1 //===- InstructionCombining.cpp - Combine multiple instructions -----------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // InstructionCombining - Combine instructions to form fewer, simple
11 // instructions. This pass does not modify the CFG. This pass is where
12 // algebraic simplification happens.
14 // This pass combines things like:
20 // This is a simple worklist driven algorithm.
22 // This pass guarantees that the following canonicalizations are performed on
24 // 1. If a binary operator has a constant operand, it is moved to the RHS
25 // 2. Bitwise operators with constant operands are always grouped so that
26 // shifts are performed first, then or's, then and's, then xor's.
27 // 3. Compare instructions are converted from <,>,<=,>= to ==,!= if possible
28 // 4. All cmp instructions on boolean values are replaced with logical ops
29 // 5. add X, X is represented as (X*2) => (X << 1)
30 // 6. Multiplies with a power-of-two constant argument are transformed into
34 //===----------------------------------------------------------------------===//
36 #define DEBUG_TYPE "instcombine"
37 #include "llvm/Transforms/Scalar.h"
38 #include "llvm/IntrinsicInst.h"
39 #include "llvm/Pass.h"
40 #include "llvm/DerivedTypes.h"
41 #include "llvm/GlobalVariable.h"
42 #include "llvm/Analysis/ConstantFolding.h"
43 #include "llvm/Analysis/ValueTracking.h"
44 #include "llvm/Target/TargetData.h"
45 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
46 #include "llvm/Transforms/Utils/Local.h"
47 #include "llvm/Support/CallSite.h"
48 #include "llvm/Support/ConstantRange.h"
49 #include "llvm/Support/Debug.h"
50 #include "llvm/Support/GetElementPtrTypeIterator.h"
51 #include "llvm/Support/InstVisitor.h"
52 #include "llvm/Support/MathExtras.h"
53 #include "llvm/Support/PatternMatch.h"
54 #include "llvm/Support/Compiler.h"
55 #include "llvm/ADT/DenseMap.h"
56 #include "llvm/ADT/SmallVector.h"
57 #include "llvm/ADT/SmallPtrSet.h"
58 #include "llvm/ADT/Statistic.h"
59 #include "llvm/ADT/STLExtras.h"
64 using namespace llvm::PatternMatch;
66 STATISTIC(NumCombined , "Number of insts combined");
67 STATISTIC(NumConstProp, "Number of constant folds");
68 STATISTIC(NumDeadInst , "Number of dead inst eliminated");
69 STATISTIC(NumDeadStore, "Number of dead stores eliminated");
70 STATISTIC(NumSunkInst , "Number of instructions sunk");
73 class VISIBILITY_HIDDEN InstCombiner
74 : public FunctionPass,
75 public InstVisitor<InstCombiner, Instruction*> {
76 // Worklist of all of the instructions that need to be simplified.
77 SmallVector<Instruction*, 256> Worklist;
78 DenseMap<Instruction*, unsigned> WorklistMap;
80 bool MustPreserveLCSSA;
82 static char ID; // Pass identification, replacement for typeid
83 InstCombiner() : FunctionPass(&ID) {}
85 /// AddToWorkList - Add the specified instruction to the worklist if it
86 /// isn't already in it.
87 void AddToWorkList(Instruction *I) {
88 if (WorklistMap.insert(std::make_pair(I, Worklist.size())).second)
89 Worklist.push_back(I);
92 // RemoveFromWorkList - remove I from the worklist if it exists.
93 void RemoveFromWorkList(Instruction *I) {
94 DenseMap<Instruction*, unsigned>::iterator It = WorklistMap.find(I);
95 if (It == WorklistMap.end()) return; // Not in worklist.
97 // Don't bother moving everything down, just null out the slot.
98 Worklist[It->second] = 0;
100 WorklistMap.erase(It);
103 Instruction *RemoveOneFromWorkList() {
104 Instruction *I = Worklist.back();
106 WorklistMap.erase(I);
111 /// AddUsersToWorkList - When an instruction is simplified, add all users of
112 /// the instruction to the work lists because they might get more simplified
115 void AddUsersToWorkList(Value &I) {
116 for (Value::use_iterator UI = I.use_begin(), UE = I.use_end();
118 AddToWorkList(cast<Instruction>(*UI));
121 /// AddUsesToWorkList - When an instruction is simplified, add operands to
122 /// the work lists because they might get more simplified now.
124 void AddUsesToWorkList(Instruction &I) {
125 for (User::op_iterator i = I.op_begin(), e = I.op_end(); i != e; ++i)
126 if (Instruction *Op = dyn_cast<Instruction>(*i))
130 /// AddSoonDeadInstToWorklist - The specified instruction is about to become
131 /// dead. Add all of its operands to the worklist, turning them into
132 /// undef's to reduce the number of uses of those instructions.
134 /// Return the specified operand before it is turned into an undef.
136 Value *AddSoonDeadInstToWorklist(Instruction &I, unsigned op) {
137 Value *R = I.getOperand(op);
139 for (User::op_iterator i = I.op_begin(), e = I.op_end(); i != e; ++i)
140 if (Instruction *Op = dyn_cast<Instruction>(*i)) {
142 // Set the operand to undef to drop the use.
143 *i = UndefValue::get(Op->getType());
150 virtual bool runOnFunction(Function &F);
152 bool DoOneIteration(Function &F, unsigned ItNum);
154 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
155 AU.addRequired<TargetData>();
156 AU.addPreservedID(LCSSAID);
157 AU.setPreservesCFG();
160 TargetData &getTargetData() const { return *TD; }
162 // Visitation implementation - Implement instruction combining for different
163 // instruction types. The semantics are as follows:
165 // null - No change was made
166 // I - Change was made, I is still valid, I may be dead though
167 // otherwise - Change was made, replace I with returned instruction
169 Instruction *visitAdd(BinaryOperator &I);
170 Instruction *visitSub(BinaryOperator &I);
171 Instruction *visitMul(BinaryOperator &I);
172 Instruction *visitURem(BinaryOperator &I);
173 Instruction *visitSRem(BinaryOperator &I);
174 Instruction *visitFRem(BinaryOperator &I);
175 bool SimplifyDivRemOfSelect(BinaryOperator &I);
176 Instruction *commonRemTransforms(BinaryOperator &I);
177 Instruction *commonIRemTransforms(BinaryOperator &I);
178 Instruction *commonDivTransforms(BinaryOperator &I);
179 Instruction *commonIDivTransforms(BinaryOperator &I);
180 Instruction *visitUDiv(BinaryOperator &I);
181 Instruction *visitSDiv(BinaryOperator &I);
182 Instruction *visitFDiv(BinaryOperator &I);
183 Instruction *FoldAndOfICmps(Instruction &I, ICmpInst *LHS, ICmpInst *RHS);
184 Instruction *visitAnd(BinaryOperator &I);
185 Instruction *FoldOrOfICmps(Instruction &I, ICmpInst *LHS, ICmpInst *RHS);
186 Instruction *visitOr (BinaryOperator &I);
187 Instruction *visitXor(BinaryOperator &I);
188 Instruction *visitShl(BinaryOperator &I);
189 Instruction *visitAShr(BinaryOperator &I);
190 Instruction *visitLShr(BinaryOperator &I);
191 Instruction *commonShiftTransforms(BinaryOperator &I);
192 Instruction *FoldFCmp_IntToFP_Cst(FCmpInst &I, Instruction *LHSI,
194 Instruction *visitFCmpInst(FCmpInst &I);
195 Instruction *visitICmpInst(ICmpInst &I);
196 Instruction *visitICmpInstWithCastAndCast(ICmpInst &ICI);
197 Instruction *visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
200 Instruction *FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
201 ConstantInt *DivRHS);
203 Instruction *FoldGEPICmp(User *GEPLHS, Value *RHS,
204 ICmpInst::Predicate Cond, Instruction &I);
205 Instruction *FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
207 Instruction *commonCastTransforms(CastInst &CI);
208 Instruction *commonIntCastTransforms(CastInst &CI);
209 Instruction *commonPointerCastTransforms(CastInst &CI);
210 Instruction *visitTrunc(TruncInst &CI);
211 Instruction *visitZExt(ZExtInst &CI);
212 Instruction *visitSExt(SExtInst &CI);
213 Instruction *visitFPTrunc(FPTruncInst &CI);
214 Instruction *visitFPExt(CastInst &CI);
215 Instruction *visitFPToUI(FPToUIInst &FI);
216 Instruction *visitFPToSI(FPToSIInst &FI);
217 Instruction *visitUIToFP(CastInst &CI);
218 Instruction *visitSIToFP(CastInst &CI);
219 Instruction *visitPtrToInt(CastInst &CI);
220 Instruction *visitIntToPtr(IntToPtrInst &CI);
221 Instruction *visitBitCast(BitCastInst &CI);
222 Instruction *FoldSelectOpOp(SelectInst &SI, Instruction *TI,
224 Instruction *visitSelectInst(SelectInst &SI);
225 Instruction *visitSelectInstWithICmp(SelectInst &SI, ICmpInst *ICI);
226 Instruction *visitCallInst(CallInst &CI);
227 Instruction *visitInvokeInst(InvokeInst &II);
228 Instruction *visitPHINode(PHINode &PN);
229 Instruction *visitGetElementPtrInst(GetElementPtrInst &GEP);
230 Instruction *visitAllocationInst(AllocationInst &AI);
231 Instruction *visitFreeInst(FreeInst &FI);
232 Instruction *visitLoadInst(LoadInst &LI);
233 Instruction *visitStoreInst(StoreInst &SI);
234 Instruction *visitBranchInst(BranchInst &BI);
235 Instruction *visitSwitchInst(SwitchInst &SI);
236 Instruction *visitInsertElementInst(InsertElementInst &IE);
237 Instruction *visitExtractElementInst(ExtractElementInst &EI);
238 Instruction *visitShuffleVectorInst(ShuffleVectorInst &SVI);
239 Instruction *visitExtractValueInst(ExtractValueInst &EV);
241 // visitInstruction - Specify what to return for unhandled instructions...
242 Instruction *visitInstruction(Instruction &I) { return 0; }
245 Instruction *visitCallSite(CallSite CS);
246 bool transformConstExprCastCall(CallSite CS);
247 Instruction *transformCallThroughTrampoline(CallSite CS);
248 Instruction *transformZExtICmp(ICmpInst *ICI, Instruction &CI,
249 bool DoXform = true);
250 bool WillNotOverflowSignedAdd(Value *LHS, Value *RHS);
253 // InsertNewInstBefore - insert an instruction New before instruction Old
254 // in the program. Add the new instruction to the worklist.
256 Instruction *InsertNewInstBefore(Instruction *New, Instruction &Old) {
257 assert(New && New->getParent() == 0 &&
258 "New instruction already inserted into a basic block!");
259 BasicBlock *BB = Old.getParent();
260 BB->getInstList().insert(&Old, New); // Insert inst
265 /// InsertCastBefore - Insert a cast of V to TY before the instruction POS.
266 /// This also adds the cast to the worklist. Finally, this returns the
268 Value *InsertCastBefore(Instruction::CastOps opc, Value *V, const Type *Ty,
270 if (V->getType() == Ty) return V;
272 if (Constant *CV = dyn_cast<Constant>(V))
273 return ConstantExpr::getCast(opc, CV, Ty);
275 Instruction *C = CastInst::Create(opc, V, Ty, V->getName(), &Pos);
280 Value *InsertBitCastBefore(Value *V, const Type *Ty, Instruction &Pos) {
281 return InsertCastBefore(Instruction::BitCast, V, Ty, Pos);
285 // ReplaceInstUsesWith - This method is to be used when an instruction is
286 // found to be dead, replacable with another preexisting expression. Here
287 // we add all uses of I to the worklist, replace all uses of I with the new
288 // value, then return I, so that the inst combiner will know that I was
291 Instruction *ReplaceInstUsesWith(Instruction &I, Value *V) {
292 AddUsersToWorkList(I); // Add all modified instrs to worklist
294 I.replaceAllUsesWith(V);
297 // If we are replacing the instruction with itself, this must be in a
298 // segment of unreachable code, so just clobber the instruction.
299 I.replaceAllUsesWith(UndefValue::get(I.getType()));
304 // UpdateValueUsesWith - This method is to be used when an value is
305 // found to be replacable with another preexisting expression or was
306 // updated. Here we add all uses of I to the worklist, replace all uses of
307 // I with the new value (unless the instruction was just updated), then
308 // return true, so that the inst combiner will know that I was modified.
310 bool UpdateValueUsesWith(Value *Old, Value *New) {
311 AddUsersToWorkList(*Old); // Add all modified instrs to worklist
313 Old->replaceAllUsesWith(New);
314 if (Instruction *I = dyn_cast<Instruction>(Old))
316 if (Instruction *I = dyn_cast<Instruction>(New))
321 // EraseInstFromFunction - When dealing with an instruction that has side
322 // effects or produces a void value, we can't rely on DCE to delete the
323 // instruction. Instead, visit methods should return the value returned by
325 Instruction *EraseInstFromFunction(Instruction &I) {
326 assert(I.use_empty() && "Cannot erase instruction that is used!");
327 AddUsesToWorkList(I);
328 RemoveFromWorkList(&I);
330 return 0; // Don't do anything with FI
333 void ComputeMaskedBits(Value *V, const APInt &Mask, APInt &KnownZero,
334 APInt &KnownOne, unsigned Depth = 0) const {
335 return llvm::ComputeMaskedBits(V, Mask, KnownZero, KnownOne, TD, Depth);
338 bool MaskedValueIsZero(Value *V, const APInt &Mask,
339 unsigned Depth = 0) const {
340 return llvm::MaskedValueIsZero(V, Mask, TD, Depth);
342 unsigned ComputeNumSignBits(Value *Op, unsigned Depth = 0) const {
343 return llvm::ComputeNumSignBits(Op, TD, Depth);
347 /// InsertOperandCastBefore - This inserts a cast of V to DestTy before the
348 /// InsertBefore instruction. This is specialized a bit to avoid inserting
349 /// casts that are known to not do anything...
351 Value *InsertOperandCastBefore(Instruction::CastOps opcode,
352 Value *V, const Type *DestTy,
353 Instruction *InsertBefore);
355 /// SimplifyCommutative - This performs a few simplifications for
356 /// commutative operators.
357 bool SimplifyCommutative(BinaryOperator &I);
359 /// SimplifyCompare - This reorders the operands of a CmpInst to get them in
360 /// most-complex to least-complex order.
361 bool SimplifyCompare(CmpInst &I);
363 /// SimplifyDemandedBits - Attempts to replace V with a simpler value based
364 /// on the demanded bits.
365 bool SimplifyDemandedBits(Value *V, APInt DemandedMask,
366 APInt& KnownZero, APInt& KnownOne,
369 Value *SimplifyDemandedVectorElts(Value *V, uint64_t DemandedElts,
370 uint64_t &UndefElts, unsigned Depth = 0);
372 // FoldOpIntoPhi - Given a binary operator or cast instruction which has a
373 // PHI node as operand #0, see if we can fold the instruction into the PHI
374 // (which is only possible if all operands to the PHI are constants).
375 Instruction *FoldOpIntoPhi(Instruction &I);
377 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
378 // operator and they all are only used by the PHI, PHI together their
379 // inputs, and do the operation once, to the result of the PHI.
380 Instruction *FoldPHIArgOpIntoPHI(PHINode &PN);
381 Instruction *FoldPHIArgBinOpIntoPHI(PHINode &PN);
384 Instruction *OptAndOp(Instruction *Op, ConstantInt *OpRHS,
385 ConstantInt *AndRHS, BinaryOperator &TheAnd);
387 Value *FoldLogicalPlusAnd(Value *LHS, Value *RHS, ConstantInt *Mask,
388 bool isSub, Instruction &I);
389 Instruction *InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
390 bool isSigned, bool Inside, Instruction &IB);
391 Instruction *PromoteCastOfAllocation(BitCastInst &CI, AllocationInst &AI);
392 Instruction *MatchBSwap(BinaryOperator &I);
393 bool SimplifyStoreAtEndOfBlock(StoreInst &SI);
394 Instruction *SimplifyMemTransfer(MemIntrinsic *MI);
395 Instruction *SimplifyMemSet(MemSetInst *MI);
398 Value *EvaluateInDifferentType(Value *V, const Type *Ty, bool isSigned);
400 bool CanEvaluateInDifferentType(Value *V, const IntegerType *Ty,
402 int &NumCastsRemoved);
403 unsigned GetOrEnforceKnownAlignment(Value *V,
404 unsigned PrefAlign = 0);
409 char InstCombiner::ID = 0;
410 static RegisterPass<InstCombiner>
411 X("instcombine", "Combine redundant instructions");
413 // getComplexity: Assign a complexity or rank value to LLVM Values...
414 // 0 -> undef, 1 -> Const, 2 -> Other, 3 -> Arg, 3 -> Unary, 4 -> OtherInst
415 static unsigned getComplexity(Value *V) {
416 if (isa<Instruction>(V)) {
417 if (BinaryOperator::isNeg(V) || BinaryOperator::isNot(V))
421 if (isa<Argument>(V)) return 3;
422 return isa<Constant>(V) ? (isa<UndefValue>(V) ? 0 : 1) : 2;
425 // isOnlyUse - Return true if this instruction will be deleted if we stop using
427 static bool isOnlyUse(Value *V) {
428 return V->hasOneUse() || isa<Constant>(V);
431 // getPromotedType - Return the specified type promoted as it would be to pass
432 // though a va_arg area...
433 static const Type *getPromotedType(const Type *Ty) {
434 if (const IntegerType* ITy = dyn_cast<IntegerType>(Ty)) {
435 if (ITy->getBitWidth() < 32)
436 return Type::Int32Ty;
441 /// getBitCastOperand - If the specified operand is a CastInst, a constant
442 /// expression bitcast, or a GetElementPtrInst with all zero indices, return the
443 /// operand value, otherwise return null.
444 static Value *getBitCastOperand(Value *V) {
445 if (BitCastInst *I = dyn_cast<BitCastInst>(V))
447 return I->getOperand(0);
448 else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(V)) {
449 // GetElementPtrInst?
450 if (GEP->hasAllZeroIndices())
451 return GEP->getOperand(0);
452 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) {
453 if (CE->getOpcode() == Instruction::BitCast)
454 // BitCast ConstantExp?
455 return CE->getOperand(0);
456 else if (CE->getOpcode() == Instruction::GetElementPtr) {
457 // GetElementPtr ConstantExp?
458 for (User::op_iterator I = CE->op_begin() + 1, E = CE->op_end();
460 ConstantInt *CI = dyn_cast<ConstantInt>(I);
461 if (!CI || !CI->isZero())
462 // Any non-zero indices? Not cast-like.
465 // All-zero indices? This is just like casting.
466 return CE->getOperand(0);
472 /// This function is a wrapper around CastInst::isEliminableCastPair. It
473 /// simply extracts arguments and returns what that function returns.
474 static Instruction::CastOps
475 isEliminableCastPair(
476 const CastInst *CI, ///< The first cast instruction
477 unsigned opcode, ///< The opcode of the second cast instruction
478 const Type *DstTy, ///< The target type for the second cast instruction
479 TargetData *TD ///< The target data for pointer size
482 const Type *SrcTy = CI->getOperand(0)->getType(); // A from above
483 const Type *MidTy = CI->getType(); // B from above
485 // Get the opcodes of the two Cast instructions
486 Instruction::CastOps firstOp = Instruction::CastOps(CI->getOpcode());
487 Instruction::CastOps secondOp = Instruction::CastOps(opcode);
489 return Instruction::CastOps(
490 CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy,
491 DstTy, TD->getIntPtrType()));
494 /// ValueRequiresCast - Return true if the cast from "V to Ty" actually results
495 /// in any code being generated. It does not require codegen if V is simple
496 /// enough or if the cast can be folded into other casts.
497 static bool ValueRequiresCast(Instruction::CastOps opcode, const Value *V,
498 const Type *Ty, TargetData *TD) {
499 if (V->getType() == Ty || isa<Constant>(V)) return false;
501 // If this is another cast that can be eliminated, it isn't codegen either.
502 if (const CastInst *CI = dyn_cast<CastInst>(V))
503 if (isEliminableCastPair(CI, opcode, Ty, TD))
508 /// InsertOperandCastBefore - This inserts a cast of V to DestTy before the
509 /// InsertBefore instruction. This is specialized a bit to avoid inserting
510 /// casts that are known to not do anything...
512 Value *InstCombiner::InsertOperandCastBefore(Instruction::CastOps opcode,
513 Value *V, const Type *DestTy,
514 Instruction *InsertBefore) {
515 if (V->getType() == DestTy) return V;
516 if (Constant *C = dyn_cast<Constant>(V))
517 return ConstantExpr::getCast(opcode, C, DestTy);
519 return InsertCastBefore(opcode, V, DestTy, *InsertBefore);
522 // SimplifyCommutative - This performs a few simplifications for commutative
525 // 1. Order operands such that they are listed from right (least complex) to
526 // left (most complex). This puts constants before unary operators before
529 // 2. Transform: (op (op V, C1), C2) ==> (op V, (op C1, C2))
530 // 3. Transform: (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
532 bool InstCombiner::SimplifyCommutative(BinaryOperator &I) {
533 bool Changed = false;
534 if (getComplexity(I.getOperand(0)) < getComplexity(I.getOperand(1)))
535 Changed = !I.swapOperands();
537 if (!I.isAssociative()) return Changed;
538 Instruction::BinaryOps Opcode = I.getOpcode();
539 if (BinaryOperator *Op = dyn_cast<BinaryOperator>(I.getOperand(0)))
540 if (Op->getOpcode() == Opcode && isa<Constant>(Op->getOperand(1))) {
541 if (isa<Constant>(I.getOperand(1))) {
542 Constant *Folded = ConstantExpr::get(I.getOpcode(),
543 cast<Constant>(I.getOperand(1)),
544 cast<Constant>(Op->getOperand(1)));
545 I.setOperand(0, Op->getOperand(0));
546 I.setOperand(1, Folded);
548 } else if (BinaryOperator *Op1=dyn_cast<BinaryOperator>(I.getOperand(1)))
549 if (Op1->getOpcode() == Opcode && isa<Constant>(Op1->getOperand(1)) &&
550 isOnlyUse(Op) && isOnlyUse(Op1)) {
551 Constant *C1 = cast<Constant>(Op->getOperand(1));
552 Constant *C2 = cast<Constant>(Op1->getOperand(1));
554 // Fold (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
555 Constant *Folded = ConstantExpr::get(I.getOpcode(), C1, C2);
556 Instruction *New = BinaryOperator::Create(Opcode, Op->getOperand(0),
560 I.setOperand(0, New);
561 I.setOperand(1, Folded);
568 /// SimplifyCompare - For a CmpInst this function just orders the operands
569 /// so that theyare listed from right (least complex) to left (most complex).
570 /// This puts constants before unary operators before binary operators.
571 bool InstCombiner::SimplifyCompare(CmpInst &I) {
572 if (getComplexity(I.getOperand(0)) >= getComplexity(I.getOperand(1)))
575 // Compare instructions are not associative so there's nothing else we can do.
579 // dyn_castNegVal - Given a 'sub' instruction, return the RHS of the instruction
580 // if the LHS is a constant zero (which is the 'negate' form).
582 static inline Value *dyn_castNegVal(Value *V) {
583 if (BinaryOperator::isNeg(V))
584 return BinaryOperator::getNegArgument(V);
586 // Constants can be considered to be negated values if they can be folded.
587 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
588 return ConstantExpr::getNeg(C);
590 if (ConstantVector *C = dyn_cast<ConstantVector>(V))
591 if (C->getType()->getElementType()->isInteger())
592 return ConstantExpr::getNeg(C);
597 static inline Value *dyn_castNotVal(Value *V) {
598 if (BinaryOperator::isNot(V))
599 return BinaryOperator::getNotArgument(V);
601 // Constants can be considered to be not'ed values...
602 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
603 return ConstantInt::get(~C->getValue());
607 // dyn_castFoldableMul - If this value is a multiply that can be folded into
608 // other computations (because it has a constant operand), return the
609 // non-constant operand of the multiply, and set CST to point to the multiplier.
610 // Otherwise, return null.
612 static inline Value *dyn_castFoldableMul(Value *V, ConstantInt *&CST) {
613 if (V->hasOneUse() && V->getType()->isInteger())
614 if (Instruction *I = dyn_cast<Instruction>(V)) {
615 if (I->getOpcode() == Instruction::Mul)
616 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1))))
617 return I->getOperand(0);
618 if (I->getOpcode() == Instruction::Shl)
619 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1)))) {
620 // The multiplier is really 1 << CST.
621 uint32_t BitWidth = cast<IntegerType>(V->getType())->getBitWidth();
622 uint32_t CSTVal = CST->getLimitedValue(BitWidth);
623 CST = ConstantInt::get(APInt(BitWidth, 1).shl(CSTVal));
624 return I->getOperand(0);
630 /// dyn_castGetElementPtr - If this is a getelementptr instruction or constant
631 /// expression, return it.
632 static User *dyn_castGetElementPtr(Value *V) {
633 if (isa<GetElementPtrInst>(V)) return cast<User>(V);
634 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
635 if (CE->getOpcode() == Instruction::GetElementPtr)
636 return cast<User>(V);
640 /// getOpcode - If this is an Instruction or a ConstantExpr, return the
641 /// opcode value. Otherwise return UserOp1.
642 static unsigned getOpcode(const Value *V) {
643 if (const Instruction *I = dyn_cast<Instruction>(V))
644 return I->getOpcode();
645 if (const ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
646 return CE->getOpcode();
647 // Use UserOp1 to mean there's no opcode.
648 return Instruction::UserOp1;
651 /// AddOne - Add one to a ConstantInt
652 static ConstantInt *AddOne(ConstantInt *C) {
653 APInt Val(C->getValue());
654 return ConstantInt::get(++Val);
656 /// SubOne - Subtract one from a ConstantInt
657 static ConstantInt *SubOne(ConstantInt *C) {
658 APInt Val(C->getValue());
659 return ConstantInt::get(--Val);
661 /// Add - Add two ConstantInts together
662 static ConstantInt *Add(ConstantInt *C1, ConstantInt *C2) {
663 return ConstantInt::get(C1->getValue() + C2->getValue());
665 /// And - Bitwise AND two ConstantInts together
666 static ConstantInt *And(ConstantInt *C1, ConstantInt *C2) {
667 return ConstantInt::get(C1->getValue() & C2->getValue());
669 /// Subtract - Subtract one ConstantInt from another
670 static ConstantInt *Subtract(ConstantInt *C1, ConstantInt *C2) {
671 return ConstantInt::get(C1->getValue() - C2->getValue());
673 /// Multiply - Multiply two ConstantInts together
674 static ConstantInt *Multiply(ConstantInt *C1, ConstantInt *C2) {
675 return ConstantInt::get(C1->getValue() * C2->getValue());
677 /// MultiplyOverflows - True if the multiply can not be expressed in an int
679 static bool MultiplyOverflows(ConstantInt *C1, ConstantInt *C2, bool sign) {
680 uint32_t W = C1->getBitWidth();
681 APInt LHSExt = C1->getValue(), RHSExt = C2->getValue();
690 APInt MulExt = LHSExt * RHSExt;
693 APInt Min = APInt::getSignedMinValue(W).sext(W * 2);
694 APInt Max = APInt::getSignedMaxValue(W).sext(W * 2);
695 return MulExt.slt(Min) || MulExt.sgt(Max);
697 return MulExt.ugt(APInt::getLowBitsSet(W * 2, W));
701 /// ShrinkDemandedConstant - Check to see if the specified operand of the
702 /// specified instruction is a constant integer. If so, check to see if there
703 /// are any bits set in the constant that are not demanded. If so, shrink the
704 /// constant and return true.
705 static bool ShrinkDemandedConstant(Instruction *I, unsigned OpNo,
707 assert(I && "No instruction?");
708 assert(OpNo < I->getNumOperands() && "Operand index too large");
710 // If the operand is not a constant integer, nothing to do.
711 ConstantInt *OpC = dyn_cast<ConstantInt>(I->getOperand(OpNo));
712 if (!OpC) return false;
714 // If there are no bits set that aren't demanded, nothing to do.
715 Demanded.zextOrTrunc(OpC->getValue().getBitWidth());
716 if ((~Demanded & OpC->getValue()) == 0)
719 // This instruction is producing bits that are not demanded. Shrink the RHS.
720 Demanded &= OpC->getValue();
721 I->setOperand(OpNo, ConstantInt::get(Demanded));
725 // ComputeSignedMinMaxValuesFromKnownBits - Given a signed integer type and a
726 // set of known zero and one bits, compute the maximum and minimum values that
727 // could have the specified known zero and known one bits, returning them in
729 static void ComputeSignedMinMaxValuesFromKnownBits(const Type *Ty,
730 const APInt& KnownZero,
731 const APInt& KnownOne,
732 APInt& Min, APInt& Max) {
733 uint32_t BitWidth = cast<IntegerType>(Ty)->getBitWidth();
734 assert(KnownZero.getBitWidth() == BitWidth &&
735 KnownOne.getBitWidth() == BitWidth &&
736 Min.getBitWidth() == BitWidth && Max.getBitWidth() == BitWidth &&
737 "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth.");
738 APInt UnknownBits = ~(KnownZero|KnownOne);
740 // The minimum value is when all unknown bits are zeros, EXCEPT for the sign
741 // bit if it is unknown.
743 Max = KnownOne|UnknownBits;
745 if (UnknownBits[BitWidth-1]) { // Sign bit is unknown
747 Max.clear(BitWidth-1);
751 // ComputeUnsignedMinMaxValuesFromKnownBits - Given an unsigned integer type and
752 // a set of known zero and one bits, compute the maximum and minimum values that
753 // could have the specified known zero and known one bits, returning them in
755 static void ComputeUnsignedMinMaxValuesFromKnownBits(const Type *Ty,
756 const APInt &KnownZero,
757 const APInt &KnownOne,
758 APInt &Min, APInt &Max) {
759 uint32_t BitWidth = cast<IntegerType>(Ty)->getBitWidth(); BitWidth = BitWidth;
760 assert(KnownZero.getBitWidth() == BitWidth &&
761 KnownOne.getBitWidth() == BitWidth &&
762 Min.getBitWidth() == BitWidth && Max.getBitWidth() &&
763 "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth.");
764 APInt UnknownBits = ~(KnownZero|KnownOne);
766 // The minimum value is when the unknown bits are all zeros.
768 // The maximum value is when the unknown bits are all ones.
769 Max = KnownOne|UnknownBits;
772 /// SimplifyDemandedBits - This function attempts to replace V with a simpler
773 /// value based on the demanded bits. When this function is called, it is known
774 /// that only the bits set in DemandedMask of the result of V are ever used
775 /// downstream. Consequently, depending on the mask and V, it may be possible
776 /// to replace V with a constant or one of its operands. In such cases, this
777 /// function does the replacement and returns true. In all other cases, it
778 /// returns false after analyzing the expression and setting KnownOne and known
779 /// to be one in the expression. KnownZero contains all the bits that are known
780 /// to be zero in the expression. These are provided to potentially allow the
781 /// caller (which might recursively be SimplifyDemandedBits itself) to simplify
782 /// the expression. KnownOne and KnownZero always follow the invariant that
783 /// KnownOne & KnownZero == 0. That is, a bit can't be both 1 and 0. Note that
784 /// the bits in KnownOne and KnownZero may only be accurate for those bits set
785 /// in DemandedMask. Note also that the bitwidth of V, DemandedMask, KnownZero
786 /// and KnownOne must all be the same.
787 bool InstCombiner::SimplifyDemandedBits(Value *V, APInt DemandedMask,
788 APInt& KnownZero, APInt& KnownOne,
790 assert(V != 0 && "Null pointer of Value???");
791 assert(Depth <= 6 && "Limit Search Depth");
792 uint32_t BitWidth = DemandedMask.getBitWidth();
793 const IntegerType *VTy = cast<IntegerType>(V->getType());
794 assert(VTy->getBitWidth() == BitWidth &&
795 KnownZero.getBitWidth() == BitWidth &&
796 KnownOne.getBitWidth() == BitWidth &&
797 "Value *V, DemandedMask, KnownZero and KnownOne \
798 must have same BitWidth");
799 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
800 // We know all of the bits for a constant!
801 KnownOne = CI->getValue() & DemandedMask;
802 KnownZero = ~KnownOne & DemandedMask;
808 if (!V->hasOneUse()) { // Other users may use these bits.
809 if (Depth != 0) { // Not at the root.
810 // Just compute the KnownZero/KnownOne bits to simplify things downstream.
811 ComputeMaskedBits(V, DemandedMask, KnownZero, KnownOne, Depth);
814 // If this is the root being simplified, allow it to have multiple uses,
815 // just set the DemandedMask to all bits.
816 DemandedMask = APInt::getAllOnesValue(BitWidth);
817 } else if (DemandedMask == 0) { // Not demanding any bits from V.
818 if (V != UndefValue::get(VTy))
819 return UpdateValueUsesWith(V, UndefValue::get(VTy));
821 } else if (Depth == 6) { // Limit search depth.
825 Instruction *I = dyn_cast<Instruction>(V);
826 if (!I) return false; // Only analyze instructions.
828 APInt LHSKnownZero(BitWidth, 0), LHSKnownOne(BitWidth, 0);
829 APInt &RHSKnownZero = KnownZero, &RHSKnownOne = KnownOne;
830 switch (I->getOpcode()) {
832 ComputeMaskedBits(V, DemandedMask, RHSKnownZero, RHSKnownOne, Depth);
834 case Instruction::And:
835 // If either the LHS or the RHS are Zero, the result is zero.
836 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
837 RHSKnownZero, RHSKnownOne, Depth+1))
839 assert((RHSKnownZero & RHSKnownOne) == 0 &&
840 "Bits known to be one AND zero?");
842 // If something is known zero on the RHS, the bits aren't demanded on the
844 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask & ~RHSKnownZero,
845 LHSKnownZero, LHSKnownOne, Depth+1))
847 assert((LHSKnownZero & LHSKnownOne) == 0 &&
848 "Bits known to be one AND zero?");
850 // If all of the demanded bits are known 1 on one side, return the other.
851 // These bits cannot contribute to the result of the 'and'.
852 if ((DemandedMask & ~LHSKnownZero & RHSKnownOne) ==
853 (DemandedMask & ~LHSKnownZero))
854 return UpdateValueUsesWith(I, I->getOperand(0));
855 if ((DemandedMask & ~RHSKnownZero & LHSKnownOne) ==
856 (DemandedMask & ~RHSKnownZero))
857 return UpdateValueUsesWith(I, I->getOperand(1));
859 // If all of the demanded bits in the inputs are known zeros, return zero.
860 if ((DemandedMask & (RHSKnownZero|LHSKnownZero)) == DemandedMask)
861 return UpdateValueUsesWith(I, Constant::getNullValue(VTy));
863 // If the RHS is a constant, see if we can simplify it.
864 if (ShrinkDemandedConstant(I, 1, DemandedMask & ~LHSKnownZero))
865 return UpdateValueUsesWith(I, I);
867 // Output known-1 bits are only known if set in both the LHS & RHS.
868 RHSKnownOne &= LHSKnownOne;
869 // Output known-0 are known to be clear if zero in either the LHS | RHS.
870 RHSKnownZero |= LHSKnownZero;
872 case Instruction::Or:
873 // If either the LHS or the RHS are One, the result is One.
874 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
875 RHSKnownZero, RHSKnownOne, Depth+1))
877 assert((RHSKnownZero & RHSKnownOne) == 0 &&
878 "Bits known to be one AND zero?");
879 // If something is known one on the RHS, the bits aren't demanded on the
881 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask & ~RHSKnownOne,
882 LHSKnownZero, LHSKnownOne, Depth+1))
884 assert((LHSKnownZero & LHSKnownOne) == 0 &&
885 "Bits known to be one AND zero?");
887 // If all of the demanded bits are known zero on one side, return the other.
888 // These bits cannot contribute to the result of the 'or'.
889 if ((DemandedMask & ~LHSKnownOne & RHSKnownZero) ==
890 (DemandedMask & ~LHSKnownOne))
891 return UpdateValueUsesWith(I, I->getOperand(0));
892 if ((DemandedMask & ~RHSKnownOne & LHSKnownZero) ==
893 (DemandedMask & ~RHSKnownOne))
894 return UpdateValueUsesWith(I, I->getOperand(1));
896 // If all of the potentially set bits on one side are known to be set on
897 // the other side, just use the 'other' side.
898 if ((DemandedMask & (~RHSKnownZero) & LHSKnownOne) ==
899 (DemandedMask & (~RHSKnownZero)))
900 return UpdateValueUsesWith(I, I->getOperand(0));
901 if ((DemandedMask & (~LHSKnownZero) & RHSKnownOne) ==
902 (DemandedMask & (~LHSKnownZero)))
903 return UpdateValueUsesWith(I, I->getOperand(1));
905 // If the RHS is a constant, see if we can simplify it.
906 if (ShrinkDemandedConstant(I, 1, DemandedMask))
907 return UpdateValueUsesWith(I, I);
909 // Output known-0 bits are only known if clear in both the LHS & RHS.
910 RHSKnownZero &= LHSKnownZero;
911 // Output known-1 are known to be set if set in either the LHS | RHS.
912 RHSKnownOne |= LHSKnownOne;
914 case Instruction::Xor: {
915 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
916 RHSKnownZero, RHSKnownOne, Depth+1))
918 assert((RHSKnownZero & RHSKnownOne) == 0 &&
919 "Bits known to be one AND zero?");
920 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
921 LHSKnownZero, LHSKnownOne, Depth+1))
923 assert((LHSKnownZero & LHSKnownOne) == 0 &&
924 "Bits known to be one AND zero?");
926 // If all of the demanded bits are known zero on one side, return the other.
927 // These bits cannot contribute to the result of the 'xor'.
928 if ((DemandedMask & RHSKnownZero) == DemandedMask)
929 return UpdateValueUsesWith(I, I->getOperand(0));
930 if ((DemandedMask & LHSKnownZero) == DemandedMask)
931 return UpdateValueUsesWith(I, I->getOperand(1));
933 // Output known-0 bits are known if clear or set in both the LHS & RHS.
934 APInt KnownZeroOut = (RHSKnownZero & LHSKnownZero) |
935 (RHSKnownOne & LHSKnownOne);
936 // Output known-1 are known to be set if set in only one of the LHS, RHS.
937 APInt KnownOneOut = (RHSKnownZero & LHSKnownOne) |
938 (RHSKnownOne & LHSKnownZero);
940 // If all of the demanded bits are known to be zero on one side or the
941 // other, turn this into an *inclusive* or.
942 // e.g. (A & C1)^(B & C2) -> (A & C1)|(B & C2) iff C1&C2 == 0
943 if ((DemandedMask & ~RHSKnownZero & ~LHSKnownZero) == 0) {
945 BinaryOperator::CreateOr(I->getOperand(0), I->getOperand(1),
947 InsertNewInstBefore(Or, *I);
948 return UpdateValueUsesWith(I, Or);
951 // If all of the demanded bits on one side are known, and all of the set
952 // bits on that side are also known to be set on the other side, turn this
953 // into an AND, as we know the bits will be cleared.
954 // e.g. (X | C1) ^ C2 --> (X | C1) & ~C2 iff (C1&C2) == C2
955 if ((DemandedMask & (RHSKnownZero|RHSKnownOne)) == DemandedMask) {
957 if ((RHSKnownOne & LHSKnownOne) == RHSKnownOne) {
958 Constant *AndC = ConstantInt::get(~RHSKnownOne & DemandedMask);
960 BinaryOperator::CreateAnd(I->getOperand(0), AndC, "tmp");
961 InsertNewInstBefore(And, *I);
962 return UpdateValueUsesWith(I, And);
966 // If the RHS is a constant, see if we can simplify it.
967 // FIXME: for XOR, we prefer to force bits to 1 if they will make a -1.
968 if (ShrinkDemandedConstant(I, 1, DemandedMask))
969 return UpdateValueUsesWith(I, I);
971 RHSKnownZero = KnownZeroOut;
972 RHSKnownOne = KnownOneOut;
975 case Instruction::Select:
976 if (SimplifyDemandedBits(I->getOperand(2), DemandedMask,
977 RHSKnownZero, RHSKnownOne, Depth+1))
979 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
980 LHSKnownZero, LHSKnownOne, Depth+1))
982 assert((RHSKnownZero & RHSKnownOne) == 0 &&
983 "Bits known to be one AND zero?");
984 assert((LHSKnownZero & LHSKnownOne) == 0 &&
985 "Bits known to be one AND zero?");
987 // If the operands are constants, see if we can simplify them.
988 if (ShrinkDemandedConstant(I, 1, DemandedMask))
989 return UpdateValueUsesWith(I, I);
990 if (ShrinkDemandedConstant(I, 2, DemandedMask))
991 return UpdateValueUsesWith(I, I);
993 // Only known if known in both the LHS and RHS.
994 RHSKnownOne &= LHSKnownOne;
995 RHSKnownZero &= LHSKnownZero;
997 case Instruction::Trunc: {
999 cast<IntegerType>(I->getOperand(0)->getType())->getBitWidth();
1000 DemandedMask.zext(truncBf);
1001 RHSKnownZero.zext(truncBf);
1002 RHSKnownOne.zext(truncBf);
1003 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1004 RHSKnownZero, RHSKnownOne, Depth+1))
1006 DemandedMask.trunc(BitWidth);
1007 RHSKnownZero.trunc(BitWidth);
1008 RHSKnownOne.trunc(BitWidth);
1009 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1010 "Bits known to be one AND zero?");
1013 case Instruction::BitCast:
1014 if (!I->getOperand(0)->getType()->isInteger())
1017 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1018 RHSKnownZero, RHSKnownOne, Depth+1))
1020 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1021 "Bits known to be one AND zero?");
1023 case Instruction::ZExt: {
1024 // Compute the bits in the result that are not present in the input.
1025 const IntegerType *SrcTy = cast<IntegerType>(I->getOperand(0)->getType());
1026 uint32_t SrcBitWidth = SrcTy->getBitWidth();
1028 DemandedMask.trunc(SrcBitWidth);
1029 RHSKnownZero.trunc(SrcBitWidth);
1030 RHSKnownOne.trunc(SrcBitWidth);
1031 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1032 RHSKnownZero, RHSKnownOne, Depth+1))
1034 DemandedMask.zext(BitWidth);
1035 RHSKnownZero.zext(BitWidth);
1036 RHSKnownOne.zext(BitWidth);
1037 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1038 "Bits known to be one AND zero?");
1039 // The top bits are known to be zero.
1040 RHSKnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
1043 case Instruction::SExt: {
1044 // Compute the bits in the result that are not present in the input.
1045 const IntegerType *SrcTy = cast<IntegerType>(I->getOperand(0)->getType());
1046 uint32_t SrcBitWidth = SrcTy->getBitWidth();
1048 APInt InputDemandedBits = DemandedMask &
1049 APInt::getLowBitsSet(BitWidth, SrcBitWidth);
1051 APInt NewBits(APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth));
1052 // If any of the sign extended bits are demanded, we know that the sign
1054 if ((NewBits & DemandedMask) != 0)
1055 InputDemandedBits.set(SrcBitWidth-1);
1057 InputDemandedBits.trunc(SrcBitWidth);
1058 RHSKnownZero.trunc(SrcBitWidth);
1059 RHSKnownOne.trunc(SrcBitWidth);
1060 if (SimplifyDemandedBits(I->getOperand(0), InputDemandedBits,
1061 RHSKnownZero, RHSKnownOne, Depth+1))
1063 InputDemandedBits.zext(BitWidth);
1064 RHSKnownZero.zext(BitWidth);
1065 RHSKnownOne.zext(BitWidth);
1066 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1067 "Bits known to be one AND zero?");
1069 // If the sign bit of the input is known set or clear, then we know the
1070 // top bits of the result.
1072 // If the input sign bit is known zero, or if the NewBits are not demanded
1073 // convert this into a zero extension.
1074 if (RHSKnownZero[SrcBitWidth-1] || (NewBits & ~DemandedMask) == NewBits)
1076 // Convert to ZExt cast
1077 CastInst *NewCast = new ZExtInst(I->getOperand(0), VTy, I->getName(), I);
1078 return UpdateValueUsesWith(I, NewCast);
1079 } else if (RHSKnownOne[SrcBitWidth-1]) { // Input sign bit known set
1080 RHSKnownOne |= NewBits;
1084 case Instruction::Add: {
1085 // Figure out what the input bits are. If the top bits of the and result
1086 // are not demanded, then the add doesn't demand them from its input
1088 uint32_t NLZ = DemandedMask.countLeadingZeros();
1090 // If there is a constant on the RHS, there are a variety of xformations
1092 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
1093 // If null, this should be simplified elsewhere. Some of the xforms here
1094 // won't work if the RHS is zero.
1098 // If the top bit of the output is demanded, demand everything from the
1099 // input. Otherwise, we demand all the input bits except NLZ top bits.
1100 APInt InDemandedBits(APInt::getLowBitsSet(BitWidth, BitWidth - NLZ));
1102 // Find information about known zero/one bits in the input.
1103 if (SimplifyDemandedBits(I->getOperand(0), InDemandedBits,
1104 LHSKnownZero, LHSKnownOne, Depth+1))
1107 // If the RHS of the add has bits set that can't affect the input, reduce
1109 if (ShrinkDemandedConstant(I, 1, InDemandedBits))
1110 return UpdateValueUsesWith(I, I);
1112 // Avoid excess work.
1113 if (LHSKnownZero == 0 && LHSKnownOne == 0)
1116 // Turn it into OR if input bits are zero.
1117 if ((LHSKnownZero & RHS->getValue()) == RHS->getValue()) {
1119 BinaryOperator::CreateOr(I->getOperand(0), I->getOperand(1),
1121 InsertNewInstBefore(Or, *I);
1122 return UpdateValueUsesWith(I, Or);
1125 // We can say something about the output known-zero and known-one bits,
1126 // depending on potential carries from the input constant and the
1127 // unknowns. For example if the LHS is known to have at most the 0x0F0F0
1128 // bits set and the RHS constant is 0x01001, then we know we have a known
1129 // one mask of 0x00001 and a known zero mask of 0xE0F0E.
1131 // To compute this, we first compute the potential carry bits. These are
1132 // the bits which may be modified. I'm not aware of a better way to do
1134 const APInt& RHSVal = RHS->getValue();
1135 APInt CarryBits((~LHSKnownZero + RHSVal) ^ (~LHSKnownZero ^ RHSVal));
1137 // Now that we know which bits have carries, compute the known-1/0 sets.
1139 // Bits are known one if they are known zero in one operand and one in the
1140 // other, and there is no input carry.
1141 RHSKnownOne = ((LHSKnownZero & RHSVal) |
1142 (LHSKnownOne & ~RHSVal)) & ~CarryBits;
1144 // Bits are known zero if they are known zero in both operands and there
1145 // is no input carry.
1146 RHSKnownZero = LHSKnownZero & ~RHSVal & ~CarryBits;
1148 // If the high-bits of this ADD are not demanded, then it does not demand
1149 // the high bits of its LHS or RHS.
1150 if (DemandedMask[BitWidth-1] == 0) {
1151 // Right fill the mask of bits for this ADD to demand the most
1152 // significant bit and all those below it.
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))
1164 case Instruction::Sub:
1165 // If the high-bits of this SUB are not demanded, then it does not demand
1166 // the high bits of its LHS or RHS.
1167 if (DemandedMask[BitWidth-1] == 0) {
1168 // Right fill the mask of bits for this SUB to demand the most
1169 // significant bit and all those below it.
1170 uint32_t NLZ = DemandedMask.countLeadingZeros();
1171 APInt DemandedFromOps(APInt::getLowBitsSet(BitWidth, BitWidth-NLZ));
1172 if (SimplifyDemandedBits(I->getOperand(0), DemandedFromOps,
1173 LHSKnownZero, LHSKnownOne, Depth+1))
1175 if (SimplifyDemandedBits(I->getOperand(1), DemandedFromOps,
1176 LHSKnownZero, LHSKnownOne, Depth+1))
1179 // Otherwise just hand the sub off to ComputeMaskedBits to fill in
1180 // the known zeros and ones.
1181 ComputeMaskedBits(V, DemandedMask, RHSKnownZero, RHSKnownOne, Depth);
1183 case Instruction::Shl:
1184 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1185 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
1186 APInt DemandedMaskIn(DemandedMask.lshr(ShiftAmt));
1187 if (SimplifyDemandedBits(I->getOperand(0), DemandedMaskIn,
1188 RHSKnownZero, RHSKnownOne, Depth+1))
1190 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1191 "Bits known to be one AND zero?");
1192 RHSKnownZero <<= ShiftAmt;
1193 RHSKnownOne <<= ShiftAmt;
1194 // low bits known zero.
1196 RHSKnownZero |= APInt::getLowBitsSet(BitWidth, ShiftAmt);
1199 case Instruction::LShr:
1200 // For a logical shift right
1201 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1202 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
1204 // Unsigned shift right.
1205 APInt DemandedMaskIn(DemandedMask.shl(ShiftAmt));
1206 if (SimplifyDemandedBits(I->getOperand(0), DemandedMaskIn,
1207 RHSKnownZero, RHSKnownOne, Depth+1))
1209 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1210 "Bits known to be one AND zero?");
1211 RHSKnownZero = APIntOps::lshr(RHSKnownZero, ShiftAmt);
1212 RHSKnownOne = APIntOps::lshr(RHSKnownOne, ShiftAmt);
1214 // Compute the new bits that are at the top now.
1215 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
1216 RHSKnownZero |= HighBits; // high bits known zero.
1220 case Instruction::AShr:
1221 // If this is an arithmetic shift right and only the low-bit is set, we can
1222 // always convert this into a logical shr, even if the shift amount is
1223 // variable. The low bit of the shift cannot be an input sign bit unless
1224 // the shift amount is >= the size of the datatype, which is undefined.
1225 if (DemandedMask == 1) {
1226 // Perform the logical shift right.
1227 Value *NewVal = BinaryOperator::CreateLShr(
1228 I->getOperand(0), I->getOperand(1), I->getName());
1229 InsertNewInstBefore(cast<Instruction>(NewVal), *I);
1230 return UpdateValueUsesWith(I, NewVal);
1233 // If the sign bit is the only bit demanded by this ashr, then there is no
1234 // need to do it, the shift doesn't change the high bit.
1235 if (DemandedMask.isSignBit())
1236 return UpdateValueUsesWith(I, I->getOperand(0));
1238 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1239 uint32_t ShiftAmt = SA->getLimitedValue(BitWidth);
1241 // Signed shift right.
1242 APInt DemandedMaskIn(DemandedMask.shl(ShiftAmt));
1243 // If any of the "high bits" are demanded, we should set the sign bit as
1245 if (DemandedMask.countLeadingZeros() <= ShiftAmt)
1246 DemandedMaskIn.set(BitWidth-1);
1247 if (SimplifyDemandedBits(I->getOperand(0),
1249 RHSKnownZero, RHSKnownOne, Depth+1))
1251 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1252 "Bits known to be one AND zero?");
1253 // Compute the new bits that are at the top now.
1254 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
1255 RHSKnownZero = APIntOps::lshr(RHSKnownZero, ShiftAmt);
1256 RHSKnownOne = APIntOps::lshr(RHSKnownOne, ShiftAmt);
1258 // Handle the sign bits.
1259 APInt SignBit(APInt::getSignBit(BitWidth));
1260 // Adjust to where it is now in the mask.
1261 SignBit = APIntOps::lshr(SignBit, ShiftAmt);
1263 // If the input sign bit is known to be zero, or if none of the top bits
1264 // are demanded, turn this into an unsigned shift right.
1265 if (BitWidth <= ShiftAmt || RHSKnownZero[BitWidth-ShiftAmt-1] ||
1266 (HighBits & ~DemandedMask) == HighBits) {
1267 // Perform the logical shift right.
1268 Value *NewVal = BinaryOperator::CreateLShr(
1269 I->getOperand(0), SA, I->getName());
1270 InsertNewInstBefore(cast<Instruction>(NewVal), *I);
1271 return UpdateValueUsesWith(I, NewVal);
1272 } else if ((RHSKnownOne & SignBit) != 0) { // New bits are known one.
1273 RHSKnownOne |= HighBits;
1277 case Instruction::SRem:
1278 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
1279 APInt RA = Rem->getValue().abs();
1280 if (RA.isPowerOf2()) {
1281 if (DemandedMask.ule(RA)) // srem won't affect demanded bits
1282 return UpdateValueUsesWith(I, I->getOperand(0));
1284 APInt LowBits = RA - 1;
1285 APInt Mask2 = LowBits | APInt::getSignBit(BitWidth);
1286 if (SimplifyDemandedBits(I->getOperand(0), Mask2,
1287 LHSKnownZero, LHSKnownOne, Depth+1))
1290 if (LHSKnownZero[BitWidth-1] || ((LHSKnownZero & LowBits) == LowBits))
1291 LHSKnownZero |= ~LowBits;
1293 KnownZero |= LHSKnownZero & DemandedMask;
1295 assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?");
1299 case Instruction::URem: {
1300 APInt KnownZero2(BitWidth, 0), KnownOne2(BitWidth, 0);
1301 APInt AllOnes = APInt::getAllOnesValue(BitWidth);
1302 if (SimplifyDemandedBits(I->getOperand(0), AllOnes,
1303 KnownZero2, KnownOne2, Depth+1))
1306 uint32_t Leaders = KnownZero2.countLeadingOnes();
1307 if (SimplifyDemandedBits(I->getOperand(1), AllOnes,
1308 KnownZero2, KnownOne2, Depth+1))
1311 Leaders = std::max(Leaders,
1312 KnownZero2.countLeadingOnes());
1313 KnownZero = APInt::getHighBitsSet(BitWidth, Leaders) & DemandedMask;
1316 case Instruction::Call:
1317 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
1318 switch (II->getIntrinsicID()) {
1320 case Intrinsic::bswap: {
1321 // If the only bits demanded come from one byte of the bswap result,
1322 // just shift the input byte into position to eliminate the bswap.
1323 unsigned NLZ = DemandedMask.countLeadingZeros();
1324 unsigned NTZ = DemandedMask.countTrailingZeros();
1326 // Round NTZ down to the next byte. If we have 11 trailing zeros, then
1327 // we need all the bits down to bit 8. Likewise, round NLZ. If we
1328 // have 14 leading zeros, round to 8.
1331 // If we need exactly one byte, we can do this transformation.
1332 if (BitWidth-NLZ-NTZ == 8) {
1333 unsigned ResultBit = NTZ;
1334 unsigned InputBit = BitWidth-NTZ-8;
1336 // Replace this with either a left or right shift to get the byte into
1338 Instruction *NewVal;
1339 if (InputBit > ResultBit)
1340 NewVal = BinaryOperator::CreateLShr(I->getOperand(1),
1341 ConstantInt::get(I->getType(), InputBit-ResultBit));
1343 NewVal = BinaryOperator::CreateShl(I->getOperand(1),
1344 ConstantInt::get(I->getType(), ResultBit-InputBit));
1345 NewVal->takeName(I);
1346 InsertNewInstBefore(NewVal, *I);
1347 return UpdateValueUsesWith(I, NewVal);
1350 // TODO: Could compute known zero/one bits based on the input.
1355 ComputeMaskedBits(V, DemandedMask, RHSKnownZero, RHSKnownOne, Depth);
1359 // If the client is only demanding bits that we know, return the known
1361 if ((DemandedMask & (RHSKnownZero|RHSKnownOne)) == DemandedMask)
1362 return UpdateValueUsesWith(I, ConstantInt::get(RHSKnownOne));
1367 /// SimplifyDemandedVectorElts - The specified value produces a vector with
1368 /// 64 or fewer elements. DemandedElts contains the set of elements that are
1369 /// actually used by the caller. This method analyzes which elements of the
1370 /// operand are undef and returns that information in UndefElts.
1372 /// If the information about demanded elements can be used to simplify the
1373 /// operation, the operation is simplified, then the resultant value is
1374 /// returned. This returns null if no change was made.
1375 Value *InstCombiner::SimplifyDemandedVectorElts(Value *V, uint64_t DemandedElts,
1376 uint64_t &UndefElts,
1378 unsigned VWidth = cast<VectorType>(V->getType())->getNumElements();
1379 assert(VWidth <= 64 && "Vector too wide to analyze!");
1380 uint64_t EltMask = ~0ULL >> (64-VWidth);
1381 assert((DemandedElts & ~EltMask) == 0 && "Invalid DemandedElts!");
1383 if (isa<UndefValue>(V)) {
1384 // If the entire vector is undefined, just return this info.
1385 UndefElts = EltMask;
1387 } else if (DemandedElts == 0) { // If nothing is demanded, provide undef.
1388 UndefElts = EltMask;
1389 return UndefValue::get(V->getType());
1393 if (ConstantVector *CP = dyn_cast<ConstantVector>(V)) {
1394 const Type *EltTy = cast<VectorType>(V->getType())->getElementType();
1395 Constant *Undef = UndefValue::get(EltTy);
1397 std::vector<Constant*> Elts;
1398 for (unsigned i = 0; i != VWidth; ++i)
1399 if (!(DemandedElts & (1ULL << i))) { // If not demanded, set to undef.
1400 Elts.push_back(Undef);
1401 UndefElts |= (1ULL << i);
1402 } else if (isa<UndefValue>(CP->getOperand(i))) { // Already undef.
1403 Elts.push_back(Undef);
1404 UndefElts |= (1ULL << i);
1405 } else { // Otherwise, defined.
1406 Elts.push_back(CP->getOperand(i));
1409 // If we changed the constant, return it.
1410 Constant *NewCP = ConstantVector::get(Elts);
1411 return NewCP != CP ? NewCP : 0;
1412 } else if (isa<ConstantAggregateZero>(V)) {
1413 // Simplify the CAZ to a ConstantVector where the non-demanded elements are
1416 // Check if this is identity. If so, return 0 since we are not simplifying
1418 if (DemandedElts == ((1ULL << VWidth) -1))
1421 const Type *EltTy = cast<VectorType>(V->getType())->getElementType();
1422 Constant *Zero = Constant::getNullValue(EltTy);
1423 Constant *Undef = UndefValue::get(EltTy);
1424 std::vector<Constant*> Elts;
1425 for (unsigned i = 0; i != VWidth; ++i)
1426 Elts.push_back((DemandedElts & (1ULL << i)) ? Zero : Undef);
1427 UndefElts = DemandedElts ^ EltMask;
1428 return ConstantVector::get(Elts);
1431 // Limit search depth.
1435 // If multiple users are using the root value, procede with
1436 // simplification conservatively assuming that all elements
1438 if (!V->hasOneUse()) {
1439 // Quit if we find multiple users of a non-root value though.
1440 // They'll be handled when it's their turn to be visited by
1441 // the main instcombine process.
1443 // TODO: Just compute the UndefElts information recursively.
1446 // Conservatively assume that all elements are needed.
1447 DemandedElts = EltMask;
1450 Instruction *I = dyn_cast<Instruction>(V);
1451 if (!I) return false; // Only analyze instructions.
1453 bool MadeChange = false;
1454 uint64_t UndefElts2;
1456 switch (I->getOpcode()) {
1459 case Instruction::InsertElement: {
1460 // If this is a variable index, we don't know which element it overwrites.
1461 // demand exactly the same input as we produce.
1462 ConstantInt *Idx = dyn_cast<ConstantInt>(I->getOperand(2));
1464 // Note that we can't propagate undef elt info, because we don't know
1465 // which elt is getting updated.
1466 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
1467 UndefElts2, Depth+1);
1468 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1472 // If this is inserting an element that isn't demanded, remove this
1474 unsigned IdxNo = Idx->getZExtValue();
1475 if (IdxNo >= VWidth || (DemandedElts & (1ULL << IdxNo)) == 0)
1476 return AddSoonDeadInstToWorklist(*I, 0);
1478 // Otherwise, the element inserted overwrites whatever was there, so the
1479 // input demanded set is simpler than the output set.
1480 TmpV = SimplifyDemandedVectorElts(I->getOperand(0),
1481 DemandedElts & ~(1ULL << IdxNo),
1482 UndefElts, Depth+1);
1483 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1485 // The inserted element is defined.
1486 UndefElts &= ~(1ULL << IdxNo);
1489 case Instruction::ShuffleVector: {
1490 ShuffleVectorInst *Shuffle = cast<ShuffleVectorInst>(I);
1491 uint64_t LHSVWidth =
1492 cast<VectorType>(Shuffle->getOperand(0)->getType())->getNumElements();
1493 uint64_t LeftDemanded = 0, RightDemanded = 0;
1494 for (unsigned i = 0; i < VWidth; i++) {
1495 if (DemandedElts & (1ULL << i)) {
1496 unsigned MaskVal = Shuffle->getMaskValue(i);
1497 if (MaskVal != -1u) {
1498 assert(MaskVal < LHSVWidth * 2 &&
1499 "shufflevector mask index out of range!");
1500 if (MaskVal < LHSVWidth)
1501 LeftDemanded |= 1ULL << MaskVal;
1503 RightDemanded |= 1ULL << (MaskVal - LHSVWidth);
1508 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), LeftDemanded,
1509 UndefElts2, Depth+1);
1510 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1512 uint64_t UndefElts3;
1513 TmpV = SimplifyDemandedVectorElts(I->getOperand(1), RightDemanded,
1514 UndefElts3, Depth+1);
1515 if (TmpV) { I->setOperand(1, TmpV); MadeChange = true; }
1517 bool NewUndefElts = false;
1518 for (unsigned i = 0; i < VWidth; i++) {
1519 unsigned MaskVal = Shuffle->getMaskValue(i);
1520 if (MaskVal == -1u) {
1521 uint64_t NewBit = 1ULL << i;
1522 UndefElts |= NewBit;
1523 } else if (MaskVal < LHSVWidth) {
1524 uint64_t NewBit = ((UndefElts2 >> MaskVal) & 1) << i;
1525 NewUndefElts |= NewBit;
1526 UndefElts |= NewBit;
1528 uint64_t NewBit = ((UndefElts3 >> (MaskVal - LHSVWidth)) & 1) << i;
1529 NewUndefElts |= NewBit;
1530 UndefElts |= NewBit;
1535 // Add additional discovered undefs.
1536 std::vector<Constant*> Elts;
1537 for (unsigned i = 0; i < VWidth; ++i) {
1538 if (UndefElts & (1ULL << i))
1539 Elts.push_back(UndefValue::get(Type::Int32Ty));
1541 Elts.push_back(ConstantInt::get(Type::Int32Ty,
1542 Shuffle->getMaskValue(i)));
1544 I->setOperand(2, ConstantVector::get(Elts));
1549 case Instruction::BitCast: {
1550 // Vector->vector casts only.
1551 const VectorType *VTy = dyn_cast<VectorType>(I->getOperand(0)->getType());
1553 unsigned InVWidth = VTy->getNumElements();
1554 uint64_t InputDemandedElts = 0;
1557 if (VWidth == InVWidth) {
1558 // If we are converting from <4 x i32> -> <4 x f32>, we demand the same
1559 // elements as are demanded of us.
1561 InputDemandedElts = DemandedElts;
1562 } else if (VWidth > InVWidth) {
1566 // If there are more elements in the result than there are in the source,
1567 // then an input element is live if any of the corresponding output
1568 // elements are live.
1569 Ratio = VWidth/InVWidth;
1570 for (unsigned OutIdx = 0; OutIdx != VWidth; ++OutIdx) {
1571 if (DemandedElts & (1ULL << OutIdx))
1572 InputDemandedElts |= 1ULL << (OutIdx/Ratio);
1578 // If there are more elements in the source than there are in the result,
1579 // then an input element is live if the corresponding output element is
1581 Ratio = InVWidth/VWidth;
1582 for (unsigned InIdx = 0; InIdx != InVWidth; ++InIdx)
1583 if (DemandedElts & (1ULL << InIdx/Ratio))
1584 InputDemandedElts |= 1ULL << InIdx;
1587 // div/rem demand all inputs, because they don't want divide by zero.
1588 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), InputDemandedElts,
1589 UndefElts2, Depth+1);
1591 I->setOperand(0, TmpV);
1595 UndefElts = UndefElts2;
1596 if (VWidth > InVWidth) {
1597 assert(0 && "Unimp");
1598 // If there are more elements in the result than there are in the source,
1599 // then an output element is undef if the corresponding input element is
1601 for (unsigned OutIdx = 0; OutIdx != VWidth; ++OutIdx)
1602 if (UndefElts2 & (1ULL << (OutIdx/Ratio)))
1603 UndefElts |= 1ULL << OutIdx;
1604 } else if (VWidth < InVWidth) {
1605 assert(0 && "Unimp");
1606 // If there are more elements in the source than there are in the result,
1607 // then a result element is undef if all of the corresponding input
1608 // elements are undef.
1609 UndefElts = ~0ULL >> (64-VWidth); // Start out all undef.
1610 for (unsigned InIdx = 0; InIdx != InVWidth; ++InIdx)
1611 if ((UndefElts2 & (1ULL << InIdx)) == 0) // Not undef?
1612 UndefElts &= ~(1ULL << (InIdx/Ratio)); // Clear undef bit.
1616 case Instruction::And:
1617 case Instruction::Or:
1618 case Instruction::Xor:
1619 case Instruction::Add:
1620 case Instruction::Sub:
1621 case Instruction::Mul:
1622 // div/rem demand all inputs, because they don't want divide by zero.
1623 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
1624 UndefElts, Depth+1);
1625 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1626 TmpV = SimplifyDemandedVectorElts(I->getOperand(1), DemandedElts,
1627 UndefElts2, Depth+1);
1628 if (TmpV) { I->setOperand(1, TmpV); MadeChange = true; }
1630 // Output elements are undefined if both are undefined. Consider things
1631 // like undef&0. The result is known zero, not undef.
1632 UndefElts &= UndefElts2;
1635 case Instruction::Call: {
1636 IntrinsicInst *II = dyn_cast<IntrinsicInst>(I);
1638 switch (II->getIntrinsicID()) {
1641 // Binary vector operations that work column-wise. A dest element is a
1642 // function of the corresponding input elements from the two inputs.
1643 case Intrinsic::x86_sse_sub_ss:
1644 case Intrinsic::x86_sse_mul_ss:
1645 case Intrinsic::x86_sse_min_ss:
1646 case Intrinsic::x86_sse_max_ss:
1647 case Intrinsic::x86_sse2_sub_sd:
1648 case Intrinsic::x86_sse2_mul_sd:
1649 case Intrinsic::x86_sse2_min_sd:
1650 case Intrinsic::x86_sse2_max_sd:
1651 TmpV = SimplifyDemandedVectorElts(II->getOperand(1), DemandedElts,
1652 UndefElts, Depth+1);
1653 if (TmpV) { II->setOperand(1, TmpV); MadeChange = true; }
1654 TmpV = SimplifyDemandedVectorElts(II->getOperand(2), DemandedElts,
1655 UndefElts2, Depth+1);
1656 if (TmpV) { II->setOperand(2, TmpV); MadeChange = true; }
1658 // If only the low elt is demanded and this is a scalarizable intrinsic,
1659 // scalarize it now.
1660 if (DemandedElts == 1) {
1661 switch (II->getIntrinsicID()) {
1663 case Intrinsic::x86_sse_sub_ss:
1664 case Intrinsic::x86_sse_mul_ss:
1665 case Intrinsic::x86_sse2_sub_sd:
1666 case Intrinsic::x86_sse2_mul_sd:
1667 // TODO: Lower MIN/MAX/ABS/etc
1668 Value *LHS = II->getOperand(1);
1669 Value *RHS = II->getOperand(2);
1670 // Extract the element as scalars.
1671 LHS = InsertNewInstBefore(new ExtractElementInst(LHS, 0U,"tmp"), *II);
1672 RHS = InsertNewInstBefore(new ExtractElementInst(RHS, 0U,"tmp"), *II);
1674 switch (II->getIntrinsicID()) {
1675 default: assert(0 && "Case stmts out of sync!");
1676 case Intrinsic::x86_sse_sub_ss:
1677 case Intrinsic::x86_sse2_sub_sd:
1678 TmpV = InsertNewInstBefore(BinaryOperator::CreateSub(LHS, RHS,
1679 II->getName()), *II);
1681 case Intrinsic::x86_sse_mul_ss:
1682 case Intrinsic::x86_sse2_mul_sd:
1683 TmpV = InsertNewInstBefore(BinaryOperator::CreateMul(LHS, RHS,
1684 II->getName()), *II);
1689 InsertElementInst::Create(UndefValue::get(II->getType()), TmpV, 0U,
1691 InsertNewInstBefore(New, *II);
1692 AddSoonDeadInstToWorklist(*II, 0);
1697 // Output elements are undefined if both are undefined. Consider things
1698 // like undef&0. The result is known zero, not undef.
1699 UndefElts &= UndefElts2;
1705 return MadeChange ? I : 0;
1709 /// AssociativeOpt - Perform an optimization on an associative operator. This
1710 /// function is designed to check a chain of associative operators for a
1711 /// potential to apply a certain optimization. Since the optimization may be
1712 /// applicable if the expression was reassociated, this checks the chain, then
1713 /// reassociates the expression as necessary to expose the optimization
1714 /// opportunity. This makes use of a special Functor, which must define
1715 /// 'shouldApply' and 'apply' methods.
1717 template<typename Functor>
1718 static Instruction *AssociativeOpt(BinaryOperator &Root, const Functor &F) {
1719 unsigned Opcode = Root.getOpcode();
1720 Value *LHS = Root.getOperand(0);
1722 // Quick check, see if the immediate LHS matches...
1723 if (F.shouldApply(LHS))
1724 return F.apply(Root);
1726 // Otherwise, if the LHS is not of the same opcode as the root, return.
1727 Instruction *LHSI = dyn_cast<Instruction>(LHS);
1728 while (LHSI && LHSI->getOpcode() == Opcode && LHSI->hasOneUse()) {
1729 // Should we apply this transform to the RHS?
1730 bool ShouldApply = F.shouldApply(LHSI->getOperand(1));
1732 // If not to the RHS, check to see if we should apply to the LHS...
1733 if (!ShouldApply && F.shouldApply(LHSI->getOperand(0))) {
1734 cast<BinaryOperator>(LHSI)->swapOperands(); // Make the LHS the RHS
1738 // If the functor wants to apply the optimization to the RHS of LHSI,
1739 // reassociate the expression from ((? op A) op B) to (? op (A op B))
1741 // Now all of the instructions are in the current basic block, go ahead
1742 // and perform the reassociation.
1743 Instruction *TmpLHSI = cast<Instruction>(Root.getOperand(0));
1745 // First move the selected RHS to the LHS of the root...
1746 Root.setOperand(0, LHSI->getOperand(1));
1748 // Make what used to be the LHS of the root be the user of the root...
1749 Value *ExtraOperand = TmpLHSI->getOperand(1);
1750 if (&Root == TmpLHSI) {
1751 Root.replaceAllUsesWith(Constant::getNullValue(TmpLHSI->getType()));
1754 Root.replaceAllUsesWith(TmpLHSI); // Users now use TmpLHSI
1755 TmpLHSI->setOperand(1, &Root); // TmpLHSI now uses the root
1756 BasicBlock::iterator ARI = &Root; ++ARI;
1757 TmpLHSI->moveBefore(ARI); // Move TmpLHSI to after Root
1760 // Now propagate the ExtraOperand down the chain of instructions until we
1762 while (TmpLHSI != LHSI) {
1763 Instruction *NextLHSI = cast<Instruction>(TmpLHSI->getOperand(0));
1764 // Move the instruction to immediately before the chain we are
1765 // constructing to avoid breaking dominance properties.
1766 NextLHSI->moveBefore(ARI);
1769 Value *NextOp = NextLHSI->getOperand(1);
1770 NextLHSI->setOperand(1, ExtraOperand);
1772 ExtraOperand = NextOp;
1775 // Now that the instructions are reassociated, have the functor perform
1776 // the transformation...
1777 return F.apply(Root);
1780 LHSI = dyn_cast<Instruction>(LHSI->getOperand(0));
1787 // AddRHS - Implements: X + X --> X << 1
1790 AddRHS(Value *rhs) : RHS(rhs) {}
1791 bool shouldApply(Value *LHS) const { return LHS == RHS; }
1792 Instruction *apply(BinaryOperator &Add) const {
1793 return BinaryOperator::CreateShl(Add.getOperand(0),
1794 ConstantInt::get(Add.getType(), 1));
1798 // AddMaskingAnd - Implements (A & C1)+(B & C2) --> (A & C1)|(B & C2)
1800 struct AddMaskingAnd {
1802 AddMaskingAnd(Constant *c) : C2(c) {}
1803 bool shouldApply(Value *LHS) const {
1805 return match(LHS, m_And(m_Value(), m_ConstantInt(C1))) &&
1806 ConstantExpr::getAnd(C1, C2)->isNullValue();
1808 Instruction *apply(BinaryOperator &Add) const {
1809 return BinaryOperator::CreateOr(Add.getOperand(0), Add.getOperand(1));
1815 static Value *FoldOperationIntoSelectOperand(Instruction &I, Value *SO,
1817 if (CastInst *CI = dyn_cast<CastInst>(&I)) {
1818 if (Constant *SOC = dyn_cast<Constant>(SO))
1819 return ConstantExpr::getCast(CI->getOpcode(), SOC, I.getType());
1821 return IC->InsertNewInstBefore(CastInst::Create(
1822 CI->getOpcode(), SO, I.getType(), SO->getName() + ".cast"), I);
1825 // Figure out if the constant is the left or the right argument.
1826 bool ConstIsRHS = isa<Constant>(I.getOperand(1));
1827 Constant *ConstOperand = cast<Constant>(I.getOperand(ConstIsRHS));
1829 if (Constant *SOC = dyn_cast<Constant>(SO)) {
1831 return ConstantExpr::get(I.getOpcode(), SOC, ConstOperand);
1832 return ConstantExpr::get(I.getOpcode(), ConstOperand, SOC);
1835 Value *Op0 = SO, *Op1 = ConstOperand;
1837 std::swap(Op0, Op1);
1839 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
1840 New = BinaryOperator::Create(BO->getOpcode(), Op0, Op1,SO->getName()+".op");
1841 else if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1842 New = CmpInst::Create(CI->getOpcode(), CI->getPredicate(), Op0, Op1,
1843 SO->getName()+".cmp");
1845 assert(0 && "Unknown binary instruction type!");
1848 return IC->InsertNewInstBefore(New, I);
1851 // FoldOpIntoSelect - Given an instruction with a select as one operand and a
1852 // constant as the other operand, try to fold the binary operator into the
1853 // select arguments. This also works for Cast instructions, which obviously do
1854 // not have a second operand.
1855 static Instruction *FoldOpIntoSelect(Instruction &Op, SelectInst *SI,
1857 // Don't modify shared select instructions
1858 if (!SI->hasOneUse()) return 0;
1859 Value *TV = SI->getOperand(1);
1860 Value *FV = SI->getOperand(2);
1862 if (isa<Constant>(TV) || isa<Constant>(FV)) {
1863 // Bool selects with constant operands can be folded to logical ops.
1864 if (SI->getType() == Type::Int1Ty) return 0;
1866 Value *SelectTrueVal = FoldOperationIntoSelectOperand(Op, TV, IC);
1867 Value *SelectFalseVal = FoldOperationIntoSelectOperand(Op, FV, IC);
1869 return SelectInst::Create(SI->getCondition(), SelectTrueVal,
1876 /// FoldOpIntoPhi - Given a binary operator or cast instruction which has a PHI
1877 /// node as operand #0, see if we can fold the instruction into the PHI (which
1878 /// is only possible if all operands to the PHI are constants).
1879 Instruction *InstCombiner::FoldOpIntoPhi(Instruction &I) {
1880 PHINode *PN = cast<PHINode>(I.getOperand(0));
1881 unsigned NumPHIValues = PN->getNumIncomingValues();
1882 if (!PN->hasOneUse() || NumPHIValues == 0) return 0;
1884 // Check to see if all of the operands of the PHI are constants. If there is
1885 // one non-constant value, remember the BB it is. If there is more than one
1886 // or if *it* is a PHI, bail out.
1887 BasicBlock *NonConstBB = 0;
1888 for (unsigned i = 0; i != NumPHIValues; ++i)
1889 if (!isa<Constant>(PN->getIncomingValue(i))) {
1890 if (NonConstBB) return 0; // More than one non-const value.
1891 if (isa<PHINode>(PN->getIncomingValue(i))) return 0; // Itself a phi.
1892 NonConstBB = PN->getIncomingBlock(i);
1894 // If the incoming non-constant value is in I's block, we have an infinite
1896 if (NonConstBB == I.getParent())
1900 // If there is exactly one non-constant value, we can insert a copy of the
1901 // operation in that block. However, if this is a critical edge, we would be
1902 // inserting the computation one some other paths (e.g. inside a loop). Only
1903 // do this if the pred block is unconditionally branching into the phi block.
1905 BranchInst *BI = dyn_cast<BranchInst>(NonConstBB->getTerminator());
1906 if (!BI || !BI->isUnconditional()) return 0;
1909 // Okay, we can do the transformation: create the new PHI node.
1910 PHINode *NewPN = PHINode::Create(I.getType(), "");
1911 NewPN->reserveOperandSpace(PN->getNumOperands()/2);
1912 InsertNewInstBefore(NewPN, *PN);
1913 NewPN->takeName(PN);
1915 // Next, add all of the operands to the PHI.
1916 if (I.getNumOperands() == 2) {
1917 Constant *C = cast<Constant>(I.getOperand(1));
1918 for (unsigned i = 0; i != NumPHIValues; ++i) {
1920 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
1921 if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1922 InV = ConstantExpr::getCompare(CI->getPredicate(), InC, C);
1924 InV = ConstantExpr::get(I.getOpcode(), InC, C);
1926 assert(PN->getIncomingBlock(i) == NonConstBB);
1927 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
1928 InV = BinaryOperator::Create(BO->getOpcode(),
1929 PN->getIncomingValue(i), C, "phitmp",
1930 NonConstBB->getTerminator());
1931 else if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1932 InV = CmpInst::Create(CI->getOpcode(),
1934 PN->getIncomingValue(i), C, "phitmp",
1935 NonConstBB->getTerminator());
1937 assert(0 && "Unknown binop!");
1939 AddToWorkList(cast<Instruction>(InV));
1941 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
1944 CastInst *CI = cast<CastInst>(&I);
1945 const Type *RetTy = CI->getType();
1946 for (unsigned i = 0; i != NumPHIValues; ++i) {
1948 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
1949 InV = ConstantExpr::getCast(CI->getOpcode(), InC, RetTy);
1951 assert(PN->getIncomingBlock(i) == NonConstBB);
1952 InV = CastInst::Create(CI->getOpcode(), PN->getIncomingValue(i),
1953 I.getType(), "phitmp",
1954 NonConstBB->getTerminator());
1955 AddToWorkList(cast<Instruction>(InV));
1957 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
1960 return ReplaceInstUsesWith(I, NewPN);
1964 /// WillNotOverflowSignedAdd - Return true if we can prove that:
1965 /// (sext (add LHS, RHS)) === (add (sext LHS), (sext RHS))
1966 /// This basically requires proving that the add in the original type would not
1967 /// overflow to change the sign bit or have a carry out.
1968 bool InstCombiner::WillNotOverflowSignedAdd(Value *LHS, Value *RHS) {
1969 // There are different heuristics we can use for this. Here are some simple
1972 // Add has the property that adding any two 2's complement numbers can only
1973 // have one carry bit which can change a sign. As such, if LHS and RHS each
1974 // have at least two sign bits, we know that the addition of the two values will
1975 // sign extend fine.
1976 if (ComputeNumSignBits(LHS) > 1 && ComputeNumSignBits(RHS) > 1)
1980 // If one of the operands only has one non-zero bit, and if the other operand
1981 // has a known-zero bit in a more significant place than it (not including the
1982 // sign bit) the ripple may go up to and fill the zero, but won't change the
1983 // sign. For example, (X & ~4) + 1.
1991 Instruction *InstCombiner::visitAdd(BinaryOperator &I) {
1992 bool Changed = SimplifyCommutative(I);
1993 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
1995 if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
1996 // X + undef -> undef
1997 if (isa<UndefValue>(RHS))
1998 return ReplaceInstUsesWith(I, RHS);
2001 if (!I.getType()->isFPOrFPVector()) { // NOTE: -0 + +0 = +0.
2002 if (RHSC->isNullValue())
2003 return ReplaceInstUsesWith(I, LHS);
2004 } else if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHSC)) {
2005 if (CFP->isExactlyValue(ConstantFP::getNegativeZero
2006 (I.getType())->getValueAPF()))
2007 return ReplaceInstUsesWith(I, LHS);
2010 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHSC)) {
2011 // X + (signbit) --> X ^ signbit
2012 const APInt& Val = CI->getValue();
2013 uint32_t BitWidth = Val.getBitWidth();
2014 if (Val == APInt::getSignBit(BitWidth))
2015 return BinaryOperator::CreateXor(LHS, RHS);
2017 // See if SimplifyDemandedBits can simplify this. This handles stuff like
2018 // (X & 254)+1 -> (X&254)|1
2019 if (!isa<VectorType>(I.getType())) {
2020 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
2021 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
2022 KnownZero, KnownOne))
2026 // zext(i1) - 1 -> select i1, 0, -1
2027 if (ZExtInst *ZI = dyn_cast<ZExtInst>(LHS))
2028 if (CI->isAllOnesValue() &&
2029 ZI->getOperand(0)->getType() == Type::Int1Ty)
2030 return SelectInst::Create(ZI->getOperand(0),
2031 Constant::getNullValue(I.getType()),
2032 ConstantInt::getAllOnesValue(I.getType()));
2035 if (isa<PHINode>(LHS))
2036 if (Instruction *NV = FoldOpIntoPhi(I))
2039 ConstantInt *XorRHS = 0;
2041 if (isa<ConstantInt>(RHSC) &&
2042 match(LHS, m_Xor(m_Value(XorLHS), m_ConstantInt(XorRHS)))) {
2043 uint32_t TySizeBits = I.getType()->getPrimitiveSizeInBits();
2044 const APInt& RHSVal = cast<ConstantInt>(RHSC)->getValue();
2046 uint32_t Size = TySizeBits / 2;
2047 APInt C0080Val(APInt(TySizeBits, 1ULL).shl(Size - 1));
2048 APInt CFF80Val(-C0080Val);
2050 if (TySizeBits > Size) {
2051 // If we have ADD(XOR(AND(X, 0xFF), 0x80), 0xF..F80), it's a sext.
2052 // If we have ADD(XOR(AND(X, 0xFF), 0xF..F80), 0x80), it's a sext.
2053 if ((RHSVal == CFF80Val && XorRHS->getValue() == C0080Val) ||
2054 (RHSVal == C0080Val && XorRHS->getValue() == CFF80Val)) {
2055 // This is a sign extend if the top bits are known zero.
2056 if (!MaskedValueIsZero(XorLHS,
2057 APInt::getHighBitsSet(TySizeBits, TySizeBits - Size)))
2058 Size = 0; // Not a sign ext, but can't be any others either.
2063 C0080Val = APIntOps::lshr(C0080Val, Size);
2064 CFF80Val = APIntOps::ashr(CFF80Val, Size);
2065 } while (Size >= 1);
2067 // FIXME: This shouldn't be necessary. When the backends can handle types
2068 // with funny bit widths then this switch statement should be removed. It
2069 // is just here to get the size of the "middle" type back up to something
2070 // that the back ends can handle.
2071 const Type *MiddleType = 0;
2074 case 32: MiddleType = Type::Int32Ty; break;
2075 case 16: MiddleType = Type::Int16Ty; break;
2076 case 8: MiddleType = Type::Int8Ty; break;
2079 Instruction *NewTrunc = new TruncInst(XorLHS, MiddleType, "sext");
2080 InsertNewInstBefore(NewTrunc, I);
2081 return new SExtInst(NewTrunc, I.getType(), I.getName());
2086 if (I.getType() == Type::Int1Ty)
2087 return BinaryOperator::CreateXor(LHS, RHS);
2090 if (I.getType()->isInteger()) {
2091 if (Instruction *Result = AssociativeOpt(I, AddRHS(RHS))) return Result;
2093 if (Instruction *RHSI = dyn_cast<Instruction>(RHS)) {
2094 if (RHSI->getOpcode() == Instruction::Sub)
2095 if (LHS == RHSI->getOperand(1)) // A + (B - A) --> B
2096 return ReplaceInstUsesWith(I, RHSI->getOperand(0));
2098 if (Instruction *LHSI = dyn_cast<Instruction>(LHS)) {
2099 if (LHSI->getOpcode() == Instruction::Sub)
2100 if (RHS == LHSI->getOperand(1)) // (B - A) + A --> B
2101 return ReplaceInstUsesWith(I, LHSI->getOperand(0));
2106 // -A + -B --> -(A + B)
2107 if (Value *LHSV = dyn_castNegVal(LHS)) {
2108 if (LHS->getType()->isIntOrIntVector()) {
2109 if (Value *RHSV = dyn_castNegVal(RHS)) {
2110 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSV, RHSV, "sum");
2111 InsertNewInstBefore(NewAdd, I);
2112 return BinaryOperator::CreateNeg(NewAdd);
2116 return BinaryOperator::CreateSub(RHS, LHSV);
2120 if (!isa<Constant>(RHS))
2121 if (Value *V = dyn_castNegVal(RHS))
2122 return BinaryOperator::CreateSub(LHS, V);
2126 if (Value *X = dyn_castFoldableMul(LHS, C2)) {
2127 if (X == RHS) // X*C + X --> X * (C+1)
2128 return BinaryOperator::CreateMul(RHS, AddOne(C2));
2130 // X*C1 + X*C2 --> X * (C1+C2)
2132 if (X == dyn_castFoldableMul(RHS, C1))
2133 return BinaryOperator::CreateMul(X, Add(C1, C2));
2136 // X + X*C --> X * (C+1)
2137 if (dyn_castFoldableMul(RHS, C2) == LHS)
2138 return BinaryOperator::CreateMul(LHS, AddOne(C2));
2140 // X + ~X --> -1 since ~X = -X-1
2141 if (dyn_castNotVal(LHS) == RHS || dyn_castNotVal(RHS) == LHS)
2142 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
2145 // (A & C1)+(B & C2) --> (A & C1)|(B & C2) iff C1&C2 == 0
2146 if (match(RHS, m_And(m_Value(), m_ConstantInt(C2))))
2147 if (Instruction *R = AssociativeOpt(I, AddMaskingAnd(C2)))
2150 // A+B --> A|B iff A and B have no bits set in common.
2151 if (const IntegerType *IT = dyn_cast<IntegerType>(I.getType())) {
2152 APInt Mask = APInt::getAllOnesValue(IT->getBitWidth());
2153 APInt LHSKnownOne(IT->getBitWidth(), 0);
2154 APInt LHSKnownZero(IT->getBitWidth(), 0);
2155 ComputeMaskedBits(LHS, Mask, LHSKnownZero, LHSKnownOne);
2156 if (LHSKnownZero != 0) {
2157 APInt RHSKnownOne(IT->getBitWidth(), 0);
2158 APInt RHSKnownZero(IT->getBitWidth(), 0);
2159 ComputeMaskedBits(RHS, Mask, RHSKnownZero, RHSKnownOne);
2161 // No bits in common -> bitwise or.
2162 if ((LHSKnownZero|RHSKnownZero).isAllOnesValue())
2163 return BinaryOperator::CreateOr(LHS, RHS);
2167 // W*X + Y*Z --> W * (X+Z) iff W == Y
2168 if (I.getType()->isIntOrIntVector()) {
2169 Value *W, *X, *Y, *Z;
2170 if (match(LHS, m_Mul(m_Value(W), m_Value(X))) &&
2171 match(RHS, m_Mul(m_Value(Y), m_Value(Z)))) {
2175 } else if (Y == X) {
2177 } else if (X == Z) {
2184 Value *NewAdd = InsertNewInstBefore(BinaryOperator::CreateAdd(X, Z,
2185 LHS->getName()), I);
2186 return BinaryOperator::CreateMul(W, NewAdd);
2191 if (ConstantInt *CRHS = dyn_cast<ConstantInt>(RHS)) {
2193 if (match(LHS, m_Not(m_Value(X)))) // ~X + C --> (C-1) - X
2194 return BinaryOperator::CreateSub(SubOne(CRHS), X);
2196 // (X & FF00) + xx00 -> (X+xx00) & FF00
2197 if (LHS->hasOneUse() && match(LHS, m_And(m_Value(X), m_ConstantInt(C2)))) {
2198 Constant *Anded = And(CRHS, C2);
2199 if (Anded == CRHS) {
2200 // See if all bits from the first bit set in the Add RHS up are included
2201 // in the mask. First, get the rightmost bit.
2202 const APInt& AddRHSV = CRHS->getValue();
2204 // Form a mask of all bits from the lowest bit added through the top.
2205 APInt AddRHSHighBits(~((AddRHSV & -AddRHSV)-1));
2207 // See if the and mask includes all of these bits.
2208 APInt AddRHSHighBitsAnd(AddRHSHighBits & C2->getValue());
2210 if (AddRHSHighBits == AddRHSHighBitsAnd) {
2211 // Okay, the xform is safe. Insert the new add pronto.
2212 Value *NewAdd = InsertNewInstBefore(BinaryOperator::CreateAdd(X, CRHS,
2213 LHS->getName()), I);
2214 return BinaryOperator::CreateAnd(NewAdd, C2);
2219 // Try to fold constant add into select arguments.
2220 if (SelectInst *SI = dyn_cast<SelectInst>(LHS))
2221 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2225 // add (cast *A to intptrtype) B ->
2226 // cast (GEP (cast *A to sbyte*) B) --> intptrtype
2228 CastInst *CI = dyn_cast<CastInst>(LHS);
2231 CI = dyn_cast<CastInst>(RHS);
2234 if (CI && CI->getType()->isSized() &&
2235 (CI->getType()->getPrimitiveSizeInBits() ==
2236 TD->getIntPtrType()->getPrimitiveSizeInBits())
2237 && isa<PointerType>(CI->getOperand(0)->getType())) {
2239 cast<PointerType>(CI->getOperand(0)->getType())->getAddressSpace();
2240 Value *I2 = InsertBitCastBefore(CI->getOperand(0),
2241 PointerType::get(Type::Int8Ty, AS), I);
2242 I2 = InsertNewInstBefore(GetElementPtrInst::Create(I2, Other, "ctg2"), I);
2243 return new PtrToIntInst(I2, CI->getType());
2247 // add (select X 0 (sub n A)) A --> select X A n
2249 SelectInst *SI = dyn_cast<SelectInst>(LHS);
2252 SI = dyn_cast<SelectInst>(RHS);
2255 if (SI && SI->hasOneUse()) {
2256 Value *TV = SI->getTrueValue();
2257 Value *FV = SI->getFalseValue();
2260 // Can we fold the add into the argument of the select?
2261 // We check both true and false select arguments for a matching subtract.
2262 if (match(FV, m_Zero()) && match(TV, m_Sub(m_Value(N), m_Specific(A))))
2263 // Fold the add into the true select value.
2264 return SelectInst::Create(SI->getCondition(), N, A);
2265 if (match(TV, m_Zero()) && match(FV, m_Sub(m_Value(N), m_Specific(A))))
2266 // Fold the add into the false select value.
2267 return SelectInst::Create(SI->getCondition(), A, N);
2271 // Check for X+0.0. Simplify it to X if we know X is not -0.0.
2272 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHS))
2273 if (CFP->getValueAPF().isPosZero() && CannotBeNegativeZero(LHS))
2274 return ReplaceInstUsesWith(I, LHS);
2276 // Check for (add (sext x), y), see if we can merge this into an
2277 // integer add followed by a sext.
2278 if (SExtInst *LHSConv = dyn_cast<SExtInst>(LHS)) {
2279 // (add (sext x), cst) --> (sext (add x, cst'))
2280 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS)) {
2282 ConstantExpr::getTrunc(RHSC, LHSConv->getOperand(0)->getType());
2283 if (LHSConv->hasOneUse() &&
2284 ConstantExpr::getSExt(CI, I.getType()) == RHSC &&
2285 WillNotOverflowSignedAdd(LHSConv->getOperand(0), CI)) {
2286 // Insert the new, smaller add.
2287 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSConv->getOperand(0),
2289 InsertNewInstBefore(NewAdd, I);
2290 return new SExtInst(NewAdd, I.getType());
2294 // (add (sext x), (sext y)) --> (sext (add int x, y))
2295 if (SExtInst *RHSConv = dyn_cast<SExtInst>(RHS)) {
2296 // Only do this if x/y have the same type, if at last one of them has a
2297 // single use (so we don't increase the number of sexts), and if the
2298 // integer add will not overflow.
2299 if (LHSConv->getOperand(0)->getType()==RHSConv->getOperand(0)->getType()&&
2300 (LHSConv->hasOneUse() || RHSConv->hasOneUse()) &&
2301 WillNotOverflowSignedAdd(LHSConv->getOperand(0),
2302 RHSConv->getOperand(0))) {
2303 // Insert the new integer add.
2304 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSConv->getOperand(0),
2305 RHSConv->getOperand(0),
2307 InsertNewInstBefore(NewAdd, I);
2308 return new SExtInst(NewAdd, I.getType());
2313 // Check for (add double (sitofp x), y), see if we can merge this into an
2314 // integer add followed by a promotion.
2315 if (SIToFPInst *LHSConv = dyn_cast<SIToFPInst>(LHS)) {
2316 // (add double (sitofp x), fpcst) --> (sitofp (add int x, intcst))
2317 // ... if the constant fits in the integer value. This is useful for things
2318 // like (double)(x & 1234) + 4.0 -> (double)((X & 1234)+4) which no longer
2319 // requires a constant pool load, and generally allows the add to be better
2321 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHS)) {
2323 ConstantExpr::getFPToSI(CFP, LHSConv->getOperand(0)->getType());
2324 if (LHSConv->hasOneUse() &&
2325 ConstantExpr::getSIToFP(CI, I.getType()) == CFP &&
2326 WillNotOverflowSignedAdd(LHSConv->getOperand(0), CI)) {
2327 // Insert the new integer add.
2328 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSConv->getOperand(0),
2330 InsertNewInstBefore(NewAdd, I);
2331 return new SIToFPInst(NewAdd, I.getType());
2335 // (add double (sitofp x), (sitofp y)) --> (sitofp (add int x, y))
2336 if (SIToFPInst *RHSConv = dyn_cast<SIToFPInst>(RHS)) {
2337 // Only do this if x/y have the same type, if at last one of them has a
2338 // single use (so we don't increase the number of int->fp conversions),
2339 // and if the integer add will not overflow.
2340 if (LHSConv->getOperand(0)->getType()==RHSConv->getOperand(0)->getType()&&
2341 (LHSConv->hasOneUse() || RHSConv->hasOneUse()) &&
2342 WillNotOverflowSignedAdd(LHSConv->getOperand(0),
2343 RHSConv->getOperand(0))) {
2344 // Insert the new integer add.
2345 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSConv->getOperand(0),
2346 RHSConv->getOperand(0),
2348 InsertNewInstBefore(NewAdd, I);
2349 return new SIToFPInst(NewAdd, I.getType());
2354 return Changed ? &I : 0;
2357 Instruction *InstCombiner::visitSub(BinaryOperator &I) {
2358 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2360 if (Op0 == Op1 && // sub X, X -> 0
2361 !I.getType()->isFPOrFPVector())
2362 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2364 // If this is a 'B = x-(-A)', change to B = x+A...
2365 if (Value *V = dyn_castNegVal(Op1))
2366 return BinaryOperator::CreateAdd(Op0, V);
2368 if (isa<UndefValue>(Op0))
2369 return ReplaceInstUsesWith(I, Op0); // undef - X -> undef
2370 if (isa<UndefValue>(Op1))
2371 return ReplaceInstUsesWith(I, Op1); // X - undef -> undef
2373 if (ConstantInt *C = dyn_cast<ConstantInt>(Op0)) {
2374 // Replace (-1 - A) with (~A)...
2375 if (C->isAllOnesValue())
2376 return BinaryOperator::CreateNot(Op1);
2378 // C - ~X == X + (1+C)
2380 if (match(Op1, m_Not(m_Value(X))))
2381 return BinaryOperator::CreateAdd(X, AddOne(C));
2383 // -(X >>u 31) -> (X >>s 31)
2384 // -(X >>s 31) -> (X >>u 31)
2386 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op1)) {
2387 if (SI->getOpcode() == Instruction::LShr) {
2388 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
2389 // Check to see if we are shifting out everything but the sign bit.
2390 if (CU->getLimitedValue(SI->getType()->getPrimitiveSizeInBits()) ==
2391 SI->getType()->getPrimitiveSizeInBits()-1) {
2392 // Ok, the transformation is safe. Insert AShr.
2393 return BinaryOperator::Create(Instruction::AShr,
2394 SI->getOperand(0), CU, SI->getName());
2398 else if (SI->getOpcode() == Instruction::AShr) {
2399 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
2400 // Check to see if we are shifting out everything but the sign bit.
2401 if (CU->getLimitedValue(SI->getType()->getPrimitiveSizeInBits()) ==
2402 SI->getType()->getPrimitiveSizeInBits()-1) {
2403 // Ok, the transformation is safe. Insert LShr.
2404 return BinaryOperator::CreateLShr(
2405 SI->getOperand(0), CU, SI->getName());
2412 // Try to fold constant sub into select arguments.
2413 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
2414 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2417 if (isa<PHINode>(Op0))
2418 if (Instruction *NV = FoldOpIntoPhi(I))
2422 if (I.getType() == Type::Int1Ty)
2423 return BinaryOperator::CreateXor(Op0, Op1);
2425 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
2426 if (Op1I->getOpcode() == Instruction::Add &&
2427 !Op0->getType()->isFPOrFPVector()) {
2428 if (Op1I->getOperand(0) == Op0) // X-(X+Y) == -Y
2429 return BinaryOperator::CreateNeg(Op1I->getOperand(1), I.getName());
2430 else if (Op1I->getOperand(1) == Op0) // X-(Y+X) == -Y
2431 return BinaryOperator::CreateNeg(Op1I->getOperand(0), I.getName());
2432 else if (ConstantInt *CI1 = dyn_cast<ConstantInt>(I.getOperand(0))) {
2433 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(Op1I->getOperand(1)))
2434 // C1-(X+C2) --> (C1-C2)-X
2435 return BinaryOperator::CreateSub(Subtract(CI1, CI2),
2436 Op1I->getOperand(0));
2440 if (Op1I->hasOneUse()) {
2441 // Replace (x - (y - z)) with (x + (z - y)) if the (y - z) subexpression
2442 // is not used by anyone else...
2444 if (Op1I->getOpcode() == Instruction::Sub &&
2445 !Op1I->getType()->isFPOrFPVector()) {
2446 // Swap the two operands of the subexpr...
2447 Value *IIOp0 = Op1I->getOperand(0), *IIOp1 = Op1I->getOperand(1);
2448 Op1I->setOperand(0, IIOp1);
2449 Op1I->setOperand(1, IIOp0);
2451 // Create the new top level add instruction...
2452 return BinaryOperator::CreateAdd(Op0, Op1);
2455 // Replace (A - (A & B)) with (A & ~B) if this is the only use of (A&B)...
2457 if (Op1I->getOpcode() == Instruction::And &&
2458 (Op1I->getOperand(0) == Op0 || Op1I->getOperand(1) == Op0)) {
2459 Value *OtherOp = Op1I->getOperand(Op1I->getOperand(0) == Op0);
2462 InsertNewInstBefore(BinaryOperator::CreateNot(OtherOp, "B.not"), I);
2463 return BinaryOperator::CreateAnd(Op0, NewNot);
2466 // 0 - (X sdiv C) -> (X sdiv -C)
2467 if (Op1I->getOpcode() == Instruction::SDiv)
2468 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
2470 if (Constant *DivRHS = dyn_cast<Constant>(Op1I->getOperand(1)))
2471 return BinaryOperator::CreateSDiv(Op1I->getOperand(0),
2472 ConstantExpr::getNeg(DivRHS));
2474 // X - X*C --> X * (1-C)
2475 ConstantInt *C2 = 0;
2476 if (dyn_castFoldableMul(Op1I, C2) == Op0) {
2477 Constant *CP1 = Subtract(ConstantInt::get(I.getType(), 1), C2);
2478 return BinaryOperator::CreateMul(Op0, CP1);
2483 if (!Op0->getType()->isFPOrFPVector())
2484 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
2485 if (Op0I->getOpcode() == Instruction::Add) {
2486 if (Op0I->getOperand(0) == Op1) // (Y+X)-Y == X
2487 return ReplaceInstUsesWith(I, Op0I->getOperand(1));
2488 else if (Op0I->getOperand(1) == Op1) // (X+Y)-Y == X
2489 return ReplaceInstUsesWith(I, Op0I->getOperand(0));
2490 } else if (Op0I->getOpcode() == Instruction::Sub) {
2491 if (Op0I->getOperand(0) == Op1) // (X-Y)-X == -Y
2492 return BinaryOperator::CreateNeg(Op0I->getOperand(1), I.getName());
2497 if (Value *X = dyn_castFoldableMul(Op0, C1)) {
2498 if (X == Op1) // X*C - X --> X * (C-1)
2499 return BinaryOperator::CreateMul(Op1, SubOne(C1));
2501 ConstantInt *C2; // X*C1 - X*C2 -> X * (C1-C2)
2502 if (X == dyn_castFoldableMul(Op1, C2))
2503 return BinaryOperator::CreateMul(X, Subtract(C1, C2));
2508 /// isSignBitCheck - Given an exploded icmp instruction, return true if the
2509 /// comparison only checks the sign bit. If it only checks the sign bit, set
2510 /// TrueIfSigned if the result of the comparison is true when the input value is
2512 static bool isSignBitCheck(ICmpInst::Predicate pred, ConstantInt *RHS,
2513 bool &TrueIfSigned) {
2515 case ICmpInst::ICMP_SLT: // True if LHS s< 0
2516 TrueIfSigned = true;
2517 return RHS->isZero();
2518 case ICmpInst::ICMP_SLE: // True if LHS s<= RHS and RHS == -1
2519 TrueIfSigned = true;
2520 return RHS->isAllOnesValue();
2521 case ICmpInst::ICMP_SGT: // True if LHS s> -1
2522 TrueIfSigned = false;
2523 return RHS->isAllOnesValue();
2524 case ICmpInst::ICMP_UGT:
2525 // True if LHS u> RHS and RHS == high-bit-mask - 1
2526 TrueIfSigned = true;
2527 return RHS->getValue() ==
2528 APInt::getSignedMaxValue(RHS->getType()->getPrimitiveSizeInBits());
2529 case ICmpInst::ICMP_UGE:
2530 // True if LHS u>= RHS and RHS == high-bit-mask (2^7, 2^15, 2^31, etc)
2531 TrueIfSigned = true;
2532 return RHS->getValue().isSignBit();
2538 Instruction *InstCombiner::visitMul(BinaryOperator &I) {
2539 bool Changed = SimplifyCommutative(I);
2540 Value *Op0 = I.getOperand(0);
2542 if (isa<UndefValue>(I.getOperand(1))) // undef * X -> 0
2543 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2545 // Simplify mul instructions with a constant RHS...
2546 if (Constant *Op1 = dyn_cast<Constant>(I.getOperand(1))) {
2547 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2549 // ((X << C1)*C2) == (X * (C2 << C1))
2550 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op0))
2551 if (SI->getOpcode() == Instruction::Shl)
2552 if (Constant *ShOp = dyn_cast<Constant>(SI->getOperand(1)))
2553 return BinaryOperator::CreateMul(SI->getOperand(0),
2554 ConstantExpr::getShl(CI, ShOp));
2557 return ReplaceInstUsesWith(I, Op1); // X * 0 == 0
2558 if (CI->equalsInt(1)) // X * 1 == X
2559 return ReplaceInstUsesWith(I, Op0);
2560 if (CI->isAllOnesValue()) // X * -1 == 0 - X
2561 return BinaryOperator::CreateNeg(Op0, I.getName());
2563 const APInt& Val = cast<ConstantInt>(CI)->getValue();
2564 if (Val.isPowerOf2()) { // Replace X*(2^C) with X << C
2565 return BinaryOperator::CreateShl(Op0,
2566 ConstantInt::get(Op0->getType(), Val.logBase2()));
2568 } else if (ConstantFP *Op1F = dyn_cast<ConstantFP>(Op1)) {
2569 if (Op1F->isNullValue())
2570 return ReplaceInstUsesWith(I, Op1);
2572 // "In IEEE floating point, x*1 is not equivalent to x for nans. However,
2573 // ANSI says we can drop signals, so we can do this anyway." (from GCC)
2574 if (Op1F->isExactlyValue(1.0))
2575 return ReplaceInstUsesWith(I, Op0); // Eliminate 'mul double %X, 1.0'
2576 } else if (isa<VectorType>(Op1->getType())) {
2577 if (isa<ConstantAggregateZero>(Op1))
2578 return ReplaceInstUsesWith(I, Op1);
2580 if (ConstantVector *Op1V = dyn_cast<ConstantVector>(Op1)) {
2581 if (Op1V->isAllOnesValue()) // X * -1 == 0 - X
2582 return BinaryOperator::CreateNeg(Op0, I.getName());
2584 // As above, vector X*splat(1.0) -> X in all defined cases.
2585 if (Constant *Splat = Op1V->getSplatValue()) {
2586 if (ConstantFP *F = dyn_cast<ConstantFP>(Splat))
2587 if (F->isExactlyValue(1.0))
2588 return ReplaceInstUsesWith(I, Op0);
2589 if (ConstantInt *CI = dyn_cast<ConstantInt>(Splat))
2590 if (CI->equalsInt(1))
2591 return ReplaceInstUsesWith(I, Op0);
2596 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
2597 if (Op0I->getOpcode() == Instruction::Add && Op0I->hasOneUse() &&
2598 isa<ConstantInt>(Op0I->getOperand(1)) && isa<ConstantInt>(Op1)) {
2599 // Canonicalize (X+C1)*C2 -> X*C2+C1*C2.
2600 Instruction *Add = BinaryOperator::CreateMul(Op0I->getOperand(0),
2602 InsertNewInstBefore(Add, I);
2603 Value *C1C2 = ConstantExpr::getMul(Op1,
2604 cast<Constant>(Op0I->getOperand(1)));
2605 return BinaryOperator::CreateAdd(Add, C1C2);
2609 // Try to fold constant mul into select arguments.
2610 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2611 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2614 if (isa<PHINode>(Op0))
2615 if (Instruction *NV = FoldOpIntoPhi(I))
2619 if (Value *Op0v = dyn_castNegVal(Op0)) // -X * -Y = X*Y
2620 if (Value *Op1v = dyn_castNegVal(I.getOperand(1)))
2621 return BinaryOperator::CreateMul(Op0v, Op1v);
2623 // (X / Y) * Y = X - (X % Y)
2624 // (X / Y) * -Y = (X % Y) - X
2626 Value *Op1 = I.getOperand(1);
2627 BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0);
2629 (BO->getOpcode() != Instruction::UDiv &&
2630 BO->getOpcode() != Instruction::SDiv)) {
2632 BO = dyn_cast<BinaryOperator>(I.getOperand(1));
2634 Value *Neg = dyn_castNegVal(Op1);
2635 if (BO && BO->hasOneUse() &&
2636 (BO->getOperand(1) == Op1 || BO->getOperand(1) == Neg) &&
2637 (BO->getOpcode() == Instruction::UDiv ||
2638 BO->getOpcode() == Instruction::SDiv)) {
2639 Value *Op0BO = BO->getOperand(0), *Op1BO = BO->getOperand(1);
2642 if (BO->getOpcode() == Instruction::UDiv)
2643 Rem = BinaryOperator::CreateURem(Op0BO, Op1BO);
2645 Rem = BinaryOperator::CreateSRem(Op0BO, Op1BO);
2647 InsertNewInstBefore(Rem, I);
2651 return BinaryOperator::CreateSub(Op0BO, Rem);
2653 return BinaryOperator::CreateSub(Rem, Op0BO);
2657 if (I.getType() == Type::Int1Ty)
2658 return BinaryOperator::CreateAnd(Op0, I.getOperand(1));
2660 // If one of the operands of the multiply is a cast from a boolean value, then
2661 // we know the bool is either zero or one, so this is a 'masking' multiply.
2662 // See if we can simplify things based on how the boolean was originally
2664 CastInst *BoolCast = 0;
2665 if (ZExtInst *CI = dyn_cast<ZExtInst>(Op0))
2666 if (CI->getOperand(0)->getType() == Type::Int1Ty)
2669 if (ZExtInst *CI = dyn_cast<ZExtInst>(I.getOperand(1)))
2670 if (CI->getOperand(0)->getType() == Type::Int1Ty)
2673 if (ICmpInst *SCI = dyn_cast<ICmpInst>(BoolCast->getOperand(0))) {
2674 Value *SCIOp0 = SCI->getOperand(0), *SCIOp1 = SCI->getOperand(1);
2675 const Type *SCOpTy = SCIOp0->getType();
2678 // If the icmp is true iff the sign bit of X is set, then convert this
2679 // multiply into a shift/and combination.
2680 if (isa<ConstantInt>(SCIOp1) &&
2681 isSignBitCheck(SCI->getPredicate(), cast<ConstantInt>(SCIOp1), TIS) &&
2683 // Shift the X value right to turn it into "all signbits".
2684 Constant *Amt = ConstantInt::get(SCIOp0->getType(),
2685 SCOpTy->getPrimitiveSizeInBits()-1);
2687 InsertNewInstBefore(
2688 BinaryOperator::Create(Instruction::AShr, SCIOp0, Amt,
2689 BoolCast->getOperand(0)->getName()+
2692 // If the multiply type is not the same as the source type, sign extend
2693 // or truncate to the multiply type.
2694 if (I.getType() != V->getType()) {
2695 uint32_t SrcBits = V->getType()->getPrimitiveSizeInBits();
2696 uint32_t DstBits = I.getType()->getPrimitiveSizeInBits();
2697 Instruction::CastOps opcode =
2698 (SrcBits == DstBits ? Instruction::BitCast :
2699 (SrcBits < DstBits ? Instruction::SExt : Instruction::Trunc));
2700 V = InsertCastBefore(opcode, V, I.getType(), I);
2703 Value *OtherOp = Op0 == BoolCast ? I.getOperand(1) : Op0;
2704 return BinaryOperator::CreateAnd(V, OtherOp);
2709 return Changed ? &I : 0;
2712 /// SimplifyDivRemOfSelect - Try to fold a divide or remainder of a select
2714 bool InstCombiner::SimplifyDivRemOfSelect(BinaryOperator &I) {
2715 SelectInst *SI = cast<SelectInst>(I.getOperand(1));
2717 // div/rem X, (Cond ? 0 : Y) -> div/rem X, Y
2718 int NonNullOperand = -1;
2719 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(1)))
2720 if (ST->isNullValue())
2722 // div/rem X, (Cond ? Y : 0) -> div/rem X, Y
2723 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(2)))
2724 if (ST->isNullValue())
2727 if (NonNullOperand == -1)
2730 Value *SelectCond = SI->getOperand(0);
2732 // Change the div/rem to use 'Y' instead of the select.
2733 I.setOperand(1, SI->getOperand(NonNullOperand));
2735 // Okay, we know we replace the operand of the div/rem with 'Y' with no
2736 // problem. However, the select, or the condition of the select may have
2737 // multiple uses. Based on our knowledge that the operand must be non-zero,
2738 // propagate the known value for the select into other uses of it, and
2739 // propagate a known value of the condition into its other users.
2741 // If the select and condition only have a single use, don't bother with this,
2743 if (SI->use_empty() && SelectCond->hasOneUse())
2746 // Scan the current block backward, looking for other uses of SI.
2747 BasicBlock::iterator BBI = &I, BBFront = I.getParent()->begin();
2749 while (BBI != BBFront) {
2751 // If we found a call to a function, we can't assume it will return, so
2752 // information from below it cannot be propagated above it.
2753 if (isa<CallInst>(BBI) && !isa<IntrinsicInst>(BBI))
2756 // Replace uses of the select or its condition with the known values.
2757 for (Instruction::op_iterator I = BBI->op_begin(), E = BBI->op_end();
2760 *I = SI->getOperand(NonNullOperand);
2762 } else if (*I == SelectCond) {
2763 *I = NonNullOperand == 1 ? ConstantInt::getTrue() :
2764 ConstantInt::getFalse();
2769 // If we past the instruction, quit looking for it.
2772 if (&*BBI == SelectCond)
2775 // If we ran out of things to eliminate, break out of the loop.
2776 if (SelectCond == 0 && SI == 0)
2784 /// This function implements the transforms on div instructions that work
2785 /// regardless of the kind of div instruction it is (udiv, sdiv, or fdiv). It is
2786 /// used by the visitors to those instructions.
2787 /// @brief Transforms common to all three div instructions
2788 Instruction *InstCombiner::commonDivTransforms(BinaryOperator &I) {
2789 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2791 // undef / X -> 0 for integer.
2792 // undef / X -> undef for FP (the undef could be a snan).
2793 if (isa<UndefValue>(Op0)) {
2794 if (Op0->getType()->isFPOrFPVector())
2795 return ReplaceInstUsesWith(I, Op0);
2796 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2799 // X / undef -> undef
2800 if (isa<UndefValue>(Op1))
2801 return ReplaceInstUsesWith(I, Op1);
2806 /// This function implements the transforms common to both integer division
2807 /// instructions (udiv and sdiv). It is called by the visitors to those integer
2808 /// division instructions.
2809 /// @brief Common integer divide transforms
2810 Instruction *InstCombiner::commonIDivTransforms(BinaryOperator &I) {
2811 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2813 // (sdiv X, X) --> 1 (udiv X, X) --> 1
2815 if (const VectorType *Ty = dyn_cast<VectorType>(I.getType())) {
2816 ConstantInt *CI = ConstantInt::get(Ty->getElementType(), 1);
2817 std::vector<Constant*> Elts(Ty->getNumElements(), CI);
2818 return ReplaceInstUsesWith(I, ConstantVector::get(Elts));
2821 ConstantInt *CI = ConstantInt::get(I.getType(), 1);
2822 return ReplaceInstUsesWith(I, CI);
2825 if (Instruction *Common = commonDivTransforms(I))
2828 // Handle cases involving: [su]div X, (select Cond, Y, Z)
2829 // This does not apply for fdiv.
2830 if (isa<SelectInst>(Op1) && SimplifyDivRemOfSelect(I))
2833 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2835 if (RHS->equalsInt(1))
2836 return ReplaceInstUsesWith(I, Op0);
2838 // (X / C1) / C2 -> X / (C1*C2)
2839 if (Instruction *LHS = dyn_cast<Instruction>(Op0))
2840 if (Instruction::BinaryOps(LHS->getOpcode()) == I.getOpcode())
2841 if (ConstantInt *LHSRHS = dyn_cast<ConstantInt>(LHS->getOperand(1))) {
2842 if (MultiplyOverflows(RHS, LHSRHS, I.getOpcode()==Instruction::SDiv))
2843 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2845 return BinaryOperator::Create(I.getOpcode(), LHS->getOperand(0),
2846 Multiply(RHS, LHSRHS));
2849 if (!RHS->isZero()) { // avoid X udiv 0
2850 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2851 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2853 if (isa<PHINode>(Op0))
2854 if (Instruction *NV = FoldOpIntoPhi(I))
2859 // 0 / X == 0, we don't need to preserve faults!
2860 if (ConstantInt *LHS = dyn_cast<ConstantInt>(Op0))
2861 if (LHS->equalsInt(0))
2862 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2864 // It can't be division by zero, hence it must be division by one.
2865 if (I.getType() == Type::Int1Ty)
2866 return ReplaceInstUsesWith(I, Op0);
2868 if (ConstantVector *Op1V = dyn_cast<ConstantVector>(Op1)) {
2869 if (ConstantInt *X = cast_or_null<ConstantInt>(Op1V->getSplatValue()))
2872 return ReplaceInstUsesWith(I, Op0);
2878 Instruction *InstCombiner::visitUDiv(BinaryOperator &I) {
2879 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2881 // Handle the integer div common cases
2882 if (Instruction *Common = commonIDivTransforms(I))
2885 // X udiv C^2 -> X >> C
2886 // Check to see if this is an unsigned division with an exact power of 2,
2887 // if so, convert to a right shift.
2888 if (ConstantInt *C = dyn_cast<ConstantInt>(Op1)) {
2889 if (C->getValue().isPowerOf2()) // 0 not included in isPowerOf2
2890 return BinaryOperator::CreateLShr(Op0,
2891 ConstantInt::get(Op0->getType(), C->getValue().logBase2()));
2894 // X udiv (C1 << N), where C1 is "1<<C2" --> X >> (N+C2)
2895 if (BinaryOperator *RHSI = dyn_cast<BinaryOperator>(I.getOperand(1))) {
2896 if (RHSI->getOpcode() == Instruction::Shl &&
2897 isa<ConstantInt>(RHSI->getOperand(0))) {
2898 const APInt& C1 = cast<ConstantInt>(RHSI->getOperand(0))->getValue();
2899 if (C1.isPowerOf2()) {
2900 Value *N = RHSI->getOperand(1);
2901 const Type *NTy = N->getType();
2902 if (uint32_t C2 = C1.logBase2()) {
2903 Constant *C2V = ConstantInt::get(NTy, C2);
2904 N = InsertNewInstBefore(BinaryOperator::CreateAdd(N, C2V, "tmp"), I);
2906 return BinaryOperator::CreateLShr(Op0, N);
2911 // udiv X, (Select Cond, C1, C2) --> Select Cond, (shr X, C1), (shr X, C2)
2912 // where C1&C2 are powers of two.
2913 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
2914 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
2915 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
2916 const APInt &TVA = STO->getValue(), &FVA = SFO->getValue();
2917 if (TVA.isPowerOf2() && FVA.isPowerOf2()) {
2918 // Compute the shift amounts
2919 uint32_t TSA = TVA.logBase2(), FSA = FVA.logBase2();
2920 // Construct the "on true" case of the select
2921 Constant *TC = ConstantInt::get(Op0->getType(), TSA);
2922 Instruction *TSI = BinaryOperator::CreateLShr(
2923 Op0, TC, SI->getName()+".t");
2924 TSI = InsertNewInstBefore(TSI, I);
2926 // Construct the "on false" case of the select
2927 Constant *FC = ConstantInt::get(Op0->getType(), FSA);
2928 Instruction *FSI = BinaryOperator::CreateLShr(
2929 Op0, FC, SI->getName()+".f");
2930 FSI = InsertNewInstBefore(FSI, I);
2932 // construct the select instruction and return it.
2933 return SelectInst::Create(SI->getOperand(0), TSI, FSI, SI->getName());
2939 Instruction *InstCombiner::visitSDiv(BinaryOperator &I) {
2940 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2942 // Handle the integer div common cases
2943 if (Instruction *Common = commonIDivTransforms(I))
2946 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2948 if (RHS->isAllOnesValue())
2949 return BinaryOperator::CreateNeg(Op0);
2952 if (Value *LHSNeg = dyn_castNegVal(Op0))
2953 return BinaryOperator::CreateSDiv(LHSNeg, ConstantExpr::getNeg(RHS));
2956 // If the sign bits of both operands are zero (i.e. we can prove they are
2957 // unsigned inputs), turn this into a udiv.
2958 if (I.getType()->isInteger()) {
2959 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()));
2960 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
2961 // X sdiv Y -> X udiv Y, iff X and Y don't have sign bit set
2962 return BinaryOperator::CreateUDiv(Op0, Op1, I.getName());
2969 Instruction *InstCombiner::visitFDiv(BinaryOperator &I) {
2970 return commonDivTransforms(I);
2973 /// This function implements the transforms on rem instructions that work
2974 /// regardless of the kind of rem instruction it is (urem, srem, or frem). It
2975 /// is used by the visitors to those instructions.
2976 /// @brief Transforms common to all three rem instructions
2977 Instruction *InstCombiner::commonRemTransforms(BinaryOperator &I) {
2978 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2980 // 0 % X == 0 for integer, we don't need to preserve faults!
2981 if (Constant *LHS = dyn_cast<Constant>(Op0))
2982 if (LHS->isNullValue())
2983 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2985 if (isa<UndefValue>(Op0)) { // undef % X -> 0
2986 if (I.getType()->isFPOrFPVector())
2987 return ReplaceInstUsesWith(I, Op0); // X % undef -> undef (could be SNaN)
2988 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2990 if (isa<UndefValue>(Op1))
2991 return ReplaceInstUsesWith(I, Op1); // X % undef -> undef
2993 // Handle cases involving: rem X, (select Cond, Y, Z)
2994 if (isa<SelectInst>(Op1) && SimplifyDivRemOfSelect(I))
3000 /// This function implements the transforms common to both integer remainder
3001 /// instructions (urem and srem). It is called by the visitors to those integer
3002 /// remainder instructions.
3003 /// @brief Common integer remainder transforms
3004 Instruction *InstCombiner::commonIRemTransforms(BinaryOperator &I) {
3005 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3007 if (Instruction *common = commonRemTransforms(I))
3010 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3011 // X % 0 == undef, we don't need to preserve faults!
3012 if (RHS->equalsInt(0))
3013 return ReplaceInstUsesWith(I, UndefValue::get(I.getType()));
3015 if (RHS->equalsInt(1)) // X % 1 == 0
3016 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3018 if (Instruction *Op0I = dyn_cast<Instruction>(Op0)) {
3019 if (SelectInst *SI = dyn_cast<SelectInst>(Op0I)) {
3020 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3022 } else if (isa<PHINode>(Op0I)) {
3023 if (Instruction *NV = FoldOpIntoPhi(I))
3027 // See if we can fold away this rem instruction.
3028 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
3029 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
3030 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
3031 KnownZero, KnownOne))
3039 Instruction *InstCombiner::visitURem(BinaryOperator &I) {
3040 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3042 if (Instruction *common = commonIRemTransforms(I))
3045 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3046 // X urem C^2 -> X and C
3047 // Check to see if this is an unsigned remainder with an exact power of 2,
3048 // if so, convert to a bitwise and.
3049 if (ConstantInt *C = dyn_cast<ConstantInt>(RHS))
3050 if (C->getValue().isPowerOf2())
3051 return BinaryOperator::CreateAnd(Op0, SubOne(C));
3054 if (Instruction *RHSI = dyn_cast<Instruction>(I.getOperand(1))) {
3055 // Turn A % (C << N), where C is 2^k, into A & ((C << N)-1)
3056 if (RHSI->getOpcode() == Instruction::Shl &&
3057 isa<ConstantInt>(RHSI->getOperand(0))) {
3058 if (cast<ConstantInt>(RHSI->getOperand(0))->getValue().isPowerOf2()) {
3059 Constant *N1 = ConstantInt::getAllOnesValue(I.getType());
3060 Value *Add = InsertNewInstBefore(BinaryOperator::CreateAdd(RHSI, N1,
3062 return BinaryOperator::CreateAnd(Op0, Add);
3067 // urem X, (select Cond, 2^C1, 2^C2) --> select Cond, (and X, C1), (and X, C2)
3068 // where C1&C2 are powers of two.
3069 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
3070 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
3071 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
3072 // STO == 0 and SFO == 0 handled above.
3073 if ((STO->getValue().isPowerOf2()) &&
3074 (SFO->getValue().isPowerOf2())) {
3075 Value *TrueAnd = InsertNewInstBefore(
3076 BinaryOperator::CreateAnd(Op0, SubOne(STO), SI->getName()+".t"), I);
3077 Value *FalseAnd = InsertNewInstBefore(
3078 BinaryOperator::CreateAnd(Op0, SubOne(SFO), SI->getName()+".f"), I);
3079 return SelectInst::Create(SI->getOperand(0), TrueAnd, FalseAnd);
3087 Instruction *InstCombiner::visitSRem(BinaryOperator &I) {
3088 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3090 // Handle the integer rem common cases
3091 if (Instruction *common = commonIRemTransforms(I))
3094 if (Value *RHSNeg = dyn_castNegVal(Op1))
3095 if (!isa<Constant>(RHSNeg) ||
3096 (isa<ConstantInt>(RHSNeg) &&
3097 cast<ConstantInt>(RHSNeg)->getValue().isStrictlyPositive())) {
3099 AddUsesToWorkList(I);
3100 I.setOperand(1, RHSNeg);
3104 // If the sign bits of both operands are zero (i.e. we can prove they are
3105 // unsigned inputs), turn this into a urem.
3106 if (I.getType()->isInteger()) {
3107 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()));
3108 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
3109 // X srem Y -> X urem Y, iff X and Y don't have sign bit set
3110 return BinaryOperator::CreateURem(Op0, Op1, I.getName());
3117 Instruction *InstCombiner::visitFRem(BinaryOperator &I) {
3118 return commonRemTransforms(I);
3121 // isOneBitSet - Return true if there is exactly one bit set in the specified
3123 static bool isOneBitSet(const ConstantInt *CI) {
3124 return CI->getValue().isPowerOf2();
3127 // isHighOnes - Return true if the constant is of the form 1+0+.
3128 // This is the same as lowones(~X).
3129 static bool isHighOnes(const ConstantInt *CI) {
3130 return (~CI->getValue() + 1).isPowerOf2();
3133 /// getICmpCode - Encode a icmp predicate into a three bit mask. These bits
3134 /// are carefully arranged to allow folding of expressions such as:
3136 /// (A < B) | (A > B) --> (A != B)
3138 /// Note that this is only valid if the first and second predicates have the
3139 /// same sign. Is illegal to do: (A u< B) | (A s> B)
3141 /// Three bits are used to represent the condition, as follows:
3146 /// <=> Value Definition
3147 /// 000 0 Always false
3154 /// 111 7 Always true
3156 static unsigned getICmpCode(const ICmpInst *ICI) {
3157 switch (ICI->getPredicate()) {
3159 case ICmpInst::ICMP_UGT: return 1; // 001
3160 case ICmpInst::ICMP_SGT: return 1; // 001
3161 case ICmpInst::ICMP_EQ: return 2; // 010
3162 case ICmpInst::ICMP_UGE: return 3; // 011
3163 case ICmpInst::ICMP_SGE: return 3; // 011
3164 case ICmpInst::ICMP_ULT: return 4; // 100
3165 case ICmpInst::ICMP_SLT: return 4; // 100
3166 case ICmpInst::ICMP_NE: return 5; // 101
3167 case ICmpInst::ICMP_ULE: return 6; // 110
3168 case ICmpInst::ICMP_SLE: return 6; // 110
3171 assert(0 && "Invalid ICmp predicate!");
3176 /// getFCmpCode - Similar to getICmpCode but for FCmpInst. This encodes a fcmp
3177 /// predicate into a three bit mask. It also returns whether it is an ordered
3178 /// predicate by reference.
3179 static unsigned getFCmpCode(FCmpInst::Predicate CC, bool &isOrdered) {
3182 case FCmpInst::FCMP_ORD: isOrdered = true; return 0; // 000
3183 case FCmpInst::FCMP_UNO: return 0; // 000
3184 case FCmpInst::FCMP_OGT: isOrdered = true; return 1; // 001
3185 case FCmpInst::FCMP_UGT: return 1; // 001
3186 case FCmpInst::FCMP_OEQ: isOrdered = true; return 2; // 010
3187 case FCmpInst::FCMP_UEQ: return 2; // 010
3188 case FCmpInst::FCMP_OGE: isOrdered = true; return 3; // 011
3189 case FCmpInst::FCMP_UGE: return 3; // 011
3190 case FCmpInst::FCMP_OLT: isOrdered = true; return 4; // 100
3191 case FCmpInst::FCMP_ULT: return 4; // 100
3192 case FCmpInst::FCMP_ONE: isOrdered = true; return 5; // 101
3193 case FCmpInst::FCMP_UNE: return 5; // 101
3194 case FCmpInst::FCMP_OLE: isOrdered = true; return 6; // 110
3195 case FCmpInst::FCMP_ULE: return 6; // 110
3198 // Not expecting FCMP_FALSE and FCMP_TRUE;
3199 assert(0 && "Unexpected FCmp predicate!");
3204 /// getICmpValue - This is the complement of getICmpCode, which turns an
3205 /// opcode and two operands into either a constant true or false, or a brand
3206 /// new ICmp instruction. The sign is passed in to determine which kind
3207 /// of predicate to use in the new icmp instruction.
3208 static Value *getICmpValue(bool sign, unsigned code, Value *LHS, Value *RHS) {
3210 default: assert(0 && "Illegal ICmp code!");
3211 case 0: return ConstantInt::getFalse();
3214 return new ICmpInst(ICmpInst::ICMP_SGT, LHS, RHS);
3216 return new ICmpInst(ICmpInst::ICMP_UGT, LHS, RHS);
3217 case 2: return new ICmpInst(ICmpInst::ICMP_EQ, LHS, RHS);
3220 return new ICmpInst(ICmpInst::ICMP_SGE, LHS, RHS);
3222 return new ICmpInst(ICmpInst::ICMP_UGE, LHS, RHS);
3225 return new ICmpInst(ICmpInst::ICMP_SLT, LHS, RHS);
3227 return new ICmpInst(ICmpInst::ICMP_ULT, LHS, RHS);
3228 case 5: return new ICmpInst(ICmpInst::ICMP_NE, LHS, RHS);
3231 return new ICmpInst(ICmpInst::ICMP_SLE, LHS, RHS);
3233 return new ICmpInst(ICmpInst::ICMP_ULE, LHS, RHS);
3234 case 7: return ConstantInt::getTrue();
3238 /// getFCmpValue - This is the complement of getFCmpCode, which turns an
3239 /// opcode and two operands into either a FCmp instruction. isordered is passed
3240 /// in to determine which kind of predicate to use in the new fcmp instruction.
3241 static Value *getFCmpValue(bool isordered, unsigned code,
3242 Value *LHS, Value *RHS) {
3244 default: assert(0 && "Illegal FCmp code!");
3247 return new FCmpInst(FCmpInst::FCMP_ORD, LHS, RHS);
3249 return new FCmpInst(FCmpInst::FCMP_UNO, LHS, RHS);
3252 return new FCmpInst(FCmpInst::FCMP_OGT, LHS, RHS);
3254 return new FCmpInst(FCmpInst::FCMP_UGT, LHS, RHS);
3257 return new FCmpInst(FCmpInst::FCMP_OEQ, LHS, RHS);
3259 return new FCmpInst(FCmpInst::FCMP_UEQ, LHS, RHS);
3262 return new FCmpInst(FCmpInst::FCMP_OGE, LHS, RHS);
3264 return new FCmpInst(FCmpInst::FCMP_UGE, LHS, RHS);
3267 return new FCmpInst(FCmpInst::FCMP_OLT, LHS, RHS);
3269 return new FCmpInst(FCmpInst::FCMP_ULT, LHS, RHS);
3272 return new FCmpInst(FCmpInst::FCMP_ONE, LHS, RHS);
3274 return new FCmpInst(FCmpInst::FCMP_UNE, LHS, RHS);
3277 return new FCmpInst(FCmpInst::FCMP_OLE, LHS, RHS);
3279 return new FCmpInst(FCmpInst::FCMP_ULE, LHS, RHS);
3280 case 7: return ConstantInt::getTrue();
3284 /// PredicatesFoldable - Return true if both predicates match sign or if at
3285 /// least one of them is an equality comparison (which is signless).
3286 static bool PredicatesFoldable(ICmpInst::Predicate p1, ICmpInst::Predicate p2) {
3287 return (ICmpInst::isSignedPredicate(p1) == ICmpInst::isSignedPredicate(p2)) ||
3288 (ICmpInst::isSignedPredicate(p1) && ICmpInst::isEquality(p2)) ||
3289 (ICmpInst::isSignedPredicate(p2) && ICmpInst::isEquality(p1));
3293 // FoldICmpLogical - Implements (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
3294 struct FoldICmpLogical {
3297 ICmpInst::Predicate pred;
3298 FoldICmpLogical(InstCombiner &ic, ICmpInst *ICI)
3299 : IC(ic), LHS(ICI->getOperand(0)), RHS(ICI->getOperand(1)),
3300 pred(ICI->getPredicate()) {}
3301 bool shouldApply(Value *V) const {
3302 if (ICmpInst *ICI = dyn_cast<ICmpInst>(V))
3303 if (PredicatesFoldable(pred, ICI->getPredicate()))
3304 return ((ICI->getOperand(0) == LHS && ICI->getOperand(1) == RHS) ||
3305 (ICI->getOperand(0) == RHS && ICI->getOperand(1) == LHS));
3308 Instruction *apply(Instruction &Log) const {
3309 ICmpInst *ICI = cast<ICmpInst>(Log.getOperand(0));
3310 if (ICI->getOperand(0) != LHS) {
3311 assert(ICI->getOperand(1) == LHS);
3312 ICI->swapOperands(); // Swap the LHS and RHS of the ICmp
3315 ICmpInst *RHSICI = cast<ICmpInst>(Log.getOperand(1));
3316 unsigned LHSCode = getICmpCode(ICI);
3317 unsigned RHSCode = getICmpCode(RHSICI);
3319 switch (Log.getOpcode()) {
3320 case Instruction::And: Code = LHSCode & RHSCode; break;
3321 case Instruction::Or: Code = LHSCode | RHSCode; break;
3322 case Instruction::Xor: Code = LHSCode ^ RHSCode; break;
3323 default: assert(0 && "Illegal logical opcode!"); return 0;
3326 bool isSigned = ICmpInst::isSignedPredicate(RHSICI->getPredicate()) ||
3327 ICmpInst::isSignedPredicate(ICI->getPredicate());
3329 Value *RV = getICmpValue(isSigned, Code, LHS, RHS);
3330 if (Instruction *I = dyn_cast<Instruction>(RV))
3332 // Otherwise, it's a constant boolean value...
3333 return IC.ReplaceInstUsesWith(Log, RV);
3336 } // end anonymous namespace
3338 // OptAndOp - This handles expressions of the form ((val OP C1) & C2). Where
3339 // the Op parameter is 'OP', OpRHS is 'C1', and AndRHS is 'C2'. Op is
3340 // guaranteed to be a binary operator.
3341 Instruction *InstCombiner::OptAndOp(Instruction *Op,
3343 ConstantInt *AndRHS,
3344 BinaryOperator &TheAnd) {
3345 Value *X = Op->getOperand(0);
3346 Constant *Together = 0;
3348 Together = And(AndRHS, OpRHS);
3350 switch (Op->getOpcode()) {
3351 case Instruction::Xor:
3352 if (Op->hasOneUse()) {
3353 // (X ^ C1) & C2 --> (X & C2) ^ (C1&C2)
3354 Instruction *And = BinaryOperator::CreateAnd(X, AndRHS);
3355 InsertNewInstBefore(And, TheAnd);
3357 return BinaryOperator::CreateXor(And, Together);
3360 case Instruction::Or:
3361 if (Together == AndRHS) // (X | C) & C --> C
3362 return ReplaceInstUsesWith(TheAnd, AndRHS);
3364 if (Op->hasOneUse() && Together != OpRHS) {
3365 // (X | C1) & C2 --> (X | (C1&C2)) & C2
3366 Instruction *Or = BinaryOperator::CreateOr(X, Together);
3367 InsertNewInstBefore(Or, TheAnd);
3369 return BinaryOperator::CreateAnd(Or, AndRHS);
3372 case Instruction::Add:
3373 if (Op->hasOneUse()) {
3374 // Adding a one to a single bit bit-field should be turned into an XOR
3375 // of the bit. First thing to check is to see if this AND is with a
3376 // single bit constant.
3377 const APInt& AndRHSV = cast<ConstantInt>(AndRHS)->getValue();
3379 // If there is only one bit set...
3380 if (isOneBitSet(cast<ConstantInt>(AndRHS))) {
3381 // Ok, at this point, we know that we are masking the result of the
3382 // ADD down to exactly one bit. If the constant we are adding has
3383 // no bits set below this bit, then we can eliminate the ADD.
3384 const APInt& AddRHS = cast<ConstantInt>(OpRHS)->getValue();
3386 // Check to see if any bits below the one bit set in AndRHSV are set.
3387 if ((AddRHS & (AndRHSV-1)) == 0) {
3388 // If not, the only thing that can effect the output of the AND is
3389 // the bit specified by AndRHSV. If that bit is set, the effect of
3390 // the XOR is to toggle the bit. If it is clear, then the ADD has
3392 if ((AddRHS & AndRHSV) == 0) { // Bit is not set, noop
3393 TheAnd.setOperand(0, X);
3396 // Pull the XOR out of the AND.
3397 Instruction *NewAnd = BinaryOperator::CreateAnd(X, AndRHS);
3398 InsertNewInstBefore(NewAnd, TheAnd);
3399 NewAnd->takeName(Op);
3400 return BinaryOperator::CreateXor(NewAnd, AndRHS);
3407 case Instruction::Shl: {
3408 // We know that the AND will not produce any of the bits shifted in, so if
3409 // the anded constant includes them, clear them now!
3411 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3412 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3413 APInt ShlMask(APInt::getHighBitsSet(BitWidth, BitWidth-OpRHSVal));
3414 ConstantInt *CI = ConstantInt::get(AndRHS->getValue() & ShlMask);
3416 if (CI->getValue() == ShlMask) {
3417 // Masking out bits that the shift already masks
3418 return ReplaceInstUsesWith(TheAnd, Op); // No need for the and.
3419 } else if (CI != AndRHS) { // Reducing bits set in and.
3420 TheAnd.setOperand(1, CI);
3425 case Instruction::LShr:
3427 // We know that the AND will not produce any of the bits shifted in, so if
3428 // the anded constant includes them, clear them now! This only applies to
3429 // unsigned shifts, because a signed shr may bring in set bits!
3431 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3432 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3433 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
3434 ConstantInt *CI = ConstantInt::get(AndRHS->getValue() & ShrMask);
3436 if (CI->getValue() == ShrMask) {
3437 // Masking out bits that the shift already masks.
3438 return ReplaceInstUsesWith(TheAnd, Op);
3439 } else if (CI != AndRHS) {
3440 TheAnd.setOperand(1, CI); // Reduce bits set in and cst.
3445 case Instruction::AShr:
3447 // See if this is shifting in some sign extension, then masking it out
3449 if (Op->hasOneUse()) {
3450 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3451 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3452 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
3453 Constant *C = ConstantInt::get(AndRHS->getValue() & ShrMask);
3454 if (C == AndRHS) { // Masking out bits shifted in.
3455 // (Val ashr C1) & C2 -> (Val lshr C1) & C2
3456 // Make the argument unsigned.
3457 Value *ShVal = Op->getOperand(0);
3458 ShVal = InsertNewInstBefore(
3459 BinaryOperator::CreateLShr(ShVal, OpRHS,
3460 Op->getName()), TheAnd);
3461 return BinaryOperator::CreateAnd(ShVal, AndRHS, TheAnd.getName());
3470 /// InsertRangeTest - Emit a computation of: (V >= Lo && V < Hi) if Inside is
3471 /// true, otherwise (V < Lo || V >= Hi). In pratice, we emit the more efficient
3472 /// (V-Lo) <u Hi-Lo. This method expects that Lo <= Hi. isSigned indicates
3473 /// whether to treat the V, Lo and HI as signed or not. IB is the location to
3474 /// insert new instructions.
3475 Instruction *InstCombiner::InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
3476 bool isSigned, bool Inside,
3478 assert(cast<ConstantInt>(ConstantExpr::getICmp((isSigned ?
3479 ICmpInst::ICMP_SLE:ICmpInst::ICMP_ULE), Lo, Hi))->getZExtValue() &&
3480 "Lo is not <= Hi in range emission code!");
3483 if (Lo == Hi) // Trivially false.
3484 return new ICmpInst(ICmpInst::ICMP_NE, V, V);
3486 // V >= Min && V < Hi --> V < Hi
3487 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
3488 ICmpInst::Predicate pred = (isSigned ?
3489 ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT);
3490 return new ICmpInst(pred, V, Hi);
3493 // Emit V-Lo <u Hi-Lo
3494 Constant *NegLo = ConstantExpr::getNeg(Lo);
3495 Instruction *Add = BinaryOperator::CreateAdd(V, NegLo, V->getName()+".off");
3496 InsertNewInstBefore(Add, IB);
3497 Constant *UpperBound = ConstantExpr::getAdd(NegLo, Hi);
3498 return new ICmpInst(ICmpInst::ICMP_ULT, Add, UpperBound);
3501 if (Lo == Hi) // Trivially true.
3502 return new ICmpInst(ICmpInst::ICMP_EQ, V, V);
3504 // V < Min || V >= Hi -> V > Hi-1
3505 Hi = SubOne(cast<ConstantInt>(Hi));
3506 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
3507 ICmpInst::Predicate pred = (isSigned ?
3508 ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT);
3509 return new ICmpInst(pred, V, Hi);
3512 // Emit V-Lo >u Hi-1-Lo
3513 // Note that Hi has already had one subtracted from it, above.
3514 ConstantInt *NegLo = cast<ConstantInt>(ConstantExpr::getNeg(Lo));
3515 Instruction *Add = BinaryOperator::CreateAdd(V, NegLo, V->getName()+".off");
3516 InsertNewInstBefore(Add, IB);
3517 Constant *LowerBound = ConstantExpr::getAdd(NegLo, Hi);
3518 return new ICmpInst(ICmpInst::ICMP_UGT, Add, LowerBound);
3521 // isRunOfOnes - Returns true iff Val consists of one contiguous run of 1s with
3522 // any number of 0s on either side. The 1s are allowed to wrap from LSB to
3523 // MSB, so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs. 0x0F0F0000 is
3524 // not, since all 1s are not contiguous.
3525 static bool isRunOfOnes(ConstantInt *Val, uint32_t &MB, uint32_t &ME) {
3526 const APInt& V = Val->getValue();
3527 uint32_t BitWidth = Val->getType()->getBitWidth();
3528 if (!APIntOps::isShiftedMask(BitWidth, V)) return false;
3530 // look for the first zero bit after the run of ones
3531 MB = BitWidth - ((V - 1) ^ V).countLeadingZeros();
3532 // look for the first non-zero bit
3533 ME = V.getActiveBits();
3537 /// FoldLogicalPlusAnd - This is part of an expression (LHS +/- RHS) & Mask,
3538 /// where isSub determines whether the operator is a sub. If we can fold one of
3539 /// the following xforms:
3541 /// ((A & N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == Mask
3542 /// ((A | N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3543 /// ((A ^ N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3545 /// return (A +/- B).
3547 Value *InstCombiner::FoldLogicalPlusAnd(Value *LHS, Value *RHS,
3548 ConstantInt *Mask, bool isSub,
3550 Instruction *LHSI = dyn_cast<Instruction>(LHS);
3551 if (!LHSI || LHSI->getNumOperands() != 2 ||
3552 !isa<ConstantInt>(LHSI->getOperand(1))) return 0;
3554 ConstantInt *N = cast<ConstantInt>(LHSI->getOperand(1));
3556 switch (LHSI->getOpcode()) {
3558 case Instruction::And:
3559 if (And(N, Mask) == Mask) {
3560 // If the AndRHS is a power of two minus one (0+1+), this is simple.
3561 if ((Mask->getValue().countLeadingZeros() +
3562 Mask->getValue().countPopulation()) ==
3563 Mask->getValue().getBitWidth())
3566 // Otherwise, if Mask is 0+1+0+, and if B is known to have the low 0+
3567 // part, we don't need any explicit masks to take them out of A. If that
3568 // is all N is, ignore it.
3569 uint32_t MB = 0, ME = 0;
3570 if (isRunOfOnes(Mask, MB, ME)) { // begin/end bit of run, inclusive
3571 uint32_t BitWidth = cast<IntegerType>(RHS->getType())->getBitWidth();
3572 APInt Mask(APInt::getLowBitsSet(BitWidth, MB-1));
3573 if (MaskedValueIsZero(RHS, Mask))
3578 case Instruction::Or:
3579 case Instruction::Xor:
3580 // If the AndRHS is a power of two minus one (0+1+), and N&Mask == 0
3581 if ((Mask->getValue().countLeadingZeros() +
3582 Mask->getValue().countPopulation()) == Mask->getValue().getBitWidth()
3583 && And(N, Mask)->isZero())
3590 New = BinaryOperator::CreateSub(LHSI->getOperand(0), RHS, "fold");
3592 New = BinaryOperator::CreateAdd(LHSI->getOperand(0), RHS, "fold");
3593 return InsertNewInstBefore(New, I);
3596 /// FoldAndOfICmps - Fold (icmp)&(icmp) if possible.
3597 Instruction *InstCombiner::FoldAndOfICmps(Instruction &I,
3598 ICmpInst *LHS, ICmpInst *RHS) {
3600 ConstantInt *LHSCst, *RHSCst;
3601 ICmpInst::Predicate LHSCC, RHSCC;
3603 // This only handles icmp of constants: (icmp1 A, C1) & (icmp2 B, C2).
3604 if (!match(LHS, m_ICmp(LHSCC, m_Value(Val), m_ConstantInt(LHSCst))) ||
3605 !match(RHS, m_ICmp(RHSCC, m_Value(Val2), m_ConstantInt(RHSCst))))
3608 // (icmp ult A, C) & (icmp ult B, C) --> (icmp ult (A|B), C)
3609 // where C is a power of 2
3610 if (LHSCst == RHSCst && LHSCC == RHSCC && LHSCC == ICmpInst::ICMP_ULT &&
3611 LHSCst->getValue().isPowerOf2()) {
3612 Instruction *NewOr = BinaryOperator::CreateOr(Val, Val2);
3613 InsertNewInstBefore(NewOr, I);
3614 return new ICmpInst(LHSCC, NewOr, LHSCst);
3617 // From here on, we only handle:
3618 // (icmp1 A, C1) & (icmp2 A, C2) --> something simpler.
3619 if (Val != Val2) return 0;
3621 // ICMP_[US][GL]E X, CST is folded to ICMP_[US][GL]T elsewhere.
3622 if (LHSCC == ICmpInst::ICMP_UGE || LHSCC == ICmpInst::ICMP_ULE ||
3623 RHSCC == ICmpInst::ICMP_UGE || RHSCC == ICmpInst::ICMP_ULE ||
3624 LHSCC == ICmpInst::ICMP_SGE || LHSCC == ICmpInst::ICMP_SLE ||
3625 RHSCC == ICmpInst::ICMP_SGE || RHSCC == ICmpInst::ICMP_SLE)
3628 // We can't fold (ugt x, C) & (sgt x, C2).
3629 if (!PredicatesFoldable(LHSCC, RHSCC))
3632 // Ensure that the larger constant is on the RHS.
3634 if (ICmpInst::isSignedPredicate(LHSCC) ||
3635 (ICmpInst::isEquality(LHSCC) &&
3636 ICmpInst::isSignedPredicate(RHSCC)))
3637 ShouldSwap = LHSCst->getValue().sgt(RHSCst->getValue());
3639 ShouldSwap = LHSCst->getValue().ugt(RHSCst->getValue());
3642 std::swap(LHS, RHS);
3643 std::swap(LHSCst, RHSCst);
3644 std::swap(LHSCC, RHSCC);
3647 // At this point, we know we have have two icmp instructions
3648 // comparing a value against two constants and and'ing the result
3649 // together. Because of the above check, we know that we only have
3650 // icmp eq, icmp ne, icmp [su]lt, and icmp [SU]gt here. We also know
3651 // (from the FoldICmpLogical check above), that the two constants
3652 // are not equal and that the larger constant is on the RHS
3653 assert(LHSCst != RHSCst && "Compares not folded above?");
3656 default: assert(0 && "Unknown integer condition code!");
3657 case ICmpInst::ICMP_EQ:
3659 default: assert(0 && "Unknown integer condition code!");
3660 case ICmpInst::ICMP_EQ: // (X == 13 & X == 15) -> false
3661 case ICmpInst::ICMP_UGT: // (X == 13 & X > 15) -> false
3662 case ICmpInst::ICMP_SGT: // (X == 13 & X > 15) -> false
3663 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3664 case ICmpInst::ICMP_NE: // (X == 13 & X != 15) -> X == 13
3665 case ICmpInst::ICMP_ULT: // (X == 13 & X < 15) -> X == 13
3666 case ICmpInst::ICMP_SLT: // (X == 13 & X < 15) -> X == 13
3667 return ReplaceInstUsesWith(I, LHS);
3669 case ICmpInst::ICMP_NE:
3671 default: assert(0 && "Unknown integer condition code!");
3672 case ICmpInst::ICMP_ULT:
3673 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X u< 14) -> X < 13
3674 return new ICmpInst(ICmpInst::ICMP_ULT, Val, LHSCst);
3675 break; // (X != 13 & X u< 15) -> no change
3676 case ICmpInst::ICMP_SLT:
3677 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X s< 14) -> X < 13
3678 return new ICmpInst(ICmpInst::ICMP_SLT, Val, LHSCst);
3679 break; // (X != 13 & X s< 15) -> no change
3680 case ICmpInst::ICMP_EQ: // (X != 13 & X == 15) -> X == 15
3681 case ICmpInst::ICMP_UGT: // (X != 13 & X u> 15) -> X u> 15
3682 case ICmpInst::ICMP_SGT: // (X != 13 & X s> 15) -> X s> 15
3683 return ReplaceInstUsesWith(I, RHS);
3684 case ICmpInst::ICMP_NE:
3685 if (LHSCst == SubOne(RHSCst)){// (X != 13 & X != 14) -> X-13 >u 1
3686 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
3687 Instruction *Add = BinaryOperator::CreateAdd(Val, AddCST,
3688 Val->getName()+".off");
3689 InsertNewInstBefore(Add, I);
3690 return new ICmpInst(ICmpInst::ICMP_UGT, Add,
3691 ConstantInt::get(Add->getType(), 1));
3693 break; // (X != 13 & X != 15) -> no change
3696 case ICmpInst::ICMP_ULT:
3698 default: assert(0 && "Unknown integer condition code!");
3699 case ICmpInst::ICMP_EQ: // (X u< 13 & X == 15) -> false
3700 case ICmpInst::ICMP_UGT: // (X u< 13 & X u> 15) -> false
3701 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3702 case ICmpInst::ICMP_SGT: // (X u< 13 & X s> 15) -> no change
3704 case ICmpInst::ICMP_NE: // (X u< 13 & X != 15) -> X u< 13
3705 case ICmpInst::ICMP_ULT: // (X u< 13 & X u< 15) -> X u< 13
3706 return ReplaceInstUsesWith(I, LHS);
3707 case ICmpInst::ICMP_SLT: // (X u< 13 & X s< 15) -> no change
3711 case ICmpInst::ICMP_SLT:
3713 default: assert(0 && "Unknown integer condition code!");
3714 case ICmpInst::ICMP_EQ: // (X s< 13 & X == 15) -> false
3715 case ICmpInst::ICMP_SGT: // (X s< 13 & X s> 15) -> false
3716 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3717 case ICmpInst::ICMP_UGT: // (X s< 13 & X u> 15) -> no change
3719 case ICmpInst::ICMP_NE: // (X s< 13 & X != 15) -> X < 13
3720 case ICmpInst::ICMP_SLT: // (X s< 13 & X s< 15) -> X < 13
3721 return ReplaceInstUsesWith(I, LHS);
3722 case ICmpInst::ICMP_ULT: // (X s< 13 & X u< 15) -> no change
3726 case ICmpInst::ICMP_UGT:
3728 default: assert(0 && "Unknown integer condition code!");
3729 case ICmpInst::ICMP_EQ: // (X u> 13 & X == 15) -> X == 15
3730 case ICmpInst::ICMP_UGT: // (X u> 13 & X u> 15) -> X u> 15
3731 return ReplaceInstUsesWith(I, RHS);
3732 case ICmpInst::ICMP_SGT: // (X u> 13 & X s> 15) -> no change
3734 case ICmpInst::ICMP_NE:
3735 if (RHSCst == AddOne(LHSCst)) // (X u> 13 & X != 14) -> X u> 14
3736 return new ICmpInst(LHSCC, Val, RHSCst);
3737 break; // (X u> 13 & X != 15) -> no change
3738 case ICmpInst::ICMP_ULT: // (X u> 13 & X u< 15) -> (X-14) <u 1
3739 return InsertRangeTest(Val, AddOne(LHSCst), RHSCst, false, true, I);
3740 case ICmpInst::ICMP_SLT: // (X u> 13 & X s< 15) -> no change
3744 case ICmpInst::ICMP_SGT:
3746 default: assert(0 && "Unknown integer condition code!");
3747 case ICmpInst::ICMP_EQ: // (X s> 13 & X == 15) -> X == 15
3748 case ICmpInst::ICMP_SGT: // (X s> 13 & X s> 15) -> X s> 15
3749 return ReplaceInstUsesWith(I, RHS);
3750 case ICmpInst::ICMP_UGT: // (X s> 13 & X u> 15) -> no change
3752 case ICmpInst::ICMP_NE:
3753 if (RHSCst == AddOne(LHSCst)) // (X s> 13 & X != 14) -> X s> 14
3754 return new ICmpInst(LHSCC, Val, RHSCst);
3755 break; // (X s> 13 & X != 15) -> no change
3756 case ICmpInst::ICMP_SLT: // (X s> 13 & X s< 15) -> (X-14) s< 1
3757 return InsertRangeTest(Val, AddOne(LHSCst), RHSCst, true, true, I);
3758 case ICmpInst::ICMP_ULT: // (X s> 13 & X u< 15) -> no change
3768 Instruction *InstCombiner::visitAnd(BinaryOperator &I) {
3769 bool Changed = SimplifyCommutative(I);
3770 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3772 if (isa<UndefValue>(Op1)) // X & undef -> 0
3773 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3777 return ReplaceInstUsesWith(I, Op1);
3779 // See if we can simplify any instructions used by the instruction whose sole
3780 // purpose is to compute bits we don't care about.
3781 if (!isa<VectorType>(I.getType())) {
3782 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
3783 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
3784 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
3785 KnownZero, KnownOne))
3788 if (ConstantVector *CP = dyn_cast<ConstantVector>(Op1)) {
3789 if (CP->isAllOnesValue()) // X & <-1,-1> -> X
3790 return ReplaceInstUsesWith(I, I.getOperand(0));
3791 } else if (isa<ConstantAggregateZero>(Op1)) {
3792 return ReplaceInstUsesWith(I, Op1); // X & <0,0> -> <0,0>
3796 if (ConstantInt *AndRHS = dyn_cast<ConstantInt>(Op1)) {
3797 const APInt& AndRHSMask = AndRHS->getValue();
3798 APInt NotAndRHS(~AndRHSMask);
3800 // Optimize a variety of ((val OP C1) & C2) combinations...
3801 if (isa<BinaryOperator>(Op0)) {
3802 Instruction *Op0I = cast<Instruction>(Op0);
3803 Value *Op0LHS = Op0I->getOperand(0);
3804 Value *Op0RHS = Op0I->getOperand(1);
3805 switch (Op0I->getOpcode()) {
3806 case Instruction::Xor:
3807 case Instruction::Or:
3808 // If the mask is only needed on one incoming arm, push it up.
3809 if (Op0I->hasOneUse()) {
3810 if (MaskedValueIsZero(Op0LHS, NotAndRHS)) {
3811 // Not masking anything out for the LHS, move to RHS.
3812 Instruction *NewRHS = BinaryOperator::CreateAnd(Op0RHS, AndRHS,
3813 Op0RHS->getName()+".masked");
3814 InsertNewInstBefore(NewRHS, I);
3815 return BinaryOperator::Create(
3816 cast<BinaryOperator>(Op0I)->getOpcode(), Op0LHS, NewRHS);
3818 if (!isa<Constant>(Op0RHS) &&
3819 MaskedValueIsZero(Op0RHS, NotAndRHS)) {
3820 // Not masking anything out for the RHS, move to LHS.
3821 Instruction *NewLHS = BinaryOperator::CreateAnd(Op0LHS, AndRHS,
3822 Op0LHS->getName()+".masked");
3823 InsertNewInstBefore(NewLHS, I);
3824 return BinaryOperator::Create(
3825 cast<BinaryOperator>(Op0I)->getOpcode(), NewLHS, Op0RHS);
3830 case Instruction::Add:
3831 // ((A & N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == AndRHS.
3832 // ((A | N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
3833 // ((A ^ N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
3834 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, false, I))
3835 return BinaryOperator::CreateAnd(V, AndRHS);
3836 if (Value *V = FoldLogicalPlusAnd(Op0RHS, Op0LHS, AndRHS, false, I))
3837 return BinaryOperator::CreateAnd(V, AndRHS); // Add commutes
3840 case Instruction::Sub:
3841 // ((A & N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == AndRHS.
3842 // ((A | N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
3843 // ((A ^ N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
3844 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, true, I))
3845 return BinaryOperator::CreateAnd(V, AndRHS);
3847 // (A - N) & AndRHS -> -N & AndRHS iff A&AndRHS==0 and AndRHS
3848 // has 1's for all bits that the subtraction with A might affect.
3849 if (Op0I->hasOneUse()) {
3850 uint32_t BitWidth = AndRHSMask.getBitWidth();
3851 uint32_t Zeros = AndRHSMask.countLeadingZeros();
3852 APInt Mask = APInt::getLowBitsSet(BitWidth, BitWidth - Zeros);
3854 ConstantInt *A = dyn_cast<ConstantInt>(Op0LHS);
3855 if (!(A && A->isZero()) && // avoid infinite recursion.
3856 MaskedValueIsZero(Op0LHS, Mask)) {
3857 Instruction *NewNeg = BinaryOperator::CreateNeg(Op0RHS);
3858 InsertNewInstBefore(NewNeg, I);
3859 return BinaryOperator::CreateAnd(NewNeg, AndRHS);
3864 case Instruction::Shl:
3865 case Instruction::LShr:
3866 // (1 << x) & 1 --> zext(x == 0)
3867 // (1 >> x) & 1 --> zext(x == 0)
3868 if (AndRHSMask == 1 && Op0LHS == AndRHS) {
3869 Instruction *NewICmp = new ICmpInst(ICmpInst::ICMP_EQ, Op0RHS,
3870 Constant::getNullValue(I.getType()));
3871 InsertNewInstBefore(NewICmp, I);
3872 return new ZExtInst(NewICmp, I.getType());
3877 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
3878 if (Instruction *Res = OptAndOp(Op0I, Op0CI, AndRHS, I))
3880 } else if (CastInst *CI = dyn_cast<CastInst>(Op0)) {
3881 // If this is an integer truncation or change from signed-to-unsigned, and
3882 // if the source is an and/or with immediate, transform it. This
3883 // frequently occurs for bitfield accesses.
3884 if (Instruction *CastOp = dyn_cast<Instruction>(CI->getOperand(0))) {
3885 if ((isa<TruncInst>(CI) || isa<BitCastInst>(CI)) &&
3886 CastOp->getNumOperands() == 2)
3887 if (ConstantInt *AndCI = dyn_cast<ConstantInt>(CastOp->getOperand(1))) {
3888 if (CastOp->getOpcode() == Instruction::And) {
3889 // Change: and (cast (and X, C1) to T), C2
3890 // into : and (cast X to T), trunc_or_bitcast(C1)&C2
3891 // This will fold the two constants together, which may allow
3892 // other simplifications.
3893 Instruction *NewCast = CastInst::CreateTruncOrBitCast(
3894 CastOp->getOperand(0), I.getType(),
3895 CastOp->getName()+".shrunk");
3896 NewCast = InsertNewInstBefore(NewCast, I);
3897 // trunc_or_bitcast(C1)&C2
3898 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
3899 C3 = ConstantExpr::getAnd(C3, AndRHS);
3900 return BinaryOperator::CreateAnd(NewCast, C3);
3901 } else if (CastOp->getOpcode() == Instruction::Or) {
3902 // Change: and (cast (or X, C1) to T), C2
3903 // into : trunc(C1)&C2 iff trunc(C1)&C2 == C2
3904 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
3905 if (ConstantExpr::getAnd(C3, AndRHS) == AndRHS) // trunc(C1)&C2
3906 return ReplaceInstUsesWith(I, AndRHS);
3912 // Try to fold constant and into select arguments.
3913 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3914 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3916 if (isa<PHINode>(Op0))
3917 if (Instruction *NV = FoldOpIntoPhi(I))
3921 Value *Op0NotVal = dyn_castNotVal(Op0);
3922 Value *Op1NotVal = dyn_castNotVal(Op1);
3924 if (Op0NotVal == Op1 || Op1NotVal == Op0) // A & ~A == ~A & A == 0
3925 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3927 // (~A & ~B) == (~(A | B)) - De Morgan's Law
3928 if (Op0NotVal && Op1NotVal && isOnlyUse(Op0) && isOnlyUse(Op1)) {
3929 Instruction *Or = BinaryOperator::CreateOr(Op0NotVal, Op1NotVal,
3930 I.getName()+".demorgan");
3931 InsertNewInstBefore(Or, I);
3932 return BinaryOperator::CreateNot(Or);
3936 Value *A = 0, *B = 0, *C = 0, *D = 0;
3937 if (match(Op0, m_Or(m_Value(A), m_Value(B)))) {
3938 if (A == Op1 || B == Op1) // (A | ?) & A --> A
3939 return ReplaceInstUsesWith(I, Op1);
3941 // (A|B) & ~(A&B) -> A^B
3942 if (match(Op1, m_Not(m_And(m_Value(C), m_Value(D))))) {
3943 if ((A == C && B == D) || (A == D && B == C))
3944 return BinaryOperator::CreateXor(A, B);
3948 if (match(Op1, m_Or(m_Value(A), m_Value(B)))) {
3949 if (A == Op0 || B == Op0) // A & (A | ?) --> A
3950 return ReplaceInstUsesWith(I, Op0);
3952 // ~(A&B) & (A|B) -> A^B
3953 if (match(Op0, m_Not(m_And(m_Value(C), m_Value(D))))) {
3954 if ((A == C && B == D) || (A == D && B == C))
3955 return BinaryOperator::CreateXor(A, B);
3959 if (Op0->hasOneUse() &&
3960 match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
3961 if (A == Op1) { // (A^B)&A -> A&(A^B)
3962 I.swapOperands(); // Simplify below
3963 std::swap(Op0, Op1);
3964 } else if (B == Op1) { // (A^B)&B -> B&(B^A)
3965 cast<BinaryOperator>(Op0)->swapOperands();
3966 I.swapOperands(); // Simplify below
3967 std::swap(Op0, Op1);
3970 if (Op1->hasOneUse() &&
3971 match(Op1, m_Xor(m_Value(A), m_Value(B)))) {
3972 if (B == Op0) { // B&(A^B) -> B&(B^A)
3973 cast<BinaryOperator>(Op1)->swapOperands();
3976 if (A == Op0) { // A&(A^B) -> A & ~B
3977 Instruction *NotB = BinaryOperator::CreateNot(B, "tmp");
3978 InsertNewInstBefore(NotB, I);
3979 return BinaryOperator::CreateAnd(A, NotB);
3984 if (ICmpInst *RHS = dyn_cast<ICmpInst>(Op1)) {
3985 // (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
3986 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
3989 if (ICmpInst *LHS = dyn_cast<ICmpInst>(Op0))
3990 if (Instruction *Res = FoldAndOfICmps(I, LHS, RHS))
3994 // fold (and (cast A), (cast B)) -> (cast (and A, B))
3995 if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
3996 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
3997 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind ?
3998 const Type *SrcTy = Op0C->getOperand(0)->getType();
3999 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
4000 // Only do this if the casts both really cause code to be generated.
4001 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4003 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4005 Instruction *NewOp = BinaryOperator::CreateAnd(Op0C->getOperand(0),
4006 Op1C->getOperand(0),
4008 InsertNewInstBefore(NewOp, I);
4009 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
4013 // (X >> Z) & (Y >> Z) -> (X&Y) >> Z for all shifts.
4014 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
4015 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
4016 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
4017 SI0->getOperand(1) == SI1->getOperand(1) &&
4018 (SI0->hasOneUse() || SI1->hasOneUse())) {
4019 Instruction *NewOp =
4020 InsertNewInstBefore(BinaryOperator::CreateAnd(SI0->getOperand(0),
4022 SI0->getName()), I);
4023 return BinaryOperator::Create(SI1->getOpcode(), NewOp,
4024 SI1->getOperand(1));
4028 // If and'ing two fcmp, try combine them into one.
4029 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0))) {
4030 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1))) {
4031 if (LHS->getPredicate() == FCmpInst::FCMP_ORD &&
4032 RHS->getPredicate() == FCmpInst::FCMP_ORD) {
4033 // (fcmp ord x, c) & (fcmp ord y, c) -> (fcmp ord x, y)
4034 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
4035 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
4036 // If either of the constants are nans, then the whole thing returns
4038 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
4039 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4040 return new FCmpInst(FCmpInst::FCMP_ORD, LHS->getOperand(0),
4041 RHS->getOperand(0));
4044 Value *Op0LHS, *Op0RHS, *Op1LHS, *Op1RHS;
4045 FCmpInst::Predicate Op0CC, Op1CC;
4046 if (match(Op0, m_FCmp(Op0CC, m_Value(Op0LHS), m_Value(Op0RHS))) &&
4047 match(Op1, m_FCmp(Op1CC, m_Value(Op1LHS), m_Value(Op1RHS)))) {
4048 if (Op0LHS == Op1RHS && Op0RHS == Op1LHS) {
4049 // Swap RHS operands to match LHS.
4050 Op1CC = FCmpInst::getSwappedPredicate(Op1CC);
4051 std::swap(Op1LHS, Op1RHS);
4053 if (Op0LHS == Op1LHS && Op0RHS == Op1RHS) {
4054 // Simplify (fcmp cc0 x, y) & (fcmp cc1 x, y).
4056 return new FCmpInst((FCmpInst::Predicate)Op0CC, Op0LHS, Op0RHS);
4057 else if (Op0CC == FCmpInst::FCMP_FALSE ||
4058 Op1CC == FCmpInst::FCMP_FALSE)
4059 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4060 else if (Op0CC == FCmpInst::FCMP_TRUE)
4061 return ReplaceInstUsesWith(I, Op1);
4062 else if (Op1CC == FCmpInst::FCMP_TRUE)
4063 return ReplaceInstUsesWith(I, Op0);
4066 unsigned Op0Pred = getFCmpCode(Op0CC, Op0Ordered);
4067 unsigned Op1Pred = getFCmpCode(Op1CC, Op1Ordered);
4069 std::swap(Op0, Op1);
4070 std::swap(Op0Pred, Op1Pred);
4071 std::swap(Op0Ordered, Op1Ordered);
4074 // uno && ueq -> uno && (uno || eq) -> ueq
4075 // ord && olt -> ord && (ord && lt) -> olt
4076 if (Op0Ordered == Op1Ordered)
4077 return ReplaceInstUsesWith(I, Op1);
4078 // uno && oeq -> uno && (ord && eq) -> false
4079 // uno && ord -> false
4081 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4082 // ord && ueq -> ord && (uno || eq) -> oeq
4083 return cast<Instruction>(getFCmpValue(true, Op1Pred,
4092 return Changed ? &I : 0;
4095 /// CollectBSwapParts - Analyze the specified subexpression and see if it is
4096 /// capable of providing pieces of a bswap. The subexpression provides pieces
4097 /// of a bswap if it is proven that each of the non-zero bytes in the output of
4098 /// the expression came from the corresponding "byte swapped" byte in some other
4099 /// value. For example, if the current subexpression is "(shl i32 %X, 24)" then
4100 /// we know that the expression deposits the low byte of %X into the high byte
4101 /// of the bswap result and that all other bytes are zero. This expression is
4102 /// accepted, the high byte of ByteValues is set to X to indicate a correct
4105 /// This function returns true if the match was unsuccessful and false if so.
4106 /// On entry to the function the "OverallLeftShift" is a signed integer value
4107 /// indicating the number of bytes that the subexpression is later shifted. For
4108 /// example, if the expression is later right shifted by 16 bits, the
4109 /// OverallLeftShift value would be -2 on entry. This is used to specify which
4110 /// byte of ByteValues is actually being set.
4112 /// Similarly, ByteMask is a bitmask where a bit is clear if its corresponding
4113 /// byte is masked to zero by a user. For example, in (X & 255), X will be
4114 /// processed with a bytemask of 1. Because bytemask is 32-bits, this limits
4115 /// this function to working on up to 32-byte (256 bit) values. ByteMask is
4116 /// always in the local (OverallLeftShift) coordinate space.
4118 static bool CollectBSwapParts(Value *V, int OverallLeftShift, uint32_t ByteMask,
4119 SmallVector<Value*, 8> &ByteValues) {
4120 if (Instruction *I = dyn_cast<Instruction>(V)) {
4121 // If this is an or instruction, it may be an inner node of the bswap.
4122 if (I->getOpcode() == Instruction::Or) {
4123 return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
4125 CollectBSwapParts(I->getOperand(1), OverallLeftShift, ByteMask,
4129 // If this is a logical shift by a constant multiple of 8, recurse with
4130 // OverallLeftShift and ByteMask adjusted.
4131 if (I->isLogicalShift() && isa<ConstantInt>(I->getOperand(1))) {
4133 cast<ConstantInt>(I->getOperand(1))->getLimitedValue(~0U);
4134 // Ensure the shift amount is defined and of a byte value.
4135 if ((ShAmt & 7) || (ShAmt > 8*ByteValues.size()))
4138 unsigned ByteShift = ShAmt >> 3;
4139 if (I->getOpcode() == Instruction::Shl) {
4140 // X << 2 -> collect(X, +2)
4141 OverallLeftShift += ByteShift;
4142 ByteMask >>= ByteShift;
4144 // X >>u 2 -> collect(X, -2)
4145 OverallLeftShift -= ByteShift;
4146 ByteMask <<= ByteShift;
4147 ByteMask &= (~0U >> (32-ByteValues.size()));
4150 if (OverallLeftShift >= (int)ByteValues.size()) return true;
4151 if (OverallLeftShift <= -(int)ByteValues.size()) return true;
4153 return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
4157 // If this is a logical 'and' with a mask that clears bytes, clear the
4158 // corresponding bytes in ByteMask.
4159 if (I->getOpcode() == Instruction::And &&
4160 isa<ConstantInt>(I->getOperand(1))) {
4161 // Scan every byte of the and mask, seeing if the byte is either 0 or 255.
4162 unsigned NumBytes = ByteValues.size();
4163 APInt Byte(I->getType()->getPrimitiveSizeInBits(), 255);
4164 const APInt &AndMask = cast<ConstantInt>(I->getOperand(1))->getValue();
4166 for (unsigned i = 0; i != NumBytes; ++i, Byte <<= 8) {
4167 // If this byte is masked out by a later operation, we don't care what
4169 if ((ByteMask & (1 << i)) == 0)
4172 // If the AndMask is all zeros for this byte, clear the bit.
4173 APInt MaskB = AndMask & Byte;
4175 ByteMask &= ~(1U << i);
4179 // If the AndMask is not all ones for this byte, it's not a bytezap.
4183 // Otherwise, this byte is kept.
4186 return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
4191 // Okay, we got to something that isn't a shift, 'or' or 'and'. This must be
4192 // the input value to the bswap. Some observations: 1) if more than one byte
4193 // is demanded from this input, then it could not be successfully assembled
4194 // into a byteswap. At least one of the two bytes would not be aligned with
4195 // their ultimate destination.
4196 if (!isPowerOf2_32(ByteMask)) return true;
4197 unsigned InputByteNo = CountTrailingZeros_32(ByteMask);
4199 // 2) The input and ultimate destinations must line up: if byte 3 of an i32
4200 // is demanded, it needs to go into byte 0 of the result. This means that the
4201 // byte needs to be shifted until it lands in the right byte bucket. The
4202 // shift amount depends on the position: if the byte is coming from the high
4203 // part of the value (e.g. byte 3) then it must be shifted right. If from the
4204 // low part, it must be shifted left.
4205 unsigned DestByteNo = InputByteNo + OverallLeftShift;
4206 if (InputByteNo < ByteValues.size()/2) {
4207 if (ByteValues.size()-1-DestByteNo != InputByteNo)
4210 if (ByteValues.size()-1-DestByteNo != InputByteNo)
4214 // If the destination byte value is already defined, the values are or'd
4215 // together, which isn't a bswap (unless it's an or of the same bits).
4216 if (ByteValues[DestByteNo] && ByteValues[DestByteNo] != V)
4218 ByteValues[DestByteNo] = V;
4222 /// MatchBSwap - Given an OR instruction, check to see if this is a bswap idiom.
4223 /// If so, insert the new bswap intrinsic and return it.
4224 Instruction *InstCombiner::MatchBSwap(BinaryOperator &I) {
4225 const IntegerType *ITy = dyn_cast<IntegerType>(I.getType());
4226 if (!ITy || ITy->getBitWidth() % 16 ||
4227 // ByteMask only allows up to 32-byte values.
4228 ITy->getBitWidth() > 32*8)
4229 return 0; // Can only bswap pairs of bytes. Can't do vectors.
4231 /// ByteValues - For each byte of the result, we keep track of which value
4232 /// defines each byte.
4233 SmallVector<Value*, 8> ByteValues;
4234 ByteValues.resize(ITy->getBitWidth()/8);
4236 // Try to find all the pieces corresponding to the bswap.
4237 uint32_t ByteMask = ~0U >> (32-ByteValues.size());
4238 if (CollectBSwapParts(&I, 0, ByteMask, ByteValues))
4241 // Check to see if all of the bytes come from the same value.
4242 Value *V = ByteValues[0];
4243 if (V == 0) return 0; // Didn't find a byte? Must be zero.
4245 // Check to make sure that all of the bytes come from the same value.
4246 for (unsigned i = 1, e = ByteValues.size(); i != e; ++i)
4247 if (ByteValues[i] != V)
4249 const Type *Tys[] = { ITy };
4250 Module *M = I.getParent()->getParent()->getParent();
4251 Function *F = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 1);
4252 return CallInst::Create(F, V);
4255 /// MatchSelectFromAndOr - We have an expression of the form (A&C)|(B&D). Check
4256 /// If A is (cond?-1:0) and either B or D is ~(cond?-1,0) or (cond?0,-1), then
4257 /// we can simplify this expression to "cond ? C : D or B".
4258 static Instruction *MatchSelectFromAndOr(Value *A, Value *B,
4259 Value *C, Value *D) {
4260 // If A is not a select of -1/0, this cannot match.
4262 if (!match(A, m_SelectCst(m_Value(Cond), -1, 0)))
4265 // ((cond?-1:0)&C) | (B&(cond?0:-1)) -> cond ? C : B.
4266 if (match(D, m_SelectCst(m_Specific(Cond), 0, -1)))
4267 return SelectInst::Create(Cond, C, B);
4268 if (match(D, m_Not(m_SelectCst(m_Specific(Cond), -1, 0))))
4269 return SelectInst::Create(Cond, C, B);
4270 // ((cond?-1:0)&C) | ((cond?0:-1)&D) -> cond ? C : D.
4271 if (match(B, m_SelectCst(m_Specific(Cond), 0, -1)))
4272 return SelectInst::Create(Cond, C, D);
4273 if (match(B, m_Not(m_SelectCst(m_Specific(Cond), -1, 0))))
4274 return SelectInst::Create(Cond, C, D);
4278 /// FoldOrOfICmps - Fold (icmp)|(icmp) if possible.
4279 Instruction *InstCombiner::FoldOrOfICmps(Instruction &I,
4280 ICmpInst *LHS, ICmpInst *RHS) {
4282 ConstantInt *LHSCst, *RHSCst;
4283 ICmpInst::Predicate LHSCC, RHSCC;
4285 // This only handles icmp of constants: (icmp1 A, C1) | (icmp2 B, C2).
4286 if (!match(LHS, m_ICmp(LHSCC, m_Value(Val), m_ConstantInt(LHSCst))) ||
4287 !match(RHS, m_ICmp(RHSCC, m_Value(Val2), m_ConstantInt(RHSCst))))
4290 // From here on, we only handle:
4291 // (icmp1 A, C1) | (icmp2 A, C2) --> something simpler.
4292 if (Val != Val2) return 0;
4294 // ICMP_[US][GL]E X, CST is folded to ICMP_[US][GL]T elsewhere.
4295 if (LHSCC == ICmpInst::ICMP_UGE || LHSCC == ICmpInst::ICMP_ULE ||
4296 RHSCC == ICmpInst::ICMP_UGE || RHSCC == ICmpInst::ICMP_ULE ||
4297 LHSCC == ICmpInst::ICMP_SGE || LHSCC == ICmpInst::ICMP_SLE ||
4298 RHSCC == ICmpInst::ICMP_SGE || RHSCC == ICmpInst::ICMP_SLE)
4301 // We can't fold (ugt x, C) | (sgt x, C2).
4302 if (!PredicatesFoldable(LHSCC, RHSCC))
4305 // Ensure that the larger constant is on the RHS.
4307 if (ICmpInst::isSignedPredicate(LHSCC) ||
4308 (ICmpInst::isEquality(LHSCC) &&
4309 ICmpInst::isSignedPredicate(RHSCC)))
4310 ShouldSwap = LHSCst->getValue().sgt(RHSCst->getValue());
4312 ShouldSwap = LHSCst->getValue().ugt(RHSCst->getValue());
4315 std::swap(LHS, RHS);
4316 std::swap(LHSCst, RHSCst);
4317 std::swap(LHSCC, RHSCC);
4320 // At this point, we know we have have two icmp instructions
4321 // comparing a value against two constants and or'ing the result
4322 // together. Because of the above check, we know that we only have
4323 // ICMP_EQ, ICMP_NE, ICMP_LT, and ICMP_GT here. We also know (from the
4324 // FoldICmpLogical check above), that the two constants are not
4326 assert(LHSCst != RHSCst && "Compares not folded above?");
4329 default: assert(0 && "Unknown integer condition code!");
4330 case ICmpInst::ICMP_EQ:
4332 default: assert(0 && "Unknown integer condition code!");
4333 case ICmpInst::ICMP_EQ:
4334 if (LHSCst == SubOne(RHSCst)) { // (X == 13 | X == 14) -> X-13 <u 2
4335 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
4336 Instruction *Add = BinaryOperator::CreateAdd(Val, AddCST,
4337 Val->getName()+".off");
4338 InsertNewInstBefore(Add, I);
4339 AddCST = Subtract(AddOne(RHSCst), LHSCst);
4340 return new ICmpInst(ICmpInst::ICMP_ULT, Add, AddCST);
4342 break; // (X == 13 | X == 15) -> no change
4343 case ICmpInst::ICMP_UGT: // (X == 13 | X u> 14) -> no change
4344 case ICmpInst::ICMP_SGT: // (X == 13 | X s> 14) -> no change
4346 case ICmpInst::ICMP_NE: // (X == 13 | X != 15) -> X != 15
4347 case ICmpInst::ICMP_ULT: // (X == 13 | X u< 15) -> X u< 15
4348 case ICmpInst::ICMP_SLT: // (X == 13 | X s< 15) -> X s< 15
4349 return ReplaceInstUsesWith(I, RHS);
4352 case ICmpInst::ICMP_NE:
4354 default: assert(0 && "Unknown integer condition code!");
4355 case ICmpInst::ICMP_EQ: // (X != 13 | X == 15) -> X != 13
4356 case ICmpInst::ICMP_UGT: // (X != 13 | X u> 15) -> X != 13
4357 case ICmpInst::ICMP_SGT: // (X != 13 | X s> 15) -> X != 13
4358 return ReplaceInstUsesWith(I, LHS);
4359 case ICmpInst::ICMP_NE: // (X != 13 | X != 15) -> true
4360 case ICmpInst::ICMP_ULT: // (X != 13 | X u< 15) -> true
4361 case ICmpInst::ICMP_SLT: // (X != 13 | X s< 15) -> true
4362 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4365 case ICmpInst::ICMP_ULT:
4367 default: assert(0 && "Unknown integer condition code!");
4368 case ICmpInst::ICMP_EQ: // (X u< 13 | X == 14) -> no change
4370 case ICmpInst::ICMP_UGT: // (X u< 13 | X u> 15) -> (X-13) u> 2
4371 // If RHSCst is [us]MAXINT, it is always false. Not handling
4372 // this can cause overflow.
4373 if (RHSCst->isMaxValue(false))
4374 return ReplaceInstUsesWith(I, LHS);
4375 return InsertRangeTest(Val, LHSCst, AddOne(RHSCst), false, false, I);
4376 case ICmpInst::ICMP_SGT: // (X u< 13 | X s> 15) -> no change
4378 case ICmpInst::ICMP_NE: // (X u< 13 | X != 15) -> X != 15
4379 case ICmpInst::ICMP_ULT: // (X u< 13 | X u< 15) -> X u< 15
4380 return ReplaceInstUsesWith(I, RHS);
4381 case ICmpInst::ICMP_SLT: // (X u< 13 | X s< 15) -> no change
4385 case ICmpInst::ICMP_SLT:
4387 default: assert(0 && "Unknown integer condition code!");
4388 case ICmpInst::ICMP_EQ: // (X s< 13 | X == 14) -> no change
4390 case ICmpInst::ICMP_SGT: // (X s< 13 | X s> 15) -> (X-13) s> 2
4391 // If RHSCst is [us]MAXINT, it is always false. Not handling
4392 // this can cause overflow.
4393 if (RHSCst->isMaxValue(true))
4394 return ReplaceInstUsesWith(I, LHS);
4395 return InsertRangeTest(Val, LHSCst, AddOne(RHSCst), true, false, I);
4396 case ICmpInst::ICMP_UGT: // (X s< 13 | X u> 15) -> no change
4398 case ICmpInst::ICMP_NE: // (X s< 13 | X != 15) -> X != 15
4399 case ICmpInst::ICMP_SLT: // (X s< 13 | X s< 15) -> X s< 15
4400 return ReplaceInstUsesWith(I, RHS);
4401 case ICmpInst::ICMP_ULT: // (X s< 13 | X u< 15) -> no change
4405 case ICmpInst::ICMP_UGT:
4407 default: assert(0 && "Unknown integer condition code!");
4408 case ICmpInst::ICMP_EQ: // (X u> 13 | X == 15) -> X u> 13
4409 case ICmpInst::ICMP_UGT: // (X u> 13 | X u> 15) -> X u> 13
4410 return ReplaceInstUsesWith(I, LHS);
4411 case ICmpInst::ICMP_SGT: // (X u> 13 | X s> 15) -> no change
4413 case ICmpInst::ICMP_NE: // (X u> 13 | X != 15) -> true
4414 case ICmpInst::ICMP_ULT: // (X u> 13 | X u< 15) -> true
4415 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4416 case ICmpInst::ICMP_SLT: // (X u> 13 | X s< 15) -> no change
4420 case ICmpInst::ICMP_SGT:
4422 default: assert(0 && "Unknown integer condition code!");
4423 case ICmpInst::ICMP_EQ: // (X s> 13 | X == 15) -> X > 13
4424 case ICmpInst::ICMP_SGT: // (X s> 13 | X s> 15) -> X > 13
4425 return ReplaceInstUsesWith(I, LHS);
4426 case ICmpInst::ICMP_UGT: // (X s> 13 | X u> 15) -> no change
4428 case ICmpInst::ICMP_NE: // (X s> 13 | X != 15) -> true
4429 case ICmpInst::ICMP_SLT: // (X s> 13 | X s< 15) -> true
4430 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4431 case ICmpInst::ICMP_ULT: // (X s> 13 | X u< 15) -> no change
4439 Instruction *InstCombiner::visitOr(BinaryOperator &I) {
4440 bool Changed = SimplifyCommutative(I);
4441 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4443 if (isa<UndefValue>(Op1)) // X | undef -> -1
4444 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4448 return ReplaceInstUsesWith(I, Op0);
4450 // See if we can simplify any instructions used by the instruction whose sole
4451 // purpose is to compute bits we don't care about.
4452 if (!isa<VectorType>(I.getType())) {
4453 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
4454 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
4455 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
4456 KnownZero, KnownOne))
4458 } else if (isa<ConstantAggregateZero>(Op1)) {
4459 return ReplaceInstUsesWith(I, Op0); // X | <0,0> -> X
4460 } else if (ConstantVector *CP = dyn_cast<ConstantVector>(Op1)) {
4461 if (CP->isAllOnesValue()) // X | <-1,-1> -> <-1,-1>
4462 return ReplaceInstUsesWith(I, I.getOperand(1));
4468 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
4469 ConstantInt *C1 = 0; Value *X = 0;
4470 // (X & C1) | C2 --> (X | C2) & (C1|C2)
4471 if (match(Op0, m_And(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
4472 Instruction *Or = BinaryOperator::CreateOr(X, RHS);
4473 InsertNewInstBefore(Or, I);
4475 return BinaryOperator::CreateAnd(Or,
4476 ConstantInt::get(RHS->getValue() | C1->getValue()));
4479 // (X ^ C1) | C2 --> (X | C2) ^ (C1&~C2)
4480 if (match(Op0, m_Xor(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
4481 Instruction *Or = BinaryOperator::CreateOr(X, RHS);
4482 InsertNewInstBefore(Or, I);
4484 return BinaryOperator::CreateXor(Or,
4485 ConstantInt::get(C1->getValue() & ~RHS->getValue()));
4488 // Try to fold constant and into select arguments.
4489 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
4490 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
4492 if (isa<PHINode>(Op0))
4493 if (Instruction *NV = FoldOpIntoPhi(I))
4497 Value *A = 0, *B = 0;
4498 ConstantInt *C1 = 0, *C2 = 0;
4500 if (match(Op0, m_And(m_Value(A), m_Value(B))))
4501 if (A == Op1 || B == Op1) // (A & ?) | A --> A
4502 return ReplaceInstUsesWith(I, Op1);
4503 if (match(Op1, m_And(m_Value(A), m_Value(B))))
4504 if (A == Op0 || B == Op0) // A | (A & ?) --> A
4505 return ReplaceInstUsesWith(I, Op0);
4507 // (A | B) | C and A | (B | C) -> bswap if possible.
4508 // (A >> B) | (C << D) and (A << B) | (B >> C) -> bswap if possible.
4509 if (match(Op0, m_Or(m_Value(), m_Value())) ||
4510 match(Op1, m_Or(m_Value(), m_Value())) ||
4511 (match(Op0, m_Shift(m_Value(), m_Value())) &&
4512 match(Op1, m_Shift(m_Value(), m_Value())))) {
4513 if (Instruction *BSwap = MatchBSwap(I))
4517 // (X^C)|Y -> (X|Y)^C iff Y&C == 0
4518 if (Op0->hasOneUse() && match(Op0, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
4519 MaskedValueIsZero(Op1, C1->getValue())) {
4520 Instruction *NOr = BinaryOperator::CreateOr(A, Op1);
4521 InsertNewInstBefore(NOr, I);
4523 return BinaryOperator::CreateXor(NOr, C1);
4526 // Y|(X^C) -> (X|Y)^C iff Y&C == 0
4527 if (Op1->hasOneUse() && match(Op1, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
4528 MaskedValueIsZero(Op0, C1->getValue())) {
4529 Instruction *NOr = BinaryOperator::CreateOr(A, Op0);
4530 InsertNewInstBefore(NOr, I);
4532 return BinaryOperator::CreateXor(NOr, C1);
4536 Value *C = 0, *D = 0;
4537 if (match(Op0, m_And(m_Value(A), m_Value(C))) &&
4538 match(Op1, m_And(m_Value(B), m_Value(D)))) {
4539 Value *V1 = 0, *V2 = 0, *V3 = 0;
4540 C1 = dyn_cast<ConstantInt>(C);
4541 C2 = dyn_cast<ConstantInt>(D);
4542 if (C1 && C2) { // (A & C1)|(B & C2)
4543 // If we have: ((V + N) & C1) | (V & C2)
4544 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
4545 // replace with V+N.
4546 if (C1->getValue() == ~C2->getValue()) {
4547 if ((C2->getValue() & (C2->getValue()+1)) == 0 && // C2 == 0+1+
4548 match(A, m_Add(m_Value(V1), m_Value(V2)))) {
4549 // Add commutes, try both ways.
4550 if (V1 == B && MaskedValueIsZero(V2, C2->getValue()))
4551 return ReplaceInstUsesWith(I, A);
4552 if (V2 == B && MaskedValueIsZero(V1, C2->getValue()))
4553 return ReplaceInstUsesWith(I, A);
4555 // Or commutes, try both ways.
4556 if ((C1->getValue() & (C1->getValue()+1)) == 0 &&
4557 match(B, m_Add(m_Value(V1), m_Value(V2)))) {
4558 // Add commutes, try both ways.
4559 if (V1 == A && MaskedValueIsZero(V2, C1->getValue()))
4560 return ReplaceInstUsesWith(I, B);
4561 if (V2 == A && MaskedValueIsZero(V1, C1->getValue()))
4562 return ReplaceInstUsesWith(I, B);
4565 V1 = 0; V2 = 0; V3 = 0;
4568 // Check to see if we have any common things being and'ed. If so, find the
4569 // terms for V1 & (V2|V3).
4570 if (isOnlyUse(Op0) || isOnlyUse(Op1)) {
4571 if (A == B) // (A & C)|(A & D) == A & (C|D)
4572 V1 = A, V2 = C, V3 = D;
4573 else if (A == D) // (A & C)|(B & A) == A & (B|C)
4574 V1 = A, V2 = B, V3 = C;
4575 else if (C == B) // (A & C)|(C & D) == C & (A|D)
4576 V1 = C, V2 = A, V3 = D;
4577 else if (C == D) // (A & C)|(B & C) == C & (A|B)
4578 V1 = C, V2 = A, V3 = B;
4582 InsertNewInstBefore(BinaryOperator::CreateOr(V2, V3, "tmp"), I);
4583 return BinaryOperator::CreateAnd(V1, Or);
4587 // (A & (C0?-1:0)) | (B & ~(C0?-1:0)) -> C0 ? A : B, and commuted variants
4588 if (Instruction *Match = MatchSelectFromAndOr(A, B, C, D))
4590 if (Instruction *Match = MatchSelectFromAndOr(B, A, D, C))
4592 if (Instruction *Match = MatchSelectFromAndOr(C, B, A, D))
4594 if (Instruction *Match = MatchSelectFromAndOr(D, A, B, C))
4598 // (X >> Z) | (Y >> Z) -> (X|Y) >> Z for all shifts.
4599 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
4600 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
4601 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
4602 SI0->getOperand(1) == SI1->getOperand(1) &&
4603 (SI0->hasOneUse() || SI1->hasOneUse())) {
4604 Instruction *NewOp =
4605 InsertNewInstBefore(BinaryOperator::CreateOr(SI0->getOperand(0),
4607 SI0->getName()), I);
4608 return BinaryOperator::Create(SI1->getOpcode(), NewOp,
4609 SI1->getOperand(1));
4613 if (match(Op0, m_Not(m_Value(A)))) { // ~A | Op1
4614 if (A == Op1) // ~A | A == -1
4615 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4619 // Note, A is still live here!
4620 if (match(Op1, m_Not(m_Value(B)))) { // Op0 | ~B
4622 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4624 // (~A | ~B) == (~(A & B)) - De Morgan's Law
4625 if (A && isOnlyUse(Op0) && isOnlyUse(Op1)) {
4626 Value *And = InsertNewInstBefore(BinaryOperator::CreateAnd(A, B,
4627 I.getName()+".demorgan"), I);
4628 return BinaryOperator::CreateNot(And);
4632 // (icmp1 A, B) | (icmp2 A, B) --> (icmp3 A, B)
4633 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1))) {
4634 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
4637 if (ICmpInst *LHS = dyn_cast<ICmpInst>(I.getOperand(0)))
4638 if (Instruction *Res = FoldOrOfICmps(I, LHS, RHS))
4642 // fold (or (cast A), (cast B)) -> (cast (or A, B))
4643 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
4644 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4645 if (Op0C->getOpcode() == Op1C->getOpcode()) {// same cast kind ?
4646 if (!isa<ICmpInst>(Op0C->getOperand(0)) ||
4647 !isa<ICmpInst>(Op1C->getOperand(0))) {
4648 const Type *SrcTy = Op0C->getOperand(0)->getType();
4649 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
4650 // Only do this if the casts both really cause code to be
4652 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4654 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4656 Instruction *NewOp = BinaryOperator::CreateOr(Op0C->getOperand(0),
4657 Op1C->getOperand(0),
4659 InsertNewInstBefore(NewOp, I);
4660 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
4667 // (fcmp uno x, c) | (fcmp uno y, c) -> (fcmp uno x, y)
4668 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0))) {
4669 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1))) {
4670 if (LHS->getPredicate() == FCmpInst::FCMP_UNO &&
4671 RHS->getPredicate() == FCmpInst::FCMP_UNO &&
4672 LHS->getOperand(0)->getType() == RHS->getOperand(0)->getType()) {
4673 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
4674 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
4675 // If either of the constants are nans, then the whole thing returns
4677 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
4678 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4680 // Otherwise, no need to compare the two constants, compare the
4682 return new FCmpInst(FCmpInst::FCMP_UNO, LHS->getOperand(0),
4683 RHS->getOperand(0));
4686 Value *Op0LHS, *Op0RHS, *Op1LHS, *Op1RHS;
4687 FCmpInst::Predicate Op0CC, Op1CC;
4688 if (match(Op0, m_FCmp(Op0CC, m_Value(Op0LHS), m_Value(Op0RHS))) &&
4689 match(Op1, m_FCmp(Op1CC, m_Value(Op1LHS), m_Value(Op1RHS)))) {
4690 if (Op0LHS == Op1RHS && Op0RHS == Op1LHS) {
4691 // Swap RHS operands to match LHS.
4692 Op1CC = FCmpInst::getSwappedPredicate(Op1CC);
4693 std::swap(Op1LHS, Op1RHS);
4695 if (Op0LHS == Op1LHS && Op0RHS == Op1RHS) {
4696 // Simplify (fcmp cc0 x, y) | (fcmp cc1 x, y).
4698 return new FCmpInst((FCmpInst::Predicate)Op0CC, Op0LHS, Op0RHS);
4699 else if (Op0CC == FCmpInst::FCMP_TRUE ||
4700 Op1CC == FCmpInst::FCMP_TRUE)
4701 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4702 else if (Op0CC == FCmpInst::FCMP_FALSE)
4703 return ReplaceInstUsesWith(I, Op1);
4704 else if (Op1CC == FCmpInst::FCMP_FALSE)
4705 return ReplaceInstUsesWith(I, Op0);
4708 unsigned Op0Pred = getFCmpCode(Op0CC, Op0Ordered);
4709 unsigned Op1Pred = getFCmpCode(Op1CC, Op1Ordered);
4710 if (Op0Ordered == Op1Ordered) {
4711 // If both are ordered or unordered, return a new fcmp with
4712 // or'ed predicates.
4713 Value *RV = getFCmpValue(Op0Ordered, Op0Pred|Op1Pred,
4715 if (Instruction *I = dyn_cast<Instruction>(RV))
4717 // Otherwise, it's a constant boolean value...
4718 return ReplaceInstUsesWith(I, RV);
4726 return Changed ? &I : 0;
4731 // XorSelf - Implements: X ^ X --> 0
4734 XorSelf(Value *rhs) : RHS(rhs) {}
4735 bool shouldApply(Value *LHS) const { return LHS == RHS; }
4736 Instruction *apply(BinaryOperator &Xor) const {
4743 Instruction *InstCombiner::visitXor(BinaryOperator &I) {
4744 bool Changed = SimplifyCommutative(I);
4745 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4747 if (isa<UndefValue>(Op1)) {
4748 if (isa<UndefValue>(Op0))
4749 // Handle undef ^ undef -> 0 special case. This is a common
4751 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
4752 return ReplaceInstUsesWith(I, Op1); // X ^ undef -> undef
4755 // xor X, X = 0, even if X is nested in a sequence of Xor's.
4756 if (Instruction *Result = AssociativeOpt(I, XorSelf(Op1))) {
4757 assert(Result == &I && "AssociativeOpt didn't work?"); Result=Result;
4758 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
4761 // See if we can simplify any instructions used by the instruction whose sole
4762 // purpose is to compute bits we don't care about.
4763 if (!isa<VectorType>(I.getType())) {
4764 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
4765 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
4766 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
4767 KnownZero, KnownOne))
4769 } else if (isa<ConstantAggregateZero>(Op1)) {
4770 return ReplaceInstUsesWith(I, Op0); // X ^ <0,0> -> X
4773 // Is this a ~ operation?
4774 if (Value *NotOp = dyn_castNotVal(&I)) {
4775 // ~(~X & Y) --> (X | ~Y) - De Morgan's Law
4776 // ~(~X | Y) === (X & ~Y) - De Morgan's Law
4777 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(NotOp)) {
4778 if (Op0I->getOpcode() == Instruction::And ||
4779 Op0I->getOpcode() == Instruction::Or) {
4780 if (dyn_castNotVal(Op0I->getOperand(1))) Op0I->swapOperands();
4781 if (Value *Op0NotVal = dyn_castNotVal(Op0I->getOperand(0))) {
4783 BinaryOperator::CreateNot(Op0I->getOperand(1),
4784 Op0I->getOperand(1)->getName()+".not");
4785 InsertNewInstBefore(NotY, I);
4786 if (Op0I->getOpcode() == Instruction::And)
4787 return BinaryOperator::CreateOr(Op0NotVal, NotY);
4789 return BinaryOperator::CreateAnd(Op0NotVal, NotY);
4796 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
4797 // xor (cmp A, B), true = not (cmp A, B) = !cmp A, B
4798 if (RHS == ConstantInt::getTrue() && Op0->hasOneUse()) {
4799 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Op0))
4800 return new ICmpInst(ICI->getInversePredicate(),
4801 ICI->getOperand(0), ICI->getOperand(1));
4803 if (FCmpInst *FCI = dyn_cast<FCmpInst>(Op0))
4804 return new FCmpInst(FCI->getInversePredicate(),
4805 FCI->getOperand(0), FCI->getOperand(1));
4808 // fold (xor(zext(cmp)), 1) and (xor(sext(cmp)), -1) to ext(!cmp).
4809 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
4810 if (CmpInst *CI = dyn_cast<CmpInst>(Op0C->getOperand(0))) {
4811 if (CI->hasOneUse() && Op0C->hasOneUse()) {
4812 Instruction::CastOps Opcode = Op0C->getOpcode();
4813 if (Opcode == Instruction::ZExt || Opcode == Instruction::SExt) {
4814 if (RHS == ConstantExpr::getCast(Opcode, ConstantInt::getTrue(),
4815 Op0C->getDestTy())) {
4816 Instruction *NewCI = InsertNewInstBefore(CmpInst::Create(
4817 CI->getOpcode(), CI->getInversePredicate(),
4818 CI->getOperand(0), CI->getOperand(1)), I);
4819 NewCI->takeName(CI);
4820 return CastInst::Create(Opcode, NewCI, Op0C->getType());
4827 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
4828 // ~(c-X) == X-c-1 == X+(-c-1)
4829 if (Op0I->getOpcode() == Instruction::Sub && RHS->isAllOnesValue())
4830 if (Constant *Op0I0C = dyn_cast<Constant>(Op0I->getOperand(0))) {
4831 Constant *NegOp0I0C = ConstantExpr::getNeg(Op0I0C);
4832 Constant *ConstantRHS = ConstantExpr::getSub(NegOp0I0C,
4833 ConstantInt::get(I.getType(), 1));
4834 return BinaryOperator::CreateAdd(Op0I->getOperand(1), ConstantRHS);
4837 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) {
4838 if (Op0I->getOpcode() == Instruction::Add) {
4839 // ~(X-c) --> (-c-1)-X
4840 if (RHS->isAllOnesValue()) {
4841 Constant *NegOp0CI = ConstantExpr::getNeg(Op0CI);
4842 return BinaryOperator::CreateSub(
4843 ConstantExpr::getSub(NegOp0CI,
4844 ConstantInt::get(I.getType(), 1)),
4845 Op0I->getOperand(0));
4846 } else if (RHS->getValue().isSignBit()) {
4847 // (X + C) ^ signbit -> (X + C + signbit)
4848 Constant *C = ConstantInt::get(RHS->getValue() + Op0CI->getValue());
4849 return BinaryOperator::CreateAdd(Op0I->getOperand(0), C);
4852 } else if (Op0I->getOpcode() == Instruction::Or) {
4853 // (X|C1)^C2 -> X^(C1|C2) iff X&~C1 == 0
4854 if (MaskedValueIsZero(Op0I->getOperand(0), Op0CI->getValue())) {
4855 Constant *NewRHS = ConstantExpr::getOr(Op0CI, RHS);
4856 // Anything in both C1 and C2 is known to be zero, remove it from
4858 Constant *CommonBits = And(Op0CI, RHS);
4859 NewRHS = ConstantExpr::getAnd(NewRHS,
4860 ConstantExpr::getNot(CommonBits));
4861 AddToWorkList(Op0I);
4862 I.setOperand(0, Op0I->getOperand(0));
4863 I.setOperand(1, NewRHS);
4870 // Try to fold constant and into select arguments.
4871 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
4872 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
4874 if (isa<PHINode>(Op0))
4875 if (Instruction *NV = FoldOpIntoPhi(I))
4879 if (Value *X = dyn_castNotVal(Op0)) // ~A ^ A == -1
4881 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4883 if (Value *X = dyn_castNotVal(Op1)) // A ^ ~A == -1
4885 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4888 BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1);
4891 if (match(Op1I, m_Or(m_Value(A), m_Value(B)))) {
4892 if (A == Op0) { // B^(B|A) == (A|B)^B
4893 Op1I->swapOperands();
4895 std::swap(Op0, Op1);
4896 } else if (B == Op0) { // B^(A|B) == (A|B)^B
4897 I.swapOperands(); // Simplified below.
4898 std::swap(Op0, Op1);
4900 } else if (match(Op1I, m_Xor(m_Specific(Op0), m_Value(B)))) {
4901 return ReplaceInstUsesWith(I, B); // A^(A^B) == B
4902 } else if (match(Op1I, m_Xor(m_Value(A), m_Specific(Op0)))) {
4903 return ReplaceInstUsesWith(I, A); // A^(B^A) == B
4904 } else if (match(Op1I, m_And(m_Value(A), m_Value(B))) && Op1I->hasOneUse()){
4905 if (A == Op0) { // A^(A&B) -> A^(B&A)
4906 Op1I->swapOperands();
4909 if (B == Op0) { // A^(B&A) -> (B&A)^A
4910 I.swapOperands(); // Simplified below.
4911 std::swap(Op0, Op1);
4916 BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0);
4919 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) && Op0I->hasOneUse()) {
4920 if (A == Op1) // (B|A)^B == (A|B)^B
4922 if (B == Op1) { // (A|B)^B == A & ~B
4924 InsertNewInstBefore(BinaryOperator::CreateNot(Op1, "tmp"), I);
4925 return BinaryOperator::CreateAnd(A, NotB);
4927 } else if (match(Op0I, m_Xor(m_Specific(Op1), m_Value(B)))) {
4928 return ReplaceInstUsesWith(I, B); // (A^B)^A == B
4929 } else if (match(Op0I, m_Xor(m_Value(A), m_Specific(Op1)))) {
4930 return ReplaceInstUsesWith(I, A); // (B^A)^A == B
4931 } else if (match(Op0I, m_And(m_Value(A), m_Value(B))) && Op0I->hasOneUse()){
4932 if (A == Op1) // (A&B)^A -> (B&A)^A
4934 if (B == Op1 && // (B&A)^A == ~B & A
4935 !isa<ConstantInt>(Op1)) { // Canonical form is (B&C)^C
4937 InsertNewInstBefore(BinaryOperator::CreateNot(A, "tmp"), I);
4938 return BinaryOperator::CreateAnd(N, Op1);
4943 // (X >> Z) ^ (Y >> Z) -> (X^Y) >> Z for all shifts.
4944 if (Op0I && Op1I && Op0I->isShift() &&
4945 Op0I->getOpcode() == Op1I->getOpcode() &&
4946 Op0I->getOperand(1) == Op1I->getOperand(1) &&
4947 (Op1I->hasOneUse() || Op1I->hasOneUse())) {
4948 Instruction *NewOp =
4949 InsertNewInstBefore(BinaryOperator::CreateXor(Op0I->getOperand(0),
4950 Op1I->getOperand(0),
4951 Op0I->getName()), I);
4952 return BinaryOperator::Create(Op1I->getOpcode(), NewOp,
4953 Op1I->getOperand(1));
4957 Value *A, *B, *C, *D;
4958 // (A & B)^(A | B) -> A ^ B
4959 if (match(Op0I, m_And(m_Value(A), m_Value(B))) &&
4960 match(Op1I, m_Or(m_Value(C), m_Value(D)))) {
4961 if ((A == C && B == D) || (A == D && B == C))
4962 return BinaryOperator::CreateXor(A, B);
4964 // (A | B)^(A & B) -> A ^ B
4965 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) &&
4966 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
4967 if ((A == C && B == D) || (A == D && B == C))
4968 return BinaryOperator::CreateXor(A, B);
4972 if ((Op0I->hasOneUse() || Op1I->hasOneUse()) &&
4973 match(Op0I, m_And(m_Value(A), m_Value(B))) &&
4974 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
4975 // (X & Y)^(X & Y) -> (Y^Z) & X
4976 Value *X = 0, *Y = 0, *Z = 0;
4978 X = A, Y = B, Z = D;
4980 X = A, Y = B, Z = C;
4982 X = B, Y = A, Z = D;
4984 X = B, Y = A, Z = C;
4987 Instruction *NewOp =
4988 InsertNewInstBefore(BinaryOperator::CreateXor(Y, Z, Op0->getName()), I);
4989 return BinaryOperator::CreateAnd(NewOp, X);
4994 // (icmp1 A, B) ^ (icmp2 A, B) --> (icmp3 A, B)
4995 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1)))
4996 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
4999 // fold (xor (cast A), (cast B)) -> (cast (xor A, B))
5000 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
5001 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
5002 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind?
5003 const Type *SrcTy = Op0C->getOperand(0)->getType();
5004 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
5005 // Only do this if the casts both really cause code to be generated.
5006 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
5008 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
5010 Instruction *NewOp = BinaryOperator::CreateXor(Op0C->getOperand(0),
5011 Op1C->getOperand(0),
5013 InsertNewInstBefore(NewOp, I);
5014 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
5019 return Changed ? &I : 0;
5022 /// AddWithOverflow - Compute Result = In1+In2, returning true if the result
5023 /// overflowed for this type.
5024 static bool AddWithOverflow(ConstantInt *&Result, ConstantInt *In1,
5025 ConstantInt *In2, bool IsSigned = false) {
5026 Result = cast<ConstantInt>(Add(In1, In2));
5029 if (In2->getValue().isNegative())
5030 return Result->getValue().sgt(In1->getValue());
5032 return Result->getValue().slt(In1->getValue());
5034 return Result->getValue().ult(In1->getValue());
5037 /// SubWithOverflow - Compute Result = In1-In2, returning true if the result
5038 /// overflowed for this type.
5039 static bool SubWithOverflow(ConstantInt *&Result, ConstantInt *In1,
5040 ConstantInt *In2, bool IsSigned = false) {
5041 Result = cast<ConstantInt>(Subtract(In1, In2));
5044 if (In2->getValue().isNegative())
5045 return Result->getValue().slt(In1->getValue());
5047 return Result->getValue().sgt(In1->getValue());
5049 return Result->getValue().ugt(In1->getValue());
5052 /// EmitGEPOffset - Given a getelementptr instruction/constantexpr, emit the
5053 /// code necessary to compute the offset from the base pointer (without adding
5054 /// in the base pointer). Return the result as a signed integer of intptr size.
5055 static Value *EmitGEPOffset(User *GEP, Instruction &I, InstCombiner &IC) {
5056 TargetData &TD = IC.getTargetData();
5057 gep_type_iterator GTI = gep_type_begin(GEP);
5058 const Type *IntPtrTy = TD.getIntPtrType();
5059 Value *Result = Constant::getNullValue(IntPtrTy);
5061 // Build a mask for high order bits.
5062 unsigned IntPtrWidth = TD.getPointerSizeInBits();
5063 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth);
5065 for (User::op_iterator i = GEP->op_begin() + 1, e = GEP->op_end(); i != e;
5068 uint64_t Size = TD.getABITypeSize(GTI.getIndexedType()) & PtrSizeMask;
5069 if (ConstantInt *OpC = dyn_cast<ConstantInt>(Op)) {
5070 if (OpC->isZero()) continue;
5072 // Handle a struct index, which adds its field offset to the pointer.
5073 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
5074 Size = TD.getStructLayout(STy)->getElementOffset(OpC->getZExtValue());
5076 if (ConstantInt *RC = dyn_cast<ConstantInt>(Result))
5077 Result = ConstantInt::get(RC->getValue() + APInt(IntPtrWidth, Size));
5079 Result = IC.InsertNewInstBefore(
5080 BinaryOperator::CreateAdd(Result,
5081 ConstantInt::get(IntPtrTy, Size),
5082 GEP->getName()+".offs"), I);
5086 Constant *Scale = ConstantInt::get(IntPtrTy, Size);
5087 Constant *OC = ConstantExpr::getIntegerCast(OpC, IntPtrTy, true /*SExt*/);
5088 Scale = ConstantExpr::getMul(OC, Scale);
5089 if (Constant *RC = dyn_cast<Constant>(Result))
5090 Result = ConstantExpr::getAdd(RC, Scale);
5092 // Emit an add instruction.
5093 Result = IC.InsertNewInstBefore(
5094 BinaryOperator::CreateAdd(Result, Scale,
5095 GEP->getName()+".offs"), I);
5099 // Convert to correct type.
5100 if (Op->getType() != IntPtrTy) {
5101 if (Constant *OpC = dyn_cast<Constant>(Op))
5102 Op = ConstantExpr::getSExt(OpC, IntPtrTy);
5104 Op = IC.InsertNewInstBefore(new SExtInst(Op, IntPtrTy,
5105 Op->getName()+".c"), I);
5108 Constant *Scale = ConstantInt::get(IntPtrTy, Size);
5109 if (Constant *OpC = dyn_cast<Constant>(Op))
5110 Op = ConstantExpr::getMul(OpC, Scale);
5111 else // We'll let instcombine(mul) convert this to a shl if possible.
5112 Op = IC.InsertNewInstBefore(BinaryOperator::CreateMul(Op, Scale,
5113 GEP->getName()+".idx"), I);
5116 // Emit an add instruction.
5117 if (isa<Constant>(Op) && isa<Constant>(Result))
5118 Result = ConstantExpr::getAdd(cast<Constant>(Op),
5119 cast<Constant>(Result));
5121 Result = IC.InsertNewInstBefore(BinaryOperator::CreateAdd(Op, Result,
5122 GEP->getName()+".offs"), I);
5128 /// EvaluateGEPOffsetExpression - Return an value that can be used to compare of
5129 /// the *offset* implied by GEP to zero. For example, if we have &A[i], we want
5130 /// to return 'i' for "icmp ne i, 0". Note that, in general, indices can be
5131 /// complex, and scales are involved. The above expression would also be legal
5132 /// to codegen as "icmp ne (i*4), 0" (assuming A is a pointer to i32). This
5133 /// later form is less amenable to optimization though, and we are allowed to
5134 /// generate the first by knowing that pointer arithmetic doesn't overflow.
5136 /// If we can't emit an optimized form for this expression, this returns null.
5138 static Value *EvaluateGEPOffsetExpression(User *GEP, Instruction &I,
5140 TargetData &TD = IC.getTargetData();
5141 gep_type_iterator GTI = gep_type_begin(GEP);
5143 // Check to see if this gep only has a single variable index. If so, and if
5144 // any constant indices are a multiple of its scale, then we can compute this
5145 // in terms of the scale of the variable index. For example, if the GEP
5146 // implies an offset of "12 + i*4", then we can codegen this as "3 + i",
5147 // because the expression will cross zero at the same point.
5148 unsigned i, e = GEP->getNumOperands();
5150 for (i = 1; i != e; ++i, ++GTI) {
5151 if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i))) {
5152 // Compute the aggregate offset of constant indices.
5153 if (CI->isZero()) continue;
5155 // Handle a struct index, which adds its field offset to the pointer.
5156 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
5157 Offset += TD.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
5159 uint64_t Size = TD.getABITypeSize(GTI.getIndexedType());
5160 Offset += Size*CI->getSExtValue();
5163 // Found our variable index.
5168 // If there are no variable indices, we must have a constant offset, just
5169 // evaluate it the general way.
5170 if (i == e) return 0;
5172 Value *VariableIdx = GEP->getOperand(i);
5173 // Determine the scale factor of the variable element. For example, this is
5174 // 4 if the variable index is into an array of i32.
5175 uint64_t VariableScale = TD.getABITypeSize(GTI.getIndexedType());
5177 // Verify that there are no other variable indices. If so, emit the hard way.
5178 for (++i, ++GTI; i != e; ++i, ++GTI) {
5179 ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i));
5182 // Compute the aggregate offset of constant indices.
5183 if (CI->isZero()) continue;
5185 // Handle a struct index, which adds its field offset to the pointer.
5186 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
5187 Offset += TD.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
5189 uint64_t Size = TD.getABITypeSize(GTI.getIndexedType());
5190 Offset += Size*CI->getSExtValue();
5194 // Okay, we know we have a single variable index, which must be a
5195 // pointer/array/vector index. If there is no offset, life is simple, return
5197 unsigned IntPtrWidth = TD.getPointerSizeInBits();
5199 // Cast to intptrty in case a truncation occurs. If an extension is needed,
5200 // we don't need to bother extending: the extension won't affect where the
5201 // computation crosses zero.
5202 if (VariableIdx->getType()->getPrimitiveSizeInBits() > IntPtrWidth)
5203 VariableIdx = new TruncInst(VariableIdx, TD.getIntPtrType(),
5204 VariableIdx->getNameStart(), &I);
5208 // Otherwise, there is an index. The computation we will do will be modulo
5209 // the pointer size, so get it.
5210 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth);
5212 Offset &= PtrSizeMask;
5213 VariableScale &= PtrSizeMask;
5215 // To do this transformation, any constant index must be a multiple of the
5216 // variable scale factor. For example, we can evaluate "12 + 4*i" as "3 + i",
5217 // but we can't evaluate "10 + 3*i" in terms of i. Check that the offset is a
5218 // multiple of the variable scale.
5219 int64_t NewOffs = Offset / (int64_t)VariableScale;
5220 if (Offset != NewOffs*(int64_t)VariableScale)
5223 // Okay, we can do this evaluation. Start by converting the index to intptr.
5224 const Type *IntPtrTy = TD.getIntPtrType();
5225 if (VariableIdx->getType() != IntPtrTy)
5226 VariableIdx = CastInst::CreateIntegerCast(VariableIdx, IntPtrTy,
5228 VariableIdx->getNameStart(), &I);
5229 Constant *OffsetVal = ConstantInt::get(IntPtrTy, NewOffs);
5230 return BinaryOperator::CreateAdd(VariableIdx, OffsetVal, "offset", &I);
5234 /// FoldGEPICmp - Fold comparisons between a GEP instruction and something
5235 /// else. At this point we know that the GEP is on the LHS of the comparison.
5236 Instruction *InstCombiner::FoldGEPICmp(User *GEPLHS, Value *RHS,
5237 ICmpInst::Predicate Cond,
5239 assert(dyn_castGetElementPtr(GEPLHS) && "LHS is not a getelementptr!");
5241 // Look through bitcasts.
5242 if (BitCastInst *BCI = dyn_cast<BitCastInst>(RHS))
5243 RHS = BCI->getOperand(0);
5245 Value *PtrBase = GEPLHS->getOperand(0);
5246 if (PtrBase == RHS) {
5247 // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
5248 // This transformation (ignoring the base and scales) is valid because we
5249 // know pointers can't overflow. See if we can output an optimized form.
5250 Value *Offset = EvaluateGEPOffsetExpression(GEPLHS, I, *this);
5252 // If not, synthesize the offset the hard way.
5254 Offset = EmitGEPOffset(GEPLHS, I, *this);
5255 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Offset,
5256 Constant::getNullValue(Offset->getType()));
5257 } else if (User *GEPRHS = dyn_castGetElementPtr(RHS)) {
5258 // If the base pointers are different, but the indices are the same, just
5259 // compare the base pointer.
5260 if (PtrBase != GEPRHS->getOperand(0)) {
5261 bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands();
5262 IndicesTheSame &= GEPLHS->getOperand(0)->getType() ==
5263 GEPRHS->getOperand(0)->getType();
5265 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
5266 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
5267 IndicesTheSame = false;
5271 // If all indices are the same, just compare the base pointers.
5273 return new ICmpInst(ICmpInst::getSignedPredicate(Cond),
5274 GEPLHS->getOperand(0), GEPRHS->getOperand(0));
5276 // Otherwise, the base pointers are different and the indices are
5277 // different, bail out.
5281 // If one of the GEPs has all zero indices, recurse.
5282 bool AllZeros = true;
5283 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
5284 if (!isa<Constant>(GEPLHS->getOperand(i)) ||
5285 !cast<Constant>(GEPLHS->getOperand(i))->isNullValue()) {
5290 return FoldGEPICmp(GEPRHS, GEPLHS->getOperand(0),
5291 ICmpInst::getSwappedPredicate(Cond), I);
5293 // If the other GEP has all zero indices, recurse.
5295 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
5296 if (!isa<Constant>(GEPRHS->getOperand(i)) ||
5297 !cast<Constant>(GEPRHS->getOperand(i))->isNullValue()) {
5302 return FoldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I);
5304 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
5305 // If the GEPs only differ by one index, compare it.
5306 unsigned NumDifferences = 0; // Keep track of # differences.
5307 unsigned DiffOperand = 0; // The operand that differs.
5308 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
5309 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
5310 if (GEPLHS->getOperand(i)->getType()->getPrimitiveSizeInBits() !=
5311 GEPRHS->getOperand(i)->getType()->getPrimitiveSizeInBits()) {
5312 // Irreconcilable differences.
5316 if (NumDifferences++) break;
5321 if (NumDifferences == 0) // SAME GEP?
5322 return ReplaceInstUsesWith(I, // No comparison is needed here.
5323 ConstantInt::get(Type::Int1Ty,
5324 ICmpInst::isTrueWhenEqual(Cond)));
5326 else if (NumDifferences == 1) {
5327 Value *LHSV = GEPLHS->getOperand(DiffOperand);
5328 Value *RHSV = GEPRHS->getOperand(DiffOperand);
5329 // Make sure we do a signed comparison here.
5330 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), LHSV, RHSV);
5334 // Only lower this if the icmp is the only user of the GEP or if we expect
5335 // the result to fold to a constant!
5336 if ((isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
5337 (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
5338 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
5339 Value *L = EmitGEPOffset(GEPLHS, I, *this);
5340 Value *R = EmitGEPOffset(GEPRHS, I, *this);
5341 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), L, R);
5347 /// FoldFCmp_IntToFP_Cst - Fold fcmp ([us]itofp x, cst) if possible.
5349 Instruction *InstCombiner::FoldFCmp_IntToFP_Cst(FCmpInst &I,
5352 if (!isa<ConstantFP>(RHSC)) return 0;
5353 const APFloat &RHS = cast<ConstantFP>(RHSC)->getValueAPF();
5355 // Get the width of the mantissa. We don't want to hack on conversions that
5356 // might lose information from the integer, e.g. "i64 -> float"
5357 int MantissaWidth = LHSI->getType()->getFPMantissaWidth();
5358 if (MantissaWidth == -1) return 0; // Unknown.
5360 // Check to see that the input is converted from an integer type that is small
5361 // enough that preserves all bits. TODO: check here for "known" sign bits.
5362 // This would allow us to handle (fptosi (x >>s 62) to float) if x is i64 f.e.
5363 unsigned InputSize = LHSI->getOperand(0)->getType()->getPrimitiveSizeInBits();
5365 // If this is a uitofp instruction, we need an extra bit to hold the sign.
5366 bool LHSUnsigned = isa<UIToFPInst>(LHSI);
5370 // If the conversion would lose info, don't hack on this.
5371 if ((int)InputSize > MantissaWidth)
5374 // Otherwise, we can potentially simplify the comparison. We know that it
5375 // will always come through as an integer value and we know the constant is
5376 // not a NAN (it would have been previously simplified).
5377 assert(!RHS.isNaN() && "NaN comparison not already folded!");
5379 ICmpInst::Predicate Pred;
5380 switch (I.getPredicate()) {
5381 default: assert(0 && "Unexpected predicate!");
5382 case FCmpInst::FCMP_UEQ:
5383 case FCmpInst::FCMP_OEQ:
5384 Pred = ICmpInst::ICMP_EQ;
5386 case FCmpInst::FCMP_UGT:
5387 case FCmpInst::FCMP_OGT:
5388 Pred = LHSUnsigned ? ICmpInst::ICMP_UGT : ICmpInst::ICMP_SGT;
5390 case FCmpInst::FCMP_UGE:
5391 case FCmpInst::FCMP_OGE:
5392 Pred = LHSUnsigned ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_SGE;
5394 case FCmpInst::FCMP_ULT:
5395 case FCmpInst::FCMP_OLT:
5396 Pred = LHSUnsigned ? ICmpInst::ICMP_ULT : ICmpInst::ICMP_SLT;
5398 case FCmpInst::FCMP_ULE:
5399 case FCmpInst::FCMP_OLE:
5400 Pred = LHSUnsigned ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_SLE;
5402 case FCmpInst::FCMP_UNE:
5403 case FCmpInst::FCMP_ONE:
5404 Pred = ICmpInst::ICMP_NE;
5406 case FCmpInst::FCMP_ORD:
5407 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
5408 case FCmpInst::FCMP_UNO:
5409 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
5412 const IntegerType *IntTy = cast<IntegerType>(LHSI->getOperand(0)->getType());
5414 // Now we know that the APFloat is a normal number, zero or inf.
5416 // See if the FP constant is too large for the integer. For example,
5417 // comparing an i8 to 300.0.
5418 unsigned IntWidth = IntTy->getPrimitiveSizeInBits();
5421 // If the RHS value is > SignedMax, fold the comparison. This handles +INF
5422 // and large values.
5423 APFloat SMax(RHS.getSemantics(), APFloat::fcZero, false);
5424 SMax.convertFromAPInt(APInt::getSignedMaxValue(IntWidth), true,
5425 APFloat::rmNearestTiesToEven);
5426 if (SMax.compare(RHS) == APFloat::cmpLessThan) { // smax < 13123.0
5427 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SLT ||
5428 Pred == ICmpInst::ICMP_SLE)
5429 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
5430 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
5433 // If the RHS value is > UnsignedMax, fold the comparison. This handles
5434 // +INF and large values.
5435 APFloat UMax(RHS.getSemantics(), APFloat::fcZero, false);
5436 UMax.convertFromAPInt(APInt::getMaxValue(IntWidth), false,
5437 APFloat::rmNearestTiesToEven);
5438 if (UMax.compare(RHS) == APFloat::cmpLessThan) { // umax < 13123.0
5439 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_ULT ||
5440 Pred == ICmpInst::ICMP_ULE)
5441 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
5442 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
5447 // See if the RHS value is < SignedMin.
5448 APFloat SMin(RHS.getSemantics(), APFloat::fcZero, false);
5449 SMin.convertFromAPInt(APInt::getSignedMinValue(IntWidth), true,
5450 APFloat::rmNearestTiesToEven);
5451 if (SMin.compare(RHS) == APFloat::cmpGreaterThan) { // smin > 12312.0
5452 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SGT ||
5453 Pred == ICmpInst::ICMP_SGE)
5454 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
5455 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
5459 // Okay, now we know that the FP constant fits in the range [SMIN, SMAX] or
5460 // [0, UMAX], but it may still be fractional. See if it is fractional by
5461 // casting the FP value to the integer value and back, checking for equality.
5462 // Don't do this for zero, because -0.0 is not fractional.
5463 Constant *RHSInt = ConstantExpr::getFPToSI(RHSC, IntTy);
5464 if (!RHS.isZero() &&
5465 ConstantExpr::getSIToFP(RHSInt, RHSC->getType()) != RHSC) {
5466 // If we had a comparison against a fractional value, we have to adjust the
5467 // compare predicate and sometimes the value. RHSC is rounded towards zero
5470 default: assert(0 && "Unexpected integer comparison!");
5471 case ICmpInst::ICMP_NE: // (float)int != 4.4 --> true
5472 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
5473 case ICmpInst::ICMP_EQ: // (float)int == 4.4 --> false
5474 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
5475 case ICmpInst::ICMP_ULE:
5476 // (float)int <= 4.4 --> int <= 4
5477 // (float)int <= -4.4 --> false
5478 if (RHS.isNegative())
5479 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
5481 case ICmpInst::ICMP_SLE:
5482 // (float)int <= 4.4 --> int <= 4
5483 // (float)int <= -4.4 --> int < -4
5484 if (RHS.isNegative())
5485 Pred = ICmpInst::ICMP_SLT;
5487 case ICmpInst::ICMP_ULT:
5488 // (float)int < -4.4 --> false
5489 // (float)int < 4.4 --> int <= 4
5490 if (RHS.isNegative())
5491 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
5492 Pred = ICmpInst::ICMP_ULE;
5494 case ICmpInst::ICMP_SLT:
5495 // (float)int < -4.4 --> int < -4
5496 // (float)int < 4.4 --> int <= 4
5497 if (!RHS.isNegative())
5498 Pred = ICmpInst::ICMP_SLE;
5500 case ICmpInst::ICMP_UGT:
5501 // (float)int > 4.4 --> int > 4
5502 // (float)int > -4.4 --> true
5503 if (RHS.isNegative())
5504 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
5506 case ICmpInst::ICMP_SGT:
5507 // (float)int > 4.4 --> int > 4
5508 // (float)int > -4.4 --> int >= -4
5509 if (RHS.isNegative())
5510 Pred = ICmpInst::ICMP_SGE;
5512 case ICmpInst::ICMP_UGE:
5513 // (float)int >= -4.4 --> true
5514 // (float)int >= 4.4 --> int > 4
5515 if (!RHS.isNegative())
5516 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
5517 Pred = ICmpInst::ICMP_UGT;
5519 case ICmpInst::ICMP_SGE:
5520 // (float)int >= -4.4 --> int >= -4
5521 // (float)int >= 4.4 --> int > 4
5522 if (!RHS.isNegative())
5523 Pred = ICmpInst::ICMP_SGT;
5528 // Lower this FP comparison into an appropriate integer version of the
5530 return new ICmpInst(Pred, LHSI->getOperand(0), RHSInt);
5533 Instruction *InstCombiner::visitFCmpInst(FCmpInst &I) {
5534 bool Changed = SimplifyCompare(I);
5535 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5537 // Fold trivial predicates.
5538 if (I.getPredicate() == FCmpInst::FCMP_FALSE)
5539 return ReplaceInstUsesWith(I, Constant::getNullValue(Type::Int1Ty));
5540 if (I.getPredicate() == FCmpInst::FCMP_TRUE)
5541 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
5543 // Simplify 'fcmp pred X, X'
5545 switch (I.getPredicate()) {
5546 default: assert(0 && "Unknown predicate!");
5547 case FCmpInst::FCMP_UEQ: // True if unordered or equal
5548 case FCmpInst::FCMP_UGE: // True if unordered, greater than, or equal
5549 case FCmpInst::FCMP_ULE: // True if unordered, less than, or equal
5550 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
5551 case FCmpInst::FCMP_OGT: // True if ordered and greater than
5552 case FCmpInst::FCMP_OLT: // True if ordered and less than
5553 case FCmpInst::FCMP_ONE: // True if ordered and operands are unequal
5554 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
5556 case FCmpInst::FCMP_UNO: // True if unordered: isnan(X) | isnan(Y)
5557 case FCmpInst::FCMP_ULT: // True if unordered or less than
5558 case FCmpInst::FCMP_UGT: // True if unordered or greater than
5559 case FCmpInst::FCMP_UNE: // True if unordered or not equal
5560 // Canonicalize these to be 'fcmp uno %X, 0.0'.
5561 I.setPredicate(FCmpInst::FCMP_UNO);
5562 I.setOperand(1, Constant::getNullValue(Op0->getType()));
5565 case FCmpInst::FCMP_ORD: // True if ordered (no nans)
5566 case FCmpInst::FCMP_OEQ: // True if ordered and equal
5567 case FCmpInst::FCMP_OGE: // True if ordered and greater than or equal
5568 case FCmpInst::FCMP_OLE: // True if ordered and less than or equal
5569 // Canonicalize these to be 'fcmp ord %X, 0.0'.
5570 I.setPredicate(FCmpInst::FCMP_ORD);
5571 I.setOperand(1, Constant::getNullValue(Op0->getType()));
5576 if (isa<UndefValue>(Op1)) // fcmp pred X, undef -> undef
5577 return ReplaceInstUsesWith(I, UndefValue::get(Type::Int1Ty));
5579 // Handle fcmp with constant RHS
5580 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
5581 // If the constant is a nan, see if we can fold the comparison based on it.
5582 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHSC)) {
5583 if (CFP->getValueAPF().isNaN()) {
5584 if (FCmpInst::isOrdered(I.getPredicate())) // True if ordered and...
5585 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
5586 assert(FCmpInst::isUnordered(I.getPredicate()) &&
5587 "Comparison must be either ordered or unordered!");
5588 // True if unordered.
5589 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
5593 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
5594 switch (LHSI->getOpcode()) {
5595 case Instruction::PHI:
5596 // Only fold fcmp into the PHI if the phi and fcmp are in the same
5597 // block. If in the same block, we're encouraging jump threading. If
5598 // not, we are just pessimizing the code by making an i1 phi.
5599 if (LHSI->getParent() == I.getParent())
5600 if (Instruction *NV = FoldOpIntoPhi(I))
5603 case Instruction::SIToFP:
5604 case Instruction::UIToFP:
5605 if (Instruction *NV = FoldFCmp_IntToFP_Cst(I, LHSI, RHSC))
5608 case Instruction::Select:
5609 // If either operand of the select is a constant, we can fold the
5610 // comparison into the select arms, which will cause one to be
5611 // constant folded and the select turned into a bitwise or.
5612 Value *Op1 = 0, *Op2 = 0;
5613 if (LHSI->hasOneUse()) {
5614 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
5615 // Fold the known value into the constant operand.
5616 Op1 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
5617 // Insert a new FCmp of the other select operand.
5618 Op2 = InsertNewInstBefore(new FCmpInst(I.getPredicate(),
5619 LHSI->getOperand(2), RHSC,
5621 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
5622 // Fold the known value into the constant operand.
5623 Op2 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
5624 // Insert a new FCmp of the other select operand.
5625 Op1 = InsertNewInstBefore(new FCmpInst(I.getPredicate(),
5626 LHSI->getOperand(1), RHSC,
5632 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
5637 return Changed ? &I : 0;
5640 Instruction *InstCombiner::visitICmpInst(ICmpInst &I) {
5641 bool Changed = SimplifyCompare(I);
5642 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5643 const Type *Ty = Op0->getType();
5647 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
5648 I.isTrueWhenEqual()));
5650 if (isa<UndefValue>(Op1)) // X icmp undef -> undef
5651 return ReplaceInstUsesWith(I, UndefValue::get(Type::Int1Ty));
5653 // icmp <global/alloca*/null>, <global/alloca*/null> - Global/Stack value
5654 // addresses never equal each other! We already know that Op0 != Op1.
5655 if ((isa<GlobalValue>(Op0) || isa<AllocaInst>(Op0) ||
5656 isa<ConstantPointerNull>(Op0)) &&
5657 (isa<GlobalValue>(Op1) || isa<AllocaInst>(Op1) ||
5658 isa<ConstantPointerNull>(Op1)))
5659 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
5660 !I.isTrueWhenEqual()));
5662 // icmp's with boolean values can always be turned into bitwise operations
5663 if (Ty == Type::Int1Ty) {
5664 switch (I.getPredicate()) {
5665 default: assert(0 && "Invalid icmp instruction!");
5666 case ICmpInst::ICMP_EQ: { // icmp eq i1 A, B -> ~(A^B)
5667 Instruction *Xor = BinaryOperator::CreateXor(Op0, Op1, I.getName()+"tmp");
5668 InsertNewInstBefore(Xor, I);
5669 return BinaryOperator::CreateNot(Xor);
5671 case ICmpInst::ICMP_NE: // icmp eq i1 A, B -> A^B
5672 return BinaryOperator::CreateXor(Op0, Op1);
5674 case ICmpInst::ICMP_UGT:
5675 std::swap(Op0, Op1); // Change icmp ugt -> icmp ult
5677 case ICmpInst::ICMP_ULT:{ // icmp ult i1 A, B -> ~A & B
5678 Instruction *Not = BinaryOperator::CreateNot(Op0, I.getName()+"tmp");
5679 InsertNewInstBefore(Not, I);
5680 return BinaryOperator::CreateAnd(Not, Op1);
5682 case ICmpInst::ICMP_SGT:
5683 std::swap(Op0, Op1); // Change icmp sgt -> icmp slt
5685 case ICmpInst::ICMP_SLT: { // icmp slt i1 A, B -> A & ~B
5686 Instruction *Not = BinaryOperator::CreateNot(Op1, I.getName()+"tmp");
5687 InsertNewInstBefore(Not, I);
5688 return BinaryOperator::CreateAnd(Not, Op0);
5690 case ICmpInst::ICMP_UGE:
5691 std::swap(Op0, Op1); // Change icmp uge -> icmp ule
5693 case ICmpInst::ICMP_ULE: { // icmp ule i1 A, B -> ~A | B
5694 Instruction *Not = BinaryOperator::CreateNot(Op0, I.getName()+"tmp");
5695 InsertNewInstBefore(Not, I);
5696 return BinaryOperator::CreateOr(Not, Op1);
5698 case ICmpInst::ICMP_SGE:
5699 std::swap(Op0, Op1); // Change icmp sge -> icmp sle
5701 case ICmpInst::ICMP_SLE: { // icmp sle i1 A, B -> A | ~B
5702 Instruction *Not = BinaryOperator::CreateNot(Op1, I.getName()+"tmp");
5703 InsertNewInstBefore(Not, I);
5704 return BinaryOperator::CreateOr(Not, Op0);
5709 // See if we are doing a comparison with a constant.
5710 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
5713 // (icmp ne/eq (sub A B) 0) -> (icmp ne/eq A, B)
5714 if (I.isEquality() && CI->isNullValue() &&
5715 match(Op0, m_Sub(m_Value(A), m_Value(B)))) {
5716 // (icmp cond A B) if cond is equality
5717 return new ICmpInst(I.getPredicate(), A, B);
5720 // If we have an icmp le or icmp ge instruction, turn it into the
5721 // appropriate icmp lt or icmp gt instruction. This allows us to rely on
5722 // them being folded in the code below.
5723 switch (I.getPredicate()) {
5725 case ICmpInst::ICMP_ULE:
5726 if (CI->isMaxValue(false)) // A <=u MAX -> TRUE
5727 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5728 return new ICmpInst(ICmpInst::ICMP_ULT, Op0, AddOne(CI));
5729 case ICmpInst::ICMP_SLE:
5730 if (CI->isMaxValue(true)) // A <=s MAX -> TRUE
5731 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5732 return new ICmpInst(ICmpInst::ICMP_SLT, Op0, AddOne(CI));
5733 case ICmpInst::ICMP_UGE:
5734 if (CI->isMinValue(false)) // A >=u MIN -> TRUE
5735 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5736 return new ICmpInst( ICmpInst::ICMP_UGT, Op0, SubOne(CI));
5737 case ICmpInst::ICMP_SGE:
5738 if (CI->isMinValue(true)) // A >=s MIN -> TRUE
5739 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5740 return new ICmpInst(ICmpInst::ICMP_SGT, Op0, SubOne(CI));
5743 // See if we can fold the comparison based on range information we can get
5744 // by checking whether bits are known to be zero or one in the input.
5745 uint32_t BitWidth = cast<IntegerType>(Ty)->getBitWidth();
5746 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
5748 // If this comparison is a normal comparison, it demands all
5749 // bits, if it is a sign bit comparison, it only demands the sign bit.
5751 bool isSignBit = isSignBitCheck(I.getPredicate(), CI, UnusedBit);
5753 if (SimplifyDemandedBits(Op0,
5754 isSignBit ? APInt::getSignBit(BitWidth)
5755 : APInt::getAllOnesValue(BitWidth),
5756 KnownZero, KnownOne, 0))
5759 // Given the known and unknown bits, compute a range that the LHS could be
5760 // in. Compute the Min, Max and RHS values based on the known bits. For the
5761 // EQ and NE we use unsigned values.
5762 APInt Min(BitWidth, 0), Max(BitWidth, 0);
5763 if (ICmpInst::isSignedPredicate(I.getPredicate()))
5764 ComputeSignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne, Min, Max);
5766 ComputeUnsignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne,Min,Max);
5768 // If Min and Max are known to be the same, then SimplifyDemandedBits
5769 // figured out that the LHS is a constant. Just constant fold this now so
5770 // that code below can assume that Min != Max.
5772 return ReplaceInstUsesWith(I, ConstantExpr::getICmp(I.getPredicate(),
5773 ConstantInt::get(Min),
5776 // Based on the range information we know about the LHS, see if we can
5777 // simplify this comparison. For example, (x&4) < 8 is always true.
5778 const APInt &RHSVal = CI->getValue();
5779 switch (I.getPredicate()) { // LE/GE have been folded already.
5780 default: assert(0 && "Unknown icmp opcode!");
5781 case ICmpInst::ICMP_EQ:
5782 if (Max.ult(RHSVal) || Min.ugt(RHSVal))
5783 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5785 case ICmpInst::ICMP_NE:
5786 if (Max.ult(RHSVal) || Min.ugt(RHSVal))
5787 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5789 case ICmpInst::ICMP_ULT:
5790 if (Max.ult(RHSVal)) // A <u C -> true iff max(A) < C
5791 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5792 if (Min.uge(RHSVal)) // A <u C -> false iff min(A) >= C
5793 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5794 if (RHSVal == Max) // A <u MAX -> A != MAX
5795 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5796 if (RHSVal == Min+1) // A <u MIN+1 -> A == MIN
5797 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, SubOne(CI));
5799 // (x <u 2147483648) -> (x >s -1) -> true if sign bit clear
5800 if (CI->isMinValue(true))
5801 return new ICmpInst(ICmpInst::ICMP_SGT, Op0,
5802 ConstantInt::getAllOnesValue(Op0->getType()));
5804 case ICmpInst::ICMP_UGT:
5805 if (Min.ugt(RHSVal)) // A >u C -> true iff min(A) > C
5806 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5807 if (Max.ule(RHSVal)) // A >u C -> false iff max(A) <= C
5808 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5810 if (RHSVal == Min) // A >u MIN -> A != MIN
5811 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5812 if (RHSVal == Max-1) // A >u MAX-1 -> A == MAX
5813 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, AddOne(CI));
5815 // (x >u 2147483647) -> (x <s 0) -> true if sign bit set
5816 if (CI->isMaxValue(true))
5817 return new ICmpInst(ICmpInst::ICMP_SLT, Op0,
5818 ConstantInt::getNullValue(Op0->getType()));
5820 case ICmpInst::ICMP_SLT:
5821 if (Max.slt(RHSVal)) // A <s C -> true iff max(A) < C
5822 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5823 if (Min.sge(RHSVal)) // A <s C -> false iff min(A) >= C
5824 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5825 if (RHSVal == Max) // A <s MAX -> A != MAX
5826 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5827 if (RHSVal == Min+1) // A <s MIN+1 -> A == MIN
5828 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, SubOne(CI));
5830 case ICmpInst::ICMP_SGT:
5831 if (Min.sgt(RHSVal)) // A >s C -> true iff min(A) > C
5832 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5833 if (Max.sle(RHSVal)) // A >s C -> false iff max(A) <= C
5834 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5836 if (RHSVal == Min) // A >s MIN -> A != MIN
5837 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5838 if (RHSVal == Max-1) // A >s MAX-1 -> A == MAX
5839 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, AddOne(CI));
5844 // Test if the ICmpInst instruction is used exclusively by a select as
5845 // part of a minimum or maximum operation. If so, refrain from doing
5846 // any other folding. This helps out other analyses which understand
5847 // non-obfuscated minimum and maximum idioms, such as ScalarEvolution
5848 // and CodeGen. And in this case, at least one of the comparison
5849 // operands has at least one user besides the compare (the select),
5850 // which would often largely negate the benefit of folding anyway.
5852 if (SelectInst *SI = dyn_cast<SelectInst>(*I.use_begin()))
5853 if ((SI->getOperand(1) == Op0 && SI->getOperand(2) == Op1) ||
5854 (SI->getOperand(2) == Op0 && SI->getOperand(1) == Op1))
5857 // See if we are doing a comparison between a constant and an instruction that
5858 // can be folded into the comparison.
5859 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
5860 // Since the RHS is a ConstantInt (CI), if the left hand side is an
5861 // instruction, see if that instruction also has constants so that the
5862 // instruction can be folded into the icmp
5863 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
5864 if (Instruction *Res = visitICmpInstWithInstAndIntCst(I, LHSI, CI))
5868 // Handle icmp with constant (but not simple integer constant) RHS
5869 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
5870 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
5871 switch (LHSI->getOpcode()) {
5872 case Instruction::GetElementPtr:
5873 if (RHSC->isNullValue()) {
5874 // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
5875 bool isAllZeros = true;
5876 for (unsigned i = 1, e = LHSI->getNumOperands(); i != e; ++i)
5877 if (!isa<Constant>(LHSI->getOperand(i)) ||
5878 !cast<Constant>(LHSI->getOperand(i))->isNullValue()) {
5883 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
5884 Constant::getNullValue(LHSI->getOperand(0)->getType()));
5888 case Instruction::PHI:
5889 // Only fold icmp into the PHI if the phi and fcmp are in the same
5890 // block. If in the same block, we're encouraging jump threading. If
5891 // not, we are just pessimizing the code by making an i1 phi.
5892 if (LHSI->getParent() == I.getParent())
5893 if (Instruction *NV = FoldOpIntoPhi(I))
5896 case Instruction::Select: {
5897 // If either operand of the select is a constant, we can fold the
5898 // comparison into the select arms, which will cause one to be
5899 // constant folded and the select turned into a bitwise or.
5900 Value *Op1 = 0, *Op2 = 0;
5901 if (LHSI->hasOneUse()) {
5902 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
5903 // Fold the known value into the constant operand.
5904 Op1 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
5905 // Insert a new ICmp of the other select operand.
5906 Op2 = InsertNewInstBefore(new ICmpInst(I.getPredicate(),
5907 LHSI->getOperand(2), RHSC,
5909 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
5910 // Fold the known value into the constant operand.
5911 Op2 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
5912 // Insert a new ICmp of the other select operand.
5913 Op1 = InsertNewInstBefore(new ICmpInst(I.getPredicate(),
5914 LHSI->getOperand(1), RHSC,
5920 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
5923 case Instruction::Malloc:
5924 // If we have (malloc != null), and if the malloc has a single use, we
5925 // can assume it is successful and remove the malloc.
5926 if (LHSI->hasOneUse() && isa<ConstantPointerNull>(RHSC)) {
5927 AddToWorkList(LHSI);
5928 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
5929 !I.isTrueWhenEqual()));
5935 // If we can optimize a 'icmp GEP, P' or 'icmp P, GEP', do so now.
5936 if (User *GEP = dyn_castGetElementPtr(Op0))
5937 if (Instruction *NI = FoldGEPICmp(GEP, Op1, I.getPredicate(), I))
5939 if (User *GEP = dyn_castGetElementPtr(Op1))
5940 if (Instruction *NI = FoldGEPICmp(GEP, Op0,
5941 ICmpInst::getSwappedPredicate(I.getPredicate()), I))
5944 // Test to see if the operands of the icmp are casted versions of other
5945 // values. If the ptr->ptr cast can be stripped off both arguments, we do so
5947 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op0)) {
5948 if (isa<PointerType>(Op0->getType()) &&
5949 (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
5950 // We keep moving the cast from the left operand over to the right
5951 // operand, where it can often be eliminated completely.
5952 Op0 = CI->getOperand(0);
5954 // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast
5955 // so eliminate it as well.
5956 if (BitCastInst *CI2 = dyn_cast<BitCastInst>(Op1))
5957 Op1 = CI2->getOperand(0);
5959 // If Op1 is a constant, we can fold the cast into the constant.
5960 if (Op0->getType() != Op1->getType()) {
5961 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
5962 Op1 = ConstantExpr::getBitCast(Op1C, Op0->getType());
5964 // Otherwise, cast the RHS right before the icmp
5965 Op1 = InsertBitCastBefore(Op1, Op0->getType(), I);
5968 return new ICmpInst(I.getPredicate(), Op0, Op1);
5972 if (isa<CastInst>(Op0)) {
5973 // Handle the special case of: icmp (cast bool to X), <cst>
5974 // This comes up when you have code like
5977 // For generality, we handle any zero-extension of any operand comparison
5978 // with a constant or another cast from the same type.
5979 if (isa<ConstantInt>(Op1) || isa<CastInst>(Op1))
5980 if (Instruction *R = visitICmpInstWithCastAndCast(I))
5984 // See if it's the same type of instruction on the left and right.
5985 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
5986 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
5987 if (Op0I->getOpcode() == Op1I->getOpcode() && Op0I->hasOneUse() &&
5988 Op1I->hasOneUse() && Op0I->getOperand(1) == Op1I->getOperand(1) &&
5990 switch (Op0I->getOpcode()) {
5992 case Instruction::Add:
5993 case Instruction::Sub:
5994 case Instruction::Xor:
5995 // a+x icmp eq/ne b+x --> a icmp b
5996 return new ICmpInst(I.getPredicate(), Op0I->getOperand(0),
5997 Op1I->getOperand(0));
5999 case Instruction::Mul:
6000 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) {
6001 // a * Cst icmp eq/ne b * Cst --> a & Mask icmp b & Mask
6002 // Mask = -1 >> count-trailing-zeros(Cst).
6003 if (!CI->isZero() && !CI->isOne()) {
6004 const APInt &AP = CI->getValue();
6005 ConstantInt *Mask = ConstantInt::get(
6006 APInt::getLowBitsSet(AP.getBitWidth(),
6008 AP.countTrailingZeros()));
6009 Instruction *And1 = BinaryOperator::CreateAnd(Op0I->getOperand(0),
6011 Instruction *And2 = BinaryOperator::CreateAnd(Op1I->getOperand(0),
6013 InsertNewInstBefore(And1, I);
6014 InsertNewInstBefore(And2, I);
6015 return new ICmpInst(I.getPredicate(), And1, And2);
6024 // ~x < ~y --> y < x
6026 if (match(Op0, m_Not(m_Value(A))) &&
6027 match(Op1, m_Not(m_Value(B))))
6028 return new ICmpInst(I.getPredicate(), B, A);
6031 if (I.isEquality()) {
6032 Value *A, *B, *C, *D;
6034 // -x == -y --> x == y
6035 if (match(Op0, m_Neg(m_Value(A))) &&
6036 match(Op1, m_Neg(m_Value(B))))
6037 return new ICmpInst(I.getPredicate(), A, B);
6039 if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
6040 if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0
6041 Value *OtherVal = A == Op1 ? B : A;
6042 return new ICmpInst(I.getPredicate(), OtherVal,
6043 Constant::getNullValue(A->getType()));
6046 if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) {
6047 // A^c1 == C^c2 --> A == C^(c1^c2)
6048 ConstantInt *C1, *C2;
6049 if (match(B, m_ConstantInt(C1)) &&
6050 match(D, m_ConstantInt(C2)) && Op1->hasOneUse()) {
6051 Constant *NC = ConstantInt::get(C1->getValue() ^ C2->getValue());
6052 Instruction *Xor = BinaryOperator::CreateXor(C, NC, "tmp");
6053 return new ICmpInst(I.getPredicate(), A,
6054 InsertNewInstBefore(Xor, I));
6057 // A^B == A^D -> B == D
6058 if (A == C) return new ICmpInst(I.getPredicate(), B, D);
6059 if (A == D) return new ICmpInst(I.getPredicate(), B, C);
6060 if (B == C) return new ICmpInst(I.getPredicate(), A, D);
6061 if (B == D) return new ICmpInst(I.getPredicate(), A, C);
6065 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
6066 (A == Op0 || B == Op0)) {
6067 // A == (A^B) -> B == 0
6068 Value *OtherVal = A == Op0 ? B : A;
6069 return new ICmpInst(I.getPredicate(), OtherVal,
6070 Constant::getNullValue(A->getType()));
6073 // (A-B) == A -> B == 0
6074 if (match(Op0, m_Sub(m_Specific(Op1), m_Value(B))))
6075 return new ICmpInst(I.getPredicate(), B,
6076 Constant::getNullValue(B->getType()));
6078 // A == (A-B) -> B == 0
6079 if (match(Op1, m_Sub(m_Specific(Op0), m_Value(B))))
6080 return new ICmpInst(I.getPredicate(), B,
6081 Constant::getNullValue(B->getType()));
6083 // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
6084 if (Op0->hasOneUse() && Op1->hasOneUse() &&
6085 match(Op0, m_And(m_Value(A), m_Value(B))) &&
6086 match(Op1, m_And(m_Value(C), m_Value(D)))) {
6087 Value *X = 0, *Y = 0, *Z = 0;
6090 X = B; Y = D; Z = A;
6091 } else if (A == D) {
6092 X = B; Y = C; Z = A;
6093 } else if (B == C) {
6094 X = A; Y = D; Z = B;
6095 } else if (B == D) {
6096 X = A; Y = C; Z = B;
6099 if (X) { // Build (X^Y) & Z
6100 Op1 = InsertNewInstBefore(BinaryOperator::CreateXor(X, Y, "tmp"), I);
6101 Op1 = InsertNewInstBefore(BinaryOperator::CreateAnd(Op1, Z, "tmp"), I);
6102 I.setOperand(0, Op1);
6103 I.setOperand(1, Constant::getNullValue(Op1->getType()));
6108 return Changed ? &I : 0;
6112 /// FoldICmpDivCst - Fold "icmp pred, ([su]div X, DivRHS), CmpRHS" where DivRHS
6113 /// and CmpRHS are both known to be integer constants.
6114 Instruction *InstCombiner::FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
6115 ConstantInt *DivRHS) {
6116 ConstantInt *CmpRHS = cast<ConstantInt>(ICI.getOperand(1));
6117 const APInt &CmpRHSV = CmpRHS->getValue();
6119 // FIXME: If the operand types don't match the type of the divide
6120 // then don't attempt this transform. The code below doesn't have the
6121 // logic to deal with a signed divide and an unsigned compare (and
6122 // vice versa). This is because (x /s C1) <s C2 produces different
6123 // results than (x /s C1) <u C2 or (x /u C1) <s C2 or even
6124 // (x /u C1) <u C2. Simply casting the operands and result won't
6125 // work. :( The if statement below tests that condition and bails
6127 bool DivIsSigned = DivI->getOpcode() == Instruction::SDiv;
6128 if (!ICI.isEquality() && DivIsSigned != ICI.isSignedPredicate())
6130 if (DivRHS->isZero())
6131 return 0; // The ProdOV computation fails on divide by zero.
6132 if (DivIsSigned && DivRHS->isAllOnesValue())
6133 return 0; // The overflow computation also screws up here
6134 if (DivRHS->isOne())
6135 return 0; // Not worth bothering, and eliminates some funny cases
6138 // Compute Prod = CI * DivRHS. We are essentially solving an equation
6139 // of form X/C1=C2. We solve for X by multiplying C1 (DivRHS) and
6140 // C2 (CI). By solving for X we can turn this into a range check
6141 // instead of computing a divide.
6142 ConstantInt *Prod = Multiply(CmpRHS, DivRHS);
6144 // Determine if the product overflows by seeing if the product is
6145 // not equal to the divide. Make sure we do the same kind of divide
6146 // as in the LHS instruction that we're folding.
6147 bool ProdOV = (DivIsSigned ? ConstantExpr::getSDiv(Prod, DivRHS) :
6148 ConstantExpr::getUDiv(Prod, DivRHS)) != CmpRHS;
6150 // Get the ICmp opcode
6151 ICmpInst::Predicate Pred = ICI.getPredicate();
6153 // Figure out the interval that is being checked. For example, a comparison
6154 // like "X /u 5 == 0" is really checking that X is in the interval [0, 5).
6155 // Compute this interval based on the constants involved and the signedness of
6156 // the compare/divide. This computes a half-open interval, keeping track of
6157 // whether either value in the interval overflows. After analysis each
6158 // overflow variable is set to 0 if it's corresponding bound variable is valid
6159 // -1 if overflowed off the bottom end, or +1 if overflowed off the top end.
6160 int LoOverflow = 0, HiOverflow = 0;
6161 ConstantInt *LoBound = 0, *HiBound = 0;
6163 if (!DivIsSigned) { // udiv
6164 // e.g. X/5 op 3 --> [15, 20)
6166 HiOverflow = LoOverflow = ProdOV;
6168 HiOverflow = AddWithOverflow(HiBound, LoBound, DivRHS, false);
6169 } else if (DivRHS->getValue().isStrictlyPositive()) { // Divisor is > 0.
6170 if (CmpRHSV == 0) { // (X / pos) op 0
6171 // Can't overflow. e.g. X/2 op 0 --> [-1, 2)
6172 LoBound = cast<ConstantInt>(ConstantExpr::getNeg(SubOne(DivRHS)));
6174 } else if (CmpRHSV.isStrictlyPositive()) { // (X / pos) op pos
6175 LoBound = Prod; // e.g. X/5 op 3 --> [15, 20)
6176 HiOverflow = LoOverflow = ProdOV;
6178 HiOverflow = AddWithOverflow(HiBound, Prod, DivRHS, true);
6179 } else { // (X / pos) op neg
6180 // e.g. X/5 op -3 --> [-15-4, -15+1) --> [-19, -14)
6181 HiBound = AddOne(Prod);
6182 LoOverflow = HiOverflow = ProdOV ? -1 : 0;
6184 ConstantInt* DivNeg = cast<ConstantInt>(ConstantExpr::getNeg(DivRHS));
6185 LoOverflow = AddWithOverflow(LoBound, HiBound, DivNeg,
6189 } else if (DivRHS->getValue().isNegative()) { // Divisor is < 0.
6190 if (CmpRHSV == 0) { // (X / neg) op 0
6191 // e.g. X/-5 op 0 --> [-4, 5)
6192 LoBound = AddOne(DivRHS);
6193 HiBound = cast<ConstantInt>(ConstantExpr::getNeg(DivRHS));
6194 if (HiBound == DivRHS) { // -INTMIN = INTMIN
6195 HiOverflow = 1; // [INTMIN+1, overflow)
6196 HiBound = 0; // e.g. X/INTMIN = 0 --> X > INTMIN
6198 } else if (CmpRHSV.isStrictlyPositive()) { // (X / neg) op pos
6199 // e.g. X/-5 op 3 --> [-19, -14)
6200 HiBound = AddOne(Prod);
6201 HiOverflow = LoOverflow = ProdOV ? -1 : 0;
6203 LoOverflow = AddWithOverflow(LoBound, HiBound, DivRHS, true) ? -1 : 0;
6204 } else { // (X / neg) op neg
6205 LoBound = Prod; // e.g. X/-5 op -3 --> [15, 20)
6206 LoOverflow = HiOverflow = ProdOV;
6208 HiOverflow = SubWithOverflow(HiBound, Prod, DivRHS, true);
6211 // Dividing by a negative swaps the condition. LT <-> GT
6212 Pred = ICmpInst::getSwappedPredicate(Pred);
6215 Value *X = DivI->getOperand(0);
6217 default: assert(0 && "Unhandled icmp opcode!");
6218 case ICmpInst::ICMP_EQ:
6219 if (LoOverflow && HiOverflow)
6220 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
6221 else if (HiOverflow)
6222 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
6223 ICmpInst::ICMP_UGE, X, LoBound);
6224 else if (LoOverflow)
6225 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
6226 ICmpInst::ICMP_ULT, X, HiBound);
6228 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned, true, ICI);
6229 case ICmpInst::ICMP_NE:
6230 if (LoOverflow && HiOverflow)
6231 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
6232 else if (HiOverflow)
6233 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
6234 ICmpInst::ICMP_ULT, X, LoBound);
6235 else if (LoOverflow)
6236 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
6237 ICmpInst::ICMP_UGE, X, HiBound);
6239 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned, false, ICI);
6240 case ICmpInst::ICMP_ULT:
6241 case ICmpInst::ICMP_SLT:
6242 if (LoOverflow == +1) // Low bound is greater than input range.
6243 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
6244 if (LoOverflow == -1) // Low bound is less than input range.
6245 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
6246 return new ICmpInst(Pred, X, LoBound);
6247 case ICmpInst::ICMP_UGT:
6248 case ICmpInst::ICMP_SGT:
6249 if (HiOverflow == +1) // High bound greater than input range.
6250 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
6251 else if (HiOverflow == -1) // High bound less than input range.
6252 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
6253 if (Pred == ICmpInst::ICMP_UGT)
6254 return new ICmpInst(ICmpInst::ICMP_UGE, X, HiBound);
6256 return new ICmpInst(ICmpInst::ICMP_SGE, X, HiBound);
6261 /// visitICmpInstWithInstAndIntCst - Handle "icmp (instr, intcst)".
6263 Instruction *InstCombiner::visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
6266 const APInt &RHSV = RHS->getValue();
6268 switch (LHSI->getOpcode()) {
6269 case Instruction::Xor: // (icmp pred (xor X, XorCST), CI)
6270 if (ConstantInt *XorCST = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
6271 // If this is a comparison that tests the signbit (X < 0) or (x > -1),
6273 if ((ICI.getPredicate() == ICmpInst::ICMP_SLT && RHSV == 0) ||
6274 (ICI.getPredicate() == ICmpInst::ICMP_SGT && RHSV.isAllOnesValue())) {
6275 Value *CompareVal = LHSI->getOperand(0);
6277 // If the sign bit of the XorCST is not set, there is no change to
6278 // the operation, just stop using the Xor.
6279 if (!XorCST->getValue().isNegative()) {
6280 ICI.setOperand(0, CompareVal);
6281 AddToWorkList(LHSI);
6285 // Was the old condition true if the operand is positive?
6286 bool isTrueIfPositive = ICI.getPredicate() == ICmpInst::ICMP_SGT;
6288 // If so, the new one isn't.
6289 isTrueIfPositive ^= true;
6291 if (isTrueIfPositive)
6292 return new ICmpInst(ICmpInst::ICMP_SGT, CompareVal, SubOne(RHS));
6294 return new ICmpInst(ICmpInst::ICMP_SLT, CompareVal, AddOne(RHS));
6298 case Instruction::And: // (icmp pred (and X, AndCST), RHS)
6299 if (LHSI->hasOneUse() && isa<ConstantInt>(LHSI->getOperand(1)) &&
6300 LHSI->getOperand(0)->hasOneUse()) {
6301 ConstantInt *AndCST = cast<ConstantInt>(LHSI->getOperand(1));
6303 // If the LHS is an AND of a truncating cast, we can widen the
6304 // and/compare to be the input width without changing the value
6305 // produced, eliminating a cast.
6306 if (TruncInst *Cast = dyn_cast<TruncInst>(LHSI->getOperand(0))) {
6307 // We can do this transformation if either the AND constant does not
6308 // have its sign bit set or if it is an equality comparison.
6309 // Extending a relational comparison when we're checking the sign
6310 // bit would not work.
6311 if (Cast->hasOneUse() &&
6312 (ICI.isEquality() ||
6313 (AndCST->getValue().isNonNegative() && RHSV.isNonNegative()))) {
6315 cast<IntegerType>(Cast->getOperand(0)->getType())->getBitWidth();
6316 APInt NewCST = AndCST->getValue();
6317 NewCST.zext(BitWidth);
6319 NewCI.zext(BitWidth);
6320 Instruction *NewAnd =
6321 BinaryOperator::CreateAnd(Cast->getOperand(0),
6322 ConstantInt::get(NewCST),LHSI->getName());
6323 InsertNewInstBefore(NewAnd, ICI);
6324 return new ICmpInst(ICI.getPredicate(), NewAnd,
6325 ConstantInt::get(NewCI));
6329 // If this is: (X >> C1) & C2 != C3 (where any shift and any compare
6330 // could exist), turn it into (X & (C2 << C1)) != (C3 << C1). This
6331 // happens a LOT in code produced by the C front-end, for bitfield
6333 BinaryOperator *Shift = dyn_cast<BinaryOperator>(LHSI->getOperand(0));
6334 if (Shift && !Shift->isShift())
6338 ShAmt = Shift ? dyn_cast<ConstantInt>(Shift->getOperand(1)) : 0;
6339 const Type *Ty = Shift ? Shift->getType() : 0; // Type of the shift.
6340 const Type *AndTy = AndCST->getType(); // Type of the and.
6342 // We can fold this as long as we can't shift unknown bits
6343 // into the mask. This can only happen with signed shift
6344 // rights, as they sign-extend.
6346 bool CanFold = Shift->isLogicalShift();
6348 // To test for the bad case of the signed shr, see if any
6349 // of the bits shifted in could be tested after the mask.
6350 uint32_t TyBits = Ty->getPrimitiveSizeInBits();
6351 int ShAmtVal = TyBits - ShAmt->getLimitedValue(TyBits);
6353 uint32_t BitWidth = AndTy->getPrimitiveSizeInBits();
6354 if ((APInt::getHighBitsSet(BitWidth, BitWidth-ShAmtVal) &
6355 AndCST->getValue()) == 0)
6361 if (Shift->getOpcode() == Instruction::Shl)
6362 NewCst = ConstantExpr::getLShr(RHS, ShAmt);
6364 NewCst = ConstantExpr::getShl(RHS, ShAmt);
6366 // Check to see if we are shifting out any of the bits being
6368 if (ConstantExpr::get(Shift->getOpcode(), NewCst, ShAmt) != RHS) {
6369 // If we shifted bits out, the fold is not going to work out.
6370 // As a special case, check to see if this means that the
6371 // result is always true or false now.
6372 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
6373 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
6374 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
6375 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
6377 ICI.setOperand(1, NewCst);
6378 Constant *NewAndCST;
6379 if (Shift->getOpcode() == Instruction::Shl)
6380 NewAndCST = ConstantExpr::getLShr(AndCST, ShAmt);
6382 NewAndCST = ConstantExpr::getShl(AndCST, ShAmt);
6383 LHSI->setOperand(1, NewAndCST);
6384 LHSI->setOperand(0, Shift->getOperand(0));
6385 AddToWorkList(Shift); // Shift is dead.
6386 AddUsesToWorkList(ICI);
6392 // Turn ((X >> Y) & C) == 0 into (X & (C << Y)) == 0. The later is
6393 // preferable because it allows the C<<Y expression to be hoisted out
6394 // of a loop if Y is invariant and X is not.
6395 if (Shift && Shift->hasOneUse() && RHSV == 0 &&
6396 ICI.isEquality() && !Shift->isArithmeticShift() &&
6397 isa<Instruction>(Shift->getOperand(0))) {
6400 if (Shift->getOpcode() == Instruction::LShr) {
6401 NS = BinaryOperator::CreateShl(AndCST,
6402 Shift->getOperand(1), "tmp");
6404 // Insert a logical shift.
6405 NS = BinaryOperator::CreateLShr(AndCST,
6406 Shift->getOperand(1), "tmp");
6408 InsertNewInstBefore(cast<Instruction>(NS), ICI);
6410 // Compute X & (C << Y).
6411 Instruction *NewAnd =
6412 BinaryOperator::CreateAnd(Shift->getOperand(0), NS, LHSI->getName());
6413 InsertNewInstBefore(NewAnd, ICI);
6415 ICI.setOperand(0, NewAnd);
6421 case Instruction::Shl: { // (icmp pred (shl X, ShAmt), CI)
6422 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
6425 uint32_t TypeBits = RHSV.getBitWidth();
6427 // Check that the shift amount is in range. If not, don't perform
6428 // undefined shifts. When the shift is visited it will be
6430 if (ShAmt->uge(TypeBits))
6433 if (ICI.isEquality()) {
6434 // If we are comparing against bits always shifted out, the
6435 // comparison cannot succeed.
6437 ConstantExpr::getShl(ConstantExpr::getLShr(RHS, ShAmt), ShAmt);
6438 if (Comp != RHS) {// Comparing against a bit that we know is zero.
6439 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
6440 Constant *Cst = ConstantInt::get(Type::Int1Ty, IsICMP_NE);
6441 return ReplaceInstUsesWith(ICI, Cst);
6444 if (LHSI->hasOneUse()) {
6445 // Otherwise strength reduce the shift into an and.
6446 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
6448 ConstantInt::get(APInt::getLowBitsSet(TypeBits, TypeBits-ShAmtVal));
6451 BinaryOperator::CreateAnd(LHSI->getOperand(0),
6452 Mask, LHSI->getName()+".mask");
6453 Value *And = InsertNewInstBefore(AndI, ICI);
6454 return new ICmpInst(ICI.getPredicate(), And,
6455 ConstantInt::get(RHSV.lshr(ShAmtVal)));
6459 // Otherwise, if this is a comparison of the sign bit, simplify to and/test.
6460 bool TrueIfSigned = false;
6461 if (LHSI->hasOneUse() &&
6462 isSignBitCheck(ICI.getPredicate(), RHS, TrueIfSigned)) {
6463 // (X << 31) <s 0 --> (X&1) != 0
6464 Constant *Mask = ConstantInt::get(APInt(TypeBits, 1) <<
6465 (TypeBits-ShAmt->getZExtValue()-1));
6467 BinaryOperator::CreateAnd(LHSI->getOperand(0),
6468 Mask, LHSI->getName()+".mask");
6469 Value *And = InsertNewInstBefore(AndI, ICI);
6471 return new ICmpInst(TrueIfSigned ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ,
6472 And, Constant::getNullValue(And->getType()));
6477 case Instruction::LShr: // (icmp pred (shr X, ShAmt), CI)
6478 case Instruction::AShr: {
6479 // Only handle equality comparisons of shift-by-constant.
6480 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
6481 if (!ShAmt || !ICI.isEquality()) break;
6483 // Check that the shift amount is in range. If not, don't perform
6484 // undefined shifts. When the shift is visited it will be
6486 uint32_t TypeBits = RHSV.getBitWidth();
6487 if (ShAmt->uge(TypeBits))
6490 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
6492 // If we are comparing against bits always shifted out, the
6493 // comparison cannot succeed.
6494 APInt Comp = RHSV << ShAmtVal;
6495 if (LHSI->getOpcode() == Instruction::LShr)
6496 Comp = Comp.lshr(ShAmtVal);
6498 Comp = Comp.ashr(ShAmtVal);
6500 if (Comp != RHSV) { // Comparing against a bit that we know is zero.
6501 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
6502 Constant *Cst = ConstantInt::get(Type::Int1Ty, IsICMP_NE);
6503 return ReplaceInstUsesWith(ICI, Cst);
6506 // Otherwise, check to see if the bits shifted out are known to be zero.
6507 // If so, we can compare against the unshifted value:
6508 // (X & 4) >> 1 == 2 --> (X & 4) == 4.
6509 if (LHSI->hasOneUse() &&
6510 MaskedValueIsZero(LHSI->getOperand(0),
6511 APInt::getLowBitsSet(Comp.getBitWidth(), ShAmtVal))) {
6512 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
6513 ConstantExpr::getShl(RHS, ShAmt));
6516 if (LHSI->hasOneUse()) {
6517 // Otherwise strength reduce the shift into an and.
6518 APInt Val(APInt::getHighBitsSet(TypeBits, TypeBits - ShAmtVal));
6519 Constant *Mask = ConstantInt::get(Val);
6522 BinaryOperator::CreateAnd(LHSI->getOperand(0),
6523 Mask, LHSI->getName()+".mask");
6524 Value *And = InsertNewInstBefore(AndI, ICI);
6525 return new ICmpInst(ICI.getPredicate(), And,
6526 ConstantExpr::getShl(RHS, ShAmt));
6531 case Instruction::SDiv:
6532 case Instruction::UDiv:
6533 // Fold: icmp pred ([us]div X, C1), C2 -> range test
6534 // Fold this div into the comparison, producing a range check.
6535 // Determine, based on the divide type, what the range is being
6536 // checked. If there is an overflow on the low or high side, remember
6537 // it, otherwise compute the range [low, hi) bounding the new value.
6538 // See: InsertRangeTest above for the kinds of replacements possible.
6539 if (ConstantInt *DivRHS = dyn_cast<ConstantInt>(LHSI->getOperand(1)))
6540 if (Instruction *R = FoldICmpDivCst(ICI, cast<BinaryOperator>(LHSI),
6545 case Instruction::Add:
6546 // Fold: icmp pred (add, X, C1), C2
6548 if (!ICI.isEquality()) {
6549 ConstantInt *LHSC = dyn_cast<ConstantInt>(LHSI->getOperand(1));
6551 const APInt &LHSV = LHSC->getValue();
6553 ConstantRange CR = ICI.makeConstantRange(ICI.getPredicate(), RHSV)
6556 if (ICI.isSignedPredicate()) {
6557 if (CR.getLower().isSignBit()) {
6558 return new ICmpInst(ICmpInst::ICMP_SLT, LHSI->getOperand(0),
6559 ConstantInt::get(CR.getUpper()));
6560 } else if (CR.getUpper().isSignBit()) {
6561 return new ICmpInst(ICmpInst::ICMP_SGE, LHSI->getOperand(0),
6562 ConstantInt::get(CR.getLower()));
6565 if (CR.getLower().isMinValue()) {
6566 return new ICmpInst(ICmpInst::ICMP_ULT, LHSI->getOperand(0),
6567 ConstantInt::get(CR.getUpper()));
6568 } else if (CR.getUpper().isMinValue()) {
6569 return new ICmpInst(ICmpInst::ICMP_UGE, LHSI->getOperand(0),
6570 ConstantInt::get(CR.getLower()));
6577 // Simplify icmp_eq and icmp_ne instructions with integer constant RHS.
6578 if (ICI.isEquality()) {
6579 bool isICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
6581 // If the first operand is (add|sub|and|or|xor|rem) with a constant, and
6582 // the second operand is a constant, simplify a bit.
6583 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(LHSI)) {
6584 switch (BO->getOpcode()) {
6585 case Instruction::SRem:
6586 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
6587 if (RHSV == 0 && isa<ConstantInt>(BO->getOperand(1)) &&BO->hasOneUse()){
6588 const APInt &V = cast<ConstantInt>(BO->getOperand(1))->getValue();
6589 if (V.sgt(APInt(V.getBitWidth(), 1)) && V.isPowerOf2()) {
6590 Instruction *NewRem =
6591 BinaryOperator::CreateURem(BO->getOperand(0), BO->getOperand(1),
6593 InsertNewInstBefore(NewRem, ICI);
6594 return new ICmpInst(ICI.getPredicate(), NewRem,
6595 Constant::getNullValue(BO->getType()));
6599 case Instruction::Add:
6600 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
6601 if (ConstantInt *BOp1C = dyn_cast<ConstantInt>(BO->getOperand(1))) {
6602 if (BO->hasOneUse())
6603 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
6604 Subtract(RHS, BOp1C));
6605 } else if (RHSV == 0) {
6606 // Replace ((add A, B) != 0) with (A != -B) if A or B is
6607 // efficiently invertible, or if the add has just this one use.
6608 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
6610 if (Value *NegVal = dyn_castNegVal(BOp1))
6611 return new ICmpInst(ICI.getPredicate(), BOp0, NegVal);
6612 else if (Value *NegVal = dyn_castNegVal(BOp0))
6613 return new ICmpInst(ICI.getPredicate(), NegVal, BOp1);
6614 else if (BO->hasOneUse()) {
6615 Instruction *Neg = BinaryOperator::CreateNeg(BOp1);
6616 InsertNewInstBefore(Neg, ICI);
6618 return new ICmpInst(ICI.getPredicate(), BOp0, Neg);
6622 case Instruction::Xor:
6623 // For the xor case, we can xor two constants together, eliminating
6624 // the explicit xor.
6625 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1)))
6626 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
6627 ConstantExpr::getXor(RHS, BOC));
6630 case Instruction::Sub:
6631 // Replace (([sub|xor] A, B) != 0) with (A != B)
6633 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
6637 case Instruction::Or:
6638 // If bits are being or'd in that are not present in the constant we
6639 // are comparing against, then the comparison could never succeed!
6640 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1))) {
6641 Constant *NotCI = ConstantExpr::getNot(RHS);
6642 if (!ConstantExpr::getAnd(BOC, NotCI)->isNullValue())
6643 return ReplaceInstUsesWith(ICI, ConstantInt::get(Type::Int1Ty,
6648 case Instruction::And:
6649 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
6650 // If bits are being compared against that are and'd out, then the
6651 // comparison can never succeed!
6652 if ((RHSV & ~BOC->getValue()) != 0)
6653 return ReplaceInstUsesWith(ICI, ConstantInt::get(Type::Int1Ty,
6656 // If we have ((X & C) == C), turn it into ((X & C) != 0).
6657 if (RHS == BOC && RHSV.isPowerOf2())
6658 return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ :
6659 ICmpInst::ICMP_NE, LHSI,
6660 Constant::getNullValue(RHS->getType()));
6662 // Replace (and X, (1 << size(X)-1) != 0) with x s< 0
6663 if (BOC->getValue().isSignBit()) {
6664 Value *X = BO->getOperand(0);
6665 Constant *Zero = Constant::getNullValue(X->getType());
6666 ICmpInst::Predicate pred = isICMP_NE ?
6667 ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
6668 return new ICmpInst(pred, X, Zero);
6671 // ((X & ~7) == 0) --> X < 8
6672 if (RHSV == 0 && isHighOnes(BOC)) {
6673 Value *X = BO->getOperand(0);
6674 Constant *NegX = ConstantExpr::getNeg(BOC);
6675 ICmpInst::Predicate pred = isICMP_NE ?
6676 ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
6677 return new ICmpInst(pred, X, NegX);
6682 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(LHSI)) {
6683 // Handle icmp {eq|ne} <intrinsic>, intcst.
6684 if (II->getIntrinsicID() == Intrinsic::bswap) {
6686 ICI.setOperand(0, II->getOperand(1));
6687 ICI.setOperand(1, ConstantInt::get(RHSV.byteSwap()));
6691 } else { // Not a ICMP_EQ/ICMP_NE
6692 // If the LHS is a cast from an integral value of the same size,
6693 // then since we know the RHS is a constant, try to simlify.
6694 if (CastInst *Cast = dyn_cast<CastInst>(LHSI)) {
6695 Value *CastOp = Cast->getOperand(0);
6696 const Type *SrcTy = CastOp->getType();
6697 uint32_t SrcTySize = SrcTy->getPrimitiveSizeInBits();
6698 if (SrcTy->isInteger() &&
6699 SrcTySize == Cast->getType()->getPrimitiveSizeInBits()) {
6700 // If this is an unsigned comparison, try to make the comparison use
6701 // smaller constant values.
6702 if (ICI.getPredicate() == ICmpInst::ICMP_ULT && RHSV.isSignBit()) {
6703 // X u< 128 => X s> -1
6704 return new ICmpInst(ICmpInst::ICMP_SGT, CastOp,
6705 ConstantInt::get(APInt::getAllOnesValue(SrcTySize)));
6706 } else if (ICI.getPredicate() == ICmpInst::ICMP_UGT &&
6707 RHSV == APInt::getSignedMaxValue(SrcTySize)) {
6708 // X u> 127 => X s< 0
6709 return new ICmpInst(ICmpInst::ICMP_SLT, CastOp,
6710 Constant::getNullValue(SrcTy));
6718 /// visitICmpInstWithCastAndCast - Handle icmp (cast x to y), (cast/cst).
6719 /// We only handle extending casts so far.
6721 Instruction *InstCombiner::visitICmpInstWithCastAndCast(ICmpInst &ICI) {
6722 const CastInst *LHSCI = cast<CastInst>(ICI.getOperand(0));
6723 Value *LHSCIOp = LHSCI->getOperand(0);
6724 const Type *SrcTy = LHSCIOp->getType();
6725 const Type *DestTy = LHSCI->getType();
6728 // Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the
6729 // integer type is the same size as the pointer type.
6730 if (LHSCI->getOpcode() == Instruction::PtrToInt &&
6731 getTargetData().getPointerSizeInBits() ==
6732 cast<IntegerType>(DestTy)->getBitWidth()) {
6734 if (Constant *RHSC = dyn_cast<Constant>(ICI.getOperand(1))) {
6735 RHSOp = ConstantExpr::getIntToPtr(RHSC, SrcTy);
6736 } else if (PtrToIntInst *RHSC = dyn_cast<PtrToIntInst>(ICI.getOperand(1))) {
6737 RHSOp = RHSC->getOperand(0);
6738 // If the pointer types don't match, insert a bitcast.
6739 if (LHSCIOp->getType() != RHSOp->getType())
6740 RHSOp = InsertBitCastBefore(RHSOp, LHSCIOp->getType(), ICI);
6744 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSOp);
6747 // The code below only handles extension cast instructions, so far.
6749 if (LHSCI->getOpcode() != Instruction::ZExt &&
6750 LHSCI->getOpcode() != Instruction::SExt)
6753 bool isSignedExt = LHSCI->getOpcode() == Instruction::SExt;
6754 bool isSignedCmp = ICI.isSignedPredicate();
6756 if (CastInst *CI = dyn_cast<CastInst>(ICI.getOperand(1))) {
6757 // Not an extension from the same type?
6758 RHSCIOp = CI->getOperand(0);
6759 if (RHSCIOp->getType() != LHSCIOp->getType())
6762 // If the signedness of the two casts doesn't agree (i.e. one is a sext
6763 // and the other is a zext), then we can't handle this.
6764 if (CI->getOpcode() != LHSCI->getOpcode())
6767 // Deal with equality cases early.
6768 if (ICI.isEquality())
6769 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
6771 // A signed comparison of sign extended values simplifies into a
6772 // signed comparison.
6773 if (isSignedCmp && isSignedExt)
6774 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
6776 // The other three cases all fold into an unsigned comparison.
6777 return new ICmpInst(ICI.getUnsignedPredicate(), LHSCIOp, RHSCIOp);
6780 // If we aren't dealing with a constant on the RHS, exit early
6781 ConstantInt *CI = dyn_cast<ConstantInt>(ICI.getOperand(1));
6785 // Compute the constant that would happen if we truncated to SrcTy then
6786 // reextended to DestTy.
6787 Constant *Res1 = ConstantExpr::getTrunc(CI, SrcTy);
6788 Constant *Res2 = ConstantExpr::getCast(LHSCI->getOpcode(), Res1, DestTy);
6790 // If the re-extended constant didn't change...
6792 // Make sure that sign of the Cmp and the sign of the Cast are the same.
6793 // For example, we might have:
6794 // %A = sext short %X to uint
6795 // %B = icmp ugt uint %A, 1330
6796 // It is incorrect to transform this into
6797 // %B = icmp ugt short %X, 1330
6798 // because %A may have negative value.
6800 // However, we allow this when the compare is EQ/NE, because they are
6802 if (isSignedExt == isSignedCmp || ICI.isEquality())
6803 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
6807 // The re-extended constant changed so the constant cannot be represented
6808 // in the shorter type. Consequently, we cannot emit a simple comparison.
6810 // First, handle some easy cases. We know the result cannot be equal at this
6811 // point so handle the ICI.isEquality() cases
6812 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
6813 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
6814 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
6815 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
6817 // Evaluate the comparison for LT (we invert for GT below). LE and GE cases
6818 // should have been folded away previously and not enter in here.
6821 // We're performing a signed comparison.
6822 if (cast<ConstantInt>(CI)->getValue().isNegative())
6823 Result = ConstantInt::getFalse(); // X < (small) --> false
6825 Result = ConstantInt::getTrue(); // X < (large) --> true
6827 // We're performing an unsigned comparison.
6829 // We're performing an unsigned comp with a sign extended value.
6830 // This is true if the input is >= 0. [aka >s -1]
6831 Constant *NegOne = ConstantInt::getAllOnesValue(SrcTy);
6832 Result = InsertNewInstBefore(new ICmpInst(ICmpInst::ICMP_SGT, LHSCIOp,
6833 NegOne, ICI.getName()), ICI);
6835 // Unsigned extend & unsigned compare -> always true.
6836 Result = ConstantInt::getTrue();
6840 // Finally, return the value computed.
6841 if (ICI.getPredicate() == ICmpInst::ICMP_ULT ||
6842 ICI.getPredicate() == ICmpInst::ICMP_SLT)
6843 return ReplaceInstUsesWith(ICI, Result);
6845 assert((ICI.getPredicate()==ICmpInst::ICMP_UGT ||
6846 ICI.getPredicate()==ICmpInst::ICMP_SGT) &&
6847 "ICmp should be folded!");
6848 if (Constant *CI = dyn_cast<Constant>(Result))
6849 return ReplaceInstUsesWith(ICI, ConstantExpr::getNot(CI));
6850 return BinaryOperator::CreateNot(Result);
6853 Instruction *InstCombiner::visitShl(BinaryOperator &I) {
6854 return commonShiftTransforms(I);
6857 Instruction *InstCombiner::visitLShr(BinaryOperator &I) {
6858 return commonShiftTransforms(I);
6861 Instruction *InstCombiner::visitAShr(BinaryOperator &I) {
6862 if (Instruction *R = commonShiftTransforms(I))
6865 Value *Op0 = I.getOperand(0);
6867 // ashr int -1, X = -1 (for any arithmetic shift rights of ~0)
6868 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
6869 if (CSI->isAllOnesValue())
6870 return ReplaceInstUsesWith(I, CSI);
6872 // See if we can turn a signed shr into an unsigned shr.
6873 if (!isa<VectorType>(I.getType()) &&
6874 MaskedValueIsZero(Op0,
6875 APInt::getSignBit(I.getType()->getPrimitiveSizeInBits())))
6876 return BinaryOperator::CreateLShr(Op0, I.getOperand(1));
6881 Instruction *InstCombiner::commonShiftTransforms(BinaryOperator &I) {
6882 assert(I.getOperand(1)->getType() == I.getOperand(0)->getType());
6883 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
6885 // shl X, 0 == X and shr X, 0 == X
6886 // shl 0, X == 0 and shr 0, X == 0
6887 if (Op1 == Constant::getNullValue(Op1->getType()) ||
6888 Op0 == Constant::getNullValue(Op0->getType()))
6889 return ReplaceInstUsesWith(I, Op0);
6891 if (isa<UndefValue>(Op0)) {
6892 if (I.getOpcode() == Instruction::AShr) // undef >>s X -> undef
6893 return ReplaceInstUsesWith(I, Op0);
6894 else // undef << X -> 0, undef >>u X -> 0
6895 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
6897 if (isa<UndefValue>(Op1)) {
6898 if (I.getOpcode() == Instruction::AShr) // X >>s undef -> X
6899 return ReplaceInstUsesWith(I, Op0);
6900 else // X << undef, X >>u undef -> 0
6901 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
6904 // Try to fold constant and into select arguments.
6905 if (isa<Constant>(Op0))
6906 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
6907 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
6910 if (ConstantInt *CUI = dyn_cast<ConstantInt>(Op1))
6911 if (Instruction *Res = FoldShiftByConstant(Op0, CUI, I))
6916 Instruction *InstCombiner::FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
6917 BinaryOperator &I) {
6918 bool isLeftShift = I.getOpcode() == Instruction::Shl;
6920 // See if we can simplify any instructions used by the instruction whose sole
6921 // purpose is to compute bits we don't care about.
6922 uint32_t TypeBits = Op0->getType()->getPrimitiveSizeInBits();
6923 APInt KnownZero(TypeBits, 0), KnownOne(TypeBits, 0);
6924 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(TypeBits),
6925 KnownZero, KnownOne))
6928 // shl uint X, 32 = 0 and shr ubyte Y, 9 = 0, ... just don't eliminate shr
6929 // of a signed value.
6931 if (Op1->uge(TypeBits)) {
6932 if (I.getOpcode() != Instruction::AShr)
6933 return ReplaceInstUsesWith(I, Constant::getNullValue(Op0->getType()));
6935 I.setOperand(1, ConstantInt::get(I.getType(), TypeBits-1));
6940 // ((X*C1) << C2) == (X * (C1 << C2))
6941 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0))
6942 if (BO->getOpcode() == Instruction::Mul && isLeftShift)
6943 if (Constant *BOOp = dyn_cast<Constant>(BO->getOperand(1)))
6944 return BinaryOperator::CreateMul(BO->getOperand(0),
6945 ConstantExpr::getShl(BOOp, Op1));
6947 // Try to fold constant and into select arguments.
6948 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
6949 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
6951 if (isa<PHINode>(Op0))
6952 if (Instruction *NV = FoldOpIntoPhi(I))
6955 // Fold shift2(trunc(shift1(x,c1)), c2) -> trunc(shift2(shift1(x,c1),c2))
6956 if (TruncInst *TI = dyn_cast<TruncInst>(Op0)) {
6957 Instruction *TrOp = dyn_cast<Instruction>(TI->getOperand(0));
6958 // If 'shift2' is an ashr, we would have to get the sign bit into a funny
6959 // place. Don't try to do this transformation in this case. Also, we
6960 // require that the input operand is a shift-by-constant so that we have
6961 // confidence that the shifts will get folded together. We could do this
6962 // xform in more cases, but it is unlikely to be profitable.
6963 if (TrOp && I.isLogicalShift() && TrOp->isShift() &&
6964 isa<ConstantInt>(TrOp->getOperand(1))) {
6965 // Okay, we'll do this xform. Make the shift of shift.
6966 Constant *ShAmt = ConstantExpr::getZExt(Op1, TrOp->getType());
6967 Instruction *NSh = BinaryOperator::Create(I.getOpcode(), TrOp, ShAmt,
6969 InsertNewInstBefore(NSh, I); // (shift2 (shift1 & 0x00FF), c2)
6971 // For logical shifts, the truncation has the effect of making the high
6972 // part of the register be zeros. Emulate this by inserting an AND to
6973 // clear the top bits as needed. This 'and' will usually be zapped by
6974 // other xforms later if dead.
6975 unsigned SrcSize = TrOp->getType()->getPrimitiveSizeInBits();
6976 unsigned DstSize = TI->getType()->getPrimitiveSizeInBits();
6977 APInt MaskV(APInt::getLowBitsSet(SrcSize, DstSize));
6979 // The mask we constructed says what the trunc would do if occurring
6980 // between the shifts. We want to know the effect *after* the second
6981 // shift. We know that it is a logical shift by a constant, so adjust the
6982 // mask as appropriate.
6983 if (I.getOpcode() == Instruction::Shl)
6984 MaskV <<= Op1->getZExtValue();
6986 assert(I.getOpcode() == Instruction::LShr && "Unknown logical shift");
6987 MaskV = MaskV.lshr(Op1->getZExtValue());
6990 Instruction *And = BinaryOperator::CreateAnd(NSh, ConstantInt::get(MaskV),
6992 InsertNewInstBefore(And, I); // shift1 & 0x00FF
6994 // Return the value truncated to the interesting size.
6995 return new TruncInst(And, I.getType());
6999 if (Op0->hasOneUse()) {
7000 if (BinaryOperator *Op0BO = dyn_cast<BinaryOperator>(Op0)) {
7001 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
7004 switch (Op0BO->getOpcode()) {
7006 case Instruction::Add:
7007 case Instruction::And:
7008 case Instruction::Or:
7009 case Instruction::Xor: {
7010 // These operators commute.
7011 // Turn (Y + (X >> C)) << C -> (X + (Y << C)) & (~0 << C)
7012 if (isLeftShift && Op0BO->getOperand(1)->hasOneUse() &&
7013 match(Op0BO->getOperand(1), m_Shr(m_Value(V1), m_Specific(Op1)))){
7014 Instruction *YS = BinaryOperator::CreateShl(
7015 Op0BO->getOperand(0), Op1,
7017 InsertNewInstBefore(YS, I); // (Y << C)
7019 BinaryOperator::Create(Op0BO->getOpcode(), YS, V1,
7020 Op0BO->getOperand(1)->getName());
7021 InsertNewInstBefore(X, I); // (X + (Y << C))
7022 uint32_t Op1Val = Op1->getLimitedValue(TypeBits);
7023 return BinaryOperator::CreateAnd(X, ConstantInt::get(
7024 APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
7027 // Turn (Y + ((X >> C) & CC)) << C -> ((X & (CC << C)) + (Y << C))
7028 Value *Op0BOOp1 = Op0BO->getOperand(1);
7029 if (isLeftShift && Op0BOOp1->hasOneUse() &&
7031 m_And(m_Shr(m_Value(V1), m_Specific(Op1)),
7032 m_ConstantInt(CC))) &&
7033 cast<BinaryOperator>(Op0BOOp1)->getOperand(0)->hasOneUse()) {
7034 Instruction *YS = BinaryOperator::CreateShl(
7035 Op0BO->getOperand(0), Op1,
7037 InsertNewInstBefore(YS, I); // (Y << C)
7039 BinaryOperator::CreateAnd(V1, ConstantExpr::getShl(CC, Op1),
7040 V1->getName()+".mask");
7041 InsertNewInstBefore(XM, I); // X & (CC << C)
7043 return BinaryOperator::Create(Op0BO->getOpcode(), YS, XM);
7048 case Instruction::Sub: {
7049 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
7050 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
7051 match(Op0BO->getOperand(0), m_Shr(m_Value(V1), m_Specific(Op1)))){
7052 Instruction *YS = BinaryOperator::CreateShl(
7053 Op0BO->getOperand(1), Op1,
7055 InsertNewInstBefore(YS, I); // (Y << C)
7057 BinaryOperator::Create(Op0BO->getOpcode(), V1, YS,
7058 Op0BO->getOperand(0)->getName());
7059 InsertNewInstBefore(X, I); // (X + (Y << C))
7060 uint32_t Op1Val = Op1->getLimitedValue(TypeBits);
7061 return BinaryOperator::CreateAnd(X, ConstantInt::get(
7062 APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
7065 // Turn (((X >> C)&CC) + Y) << C -> (X + (Y << C)) & (CC << C)
7066 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
7067 match(Op0BO->getOperand(0),
7068 m_And(m_Shr(m_Value(V1), m_Value(V2)),
7069 m_ConstantInt(CC))) && V2 == Op1 &&
7070 cast<BinaryOperator>(Op0BO->getOperand(0))
7071 ->getOperand(0)->hasOneUse()) {
7072 Instruction *YS = BinaryOperator::CreateShl(
7073 Op0BO->getOperand(1), Op1,
7075 InsertNewInstBefore(YS, I); // (Y << C)
7077 BinaryOperator::CreateAnd(V1, ConstantExpr::getShl(CC, Op1),
7078 V1->getName()+".mask");
7079 InsertNewInstBefore(XM, I); // X & (CC << C)
7081 return BinaryOperator::Create(Op0BO->getOpcode(), XM, YS);
7089 // If the operand is an bitwise operator with a constant RHS, and the
7090 // shift is the only use, we can pull it out of the shift.
7091 if (ConstantInt *Op0C = dyn_cast<ConstantInt>(Op0BO->getOperand(1))) {
7092 bool isValid = true; // Valid only for And, Or, Xor
7093 bool highBitSet = false; // Transform if high bit of constant set?
7095 switch (Op0BO->getOpcode()) {
7096 default: isValid = false; break; // Do not perform transform!
7097 case Instruction::Add:
7098 isValid = isLeftShift;
7100 case Instruction::Or:
7101 case Instruction::Xor:
7104 case Instruction::And:
7109 // If this is a signed shift right, and the high bit is modified
7110 // by the logical operation, do not perform the transformation.
7111 // The highBitSet boolean indicates the value of the high bit of
7112 // the constant which would cause it to be modified for this
7115 if (isValid && I.getOpcode() == Instruction::AShr)
7116 isValid = Op0C->getValue()[TypeBits-1] == highBitSet;
7119 Constant *NewRHS = ConstantExpr::get(I.getOpcode(), Op0C, Op1);
7121 Instruction *NewShift =
7122 BinaryOperator::Create(I.getOpcode(), Op0BO->getOperand(0), Op1);
7123 InsertNewInstBefore(NewShift, I);
7124 NewShift->takeName(Op0BO);
7126 return BinaryOperator::Create(Op0BO->getOpcode(), NewShift,
7133 // Find out if this is a shift of a shift by a constant.
7134 BinaryOperator *ShiftOp = dyn_cast<BinaryOperator>(Op0);
7135 if (ShiftOp && !ShiftOp->isShift())
7138 if (ShiftOp && isa<ConstantInt>(ShiftOp->getOperand(1))) {
7139 ConstantInt *ShiftAmt1C = cast<ConstantInt>(ShiftOp->getOperand(1));
7140 uint32_t ShiftAmt1 = ShiftAmt1C->getLimitedValue(TypeBits);
7141 uint32_t ShiftAmt2 = Op1->getLimitedValue(TypeBits);
7142 assert(ShiftAmt2 != 0 && "Should have been simplified earlier");
7143 if (ShiftAmt1 == 0) return 0; // Will be simplified in the future.
7144 Value *X = ShiftOp->getOperand(0);
7146 uint32_t AmtSum = ShiftAmt1+ShiftAmt2; // Fold into one big shift.
7147 if (AmtSum > TypeBits)
7150 const IntegerType *Ty = cast<IntegerType>(I.getType());
7152 // Check for (X << c1) << c2 and (X >> c1) >> c2
7153 if (I.getOpcode() == ShiftOp->getOpcode()) {
7154 return BinaryOperator::Create(I.getOpcode(), X,
7155 ConstantInt::get(Ty, AmtSum));
7156 } else if (ShiftOp->getOpcode() == Instruction::LShr &&
7157 I.getOpcode() == Instruction::AShr) {
7158 // ((X >>u C1) >>s C2) -> (X >>u (C1+C2)) since C1 != 0.
7159 return BinaryOperator::CreateLShr(X, ConstantInt::get(Ty, AmtSum));
7160 } else if (ShiftOp->getOpcode() == Instruction::AShr &&
7161 I.getOpcode() == Instruction::LShr) {
7162 // ((X >>s C1) >>u C2) -> ((X >>s (C1+C2)) & mask) since C1 != 0.
7163 Instruction *Shift =
7164 BinaryOperator::CreateAShr(X, ConstantInt::get(Ty, AmtSum));
7165 InsertNewInstBefore(Shift, I);
7167 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
7168 return BinaryOperator::CreateAnd(Shift, ConstantInt::get(Mask));
7171 // Okay, if we get here, one shift must be left, and the other shift must be
7172 // right. See if the amounts are equal.
7173 if (ShiftAmt1 == ShiftAmt2) {
7174 // If we have ((X >>? C) << C), turn this into X & (-1 << C).
7175 if (I.getOpcode() == Instruction::Shl) {
7176 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt1));
7177 return BinaryOperator::CreateAnd(X, ConstantInt::get(Mask));
7179 // If we have ((X << C) >>u C), turn this into X & (-1 >>u C).
7180 if (I.getOpcode() == Instruction::LShr) {
7181 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt1));
7182 return BinaryOperator::CreateAnd(X, ConstantInt::get(Mask));
7184 // We can simplify ((X << C) >>s C) into a trunc + sext.
7185 // NOTE: we could do this for any C, but that would make 'unusual' integer
7186 // types. For now, just stick to ones well-supported by the code
7188 const Type *SExtType = 0;
7189 switch (Ty->getBitWidth() - ShiftAmt1) {
7196 SExtType = IntegerType::get(Ty->getBitWidth() - ShiftAmt1);
7201 Instruction *NewTrunc = new TruncInst(X, SExtType, "sext");
7202 InsertNewInstBefore(NewTrunc, I);
7203 return new SExtInst(NewTrunc, Ty);
7205 // Otherwise, we can't handle it yet.
7206 } else if (ShiftAmt1 < ShiftAmt2) {
7207 uint32_t ShiftDiff = ShiftAmt2-ShiftAmt1;
7209 // (X >>? C1) << C2 --> X << (C2-C1) & (-1 << C2)
7210 if (I.getOpcode() == Instruction::Shl) {
7211 assert(ShiftOp->getOpcode() == Instruction::LShr ||
7212 ShiftOp->getOpcode() == Instruction::AShr);
7213 Instruction *Shift =
7214 BinaryOperator::CreateShl(X, ConstantInt::get(Ty, ShiftDiff));
7215 InsertNewInstBefore(Shift, I);
7217 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
7218 return BinaryOperator::CreateAnd(Shift, ConstantInt::get(Mask));
7221 // (X << C1) >>u C2 --> X >>u (C2-C1) & (-1 >> C2)
7222 if (I.getOpcode() == Instruction::LShr) {
7223 assert(ShiftOp->getOpcode() == Instruction::Shl);
7224 Instruction *Shift =
7225 BinaryOperator::CreateLShr(X, ConstantInt::get(Ty, ShiftDiff));
7226 InsertNewInstBefore(Shift, I);
7228 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
7229 return BinaryOperator::CreateAnd(Shift, ConstantInt::get(Mask));
7232 // We can't handle (X << C1) >>s C2, it shifts arbitrary bits in.
7234 assert(ShiftAmt2 < ShiftAmt1);
7235 uint32_t ShiftDiff = ShiftAmt1-ShiftAmt2;
7237 // (X >>? C1) << C2 --> X >>? (C1-C2) & (-1 << C2)
7238 if (I.getOpcode() == Instruction::Shl) {
7239 assert(ShiftOp->getOpcode() == Instruction::LShr ||
7240 ShiftOp->getOpcode() == Instruction::AShr);
7241 Instruction *Shift =
7242 BinaryOperator::Create(ShiftOp->getOpcode(), X,
7243 ConstantInt::get(Ty, ShiftDiff));
7244 InsertNewInstBefore(Shift, I);
7246 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
7247 return BinaryOperator::CreateAnd(Shift, ConstantInt::get(Mask));
7250 // (X << C1) >>u C2 --> X << (C1-C2) & (-1 >> C2)
7251 if (I.getOpcode() == Instruction::LShr) {
7252 assert(ShiftOp->getOpcode() == Instruction::Shl);
7253 Instruction *Shift =
7254 BinaryOperator::CreateShl(X, ConstantInt::get(Ty, ShiftDiff));
7255 InsertNewInstBefore(Shift, I);
7257 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
7258 return BinaryOperator::CreateAnd(Shift, ConstantInt::get(Mask));
7261 // We can't handle (X << C1) >>a C2, it shifts arbitrary bits in.
7268 /// DecomposeSimpleLinearExpr - Analyze 'Val', seeing if it is a simple linear
7269 /// expression. If so, decompose it, returning some value X, such that Val is
7272 static Value *DecomposeSimpleLinearExpr(Value *Val, unsigned &Scale,
7274 assert(Val->getType() == Type::Int32Ty && "Unexpected allocation size type!");
7275 if (ConstantInt *CI = dyn_cast<ConstantInt>(Val)) {
7276 Offset = CI->getZExtValue();
7278 return ConstantInt::get(Type::Int32Ty, 0);
7279 } else if (BinaryOperator *I = dyn_cast<BinaryOperator>(Val)) {
7280 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
7281 if (I->getOpcode() == Instruction::Shl) {
7282 // This is a value scaled by '1 << the shift amt'.
7283 Scale = 1U << RHS->getZExtValue();
7285 return I->getOperand(0);
7286 } else if (I->getOpcode() == Instruction::Mul) {
7287 // This value is scaled by 'RHS'.
7288 Scale = RHS->getZExtValue();
7290 return I->getOperand(0);
7291 } else if (I->getOpcode() == Instruction::Add) {
7292 // We have X+C. Check to see if we really have (X*C2)+C1,
7293 // where C1 is divisible by C2.
7296 DecomposeSimpleLinearExpr(I->getOperand(0), SubScale, Offset);
7297 Offset += RHS->getZExtValue();
7304 // Otherwise, we can't look past this.
7311 /// PromoteCastOfAllocation - If we find a cast of an allocation instruction,
7312 /// try to eliminate the cast by moving the type information into the alloc.
7313 Instruction *InstCombiner::PromoteCastOfAllocation(BitCastInst &CI,
7314 AllocationInst &AI) {
7315 const PointerType *PTy = cast<PointerType>(CI.getType());
7317 // Remove any uses of AI that are dead.
7318 assert(!CI.use_empty() && "Dead instructions should be removed earlier!");
7320 for (Value::use_iterator UI = AI.use_begin(), E = AI.use_end(); UI != E; ) {
7321 Instruction *User = cast<Instruction>(*UI++);
7322 if (isInstructionTriviallyDead(User)) {
7323 while (UI != E && *UI == User)
7324 ++UI; // If this instruction uses AI more than once, don't break UI.
7327 DOUT << "IC: DCE: " << *User;
7328 EraseInstFromFunction(*User);
7332 // Get the type really allocated and the type casted to.
7333 const Type *AllocElTy = AI.getAllocatedType();
7334 const Type *CastElTy = PTy->getElementType();
7335 if (!AllocElTy->isSized() || !CastElTy->isSized()) return 0;
7337 unsigned AllocElTyAlign = TD->getABITypeAlignment(AllocElTy);
7338 unsigned CastElTyAlign = TD->getABITypeAlignment(CastElTy);
7339 if (CastElTyAlign < AllocElTyAlign) return 0;
7341 // If the allocation has multiple uses, only promote it if we are strictly
7342 // increasing the alignment of the resultant allocation. If we keep it the
7343 // same, we open the door to infinite loops of various kinds.
7344 if (!AI.hasOneUse() && CastElTyAlign == AllocElTyAlign) return 0;
7346 uint64_t AllocElTySize = TD->getABITypeSize(AllocElTy);
7347 uint64_t CastElTySize = TD->getABITypeSize(CastElTy);
7348 if (CastElTySize == 0 || AllocElTySize == 0) return 0;
7350 // See if we can satisfy the modulus by pulling a scale out of the array
7352 unsigned ArraySizeScale;
7354 Value *NumElements = // See if the array size is a decomposable linear expr.
7355 DecomposeSimpleLinearExpr(AI.getOperand(0), ArraySizeScale, ArrayOffset);
7357 // If we can now satisfy the modulus, by using a non-1 scale, we really can
7359 if ((AllocElTySize*ArraySizeScale) % CastElTySize != 0 ||
7360 (AllocElTySize*ArrayOffset ) % CastElTySize != 0) return 0;
7362 unsigned Scale = (AllocElTySize*ArraySizeScale)/CastElTySize;
7367 // If the allocation size is constant, form a constant mul expression
7368 Amt = ConstantInt::get(Type::Int32Ty, Scale);
7369 if (isa<ConstantInt>(NumElements))
7370 Amt = Multiply(cast<ConstantInt>(NumElements), cast<ConstantInt>(Amt));
7371 // otherwise multiply the amount and the number of elements
7372 else if (Scale != 1) {
7373 Instruction *Tmp = BinaryOperator::CreateMul(Amt, NumElements, "tmp");
7374 Amt = InsertNewInstBefore(Tmp, AI);
7378 if (int Offset = (AllocElTySize*ArrayOffset)/CastElTySize) {
7379 Value *Off = ConstantInt::get(Type::Int32Ty, Offset, true);
7380 Instruction *Tmp = BinaryOperator::CreateAdd(Amt, Off, "tmp");
7381 Amt = InsertNewInstBefore(Tmp, AI);
7384 AllocationInst *New;
7385 if (isa<MallocInst>(AI))
7386 New = new MallocInst(CastElTy, Amt, AI.getAlignment());
7388 New = new AllocaInst(CastElTy, Amt, AI.getAlignment());
7389 InsertNewInstBefore(New, AI);
7392 // If the allocation has multiple uses, insert a cast and change all things
7393 // that used it to use the new cast. This will also hack on CI, but it will
7395 if (!AI.hasOneUse()) {
7396 AddUsesToWorkList(AI);
7397 // New is the allocation instruction, pointer typed. AI is the original
7398 // allocation instruction, also pointer typed. Thus, cast to use is BitCast.
7399 CastInst *NewCast = new BitCastInst(New, AI.getType(), "tmpcast");
7400 InsertNewInstBefore(NewCast, AI);
7401 AI.replaceAllUsesWith(NewCast);
7403 return ReplaceInstUsesWith(CI, New);
7406 /// CanEvaluateInDifferentType - Return true if we can take the specified value
7407 /// and return it as type Ty without inserting any new casts and without
7408 /// changing the computed value. This is used by code that tries to decide
7409 /// whether promoting or shrinking integer operations to wider or smaller types
7410 /// will allow us to eliminate a truncate or extend.
7412 /// This is a truncation operation if Ty is smaller than V->getType(), or an
7413 /// extension operation if Ty is larger.
7415 /// If CastOpc is a truncation, then Ty will be a type smaller than V. We
7416 /// should return true if trunc(V) can be computed by computing V in the smaller
7417 /// type. If V is an instruction, then trunc(inst(x,y)) can be computed as
7418 /// inst(trunc(x),trunc(y)), which only makes sense if x and y can be
7419 /// efficiently truncated.
7421 /// If CastOpc is a sext or zext, we are asking if the low bits of the value can
7422 /// bit computed in a larger type, which is then and'd or sext_in_reg'd to get
7423 /// the final result.
7424 bool InstCombiner::CanEvaluateInDifferentType(Value *V, const IntegerType *Ty,
7426 int &NumCastsRemoved) {
7427 // We can always evaluate constants in another type.
7428 if (isa<ConstantInt>(V))
7431 Instruction *I = dyn_cast<Instruction>(V);
7432 if (!I) return false;
7434 const IntegerType *OrigTy = cast<IntegerType>(V->getType());
7436 // If this is an extension or truncate, we can often eliminate it.
7437 if (isa<TruncInst>(I) || isa<ZExtInst>(I) || isa<SExtInst>(I)) {
7438 // If this is a cast from the destination type, we can trivially eliminate
7439 // it, and this will remove a cast overall.
7440 if (I->getOperand(0)->getType() == Ty) {
7441 // If the first operand is itself a cast, and is eliminable, do not count
7442 // this as an eliminable cast. We would prefer to eliminate those two
7444 if (!isa<CastInst>(I->getOperand(0)) && I->hasOneUse())
7450 // We can't extend or shrink something that has multiple uses: doing so would
7451 // require duplicating the instruction in general, which isn't profitable.
7452 if (!I->hasOneUse()) return false;
7454 switch (I->getOpcode()) {
7455 case Instruction::Add:
7456 case Instruction::Sub:
7457 case Instruction::Mul:
7458 case Instruction::And:
7459 case Instruction::Or:
7460 case Instruction::Xor:
7461 // These operators can all arbitrarily be extended or truncated.
7462 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
7464 CanEvaluateInDifferentType(I->getOperand(1), Ty, CastOpc,
7467 case Instruction::Shl:
7468 // If we are truncating the result of this SHL, and if it's a shift of a
7469 // constant amount, we can always perform a SHL in a smaller type.
7470 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
7471 uint32_t BitWidth = Ty->getBitWidth();
7472 if (BitWidth < OrigTy->getBitWidth() &&
7473 CI->getLimitedValue(BitWidth) < BitWidth)
7474 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
7478 case Instruction::LShr:
7479 // If this is a truncate of a logical shr, we can truncate it to a smaller
7480 // lshr iff we know that the bits we would otherwise be shifting in are
7482 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
7483 uint32_t OrigBitWidth = OrigTy->getBitWidth();
7484 uint32_t BitWidth = Ty->getBitWidth();
7485 if (BitWidth < OrigBitWidth &&
7486 MaskedValueIsZero(I->getOperand(0),
7487 APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth)) &&
7488 CI->getLimitedValue(BitWidth) < BitWidth) {
7489 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
7494 case Instruction::ZExt:
7495 case Instruction::SExt:
7496 case Instruction::Trunc:
7497 // If this is the same kind of case as our original (e.g. zext+zext), we
7498 // can safely replace it. Note that replacing it does not reduce the number
7499 // of casts in the input.
7500 if (I->getOpcode() == CastOpc)
7503 case Instruction::Select: {
7504 SelectInst *SI = cast<SelectInst>(I);
7505 return CanEvaluateInDifferentType(SI->getTrueValue(), Ty, CastOpc,
7507 CanEvaluateInDifferentType(SI->getFalseValue(), Ty, CastOpc,
7510 case Instruction::PHI: {
7511 // We can change a phi if we can change all operands.
7512 PHINode *PN = cast<PHINode>(I);
7513 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
7514 if (!CanEvaluateInDifferentType(PN->getIncomingValue(i), Ty, CastOpc,
7520 // TODO: Can handle more cases here.
7527 /// EvaluateInDifferentType - Given an expression that
7528 /// CanEvaluateInDifferentType returns true for, actually insert the code to
7529 /// evaluate the expression.
7530 Value *InstCombiner::EvaluateInDifferentType(Value *V, const Type *Ty,
7532 if (Constant *C = dyn_cast<Constant>(V))
7533 return ConstantExpr::getIntegerCast(C, Ty, isSigned /*Sext or ZExt*/);
7535 // Otherwise, it must be an instruction.
7536 Instruction *I = cast<Instruction>(V);
7537 Instruction *Res = 0;
7538 switch (I->getOpcode()) {
7539 case Instruction::Add:
7540 case Instruction::Sub:
7541 case Instruction::Mul:
7542 case Instruction::And:
7543 case Instruction::Or:
7544 case Instruction::Xor:
7545 case Instruction::AShr:
7546 case Instruction::LShr:
7547 case Instruction::Shl: {
7548 Value *LHS = EvaluateInDifferentType(I->getOperand(0), Ty, isSigned);
7549 Value *RHS = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
7550 Res = BinaryOperator::Create((Instruction::BinaryOps)I->getOpcode(),
7554 case Instruction::Trunc:
7555 case Instruction::ZExt:
7556 case Instruction::SExt:
7557 // If the source type of the cast is the type we're trying for then we can
7558 // just return the source. There's no need to insert it because it is not
7560 if (I->getOperand(0)->getType() == Ty)
7561 return I->getOperand(0);
7563 // Otherwise, must be the same type of cast, so just reinsert a new one.
7564 Res = CastInst::Create(cast<CastInst>(I)->getOpcode(), I->getOperand(0),
7567 case Instruction::Select: {
7568 Value *True = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
7569 Value *False = EvaluateInDifferentType(I->getOperand(2), Ty, isSigned);
7570 Res = SelectInst::Create(I->getOperand(0), True, False);
7573 case Instruction::PHI: {
7574 PHINode *OPN = cast<PHINode>(I);
7575 PHINode *NPN = PHINode::Create(Ty);
7576 for (unsigned i = 0, e = OPN->getNumIncomingValues(); i != e; ++i) {
7577 Value *V =EvaluateInDifferentType(OPN->getIncomingValue(i), Ty, isSigned);
7578 NPN->addIncoming(V, OPN->getIncomingBlock(i));
7584 // TODO: Can handle more cases here.
7585 assert(0 && "Unreachable!");
7590 return InsertNewInstBefore(Res, *I);
7593 /// @brief Implement the transforms common to all CastInst visitors.
7594 Instruction *InstCombiner::commonCastTransforms(CastInst &CI) {
7595 Value *Src = CI.getOperand(0);
7597 // Many cases of "cast of a cast" are eliminable. If it's eliminable we just
7598 // eliminate it now.
7599 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
7600 if (Instruction::CastOps opc =
7601 isEliminableCastPair(CSrc, CI.getOpcode(), CI.getType(), TD)) {
7602 // The first cast (CSrc) is eliminable so we need to fix up or replace
7603 // the second cast (CI). CSrc will then have a good chance of being dead.
7604 return CastInst::Create(opc, CSrc->getOperand(0), CI.getType());
7608 // If we are casting a select then fold the cast into the select
7609 if (SelectInst *SI = dyn_cast<SelectInst>(Src))
7610 if (Instruction *NV = FoldOpIntoSelect(CI, SI, this))
7613 // If we are casting a PHI then fold the cast into the PHI
7614 if (isa<PHINode>(Src))
7615 if (Instruction *NV = FoldOpIntoPhi(CI))
7621 /// @brief Implement the transforms for cast of pointer (bitcast/ptrtoint)
7622 Instruction *InstCombiner::commonPointerCastTransforms(CastInst &CI) {
7623 Value *Src = CI.getOperand(0);
7625 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Src)) {
7626 // If casting the result of a getelementptr instruction with no offset, turn
7627 // this into a cast of the original pointer!
7628 if (GEP->hasAllZeroIndices()) {
7629 // Changing the cast operand is usually not a good idea but it is safe
7630 // here because the pointer operand is being replaced with another
7631 // pointer operand so the opcode doesn't need to change.
7633 CI.setOperand(0, GEP->getOperand(0));
7637 // If the GEP has a single use, and the base pointer is a bitcast, and the
7638 // GEP computes a constant offset, see if we can convert these three
7639 // instructions into fewer. This typically happens with unions and other
7640 // non-type-safe code.
7641 if (GEP->hasOneUse() && isa<BitCastInst>(GEP->getOperand(0))) {
7642 if (GEP->hasAllConstantIndices()) {
7643 // We are guaranteed to get a constant from EmitGEPOffset.
7644 ConstantInt *OffsetV = cast<ConstantInt>(EmitGEPOffset(GEP, CI, *this));
7645 int64_t Offset = OffsetV->getSExtValue();
7647 // Get the base pointer input of the bitcast, and the type it points to.
7648 Value *OrigBase = cast<BitCastInst>(GEP->getOperand(0))->getOperand(0);
7649 const Type *GEPIdxTy =
7650 cast<PointerType>(OrigBase->getType())->getElementType();
7651 if (GEPIdxTy->isSized()) {
7652 SmallVector<Value*, 8> NewIndices;
7654 // Start with the index over the outer type. Note that the type size
7655 // might be zero (even if the offset isn't zero) if the indexed type
7656 // is something like [0 x {int, int}]
7657 const Type *IntPtrTy = TD->getIntPtrType();
7658 int64_t FirstIdx = 0;
7659 if (int64_t TySize = TD->getABITypeSize(GEPIdxTy)) {
7660 FirstIdx = Offset/TySize;
7663 // Handle silly modulus not returning values values [0..TySize).
7667 assert(Offset >= 0);
7669 assert((uint64_t)Offset < (uint64_t)TySize &&"Out of range offset");
7672 NewIndices.push_back(ConstantInt::get(IntPtrTy, FirstIdx));
7674 // Index into the types. If we fail, set OrigBase to null.
7676 if (const StructType *STy = dyn_cast<StructType>(GEPIdxTy)) {
7677 const StructLayout *SL = TD->getStructLayout(STy);
7678 if (Offset < (int64_t)SL->getSizeInBytes()) {
7679 unsigned Elt = SL->getElementContainingOffset(Offset);
7680 NewIndices.push_back(ConstantInt::get(Type::Int32Ty, Elt));
7682 Offset -= SL->getElementOffset(Elt);
7683 GEPIdxTy = STy->getElementType(Elt);
7685 // Otherwise, we can't index into this, bail out.
7689 } else if (isa<ArrayType>(GEPIdxTy) || isa<VectorType>(GEPIdxTy)) {
7690 const SequentialType *STy = cast<SequentialType>(GEPIdxTy);
7691 if (uint64_t EltSize = TD->getABITypeSize(STy->getElementType())){
7692 NewIndices.push_back(ConstantInt::get(IntPtrTy,Offset/EltSize));
7695 NewIndices.push_back(ConstantInt::get(IntPtrTy, 0));
7697 GEPIdxTy = STy->getElementType();
7699 // Otherwise, we can't index into this, bail out.
7705 // If we were able to index down into an element, create the GEP
7706 // and bitcast the result. This eliminates one bitcast, potentially
7708 Instruction *NGEP = GetElementPtrInst::Create(OrigBase,
7710 NewIndices.end(), "");
7711 InsertNewInstBefore(NGEP, CI);
7712 NGEP->takeName(GEP);
7714 if (isa<BitCastInst>(CI))
7715 return new BitCastInst(NGEP, CI.getType());
7716 assert(isa<PtrToIntInst>(CI));
7717 return new PtrToIntInst(NGEP, CI.getType());
7724 return commonCastTransforms(CI);
7729 /// Only the TRUNC, ZEXT, SEXT, and BITCAST can both operand and result as
7730 /// integer types. This function implements the common transforms for all those
7732 /// @brief Implement the transforms common to CastInst with integer operands
7733 Instruction *InstCombiner::commonIntCastTransforms(CastInst &CI) {
7734 if (Instruction *Result = commonCastTransforms(CI))
7737 Value *Src = CI.getOperand(0);
7738 const Type *SrcTy = Src->getType();
7739 const Type *DestTy = CI.getType();
7740 uint32_t SrcBitSize = SrcTy->getPrimitiveSizeInBits();
7741 uint32_t DestBitSize = DestTy->getPrimitiveSizeInBits();
7743 // See if we can simplify any instructions used by the LHS whose sole
7744 // purpose is to compute bits we don't care about.
7745 APInt KnownZero(DestBitSize, 0), KnownOne(DestBitSize, 0);
7746 if (SimplifyDemandedBits(&CI, APInt::getAllOnesValue(DestBitSize),
7747 KnownZero, KnownOne))
7750 // If the source isn't an instruction or has more than one use then we
7751 // can't do anything more.
7752 Instruction *SrcI = dyn_cast<Instruction>(Src);
7753 if (!SrcI || !Src->hasOneUse())
7756 // Attempt to propagate the cast into the instruction for int->int casts.
7757 int NumCastsRemoved = 0;
7758 if (!isa<BitCastInst>(CI) &&
7759 CanEvaluateInDifferentType(SrcI, cast<IntegerType>(DestTy),
7760 CI.getOpcode(), NumCastsRemoved)) {
7761 // If this cast is a truncate, evaluting in a different type always
7762 // eliminates the cast, so it is always a win. If this is a zero-extension,
7763 // we need to do an AND to maintain the clear top-part of the computation,
7764 // so we require that the input have eliminated at least one cast. If this
7765 // is a sign extension, we insert two new casts (to do the extension) so we
7766 // require that two casts have been eliminated.
7768 switch (CI.getOpcode()) {
7770 // All the others use floating point so we shouldn't actually
7771 // get here because of the check above.
7772 assert(0 && "Unknown cast type");
7773 case Instruction::Trunc:
7776 case Instruction::ZExt:
7777 DoXForm = NumCastsRemoved >= 1;
7779 case Instruction::SExt:
7780 DoXForm = NumCastsRemoved >= 2;
7785 Value *Res = EvaluateInDifferentType(SrcI, DestTy,
7786 CI.getOpcode() == Instruction::SExt);
7787 assert(Res->getType() == DestTy);
7788 switch (CI.getOpcode()) {
7789 default: assert(0 && "Unknown cast type!");
7790 case Instruction::Trunc:
7791 case Instruction::BitCast:
7792 // Just replace this cast with the result.
7793 return ReplaceInstUsesWith(CI, Res);
7794 case Instruction::ZExt: {
7795 // We need to emit an AND to clear the high bits.
7796 assert(SrcBitSize < DestBitSize && "Not a zext?");
7797 Constant *C = ConstantInt::get(APInt::getLowBitsSet(DestBitSize,
7799 return BinaryOperator::CreateAnd(Res, C);
7801 case Instruction::SExt:
7802 // We need to emit a cast to truncate, then a cast to sext.
7803 return CastInst::Create(Instruction::SExt,
7804 InsertCastBefore(Instruction::Trunc, Res, Src->getType(),
7810 Value *Op0 = SrcI->getNumOperands() > 0 ? SrcI->getOperand(0) : 0;
7811 Value *Op1 = SrcI->getNumOperands() > 1 ? SrcI->getOperand(1) : 0;
7813 switch (SrcI->getOpcode()) {
7814 case Instruction::Add:
7815 case Instruction::Mul:
7816 case Instruction::And:
7817 case Instruction::Or:
7818 case Instruction::Xor:
7819 // If we are discarding information, rewrite.
7820 if (DestBitSize <= SrcBitSize && DestBitSize != 1) {
7821 // Don't insert two casts if they cannot be eliminated. We allow
7822 // two casts to be inserted if the sizes are the same. This could
7823 // only be converting signedness, which is a noop.
7824 if (DestBitSize == SrcBitSize ||
7825 !ValueRequiresCast(CI.getOpcode(), Op1, DestTy,TD) ||
7826 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
7827 Instruction::CastOps opcode = CI.getOpcode();
7828 Value *Op0c = InsertOperandCastBefore(opcode, Op0, DestTy, SrcI);
7829 Value *Op1c = InsertOperandCastBefore(opcode, Op1, DestTy, SrcI);
7830 return BinaryOperator::Create(
7831 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
7835 // cast (xor bool X, true) to int --> xor (cast bool X to int), 1
7836 if (isa<ZExtInst>(CI) && SrcBitSize == 1 &&
7837 SrcI->getOpcode() == Instruction::Xor &&
7838 Op1 == ConstantInt::getTrue() &&
7839 (!Op0->hasOneUse() || !isa<CmpInst>(Op0))) {
7840 Value *New = InsertOperandCastBefore(Instruction::ZExt, Op0, DestTy, &CI);
7841 return BinaryOperator::CreateXor(New, ConstantInt::get(CI.getType(), 1));
7844 case Instruction::SDiv:
7845 case Instruction::UDiv:
7846 case Instruction::SRem:
7847 case Instruction::URem:
7848 // If we are just changing the sign, rewrite.
7849 if (DestBitSize == SrcBitSize) {
7850 // Don't insert two casts if they cannot be eliminated. We allow
7851 // two casts to be inserted if the sizes are the same. This could
7852 // only be converting signedness, which is a noop.
7853 if (!ValueRequiresCast(CI.getOpcode(), Op1, DestTy, TD) ||
7854 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
7855 Value *Op0c = InsertOperandCastBefore(Instruction::BitCast,
7857 Value *Op1c = InsertOperandCastBefore(Instruction::BitCast,
7859 return BinaryOperator::Create(
7860 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
7865 case Instruction::Shl:
7866 // Allow changing the sign of the source operand. Do not allow
7867 // changing the size of the shift, UNLESS the shift amount is a
7868 // constant. We must not change variable sized shifts to a smaller
7869 // size, because it is undefined to shift more bits out than exist
7871 if (DestBitSize == SrcBitSize ||
7872 (DestBitSize < SrcBitSize && isa<Constant>(Op1))) {
7873 Instruction::CastOps opcode = (DestBitSize == SrcBitSize ?
7874 Instruction::BitCast : Instruction::Trunc);
7875 Value *Op0c = InsertOperandCastBefore(opcode, Op0, DestTy, SrcI);
7876 Value *Op1c = InsertOperandCastBefore(opcode, Op1, DestTy, SrcI);
7877 return BinaryOperator::CreateShl(Op0c, Op1c);
7880 case Instruction::AShr:
7881 // If this is a signed shr, and if all bits shifted in are about to be
7882 // truncated off, turn it into an unsigned shr to allow greater
7884 if (DestBitSize < SrcBitSize &&
7885 isa<ConstantInt>(Op1)) {
7886 uint32_t ShiftAmt = cast<ConstantInt>(Op1)->getLimitedValue(SrcBitSize);
7887 if (SrcBitSize > ShiftAmt && SrcBitSize-ShiftAmt >= DestBitSize) {
7888 // Insert the new logical shift right.
7889 return BinaryOperator::CreateLShr(Op0, Op1);
7897 Instruction *InstCombiner::visitTrunc(TruncInst &CI) {
7898 if (Instruction *Result = commonIntCastTransforms(CI))
7901 Value *Src = CI.getOperand(0);
7902 const Type *Ty = CI.getType();
7903 uint32_t DestBitWidth = Ty->getPrimitiveSizeInBits();
7904 uint32_t SrcBitWidth = cast<IntegerType>(Src->getType())->getBitWidth();
7906 if (Instruction *SrcI = dyn_cast<Instruction>(Src)) {
7907 switch (SrcI->getOpcode()) {
7909 case Instruction::LShr:
7910 // We can shrink lshr to something smaller if we know the bits shifted in
7911 // are already zeros.
7912 if (ConstantInt *ShAmtV = dyn_cast<ConstantInt>(SrcI->getOperand(1))) {
7913 uint32_t ShAmt = ShAmtV->getLimitedValue(SrcBitWidth);
7915 // Get a mask for the bits shifting in.
7916 APInt Mask(APInt::getLowBitsSet(SrcBitWidth, ShAmt).shl(DestBitWidth));
7917 Value* SrcIOp0 = SrcI->getOperand(0);
7918 if (SrcI->hasOneUse() && MaskedValueIsZero(SrcIOp0, Mask)) {
7919 if (ShAmt >= DestBitWidth) // All zeros.
7920 return ReplaceInstUsesWith(CI, Constant::getNullValue(Ty));
7922 // Okay, we can shrink this. Truncate the input, then return a new
7924 Value *V1 = InsertCastBefore(Instruction::Trunc, SrcIOp0, Ty, CI);
7925 Value *V2 = InsertCastBefore(Instruction::Trunc, SrcI->getOperand(1),
7927 return BinaryOperator::CreateLShr(V1, V2);
7929 } else { // This is a variable shr.
7931 // Turn 'trunc (lshr X, Y) to bool' into '(X & (1 << Y)) != 0'. This is
7932 // more LLVM instructions, but allows '1 << Y' to be hoisted if
7933 // loop-invariant and CSE'd.
7934 if (CI.getType() == Type::Int1Ty && SrcI->hasOneUse()) {
7935 Value *One = ConstantInt::get(SrcI->getType(), 1);
7937 Value *V = InsertNewInstBefore(
7938 BinaryOperator::CreateShl(One, SrcI->getOperand(1),
7940 V = InsertNewInstBefore(BinaryOperator::CreateAnd(V,
7941 SrcI->getOperand(0),
7943 Value *Zero = Constant::getNullValue(V->getType());
7944 return new ICmpInst(ICmpInst::ICMP_NE, V, Zero);
7954 /// transformZExtICmp - Transform (zext icmp) to bitwise / integer operations
7955 /// in order to eliminate the icmp.
7956 Instruction *InstCombiner::transformZExtICmp(ICmpInst *ICI, Instruction &CI,
7958 // If we are just checking for a icmp eq of a single bit and zext'ing it
7959 // to an integer, then shift the bit to the appropriate place and then
7960 // cast to integer to avoid the comparison.
7961 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(ICI->getOperand(1))) {
7962 const APInt &Op1CV = Op1C->getValue();
7964 // zext (x <s 0) to i32 --> x>>u31 true if signbit set.
7965 // zext (x >s -1) to i32 --> (x>>u31)^1 true if signbit clear.
7966 if ((ICI->getPredicate() == ICmpInst::ICMP_SLT && Op1CV == 0) ||
7967 (ICI->getPredicate() == ICmpInst::ICMP_SGT &&Op1CV.isAllOnesValue())) {
7968 if (!DoXform) return ICI;
7970 Value *In = ICI->getOperand(0);
7971 Value *Sh = ConstantInt::get(In->getType(),
7972 In->getType()->getPrimitiveSizeInBits()-1);
7973 In = InsertNewInstBefore(BinaryOperator::CreateLShr(In, Sh,
7974 In->getName()+".lobit"),
7976 if (In->getType() != CI.getType())
7977 In = CastInst::CreateIntegerCast(In, CI.getType(),
7978 false/*ZExt*/, "tmp", &CI);
7980 if (ICI->getPredicate() == ICmpInst::ICMP_SGT) {
7981 Constant *One = ConstantInt::get(In->getType(), 1);
7982 In = InsertNewInstBefore(BinaryOperator::CreateXor(In, One,
7983 In->getName()+".not"),
7987 return ReplaceInstUsesWith(CI, In);
7992 // zext (X == 0) to i32 --> X^1 iff X has only the low bit set.
7993 // zext (X == 0) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
7994 // zext (X == 1) to i32 --> X iff X has only the low bit set.
7995 // zext (X == 2) to i32 --> X>>1 iff X has only the 2nd bit set.
7996 // zext (X != 0) to i32 --> X iff X has only the low bit set.
7997 // zext (X != 0) to i32 --> X>>1 iff X has only the 2nd bit set.
7998 // zext (X != 1) to i32 --> X^1 iff X has only the low bit set.
7999 // zext (X != 2) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
8000 if ((Op1CV == 0 || Op1CV.isPowerOf2()) &&
8001 // This only works for EQ and NE
8002 ICI->isEquality()) {
8003 // If Op1C some other power of two, convert:
8004 uint32_t BitWidth = Op1C->getType()->getBitWidth();
8005 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
8006 APInt TypeMask(APInt::getAllOnesValue(BitWidth));
8007 ComputeMaskedBits(ICI->getOperand(0), TypeMask, KnownZero, KnownOne);
8009 APInt KnownZeroMask(~KnownZero);
8010 if (KnownZeroMask.isPowerOf2()) { // Exactly 1 possible 1?
8011 if (!DoXform) return ICI;
8013 bool isNE = ICI->getPredicate() == ICmpInst::ICMP_NE;
8014 if (Op1CV != 0 && (Op1CV != KnownZeroMask)) {
8015 // (X&4) == 2 --> false
8016 // (X&4) != 2 --> true
8017 Constant *Res = ConstantInt::get(Type::Int1Ty, isNE);
8018 Res = ConstantExpr::getZExt(Res, CI.getType());
8019 return ReplaceInstUsesWith(CI, Res);
8022 uint32_t ShiftAmt = KnownZeroMask.logBase2();
8023 Value *In = ICI->getOperand(0);
8025 // Perform a logical shr by shiftamt.
8026 // Insert the shift to put the result in the low bit.
8027 In = InsertNewInstBefore(BinaryOperator::CreateLShr(In,
8028 ConstantInt::get(In->getType(), ShiftAmt),
8029 In->getName()+".lobit"), CI);
8032 if ((Op1CV != 0) == isNE) { // Toggle the low bit.
8033 Constant *One = ConstantInt::get(In->getType(), 1);
8034 In = BinaryOperator::CreateXor(In, One, "tmp");
8035 InsertNewInstBefore(cast<Instruction>(In), CI);
8038 if (CI.getType() == In->getType())
8039 return ReplaceInstUsesWith(CI, In);
8041 return CastInst::CreateIntegerCast(In, CI.getType(), false/*ZExt*/);
8049 Instruction *InstCombiner::visitZExt(ZExtInst &CI) {
8050 // If one of the common conversion will work ..
8051 if (Instruction *Result = commonIntCastTransforms(CI))
8054 Value *Src = CI.getOperand(0);
8056 // If this is a cast of a cast
8057 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
8058 // If this is a TRUNC followed by a ZEXT then we are dealing with integral
8059 // types and if the sizes are just right we can convert this into a logical
8060 // 'and' which will be much cheaper than the pair of casts.
8061 if (isa<TruncInst>(CSrc)) {
8062 // Get the sizes of the types involved
8063 Value *A = CSrc->getOperand(0);
8064 uint32_t SrcSize = A->getType()->getPrimitiveSizeInBits();
8065 uint32_t MidSize = CSrc->getType()->getPrimitiveSizeInBits();
8066 uint32_t DstSize = CI.getType()->getPrimitiveSizeInBits();
8067 // If we're actually extending zero bits and the trunc is a no-op
8068 if (MidSize < DstSize && SrcSize == DstSize) {
8069 // Replace both of the casts with an And of the type mask.
8070 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
8071 Constant *AndConst = ConstantInt::get(AndValue);
8073 BinaryOperator::CreateAnd(CSrc->getOperand(0), AndConst);
8074 // Unfortunately, if the type changed, we need to cast it back.
8075 if (And->getType() != CI.getType()) {
8076 And->setName(CSrc->getName()+".mask");
8077 InsertNewInstBefore(And, CI);
8078 And = CastInst::CreateIntegerCast(And, CI.getType(), false/*ZExt*/);
8085 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src))
8086 return transformZExtICmp(ICI, CI);
8088 BinaryOperator *SrcI = dyn_cast<BinaryOperator>(Src);
8089 if (SrcI && SrcI->getOpcode() == Instruction::Or) {
8090 // zext (or icmp, icmp) --> or (zext icmp), (zext icmp) if at least one
8091 // of the (zext icmp) will be transformed.
8092 ICmpInst *LHS = dyn_cast<ICmpInst>(SrcI->getOperand(0));
8093 ICmpInst *RHS = dyn_cast<ICmpInst>(SrcI->getOperand(1));
8094 if (LHS && RHS && LHS->hasOneUse() && RHS->hasOneUse() &&
8095 (transformZExtICmp(LHS, CI, false) ||
8096 transformZExtICmp(RHS, CI, false))) {
8097 Value *LCast = InsertCastBefore(Instruction::ZExt, LHS, CI.getType(), CI);
8098 Value *RCast = InsertCastBefore(Instruction::ZExt, RHS, CI.getType(), CI);
8099 return BinaryOperator::Create(Instruction::Or, LCast, RCast);
8106 Instruction *InstCombiner::visitSExt(SExtInst &CI) {
8107 if (Instruction *I = commonIntCastTransforms(CI))
8110 Value *Src = CI.getOperand(0);
8112 // Canonicalize sign-extend from i1 to a select.
8113 if (Src->getType() == Type::Int1Ty)
8114 return SelectInst::Create(Src,
8115 ConstantInt::getAllOnesValue(CI.getType()),
8116 Constant::getNullValue(CI.getType()));
8118 // See if the value being truncated is already sign extended. If so, just
8119 // eliminate the trunc/sext pair.
8120 if (getOpcode(Src) == Instruction::Trunc) {
8121 Value *Op = cast<User>(Src)->getOperand(0);
8122 unsigned OpBits = cast<IntegerType>(Op->getType())->getBitWidth();
8123 unsigned MidBits = cast<IntegerType>(Src->getType())->getBitWidth();
8124 unsigned DestBits = cast<IntegerType>(CI.getType())->getBitWidth();
8125 unsigned NumSignBits = ComputeNumSignBits(Op);
8127 if (OpBits == DestBits) {
8128 // Op is i32, Mid is i8, and Dest is i32. If Op has more than 24 sign
8129 // bits, it is already ready.
8130 if (NumSignBits > DestBits-MidBits)
8131 return ReplaceInstUsesWith(CI, Op);
8132 } else if (OpBits < DestBits) {
8133 // Op is i32, Mid is i8, and Dest is i64. If Op has more than 24 sign
8134 // bits, just sext from i32.
8135 if (NumSignBits > OpBits-MidBits)
8136 return new SExtInst(Op, CI.getType(), "tmp");
8138 // Op is i64, Mid is i8, and Dest is i32. If Op has more than 56 sign
8139 // bits, just truncate to i32.
8140 if (NumSignBits > OpBits-MidBits)
8141 return new TruncInst(Op, CI.getType(), "tmp");
8145 // If the input is a shl/ashr pair of a same constant, then this is a sign
8146 // extension from a smaller value. If we could trust arbitrary bitwidth
8147 // integers, we could turn this into a truncate to the smaller bit and then
8148 // use a sext for the whole extension. Since we don't, look deeper and check
8149 // for a truncate. If the source and dest are the same type, eliminate the
8150 // trunc and extend and just do shifts. For example, turn:
8151 // %a = trunc i32 %i to i8
8152 // %b = shl i8 %a, 6
8153 // %c = ashr i8 %b, 6
8154 // %d = sext i8 %c to i32
8156 // %a = shl i32 %i, 30
8157 // %d = ashr i32 %a, 30
8159 ConstantInt *BA = 0, *CA = 0;
8160 if (match(Src, m_AShr(m_Shl(m_Value(A), m_ConstantInt(BA)),
8161 m_ConstantInt(CA))) &&
8162 BA == CA && isa<TruncInst>(A)) {
8163 Value *I = cast<TruncInst>(A)->getOperand(0);
8164 if (I->getType() == CI.getType()) {
8165 unsigned MidSize = Src->getType()->getPrimitiveSizeInBits();
8166 unsigned SrcDstSize = CI.getType()->getPrimitiveSizeInBits();
8167 unsigned ShAmt = CA->getZExtValue()+SrcDstSize-MidSize;
8168 Constant *ShAmtV = ConstantInt::get(CI.getType(), ShAmt);
8169 I = InsertNewInstBefore(BinaryOperator::CreateShl(I, ShAmtV,
8171 return BinaryOperator::CreateAShr(I, ShAmtV);
8178 /// FitsInFPType - Return a Constant* for the specified FP constant if it fits
8179 /// in the specified FP type without changing its value.
8180 static Constant *FitsInFPType(ConstantFP *CFP, const fltSemantics &Sem) {
8182 APFloat F = CFP->getValueAPF();
8183 (void)F.convert(Sem, APFloat::rmNearestTiesToEven, &losesInfo);
8185 return ConstantFP::get(F);
8189 /// LookThroughFPExtensions - If this is an fp extension instruction, look
8190 /// through it until we get the source value.
8191 static Value *LookThroughFPExtensions(Value *V) {
8192 if (Instruction *I = dyn_cast<Instruction>(V))
8193 if (I->getOpcode() == Instruction::FPExt)
8194 return LookThroughFPExtensions(I->getOperand(0));
8196 // If this value is a constant, return the constant in the smallest FP type
8197 // that can accurately represent it. This allows us to turn
8198 // (float)((double)X+2.0) into x+2.0f.
8199 if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
8200 if (CFP->getType() == Type::PPC_FP128Ty)
8201 return V; // No constant folding of this.
8202 // See if the value can be truncated to float and then reextended.
8203 if (Value *V = FitsInFPType(CFP, APFloat::IEEEsingle))
8205 if (CFP->getType() == Type::DoubleTy)
8206 return V; // Won't shrink.
8207 if (Value *V = FitsInFPType(CFP, APFloat::IEEEdouble))
8209 // Don't try to shrink to various long double types.
8215 Instruction *InstCombiner::visitFPTrunc(FPTruncInst &CI) {
8216 if (Instruction *I = commonCastTransforms(CI))
8219 // If we have fptrunc(add (fpextend x), (fpextend y)), where x and y are
8220 // smaller than the destination type, we can eliminate the truncate by doing
8221 // the add as the smaller type. This applies to add/sub/mul/div as well as
8222 // many builtins (sqrt, etc).
8223 BinaryOperator *OpI = dyn_cast<BinaryOperator>(CI.getOperand(0));
8224 if (OpI && OpI->hasOneUse()) {
8225 switch (OpI->getOpcode()) {
8227 case Instruction::Add:
8228 case Instruction::Sub:
8229 case Instruction::Mul:
8230 case Instruction::FDiv:
8231 case Instruction::FRem:
8232 const Type *SrcTy = OpI->getType();
8233 Value *LHSTrunc = LookThroughFPExtensions(OpI->getOperand(0));
8234 Value *RHSTrunc = LookThroughFPExtensions(OpI->getOperand(1));
8235 if (LHSTrunc->getType() != SrcTy &&
8236 RHSTrunc->getType() != SrcTy) {
8237 unsigned DstSize = CI.getType()->getPrimitiveSizeInBits();
8238 // If the source types were both smaller than the destination type of
8239 // the cast, do this xform.
8240 if (LHSTrunc->getType()->getPrimitiveSizeInBits() <= DstSize &&
8241 RHSTrunc->getType()->getPrimitiveSizeInBits() <= DstSize) {
8242 LHSTrunc = InsertCastBefore(Instruction::FPExt, LHSTrunc,
8244 RHSTrunc = InsertCastBefore(Instruction::FPExt, RHSTrunc,
8246 return BinaryOperator::Create(OpI->getOpcode(), LHSTrunc, RHSTrunc);
8255 Instruction *InstCombiner::visitFPExt(CastInst &CI) {
8256 return commonCastTransforms(CI);
8259 Instruction *InstCombiner::visitFPToUI(FPToUIInst &FI) {
8260 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0));
8262 return commonCastTransforms(FI);
8264 // fptoui(uitofp(X)) --> X
8265 // fptoui(sitofp(X)) --> X
8266 // This is safe if the intermediate type has enough bits in its mantissa to
8267 // accurately represent all values of X. For example, do not do this with
8268 // i64->float->i64. This is also safe for sitofp case, because any negative
8269 // 'X' value would cause an undefined result for the fptoui.
8270 if ((isa<UIToFPInst>(OpI) || isa<SIToFPInst>(OpI)) &&
8271 OpI->getOperand(0)->getType() == FI.getType() &&
8272 (int)FI.getType()->getPrimitiveSizeInBits() < /*extra bit for sign */
8273 OpI->getType()->getFPMantissaWidth())
8274 return ReplaceInstUsesWith(FI, OpI->getOperand(0));
8276 return commonCastTransforms(FI);
8279 Instruction *InstCombiner::visitFPToSI(FPToSIInst &FI) {
8280 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0));
8282 return commonCastTransforms(FI);
8284 // fptosi(sitofp(X)) --> X
8285 // fptosi(uitofp(X)) --> X
8286 // This is safe if the intermediate type has enough bits in its mantissa to
8287 // accurately represent all values of X. For example, do not do this with
8288 // i64->float->i64. This is also safe for sitofp case, because any negative
8289 // 'X' value would cause an undefined result for the fptoui.
8290 if ((isa<UIToFPInst>(OpI) || isa<SIToFPInst>(OpI)) &&
8291 OpI->getOperand(0)->getType() == FI.getType() &&
8292 (int)FI.getType()->getPrimitiveSizeInBits() <=
8293 OpI->getType()->getFPMantissaWidth())
8294 return ReplaceInstUsesWith(FI, OpI->getOperand(0));
8296 return commonCastTransforms(FI);
8299 Instruction *InstCombiner::visitUIToFP(CastInst &CI) {
8300 return commonCastTransforms(CI);
8303 Instruction *InstCombiner::visitSIToFP(CastInst &CI) {
8304 return commonCastTransforms(CI);
8307 Instruction *InstCombiner::visitPtrToInt(CastInst &CI) {
8308 return commonPointerCastTransforms(CI);
8311 Instruction *InstCombiner::visitIntToPtr(IntToPtrInst &CI) {
8312 if (Instruction *I = commonCastTransforms(CI))
8315 const Type *DestPointee = cast<PointerType>(CI.getType())->getElementType();
8316 if (!DestPointee->isSized()) return 0;
8318 // If this is inttoptr(add (ptrtoint x), cst), try to turn this into a GEP.
8321 if (match(CI.getOperand(0), m_Add(m_Cast<PtrToIntInst>(m_Value(X)),
8322 m_ConstantInt(Cst)))) {
8323 // If the source and destination operands have the same type, see if this
8324 // is a single-index GEP.
8325 if (X->getType() == CI.getType()) {
8326 // Get the size of the pointee type.
8327 uint64_t Size = TD->getABITypeSize(DestPointee);
8329 // Convert the constant to intptr type.
8330 APInt Offset = Cst->getValue();
8331 Offset.sextOrTrunc(TD->getPointerSizeInBits());
8333 // If Offset is evenly divisible by Size, we can do this xform.
8334 if (Size && !APIntOps::srem(Offset, APInt(Offset.getBitWidth(), Size))){
8335 Offset = APIntOps::sdiv(Offset, APInt(Offset.getBitWidth(), Size));
8336 return GetElementPtrInst::Create(X, ConstantInt::get(Offset));
8339 // TODO: Could handle other cases, e.g. where add is indexing into field of
8341 } else if (CI.getOperand(0)->hasOneUse() &&
8342 match(CI.getOperand(0), m_Add(m_Value(X), m_ConstantInt(Cst)))) {
8343 // Otherwise, if this is inttoptr(add x, cst), try to turn this into an
8344 // "inttoptr+GEP" instead of "add+intptr".
8346 // Get the size of the pointee type.
8347 uint64_t Size = TD->getABITypeSize(DestPointee);
8349 // Convert the constant to intptr type.
8350 APInt Offset = Cst->getValue();
8351 Offset.sextOrTrunc(TD->getPointerSizeInBits());
8353 // If Offset is evenly divisible by Size, we can do this xform.
8354 if (Size && !APIntOps::srem(Offset, APInt(Offset.getBitWidth(), Size))){
8355 Offset = APIntOps::sdiv(Offset, APInt(Offset.getBitWidth(), Size));
8357 Instruction *P = InsertNewInstBefore(new IntToPtrInst(X, CI.getType(),
8359 return GetElementPtrInst::Create(P, ConstantInt::get(Offset), "tmp");
8365 Instruction *InstCombiner::visitBitCast(BitCastInst &CI) {
8366 // If the operands are integer typed then apply the integer transforms,
8367 // otherwise just apply the common ones.
8368 Value *Src = CI.getOperand(0);
8369 const Type *SrcTy = Src->getType();
8370 const Type *DestTy = CI.getType();
8372 if (SrcTy->isInteger() && DestTy->isInteger()) {
8373 if (Instruction *Result = commonIntCastTransforms(CI))
8375 } else if (isa<PointerType>(SrcTy)) {
8376 if (Instruction *I = commonPointerCastTransforms(CI))
8379 if (Instruction *Result = commonCastTransforms(CI))
8384 // Get rid of casts from one type to the same type. These are useless and can
8385 // be replaced by the operand.
8386 if (DestTy == Src->getType())
8387 return ReplaceInstUsesWith(CI, Src);
8389 if (const PointerType *DstPTy = dyn_cast<PointerType>(DestTy)) {
8390 const PointerType *SrcPTy = cast<PointerType>(SrcTy);
8391 const Type *DstElTy = DstPTy->getElementType();
8392 const Type *SrcElTy = SrcPTy->getElementType();
8394 // If the address spaces don't match, don't eliminate the bitcast, which is
8395 // required for changing types.
8396 if (SrcPTy->getAddressSpace() != DstPTy->getAddressSpace())
8399 // If we are casting a malloc or alloca to a pointer to a type of the same
8400 // size, rewrite the allocation instruction to allocate the "right" type.
8401 if (AllocationInst *AI = dyn_cast<AllocationInst>(Src))
8402 if (Instruction *V = PromoteCastOfAllocation(CI, *AI))
8405 // If the source and destination are pointers, and this cast is equivalent
8406 // to a getelementptr X, 0, 0, 0... turn it into the appropriate gep.
8407 // This can enhance SROA and other transforms that want type-safe pointers.
8408 Constant *ZeroUInt = Constant::getNullValue(Type::Int32Ty);
8409 unsigned NumZeros = 0;
8410 while (SrcElTy != DstElTy &&
8411 isa<CompositeType>(SrcElTy) && !isa<PointerType>(SrcElTy) &&
8412 SrcElTy->getNumContainedTypes() /* not "{}" */) {
8413 SrcElTy = cast<CompositeType>(SrcElTy)->getTypeAtIndex(ZeroUInt);
8417 // If we found a path from the src to dest, create the getelementptr now.
8418 if (SrcElTy == DstElTy) {
8419 SmallVector<Value*, 8> Idxs(NumZeros+1, ZeroUInt);
8420 return GetElementPtrInst::Create(Src, Idxs.begin(), Idxs.end(), "",
8421 ((Instruction*) NULL));
8425 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(Src)) {
8426 if (SVI->hasOneUse()) {
8427 // Okay, we have (bitconvert (shuffle ..)). Check to see if this is
8428 // a bitconvert to a vector with the same # elts.
8429 if (isa<VectorType>(DestTy) &&
8430 cast<VectorType>(DestTy)->getNumElements() ==
8431 SVI->getType()->getNumElements() &&
8432 SVI->getType()->getNumElements() ==
8433 cast<VectorType>(SVI->getOperand(0)->getType())->getNumElements()) {
8435 // If either of the operands is a cast from CI.getType(), then
8436 // evaluating the shuffle in the casted destination's type will allow
8437 // us to eliminate at least one cast.
8438 if (((Tmp = dyn_cast<CastInst>(SVI->getOperand(0))) &&
8439 Tmp->getOperand(0)->getType() == DestTy) ||
8440 ((Tmp = dyn_cast<CastInst>(SVI->getOperand(1))) &&
8441 Tmp->getOperand(0)->getType() == DestTy)) {
8442 Value *LHS = InsertOperandCastBefore(Instruction::BitCast,
8443 SVI->getOperand(0), DestTy, &CI);
8444 Value *RHS = InsertOperandCastBefore(Instruction::BitCast,
8445 SVI->getOperand(1), DestTy, &CI);
8446 // Return a new shuffle vector. Use the same element ID's, as we
8447 // know the vector types match #elts.
8448 return new ShuffleVectorInst(LHS, RHS, SVI->getOperand(2));
8456 /// GetSelectFoldableOperands - We want to turn code that looks like this:
8458 /// %D = select %cond, %C, %A
8460 /// %C = select %cond, %B, 0
8463 /// Assuming that the specified instruction is an operand to the select, return
8464 /// a bitmask indicating which operands of this instruction are foldable if they
8465 /// equal the other incoming value of the select.
8467 static unsigned GetSelectFoldableOperands(Instruction *I) {
8468 switch (I->getOpcode()) {
8469 case Instruction::Add:
8470 case Instruction::Mul:
8471 case Instruction::And:
8472 case Instruction::Or:
8473 case Instruction::Xor:
8474 return 3; // Can fold through either operand.
8475 case Instruction::Sub: // Can only fold on the amount subtracted.
8476 case Instruction::Shl: // Can only fold on the shift amount.
8477 case Instruction::LShr:
8478 case Instruction::AShr:
8481 return 0; // Cannot fold
8485 /// GetSelectFoldableConstant - For the same transformation as the previous
8486 /// function, return the identity constant that goes into the select.
8487 static Constant *GetSelectFoldableConstant(Instruction *I) {
8488 switch (I->getOpcode()) {
8489 default: assert(0 && "This cannot happen!"); abort();
8490 case Instruction::Add:
8491 case Instruction::Sub:
8492 case Instruction::Or:
8493 case Instruction::Xor:
8494 case Instruction::Shl:
8495 case Instruction::LShr:
8496 case Instruction::AShr:
8497 return Constant::getNullValue(I->getType());
8498 case Instruction::And:
8499 return Constant::getAllOnesValue(I->getType());
8500 case Instruction::Mul:
8501 return ConstantInt::get(I->getType(), 1);
8505 /// FoldSelectOpOp - Here we have (select c, TI, FI), and we know that TI and FI
8506 /// have the same opcode and only one use each. Try to simplify this.
8507 Instruction *InstCombiner::FoldSelectOpOp(SelectInst &SI, Instruction *TI,
8509 if (TI->getNumOperands() == 1) {
8510 // If this is a non-volatile load or a cast from the same type,
8513 if (TI->getOperand(0)->getType() != FI->getOperand(0)->getType())
8516 return 0; // unknown unary op.
8519 // Fold this by inserting a select from the input values.
8520 SelectInst *NewSI = SelectInst::Create(SI.getCondition(), TI->getOperand(0),
8521 FI->getOperand(0), SI.getName()+".v");
8522 InsertNewInstBefore(NewSI, SI);
8523 return CastInst::Create(Instruction::CastOps(TI->getOpcode()), NewSI,
8527 // Only handle binary operators here.
8528 if (!isa<BinaryOperator>(TI))
8531 // Figure out if the operations have any operands in common.
8532 Value *MatchOp, *OtherOpT, *OtherOpF;
8534 if (TI->getOperand(0) == FI->getOperand(0)) {
8535 MatchOp = TI->getOperand(0);
8536 OtherOpT = TI->getOperand(1);
8537 OtherOpF = FI->getOperand(1);
8538 MatchIsOpZero = true;
8539 } else if (TI->getOperand(1) == FI->getOperand(1)) {
8540 MatchOp = TI->getOperand(1);
8541 OtherOpT = TI->getOperand(0);
8542 OtherOpF = FI->getOperand(0);
8543 MatchIsOpZero = false;
8544 } else if (!TI->isCommutative()) {
8546 } else if (TI->getOperand(0) == FI->getOperand(1)) {
8547 MatchOp = TI->getOperand(0);
8548 OtherOpT = TI->getOperand(1);
8549 OtherOpF = FI->getOperand(0);
8550 MatchIsOpZero = true;
8551 } else if (TI->getOperand(1) == FI->getOperand(0)) {
8552 MatchOp = TI->getOperand(1);
8553 OtherOpT = TI->getOperand(0);
8554 OtherOpF = FI->getOperand(1);
8555 MatchIsOpZero = true;
8560 // If we reach here, they do have operations in common.
8561 SelectInst *NewSI = SelectInst::Create(SI.getCondition(), OtherOpT,
8562 OtherOpF, SI.getName()+".v");
8563 InsertNewInstBefore(NewSI, SI);
8565 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TI)) {
8567 return BinaryOperator::Create(BO->getOpcode(), MatchOp, NewSI);
8569 return BinaryOperator::Create(BO->getOpcode(), NewSI, MatchOp);
8571 assert(0 && "Shouldn't get here");
8575 /// visitSelectInstWithICmp - Visit a SelectInst that has an
8576 /// ICmpInst as its first operand.
8578 Instruction *InstCombiner::visitSelectInstWithICmp(SelectInst &SI,
8580 bool Changed = false;
8581 ICmpInst::Predicate Pred = ICI->getPredicate();
8582 Value *CmpLHS = ICI->getOperand(0);
8583 Value *CmpRHS = ICI->getOperand(1);
8584 Value *TrueVal = SI.getTrueValue();
8585 Value *FalseVal = SI.getFalseValue();
8587 // Check cases where the comparison is with a constant that
8588 // can be adjusted to fit the min/max idiom. We may edit ICI in
8589 // place here, so make sure the select is the only user.
8590 if (ICI->hasOneUse())
8591 if (ConstantInt *CI = dyn_cast<ConstantInt>(CmpRHS)) {
8594 case ICmpInst::ICMP_ULT:
8595 case ICmpInst::ICMP_SLT: {
8596 // X < MIN ? T : F --> F
8597 if (CI->isMinValue(Pred == ICmpInst::ICMP_SLT))
8598 return ReplaceInstUsesWith(SI, FalseVal);
8599 // X < C ? X : C-1 --> X > C-1 ? C-1 : X
8600 Constant *AdjustedRHS = SubOne(CI);
8601 if ((CmpLHS == TrueVal && AdjustedRHS == FalseVal) ||
8602 (CmpLHS == FalseVal && AdjustedRHS == TrueVal)) {
8603 Pred = ICmpInst::getSwappedPredicate(Pred);
8604 CmpRHS = AdjustedRHS;
8605 std::swap(FalseVal, TrueVal);
8606 ICI->setPredicate(Pred);
8607 ICI->setOperand(1, CmpRHS);
8608 SI.setOperand(1, TrueVal);
8609 SI.setOperand(2, FalseVal);
8614 case ICmpInst::ICMP_UGT:
8615 case ICmpInst::ICMP_SGT: {
8616 // X > MAX ? T : F --> F
8617 if (CI->isMaxValue(Pred == ICmpInst::ICMP_SGT))
8618 return ReplaceInstUsesWith(SI, FalseVal);
8619 // X > C ? X : C+1 --> X < C+1 ? C+1 : X
8620 Constant *AdjustedRHS = AddOne(CI);
8621 if ((CmpLHS == TrueVal && AdjustedRHS == FalseVal) ||
8622 (CmpLHS == FalseVal && AdjustedRHS == TrueVal)) {
8623 Pred = ICmpInst::getSwappedPredicate(Pred);
8624 CmpRHS = AdjustedRHS;
8625 std::swap(FalseVal, TrueVal);
8626 ICI->setPredicate(Pred);
8627 ICI->setOperand(1, CmpRHS);
8628 SI.setOperand(1, TrueVal);
8629 SI.setOperand(2, FalseVal);
8636 // (x <s 0) ? -1 : 0 -> ashr x, 31 -> all ones if signed
8637 // (x >s -1) ? -1 : 0 -> ashr x, 31 -> all ones if not signed
8638 CmpInst::Predicate Pred = CmpInst::BAD_ICMP_PREDICATE;
8639 if (match(TrueVal, m_ConstantInt(-1)) &&
8640 match(FalseVal, m_ConstantInt(0)))
8641 Pred = ICI->getPredicate();
8642 else if (match(TrueVal, m_ConstantInt(0)) &&
8643 match(FalseVal, m_ConstantInt(-1)))
8644 Pred = CmpInst::getInversePredicate(ICI->getPredicate());
8646 if (Pred != CmpInst::BAD_ICMP_PREDICATE) {
8647 // If we are just checking for a icmp eq of a single bit and zext'ing it
8648 // to an integer, then shift the bit to the appropriate place and then
8649 // cast to integer to avoid the comparison.
8650 const APInt &Op1CV = CI->getValue();
8652 // sext (x <s 0) to i32 --> x>>s31 true if signbit set.
8653 // sext (x >s -1) to i32 --> (x>>s31)^-1 true if signbit clear.
8654 if ((Pred == ICmpInst::ICMP_SLT && Op1CV == 0) ||
8655 (Pred == ICmpInst::ICMP_SGT && Op1CV.isAllOnesValue())) {
8656 Value *In = ICI->getOperand(0);
8657 Value *Sh = ConstantInt::get(In->getType(),
8658 In->getType()->getPrimitiveSizeInBits()-1);
8659 In = InsertNewInstBefore(BinaryOperator::CreateAShr(In, Sh,
8660 In->getName()+".lobit"),
8662 if (In->getType() != SI.getType())
8663 In = CastInst::CreateIntegerCast(In, SI.getType(),
8664 true/*SExt*/, "tmp", ICI);
8666 if (Pred == ICmpInst::ICMP_SGT)
8667 In = InsertNewInstBefore(BinaryOperator::CreateNot(In,
8668 In->getName()+".not"), *ICI);
8670 return ReplaceInstUsesWith(SI, In);
8675 if (CmpLHS == TrueVal && CmpRHS == FalseVal) {
8676 // Transform (X == Y) ? X : Y -> Y
8677 if (Pred == ICmpInst::ICMP_EQ)
8678 return ReplaceInstUsesWith(SI, FalseVal);
8679 // Transform (X != Y) ? X : Y -> X
8680 if (Pred == ICmpInst::ICMP_NE)
8681 return ReplaceInstUsesWith(SI, TrueVal);
8682 /// NOTE: if we wanted to, this is where to detect integer MIN/MAX
8684 } else if (CmpLHS == FalseVal && CmpRHS == TrueVal) {
8685 // Transform (X == Y) ? Y : X -> X
8686 if (Pred == ICmpInst::ICMP_EQ)
8687 return ReplaceInstUsesWith(SI, FalseVal);
8688 // Transform (X != Y) ? Y : X -> Y
8689 if (Pred == ICmpInst::ICMP_NE)
8690 return ReplaceInstUsesWith(SI, TrueVal);
8691 /// NOTE: if we wanted to, this is where to detect integer MIN/MAX
8694 /// NOTE: if we wanted to, this is where to detect integer ABS
8696 return Changed ? &SI : 0;
8699 Instruction *InstCombiner::visitSelectInst(SelectInst &SI) {
8700 Value *CondVal = SI.getCondition();
8701 Value *TrueVal = SI.getTrueValue();
8702 Value *FalseVal = SI.getFalseValue();
8704 // select true, X, Y -> X
8705 // select false, X, Y -> Y
8706 if (ConstantInt *C = dyn_cast<ConstantInt>(CondVal))
8707 return ReplaceInstUsesWith(SI, C->getZExtValue() ? TrueVal : FalseVal);
8709 // select C, X, X -> X
8710 if (TrueVal == FalseVal)
8711 return ReplaceInstUsesWith(SI, TrueVal);
8713 if (isa<UndefValue>(TrueVal)) // select C, undef, X -> X
8714 return ReplaceInstUsesWith(SI, FalseVal);
8715 if (isa<UndefValue>(FalseVal)) // select C, X, undef -> X
8716 return ReplaceInstUsesWith(SI, TrueVal);
8717 if (isa<UndefValue>(CondVal)) { // select undef, X, Y -> X or Y
8718 if (isa<Constant>(TrueVal))
8719 return ReplaceInstUsesWith(SI, TrueVal);
8721 return ReplaceInstUsesWith(SI, FalseVal);
8724 if (SI.getType() == Type::Int1Ty) {
8725 if (ConstantInt *C = dyn_cast<ConstantInt>(TrueVal)) {
8726 if (C->getZExtValue()) {
8727 // Change: A = select B, true, C --> A = or B, C
8728 return BinaryOperator::CreateOr(CondVal, FalseVal);
8730 // Change: A = select B, false, C --> A = and !B, C
8732 InsertNewInstBefore(BinaryOperator::CreateNot(CondVal,
8733 "not."+CondVal->getName()), SI);
8734 return BinaryOperator::CreateAnd(NotCond, FalseVal);
8736 } else if (ConstantInt *C = dyn_cast<ConstantInt>(FalseVal)) {
8737 if (C->getZExtValue() == false) {
8738 // Change: A = select B, C, false --> A = and B, C
8739 return BinaryOperator::CreateAnd(CondVal, TrueVal);
8741 // Change: A = select B, C, true --> A = or !B, C
8743 InsertNewInstBefore(BinaryOperator::CreateNot(CondVal,
8744 "not."+CondVal->getName()), SI);
8745 return BinaryOperator::CreateOr(NotCond, TrueVal);
8749 // select a, b, a -> a&b
8750 // select a, a, b -> a|b
8751 if (CondVal == TrueVal)
8752 return BinaryOperator::CreateOr(CondVal, FalseVal);
8753 else if (CondVal == FalseVal)
8754 return BinaryOperator::CreateAnd(CondVal, TrueVal);
8757 // Selecting between two integer constants?
8758 if (ConstantInt *TrueValC = dyn_cast<ConstantInt>(TrueVal))
8759 if (ConstantInt *FalseValC = dyn_cast<ConstantInt>(FalseVal)) {
8760 // select C, 1, 0 -> zext C to int
8761 if (FalseValC->isZero() && TrueValC->getValue() == 1) {
8762 return CastInst::Create(Instruction::ZExt, CondVal, SI.getType());
8763 } else if (TrueValC->isZero() && FalseValC->getValue() == 1) {
8764 // select C, 0, 1 -> zext !C to int
8766 InsertNewInstBefore(BinaryOperator::CreateNot(CondVal,
8767 "not."+CondVal->getName()), SI);
8768 return CastInst::Create(Instruction::ZExt, NotCond, SI.getType());
8771 // FIXME: Turn select 0/-1 and -1/0 into sext from condition!
8773 if (ICmpInst *IC = dyn_cast<ICmpInst>(SI.getCondition())) {
8775 // (x <s 0) ? -1 : 0 -> ashr x, 31
8776 if (TrueValC->isAllOnesValue() && FalseValC->isZero())
8777 if (ConstantInt *CmpCst = dyn_cast<ConstantInt>(IC->getOperand(1))) {
8778 if (IC->getPredicate() == ICmpInst::ICMP_SLT && CmpCst->isZero()) {
8779 // The comparison constant and the result are not neccessarily the
8780 // same width. Make an all-ones value by inserting a AShr.
8781 Value *X = IC->getOperand(0);
8782 uint32_t Bits = X->getType()->getPrimitiveSizeInBits();
8783 Constant *ShAmt = ConstantInt::get(X->getType(), Bits-1);
8784 Instruction *SRA = BinaryOperator::Create(Instruction::AShr, X,
8786 InsertNewInstBefore(SRA, SI);
8788 // Finally, convert to the type of the select RHS. We figure out
8789 // if this requires a SExt, Trunc or BitCast based on the sizes.
8790 Instruction::CastOps opc = Instruction::BitCast;
8791 uint32_t SRASize = SRA->getType()->getPrimitiveSizeInBits();
8792 uint32_t SISize = SI.getType()->getPrimitiveSizeInBits();
8793 if (SRASize < SISize)
8794 opc = Instruction::SExt;
8795 else if (SRASize > SISize)
8796 opc = Instruction::Trunc;
8797 return CastInst::Create(opc, SRA, SI.getType());
8802 // If one of the constants is zero (we know they can't both be) and we
8803 // have an icmp instruction with zero, and we have an 'and' with the
8804 // non-constant value, eliminate this whole mess. This corresponds to
8805 // cases like this: ((X & 27) ? 27 : 0)
8806 if (TrueValC->isZero() || FalseValC->isZero())
8807 if (IC->isEquality() && isa<ConstantInt>(IC->getOperand(1)) &&
8808 cast<Constant>(IC->getOperand(1))->isNullValue())
8809 if (Instruction *ICA = dyn_cast<Instruction>(IC->getOperand(0)))
8810 if (ICA->getOpcode() == Instruction::And &&
8811 isa<ConstantInt>(ICA->getOperand(1)) &&
8812 (ICA->getOperand(1) == TrueValC ||
8813 ICA->getOperand(1) == FalseValC) &&
8814 isOneBitSet(cast<ConstantInt>(ICA->getOperand(1)))) {
8815 // Okay, now we know that everything is set up, we just don't
8816 // know whether we have a icmp_ne or icmp_eq and whether the
8817 // true or false val is the zero.
8818 bool ShouldNotVal = !TrueValC->isZero();
8819 ShouldNotVal ^= IC->getPredicate() == ICmpInst::ICMP_NE;
8822 V = InsertNewInstBefore(BinaryOperator::Create(
8823 Instruction::Xor, V, ICA->getOperand(1)), SI);
8824 return ReplaceInstUsesWith(SI, V);
8829 // See if we are selecting two values based on a comparison of the two values.
8830 if (FCmpInst *FCI = dyn_cast<FCmpInst>(CondVal)) {
8831 if (FCI->getOperand(0) == TrueVal && FCI->getOperand(1) == FalseVal) {
8832 // Transform (X == Y) ? X : Y -> Y
8833 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ) {
8834 // This is not safe in general for floating point:
8835 // consider X== -0, Y== +0.
8836 // It becomes safe if either operand is a nonzero constant.
8837 ConstantFP *CFPt, *CFPf;
8838 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) &&
8839 !CFPt->getValueAPF().isZero()) ||
8840 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) &&
8841 !CFPf->getValueAPF().isZero()))
8842 return ReplaceInstUsesWith(SI, FalseVal);
8844 // Transform (X != Y) ? X : Y -> X
8845 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
8846 return ReplaceInstUsesWith(SI, TrueVal);
8847 // NOTE: if we wanted to, this is where to detect MIN/MAX
8849 } else if (FCI->getOperand(0) == FalseVal && FCI->getOperand(1) == TrueVal){
8850 // Transform (X == Y) ? Y : X -> X
8851 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ) {
8852 // This is not safe in general for floating point:
8853 // consider X== -0, Y== +0.
8854 // It becomes safe if either operand is a nonzero constant.
8855 ConstantFP *CFPt, *CFPf;
8856 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) &&
8857 !CFPt->getValueAPF().isZero()) ||
8858 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) &&
8859 !CFPf->getValueAPF().isZero()))
8860 return ReplaceInstUsesWith(SI, FalseVal);
8862 // Transform (X != Y) ? Y : X -> Y
8863 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
8864 return ReplaceInstUsesWith(SI, TrueVal);
8865 // NOTE: if we wanted to, this is where to detect MIN/MAX
8867 // NOTE: if we wanted to, this is where to detect ABS
8870 // See if we are selecting two values based on a comparison of the two values.
8871 if (ICmpInst *ICI = dyn_cast<ICmpInst>(CondVal))
8872 if (Instruction *Result = visitSelectInstWithICmp(SI, ICI))
8875 if (Instruction *TI = dyn_cast<Instruction>(TrueVal))
8876 if (Instruction *FI = dyn_cast<Instruction>(FalseVal))
8877 if (TI->hasOneUse() && FI->hasOneUse()) {
8878 Instruction *AddOp = 0, *SubOp = 0;
8880 // Turn (select C, (op X, Y), (op X, Z)) -> (op X, (select C, Y, Z))
8881 if (TI->getOpcode() == FI->getOpcode())
8882 if (Instruction *IV = FoldSelectOpOp(SI, TI, FI))
8885 // Turn select C, (X+Y), (X-Y) --> (X+(select C, Y, (-Y))). This is
8886 // even legal for FP.
8887 if (TI->getOpcode() == Instruction::Sub &&
8888 FI->getOpcode() == Instruction::Add) {
8889 AddOp = FI; SubOp = TI;
8890 } else if (FI->getOpcode() == Instruction::Sub &&
8891 TI->getOpcode() == Instruction::Add) {
8892 AddOp = TI; SubOp = FI;
8896 Value *OtherAddOp = 0;
8897 if (SubOp->getOperand(0) == AddOp->getOperand(0)) {
8898 OtherAddOp = AddOp->getOperand(1);
8899 } else if (SubOp->getOperand(0) == AddOp->getOperand(1)) {
8900 OtherAddOp = AddOp->getOperand(0);
8904 // So at this point we know we have (Y -> OtherAddOp):
8905 // select C, (add X, Y), (sub X, Z)
8906 Value *NegVal; // Compute -Z
8907 if (Constant *C = dyn_cast<Constant>(SubOp->getOperand(1))) {
8908 NegVal = ConstantExpr::getNeg(C);
8910 NegVal = InsertNewInstBefore(
8911 BinaryOperator::CreateNeg(SubOp->getOperand(1), "tmp"), SI);
8914 Value *NewTrueOp = OtherAddOp;
8915 Value *NewFalseOp = NegVal;
8917 std::swap(NewTrueOp, NewFalseOp);
8918 Instruction *NewSel =
8919 SelectInst::Create(CondVal, NewTrueOp,
8920 NewFalseOp, SI.getName() + ".p");
8922 NewSel = InsertNewInstBefore(NewSel, SI);
8923 return BinaryOperator::CreateAdd(SubOp->getOperand(0), NewSel);
8928 // See if we can fold the select into one of our operands.
8929 if (SI.getType()->isInteger()) {
8930 // See the comment above GetSelectFoldableOperands for a description of the
8931 // transformation we are doing here.
8932 if (Instruction *TVI = dyn_cast<Instruction>(TrueVal))
8933 if (TVI->hasOneUse() && TVI->getNumOperands() == 2 &&
8934 !isa<Constant>(FalseVal))
8935 if (unsigned SFO = GetSelectFoldableOperands(TVI)) {
8936 unsigned OpToFold = 0;
8937 if ((SFO & 1) && FalseVal == TVI->getOperand(0)) {
8939 } else if ((SFO & 2) && FalseVal == TVI->getOperand(1)) {
8944 Constant *C = GetSelectFoldableConstant(TVI);
8945 Instruction *NewSel =
8946 SelectInst::Create(SI.getCondition(),
8947 TVI->getOperand(2-OpToFold), C);
8948 InsertNewInstBefore(NewSel, SI);
8949 NewSel->takeName(TVI);
8950 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TVI))
8951 return BinaryOperator::Create(BO->getOpcode(), FalseVal, NewSel);
8953 assert(0 && "Unknown instruction!!");
8958 if (Instruction *FVI = dyn_cast<Instruction>(FalseVal))
8959 if (FVI->hasOneUse() && FVI->getNumOperands() == 2 &&
8960 !isa<Constant>(TrueVal))
8961 if (unsigned SFO = GetSelectFoldableOperands(FVI)) {
8962 unsigned OpToFold = 0;
8963 if ((SFO & 1) && TrueVal == FVI->getOperand(0)) {
8965 } else if ((SFO & 2) && TrueVal == FVI->getOperand(1)) {
8970 Constant *C = GetSelectFoldableConstant(FVI);
8971 Instruction *NewSel =
8972 SelectInst::Create(SI.getCondition(), C,
8973 FVI->getOperand(2-OpToFold));
8974 InsertNewInstBefore(NewSel, SI);
8975 NewSel->takeName(FVI);
8976 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(FVI))
8977 return BinaryOperator::Create(BO->getOpcode(), TrueVal, NewSel);
8979 assert(0 && "Unknown instruction!!");
8984 if (BinaryOperator::isNot(CondVal)) {
8985 SI.setOperand(0, BinaryOperator::getNotArgument(CondVal));
8986 SI.setOperand(1, FalseVal);
8987 SI.setOperand(2, TrueVal);
8994 /// EnforceKnownAlignment - If the specified pointer points to an object that
8995 /// we control, modify the object's alignment to PrefAlign. This isn't
8996 /// often possible though. If alignment is important, a more reliable approach
8997 /// is to simply align all global variables and allocation instructions to
8998 /// their preferred alignment from the beginning.
9000 static unsigned EnforceKnownAlignment(Value *V,
9001 unsigned Align, unsigned PrefAlign) {
9003 User *U = dyn_cast<User>(V);
9004 if (!U) return Align;
9006 switch (getOpcode(U)) {
9008 case Instruction::BitCast:
9009 return EnforceKnownAlignment(U->getOperand(0), Align, PrefAlign);
9010 case Instruction::GetElementPtr: {
9011 // If all indexes are zero, it is just the alignment of the base pointer.
9012 bool AllZeroOperands = true;
9013 for (User::op_iterator i = U->op_begin() + 1, e = U->op_end(); i != e; ++i)
9014 if (!isa<Constant>(*i) ||
9015 !cast<Constant>(*i)->isNullValue()) {
9016 AllZeroOperands = false;
9020 if (AllZeroOperands) {
9021 // Treat this like a bitcast.
9022 return EnforceKnownAlignment(U->getOperand(0), Align, PrefAlign);
9028 if (GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
9029 // If there is a large requested alignment and we can, bump up the alignment
9031 if (!GV->isDeclaration()) {
9032 GV->setAlignment(PrefAlign);
9035 } else if (AllocationInst *AI = dyn_cast<AllocationInst>(V)) {
9036 // If there is a requested alignment and if this is an alloca, round up. We
9037 // don't do this for malloc, because some systems can't respect the request.
9038 if (isa<AllocaInst>(AI)) {
9039 AI->setAlignment(PrefAlign);
9047 /// GetOrEnforceKnownAlignment - If the specified pointer has an alignment that
9048 /// we can determine, return it, otherwise return 0. If PrefAlign is specified,
9049 /// and it is more than the alignment of the ultimate object, see if we can
9050 /// increase the alignment of the ultimate object, making this check succeed.
9051 unsigned InstCombiner::GetOrEnforceKnownAlignment(Value *V,
9052 unsigned PrefAlign) {
9053 unsigned BitWidth = TD ? TD->getTypeSizeInBits(V->getType()) :
9054 sizeof(PrefAlign) * CHAR_BIT;
9055 APInt Mask = APInt::getAllOnesValue(BitWidth);
9056 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
9057 ComputeMaskedBits(V, Mask, KnownZero, KnownOne);
9058 unsigned TrailZ = KnownZero.countTrailingOnes();
9059 unsigned Align = 1u << std::min(BitWidth - 1, TrailZ);
9061 if (PrefAlign > Align)
9062 Align = EnforceKnownAlignment(V, Align, PrefAlign);
9064 // We don't need to make any adjustment.
9068 Instruction *InstCombiner::SimplifyMemTransfer(MemIntrinsic *MI) {
9069 unsigned DstAlign = GetOrEnforceKnownAlignment(MI->getOperand(1));
9070 unsigned SrcAlign = GetOrEnforceKnownAlignment(MI->getOperand(2));
9071 unsigned MinAlign = std::min(DstAlign, SrcAlign);
9072 unsigned CopyAlign = MI->getAlignment()->getZExtValue();
9074 if (CopyAlign < MinAlign) {
9075 MI->setAlignment(ConstantInt::get(Type::Int32Ty, MinAlign));
9079 // If MemCpyInst length is 1/2/4/8 bytes then replace memcpy with
9081 ConstantInt *MemOpLength = dyn_cast<ConstantInt>(MI->getOperand(3));
9082 if (MemOpLength == 0) return 0;
9084 // Source and destination pointer types are always "i8*" for intrinsic. See
9085 // if the size is something we can handle with a single primitive load/store.
9086 // A single load+store correctly handles overlapping memory in the memmove
9088 unsigned Size = MemOpLength->getZExtValue();
9089 if (Size == 0) return MI; // Delete this mem transfer.
9091 if (Size > 8 || (Size&(Size-1)))
9092 return 0; // If not 1/2/4/8 bytes, exit.
9094 // Use an integer load+store unless we can find something better.
9095 Type *NewPtrTy = PointerType::getUnqual(IntegerType::get(Size<<3));
9097 // Memcpy forces the use of i8* for the source and destination. That means
9098 // that if you're using memcpy to move one double around, you'll get a cast
9099 // from double* to i8*. We'd much rather use a double load+store rather than
9100 // an i64 load+store, here because this improves the odds that the source or
9101 // dest address will be promotable. See if we can find a better type than the
9102 // integer datatype.
9103 if (Value *Op = getBitCastOperand(MI->getOperand(1))) {
9104 const Type *SrcETy = cast<PointerType>(Op->getType())->getElementType();
9105 if (SrcETy->isSized() && TD->getTypeStoreSize(SrcETy) == Size) {
9106 // The SrcETy might be something like {{{double}}} or [1 x double]. Rip
9107 // down through these levels if so.
9108 while (!SrcETy->isSingleValueType()) {
9109 if (const StructType *STy = dyn_cast<StructType>(SrcETy)) {
9110 if (STy->getNumElements() == 1)
9111 SrcETy = STy->getElementType(0);
9114 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(SrcETy)) {
9115 if (ATy->getNumElements() == 1)
9116 SrcETy = ATy->getElementType();
9123 if (SrcETy->isSingleValueType())
9124 NewPtrTy = PointerType::getUnqual(SrcETy);
9129 // If the memcpy/memmove provides better alignment info than we can
9131 SrcAlign = std::max(SrcAlign, CopyAlign);
9132 DstAlign = std::max(DstAlign, CopyAlign);
9134 Value *Src = InsertBitCastBefore(MI->getOperand(2), NewPtrTy, *MI);
9135 Value *Dest = InsertBitCastBefore(MI->getOperand(1), NewPtrTy, *MI);
9136 Instruction *L = new LoadInst(Src, "tmp", false, SrcAlign);
9137 InsertNewInstBefore(L, *MI);
9138 InsertNewInstBefore(new StoreInst(L, Dest, false, DstAlign), *MI);
9140 // Set the size of the copy to 0, it will be deleted on the next iteration.
9141 MI->setOperand(3, Constant::getNullValue(MemOpLength->getType()));
9145 Instruction *InstCombiner::SimplifyMemSet(MemSetInst *MI) {
9146 unsigned Alignment = GetOrEnforceKnownAlignment(MI->getDest());
9147 if (MI->getAlignment()->getZExtValue() < Alignment) {
9148 MI->setAlignment(ConstantInt::get(Type::Int32Ty, Alignment));
9152 // Extract the length and alignment and fill if they are constant.
9153 ConstantInt *LenC = dyn_cast<ConstantInt>(MI->getLength());
9154 ConstantInt *FillC = dyn_cast<ConstantInt>(MI->getValue());
9155 if (!LenC || !FillC || FillC->getType() != Type::Int8Ty)
9157 uint64_t Len = LenC->getZExtValue();
9158 Alignment = MI->getAlignment()->getZExtValue();
9160 // If the length is zero, this is a no-op
9161 if (Len == 0) return MI; // memset(d,c,0,a) -> noop
9163 // memset(s,c,n) -> store s, c (for n=1,2,4,8)
9164 if (Len <= 8 && isPowerOf2_32((uint32_t)Len)) {
9165 const Type *ITy = IntegerType::get(Len*8); // n=1 -> i8.
9167 Value *Dest = MI->getDest();
9168 Dest = InsertBitCastBefore(Dest, PointerType::getUnqual(ITy), *MI);
9170 // Alignment 0 is identity for alignment 1 for memset, but not store.
9171 if (Alignment == 0) Alignment = 1;
9173 // Extract the fill value and store.
9174 uint64_t Fill = FillC->getZExtValue()*0x0101010101010101ULL;
9175 InsertNewInstBefore(new StoreInst(ConstantInt::get(ITy, Fill), Dest, false,
9178 // Set the size of the copy to 0, it will be deleted on the next iteration.
9179 MI->setLength(Constant::getNullValue(LenC->getType()));
9187 /// visitCallInst - CallInst simplification. This mostly only handles folding
9188 /// of intrinsic instructions. For normal calls, it allows visitCallSite to do
9189 /// the heavy lifting.
9191 Instruction *InstCombiner::visitCallInst(CallInst &CI) {
9192 IntrinsicInst *II = dyn_cast<IntrinsicInst>(&CI);
9193 if (!II) return visitCallSite(&CI);
9195 // Intrinsics cannot occur in an invoke, so handle them here instead of in
9197 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(II)) {
9198 bool Changed = false;
9200 // memmove/cpy/set of zero bytes is a noop.
9201 if (Constant *NumBytes = dyn_cast<Constant>(MI->getLength())) {
9202 if (NumBytes->isNullValue()) return EraseInstFromFunction(CI);
9204 if (ConstantInt *CI = dyn_cast<ConstantInt>(NumBytes))
9205 if (CI->getZExtValue() == 1) {
9206 // Replace the instruction with just byte operations. We would
9207 // transform other cases to loads/stores, but we don't know if
9208 // alignment is sufficient.
9212 // If we have a memmove and the source operation is a constant global,
9213 // then the source and dest pointers can't alias, so we can change this
9214 // into a call to memcpy.
9215 if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(MI)) {
9216 if (GlobalVariable *GVSrc = dyn_cast<GlobalVariable>(MMI->getSource()))
9217 if (GVSrc->isConstant()) {
9218 Module *M = CI.getParent()->getParent()->getParent();
9219 Intrinsic::ID MemCpyID = Intrinsic::memcpy;
9221 Tys[0] = CI.getOperand(3)->getType();
9223 Intrinsic::getDeclaration(M, MemCpyID, Tys, 1));
9227 // memmove(x,x,size) -> noop.
9228 if (MMI->getSource() == MMI->getDest())
9229 return EraseInstFromFunction(CI);
9232 // If we can determine a pointer alignment that is bigger than currently
9233 // set, update the alignment.
9234 if (isa<MemCpyInst>(MI) || isa<MemMoveInst>(MI)) {
9235 if (Instruction *I = SimplifyMemTransfer(MI))
9237 } else if (MemSetInst *MSI = dyn_cast<MemSetInst>(MI)) {
9238 if (Instruction *I = SimplifyMemSet(MSI))
9242 if (Changed) return II;
9245 switch (II->getIntrinsicID()) {
9247 case Intrinsic::bswap:
9248 // bswap(bswap(x)) -> x
9249 if (IntrinsicInst *Operand = dyn_cast<IntrinsicInst>(II->getOperand(1)))
9250 if (Operand->getIntrinsicID() == Intrinsic::bswap)
9251 return ReplaceInstUsesWith(CI, Operand->getOperand(1));
9253 case Intrinsic::ppc_altivec_lvx:
9254 case Intrinsic::ppc_altivec_lvxl:
9255 case Intrinsic::x86_sse_loadu_ps:
9256 case Intrinsic::x86_sse2_loadu_pd:
9257 case Intrinsic::x86_sse2_loadu_dq:
9258 // Turn PPC lvx -> load if the pointer is known aligned.
9259 // Turn X86 loadups -> load if the pointer is known aligned.
9260 if (GetOrEnforceKnownAlignment(II->getOperand(1), 16) >= 16) {
9261 Value *Ptr = InsertBitCastBefore(II->getOperand(1),
9262 PointerType::getUnqual(II->getType()),
9264 return new LoadInst(Ptr);
9267 case Intrinsic::ppc_altivec_stvx:
9268 case Intrinsic::ppc_altivec_stvxl:
9269 // Turn stvx -> store if the pointer is known aligned.
9270 if (GetOrEnforceKnownAlignment(II->getOperand(2), 16) >= 16) {
9271 const Type *OpPtrTy =
9272 PointerType::getUnqual(II->getOperand(1)->getType());
9273 Value *Ptr = InsertBitCastBefore(II->getOperand(2), OpPtrTy, CI);
9274 return new StoreInst(II->getOperand(1), Ptr);
9277 case Intrinsic::x86_sse_storeu_ps:
9278 case Intrinsic::x86_sse2_storeu_pd:
9279 case Intrinsic::x86_sse2_storeu_dq:
9280 // Turn X86 storeu -> store if the pointer is known aligned.
9281 if (GetOrEnforceKnownAlignment(II->getOperand(1), 16) >= 16) {
9282 const Type *OpPtrTy =
9283 PointerType::getUnqual(II->getOperand(2)->getType());
9284 Value *Ptr = InsertBitCastBefore(II->getOperand(1), OpPtrTy, CI);
9285 return new StoreInst(II->getOperand(2), Ptr);
9289 case Intrinsic::x86_sse_cvttss2si: {
9290 // These intrinsics only demands the 0th element of its input vector. If
9291 // we can simplify the input based on that, do so now.
9293 if (Value *V = SimplifyDemandedVectorElts(II->getOperand(1), 1,
9295 II->setOperand(1, V);
9301 case Intrinsic::ppc_altivec_vperm:
9302 // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant.
9303 if (ConstantVector *Mask = dyn_cast<ConstantVector>(II->getOperand(3))) {
9304 assert(Mask->getNumOperands() == 16 && "Bad type for intrinsic!");
9306 // Check that all of the elements are integer constants or undefs.
9307 bool AllEltsOk = true;
9308 for (unsigned i = 0; i != 16; ++i) {
9309 if (!isa<ConstantInt>(Mask->getOperand(i)) &&
9310 !isa<UndefValue>(Mask->getOperand(i))) {
9317 // Cast the input vectors to byte vectors.
9318 Value *Op0 =InsertBitCastBefore(II->getOperand(1),Mask->getType(),CI);
9319 Value *Op1 =InsertBitCastBefore(II->getOperand(2),Mask->getType(),CI);
9320 Value *Result = UndefValue::get(Op0->getType());
9322 // Only extract each element once.
9323 Value *ExtractedElts[32];
9324 memset(ExtractedElts, 0, sizeof(ExtractedElts));
9326 for (unsigned i = 0; i != 16; ++i) {
9327 if (isa<UndefValue>(Mask->getOperand(i)))
9329 unsigned Idx=cast<ConstantInt>(Mask->getOperand(i))->getZExtValue();
9330 Idx &= 31; // Match the hardware behavior.
9332 if (ExtractedElts[Idx] == 0) {
9334 new ExtractElementInst(Idx < 16 ? Op0 : Op1, Idx&15, "tmp");
9335 InsertNewInstBefore(Elt, CI);
9336 ExtractedElts[Idx] = Elt;
9339 // Insert this value into the result vector.
9340 Result = InsertElementInst::Create(Result, ExtractedElts[Idx],
9342 InsertNewInstBefore(cast<Instruction>(Result), CI);
9344 return CastInst::Create(Instruction::BitCast, Result, CI.getType());
9349 case Intrinsic::stackrestore: {
9350 // If the save is right next to the restore, remove the restore. This can
9351 // happen when variable allocas are DCE'd.
9352 if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getOperand(1))) {
9353 if (SS->getIntrinsicID() == Intrinsic::stacksave) {
9354 BasicBlock::iterator BI = SS;
9356 return EraseInstFromFunction(CI);
9360 // Scan down this block to see if there is another stack restore in the
9361 // same block without an intervening call/alloca.
9362 BasicBlock::iterator BI = II;
9363 TerminatorInst *TI = II->getParent()->getTerminator();
9364 bool CannotRemove = false;
9365 for (++BI; &*BI != TI; ++BI) {
9366 if (isa<AllocaInst>(BI)) {
9367 CannotRemove = true;
9370 if (CallInst *BCI = dyn_cast<CallInst>(BI)) {
9371 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(BCI)) {
9372 // If there is a stackrestore below this one, remove this one.
9373 if (II->getIntrinsicID() == Intrinsic::stackrestore)
9374 return EraseInstFromFunction(CI);
9375 // Otherwise, ignore the intrinsic.
9377 // If we found a non-intrinsic call, we can't remove the stack
9379 CannotRemove = true;
9385 // If the stack restore is in a return/unwind block and if there are no
9386 // allocas or calls between the restore and the return, nuke the restore.
9387 if (!CannotRemove && (isa<ReturnInst>(TI) || isa<UnwindInst>(TI)))
9388 return EraseInstFromFunction(CI);
9393 return visitCallSite(II);
9396 // InvokeInst simplification
9398 Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) {
9399 return visitCallSite(&II);
9402 /// isSafeToEliminateVarargsCast - If this cast does not affect the value
9403 /// passed through the varargs area, we can eliminate the use of the cast.
9404 static bool isSafeToEliminateVarargsCast(const CallSite CS,
9405 const CastInst * const CI,
9406 const TargetData * const TD,
9408 if (!CI->isLosslessCast())
9411 // The size of ByVal arguments is derived from the type, so we
9412 // can't change to a type with a different size. If the size were
9413 // passed explicitly we could avoid this check.
9414 if (!CS.paramHasAttr(ix, Attribute::ByVal))
9418 cast<PointerType>(CI->getOperand(0)->getType())->getElementType();
9419 const Type* DstTy = cast<PointerType>(CI->getType())->getElementType();
9420 if (!SrcTy->isSized() || !DstTy->isSized())
9422 if (TD->getABITypeSize(SrcTy) != TD->getABITypeSize(DstTy))
9427 // visitCallSite - Improvements for call and invoke instructions.
9429 Instruction *InstCombiner::visitCallSite(CallSite CS) {
9430 bool Changed = false;
9432 // If the callee is a constexpr cast of a function, attempt to move the cast
9433 // to the arguments of the call/invoke.
9434 if (transformConstExprCastCall(CS)) return 0;
9436 Value *Callee = CS.getCalledValue();
9438 if (Function *CalleeF = dyn_cast<Function>(Callee))
9439 if (CalleeF->getCallingConv() != CS.getCallingConv()) {
9440 Instruction *OldCall = CS.getInstruction();
9441 // If the call and callee calling conventions don't match, this call must
9442 // be unreachable, as the call is undefined.
9443 new StoreInst(ConstantInt::getTrue(),
9444 UndefValue::get(PointerType::getUnqual(Type::Int1Ty)),
9446 if (!OldCall->use_empty())
9447 OldCall->replaceAllUsesWith(UndefValue::get(OldCall->getType()));
9448 if (isa<CallInst>(OldCall)) // Not worth removing an invoke here.
9449 return EraseInstFromFunction(*OldCall);
9453 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) {
9454 // This instruction is not reachable, just remove it. We insert a store to
9455 // undef so that we know that this code is not reachable, despite the fact
9456 // that we can't modify the CFG here.
9457 new StoreInst(ConstantInt::getTrue(),
9458 UndefValue::get(PointerType::getUnqual(Type::Int1Ty)),
9459 CS.getInstruction());
9461 if (!CS.getInstruction()->use_empty())
9462 CS.getInstruction()->
9463 replaceAllUsesWith(UndefValue::get(CS.getInstruction()->getType()));
9465 if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) {
9466 // Don't break the CFG, insert a dummy cond branch.
9467 BranchInst::Create(II->getNormalDest(), II->getUnwindDest(),
9468 ConstantInt::getTrue(), II);
9470 return EraseInstFromFunction(*CS.getInstruction());
9473 if (BitCastInst *BC = dyn_cast<BitCastInst>(Callee))
9474 if (IntrinsicInst *In = dyn_cast<IntrinsicInst>(BC->getOperand(0)))
9475 if (In->getIntrinsicID() == Intrinsic::init_trampoline)
9476 return transformCallThroughTrampoline(CS);
9478 const PointerType *PTy = cast<PointerType>(Callee->getType());
9479 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
9480 if (FTy->isVarArg()) {
9481 int ix = FTy->getNumParams() + (isa<InvokeInst>(Callee) ? 3 : 1);
9482 // See if we can optimize any arguments passed through the varargs area of
9484 for (CallSite::arg_iterator I = CS.arg_begin()+FTy->getNumParams(),
9485 E = CS.arg_end(); I != E; ++I, ++ix) {
9486 CastInst *CI = dyn_cast<CastInst>(*I);
9487 if (CI && isSafeToEliminateVarargsCast(CS, CI, TD, ix)) {
9488 *I = CI->getOperand(0);
9494 if (isa<InlineAsm>(Callee) && !CS.doesNotThrow()) {
9495 // Inline asm calls cannot throw - mark them 'nounwind'.
9496 CS.setDoesNotThrow();
9500 return Changed ? CS.getInstruction() : 0;
9503 // transformConstExprCastCall - If the callee is a constexpr cast of a function,
9504 // attempt to move the cast to the arguments of the call/invoke.
9506 bool InstCombiner::transformConstExprCastCall(CallSite CS) {
9507 if (!isa<ConstantExpr>(CS.getCalledValue())) return false;
9508 ConstantExpr *CE = cast<ConstantExpr>(CS.getCalledValue());
9509 if (CE->getOpcode() != Instruction::BitCast ||
9510 !isa<Function>(CE->getOperand(0)))
9512 Function *Callee = cast<Function>(CE->getOperand(0));
9513 Instruction *Caller = CS.getInstruction();
9514 const AttrListPtr &CallerPAL = CS.getAttributes();
9516 // Okay, this is a cast from a function to a different type. Unless doing so
9517 // would cause a type conversion of one of our arguments, change this call to
9518 // be a direct call with arguments casted to the appropriate types.
9520 const FunctionType *FT = Callee->getFunctionType();
9521 const Type *OldRetTy = Caller->getType();
9522 const Type *NewRetTy = FT->getReturnType();
9524 if (isa<StructType>(NewRetTy))
9525 return false; // TODO: Handle multiple return values.
9527 // Check to see if we are changing the return type...
9528 if (OldRetTy != NewRetTy) {
9529 if (Callee->isDeclaration() &&
9530 // Conversion is ok if changing from one pointer type to another or from
9531 // a pointer to an integer of the same size.
9532 !((isa<PointerType>(OldRetTy) || OldRetTy == TD->getIntPtrType()) &&
9533 (isa<PointerType>(NewRetTy) || NewRetTy == TD->getIntPtrType())))
9534 return false; // Cannot transform this return value.
9536 if (!Caller->use_empty() &&
9537 // void -> non-void is handled specially
9538 NewRetTy != Type::VoidTy && !CastInst::isCastable(NewRetTy, OldRetTy))
9539 return false; // Cannot transform this return value.
9541 if (!CallerPAL.isEmpty() && !Caller->use_empty()) {
9542 Attributes RAttrs = CallerPAL.getRetAttributes();
9543 if (RAttrs & Attribute::typeIncompatible(NewRetTy))
9544 return false; // Attribute not compatible with transformed value.
9547 // If the callsite is an invoke instruction, and the return value is used by
9548 // a PHI node in a successor, we cannot change the return type of the call
9549 // because there is no place to put the cast instruction (without breaking
9550 // the critical edge). Bail out in this case.
9551 if (!Caller->use_empty())
9552 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller))
9553 for (Value::use_iterator UI = II->use_begin(), E = II->use_end();
9555 if (PHINode *PN = dyn_cast<PHINode>(*UI))
9556 if (PN->getParent() == II->getNormalDest() ||
9557 PN->getParent() == II->getUnwindDest())
9561 unsigned NumActualArgs = unsigned(CS.arg_end()-CS.arg_begin());
9562 unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs);
9564 CallSite::arg_iterator AI = CS.arg_begin();
9565 for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) {
9566 const Type *ParamTy = FT->getParamType(i);
9567 const Type *ActTy = (*AI)->getType();
9569 if (!CastInst::isCastable(ActTy, ParamTy))
9570 return false; // Cannot transform this parameter value.
9572 if (CallerPAL.getParamAttributes(i + 1)
9573 & Attribute::typeIncompatible(ParamTy))
9574 return false; // Attribute not compatible with transformed value.
9576 // Converting from one pointer type to another or between a pointer and an
9577 // integer of the same size is safe even if we do not have a body.
9578 bool isConvertible = ActTy == ParamTy ||
9579 ((isa<PointerType>(ParamTy) || ParamTy == TD->getIntPtrType()) &&
9580 (isa<PointerType>(ActTy) || ActTy == TD->getIntPtrType()));
9581 if (Callee->isDeclaration() && !isConvertible) return false;
9584 if (FT->getNumParams() < NumActualArgs && !FT->isVarArg() &&
9585 Callee->isDeclaration())
9586 return false; // Do not delete arguments unless we have a function body.
9588 if (FT->getNumParams() < NumActualArgs && FT->isVarArg() &&
9589 !CallerPAL.isEmpty())
9590 // In this case we have more arguments than the new function type, but we
9591 // won't be dropping them. Check that these extra arguments have attributes
9592 // that are compatible with being a vararg call argument.
9593 for (unsigned i = CallerPAL.getNumSlots(); i; --i) {
9594 if (CallerPAL.getSlot(i - 1).Index <= FT->getNumParams())
9596 Attributes PAttrs = CallerPAL.getSlot(i - 1).Attrs;
9597 if (PAttrs & Attribute::VarArgsIncompatible)
9601 // Okay, we decided that this is a safe thing to do: go ahead and start
9602 // inserting cast instructions as necessary...
9603 std::vector<Value*> Args;
9604 Args.reserve(NumActualArgs);
9605 SmallVector<AttributeWithIndex, 8> attrVec;
9606 attrVec.reserve(NumCommonArgs);
9608 // Get any return attributes.
9609 Attributes RAttrs = CallerPAL.getRetAttributes();
9611 // If the return value is not being used, the type may not be compatible
9612 // with the existing attributes. Wipe out any problematic attributes.
9613 RAttrs &= ~Attribute::typeIncompatible(NewRetTy);
9615 // Add the new return attributes.
9617 attrVec.push_back(AttributeWithIndex::get(0, RAttrs));
9619 AI = CS.arg_begin();
9620 for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) {
9621 const Type *ParamTy = FT->getParamType(i);
9622 if ((*AI)->getType() == ParamTy) {
9623 Args.push_back(*AI);
9625 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI,
9626 false, ParamTy, false);
9627 CastInst *NewCast = CastInst::Create(opcode, *AI, ParamTy, "tmp");
9628 Args.push_back(InsertNewInstBefore(NewCast, *Caller));
9631 // Add any parameter attributes.
9632 if (Attributes PAttrs = CallerPAL.getParamAttributes(i + 1))
9633 attrVec.push_back(AttributeWithIndex::get(i + 1, PAttrs));
9636 // If the function takes more arguments than the call was taking, add them
9638 for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i)
9639 Args.push_back(Constant::getNullValue(FT->getParamType(i)));
9641 // If we are removing arguments to the function, emit an obnoxious warning...
9642 if (FT->getNumParams() < NumActualArgs) {
9643 if (!FT->isVarArg()) {
9644 cerr << "WARNING: While resolving call to function '"
9645 << Callee->getName() << "' arguments were dropped!\n";
9647 // Add all of the arguments in their promoted form to the arg list...
9648 for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) {
9649 const Type *PTy = getPromotedType((*AI)->getType());
9650 if (PTy != (*AI)->getType()) {
9651 // Must promote to pass through va_arg area!
9652 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI, false,
9654 Instruction *Cast = CastInst::Create(opcode, *AI, PTy, "tmp");
9655 InsertNewInstBefore(Cast, *Caller);
9656 Args.push_back(Cast);
9658 Args.push_back(*AI);
9661 // Add any parameter attributes.
9662 if (Attributes PAttrs = CallerPAL.getParamAttributes(i + 1))
9663 attrVec.push_back(AttributeWithIndex::get(i + 1, PAttrs));
9668 if (Attributes FnAttrs = CallerPAL.getFnAttributes())
9669 attrVec.push_back(AttributeWithIndex::get(~0, FnAttrs));
9671 if (NewRetTy == Type::VoidTy)
9672 Caller->setName(""); // Void type should not have a name.
9674 const AttrListPtr &NewCallerPAL = AttrListPtr::get(attrVec.begin(),attrVec.end());
9677 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
9678 NC = InvokeInst::Create(Callee, II->getNormalDest(), II->getUnwindDest(),
9679 Args.begin(), Args.end(),
9680 Caller->getName(), Caller);
9681 cast<InvokeInst>(NC)->setCallingConv(II->getCallingConv());
9682 cast<InvokeInst>(NC)->setAttributes(NewCallerPAL);
9684 NC = CallInst::Create(Callee, Args.begin(), Args.end(),
9685 Caller->getName(), Caller);
9686 CallInst *CI = cast<CallInst>(Caller);
9687 if (CI->isTailCall())
9688 cast<CallInst>(NC)->setTailCall();
9689 cast<CallInst>(NC)->setCallingConv(CI->getCallingConv());
9690 cast<CallInst>(NC)->setAttributes(NewCallerPAL);
9693 // Insert a cast of the return type as necessary.
9695 if (OldRetTy != NV->getType() && !Caller->use_empty()) {
9696 if (NV->getType() != Type::VoidTy) {
9697 Instruction::CastOps opcode = CastInst::getCastOpcode(NC, false,
9699 NV = NC = CastInst::Create(opcode, NC, OldRetTy, "tmp");
9701 // If this is an invoke instruction, we should insert it after the first
9702 // non-phi, instruction in the normal successor block.
9703 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
9704 BasicBlock::iterator I = II->getNormalDest()->getFirstNonPHI();
9705 InsertNewInstBefore(NC, *I);
9707 // Otherwise, it's a call, just insert cast right after the call instr
9708 InsertNewInstBefore(NC, *Caller);
9710 AddUsersToWorkList(*Caller);
9712 NV = UndefValue::get(Caller->getType());
9716 if (Caller->getType() != Type::VoidTy && !Caller->use_empty())
9717 Caller->replaceAllUsesWith(NV);
9718 Caller->eraseFromParent();
9719 RemoveFromWorkList(Caller);
9723 // transformCallThroughTrampoline - Turn a call to a function created by the
9724 // init_trampoline intrinsic into a direct call to the underlying function.
9726 Instruction *InstCombiner::transformCallThroughTrampoline(CallSite CS) {
9727 Value *Callee = CS.getCalledValue();
9728 const PointerType *PTy = cast<PointerType>(Callee->getType());
9729 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
9730 const AttrListPtr &Attrs = CS.getAttributes();
9732 // If the call already has the 'nest' attribute somewhere then give up -
9733 // otherwise 'nest' would occur twice after splicing in the chain.
9734 if (Attrs.hasAttrSomewhere(Attribute::Nest))
9737 IntrinsicInst *Tramp =
9738 cast<IntrinsicInst>(cast<BitCastInst>(Callee)->getOperand(0));
9740 Function *NestF = cast<Function>(Tramp->getOperand(2)->stripPointerCasts());
9741 const PointerType *NestFPTy = cast<PointerType>(NestF->getType());
9742 const FunctionType *NestFTy = cast<FunctionType>(NestFPTy->getElementType());
9744 const AttrListPtr &NestAttrs = NestF->getAttributes();
9745 if (!NestAttrs.isEmpty()) {
9746 unsigned NestIdx = 1;
9747 const Type *NestTy = 0;
9748 Attributes NestAttr = Attribute::None;
9750 // Look for a parameter marked with the 'nest' attribute.
9751 for (FunctionType::param_iterator I = NestFTy->param_begin(),
9752 E = NestFTy->param_end(); I != E; ++NestIdx, ++I)
9753 if (NestAttrs.paramHasAttr(NestIdx, Attribute::Nest)) {
9754 // Record the parameter type and any other attributes.
9756 NestAttr = NestAttrs.getParamAttributes(NestIdx);
9761 Instruction *Caller = CS.getInstruction();
9762 std::vector<Value*> NewArgs;
9763 NewArgs.reserve(unsigned(CS.arg_end()-CS.arg_begin())+1);
9765 SmallVector<AttributeWithIndex, 8> NewAttrs;
9766 NewAttrs.reserve(Attrs.getNumSlots() + 1);
9768 // Insert the nest argument into the call argument list, which may
9769 // mean appending it. Likewise for attributes.
9771 // Add any result attributes.
9772 if (Attributes Attr = Attrs.getRetAttributes())
9773 NewAttrs.push_back(AttributeWithIndex::get(0, Attr));
9777 CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
9779 if (Idx == NestIdx) {
9780 // Add the chain argument and attributes.
9781 Value *NestVal = Tramp->getOperand(3);
9782 if (NestVal->getType() != NestTy)
9783 NestVal = new BitCastInst(NestVal, NestTy, "nest", Caller);
9784 NewArgs.push_back(NestVal);
9785 NewAttrs.push_back(AttributeWithIndex::get(NestIdx, NestAttr));
9791 // Add the original argument and attributes.
9792 NewArgs.push_back(*I);
9793 if (Attributes Attr = Attrs.getParamAttributes(Idx))
9795 (AttributeWithIndex::get(Idx + (Idx >= NestIdx), Attr));
9801 // Add any function attributes.
9802 if (Attributes Attr = Attrs.getFnAttributes())
9803 NewAttrs.push_back(AttributeWithIndex::get(~0, Attr));
9805 // The trampoline may have been bitcast to a bogus type (FTy).
9806 // Handle this by synthesizing a new function type, equal to FTy
9807 // with the chain parameter inserted.
9809 std::vector<const Type*> NewTypes;
9810 NewTypes.reserve(FTy->getNumParams()+1);
9812 // Insert the chain's type into the list of parameter types, which may
9813 // mean appending it.
9816 FunctionType::param_iterator I = FTy->param_begin(),
9817 E = FTy->param_end();
9821 // Add the chain's type.
9822 NewTypes.push_back(NestTy);
9827 // Add the original type.
9828 NewTypes.push_back(*I);
9834 // Replace the trampoline call with a direct call. Let the generic
9835 // code sort out any function type mismatches.
9836 FunctionType *NewFTy =
9837 FunctionType::get(FTy->getReturnType(), NewTypes, FTy->isVarArg());
9838 Constant *NewCallee = NestF->getType() == PointerType::getUnqual(NewFTy) ?
9839 NestF : ConstantExpr::getBitCast(NestF, PointerType::getUnqual(NewFTy));
9840 const AttrListPtr &NewPAL = AttrListPtr::get(NewAttrs.begin(),NewAttrs.end());
9842 Instruction *NewCaller;
9843 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
9844 NewCaller = InvokeInst::Create(NewCallee,
9845 II->getNormalDest(), II->getUnwindDest(),
9846 NewArgs.begin(), NewArgs.end(),
9847 Caller->getName(), Caller);
9848 cast<InvokeInst>(NewCaller)->setCallingConv(II->getCallingConv());
9849 cast<InvokeInst>(NewCaller)->setAttributes(NewPAL);
9851 NewCaller = CallInst::Create(NewCallee, NewArgs.begin(), NewArgs.end(),
9852 Caller->getName(), Caller);
9853 if (cast<CallInst>(Caller)->isTailCall())
9854 cast<CallInst>(NewCaller)->setTailCall();
9855 cast<CallInst>(NewCaller)->
9856 setCallingConv(cast<CallInst>(Caller)->getCallingConv());
9857 cast<CallInst>(NewCaller)->setAttributes(NewPAL);
9859 if (Caller->getType() != Type::VoidTy && !Caller->use_empty())
9860 Caller->replaceAllUsesWith(NewCaller);
9861 Caller->eraseFromParent();
9862 RemoveFromWorkList(Caller);
9867 // Replace the trampoline call with a direct call. Since there is no 'nest'
9868 // parameter, there is no need to adjust the argument list. Let the generic
9869 // code sort out any function type mismatches.
9870 Constant *NewCallee =
9871 NestF->getType() == PTy ? NestF : ConstantExpr::getBitCast(NestF, PTy);
9872 CS.setCalledFunction(NewCallee);
9873 return CS.getInstruction();
9876 /// FoldPHIArgBinOpIntoPHI - If we have something like phi [add (a,b), add(c,d)]
9877 /// and if a/b/c/d and the add's all have a single use, turn this into two phi's
9878 /// and a single binop.
9879 Instruction *InstCombiner::FoldPHIArgBinOpIntoPHI(PHINode &PN) {
9880 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
9881 assert(isa<BinaryOperator>(FirstInst) || isa<GetElementPtrInst>(FirstInst) ||
9882 isa<CmpInst>(FirstInst));
9883 unsigned Opc = FirstInst->getOpcode();
9884 Value *LHSVal = FirstInst->getOperand(0);
9885 Value *RHSVal = FirstInst->getOperand(1);
9887 const Type *LHSType = LHSVal->getType();
9888 const Type *RHSType = RHSVal->getType();
9890 // Scan to see if all operands are the same opcode, all have one use, and all
9891 // kill their operands (i.e. the operands have one use).
9892 for (unsigned i = 0; i != PN.getNumIncomingValues(); ++i) {
9893 Instruction *I = dyn_cast<Instruction>(PN.getIncomingValue(i));
9894 if (!I || I->getOpcode() != Opc || !I->hasOneUse() ||
9895 // Verify type of the LHS matches so we don't fold cmp's of different
9896 // types or GEP's with different index types.
9897 I->getOperand(0)->getType() != LHSType ||
9898 I->getOperand(1)->getType() != RHSType)
9901 // If they are CmpInst instructions, check their predicates
9902 if (Opc == Instruction::ICmp || Opc == Instruction::FCmp)
9903 if (cast<CmpInst>(I)->getPredicate() !=
9904 cast<CmpInst>(FirstInst)->getPredicate())
9907 // Keep track of which operand needs a phi node.
9908 if (I->getOperand(0) != LHSVal) LHSVal = 0;
9909 if (I->getOperand(1) != RHSVal) RHSVal = 0;
9912 // Otherwise, this is safe to transform, determine if it is profitable.
9914 // If this is a GEP, and if the index (not the pointer) needs a PHI, bail out.
9915 // Indexes are often folded into load/store instructions, so we don't want to
9916 // hide them behind a phi.
9917 if (isa<GetElementPtrInst>(FirstInst) && RHSVal == 0)
9920 Value *InLHS = FirstInst->getOperand(0);
9921 Value *InRHS = FirstInst->getOperand(1);
9922 PHINode *NewLHS = 0, *NewRHS = 0;
9924 NewLHS = PHINode::Create(LHSType,
9925 FirstInst->getOperand(0)->getName() + ".pn");
9926 NewLHS->reserveOperandSpace(PN.getNumOperands()/2);
9927 NewLHS->addIncoming(InLHS, PN.getIncomingBlock(0));
9928 InsertNewInstBefore(NewLHS, PN);
9933 NewRHS = PHINode::Create(RHSType,
9934 FirstInst->getOperand(1)->getName() + ".pn");
9935 NewRHS->reserveOperandSpace(PN.getNumOperands()/2);
9936 NewRHS->addIncoming(InRHS, PN.getIncomingBlock(0));
9937 InsertNewInstBefore(NewRHS, PN);
9941 // Add all operands to the new PHIs.
9942 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
9944 Value *NewInLHS =cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
9945 NewLHS->addIncoming(NewInLHS, PN.getIncomingBlock(i));
9948 Value *NewInRHS =cast<Instruction>(PN.getIncomingValue(i))->getOperand(1);
9949 NewRHS->addIncoming(NewInRHS, PN.getIncomingBlock(i));
9953 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
9954 return BinaryOperator::Create(BinOp->getOpcode(), LHSVal, RHSVal);
9955 else if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst))
9956 return CmpInst::Create(CIOp->getOpcode(), CIOp->getPredicate(), LHSVal,
9959 assert(isa<GetElementPtrInst>(FirstInst));
9960 return GetElementPtrInst::Create(LHSVal, RHSVal);
9964 /// isSafeToSinkLoad - Return true if we know that it is safe sink the load out
9965 /// of the block that defines it. This means that it must be obvious the value
9966 /// of the load is not changed from the point of the load to the end of the
9969 /// Finally, it is safe, but not profitable, to sink a load targetting a
9970 /// non-address-taken alloca. Doing so will cause us to not promote the alloca
9972 static bool isSafeToSinkLoad(LoadInst *L) {
9973 BasicBlock::iterator BBI = L, E = L->getParent()->end();
9975 for (++BBI; BBI != E; ++BBI)
9976 if (BBI->mayWriteToMemory())
9979 // Check for non-address taken alloca. If not address-taken already, it isn't
9980 // profitable to do this xform.
9981 if (AllocaInst *AI = dyn_cast<AllocaInst>(L->getOperand(0))) {
9982 bool isAddressTaken = false;
9983 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
9985 if (isa<LoadInst>(UI)) continue;
9986 if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
9987 // If storing TO the alloca, then the address isn't taken.
9988 if (SI->getOperand(1) == AI) continue;
9990 isAddressTaken = true;
9994 if (!isAddressTaken)
10002 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
10003 // operator and they all are only used by the PHI, PHI together their
10004 // inputs, and do the operation once, to the result of the PHI.
10005 Instruction *InstCombiner::FoldPHIArgOpIntoPHI(PHINode &PN) {
10006 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
10008 // Scan the instruction, looking for input operations that can be folded away.
10009 // If all input operands to the phi are the same instruction (e.g. a cast from
10010 // the same type or "+42") we can pull the operation through the PHI, reducing
10011 // code size and simplifying code.
10012 Constant *ConstantOp = 0;
10013 const Type *CastSrcTy = 0;
10014 bool isVolatile = false;
10015 if (isa<CastInst>(FirstInst)) {
10016 CastSrcTy = FirstInst->getOperand(0)->getType();
10017 } else if (isa<BinaryOperator>(FirstInst) || isa<CmpInst>(FirstInst)) {
10018 // Can fold binop, compare or shift here if the RHS is a constant,
10019 // otherwise call FoldPHIArgBinOpIntoPHI.
10020 ConstantOp = dyn_cast<Constant>(FirstInst->getOperand(1));
10021 if (ConstantOp == 0)
10022 return FoldPHIArgBinOpIntoPHI(PN);
10023 } else if (LoadInst *LI = dyn_cast<LoadInst>(FirstInst)) {
10024 isVolatile = LI->isVolatile();
10025 // We can't sink the load if the loaded value could be modified between the
10026 // load and the PHI.
10027 if (LI->getParent() != PN.getIncomingBlock(0) ||
10028 !isSafeToSinkLoad(LI))
10031 // If the PHI is of volatile loads and the load block has multiple
10032 // successors, sinking it would remove a load of the volatile value from
10033 // the path through the other successor.
10035 LI->getParent()->getTerminator()->getNumSuccessors() != 1)
10038 } else if (isa<GetElementPtrInst>(FirstInst)) {
10039 if (FirstInst->getNumOperands() == 2)
10040 return FoldPHIArgBinOpIntoPHI(PN);
10041 // Can't handle general GEPs yet.
10044 return 0; // Cannot fold this operation.
10047 // Check to see if all arguments are the same operation.
10048 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
10049 if (!isa<Instruction>(PN.getIncomingValue(i))) return 0;
10050 Instruction *I = cast<Instruction>(PN.getIncomingValue(i));
10051 if (!I->hasOneUse() || !I->isSameOperationAs(FirstInst))
10054 if (I->getOperand(0)->getType() != CastSrcTy)
10055 return 0; // Cast operation must match.
10056 } else if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
10057 // We can't sink the load if the loaded value could be modified between
10058 // the load and the PHI.
10059 if (LI->isVolatile() != isVolatile ||
10060 LI->getParent() != PN.getIncomingBlock(i) ||
10061 !isSafeToSinkLoad(LI))
10064 // If the PHI is of volatile loads and the load block has multiple
10065 // successors, sinking it would remove a load of the volatile value from
10066 // the path through the other successor.
10068 LI->getParent()->getTerminator()->getNumSuccessors() != 1)
10072 } else if (I->getOperand(1) != ConstantOp) {
10077 // Okay, they are all the same operation. Create a new PHI node of the
10078 // correct type, and PHI together all of the LHS's of the instructions.
10079 PHINode *NewPN = PHINode::Create(FirstInst->getOperand(0)->getType(),
10080 PN.getName()+".in");
10081 NewPN->reserveOperandSpace(PN.getNumOperands()/2);
10083 Value *InVal = FirstInst->getOperand(0);
10084 NewPN->addIncoming(InVal, PN.getIncomingBlock(0));
10086 // Add all operands to the new PHI.
10087 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
10088 Value *NewInVal = cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
10089 if (NewInVal != InVal)
10091 NewPN->addIncoming(NewInVal, PN.getIncomingBlock(i));
10096 // The new PHI unions all of the same values together. This is really
10097 // common, so we handle it intelligently here for compile-time speed.
10101 InsertNewInstBefore(NewPN, PN);
10105 // Insert and return the new operation.
10106 if (CastInst* FirstCI = dyn_cast<CastInst>(FirstInst))
10107 return CastInst::Create(FirstCI->getOpcode(), PhiVal, PN.getType());
10108 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
10109 return BinaryOperator::Create(BinOp->getOpcode(), PhiVal, ConstantOp);
10110 if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst))
10111 return CmpInst::Create(CIOp->getOpcode(), CIOp->getPredicate(),
10112 PhiVal, ConstantOp);
10113 assert(isa<LoadInst>(FirstInst) && "Unknown operation");
10115 // If this was a volatile load that we are merging, make sure to loop through
10116 // and mark all the input loads as non-volatile. If we don't do this, we will
10117 // insert a new volatile load and the old ones will not be deletable.
10119 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i)
10120 cast<LoadInst>(PN.getIncomingValue(i))->setVolatile(false);
10122 return new LoadInst(PhiVal, "", isVolatile);
10125 /// DeadPHICycle - Return true if this PHI node is only used by a PHI node cycle
10127 static bool DeadPHICycle(PHINode *PN,
10128 SmallPtrSet<PHINode*, 16> &PotentiallyDeadPHIs) {
10129 if (PN->use_empty()) return true;
10130 if (!PN->hasOneUse()) return false;
10132 // Remember this node, and if we find the cycle, return.
10133 if (!PotentiallyDeadPHIs.insert(PN))
10136 // Don't scan crazily complex things.
10137 if (PotentiallyDeadPHIs.size() == 16)
10140 if (PHINode *PU = dyn_cast<PHINode>(PN->use_back()))
10141 return DeadPHICycle(PU, PotentiallyDeadPHIs);
10146 /// PHIsEqualValue - Return true if this phi node is always equal to
10147 /// NonPhiInVal. This happens with mutually cyclic phi nodes like:
10148 /// z = some value; x = phi (y, z); y = phi (x, z)
10149 static bool PHIsEqualValue(PHINode *PN, Value *NonPhiInVal,
10150 SmallPtrSet<PHINode*, 16> &ValueEqualPHIs) {
10151 // See if we already saw this PHI node.
10152 if (!ValueEqualPHIs.insert(PN))
10155 // Don't scan crazily complex things.
10156 if (ValueEqualPHIs.size() == 16)
10159 // Scan the operands to see if they are either phi nodes or are equal to
10161 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
10162 Value *Op = PN->getIncomingValue(i);
10163 if (PHINode *OpPN = dyn_cast<PHINode>(Op)) {
10164 if (!PHIsEqualValue(OpPN, NonPhiInVal, ValueEqualPHIs))
10166 } else if (Op != NonPhiInVal)
10174 // PHINode simplification
10176 Instruction *InstCombiner::visitPHINode(PHINode &PN) {
10177 // If LCSSA is around, don't mess with Phi nodes
10178 if (MustPreserveLCSSA) return 0;
10180 if (Value *V = PN.hasConstantValue())
10181 return ReplaceInstUsesWith(PN, V);
10183 // If all PHI operands are the same operation, pull them through the PHI,
10184 // reducing code size.
10185 if (isa<Instruction>(PN.getIncomingValue(0)) &&
10186 PN.getIncomingValue(0)->hasOneUse())
10187 if (Instruction *Result = FoldPHIArgOpIntoPHI(PN))
10190 // If this is a trivial cycle in the PHI node graph, remove it. Basically, if
10191 // this PHI only has a single use (a PHI), and if that PHI only has one use (a
10192 // PHI)... break the cycle.
10193 if (PN.hasOneUse()) {
10194 Instruction *PHIUser = cast<Instruction>(PN.use_back());
10195 if (PHINode *PU = dyn_cast<PHINode>(PHIUser)) {
10196 SmallPtrSet<PHINode*, 16> PotentiallyDeadPHIs;
10197 PotentiallyDeadPHIs.insert(&PN);
10198 if (DeadPHICycle(PU, PotentiallyDeadPHIs))
10199 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
10202 // If this phi has a single use, and if that use just computes a value for
10203 // the next iteration of a loop, delete the phi. This occurs with unused
10204 // induction variables, e.g. "for (int j = 0; ; ++j);". Detecting this
10205 // common case here is good because the only other things that catch this
10206 // are induction variable analysis (sometimes) and ADCE, which is only run
10208 if (PHIUser->hasOneUse() &&
10209 (isa<BinaryOperator>(PHIUser) || isa<GetElementPtrInst>(PHIUser)) &&
10210 PHIUser->use_back() == &PN) {
10211 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
10215 // We sometimes end up with phi cycles that non-obviously end up being the
10216 // same value, for example:
10217 // z = some value; x = phi (y, z); y = phi (x, z)
10218 // where the phi nodes don't necessarily need to be in the same block. Do a
10219 // quick check to see if the PHI node only contains a single non-phi value, if
10220 // so, scan to see if the phi cycle is actually equal to that value.
10222 unsigned InValNo = 0, NumOperandVals = PN.getNumIncomingValues();
10223 // Scan for the first non-phi operand.
10224 while (InValNo != NumOperandVals &&
10225 isa<PHINode>(PN.getIncomingValue(InValNo)))
10228 if (InValNo != NumOperandVals) {
10229 Value *NonPhiInVal = PN.getOperand(InValNo);
10231 // Scan the rest of the operands to see if there are any conflicts, if so
10232 // there is no need to recursively scan other phis.
10233 for (++InValNo; InValNo != NumOperandVals; ++InValNo) {
10234 Value *OpVal = PN.getIncomingValue(InValNo);
10235 if (OpVal != NonPhiInVal && !isa<PHINode>(OpVal))
10239 // If we scanned over all operands, then we have one unique value plus
10240 // phi values. Scan PHI nodes to see if they all merge in each other or
10242 if (InValNo == NumOperandVals) {
10243 SmallPtrSet<PHINode*, 16> ValueEqualPHIs;
10244 if (PHIsEqualValue(&PN, NonPhiInVal, ValueEqualPHIs))
10245 return ReplaceInstUsesWith(PN, NonPhiInVal);
10252 static Value *InsertCastToIntPtrTy(Value *V, const Type *DTy,
10253 Instruction *InsertPoint,
10254 InstCombiner *IC) {
10255 unsigned PtrSize = DTy->getPrimitiveSizeInBits();
10256 unsigned VTySize = V->getType()->getPrimitiveSizeInBits();
10257 // We must cast correctly to the pointer type. Ensure that we
10258 // sign extend the integer value if it is smaller as this is
10259 // used for address computation.
10260 Instruction::CastOps opcode =
10261 (VTySize < PtrSize ? Instruction::SExt :
10262 (VTySize == PtrSize ? Instruction::BitCast : Instruction::Trunc));
10263 return IC->InsertCastBefore(opcode, V, DTy, *InsertPoint);
10267 Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) {
10268 Value *PtrOp = GEP.getOperand(0);
10269 // Is it 'getelementptr %P, i32 0' or 'getelementptr %P'
10270 // If so, eliminate the noop.
10271 if (GEP.getNumOperands() == 1)
10272 return ReplaceInstUsesWith(GEP, PtrOp);
10274 if (isa<UndefValue>(GEP.getOperand(0)))
10275 return ReplaceInstUsesWith(GEP, UndefValue::get(GEP.getType()));
10277 bool HasZeroPointerIndex = false;
10278 if (Constant *C = dyn_cast<Constant>(GEP.getOperand(1)))
10279 HasZeroPointerIndex = C->isNullValue();
10281 if (GEP.getNumOperands() == 2 && HasZeroPointerIndex)
10282 return ReplaceInstUsesWith(GEP, PtrOp);
10284 // Eliminate unneeded casts for indices.
10285 bool MadeChange = false;
10287 gep_type_iterator GTI = gep_type_begin(GEP);
10288 for (User::op_iterator i = GEP.op_begin() + 1, e = GEP.op_end();
10289 i != e; ++i, ++GTI) {
10290 if (isa<SequentialType>(*GTI)) {
10291 if (CastInst *CI = dyn_cast<CastInst>(*i)) {
10292 if (CI->getOpcode() == Instruction::ZExt ||
10293 CI->getOpcode() == Instruction::SExt) {
10294 const Type *SrcTy = CI->getOperand(0)->getType();
10295 // We can eliminate a cast from i32 to i64 iff the target
10296 // is a 32-bit pointer target.
10297 if (SrcTy->getPrimitiveSizeInBits() >= TD->getPointerSizeInBits()) {
10299 *i = CI->getOperand(0);
10303 // If we are using a wider index than needed for this platform, shrink it
10304 // to what we need. If narrower, sign-extend it to what we need.
10305 // If the incoming value needs a cast instruction,
10306 // insert it. This explicit cast can make subsequent optimizations more
10309 if (TD->getTypeSizeInBits(Op->getType()) > TD->getPointerSizeInBits()) {
10310 if (Constant *C = dyn_cast<Constant>(Op)) {
10311 *i = ConstantExpr::getTrunc(C, TD->getIntPtrType());
10314 Op = InsertCastBefore(Instruction::Trunc, Op, TD->getIntPtrType(),
10319 } else if (TD->getTypeSizeInBits(Op->getType()) < TD->getPointerSizeInBits()) {
10320 if (Constant *C = dyn_cast<Constant>(Op)) {
10321 *i = ConstantExpr::getSExt(C, TD->getIntPtrType());
10324 Op = InsertCastBefore(Instruction::SExt, Op, TD->getIntPtrType(),
10332 if (MadeChange) return &GEP;
10334 // If this GEP instruction doesn't move the pointer, and if the input operand
10335 // is a bitcast of another pointer, just replace the GEP with a bitcast of the
10336 // real input to the dest type.
10337 if (GEP.hasAllZeroIndices()) {
10338 if (BitCastInst *BCI = dyn_cast<BitCastInst>(GEP.getOperand(0))) {
10339 // If the bitcast is of an allocation, and the allocation will be
10340 // converted to match the type of the cast, don't touch this.
10341 if (isa<AllocationInst>(BCI->getOperand(0))) {
10342 // See if the bitcast simplifies, if so, don't nuke this GEP yet.
10343 if (Instruction *I = visitBitCast(*BCI)) {
10346 BCI->getParent()->getInstList().insert(BCI, I);
10347 ReplaceInstUsesWith(*BCI, I);
10352 return new BitCastInst(BCI->getOperand(0), GEP.getType());
10356 // Combine Indices - If the source pointer to this getelementptr instruction
10357 // is a getelementptr instruction, combine the indices of the two
10358 // getelementptr instructions into a single instruction.
10360 SmallVector<Value*, 8> SrcGEPOperands;
10361 if (User *Src = dyn_castGetElementPtr(PtrOp))
10362 SrcGEPOperands.append(Src->op_begin(), Src->op_end());
10364 if (!SrcGEPOperands.empty()) {
10365 // Note that if our source is a gep chain itself that we wait for that
10366 // chain to be resolved before we perform this transformation. This
10367 // avoids us creating a TON of code in some cases.
10369 if (isa<GetElementPtrInst>(SrcGEPOperands[0]) &&
10370 cast<Instruction>(SrcGEPOperands[0])->getNumOperands() == 2)
10371 return 0; // Wait until our source is folded to completion.
10373 SmallVector<Value*, 8> Indices;
10375 // Find out whether the last index in the source GEP is a sequential idx.
10376 bool EndsWithSequential = false;
10377 for (gep_type_iterator I = gep_type_begin(*cast<User>(PtrOp)),
10378 E = gep_type_end(*cast<User>(PtrOp)); I != E; ++I)
10379 EndsWithSequential = !isa<StructType>(*I);
10381 // Can we combine the two pointer arithmetics offsets?
10382 if (EndsWithSequential) {
10383 // Replace: gep (gep %P, long B), long A, ...
10384 // With: T = long A+B; gep %P, T, ...
10386 Value *Sum, *SO1 = SrcGEPOperands.back(), *GO1 = GEP.getOperand(1);
10387 if (SO1 == Constant::getNullValue(SO1->getType())) {
10389 } else if (GO1 == Constant::getNullValue(GO1->getType())) {
10392 // If they aren't the same type, convert both to an integer of the
10393 // target's pointer size.
10394 if (SO1->getType() != GO1->getType()) {
10395 if (Constant *SO1C = dyn_cast<Constant>(SO1)) {
10396 SO1 = ConstantExpr::getIntegerCast(SO1C, GO1->getType(), true);
10397 } else if (Constant *GO1C = dyn_cast<Constant>(GO1)) {
10398 GO1 = ConstantExpr::getIntegerCast(GO1C, SO1->getType(), true);
10400 unsigned PS = TD->getPointerSizeInBits();
10401 if (TD->getTypeSizeInBits(SO1->getType()) == PS) {
10402 // Convert GO1 to SO1's type.
10403 GO1 = InsertCastToIntPtrTy(GO1, SO1->getType(), &GEP, this);
10405 } else if (TD->getTypeSizeInBits(GO1->getType()) == PS) {
10406 // Convert SO1 to GO1's type.
10407 SO1 = InsertCastToIntPtrTy(SO1, GO1->getType(), &GEP, this);
10409 const Type *PT = TD->getIntPtrType();
10410 SO1 = InsertCastToIntPtrTy(SO1, PT, &GEP, this);
10411 GO1 = InsertCastToIntPtrTy(GO1, PT, &GEP, this);
10415 if (isa<Constant>(SO1) && isa<Constant>(GO1))
10416 Sum = ConstantExpr::getAdd(cast<Constant>(SO1), cast<Constant>(GO1));
10418 Sum = BinaryOperator::CreateAdd(SO1, GO1, PtrOp->getName()+".sum");
10419 InsertNewInstBefore(cast<Instruction>(Sum), GEP);
10423 // Recycle the GEP we already have if possible.
10424 if (SrcGEPOperands.size() == 2) {
10425 GEP.setOperand(0, SrcGEPOperands[0]);
10426 GEP.setOperand(1, Sum);
10429 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
10430 SrcGEPOperands.end()-1);
10431 Indices.push_back(Sum);
10432 Indices.insert(Indices.end(), GEP.op_begin()+2, GEP.op_end());
10434 } else if (isa<Constant>(*GEP.idx_begin()) &&
10435 cast<Constant>(*GEP.idx_begin())->isNullValue() &&
10436 SrcGEPOperands.size() != 1) {
10437 // Otherwise we can do the fold if the first index of the GEP is a zero
10438 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
10439 SrcGEPOperands.end());
10440 Indices.insert(Indices.end(), GEP.idx_begin()+1, GEP.idx_end());
10443 if (!Indices.empty())
10444 return GetElementPtrInst::Create(SrcGEPOperands[0], Indices.begin(),
10445 Indices.end(), GEP.getName());
10447 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(PtrOp)) {
10448 // GEP of global variable. If all of the indices for this GEP are
10449 // constants, we can promote this to a constexpr instead of an instruction.
10451 // Scan for nonconstants...
10452 SmallVector<Constant*, 8> Indices;
10453 User::op_iterator I = GEP.idx_begin(), E = GEP.idx_end();
10454 for (; I != E && isa<Constant>(*I); ++I)
10455 Indices.push_back(cast<Constant>(*I));
10457 if (I == E) { // If they are all constants...
10458 Constant *CE = ConstantExpr::getGetElementPtr(GV,
10459 &Indices[0],Indices.size());
10461 // Replace all uses of the GEP with the new constexpr...
10462 return ReplaceInstUsesWith(GEP, CE);
10464 } else if (Value *X = getBitCastOperand(PtrOp)) { // Is the operand a cast?
10465 if (!isa<PointerType>(X->getType())) {
10466 // Not interesting. Source pointer must be a cast from pointer.
10467 } else if (HasZeroPointerIndex) {
10468 // transform: GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ...
10469 // into : GEP [10 x i8]* X, i32 0, ...
10471 // This occurs when the program declares an array extern like "int X[];"
10473 const PointerType *CPTy = cast<PointerType>(PtrOp->getType());
10474 const PointerType *XTy = cast<PointerType>(X->getType());
10475 if (const ArrayType *XATy =
10476 dyn_cast<ArrayType>(XTy->getElementType()))
10477 if (const ArrayType *CATy =
10478 dyn_cast<ArrayType>(CPTy->getElementType()))
10479 if (CATy->getElementType() == XATy->getElementType()) {
10480 // At this point, we know that the cast source type is a pointer
10481 // to an array of the same type as the destination pointer
10482 // array. Because the array type is never stepped over (there
10483 // is a leading zero) we can fold the cast into this GEP.
10484 GEP.setOperand(0, X);
10487 } else if (GEP.getNumOperands() == 2) {
10488 // Transform things like:
10489 // %t = getelementptr i32* bitcast ([2 x i32]* %str to i32*), i32 %V
10490 // into: %t1 = getelementptr [2 x i32]* %str, i32 0, i32 %V; bitcast
10491 const Type *SrcElTy = cast<PointerType>(X->getType())->getElementType();
10492 const Type *ResElTy=cast<PointerType>(PtrOp->getType())->getElementType();
10493 if (isa<ArrayType>(SrcElTy) &&
10494 TD->getABITypeSize(cast<ArrayType>(SrcElTy)->getElementType()) ==
10495 TD->getABITypeSize(ResElTy)) {
10497 Idx[0] = Constant::getNullValue(Type::Int32Ty);
10498 Idx[1] = GEP.getOperand(1);
10499 Value *V = InsertNewInstBefore(
10500 GetElementPtrInst::Create(X, Idx, Idx + 2, GEP.getName()), GEP);
10501 // V and GEP are both pointer types --> BitCast
10502 return new BitCastInst(V, GEP.getType());
10505 // Transform things like:
10506 // getelementptr i8* bitcast ([100 x double]* X to i8*), i32 %tmp
10507 // (where tmp = 8*tmp2) into:
10508 // getelementptr [100 x double]* %arr, i32 0, i32 %tmp2; bitcast
10510 if (isa<ArrayType>(SrcElTy) && ResElTy == Type::Int8Ty) {
10511 uint64_t ArrayEltSize =
10512 TD->getABITypeSize(cast<ArrayType>(SrcElTy)->getElementType());
10514 // Check to see if "tmp" is a scale by a multiple of ArrayEltSize. We
10515 // allow either a mul, shift, or constant here.
10517 ConstantInt *Scale = 0;
10518 if (ArrayEltSize == 1) {
10519 NewIdx = GEP.getOperand(1);
10520 Scale = ConstantInt::get(NewIdx->getType(), 1);
10521 } else if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP.getOperand(1))) {
10522 NewIdx = ConstantInt::get(CI->getType(), 1);
10524 } else if (Instruction *Inst =dyn_cast<Instruction>(GEP.getOperand(1))){
10525 if (Inst->getOpcode() == Instruction::Shl &&
10526 isa<ConstantInt>(Inst->getOperand(1))) {
10527 ConstantInt *ShAmt = cast<ConstantInt>(Inst->getOperand(1));
10528 uint32_t ShAmtVal = ShAmt->getLimitedValue(64);
10529 Scale = ConstantInt::get(Inst->getType(), 1ULL << ShAmtVal);
10530 NewIdx = Inst->getOperand(0);
10531 } else if (Inst->getOpcode() == Instruction::Mul &&
10532 isa<ConstantInt>(Inst->getOperand(1))) {
10533 Scale = cast<ConstantInt>(Inst->getOperand(1));
10534 NewIdx = Inst->getOperand(0);
10538 // If the index will be to exactly the right offset with the scale taken
10539 // out, perform the transformation. Note, we don't know whether Scale is
10540 // signed or not. We'll use unsigned version of division/modulo
10541 // operation after making sure Scale doesn't have the sign bit set.
10542 if (Scale && Scale->getSExtValue() >= 0LL &&
10543 Scale->getZExtValue() % ArrayEltSize == 0) {
10544 Scale = ConstantInt::get(Scale->getType(),
10545 Scale->getZExtValue() / ArrayEltSize);
10546 if (Scale->getZExtValue() != 1) {
10547 Constant *C = ConstantExpr::getIntegerCast(Scale, NewIdx->getType(),
10549 Instruction *Sc = BinaryOperator::CreateMul(NewIdx, C, "idxscale");
10550 NewIdx = InsertNewInstBefore(Sc, GEP);
10553 // Insert the new GEP instruction.
10555 Idx[0] = Constant::getNullValue(Type::Int32Ty);
10557 Instruction *NewGEP =
10558 GetElementPtrInst::Create(X, Idx, Idx + 2, GEP.getName());
10559 NewGEP = InsertNewInstBefore(NewGEP, GEP);
10560 // The NewGEP must be pointer typed, so must the old one -> BitCast
10561 return new BitCastInst(NewGEP, GEP.getType());
10570 Instruction *InstCombiner::visitAllocationInst(AllocationInst &AI) {
10571 // Convert: malloc Ty, C - where C is a constant != 1 into: malloc [C x Ty], 1
10572 if (AI.isArrayAllocation()) { // Check C != 1
10573 if (const ConstantInt *C = dyn_cast<ConstantInt>(AI.getArraySize())) {
10574 const Type *NewTy =
10575 ArrayType::get(AI.getAllocatedType(), C->getZExtValue());
10576 AllocationInst *New = 0;
10578 // Create and insert the replacement instruction...
10579 if (isa<MallocInst>(AI))
10580 New = new MallocInst(NewTy, 0, AI.getAlignment(), AI.getName());
10582 assert(isa<AllocaInst>(AI) && "Unknown type of allocation inst!");
10583 New = new AllocaInst(NewTy, 0, AI.getAlignment(), AI.getName());
10586 InsertNewInstBefore(New, AI);
10588 // Scan to the end of the allocation instructions, to skip over a block of
10589 // allocas if possible...
10591 BasicBlock::iterator It = New;
10592 while (isa<AllocationInst>(*It)) ++It;
10594 // Now that I is pointing to the first non-allocation-inst in the block,
10595 // insert our getelementptr instruction...
10597 Value *NullIdx = Constant::getNullValue(Type::Int32Ty);
10601 Value *V = GetElementPtrInst::Create(New, Idx, Idx + 2,
10602 New->getName()+".sub", It);
10604 // Now make everything use the getelementptr instead of the original
10606 return ReplaceInstUsesWith(AI, V);
10607 } else if (isa<UndefValue>(AI.getArraySize())) {
10608 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
10612 // If alloca'ing a zero byte object, replace the alloca with a null pointer.
10613 // Note that we only do this for alloca's, because malloc should allocate and
10614 // return a unique pointer, even for a zero byte allocation.
10615 if (isa<AllocaInst>(AI) && AI.getAllocatedType()->isSized() &&
10616 TD->getABITypeSize(AI.getAllocatedType()) == 0)
10617 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
10622 Instruction *InstCombiner::visitFreeInst(FreeInst &FI) {
10623 Value *Op = FI.getOperand(0);
10625 // free undef -> unreachable.
10626 if (isa<UndefValue>(Op)) {
10627 // Insert a new store to null because we cannot modify the CFG here.
10628 new StoreInst(ConstantInt::getTrue(),
10629 UndefValue::get(PointerType::getUnqual(Type::Int1Ty)), &FI);
10630 return EraseInstFromFunction(FI);
10633 // If we have 'free null' delete the instruction. This can happen in stl code
10634 // when lots of inlining happens.
10635 if (isa<ConstantPointerNull>(Op))
10636 return EraseInstFromFunction(FI);
10638 // Change free <ty>* (cast <ty2>* X to <ty>*) into free <ty2>* X
10639 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op)) {
10640 FI.setOperand(0, CI->getOperand(0));
10644 // Change free (gep X, 0,0,0,0) into free(X)
10645 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op)) {
10646 if (GEPI->hasAllZeroIndices()) {
10647 AddToWorkList(GEPI);
10648 FI.setOperand(0, GEPI->getOperand(0));
10653 // Change free(malloc) into nothing, if the malloc has a single use.
10654 if (MallocInst *MI = dyn_cast<MallocInst>(Op))
10655 if (MI->hasOneUse()) {
10656 EraseInstFromFunction(FI);
10657 return EraseInstFromFunction(*MI);
10664 /// InstCombineLoadCast - Fold 'load (cast P)' -> cast (load P)' when possible.
10665 static Instruction *InstCombineLoadCast(InstCombiner &IC, LoadInst &LI,
10666 const TargetData *TD) {
10667 User *CI = cast<User>(LI.getOperand(0));
10668 Value *CastOp = CI->getOperand(0);
10670 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(CI)) {
10671 // Instead of loading constant c string, use corresponding integer value
10672 // directly if string length is small enough.
10674 if (GetConstantStringInfo(CE->getOperand(0), Str) && !Str.empty()) {
10675 unsigned len = Str.length();
10676 const Type *Ty = cast<PointerType>(CE->getType())->getElementType();
10677 unsigned numBits = Ty->getPrimitiveSizeInBits();
10678 // Replace LI with immediate integer store.
10679 if ((numBits >> 3) == len + 1) {
10680 APInt StrVal(numBits, 0);
10681 APInt SingleChar(numBits, 0);
10682 if (TD->isLittleEndian()) {
10683 for (signed i = len-1; i >= 0; i--) {
10684 SingleChar = (uint64_t) Str[i];
10685 StrVal = (StrVal << 8) | SingleChar;
10688 for (unsigned i = 0; i < len; i++) {
10689 SingleChar = (uint64_t) Str[i];
10690 StrVal = (StrVal << 8) | SingleChar;
10692 // Append NULL at the end.
10694 StrVal = (StrVal << 8) | SingleChar;
10696 Value *NL = ConstantInt::get(StrVal);
10697 return IC.ReplaceInstUsesWith(LI, NL);
10702 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
10703 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
10704 const Type *SrcPTy = SrcTy->getElementType();
10706 if (DestPTy->isInteger() || isa<PointerType>(DestPTy) ||
10707 isa<VectorType>(DestPTy)) {
10708 // If the source is an array, the code below will not succeed. Check to
10709 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
10711 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
10712 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
10713 if (ASrcTy->getNumElements() != 0) {
10715 Idxs[0] = Idxs[1] = Constant::getNullValue(Type::Int32Ty);
10716 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs, 2);
10717 SrcTy = cast<PointerType>(CastOp->getType());
10718 SrcPTy = SrcTy->getElementType();
10721 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy) ||
10722 isa<VectorType>(SrcPTy)) &&
10723 // Do not allow turning this into a load of an integer, which is then
10724 // casted to a pointer, this pessimizes pointer analysis a lot.
10725 (isa<PointerType>(SrcPTy) == isa<PointerType>(LI.getType())) &&
10726 IC.getTargetData().getTypeSizeInBits(SrcPTy) ==
10727 IC.getTargetData().getTypeSizeInBits(DestPTy)) {
10729 // Okay, we are casting from one integer or pointer type to another of
10730 // the same size. Instead of casting the pointer before the load, cast
10731 // the result of the loaded value.
10732 Value *NewLoad = IC.InsertNewInstBefore(new LoadInst(CastOp,
10734 LI.isVolatile()),LI);
10735 // Now cast the result of the load.
10736 return new BitCastInst(NewLoad, LI.getType());
10743 /// isSafeToLoadUnconditionally - Return true if we know that executing a load
10744 /// from this value cannot trap. If it is not obviously safe to load from the
10745 /// specified pointer, we do a quick local scan of the basic block containing
10746 /// ScanFrom, to determine if the address is already accessed.
10747 static bool isSafeToLoadUnconditionally(Value *V, Instruction *ScanFrom) {
10748 // If it is an alloca it is always safe to load from.
10749 if (isa<AllocaInst>(V)) return true;
10751 // If it is a global variable it is mostly safe to load from.
10752 if (const GlobalValue *GV = dyn_cast<GlobalVariable>(V))
10753 // Don't try to evaluate aliases. External weak GV can be null.
10754 return !isa<GlobalAlias>(GV) && !GV->hasExternalWeakLinkage();
10756 // Otherwise, be a little bit agressive by scanning the local block where we
10757 // want to check to see if the pointer is already being loaded or stored
10758 // from/to. If so, the previous load or store would have already trapped,
10759 // so there is no harm doing an extra load (also, CSE will later eliminate
10760 // the load entirely).
10761 BasicBlock::iterator BBI = ScanFrom, E = ScanFrom->getParent()->begin();
10766 // If we see a free or a call (which might do a free) the pointer could be
10768 if (isa<FreeInst>(BBI) || isa<CallInst>(BBI))
10771 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
10772 if (LI->getOperand(0) == V) return true;
10773 } else if (StoreInst *SI = dyn_cast<StoreInst>(BBI)) {
10774 if (SI->getOperand(1) == V) return true;
10781 Instruction *InstCombiner::visitLoadInst(LoadInst &LI) {
10782 Value *Op = LI.getOperand(0);
10784 // Attempt to improve the alignment.
10785 unsigned KnownAlign = GetOrEnforceKnownAlignment(Op);
10787 (LI.getAlignment() == 0 ? TD->getABITypeAlignment(LI.getType()) :
10788 LI.getAlignment()))
10789 LI.setAlignment(KnownAlign);
10791 // load (cast X) --> cast (load X) iff safe
10792 if (isa<CastInst>(Op))
10793 if (Instruction *Res = InstCombineLoadCast(*this, LI, TD))
10796 // None of the following transforms are legal for volatile loads.
10797 if (LI.isVolatile()) return 0;
10799 // Do really simple store-to-load forwarding and load CSE, to catch cases
10800 // where there are several consequtive memory accesses to the same location,
10801 // separated by a few arithmetic operations.
10802 BasicBlock::iterator BBI = &LI;
10803 if (Value *AvailableVal = FindAvailableLoadedValue(Op, LI.getParent(), BBI,6))
10804 return ReplaceInstUsesWith(LI, AvailableVal);
10806 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op)) {
10807 const Value *GEPI0 = GEPI->getOperand(0);
10808 // TODO: Consider a target hook for valid address spaces for this xform.
10809 if (isa<ConstantPointerNull>(GEPI0) &&
10810 cast<PointerType>(GEPI0->getType())->getAddressSpace() == 0) {
10811 // Insert a new store to null instruction before the load to indicate
10812 // that this code is not reachable. We do this instead of inserting
10813 // an unreachable instruction directly because we cannot modify the
10815 new StoreInst(UndefValue::get(LI.getType()),
10816 Constant::getNullValue(Op->getType()), &LI);
10817 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
10821 if (Constant *C = dyn_cast<Constant>(Op)) {
10822 // load null/undef -> undef
10823 // TODO: Consider a target hook for valid address spaces for this xform.
10824 if (isa<UndefValue>(C) || (C->isNullValue() &&
10825 cast<PointerType>(Op->getType())->getAddressSpace() == 0)) {
10826 // Insert a new store to null instruction before the load to indicate that
10827 // this code is not reachable. We do this instead of inserting an
10828 // unreachable instruction directly because we cannot modify the CFG.
10829 new StoreInst(UndefValue::get(LI.getType()),
10830 Constant::getNullValue(Op->getType()), &LI);
10831 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
10834 // Instcombine load (constant global) into the value loaded.
10835 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Op))
10836 if (GV->isConstant() && !GV->isDeclaration())
10837 return ReplaceInstUsesWith(LI, GV->getInitializer());
10839 // Instcombine load (constantexpr_GEP global, 0, ...) into the value loaded.
10840 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Op)) {
10841 if (CE->getOpcode() == Instruction::GetElementPtr) {
10842 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(CE->getOperand(0)))
10843 if (GV->isConstant() && !GV->isDeclaration())
10845 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE))
10846 return ReplaceInstUsesWith(LI, V);
10847 if (CE->getOperand(0)->isNullValue()) {
10848 // Insert a new store to null instruction before the load to indicate
10849 // that this code is not reachable. We do this instead of inserting
10850 // an unreachable instruction directly because we cannot modify the
10852 new StoreInst(UndefValue::get(LI.getType()),
10853 Constant::getNullValue(Op->getType()), &LI);
10854 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
10857 } else if (CE->isCast()) {
10858 if (Instruction *Res = InstCombineLoadCast(*this, LI, TD))
10864 // If this load comes from anywhere in a constant global, and if the global
10865 // is all undef or zero, we know what it loads.
10866 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Op->getUnderlyingObject())){
10867 if (GV->isConstant() && GV->hasInitializer()) {
10868 if (GV->getInitializer()->isNullValue())
10869 return ReplaceInstUsesWith(LI, Constant::getNullValue(LI.getType()));
10870 else if (isa<UndefValue>(GV->getInitializer()))
10871 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
10875 if (Op->hasOneUse()) {
10876 // Change select and PHI nodes to select values instead of addresses: this
10877 // helps alias analysis out a lot, allows many others simplifications, and
10878 // exposes redundancy in the code.
10880 // Note that we cannot do the transformation unless we know that the
10881 // introduced loads cannot trap! Something like this is valid as long as
10882 // the condition is always false: load (select bool %C, int* null, int* %G),
10883 // but it would not be valid if we transformed it to load from null
10884 // unconditionally.
10886 if (SelectInst *SI = dyn_cast<SelectInst>(Op)) {
10887 // load (select (Cond, &V1, &V2)) --> select(Cond, load &V1, load &V2).
10888 if (isSafeToLoadUnconditionally(SI->getOperand(1), SI) &&
10889 isSafeToLoadUnconditionally(SI->getOperand(2), SI)) {
10890 Value *V1 = InsertNewInstBefore(new LoadInst(SI->getOperand(1),
10891 SI->getOperand(1)->getName()+".val"), LI);
10892 Value *V2 = InsertNewInstBefore(new LoadInst(SI->getOperand(2),
10893 SI->getOperand(2)->getName()+".val"), LI);
10894 return SelectInst::Create(SI->getCondition(), V1, V2);
10897 // load (select (cond, null, P)) -> load P
10898 if (Constant *C = dyn_cast<Constant>(SI->getOperand(1)))
10899 if (C->isNullValue()) {
10900 LI.setOperand(0, SI->getOperand(2));
10904 // load (select (cond, P, null)) -> load P
10905 if (Constant *C = dyn_cast<Constant>(SI->getOperand(2)))
10906 if (C->isNullValue()) {
10907 LI.setOperand(0, SI->getOperand(1));
10915 /// InstCombineStoreToCast - Fold store V, (cast P) -> store (cast V), P
10917 static Instruction *InstCombineStoreToCast(InstCombiner &IC, StoreInst &SI) {
10918 User *CI = cast<User>(SI.getOperand(1));
10919 Value *CastOp = CI->getOperand(0);
10921 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
10922 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
10923 const Type *SrcPTy = SrcTy->getElementType();
10925 if (DestPTy->isInteger() || isa<PointerType>(DestPTy)) {
10926 // If the source is an array, the code below will not succeed. Check to
10927 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
10929 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
10930 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
10931 if (ASrcTy->getNumElements() != 0) {
10933 Idxs[0] = Idxs[1] = Constant::getNullValue(Type::Int32Ty);
10934 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs, 2);
10935 SrcTy = cast<PointerType>(CastOp->getType());
10936 SrcPTy = SrcTy->getElementType();
10939 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy)) &&
10940 IC.getTargetData().getTypeSizeInBits(SrcPTy) ==
10941 IC.getTargetData().getTypeSizeInBits(DestPTy)) {
10943 // Okay, we are casting from one integer or pointer type to another of
10944 // the same size. Instead of casting the pointer before
10945 // the store, cast the value to be stored.
10947 Value *SIOp0 = SI.getOperand(0);
10948 Instruction::CastOps opcode = Instruction::BitCast;
10949 const Type* CastSrcTy = SIOp0->getType();
10950 const Type* CastDstTy = SrcPTy;
10951 if (isa<PointerType>(CastDstTy)) {
10952 if (CastSrcTy->isInteger())
10953 opcode = Instruction::IntToPtr;
10954 } else if (isa<IntegerType>(CastDstTy)) {
10955 if (isa<PointerType>(SIOp0->getType()))
10956 opcode = Instruction::PtrToInt;
10958 if (Constant *C = dyn_cast<Constant>(SIOp0))
10959 NewCast = ConstantExpr::getCast(opcode, C, CastDstTy);
10961 NewCast = IC.InsertNewInstBefore(
10962 CastInst::Create(opcode, SIOp0, CastDstTy, SIOp0->getName()+".c"),
10964 return new StoreInst(NewCast, CastOp);
10971 /// equivalentAddressValues - Test if A and B will obviously have the same
10972 /// value. This includes recognizing that %t0 and %t1 will have the same
10973 /// value in code like this:
10974 /// %t0 = getelementptr @a, 0, 3
10975 /// store i32 0, i32* %t0
10976 /// %t1 = getelementptr @a, 0, 3
10977 /// %t2 = load i32* %t1
10979 static bool equivalentAddressValues(Value *A, Value *B) {
10980 // Test if the values are trivially equivalent.
10981 if (A == B) return true;
10983 // Test if the values come form identical arithmetic instructions.
10984 if (isa<BinaryOperator>(A) ||
10985 isa<CastInst>(A) ||
10987 isa<GetElementPtrInst>(A))
10988 if (Instruction *BI = dyn_cast<Instruction>(B))
10989 if (cast<Instruction>(A)->isIdenticalTo(BI))
10992 // Otherwise they may not be equivalent.
10996 Instruction *InstCombiner::visitStoreInst(StoreInst &SI) {
10997 Value *Val = SI.getOperand(0);
10998 Value *Ptr = SI.getOperand(1);
11000 if (isa<UndefValue>(Ptr)) { // store X, undef -> noop (even if volatile)
11001 EraseInstFromFunction(SI);
11006 // If the RHS is an alloca with a single use, zapify the store, making the
11008 if (Ptr->hasOneUse() && !SI.isVolatile()) {
11009 if (isa<AllocaInst>(Ptr)) {
11010 EraseInstFromFunction(SI);
11015 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr))
11016 if (isa<AllocaInst>(GEP->getOperand(0)) &&
11017 GEP->getOperand(0)->hasOneUse()) {
11018 EraseInstFromFunction(SI);
11024 // Attempt to improve the alignment.
11025 unsigned KnownAlign = GetOrEnforceKnownAlignment(Ptr);
11027 (SI.getAlignment() == 0 ? TD->getABITypeAlignment(Val->getType()) :
11028 SI.getAlignment()))
11029 SI.setAlignment(KnownAlign);
11031 // Do really simple DSE, to catch cases where there are several consequtive
11032 // stores to the same location, separated by a few arithmetic operations. This
11033 // situation often occurs with bitfield accesses.
11034 BasicBlock::iterator BBI = &SI;
11035 for (unsigned ScanInsts = 6; BBI != SI.getParent()->begin() && ScanInsts;
11039 if (StoreInst *PrevSI = dyn_cast<StoreInst>(BBI)) {
11040 // Prev store isn't volatile, and stores to the same location?
11041 if (!PrevSI->isVolatile() &&equivalentAddressValues(PrevSI->getOperand(1),
11042 SI.getOperand(1))) {
11045 EraseInstFromFunction(*PrevSI);
11051 // If this is a load, we have to stop. However, if the loaded value is from
11052 // the pointer we're loading and is producing the pointer we're storing,
11053 // then *this* store is dead (X = load P; store X -> P).
11054 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
11055 if (LI == Val && equivalentAddressValues(LI->getOperand(0), Ptr) &&
11056 !SI.isVolatile()) {
11057 EraseInstFromFunction(SI);
11061 // Otherwise, this is a load from some other location. Stores before it
11062 // may not be dead.
11066 // Don't skip over loads or things that can modify memory.
11067 if (BBI->mayWriteToMemory() || BBI->mayReadFromMemory())
11072 if (SI.isVolatile()) return 0; // Don't hack volatile stores.
11074 // store X, null -> turns into 'unreachable' in SimplifyCFG
11075 if (isa<ConstantPointerNull>(Ptr)) {
11076 if (!isa<UndefValue>(Val)) {
11077 SI.setOperand(0, UndefValue::get(Val->getType()));
11078 if (Instruction *U = dyn_cast<Instruction>(Val))
11079 AddToWorkList(U); // Dropped a use.
11082 return 0; // Do not modify these!
11085 // store undef, Ptr -> noop
11086 if (isa<UndefValue>(Val)) {
11087 EraseInstFromFunction(SI);
11092 // If the pointer destination is a cast, see if we can fold the cast into the
11094 if (isa<CastInst>(Ptr))
11095 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
11097 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
11099 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
11103 // If this store is the last instruction in the basic block, and if the block
11104 // ends with an unconditional branch, try to move it to the successor block.
11106 if (BranchInst *BI = dyn_cast<BranchInst>(BBI))
11107 if (BI->isUnconditional())
11108 if (SimplifyStoreAtEndOfBlock(SI))
11109 return 0; // xform done!
11114 /// SimplifyStoreAtEndOfBlock - Turn things like:
11115 /// if () { *P = v1; } else { *P = v2 }
11116 /// into a phi node with a store in the successor.
11118 /// Simplify things like:
11119 /// *P = v1; if () { *P = v2; }
11120 /// into a phi node with a store in the successor.
11122 bool InstCombiner::SimplifyStoreAtEndOfBlock(StoreInst &SI) {
11123 BasicBlock *StoreBB = SI.getParent();
11125 // Check to see if the successor block has exactly two incoming edges. If
11126 // so, see if the other predecessor contains a store to the same location.
11127 // if so, insert a PHI node (if needed) and move the stores down.
11128 BasicBlock *DestBB = StoreBB->getTerminator()->getSuccessor(0);
11130 // Determine whether Dest has exactly two predecessors and, if so, compute
11131 // the other predecessor.
11132 pred_iterator PI = pred_begin(DestBB);
11133 BasicBlock *OtherBB = 0;
11134 if (*PI != StoreBB)
11137 if (PI == pred_end(DestBB))
11140 if (*PI != StoreBB) {
11145 if (++PI != pred_end(DestBB))
11148 // Bail out if all the relevant blocks aren't distinct (this can happen,
11149 // for example, if SI is in an infinite loop)
11150 if (StoreBB == DestBB || OtherBB == DestBB)
11153 // Verify that the other block ends in a branch and is not otherwise empty.
11154 BasicBlock::iterator BBI = OtherBB->getTerminator();
11155 BranchInst *OtherBr = dyn_cast<BranchInst>(BBI);
11156 if (!OtherBr || BBI == OtherBB->begin())
11159 // If the other block ends in an unconditional branch, check for the 'if then
11160 // else' case. there is an instruction before the branch.
11161 StoreInst *OtherStore = 0;
11162 if (OtherBr->isUnconditional()) {
11163 // If this isn't a store, or isn't a store to the same location, bail out.
11165 OtherStore = dyn_cast<StoreInst>(BBI);
11166 if (!OtherStore || OtherStore->getOperand(1) != SI.getOperand(1))
11169 // Otherwise, the other block ended with a conditional branch. If one of the
11170 // destinations is StoreBB, then we have the if/then case.
11171 if (OtherBr->getSuccessor(0) != StoreBB &&
11172 OtherBr->getSuccessor(1) != StoreBB)
11175 // Okay, we know that OtherBr now goes to Dest and StoreBB, so this is an
11176 // if/then triangle. See if there is a store to the same ptr as SI that
11177 // lives in OtherBB.
11179 // Check to see if we find the matching store.
11180 if ((OtherStore = dyn_cast<StoreInst>(BBI))) {
11181 if (OtherStore->getOperand(1) != SI.getOperand(1))
11185 // If we find something that may be using or overwriting the stored
11186 // value, or if we run out of instructions, we can't do the xform.
11187 if (BBI->mayReadFromMemory() || BBI->mayWriteToMemory() ||
11188 BBI == OtherBB->begin())
11192 // In order to eliminate the store in OtherBr, we have to
11193 // make sure nothing reads or overwrites the stored value in
11195 for (BasicBlock::iterator I = StoreBB->begin(); &*I != &SI; ++I) {
11196 // FIXME: This should really be AA driven.
11197 if (I->mayReadFromMemory() || I->mayWriteToMemory())
11202 // Insert a PHI node now if we need it.
11203 Value *MergedVal = OtherStore->getOperand(0);
11204 if (MergedVal != SI.getOperand(0)) {
11205 PHINode *PN = PHINode::Create(MergedVal->getType(), "storemerge");
11206 PN->reserveOperandSpace(2);
11207 PN->addIncoming(SI.getOperand(0), SI.getParent());
11208 PN->addIncoming(OtherStore->getOperand(0), OtherBB);
11209 MergedVal = InsertNewInstBefore(PN, DestBB->front());
11212 // Advance to a place where it is safe to insert the new store and
11214 BBI = DestBB->getFirstNonPHI();
11215 InsertNewInstBefore(new StoreInst(MergedVal, SI.getOperand(1),
11216 OtherStore->isVolatile()), *BBI);
11218 // Nuke the old stores.
11219 EraseInstFromFunction(SI);
11220 EraseInstFromFunction(*OtherStore);
11226 Instruction *InstCombiner::visitBranchInst(BranchInst &BI) {
11227 // Change br (not X), label True, label False to: br X, label False, True
11229 BasicBlock *TrueDest;
11230 BasicBlock *FalseDest;
11231 if (match(&BI, m_Br(m_Not(m_Value(X)), TrueDest, FalseDest)) &&
11232 !isa<Constant>(X)) {
11233 // Swap Destinations and condition...
11234 BI.setCondition(X);
11235 BI.setSuccessor(0, FalseDest);
11236 BI.setSuccessor(1, TrueDest);
11240 // Cannonicalize fcmp_one -> fcmp_oeq
11241 FCmpInst::Predicate FPred; Value *Y;
11242 if (match(&BI, m_Br(m_FCmp(FPred, m_Value(X), m_Value(Y)),
11243 TrueDest, FalseDest)))
11244 if ((FPred == FCmpInst::FCMP_ONE || FPred == FCmpInst::FCMP_OLE ||
11245 FPred == FCmpInst::FCMP_OGE) && BI.getCondition()->hasOneUse()) {
11246 FCmpInst *I = cast<FCmpInst>(BI.getCondition());
11247 FCmpInst::Predicate NewPred = FCmpInst::getInversePredicate(FPred);
11248 Instruction *NewSCC = new FCmpInst(NewPred, X, Y, "", I);
11249 NewSCC->takeName(I);
11250 // Swap Destinations and condition...
11251 BI.setCondition(NewSCC);
11252 BI.setSuccessor(0, FalseDest);
11253 BI.setSuccessor(1, TrueDest);
11254 RemoveFromWorkList(I);
11255 I->eraseFromParent();
11256 AddToWorkList(NewSCC);
11260 // Cannonicalize icmp_ne -> icmp_eq
11261 ICmpInst::Predicate IPred;
11262 if (match(&BI, m_Br(m_ICmp(IPred, m_Value(X), m_Value(Y)),
11263 TrueDest, FalseDest)))
11264 if ((IPred == ICmpInst::ICMP_NE || IPred == ICmpInst::ICMP_ULE ||
11265 IPred == ICmpInst::ICMP_SLE || IPred == ICmpInst::ICMP_UGE ||
11266 IPred == ICmpInst::ICMP_SGE) && BI.getCondition()->hasOneUse()) {
11267 ICmpInst *I = cast<ICmpInst>(BI.getCondition());
11268 ICmpInst::Predicate NewPred = ICmpInst::getInversePredicate(IPred);
11269 Instruction *NewSCC = new ICmpInst(NewPred, X, Y, "", I);
11270 NewSCC->takeName(I);
11271 // Swap Destinations and condition...
11272 BI.setCondition(NewSCC);
11273 BI.setSuccessor(0, FalseDest);
11274 BI.setSuccessor(1, TrueDest);
11275 RemoveFromWorkList(I);
11276 I->eraseFromParent();;
11277 AddToWorkList(NewSCC);
11284 Instruction *InstCombiner::visitSwitchInst(SwitchInst &SI) {
11285 Value *Cond = SI.getCondition();
11286 if (Instruction *I = dyn_cast<Instruction>(Cond)) {
11287 if (I->getOpcode() == Instruction::Add)
11288 if (ConstantInt *AddRHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
11289 // change 'switch (X+4) case 1:' into 'switch (X) case -3'
11290 for (unsigned i = 2, e = SI.getNumOperands(); i != e; i += 2)
11291 SI.setOperand(i,ConstantExpr::getSub(cast<Constant>(SI.getOperand(i)),
11293 SI.setOperand(0, I->getOperand(0));
11301 Instruction *InstCombiner::visitExtractValueInst(ExtractValueInst &EV) {
11302 Value *Agg = EV.getAggregateOperand();
11304 if (!EV.hasIndices())
11305 return ReplaceInstUsesWith(EV, Agg);
11307 if (Constant *C = dyn_cast<Constant>(Agg)) {
11308 if (isa<UndefValue>(C))
11309 return ReplaceInstUsesWith(EV, UndefValue::get(EV.getType()));
11311 if (isa<ConstantAggregateZero>(C))
11312 return ReplaceInstUsesWith(EV, Constant::getNullValue(EV.getType()));
11314 if (isa<ConstantArray>(C) || isa<ConstantStruct>(C)) {
11315 // Extract the element indexed by the first index out of the constant
11316 Value *V = C->getOperand(*EV.idx_begin());
11317 if (EV.getNumIndices() > 1)
11318 // Extract the remaining indices out of the constant indexed by the
11320 return ExtractValueInst::Create(V, EV.idx_begin() + 1, EV.idx_end());
11322 return ReplaceInstUsesWith(EV, V);
11324 return 0; // Can't handle other constants
11326 if (InsertValueInst *IV = dyn_cast<InsertValueInst>(Agg)) {
11327 // We're extracting from an insertvalue instruction, compare the indices
11328 const unsigned *exti, *exte, *insi, *inse;
11329 for (exti = EV.idx_begin(), insi = IV->idx_begin(),
11330 exte = EV.idx_end(), inse = IV->idx_end();
11331 exti != exte && insi != inse;
11333 if (*insi != *exti)
11334 // The insert and extract both reference distinctly different elements.
11335 // This means the extract is not influenced by the insert, and we can
11336 // replace the aggregate operand of the extract with the aggregate
11337 // operand of the insert. i.e., replace
11338 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1
11339 // %E = extractvalue { i32, { i32 } } %I, 0
11341 // %E = extractvalue { i32, { i32 } } %A, 0
11342 return ExtractValueInst::Create(IV->getAggregateOperand(),
11343 EV.idx_begin(), EV.idx_end());
11345 if (exti == exte && insi == inse)
11346 // Both iterators are at the end: Index lists are identical. Replace
11347 // %B = insertvalue { i32, { i32 } } %A, i32 42, 1, 0
11348 // %C = extractvalue { i32, { i32 } } %B, 1, 0
11350 return ReplaceInstUsesWith(EV, IV->getInsertedValueOperand());
11351 if (exti == exte) {
11352 // The extract list is a prefix of the insert list. i.e. replace
11353 // %I = insertvalue { i32, { i32 } } %A, i32 42, 1, 0
11354 // %E = extractvalue { i32, { i32 } } %I, 1
11356 // %X = extractvalue { i32, { i32 } } %A, 1
11357 // %E = insertvalue { i32 } %X, i32 42, 0
11358 // by switching the order of the insert and extract (though the
11359 // insertvalue should be left in, since it may have other uses).
11360 Value *NewEV = InsertNewInstBefore(
11361 ExtractValueInst::Create(IV->getAggregateOperand(),
11362 EV.idx_begin(), EV.idx_end()),
11364 return InsertValueInst::Create(NewEV, IV->getInsertedValueOperand(),
11368 // The insert list is a prefix of the extract list
11369 // We can simply remove the common indices from the extract and make it
11370 // operate on the inserted value instead of the insertvalue result.
11372 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1
11373 // %E = extractvalue { i32, { i32 } } %I, 1, 0
11375 // %E extractvalue { i32 } { i32 42 }, 0
11376 return ExtractValueInst::Create(IV->getInsertedValueOperand(),
11379 // Can't simplify extracts from other values. Note that nested extracts are
11380 // already simplified implicitely by the above (extract ( extract (insert) )
11381 // will be translated into extract ( insert ( extract ) ) first and then just
11382 // the value inserted, if appropriate).
11386 /// CheapToScalarize - Return true if the value is cheaper to scalarize than it
11387 /// is to leave as a vector operation.
11388 static bool CheapToScalarize(Value *V, bool isConstant) {
11389 if (isa<ConstantAggregateZero>(V))
11391 if (ConstantVector *C = dyn_cast<ConstantVector>(V)) {
11392 if (isConstant) return true;
11393 // If all elts are the same, we can extract.
11394 Constant *Op0 = C->getOperand(0);
11395 for (unsigned i = 1; i < C->getNumOperands(); ++i)
11396 if (C->getOperand(i) != Op0)
11400 Instruction *I = dyn_cast<Instruction>(V);
11401 if (!I) return false;
11403 // Insert element gets simplified to the inserted element or is deleted if
11404 // this is constant idx extract element and its a constant idx insertelt.
11405 if (I->getOpcode() == Instruction::InsertElement && isConstant &&
11406 isa<ConstantInt>(I->getOperand(2)))
11408 if (I->getOpcode() == Instruction::Load && I->hasOneUse())
11410 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I))
11411 if (BO->hasOneUse() &&
11412 (CheapToScalarize(BO->getOperand(0), isConstant) ||
11413 CheapToScalarize(BO->getOperand(1), isConstant)))
11415 if (CmpInst *CI = dyn_cast<CmpInst>(I))
11416 if (CI->hasOneUse() &&
11417 (CheapToScalarize(CI->getOperand(0), isConstant) ||
11418 CheapToScalarize(CI->getOperand(1), isConstant)))
11424 /// Read and decode a shufflevector mask.
11426 /// It turns undef elements into values that are larger than the number of
11427 /// elements in the input.
11428 static std::vector<unsigned> getShuffleMask(const ShuffleVectorInst *SVI) {
11429 unsigned NElts = SVI->getType()->getNumElements();
11430 if (isa<ConstantAggregateZero>(SVI->getOperand(2)))
11431 return std::vector<unsigned>(NElts, 0);
11432 if (isa<UndefValue>(SVI->getOperand(2)))
11433 return std::vector<unsigned>(NElts, 2*NElts);
11435 std::vector<unsigned> Result;
11436 const ConstantVector *CP = cast<ConstantVector>(SVI->getOperand(2));
11437 for (User::const_op_iterator i = CP->op_begin(), e = CP->op_end(); i!=e; ++i)
11438 if (isa<UndefValue>(*i))
11439 Result.push_back(NElts*2); // undef -> 8
11441 Result.push_back(cast<ConstantInt>(*i)->getZExtValue());
11445 /// FindScalarElement - Given a vector and an element number, see if the scalar
11446 /// value is already around as a register, for example if it were inserted then
11447 /// extracted from the vector.
11448 static Value *FindScalarElement(Value *V, unsigned EltNo) {
11449 assert(isa<VectorType>(V->getType()) && "Not looking at a vector?");
11450 const VectorType *PTy = cast<VectorType>(V->getType());
11451 unsigned Width = PTy->getNumElements();
11452 if (EltNo >= Width) // Out of range access.
11453 return UndefValue::get(PTy->getElementType());
11455 if (isa<UndefValue>(V))
11456 return UndefValue::get(PTy->getElementType());
11457 else if (isa<ConstantAggregateZero>(V))
11458 return Constant::getNullValue(PTy->getElementType());
11459 else if (ConstantVector *CP = dyn_cast<ConstantVector>(V))
11460 return CP->getOperand(EltNo);
11461 else if (InsertElementInst *III = dyn_cast<InsertElementInst>(V)) {
11462 // If this is an insert to a variable element, we don't know what it is.
11463 if (!isa<ConstantInt>(III->getOperand(2)))
11465 unsigned IIElt = cast<ConstantInt>(III->getOperand(2))->getZExtValue();
11467 // If this is an insert to the element we are looking for, return the
11469 if (EltNo == IIElt)
11470 return III->getOperand(1);
11472 // Otherwise, the insertelement doesn't modify the value, recurse on its
11474 return FindScalarElement(III->getOperand(0), EltNo);
11475 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(V)) {
11476 unsigned LHSWidth =
11477 cast<VectorType>(SVI->getOperand(0)->getType())->getNumElements();
11478 unsigned InEl = getShuffleMask(SVI)[EltNo];
11479 if (InEl < LHSWidth)
11480 return FindScalarElement(SVI->getOperand(0), InEl);
11481 else if (InEl < LHSWidth*2)
11482 return FindScalarElement(SVI->getOperand(1), InEl - LHSWidth);
11484 return UndefValue::get(PTy->getElementType());
11487 // Otherwise, we don't know.
11491 Instruction *InstCombiner::visitExtractElementInst(ExtractElementInst &EI) {
11492 // If vector val is undef, replace extract with scalar undef.
11493 if (isa<UndefValue>(EI.getOperand(0)))
11494 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
11496 // If vector val is constant 0, replace extract with scalar 0.
11497 if (isa<ConstantAggregateZero>(EI.getOperand(0)))
11498 return ReplaceInstUsesWith(EI, Constant::getNullValue(EI.getType()));
11500 if (ConstantVector *C = dyn_cast<ConstantVector>(EI.getOperand(0))) {
11501 // If vector val is constant with all elements the same, replace EI with
11502 // that element. When the elements are not identical, we cannot replace yet
11503 // (we do that below, but only when the index is constant).
11504 Constant *op0 = C->getOperand(0);
11505 for (unsigned i = 1; i < C->getNumOperands(); ++i)
11506 if (C->getOperand(i) != op0) {
11511 return ReplaceInstUsesWith(EI, op0);
11514 // If extracting a specified index from the vector, see if we can recursively
11515 // find a previously computed scalar that was inserted into the vector.
11516 if (ConstantInt *IdxC = dyn_cast<ConstantInt>(EI.getOperand(1))) {
11517 unsigned IndexVal = IdxC->getZExtValue();
11518 unsigned VectorWidth =
11519 cast<VectorType>(EI.getOperand(0)->getType())->getNumElements();
11521 // If this is extracting an invalid index, turn this into undef, to avoid
11522 // crashing the code below.
11523 if (IndexVal >= VectorWidth)
11524 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
11526 // This instruction only demands the single element from the input vector.
11527 // If the input vector has a single use, simplify it based on this use
11529 if (EI.getOperand(0)->hasOneUse() && VectorWidth != 1) {
11530 uint64_t UndefElts;
11531 if (Value *V = SimplifyDemandedVectorElts(EI.getOperand(0),
11534 EI.setOperand(0, V);
11539 if (Value *Elt = FindScalarElement(EI.getOperand(0), IndexVal))
11540 return ReplaceInstUsesWith(EI, Elt);
11542 // If the this extractelement is directly using a bitcast from a vector of
11543 // the same number of elements, see if we can find the source element from
11544 // it. In this case, we will end up needing to bitcast the scalars.
11545 if (BitCastInst *BCI = dyn_cast<BitCastInst>(EI.getOperand(0))) {
11546 if (const VectorType *VT =
11547 dyn_cast<VectorType>(BCI->getOperand(0)->getType()))
11548 if (VT->getNumElements() == VectorWidth)
11549 if (Value *Elt = FindScalarElement(BCI->getOperand(0), IndexVal))
11550 return new BitCastInst(Elt, EI.getType());
11554 if (Instruction *I = dyn_cast<Instruction>(EI.getOperand(0))) {
11555 if (I->hasOneUse()) {
11556 // Push extractelement into predecessor operation if legal and
11557 // profitable to do so
11558 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
11559 bool isConstantElt = isa<ConstantInt>(EI.getOperand(1));
11560 if (CheapToScalarize(BO, isConstantElt)) {
11561 ExtractElementInst *newEI0 =
11562 new ExtractElementInst(BO->getOperand(0), EI.getOperand(1),
11563 EI.getName()+".lhs");
11564 ExtractElementInst *newEI1 =
11565 new ExtractElementInst(BO->getOperand(1), EI.getOperand(1),
11566 EI.getName()+".rhs");
11567 InsertNewInstBefore(newEI0, EI);
11568 InsertNewInstBefore(newEI1, EI);
11569 return BinaryOperator::Create(BO->getOpcode(), newEI0, newEI1);
11571 } else if (isa<LoadInst>(I)) {
11573 cast<PointerType>(I->getOperand(0)->getType())->getAddressSpace();
11574 Value *Ptr = InsertBitCastBefore(I->getOperand(0),
11575 PointerType::get(EI.getType(), AS),EI);
11576 GetElementPtrInst *GEP =
11577 GetElementPtrInst::Create(Ptr, EI.getOperand(1), I->getName()+".gep");
11578 InsertNewInstBefore(GEP, EI);
11579 return new LoadInst(GEP);
11582 if (InsertElementInst *IE = dyn_cast<InsertElementInst>(I)) {
11583 // Extracting the inserted element?
11584 if (IE->getOperand(2) == EI.getOperand(1))
11585 return ReplaceInstUsesWith(EI, IE->getOperand(1));
11586 // If the inserted and extracted elements are constants, they must not
11587 // be the same value, extract from the pre-inserted value instead.
11588 if (isa<Constant>(IE->getOperand(2)) &&
11589 isa<Constant>(EI.getOperand(1))) {
11590 AddUsesToWorkList(EI);
11591 EI.setOperand(0, IE->getOperand(0));
11594 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I)) {
11595 // If this is extracting an element from a shufflevector, figure out where
11596 // it came from and extract from the appropriate input element instead.
11597 if (ConstantInt *Elt = dyn_cast<ConstantInt>(EI.getOperand(1))) {
11598 unsigned SrcIdx = getShuffleMask(SVI)[Elt->getZExtValue()];
11600 unsigned LHSWidth =
11601 cast<VectorType>(SVI->getOperand(0)->getType())->getNumElements();
11603 if (SrcIdx < LHSWidth)
11604 Src = SVI->getOperand(0);
11605 else if (SrcIdx < LHSWidth*2) {
11606 SrcIdx -= LHSWidth;
11607 Src = SVI->getOperand(1);
11609 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
11611 return new ExtractElementInst(Src, SrcIdx);
11618 /// CollectSingleShuffleElements - If V is a shuffle of values that ONLY returns
11619 /// elements from either LHS or RHS, return the shuffle mask and true.
11620 /// Otherwise, return false.
11621 static bool CollectSingleShuffleElements(Value *V, Value *LHS, Value *RHS,
11622 std::vector<Constant*> &Mask) {
11623 assert(V->getType() == LHS->getType() && V->getType() == RHS->getType() &&
11624 "Invalid CollectSingleShuffleElements");
11625 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
11627 if (isa<UndefValue>(V)) {
11628 Mask.assign(NumElts, UndefValue::get(Type::Int32Ty));
11630 } else if (V == LHS) {
11631 for (unsigned i = 0; i != NumElts; ++i)
11632 Mask.push_back(ConstantInt::get(Type::Int32Ty, i));
11634 } else if (V == RHS) {
11635 for (unsigned i = 0; i != NumElts; ++i)
11636 Mask.push_back(ConstantInt::get(Type::Int32Ty, i+NumElts));
11638 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
11639 // If this is an insert of an extract from some other vector, include it.
11640 Value *VecOp = IEI->getOperand(0);
11641 Value *ScalarOp = IEI->getOperand(1);
11642 Value *IdxOp = IEI->getOperand(2);
11644 if (!isa<ConstantInt>(IdxOp))
11646 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
11648 if (isa<UndefValue>(ScalarOp)) { // inserting undef into vector.
11649 // Okay, we can handle this if the vector we are insertinting into is
11650 // transitively ok.
11651 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
11652 // If so, update the mask to reflect the inserted undef.
11653 Mask[InsertedIdx] = UndefValue::get(Type::Int32Ty);
11656 } else if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)){
11657 if (isa<ConstantInt>(EI->getOperand(1)) &&
11658 EI->getOperand(0)->getType() == V->getType()) {
11659 unsigned ExtractedIdx =
11660 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
11662 // This must be extracting from either LHS or RHS.
11663 if (EI->getOperand(0) == LHS || EI->getOperand(0) == RHS) {
11664 // Okay, we can handle this if the vector we are insertinting into is
11665 // transitively ok.
11666 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
11667 // If so, update the mask to reflect the inserted value.
11668 if (EI->getOperand(0) == LHS) {
11669 Mask[InsertedIdx % NumElts] =
11670 ConstantInt::get(Type::Int32Ty, ExtractedIdx);
11672 assert(EI->getOperand(0) == RHS);
11673 Mask[InsertedIdx % NumElts] =
11674 ConstantInt::get(Type::Int32Ty, ExtractedIdx+NumElts);
11683 // TODO: Handle shufflevector here!
11688 /// CollectShuffleElements - We are building a shuffle of V, using RHS as the
11689 /// RHS of the shuffle instruction, if it is not null. Return a shuffle mask
11690 /// that computes V and the LHS value of the shuffle.
11691 static Value *CollectShuffleElements(Value *V, std::vector<Constant*> &Mask,
11693 assert(isa<VectorType>(V->getType()) &&
11694 (RHS == 0 || V->getType() == RHS->getType()) &&
11695 "Invalid shuffle!");
11696 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
11698 if (isa<UndefValue>(V)) {
11699 Mask.assign(NumElts, UndefValue::get(Type::Int32Ty));
11701 } else if (isa<ConstantAggregateZero>(V)) {
11702 Mask.assign(NumElts, ConstantInt::get(Type::Int32Ty, 0));
11704 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
11705 // If this is an insert of an extract from some other vector, include it.
11706 Value *VecOp = IEI->getOperand(0);
11707 Value *ScalarOp = IEI->getOperand(1);
11708 Value *IdxOp = IEI->getOperand(2);
11710 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
11711 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
11712 EI->getOperand(0)->getType() == V->getType()) {
11713 unsigned ExtractedIdx =
11714 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
11715 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
11717 // Either the extracted from or inserted into vector must be RHSVec,
11718 // otherwise we'd end up with a shuffle of three inputs.
11719 if (EI->getOperand(0) == RHS || RHS == 0) {
11720 RHS = EI->getOperand(0);
11721 Value *V = CollectShuffleElements(VecOp, Mask, RHS);
11722 Mask[InsertedIdx % NumElts] =
11723 ConstantInt::get(Type::Int32Ty, NumElts+ExtractedIdx);
11727 if (VecOp == RHS) {
11728 Value *V = CollectShuffleElements(EI->getOperand(0), Mask, RHS);
11729 // Everything but the extracted element is replaced with the RHS.
11730 for (unsigned i = 0; i != NumElts; ++i) {
11731 if (i != InsertedIdx)
11732 Mask[i] = ConstantInt::get(Type::Int32Ty, NumElts+i);
11737 // If this insertelement is a chain that comes from exactly these two
11738 // vectors, return the vector and the effective shuffle.
11739 if (CollectSingleShuffleElements(IEI, EI->getOperand(0), RHS, Mask))
11740 return EI->getOperand(0);
11745 // TODO: Handle shufflevector here!
11747 // Otherwise, can't do anything fancy. Return an identity vector.
11748 for (unsigned i = 0; i != NumElts; ++i)
11749 Mask.push_back(ConstantInt::get(Type::Int32Ty, i));
11753 Instruction *InstCombiner::visitInsertElementInst(InsertElementInst &IE) {
11754 Value *VecOp = IE.getOperand(0);
11755 Value *ScalarOp = IE.getOperand(1);
11756 Value *IdxOp = IE.getOperand(2);
11758 // Inserting an undef or into an undefined place, remove this.
11759 if (isa<UndefValue>(ScalarOp) || isa<UndefValue>(IdxOp))
11760 ReplaceInstUsesWith(IE, VecOp);
11762 // If the inserted element was extracted from some other vector, and if the
11763 // indexes are constant, try to turn this into a shufflevector operation.
11764 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
11765 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
11766 EI->getOperand(0)->getType() == IE.getType()) {
11767 unsigned NumVectorElts = IE.getType()->getNumElements();
11768 unsigned ExtractedIdx =
11769 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
11770 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
11772 if (ExtractedIdx >= NumVectorElts) // Out of range extract.
11773 return ReplaceInstUsesWith(IE, VecOp);
11775 if (InsertedIdx >= NumVectorElts) // Out of range insert.
11776 return ReplaceInstUsesWith(IE, UndefValue::get(IE.getType()));
11778 // If we are extracting a value from a vector, then inserting it right
11779 // back into the same place, just use the input vector.
11780 if (EI->getOperand(0) == VecOp && ExtractedIdx == InsertedIdx)
11781 return ReplaceInstUsesWith(IE, VecOp);
11783 // We could theoretically do this for ANY input. However, doing so could
11784 // turn chains of insertelement instructions into a chain of shufflevector
11785 // instructions, and right now we do not merge shufflevectors. As such,
11786 // only do this in a situation where it is clear that there is benefit.
11787 if (isa<UndefValue>(VecOp) || isa<ConstantAggregateZero>(VecOp)) {
11788 // Turn this into shuffle(EIOp0, VecOp, Mask). The result has all of
11789 // the values of VecOp, except then one read from EIOp0.
11790 // Build a new shuffle mask.
11791 std::vector<Constant*> Mask;
11792 if (isa<UndefValue>(VecOp))
11793 Mask.assign(NumVectorElts, UndefValue::get(Type::Int32Ty));
11795 assert(isa<ConstantAggregateZero>(VecOp) && "Unknown thing");
11796 Mask.assign(NumVectorElts, ConstantInt::get(Type::Int32Ty,
11799 Mask[InsertedIdx] = ConstantInt::get(Type::Int32Ty, ExtractedIdx);
11800 return new ShuffleVectorInst(EI->getOperand(0), VecOp,
11801 ConstantVector::get(Mask));
11804 // If this insertelement isn't used by some other insertelement, turn it
11805 // (and any insertelements it points to), into one big shuffle.
11806 if (!IE.hasOneUse() || !isa<InsertElementInst>(IE.use_back())) {
11807 std::vector<Constant*> Mask;
11809 Value *LHS = CollectShuffleElements(&IE, Mask, RHS);
11810 if (RHS == 0) RHS = UndefValue::get(LHS->getType());
11811 // We now have a shuffle of LHS, RHS, Mask.
11812 return new ShuffleVectorInst(LHS, RHS, ConstantVector::get(Mask));
11821 Instruction *InstCombiner::visitShuffleVectorInst(ShuffleVectorInst &SVI) {
11822 Value *LHS = SVI.getOperand(0);
11823 Value *RHS = SVI.getOperand(1);
11824 std::vector<unsigned> Mask = getShuffleMask(&SVI);
11826 bool MadeChange = false;
11828 // Undefined shuffle mask -> undefined value.
11829 if (isa<UndefValue>(SVI.getOperand(2)))
11830 return ReplaceInstUsesWith(SVI, UndefValue::get(SVI.getType()));
11832 uint64_t UndefElts;
11833 unsigned VWidth = cast<VectorType>(SVI.getType())->getNumElements();
11835 if (VWidth != cast<VectorType>(LHS->getType())->getNumElements())
11838 uint64_t AllOnesEltMask = ~0ULL >> (64-VWidth);
11839 if (VWidth <= 64 &&
11840 SimplifyDemandedVectorElts(&SVI, AllOnesEltMask, UndefElts)) {
11841 LHS = SVI.getOperand(0);
11842 RHS = SVI.getOperand(1);
11846 // Canonicalize shuffle(x ,x,mask) -> shuffle(x, undef,mask')
11847 // Canonicalize shuffle(undef,x,mask) -> shuffle(x, undef,mask').
11848 if (LHS == RHS || isa<UndefValue>(LHS)) {
11849 if (isa<UndefValue>(LHS) && LHS == RHS) {
11850 // shuffle(undef,undef,mask) -> undef.
11851 return ReplaceInstUsesWith(SVI, LHS);
11854 // Remap any references to RHS to use LHS.
11855 std::vector<Constant*> Elts;
11856 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
11857 if (Mask[i] >= 2*e)
11858 Elts.push_back(UndefValue::get(Type::Int32Ty));
11860 if ((Mask[i] >= e && isa<UndefValue>(RHS)) ||
11861 (Mask[i] < e && isa<UndefValue>(LHS))) {
11862 Mask[i] = 2*e; // Turn into undef.
11863 Elts.push_back(UndefValue::get(Type::Int32Ty));
11865 Mask[i] = Mask[i] % e; // Force to LHS.
11866 Elts.push_back(ConstantInt::get(Type::Int32Ty, Mask[i]));
11870 SVI.setOperand(0, SVI.getOperand(1));
11871 SVI.setOperand(1, UndefValue::get(RHS->getType()));
11872 SVI.setOperand(2, ConstantVector::get(Elts));
11873 LHS = SVI.getOperand(0);
11874 RHS = SVI.getOperand(1);
11878 // Analyze the shuffle, are the LHS or RHS and identity shuffles?
11879 bool isLHSID = true, isRHSID = true;
11881 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
11882 if (Mask[i] >= e*2) continue; // Ignore undef values.
11883 // Is this an identity shuffle of the LHS value?
11884 isLHSID &= (Mask[i] == i);
11886 // Is this an identity shuffle of the RHS value?
11887 isRHSID &= (Mask[i]-e == i);
11890 // Eliminate identity shuffles.
11891 if (isLHSID) return ReplaceInstUsesWith(SVI, LHS);
11892 if (isRHSID) return ReplaceInstUsesWith(SVI, RHS);
11894 // If the LHS is a shufflevector itself, see if we can combine it with this
11895 // one without producing an unusual shuffle. Here we are really conservative:
11896 // we are absolutely afraid of producing a shuffle mask not in the input
11897 // program, because the code gen may not be smart enough to turn a merged
11898 // shuffle into two specific shuffles: it may produce worse code. As such,
11899 // we only merge two shuffles if the result is one of the two input shuffle
11900 // masks. In this case, merging the shuffles just removes one instruction,
11901 // which we know is safe. This is good for things like turning:
11902 // (splat(splat)) -> splat.
11903 if (ShuffleVectorInst *LHSSVI = dyn_cast<ShuffleVectorInst>(LHS)) {
11904 if (isa<UndefValue>(RHS)) {
11905 std::vector<unsigned> LHSMask = getShuffleMask(LHSSVI);
11907 std::vector<unsigned> NewMask;
11908 for (unsigned i = 0, e = Mask.size(); i != e; ++i)
11909 if (Mask[i] >= 2*e)
11910 NewMask.push_back(2*e);
11912 NewMask.push_back(LHSMask[Mask[i]]);
11914 // If the result mask is equal to the src shuffle or this shuffle mask, do
11915 // the replacement.
11916 if (NewMask == LHSMask || NewMask == Mask) {
11917 std::vector<Constant*> Elts;
11918 for (unsigned i = 0, e = NewMask.size(); i != e; ++i) {
11919 if (NewMask[i] >= e*2) {
11920 Elts.push_back(UndefValue::get(Type::Int32Ty));
11922 Elts.push_back(ConstantInt::get(Type::Int32Ty, NewMask[i]));
11925 return new ShuffleVectorInst(LHSSVI->getOperand(0),
11926 LHSSVI->getOperand(1),
11927 ConstantVector::get(Elts));
11932 return MadeChange ? &SVI : 0;
11938 /// TryToSinkInstruction - Try to move the specified instruction from its
11939 /// current block into the beginning of DestBlock, which can only happen if it's
11940 /// safe to move the instruction past all of the instructions between it and the
11941 /// end of its block.
11942 static bool TryToSinkInstruction(Instruction *I, BasicBlock *DestBlock) {
11943 assert(I->hasOneUse() && "Invariants didn't hold!");
11945 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
11946 if (isa<PHINode>(I) || I->mayWriteToMemory() || isa<TerminatorInst>(I))
11949 // Do not sink alloca instructions out of the entry block.
11950 if (isa<AllocaInst>(I) && I->getParent() ==
11951 &DestBlock->getParent()->getEntryBlock())
11954 // We can only sink load instructions if there is nothing between the load and
11955 // the end of block that could change the value.
11956 if (I->mayReadFromMemory()) {
11957 for (BasicBlock::iterator Scan = I, E = I->getParent()->end();
11959 if (Scan->mayWriteToMemory())
11963 BasicBlock::iterator InsertPos = DestBlock->getFirstNonPHI();
11965 I->moveBefore(InsertPos);
11971 /// AddReachableCodeToWorklist - Walk the function in depth-first order, adding
11972 /// all reachable code to the worklist.
11974 /// This has a couple of tricks to make the code faster and more powerful. In
11975 /// particular, we constant fold and DCE instructions as we go, to avoid adding
11976 /// them to the worklist (this significantly speeds up instcombine on code where
11977 /// many instructions are dead or constant). Additionally, if we find a branch
11978 /// whose condition is a known constant, we only visit the reachable successors.
11980 static void AddReachableCodeToWorklist(BasicBlock *BB,
11981 SmallPtrSet<BasicBlock*, 64> &Visited,
11983 const TargetData *TD) {
11984 SmallVector<BasicBlock*, 256> Worklist;
11985 Worklist.push_back(BB);
11987 while (!Worklist.empty()) {
11988 BB = Worklist.back();
11989 Worklist.pop_back();
11991 // We have now visited this block! If we've already been here, ignore it.
11992 if (!Visited.insert(BB)) continue;
11994 DbgInfoIntrinsic *DBI_Prev = NULL;
11995 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) {
11996 Instruction *Inst = BBI++;
11998 // DCE instruction if trivially dead.
11999 if (isInstructionTriviallyDead(Inst)) {
12001 DOUT << "IC: DCE: " << *Inst;
12002 Inst->eraseFromParent();
12006 // ConstantProp instruction if trivially constant.
12007 if (Constant *C = ConstantFoldInstruction(Inst, TD)) {
12008 DOUT << "IC: ConstFold to: " << *C << " from: " << *Inst;
12009 Inst->replaceAllUsesWith(C);
12011 Inst->eraseFromParent();
12015 // If there are two consecutive llvm.dbg.stoppoint calls then
12016 // it is likely that the optimizer deleted code in between these
12018 DbgInfoIntrinsic *DBI_Next = dyn_cast<DbgInfoIntrinsic>(Inst);
12021 && DBI_Prev->getIntrinsicID() == llvm::Intrinsic::dbg_stoppoint
12022 && DBI_Next->getIntrinsicID() == llvm::Intrinsic::dbg_stoppoint) {
12023 IC.RemoveFromWorkList(DBI_Prev);
12024 DBI_Prev->eraseFromParent();
12026 DBI_Prev = DBI_Next;
12029 IC.AddToWorkList(Inst);
12032 // Recursively visit successors. If this is a branch or switch on a
12033 // constant, only visit the reachable successor.
12034 TerminatorInst *TI = BB->getTerminator();
12035 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
12036 if (BI->isConditional() && isa<ConstantInt>(BI->getCondition())) {
12037 bool CondVal = cast<ConstantInt>(BI->getCondition())->getZExtValue();
12038 BasicBlock *ReachableBB = BI->getSuccessor(!CondVal);
12039 Worklist.push_back(ReachableBB);
12042 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
12043 if (ConstantInt *Cond = dyn_cast<ConstantInt>(SI->getCondition())) {
12044 // See if this is an explicit destination.
12045 for (unsigned i = 1, e = SI->getNumSuccessors(); i != e; ++i)
12046 if (SI->getCaseValue(i) == Cond) {
12047 BasicBlock *ReachableBB = SI->getSuccessor(i);
12048 Worklist.push_back(ReachableBB);
12052 // Otherwise it is the default destination.
12053 Worklist.push_back(SI->getSuccessor(0));
12058 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
12059 Worklist.push_back(TI->getSuccessor(i));
12063 bool InstCombiner::DoOneIteration(Function &F, unsigned Iteration) {
12064 bool Changed = false;
12065 TD = &getAnalysis<TargetData>();
12067 DEBUG(DOUT << "\n\nINSTCOMBINE ITERATION #" << Iteration << " on "
12068 << F.getNameStr() << "\n");
12071 // Do a depth-first traversal of the function, populate the worklist with
12072 // the reachable instructions. Ignore blocks that are not reachable. Keep
12073 // track of which blocks we visit.
12074 SmallPtrSet<BasicBlock*, 64> Visited;
12075 AddReachableCodeToWorklist(F.begin(), Visited, *this, TD);
12077 // Do a quick scan over the function. If we find any blocks that are
12078 // unreachable, remove any instructions inside of them. This prevents
12079 // the instcombine code from having to deal with some bad special cases.
12080 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
12081 if (!Visited.count(BB)) {
12082 Instruction *Term = BB->getTerminator();
12083 while (Term != BB->begin()) { // Remove instrs bottom-up
12084 BasicBlock::iterator I = Term; --I;
12086 DOUT << "IC: DCE: " << *I;
12089 if (!I->use_empty())
12090 I->replaceAllUsesWith(UndefValue::get(I->getType()));
12091 I->eraseFromParent();
12096 while (!Worklist.empty()) {
12097 Instruction *I = RemoveOneFromWorkList();
12098 if (I == 0) continue; // skip null values.
12100 // Check to see if we can DCE the instruction.
12101 if (isInstructionTriviallyDead(I)) {
12102 // Add operands to the worklist.
12103 if (I->getNumOperands() < 4)
12104 AddUsesToWorkList(*I);
12107 DOUT << "IC: DCE: " << *I;
12109 I->eraseFromParent();
12110 RemoveFromWorkList(I);
12114 // Instruction isn't dead, see if we can constant propagate it.
12115 if (Constant *C = ConstantFoldInstruction(I, TD)) {
12116 DOUT << "IC: ConstFold to: " << *C << " from: " << *I;
12118 // Add operands to the worklist.
12119 AddUsesToWorkList(*I);
12120 ReplaceInstUsesWith(*I, C);
12123 I->eraseFromParent();
12124 RemoveFromWorkList(I);
12128 if (TD && I->getType()->getTypeID() == Type::VoidTyID) {
12129 // See if we can constant fold its operands.
12130 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i) {
12131 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(i)) {
12132 if (Constant *NewC = ConstantFoldConstantExpression(CE, TD))
12138 // See if we can trivially sink this instruction to a successor basic block.
12139 if (I->hasOneUse()) {
12140 BasicBlock *BB = I->getParent();
12141 BasicBlock *UserParent = cast<Instruction>(I->use_back())->getParent();
12142 if (UserParent != BB) {
12143 bool UserIsSuccessor = false;
12144 // See if the user is one of our successors.
12145 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI)
12146 if (*SI == UserParent) {
12147 UserIsSuccessor = true;
12151 // If the user is one of our immediate successors, and if that successor
12152 // only has us as a predecessors (we'd have to split the critical edge
12153 // otherwise), we can keep going.
12154 if (UserIsSuccessor && !isa<PHINode>(I->use_back()) &&
12155 next(pred_begin(UserParent)) == pred_end(UserParent))
12156 // Okay, the CFG is simple enough, try to sink this instruction.
12157 Changed |= TryToSinkInstruction(I, UserParent);
12161 // Now that we have an instruction, try combining it to simplify it...
12165 DEBUG(std::ostringstream SS; I->print(SS); OrigI = SS.str(););
12166 if (Instruction *Result = visit(*I)) {
12168 // Should we replace the old instruction with a new one?
12170 DOUT << "IC: Old = " << *I
12171 << " New = " << *Result;
12173 // Everything uses the new instruction now.
12174 I->replaceAllUsesWith(Result);
12176 // Push the new instruction and any users onto the worklist.
12177 AddToWorkList(Result);
12178 AddUsersToWorkList(*Result);
12180 // Move the name to the new instruction first.
12181 Result->takeName(I);
12183 // Insert the new instruction into the basic block...
12184 BasicBlock *InstParent = I->getParent();
12185 BasicBlock::iterator InsertPos = I;
12187 if (!isa<PHINode>(Result)) // If combining a PHI, don't insert
12188 while (isa<PHINode>(InsertPos)) // middle of a block of PHIs.
12191 InstParent->getInstList().insert(InsertPos, Result);
12193 // Make sure that we reprocess all operands now that we reduced their
12195 AddUsesToWorkList(*I);
12197 // Instructions can end up on the worklist more than once. Make sure
12198 // we do not process an instruction that has been deleted.
12199 RemoveFromWorkList(I);
12201 // Erase the old instruction.
12202 InstParent->getInstList().erase(I);
12205 DOUT << "IC: Mod = " << OrigI
12206 << " New = " << *I;
12209 // If the instruction was modified, it's possible that it is now dead.
12210 // if so, remove it.
12211 if (isInstructionTriviallyDead(I)) {
12212 // Make sure we process all operands now that we are reducing their
12214 AddUsesToWorkList(*I);
12216 // Instructions may end up in the worklist more than once. Erase all
12217 // occurrences of this instruction.
12218 RemoveFromWorkList(I);
12219 I->eraseFromParent();
12222 AddUsersToWorkList(*I);
12229 assert(WorklistMap.empty() && "Worklist empty, but map not?");
12231 // Do an explicit clear, this shrinks the map if needed.
12232 WorklistMap.clear();
12237 bool InstCombiner::runOnFunction(Function &F) {
12238 MustPreserveLCSSA = mustPreserveAnalysisID(LCSSAID);
12240 bool EverMadeChange = false;
12242 // Iterate while there is work to do.
12243 unsigned Iteration = 0;
12244 while (DoOneIteration(F, Iteration++))
12245 EverMadeChange = true;
12246 return EverMadeChange;
12249 FunctionPass *llvm::createInstructionCombiningPass() {
12250 return new InstCombiner();