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
11001 } else if (CE->isCast()) {
11002 if (Instruction *Res = InstCombineLoadCast(*this, LI, TD))
11008 // If this load comes from anywhere in a constant global, and if the global
11009 // is all undef or zero, we know what it loads.
11010 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Op->getUnderlyingObject())){
11011 if (GV->isConstant() && GV->hasInitializer()) {
11012 if (GV->getInitializer()->isNullValue())
11013 return ReplaceInstUsesWith(LI, Constant::getNullValue(LI.getType()));
11014 else if (isa<UndefValue>(GV->getInitializer()))
11015 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
11019 if (Op->hasOneUse()) {
11020 // Change select and PHI nodes to select values instead of addresses: this
11021 // helps alias analysis out a lot, allows many others simplifications, and
11022 // exposes redundancy in the code.
11024 // Note that we cannot do the transformation unless we know that the
11025 // introduced loads cannot trap! Something like this is valid as long as
11026 // the condition is always false: load (select bool %C, int* null, int* %G),
11027 // but it would not be valid if we transformed it to load from null
11028 // unconditionally.
11030 if (SelectInst *SI = dyn_cast<SelectInst>(Op)) {
11031 // load (select (Cond, &V1, &V2)) --> select(Cond, load &V1, load &V2).
11032 if (isSafeToLoadUnconditionally(SI->getOperand(1), SI) &&
11033 isSafeToLoadUnconditionally(SI->getOperand(2), SI)) {
11034 Value *V1 = InsertNewInstBefore(new LoadInst(SI->getOperand(1),
11035 SI->getOperand(1)->getName()+".val"), LI);
11036 Value *V2 = InsertNewInstBefore(new LoadInst(SI->getOperand(2),
11037 SI->getOperand(2)->getName()+".val"), LI);
11038 return SelectInst::Create(SI->getCondition(), V1, V2);
11041 // load (select (cond, null, P)) -> load P
11042 if (Constant *C = dyn_cast<Constant>(SI->getOperand(1)))
11043 if (C->isNullValue()) {
11044 LI.setOperand(0, SI->getOperand(2));
11048 // load (select (cond, P, null)) -> load P
11049 if (Constant *C = dyn_cast<Constant>(SI->getOperand(2)))
11050 if (C->isNullValue()) {
11051 LI.setOperand(0, SI->getOperand(1));
11059 /// InstCombineStoreToCast - Fold store V, (cast P) -> store (cast V), P
11061 static Instruction *InstCombineStoreToCast(InstCombiner &IC, StoreInst &SI) {
11062 User *CI = cast<User>(SI.getOperand(1));
11063 Value *CastOp = CI->getOperand(0);
11065 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
11066 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
11067 const Type *SrcPTy = SrcTy->getElementType();
11069 if (DestPTy->isInteger() || isa<PointerType>(DestPTy)) {
11070 // If the source is an array, the code below will not succeed. Check to
11071 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
11073 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
11074 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
11075 if (ASrcTy->getNumElements() != 0) {
11077 Idxs[0] = Idxs[1] = Constant::getNullValue(Type::Int32Ty);
11078 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs, 2);
11079 SrcTy = cast<PointerType>(CastOp->getType());
11080 SrcPTy = SrcTy->getElementType();
11083 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy)) &&
11084 IC.getTargetData().getTypeSizeInBits(SrcPTy) ==
11085 IC.getTargetData().getTypeSizeInBits(DestPTy)) {
11087 // Okay, we are casting from one integer or pointer type to another of
11088 // the same size. Instead of casting the pointer before
11089 // the store, cast the value to be stored.
11091 Value *SIOp0 = SI.getOperand(0);
11092 Instruction::CastOps opcode = Instruction::BitCast;
11093 const Type* CastSrcTy = SIOp0->getType();
11094 const Type* CastDstTy = SrcPTy;
11095 if (isa<PointerType>(CastDstTy)) {
11096 if (CastSrcTy->isInteger())
11097 opcode = Instruction::IntToPtr;
11098 } else if (isa<IntegerType>(CastDstTy)) {
11099 if (isa<PointerType>(SIOp0->getType()))
11100 opcode = Instruction::PtrToInt;
11102 if (Constant *C = dyn_cast<Constant>(SIOp0))
11103 NewCast = ConstantExpr::getCast(opcode, C, CastDstTy);
11105 NewCast = IC.InsertNewInstBefore(
11106 CastInst::Create(opcode, SIOp0, CastDstTy, SIOp0->getName()+".c"),
11108 return new StoreInst(NewCast, CastOp);
11115 /// equivalentAddressValues - Test if A and B will obviously have the same
11116 /// value. This includes recognizing that %t0 and %t1 will have the same
11117 /// value in code like this:
11118 /// %t0 = getelementptr @a, 0, 3
11119 /// store i32 0, i32* %t0
11120 /// %t1 = getelementptr @a, 0, 3
11121 /// %t2 = load i32* %t1
11123 static bool equivalentAddressValues(Value *A, Value *B) {
11124 // Test if the values are trivially equivalent.
11125 if (A == B) return true;
11127 // Test if the values come form identical arithmetic instructions.
11128 if (isa<BinaryOperator>(A) ||
11129 isa<CastInst>(A) ||
11131 isa<GetElementPtrInst>(A))
11132 if (Instruction *BI = dyn_cast<Instruction>(B))
11133 if (cast<Instruction>(A)->isIdenticalTo(BI))
11136 // Otherwise they may not be equivalent.
11140 Instruction *InstCombiner::visitStoreInst(StoreInst &SI) {
11141 Value *Val = SI.getOperand(0);
11142 Value *Ptr = SI.getOperand(1);
11144 if (isa<UndefValue>(Ptr)) { // store X, undef -> noop (even if volatile)
11145 EraseInstFromFunction(SI);
11150 // If the RHS is an alloca with a single use, zapify the store, making the
11152 if (Ptr->hasOneUse() && !SI.isVolatile()) {
11153 if (isa<AllocaInst>(Ptr)) {
11154 EraseInstFromFunction(SI);
11159 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr))
11160 if (isa<AllocaInst>(GEP->getOperand(0)) &&
11161 GEP->getOperand(0)->hasOneUse()) {
11162 EraseInstFromFunction(SI);
11168 // Attempt to improve the alignment.
11169 unsigned KnownAlign = GetOrEnforceKnownAlignment(Ptr);
11171 (SI.getAlignment() == 0 ? TD->getABITypeAlignment(Val->getType()) :
11172 SI.getAlignment()))
11173 SI.setAlignment(KnownAlign);
11175 // Do really simple DSE, to catch cases where there are several consequtive
11176 // stores to the same location, separated by a few arithmetic operations. This
11177 // situation often occurs with bitfield accesses.
11178 BasicBlock::iterator BBI = &SI;
11179 for (unsigned ScanInsts = 6; BBI != SI.getParent()->begin() && ScanInsts;
11183 if (StoreInst *PrevSI = dyn_cast<StoreInst>(BBI)) {
11184 // Prev store isn't volatile, and stores to the same location?
11185 if (!PrevSI->isVolatile() &&equivalentAddressValues(PrevSI->getOperand(1),
11186 SI.getOperand(1))) {
11189 EraseInstFromFunction(*PrevSI);
11195 // If this is a load, we have to stop. However, if the loaded value is from
11196 // the pointer we're loading and is producing the pointer we're storing,
11197 // then *this* store is dead (X = load P; store X -> P).
11198 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
11199 if (LI == Val && equivalentAddressValues(LI->getOperand(0), Ptr) &&
11200 !SI.isVolatile()) {
11201 EraseInstFromFunction(SI);
11205 // Otherwise, this is a load from some other location. Stores before it
11206 // may not be dead.
11210 // Don't skip over loads or things that can modify memory.
11211 if (BBI->mayWriteToMemory() || BBI->mayReadFromMemory())
11216 if (SI.isVolatile()) return 0; // Don't hack volatile stores.
11218 // store X, null -> turns into 'unreachable' in SimplifyCFG
11219 if (isa<ConstantPointerNull>(Ptr)) {
11220 if (!isa<UndefValue>(Val)) {
11221 SI.setOperand(0, UndefValue::get(Val->getType()));
11222 if (Instruction *U = dyn_cast<Instruction>(Val))
11223 AddToWorkList(U); // Dropped a use.
11226 return 0; // Do not modify these!
11229 // store undef, Ptr -> noop
11230 if (isa<UndefValue>(Val)) {
11231 EraseInstFromFunction(SI);
11236 // If the pointer destination is a cast, see if we can fold the cast into the
11238 if (isa<CastInst>(Ptr))
11239 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
11241 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
11243 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
11247 // If this store is the last instruction in the basic block, and if the block
11248 // ends with an unconditional branch, try to move it to the successor block.
11250 if (BranchInst *BI = dyn_cast<BranchInst>(BBI))
11251 if (BI->isUnconditional())
11252 if (SimplifyStoreAtEndOfBlock(SI))
11253 return 0; // xform done!
11258 /// SimplifyStoreAtEndOfBlock - Turn things like:
11259 /// if () { *P = v1; } else { *P = v2 }
11260 /// into a phi node with a store in the successor.
11262 /// Simplify things like:
11263 /// *P = v1; if () { *P = v2; }
11264 /// into a phi node with a store in the successor.
11266 bool InstCombiner::SimplifyStoreAtEndOfBlock(StoreInst &SI) {
11267 BasicBlock *StoreBB = SI.getParent();
11269 // Check to see if the successor block has exactly two incoming edges. If
11270 // so, see if the other predecessor contains a store to the same location.
11271 // if so, insert a PHI node (if needed) and move the stores down.
11272 BasicBlock *DestBB = StoreBB->getTerminator()->getSuccessor(0);
11274 // Determine whether Dest has exactly two predecessors and, if so, compute
11275 // the other predecessor.
11276 pred_iterator PI = pred_begin(DestBB);
11277 BasicBlock *OtherBB = 0;
11278 if (*PI != StoreBB)
11281 if (PI == pred_end(DestBB))
11284 if (*PI != StoreBB) {
11289 if (++PI != pred_end(DestBB))
11292 // Bail out if all the relevant blocks aren't distinct (this can happen,
11293 // for example, if SI is in an infinite loop)
11294 if (StoreBB == DestBB || OtherBB == DestBB)
11297 // Verify that the other block ends in a branch and is not otherwise empty.
11298 BasicBlock::iterator BBI = OtherBB->getTerminator();
11299 BranchInst *OtherBr = dyn_cast<BranchInst>(BBI);
11300 if (!OtherBr || BBI == OtherBB->begin())
11303 // If the other block ends in an unconditional branch, check for the 'if then
11304 // else' case. there is an instruction before the branch.
11305 StoreInst *OtherStore = 0;
11306 if (OtherBr->isUnconditional()) {
11307 // If this isn't a store, or isn't a store to the same location, bail out.
11309 OtherStore = dyn_cast<StoreInst>(BBI);
11310 if (!OtherStore || OtherStore->getOperand(1) != SI.getOperand(1))
11313 // Otherwise, the other block ended with a conditional branch. If one of the
11314 // destinations is StoreBB, then we have the if/then case.
11315 if (OtherBr->getSuccessor(0) != StoreBB &&
11316 OtherBr->getSuccessor(1) != StoreBB)
11319 // Okay, we know that OtherBr now goes to Dest and StoreBB, so this is an
11320 // if/then triangle. See if there is a store to the same ptr as SI that
11321 // lives in OtherBB.
11323 // Check to see if we find the matching store.
11324 if ((OtherStore = dyn_cast<StoreInst>(BBI))) {
11325 if (OtherStore->getOperand(1) != SI.getOperand(1))
11329 // If we find something that may be using or overwriting the stored
11330 // value, or if we run out of instructions, we can't do the xform.
11331 if (BBI->mayReadFromMemory() || BBI->mayWriteToMemory() ||
11332 BBI == OtherBB->begin())
11336 // In order to eliminate the store in OtherBr, we have to
11337 // make sure nothing reads or overwrites the stored value in
11339 for (BasicBlock::iterator I = StoreBB->begin(); &*I != &SI; ++I) {
11340 // FIXME: This should really be AA driven.
11341 if (I->mayReadFromMemory() || I->mayWriteToMemory())
11346 // Insert a PHI node now if we need it.
11347 Value *MergedVal = OtherStore->getOperand(0);
11348 if (MergedVal != SI.getOperand(0)) {
11349 PHINode *PN = PHINode::Create(MergedVal->getType(), "storemerge");
11350 PN->reserveOperandSpace(2);
11351 PN->addIncoming(SI.getOperand(0), SI.getParent());
11352 PN->addIncoming(OtherStore->getOperand(0), OtherBB);
11353 MergedVal = InsertNewInstBefore(PN, DestBB->front());
11356 // Advance to a place where it is safe to insert the new store and
11358 BBI = DestBB->getFirstNonPHI();
11359 InsertNewInstBefore(new StoreInst(MergedVal, SI.getOperand(1),
11360 OtherStore->isVolatile()), *BBI);
11362 // Nuke the old stores.
11363 EraseInstFromFunction(SI);
11364 EraseInstFromFunction(*OtherStore);
11370 Instruction *InstCombiner::visitBranchInst(BranchInst &BI) {
11371 // Change br (not X), label True, label False to: br X, label False, True
11373 BasicBlock *TrueDest;
11374 BasicBlock *FalseDest;
11375 if (match(&BI, m_Br(m_Not(m_Value(X)), TrueDest, FalseDest)) &&
11376 !isa<Constant>(X)) {
11377 // Swap Destinations and condition...
11378 BI.setCondition(X);
11379 BI.setSuccessor(0, FalseDest);
11380 BI.setSuccessor(1, TrueDest);
11384 // Cannonicalize fcmp_one -> fcmp_oeq
11385 FCmpInst::Predicate FPred; Value *Y;
11386 if (match(&BI, m_Br(m_FCmp(FPred, m_Value(X), m_Value(Y)),
11387 TrueDest, FalseDest)))
11388 if ((FPred == FCmpInst::FCMP_ONE || FPred == FCmpInst::FCMP_OLE ||
11389 FPred == FCmpInst::FCMP_OGE) && BI.getCondition()->hasOneUse()) {
11390 FCmpInst *I = cast<FCmpInst>(BI.getCondition());
11391 FCmpInst::Predicate NewPred = FCmpInst::getInversePredicate(FPred);
11392 Instruction *NewSCC = new FCmpInst(NewPred, X, Y, "", I);
11393 NewSCC->takeName(I);
11394 // Swap Destinations and condition...
11395 BI.setCondition(NewSCC);
11396 BI.setSuccessor(0, FalseDest);
11397 BI.setSuccessor(1, TrueDest);
11398 RemoveFromWorkList(I);
11399 I->eraseFromParent();
11400 AddToWorkList(NewSCC);
11404 // Cannonicalize icmp_ne -> icmp_eq
11405 ICmpInst::Predicate IPred;
11406 if (match(&BI, m_Br(m_ICmp(IPred, m_Value(X), m_Value(Y)),
11407 TrueDest, FalseDest)))
11408 if ((IPred == ICmpInst::ICMP_NE || IPred == ICmpInst::ICMP_ULE ||
11409 IPred == ICmpInst::ICMP_SLE || IPred == ICmpInst::ICMP_UGE ||
11410 IPred == ICmpInst::ICMP_SGE) && BI.getCondition()->hasOneUse()) {
11411 ICmpInst *I = cast<ICmpInst>(BI.getCondition());
11412 ICmpInst::Predicate NewPred = ICmpInst::getInversePredicate(IPred);
11413 Instruction *NewSCC = new ICmpInst(NewPred, X, Y, "", I);
11414 NewSCC->takeName(I);
11415 // Swap Destinations and condition...
11416 BI.setCondition(NewSCC);
11417 BI.setSuccessor(0, FalseDest);
11418 BI.setSuccessor(1, TrueDest);
11419 RemoveFromWorkList(I);
11420 I->eraseFromParent();;
11421 AddToWorkList(NewSCC);
11428 Instruction *InstCombiner::visitSwitchInst(SwitchInst &SI) {
11429 Value *Cond = SI.getCondition();
11430 if (Instruction *I = dyn_cast<Instruction>(Cond)) {
11431 if (I->getOpcode() == Instruction::Add)
11432 if (ConstantInt *AddRHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
11433 // change 'switch (X+4) case 1:' into 'switch (X) case -3'
11434 for (unsigned i = 2, e = SI.getNumOperands(); i != e; i += 2)
11435 SI.setOperand(i,ConstantExpr::getSub(cast<Constant>(SI.getOperand(i)),
11437 SI.setOperand(0, I->getOperand(0));
11445 Instruction *InstCombiner::visitExtractValueInst(ExtractValueInst &EV) {
11446 Value *Agg = EV.getAggregateOperand();
11448 if (!EV.hasIndices())
11449 return ReplaceInstUsesWith(EV, Agg);
11451 if (Constant *C = dyn_cast<Constant>(Agg)) {
11452 if (isa<UndefValue>(C))
11453 return ReplaceInstUsesWith(EV, UndefValue::get(EV.getType()));
11455 if (isa<ConstantAggregateZero>(C))
11456 return ReplaceInstUsesWith(EV, Constant::getNullValue(EV.getType()));
11458 if (isa<ConstantArray>(C) || isa<ConstantStruct>(C)) {
11459 // Extract the element indexed by the first index out of the constant
11460 Value *V = C->getOperand(*EV.idx_begin());
11461 if (EV.getNumIndices() > 1)
11462 // Extract the remaining indices out of the constant indexed by the
11464 return ExtractValueInst::Create(V, EV.idx_begin() + 1, EV.idx_end());
11466 return ReplaceInstUsesWith(EV, V);
11468 return 0; // Can't handle other constants
11470 if (InsertValueInst *IV = dyn_cast<InsertValueInst>(Agg)) {
11471 // We're extracting from an insertvalue instruction, compare the indices
11472 const unsigned *exti, *exte, *insi, *inse;
11473 for (exti = EV.idx_begin(), insi = IV->idx_begin(),
11474 exte = EV.idx_end(), inse = IV->idx_end();
11475 exti != exte && insi != inse;
11477 if (*insi != *exti)
11478 // The insert and extract both reference distinctly different elements.
11479 // This means the extract is not influenced by the insert, and we can
11480 // replace the aggregate operand of the extract with the aggregate
11481 // operand of the insert. i.e., replace
11482 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1
11483 // %E = extractvalue { i32, { i32 } } %I, 0
11485 // %E = extractvalue { i32, { i32 } } %A, 0
11486 return ExtractValueInst::Create(IV->getAggregateOperand(),
11487 EV.idx_begin(), EV.idx_end());
11489 if (exti == exte && insi == inse)
11490 // Both iterators are at the end: Index lists are identical. Replace
11491 // %B = insertvalue { i32, { i32 } } %A, i32 42, 1, 0
11492 // %C = extractvalue { i32, { i32 } } %B, 1, 0
11494 return ReplaceInstUsesWith(EV, IV->getInsertedValueOperand());
11495 if (exti == exte) {
11496 // The extract list is a prefix of the insert list. i.e. replace
11497 // %I = insertvalue { i32, { i32 } } %A, i32 42, 1, 0
11498 // %E = extractvalue { i32, { i32 } } %I, 1
11500 // %X = extractvalue { i32, { i32 } } %A, 1
11501 // %E = insertvalue { i32 } %X, i32 42, 0
11502 // by switching the order of the insert and extract (though the
11503 // insertvalue should be left in, since it may have other uses).
11504 Value *NewEV = InsertNewInstBefore(
11505 ExtractValueInst::Create(IV->getAggregateOperand(),
11506 EV.idx_begin(), EV.idx_end()),
11508 return InsertValueInst::Create(NewEV, IV->getInsertedValueOperand(),
11512 // The insert list is a prefix of the extract list
11513 // We can simply remove the common indices from the extract and make it
11514 // operate on the inserted value instead of the insertvalue result.
11516 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1
11517 // %E = extractvalue { i32, { i32 } } %I, 1, 0
11519 // %E extractvalue { i32 } { i32 42 }, 0
11520 return ExtractValueInst::Create(IV->getInsertedValueOperand(),
11523 // Can't simplify extracts from other values. Note that nested extracts are
11524 // already simplified implicitely by the above (extract ( extract (insert) )
11525 // will be translated into extract ( insert ( extract ) ) first and then just
11526 // the value inserted, if appropriate).
11530 /// CheapToScalarize - Return true if the value is cheaper to scalarize than it
11531 /// is to leave as a vector operation.
11532 static bool CheapToScalarize(Value *V, bool isConstant) {
11533 if (isa<ConstantAggregateZero>(V))
11535 if (ConstantVector *C = dyn_cast<ConstantVector>(V)) {
11536 if (isConstant) return true;
11537 // If all elts are the same, we can extract.
11538 Constant *Op0 = C->getOperand(0);
11539 for (unsigned i = 1; i < C->getNumOperands(); ++i)
11540 if (C->getOperand(i) != Op0)
11544 Instruction *I = dyn_cast<Instruction>(V);
11545 if (!I) return false;
11547 // Insert element gets simplified to the inserted element or is deleted if
11548 // this is constant idx extract element and its a constant idx insertelt.
11549 if (I->getOpcode() == Instruction::InsertElement && isConstant &&
11550 isa<ConstantInt>(I->getOperand(2)))
11552 if (I->getOpcode() == Instruction::Load && I->hasOneUse())
11554 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I))
11555 if (BO->hasOneUse() &&
11556 (CheapToScalarize(BO->getOperand(0), isConstant) ||
11557 CheapToScalarize(BO->getOperand(1), isConstant)))
11559 if (CmpInst *CI = dyn_cast<CmpInst>(I))
11560 if (CI->hasOneUse() &&
11561 (CheapToScalarize(CI->getOperand(0), isConstant) ||
11562 CheapToScalarize(CI->getOperand(1), isConstant)))
11568 /// Read and decode a shufflevector mask.
11570 /// It turns undef elements into values that are larger than the number of
11571 /// elements in the input.
11572 static std::vector<unsigned> getShuffleMask(const ShuffleVectorInst *SVI) {
11573 unsigned NElts = SVI->getType()->getNumElements();
11574 if (isa<ConstantAggregateZero>(SVI->getOperand(2)))
11575 return std::vector<unsigned>(NElts, 0);
11576 if (isa<UndefValue>(SVI->getOperand(2)))
11577 return std::vector<unsigned>(NElts, 2*NElts);
11579 std::vector<unsigned> Result;
11580 const ConstantVector *CP = cast<ConstantVector>(SVI->getOperand(2));
11581 for (User::const_op_iterator i = CP->op_begin(), e = CP->op_end(); i!=e; ++i)
11582 if (isa<UndefValue>(*i))
11583 Result.push_back(NElts*2); // undef -> 8
11585 Result.push_back(cast<ConstantInt>(*i)->getZExtValue());
11589 /// FindScalarElement - Given a vector and an element number, see if the scalar
11590 /// value is already around as a register, for example if it were inserted then
11591 /// extracted from the vector.
11592 static Value *FindScalarElement(Value *V, unsigned EltNo) {
11593 assert(isa<VectorType>(V->getType()) && "Not looking at a vector?");
11594 const VectorType *PTy = cast<VectorType>(V->getType());
11595 unsigned Width = PTy->getNumElements();
11596 if (EltNo >= Width) // Out of range access.
11597 return UndefValue::get(PTy->getElementType());
11599 if (isa<UndefValue>(V))
11600 return UndefValue::get(PTy->getElementType());
11601 else if (isa<ConstantAggregateZero>(V))
11602 return Constant::getNullValue(PTy->getElementType());
11603 else if (ConstantVector *CP = dyn_cast<ConstantVector>(V))
11604 return CP->getOperand(EltNo);
11605 else if (InsertElementInst *III = dyn_cast<InsertElementInst>(V)) {
11606 // If this is an insert to a variable element, we don't know what it is.
11607 if (!isa<ConstantInt>(III->getOperand(2)))
11609 unsigned IIElt = cast<ConstantInt>(III->getOperand(2))->getZExtValue();
11611 // If this is an insert to the element we are looking for, return the
11613 if (EltNo == IIElt)
11614 return III->getOperand(1);
11616 // Otherwise, the insertelement doesn't modify the value, recurse on its
11618 return FindScalarElement(III->getOperand(0), EltNo);
11619 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(V)) {
11620 unsigned LHSWidth =
11621 cast<VectorType>(SVI->getOperand(0)->getType())->getNumElements();
11622 unsigned InEl = getShuffleMask(SVI)[EltNo];
11623 if (InEl < LHSWidth)
11624 return FindScalarElement(SVI->getOperand(0), InEl);
11625 else if (InEl < LHSWidth*2)
11626 return FindScalarElement(SVI->getOperand(1), InEl - LHSWidth);
11628 return UndefValue::get(PTy->getElementType());
11631 // Otherwise, we don't know.
11635 Instruction *InstCombiner::visitExtractElementInst(ExtractElementInst &EI) {
11636 // If vector val is undef, replace extract with scalar undef.
11637 if (isa<UndefValue>(EI.getOperand(0)))
11638 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
11640 // If vector val is constant 0, replace extract with scalar 0.
11641 if (isa<ConstantAggregateZero>(EI.getOperand(0)))
11642 return ReplaceInstUsesWith(EI, Constant::getNullValue(EI.getType()));
11644 if (ConstantVector *C = dyn_cast<ConstantVector>(EI.getOperand(0))) {
11645 // If vector val is constant with all elements the same, replace EI with
11646 // that element. When the elements are not identical, we cannot replace yet
11647 // (we do that below, but only when the index is constant).
11648 Constant *op0 = C->getOperand(0);
11649 for (unsigned i = 1; i < C->getNumOperands(); ++i)
11650 if (C->getOperand(i) != op0) {
11655 return ReplaceInstUsesWith(EI, op0);
11658 // If extracting a specified index from the vector, see if we can recursively
11659 // find a previously computed scalar that was inserted into the vector.
11660 if (ConstantInt *IdxC = dyn_cast<ConstantInt>(EI.getOperand(1))) {
11661 unsigned IndexVal = IdxC->getZExtValue();
11662 unsigned VectorWidth =
11663 cast<VectorType>(EI.getOperand(0)->getType())->getNumElements();
11665 // If this is extracting an invalid index, turn this into undef, to avoid
11666 // crashing the code below.
11667 if (IndexVal >= VectorWidth)
11668 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
11670 // This instruction only demands the single element from the input vector.
11671 // If the input vector has a single use, simplify it based on this use
11673 if (EI.getOperand(0)->hasOneUse() && VectorWidth != 1) {
11674 uint64_t UndefElts;
11675 if (Value *V = SimplifyDemandedVectorElts(EI.getOperand(0),
11678 EI.setOperand(0, V);
11683 if (Value *Elt = FindScalarElement(EI.getOperand(0), IndexVal))
11684 return ReplaceInstUsesWith(EI, Elt);
11686 // If the this extractelement is directly using a bitcast from a vector of
11687 // the same number of elements, see if we can find the source element from
11688 // it. In this case, we will end up needing to bitcast the scalars.
11689 if (BitCastInst *BCI = dyn_cast<BitCastInst>(EI.getOperand(0))) {
11690 if (const VectorType *VT =
11691 dyn_cast<VectorType>(BCI->getOperand(0)->getType()))
11692 if (VT->getNumElements() == VectorWidth)
11693 if (Value *Elt = FindScalarElement(BCI->getOperand(0), IndexVal))
11694 return new BitCastInst(Elt, EI.getType());
11698 if (Instruction *I = dyn_cast<Instruction>(EI.getOperand(0))) {
11699 if (I->hasOneUse()) {
11700 // Push extractelement into predecessor operation if legal and
11701 // profitable to do so
11702 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
11703 bool isConstantElt = isa<ConstantInt>(EI.getOperand(1));
11704 if (CheapToScalarize(BO, isConstantElt)) {
11705 ExtractElementInst *newEI0 =
11706 new ExtractElementInst(BO->getOperand(0), EI.getOperand(1),
11707 EI.getName()+".lhs");
11708 ExtractElementInst *newEI1 =
11709 new ExtractElementInst(BO->getOperand(1), EI.getOperand(1),
11710 EI.getName()+".rhs");
11711 InsertNewInstBefore(newEI0, EI);
11712 InsertNewInstBefore(newEI1, EI);
11713 return BinaryOperator::Create(BO->getOpcode(), newEI0, newEI1);
11715 } else if (isa<LoadInst>(I)) {
11717 cast<PointerType>(I->getOperand(0)->getType())->getAddressSpace();
11718 Value *Ptr = InsertBitCastBefore(I->getOperand(0),
11719 PointerType::get(EI.getType(), AS),EI);
11720 GetElementPtrInst *GEP =
11721 GetElementPtrInst::Create(Ptr, EI.getOperand(1), I->getName()+".gep");
11722 InsertNewInstBefore(GEP, EI);
11723 return new LoadInst(GEP);
11726 if (InsertElementInst *IE = dyn_cast<InsertElementInst>(I)) {
11727 // Extracting the inserted element?
11728 if (IE->getOperand(2) == EI.getOperand(1))
11729 return ReplaceInstUsesWith(EI, IE->getOperand(1));
11730 // If the inserted and extracted elements are constants, they must not
11731 // be the same value, extract from the pre-inserted value instead.
11732 if (isa<Constant>(IE->getOperand(2)) &&
11733 isa<Constant>(EI.getOperand(1))) {
11734 AddUsesToWorkList(EI);
11735 EI.setOperand(0, IE->getOperand(0));
11738 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I)) {
11739 // If this is extracting an element from a shufflevector, figure out where
11740 // it came from and extract from the appropriate input element instead.
11741 if (ConstantInt *Elt = dyn_cast<ConstantInt>(EI.getOperand(1))) {
11742 unsigned SrcIdx = getShuffleMask(SVI)[Elt->getZExtValue()];
11744 unsigned LHSWidth =
11745 cast<VectorType>(SVI->getOperand(0)->getType())->getNumElements();
11747 if (SrcIdx < LHSWidth)
11748 Src = SVI->getOperand(0);
11749 else if (SrcIdx < LHSWidth*2) {
11750 SrcIdx -= LHSWidth;
11751 Src = SVI->getOperand(1);
11753 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
11755 return new ExtractElementInst(Src, SrcIdx);
11762 /// CollectSingleShuffleElements - If V is a shuffle of values that ONLY returns
11763 /// elements from either LHS or RHS, return the shuffle mask and true.
11764 /// Otherwise, return false.
11765 static bool CollectSingleShuffleElements(Value *V, Value *LHS, Value *RHS,
11766 std::vector<Constant*> &Mask) {
11767 assert(V->getType() == LHS->getType() && V->getType() == RHS->getType() &&
11768 "Invalid CollectSingleShuffleElements");
11769 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
11771 if (isa<UndefValue>(V)) {
11772 Mask.assign(NumElts, UndefValue::get(Type::Int32Ty));
11774 } else if (V == LHS) {
11775 for (unsigned i = 0; i != NumElts; ++i)
11776 Mask.push_back(ConstantInt::get(Type::Int32Ty, i));
11778 } else if (V == RHS) {
11779 for (unsigned i = 0; i != NumElts; ++i)
11780 Mask.push_back(ConstantInt::get(Type::Int32Ty, i+NumElts));
11782 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
11783 // If this is an insert of an extract from some other vector, include it.
11784 Value *VecOp = IEI->getOperand(0);
11785 Value *ScalarOp = IEI->getOperand(1);
11786 Value *IdxOp = IEI->getOperand(2);
11788 if (!isa<ConstantInt>(IdxOp))
11790 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
11792 if (isa<UndefValue>(ScalarOp)) { // inserting undef into vector.
11793 // Okay, we can handle this if the vector we are insertinting into is
11794 // transitively ok.
11795 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
11796 // If so, update the mask to reflect the inserted undef.
11797 Mask[InsertedIdx] = UndefValue::get(Type::Int32Ty);
11800 } else if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)){
11801 if (isa<ConstantInt>(EI->getOperand(1)) &&
11802 EI->getOperand(0)->getType() == V->getType()) {
11803 unsigned ExtractedIdx =
11804 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
11806 // This must be extracting from either LHS or RHS.
11807 if (EI->getOperand(0) == LHS || EI->getOperand(0) == RHS) {
11808 // Okay, we can handle this if the vector we are insertinting into is
11809 // transitively ok.
11810 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
11811 // If so, update the mask to reflect the inserted value.
11812 if (EI->getOperand(0) == LHS) {
11813 Mask[InsertedIdx % NumElts] =
11814 ConstantInt::get(Type::Int32Ty, ExtractedIdx);
11816 assert(EI->getOperand(0) == RHS);
11817 Mask[InsertedIdx % NumElts] =
11818 ConstantInt::get(Type::Int32Ty, ExtractedIdx+NumElts);
11827 // TODO: Handle shufflevector here!
11832 /// CollectShuffleElements - We are building a shuffle of V, using RHS as the
11833 /// RHS of the shuffle instruction, if it is not null. Return a shuffle mask
11834 /// that computes V and the LHS value of the shuffle.
11835 static Value *CollectShuffleElements(Value *V, std::vector<Constant*> &Mask,
11837 assert(isa<VectorType>(V->getType()) &&
11838 (RHS == 0 || V->getType() == RHS->getType()) &&
11839 "Invalid shuffle!");
11840 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
11842 if (isa<UndefValue>(V)) {
11843 Mask.assign(NumElts, UndefValue::get(Type::Int32Ty));
11845 } else if (isa<ConstantAggregateZero>(V)) {
11846 Mask.assign(NumElts, ConstantInt::get(Type::Int32Ty, 0));
11848 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
11849 // If this is an insert of an extract from some other vector, include it.
11850 Value *VecOp = IEI->getOperand(0);
11851 Value *ScalarOp = IEI->getOperand(1);
11852 Value *IdxOp = IEI->getOperand(2);
11854 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
11855 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
11856 EI->getOperand(0)->getType() == V->getType()) {
11857 unsigned ExtractedIdx =
11858 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
11859 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
11861 // Either the extracted from or inserted into vector must be RHSVec,
11862 // otherwise we'd end up with a shuffle of three inputs.
11863 if (EI->getOperand(0) == RHS || RHS == 0) {
11864 RHS = EI->getOperand(0);
11865 Value *V = CollectShuffleElements(VecOp, Mask, RHS);
11866 Mask[InsertedIdx % NumElts] =
11867 ConstantInt::get(Type::Int32Ty, NumElts+ExtractedIdx);
11871 if (VecOp == RHS) {
11872 Value *V = CollectShuffleElements(EI->getOperand(0), Mask, RHS);
11873 // Everything but the extracted element is replaced with the RHS.
11874 for (unsigned i = 0; i != NumElts; ++i) {
11875 if (i != InsertedIdx)
11876 Mask[i] = ConstantInt::get(Type::Int32Ty, NumElts+i);
11881 // If this insertelement is a chain that comes from exactly these two
11882 // vectors, return the vector and the effective shuffle.
11883 if (CollectSingleShuffleElements(IEI, EI->getOperand(0), RHS, Mask))
11884 return EI->getOperand(0);
11889 // TODO: Handle shufflevector here!
11891 // Otherwise, can't do anything fancy. Return an identity vector.
11892 for (unsigned i = 0; i != NumElts; ++i)
11893 Mask.push_back(ConstantInt::get(Type::Int32Ty, i));
11897 Instruction *InstCombiner::visitInsertElementInst(InsertElementInst &IE) {
11898 Value *VecOp = IE.getOperand(0);
11899 Value *ScalarOp = IE.getOperand(1);
11900 Value *IdxOp = IE.getOperand(2);
11902 // Inserting an undef or into an undefined place, remove this.
11903 if (isa<UndefValue>(ScalarOp) || isa<UndefValue>(IdxOp))
11904 ReplaceInstUsesWith(IE, VecOp);
11906 // If the inserted element was extracted from some other vector, and if the
11907 // indexes are constant, try to turn this into a shufflevector operation.
11908 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
11909 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
11910 EI->getOperand(0)->getType() == IE.getType()) {
11911 unsigned NumVectorElts = IE.getType()->getNumElements();
11912 unsigned ExtractedIdx =
11913 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
11914 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
11916 if (ExtractedIdx >= NumVectorElts) // Out of range extract.
11917 return ReplaceInstUsesWith(IE, VecOp);
11919 if (InsertedIdx >= NumVectorElts) // Out of range insert.
11920 return ReplaceInstUsesWith(IE, UndefValue::get(IE.getType()));
11922 // If we are extracting a value from a vector, then inserting it right
11923 // back into the same place, just use the input vector.
11924 if (EI->getOperand(0) == VecOp && ExtractedIdx == InsertedIdx)
11925 return ReplaceInstUsesWith(IE, VecOp);
11927 // We could theoretically do this for ANY input. However, doing so could
11928 // turn chains of insertelement instructions into a chain of shufflevector
11929 // instructions, and right now we do not merge shufflevectors. As such,
11930 // only do this in a situation where it is clear that there is benefit.
11931 if (isa<UndefValue>(VecOp) || isa<ConstantAggregateZero>(VecOp)) {
11932 // Turn this into shuffle(EIOp0, VecOp, Mask). The result has all of
11933 // the values of VecOp, except then one read from EIOp0.
11934 // Build a new shuffle mask.
11935 std::vector<Constant*> Mask;
11936 if (isa<UndefValue>(VecOp))
11937 Mask.assign(NumVectorElts, UndefValue::get(Type::Int32Ty));
11939 assert(isa<ConstantAggregateZero>(VecOp) && "Unknown thing");
11940 Mask.assign(NumVectorElts, ConstantInt::get(Type::Int32Ty,
11943 Mask[InsertedIdx] = ConstantInt::get(Type::Int32Ty, ExtractedIdx);
11944 return new ShuffleVectorInst(EI->getOperand(0), VecOp,
11945 ConstantVector::get(Mask));
11948 // If this insertelement isn't used by some other insertelement, turn it
11949 // (and any insertelements it points to), into one big shuffle.
11950 if (!IE.hasOneUse() || !isa<InsertElementInst>(IE.use_back())) {
11951 std::vector<Constant*> Mask;
11953 Value *LHS = CollectShuffleElements(&IE, Mask, RHS);
11954 if (RHS == 0) RHS = UndefValue::get(LHS->getType());
11955 // We now have a shuffle of LHS, RHS, Mask.
11956 return new ShuffleVectorInst(LHS, RHS, ConstantVector::get(Mask));
11965 Instruction *InstCombiner::visitShuffleVectorInst(ShuffleVectorInst &SVI) {
11966 Value *LHS = SVI.getOperand(0);
11967 Value *RHS = SVI.getOperand(1);
11968 std::vector<unsigned> Mask = getShuffleMask(&SVI);
11970 bool MadeChange = false;
11972 // Undefined shuffle mask -> undefined value.
11973 if (isa<UndefValue>(SVI.getOperand(2)))
11974 return ReplaceInstUsesWith(SVI, UndefValue::get(SVI.getType()));
11976 uint64_t UndefElts;
11977 unsigned VWidth = cast<VectorType>(SVI.getType())->getNumElements();
11979 if (VWidth != cast<VectorType>(LHS->getType())->getNumElements())
11982 uint64_t AllOnesEltMask = ~0ULL >> (64-VWidth);
11983 if (VWidth <= 64 &&
11984 SimplifyDemandedVectorElts(&SVI, AllOnesEltMask, UndefElts)) {
11985 LHS = SVI.getOperand(0);
11986 RHS = SVI.getOperand(1);
11990 // Canonicalize shuffle(x ,x,mask) -> shuffle(x, undef,mask')
11991 // Canonicalize shuffle(undef,x,mask) -> shuffle(x, undef,mask').
11992 if (LHS == RHS || isa<UndefValue>(LHS)) {
11993 if (isa<UndefValue>(LHS) && LHS == RHS) {
11994 // shuffle(undef,undef,mask) -> undef.
11995 return ReplaceInstUsesWith(SVI, LHS);
11998 // Remap any references to RHS to use LHS.
11999 std::vector<Constant*> Elts;
12000 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
12001 if (Mask[i] >= 2*e)
12002 Elts.push_back(UndefValue::get(Type::Int32Ty));
12004 if ((Mask[i] >= e && isa<UndefValue>(RHS)) ||
12005 (Mask[i] < e && isa<UndefValue>(LHS))) {
12006 Mask[i] = 2*e; // Turn into undef.
12007 Elts.push_back(UndefValue::get(Type::Int32Ty));
12009 Mask[i] = Mask[i] % e; // Force to LHS.
12010 Elts.push_back(ConstantInt::get(Type::Int32Ty, Mask[i]));
12014 SVI.setOperand(0, SVI.getOperand(1));
12015 SVI.setOperand(1, UndefValue::get(RHS->getType()));
12016 SVI.setOperand(2, ConstantVector::get(Elts));
12017 LHS = SVI.getOperand(0);
12018 RHS = SVI.getOperand(1);
12022 // Analyze the shuffle, are the LHS or RHS and identity shuffles?
12023 bool isLHSID = true, isRHSID = true;
12025 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
12026 if (Mask[i] >= e*2) continue; // Ignore undef values.
12027 // Is this an identity shuffle of the LHS value?
12028 isLHSID &= (Mask[i] == i);
12030 // Is this an identity shuffle of the RHS value?
12031 isRHSID &= (Mask[i]-e == i);
12034 // Eliminate identity shuffles.
12035 if (isLHSID) return ReplaceInstUsesWith(SVI, LHS);
12036 if (isRHSID) return ReplaceInstUsesWith(SVI, RHS);
12038 // If the LHS is a shufflevector itself, see if we can combine it with this
12039 // one without producing an unusual shuffle. Here we are really conservative:
12040 // we are absolutely afraid of producing a shuffle mask not in the input
12041 // program, because the code gen may not be smart enough to turn a merged
12042 // shuffle into two specific shuffles: it may produce worse code. As such,
12043 // we only merge two shuffles if the result is one of the two input shuffle
12044 // masks. In this case, merging the shuffles just removes one instruction,
12045 // which we know is safe. This is good for things like turning:
12046 // (splat(splat)) -> splat.
12047 if (ShuffleVectorInst *LHSSVI = dyn_cast<ShuffleVectorInst>(LHS)) {
12048 if (isa<UndefValue>(RHS)) {
12049 std::vector<unsigned> LHSMask = getShuffleMask(LHSSVI);
12051 std::vector<unsigned> NewMask;
12052 for (unsigned i = 0, e = Mask.size(); i != e; ++i)
12053 if (Mask[i] >= 2*e)
12054 NewMask.push_back(2*e);
12056 NewMask.push_back(LHSMask[Mask[i]]);
12058 // If the result mask is equal to the src shuffle or this shuffle mask, do
12059 // the replacement.
12060 if (NewMask == LHSMask || NewMask == Mask) {
12061 std::vector<Constant*> Elts;
12062 for (unsigned i = 0, e = NewMask.size(); i != e; ++i) {
12063 if (NewMask[i] >= e*2) {
12064 Elts.push_back(UndefValue::get(Type::Int32Ty));
12066 Elts.push_back(ConstantInt::get(Type::Int32Ty, NewMask[i]));
12069 return new ShuffleVectorInst(LHSSVI->getOperand(0),
12070 LHSSVI->getOperand(1),
12071 ConstantVector::get(Elts));
12076 return MadeChange ? &SVI : 0;
12082 /// TryToSinkInstruction - Try to move the specified instruction from its
12083 /// current block into the beginning of DestBlock, which can only happen if it's
12084 /// safe to move the instruction past all of the instructions between it and the
12085 /// end of its block.
12086 static bool TryToSinkInstruction(Instruction *I, BasicBlock *DestBlock) {
12087 assert(I->hasOneUse() && "Invariants didn't hold!");
12089 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
12090 if (isa<PHINode>(I) || I->mayWriteToMemory() || isa<TerminatorInst>(I))
12093 // Do not sink alloca instructions out of the entry block.
12094 if (isa<AllocaInst>(I) && I->getParent() ==
12095 &DestBlock->getParent()->getEntryBlock())
12098 // We can only sink load instructions if there is nothing between the load and
12099 // the end of block that could change the value.
12100 if (I->mayReadFromMemory()) {
12101 for (BasicBlock::iterator Scan = I, E = I->getParent()->end();
12103 if (Scan->mayWriteToMemory())
12107 BasicBlock::iterator InsertPos = DestBlock->getFirstNonPHI();
12109 I->moveBefore(InsertPos);
12115 /// AddReachableCodeToWorklist - Walk the function in depth-first order, adding
12116 /// all reachable code to the worklist.
12118 /// This has a couple of tricks to make the code faster and more powerful. In
12119 /// particular, we constant fold and DCE instructions as we go, to avoid adding
12120 /// them to the worklist (this significantly speeds up instcombine on code where
12121 /// many instructions are dead or constant). Additionally, if we find a branch
12122 /// whose condition is a known constant, we only visit the reachable successors.
12124 static void AddReachableCodeToWorklist(BasicBlock *BB,
12125 SmallPtrSet<BasicBlock*, 64> &Visited,
12127 const TargetData *TD) {
12128 SmallVector<BasicBlock*, 256> Worklist;
12129 Worklist.push_back(BB);
12131 while (!Worklist.empty()) {
12132 BB = Worklist.back();
12133 Worklist.pop_back();
12135 // We have now visited this block! If we've already been here, ignore it.
12136 if (!Visited.insert(BB)) continue;
12138 DbgInfoIntrinsic *DBI_Prev = NULL;
12139 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) {
12140 Instruction *Inst = BBI++;
12142 // DCE instruction if trivially dead.
12143 if (isInstructionTriviallyDead(Inst)) {
12145 DOUT << "IC: DCE: " << *Inst;
12146 Inst->eraseFromParent();
12150 // ConstantProp instruction if trivially constant.
12151 if (Constant *C = ConstantFoldInstruction(Inst, TD)) {
12152 DOUT << "IC: ConstFold to: " << *C << " from: " << *Inst;
12153 Inst->replaceAllUsesWith(C);
12155 Inst->eraseFromParent();
12159 // If there are two consecutive llvm.dbg.stoppoint calls then
12160 // it is likely that the optimizer deleted code in between these
12162 DbgInfoIntrinsic *DBI_Next = dyn_cast<DbgInfoIntrinsic>(Inst);
12165 && DBI_Prev->getIntrinsicID() == llvm::Intrinsic::dbg_stoppoint
12166 && DBI_Next->getIntrinsicID() == llvm::Intrinsic::dbg_stoppoint) {
12167 IC.RemoveFromWorkList(DBI_Prev);
12168 DBI_Prev->eraseFromParent();
12170 DBI_Prev = DBI_Next;
12173 IC.AddToWorkList(Inst);
12176 // Recursively visit successors. If this is a branch or switch on a
12177 // constant, only visit the reachable successor.
12178 TerminatorInst *TI = BB->getTerminator();
12179 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
12180 if (BI->isConditional() && isa<ConstantInt>(BI->getCondition())) {
12181 bool CondVal = cast<ConstantInt>(BI->getCondition())->getZExtValue();
12182 BasicBlock *ReachableBB = BI->getSuccessor(!CondVal);
12183 Worklist.push_back(ReachableBB);
12186 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
12187 if (ConstantInt *Cond = dyn_cast<ConstantInt>(SI->getCondition())) {
12188 // See if this is an explicit destination.
12189 for (unsigned i = 1, e = SI->getNumSuccessors(); i != e; ++i)
12190 if (SI->getCaseValue(i) == Cond) {
12191 BasicBlock *ReachableBB = SI->getSuccessor(i);
12192 Worklist.push_back(ReachableBB);
12196 // Otherwise it is the default destination.
12197 Worklist.push_back(SI->getSuccessor(0));
12202 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
12203 Worklist.push_back(TI->getSuccessor(i));
12207 bool InstCombiner::DoOneIteration(Function &F, unsigned Iteration) {
12208 bool Changed = false;
12209 TD = &getAnalysis<TargetData>();
12211 DEBUG(DOUT << "\n\nINSTCOMBINE ITERATION #" << Iteration << " on "
12212 << F.getNameStr() << "\n");
12215 // Do a depth-first traversal of the function, populate the worklist with
12216 // the reachable instructions. Ignore blocks that are not reachable. Keep
12217 // track of which blocks we visit.
12218 SmallPtrSet<BasicBlock*, 64> Visited;
12219 AddReachableCodeToWorklist(F.begin(), Visited, *this, TD);
12221 // Do a quick scan over the function. If we find any blocks that are
12222 // unreachable, remove any instructions inside of them. This prevents
12223 // the instcombine code from having to deal with some bad special cases.
12224 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
12225 if (!Visited.count(BB)) {
12226 Instruction *Term = BB->getTerminator();
12227 while (Term != BB->begin()) { // Remove instrs bottom-up
12228 BasicBlock::iterator I = Term; --I;
12230 DOUT << "IC: DCE: " << *I;
12233 if (!I->use_empty())
12234 I->replaceAllUsesWith(UndefValue::get(I->getType()));
12235 I->eraseFromParent();
12240 while (!Worklist.empty()) {
12241 Instruction *I = RemoveOneFromWorkList();
12242 if (I == 0) continue; // skip null values.
12244 // Check to see if we can DCE the instruction.
12245 if (isInstructionTriviallyDead(I)) {
12246 // Add operands to the worklist.
12247 if (I->getNumOperands() < 4)
12248 AddUsesToWorkList(*I);
12251 DOUT << "IC: DCE: " << *I;
12253 I->eraseFromParent();
12254 RemoveFromWorkList(I);
12258 // Instruction isn't dead, see if we can constant propagate it.
12259 if (Constant *C = ConstantFoldInstruction(I, TD)) {
12260 DOUT << "IC: ConstFold to: " << *C << " from: " << *I;
12262 // Add operands to the worklist.
12263 AddUsesToWorkList(*I);
12264 ReplaceInstUsesWith(*I, C);
12267 I->eraseFromParent();
12268 RemoveFromWorkList(I);
12272 if (TD && I->getType()->getTypeID() == Type::VoidTyID) {
12273 // See if we can constant fold its operands.
12274 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i) {
12275 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(i)) {
12276 if (Constant *NewC = ConstantFoldConstantExpression(CE, TD))
12282 // See if we can trivially sink this instruction to a successor basic block.
12283 if (I->hasOneUse()) {
12284 BasicBlock *BB = I->getParent();
12285 BasicBlock *UserParent = cast<Instruction>(I->use_back())->getParent();
12286 if (UserParent != BB) {
12287 bool UserIsSuccessor = false;
12288 // See if the user is one of our successors.
12289 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI)
12290 if (*SI == UserParent) {
12291 UserIsSuccessor = true;
12295 // If the user is one of our immediate successors, and if that successor
12296 // only has us as a predecessors (we'd have to split the critical edge
12297 // otherwise), we can keep going.
12298 if (UserIsSuccessor && !isa<PHINode>(I->use_back()) &&
12299 next(pred_begin(UserParent)) == pred_end(UserParent))
12300 // Okay, the CFG is simple enough, try to sink this instruction.
12301 Changed |= TryToSinkInstruction(I, UserParent);
12305 // Now that we have an instruction, try combining it to simplify it...
12309 DEBUG(std::ostringstream SS; I->print(SS); OrigI = SS.str(););
12310 if (Instruction *Result = visit(*I)) {
12312 // Should we replace the old instruction with a new one?
12314 DOUT << "IC: Old = " << *I
12315 << " New = " << *Result;
12317 // Everything uses the new instruction now.
12318 I->replaceAllUsesWith(Result);
12320 // Push the new instruction and any users onto the worklist.
12321 AddToWorkList(Result);
12322 AddUsersToWorkList(*Result);
12324 // Move the name to the new instruction first.
12325 Result->takeName(I);
12327 // Insert the new instruction into the basic block...
12328 BasicBlock *InstParent = I->getParent();
12329 BasicBlock::iterator InsertPos = I;
12331 if (!isa<PHINode>(Result)) // If combining a PHI, don't insert
12332 while (isa<PHINode>(InsertPos)) // middle of a block of PHIs.
12335 InstParent->getInstList().insert(InsertPos, Result);
12337 // Make sure that we reprocess all operands now that we reduced their
12339 AddUsesToWorkList(*I);
12341 // Instructions can end up on the worklist more than once. Make sure
12342 // we do not process an instruction that has been deleted.
12343 RemoveFromWorkList(I);
12345 // Erase the old instruction.
12346 InstParent->getInstList().erase(I);
12349 DOUT << "IC: Mod = " << OrigI
12350 << " New = " << *I;
12353 // If the instruction was modified, it's possible that it is now dead.
12354 // if so, remove it.
12355 if (isInstructionTriviallyDead(I)) {
12356 // Make sure we process all operands now that we are reducing their
12358 AddUsesToWorkList(*I);
12360 // Instructions may end up in the worklist more than once. Erase all
12361 // occurrences of this instruction.
12362 RemoveFromWorkList(I);
12363 I->eraseFromParent();
12366 AddUsersToWorkList(*I);
12373 assert(WorklistMap.empty() && "Worklist empty, but map not?");
12375 // Do an explicit clear, this shrinks the map if needed.
12376 WorklistMap.clear();
12381 bool InstCombiner::runOnFunction(Function &F) {
12382 MustPreserveLCSSA = mustPreserveAnalysisID(LCSSAID);
12384 bool EverMadeChange = false;
12386 // Iterate while there is work to do.
12387 unsigned Iteration = 0;
12388 while (DoOneIteration(F, Iteration++))
12389 EverMadeChange = true;
12390 return EverMadeChange;
12393 FunctionPass *llvm::createInstructionCombiningPass() {
12394 return new InstCombiner();