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 if (ConstantInt *C = dyn_cast<ConstantInt>(Op1)) {
2886 // X udiv C^2 -> X >> C
2887 // Check to see if this is an unsigned division with an exact power of 2,
2888 // if so, convert to a right shift.
2889 if (C->getValue().isPowerOf2()) // 0 not included in isPowerOf2
2890 return BinaryOperator::CreateLShr(Op0,
2891 ConstantInt::get(Op0->getType(), C->getValue().logBase2()));
2893 // X udiv C, where C >= signbit
2894 if (C->getValue().isNegative()) {
2895 Value *IC = InsertNewInstBefore(new ICmpInst(ICmpInst::ICMP_ULT, Op0, C),
2897 return SelectInst::Create(IC, Constant::getNullValue(I.getType()),
2898 ConstantInt::get(I.getType(), 1));
2902 // X udiv (C1 << N), where C1 is "1<<C2" --> X >> (N+C2)
2903 if (BinaryOperator *RHSI = dyn_cast<BinaryOperator>(I.getOperand(1))) {
2904 if (RHSI->getOpcode() == Instruction::Shl &&
2905 isa<ConstantInt>(RHSI->getOperand(0))) {
2906 const APInt& C1 = cast<ConstantInt>(RHSI->getOperand(0))->getValue();
2907 if (C1.isPowerOf2()) {
2908 Value *N = RHSI->getOperand(1);
2909 const Type *NTy = N->getType();
2910 if (uint32_t C2 = C1.logBase2()) {
2911 Constant *C2V = ConstantInt::get(NTy, C2);
2912 N = InsertNewInstBefore(BinaryOperator::CreateAdd(N, C2V, "tmp"), I);
2914 return BinaryOperator::CreateLShr(Op0, N);
2919 // udiv X, (Select Cond, C1, C2) --> Select Cond, (shr X, C1), (shr X, C2)
2920 // where C1&C2 are powers of two.
2921 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
2922 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
2923 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
2924 const APInt &TVA = STO->getValue(), &FVA = SFO->getValue();
2925 if (TVA.isPowerOf2() && FVA.isPowerOf2()) {
2926 // Compute the shift amounts
2927 uint32_t TSA = TVA.logBase2(), FSA = FVA.logBase2();
2928 // Construct the "on true" case of the select
2929 Constant *TC = ConstantInt::get(Op0->getType(), TSA);
2930 Instruction *TSI = BinaryOperator::CreateLShr(
2931 Op0, TC, SI->getName()+".t");
2932 TSI = InsertNewInstBefore(TSI, I);
2934 // Construct the "on false" case of the select
2935 Constant *FC = ConstantInt::get(Op0->getType(), FSA);
2936 Instruction *FSI = BinaryOperator::CreateLShr(
2937 Op0, FC, SI->getName()+".f");
2938 FSI = InsertNewInstBefore(FSI, I);
2940 // construct the select instruction and return it.
2941 return SelectInst::Create(SI->getOperand(0), TSI, FSI, SI->getName());
2947 Instruction *InstCombiner::visitSDiv(BinaryOperator &I) {
2948 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2950 // Handle the integer div common cases
2951 if (Instruction *Common = commonIDivTransforms(I))
2954 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2956 if (RHS->isAllOnesValue())
2957 return BinaryOperator::CreateNeg(Op0);
2959 ConstantInt *RHSNeg = cast<ConstantInt>(ConstantExpr::getNeg(RHS));
2960 APInt RHSNegAPI(RHSNeg->getBitWidth(), RHSNeg->getSExtValue(), true);
2962 APInt NegOne = -APInt(RHSNeg->getBitWidth(), 1, true);
2963 APInt TwoToExp(RHSNeg->getBitWidth(), 1 << (RHSNeg->getBitWidth() - 1),
2966 // -X/C -> X/-C, if and only if negation doesn't overflow.
2967 if ((RHS->getSExtValue() < 0 && RHSNegAPI.slt(TwoToExp - 1)) ||
2968 (RHS->getSExtValue() > 0 && RHSNegAPI.sgt(TwoToExp * NegOne))) {
2969 if (Value *LHSNeg = dyn_castNegVal(Op0)) {
2970 if (ConstantInt *CI = dyn_cast<ConstantInt>(LHSNeg)) {
2971 ConstantInt *CINeg = cast<ConstantInt>(ConstantExpr::getNeg(CI));
2972 APInt CINegAPI(CINeg->getBitWidth(), CINeg->getSExtValue(), true);
2974 if ((CI->getSExtValue() < 0 && CINegAPI.slt(TwoToExp - 1)) ||
2975 (CI->getSExtValue() > 0 && CINegAPI.sgt(TwoToExp * NegOne)))
2976 return BinaryOperator::CreateSDiv(LHSNeg,
2977 ConstantExpr::getNeg(RHS));
2983 // If the sign bits of both operands are zero (i.e. we can prove they are
2984 // unsigned inputs), turn this into a udiv.
2985 if (I.getType()->isInteger()) {
2986 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()));
2987 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
2988 // X sdiv Y -> X udiv Y, iff X and Y don't have sign bit set
2989 return BinaryOperator::CreateUDiv(Op0, Op1, I.getName());
2996 Instruction *InstCombiner::visitFDiv(BinaryOperator &I) {
2997 return commonDivTransforms(I);
3000 /// This function implements the transforms on rem instructions that work
3001 /// regardless of the kind of rem instruction it is (urem, srem, or frem). It
3002 /// is used by the visitors to those instructions.
3003 /// @brief Transforms common to all three rem instructions
3004 Instruction *InstCombiner::commonRemTransforms(BinaryOperator &I) {
3005 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3007 // 0 % X == 0 for integer, we don't need to preserve faults!
3008 if (Constant *LHS = dyn_cast<Constant>(Op0))
3009 if (LHS->isNullValue())
3010 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3012 if (isa<UndefValue>(Op0)) { // undef % X -> 0
3013 if (I.getType()->isFPOrFPVector())
3014 return ReplaceInstUsesWith(I, Op0); // X % undef -> undef (could be SNaN)
3015 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3017 if (isa<UndefValue>(Op1))
3018 return ReplaceInstUsesWith(I, Op1); // X % undef -> undef
3020 // Handle cases involving: rem X, (select Cond, Y, Z)
3021 if (isa<SelectInst>(Op1) && SimplifyDivRemOfSelect(I))
3027 /// This function implements the transforms common to both integer remainder
3028 /// instructions (urem and srem). It is called by the visitors to those integer
3029 /// remainder instructions.
3030 /// @brief Common integer remainder transforms
3031 Instruction *InstCombiner::commonIRemTransforms(BinaryOperator &I) {
3032 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3034 if (Instruction *common = commonRemTransforms(I))
3037 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3038 // X % 0 == undef, we don't need to preserve faults!
3039 if (RHS->equalsInt(0))
3040 return ReplaceInstUsesWith(I, UndefValue::get(I.getType()));
3042 if (RHS->equalsInt(1)) // X % 1 == 0
3043 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3045 if (Instruction *Op0I = dyn_cast<Instruction>(Op0)) {
3046 if (SelectInst *SI = dyn_cast<SelectInst>(Op0I)) {
3047 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3049 } else if (isa<PHINode>(Op0I)) {
3050 if (Instruction *NV = FoldOpIntoPhi(I))
3054 // See if we can fold away this rem instruction.
3055 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
3056 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
3057 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
3058 KnownZero, KnownOne))
3066 Instruction *InstCombiner::visitURem(BinaryOperator &I) {
3067 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3069 if (Instruction *common = commonIRemTransforms(I))
3072 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3073 // X urem C^2 -> X and C
3074 // Check to see if this is an unsigned remainder with an exact power of 2,
3075 // if so, convert to a bitwise and.
3076 if (ConstantInt *C = dyn_cast<ConstantInt>(RHS))
3077 if (C->getValue().isPowerOf2())
3078 return BinaryOperator::CreateAnd(Op0, SubOne(C));
3081 if (Instruction *RHSI = dyn_cast<Instruction>(I.getOperand(1))) {
3082 // Turn A % (C << N), where C is 2^k, into A & ((C << N)-1)
3083 if (RHSI->getOpcode() == Instruction::Shl &&
3084 isa<ConstantInt>(RHSI->getOperand(0))) {
3085 if (cast<ConstantInt>(RHSI->getOperand(0))->getValue().isPowerOf2()) {
3086 Constant *N1 = ConstantInt::getAllOnesValue(I.getType());
3087 Value *Add = InsertNewInstBefore(BinaryOperator::CreateAdd(RHSI, N1,
3089 return BinaryOperator::CreateAnd(Op0, Add);
3094 // urem X, (select Cond, 2^C1, 2^C2) --> select Cond, (and X, C1), (and X, C2)
3095 // where C1&C2 are powers of two.
3096 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
3097 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
3098 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
3099 // STO == 0 and SFO == 0 handled above.
3100 if ((STO->getValue().isPowerOf2()) &&
3101 (SFO->getValue().isPowerOf2())) {
3102 Value *TrueAnd = InsertNewInstBefore(
3103 BinaryOperator::CreateAnd(Op0, SubOne(STO), SI->getName()+".t"), I);
3104 Value *FalseAnd = InsertNewInstBefore(
3105 BinaryOperator::CreateAnd(Op0, SubOne(SFO), SI->getName()+".f"), I);
3106 return SelectInst::Create(SI->getOperand(0), TrueAnd, FalseAnd);
3114 Instruction *InstCombiner::visitSRem(BinaryOperator &I) {
3115 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3117 // Handle the integer rem common cases
3118 if (Instruction *common = commonIRemTransforms(I))
3121 if (Value *RHSNeg = dyn_castNegVal(Op1))
3122 if (!isa<Constant>(RHSNeg) ||
3123 (isa<ConstantInt>(RHSNeg) &&
3124 cast<ConstantInt>(RHSNeg)->getValue().isStrictlyPositive())) {
3126 AddUsesToWorkList(I);
3127 I.setOperand(1, RHSNeg);
3131 // If the sign bits of both operands are zero (i.e. we can prove they are
3132 // unsigned inputs), turn this into a urem.
3133 if (I.getType()->isInteger()) {
3134 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()));
3135 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
3136 // X srem Y -> X urem Y, iff X and Y don't have sign bit set
3137 return BinaryOperator::CreateURem(Op0, Op1, I.getName());
3144 Instruction *InstCombiner::visitFRem(BinaryOperator &I) {
3145 return commonRemTransforms(I);
3148 // isOneBitSet - Return true if there is exactly one bit set in the specified
3150 static bool isOneBitSet(const ConstantInt *CI) {
3151 return CI->getValue().isPowerOf2();
3154 // isHighOnes - Return true if the constant is of the form 1+0+.
3155 // This is the same as lowones(~X).
3156 static bool isHighOnes(const ConstantInt *CI) {
3157 return (~CI->getValue() + 1).isPowerOf2();
3160 /// getICmpCode - Encode a icmp predicate into a three bit mask. These bits
3161 /// are carefully arranged to allow folding of expressions such as:
3163 /// (A < B) | (A > B) --> (A != B)
3165 /// Note that this is only valid if the first and second predicates have the
3166 /// same sign. Is illegal to do: (A u< B) | (A s> B)
3168 /// Three bits are used to represent the condition, as follows:
3173 /// <=> Value Definition
3174 /// 000 0 Always false
3181 /// 111 7 Always true
3183 static unsigned getICmpCode(const ICmpInst *ICI) {
3184 switch (ICI->getPredicate()) {
3186 case ICmpInst::ICMP_UGT: return 1; // 001
3187 case ICmpInst::ICMP_SGT: return 1; // 001
3188 case ICmpInst::ICMP_EQ: return 2; // 010
3189 case ICmpInst::ICMP_UGE: return 3; // 011
3190 case ICmpInst::ICMP_SGE: return 3; // 011
3191 case ICmpInst::ICMP_ULT: return 4; // 100
3192 case ICmpInst::ICMP_SLT: return 4; // 100
3193 case ICmpInst::ICMP_NE: return 5; // 101
3194 case ICmpInst::ICMP_ULE: return 6; // 110
3195 case ICmpInst::ICMP_SLE: return 6; // 110
3198 assert(0 && "Invalid ICmp predicate!");
3203 /// getFCmpCode - Similar to getICmpCode but for FCmpInst. This encodes a fcmp
3204 /// predicate into a three bit mask. It also returns whether it is an ordered
3205 /// predicate by reference.
3206 static unsigned getFCmpCode(FCmpInst::Predicate CC, bool &isOrdered) {
3209 case FCmpInst::FCMP_ORD: isOrdered = true; return 0; // 000
3210 case FCmpInst::FCMP_UNO: return 0; // 000
3211 case FCmpInst::FCMP_OGT: isOrdered = true; return 1; // 001
3212 case FCmpInst::FCMP_UGT: return 1; // 001
3213 case FCmpInst::FCMP_OEQ: isOrdered = true; return 2; // 010
3214 case FCmpInst::FCMP_UEQ: return 2; // 010
3215 case FCmpInst::FCMP_OGE: isOrdered = true; return 3; // 011
3216 case FCmpInst::FCMP_UGE: return 3; // 011
3217 case FCmpInst::FCMP_OLT: isOrdered = true; return 4; // 100
3218 case FCmpInst::FCMP_ULT: return 4; // 100
3219 case FCmpInst::FCMP_ONE: isOrdered = true; return 5; // 101
3220 case FCmpInst::FCMP_UNE: return 5; // 101
3221 case FCmpInst::FCMP_OLE: isOrdered = true; return 6; // 110
3222 case FCmpInst::FCMP_ULE: return 6; // 110
3225 // Not expecting FCMP_FALSE and FCMP_TRUE;
3226 assert(0 && "Unexpected FCmp predicate!");
3231 /// getICmpValue - This is the complement of getICmpCode, which turns an
3232 /// opcode and two operands into either a constant true or false, or a brand
3233 /// new ICmp instruction. The sign is passed in to determine which kind
3234 /// of predicate to use in the new icmp instruction.
3235 static Value *getICmpValue(bool sign, unsigned code, Value *LHS, Value *RHS) {
3237 default: assert(0 && "Illegal ICmp code!");
3238 case 0: return ConstantInt::getFalse();
3241 return new ICmpInst(ICmpInst::ICMP_SGT, LHS, RHS);
3243 return new ICmpInst(ICmpInst::ICMP_UGT, LHS, RHS);
3244 case 2: return new ICmpInst(ICmpInst::ICMP_EQ, LHS, RHS);
3247 return new ICmpInst(ICmpInst::ICMP_SGE, LHS, RHS);
3249 return new ICmpInst(ICmpInst::ICMP_UGE, LHS, RHS);
3252 return new ICmpInst(ICmpInst::ICMP_SLT, LHS, RHS);
3254 return new ICmpInst(ICmpInst::ICMP_ULT, LHS, RHS);
3255 case 5: return new ICmpInst(ICmpInst::ICMP_NE, LHS, RHS);
3258 return new ICmpInst(ICmpInst::ICMP_SLE, LHS, RHS);
3260 return new ICmpInst(ICmpInst::ICMP_ULE, LHS, RHS);
3261 case 7: return ConstantInt::getTrue();
3265 /// getFCmpValue - This is the complement of getFCmpCode, which turns an
3266 /// opcode and two operands into either a FCmp instruction. isordered is passed
3267 /// in to determine which kind of predicate to use in the new fcmp instruction.
3268 static Value *getFCmpValue(bool isordered, unsigned code,
3269 Value *LHS, Value *RHS) {
3271 default: assert(0 && "Illegal FCmp code!");
3274 return new FCmpInst(FCmpInst::FCMP_ORD, LHS, RHS);
3276 return new FCmpInst(FCmpInst::FCMP_UNO, LHS, RHS);
3279 return new FCmpInst(FCmpInst::FCMP_OGT, LHS, RHS);
3281 return new FCmpInst(FCmpInst::FCMP_UGT, LHS, RHS);
3284 return new FCmpInst(FCmpInst::FCMP_OEQ, LHS, RHS);
3286 return new FCmpInst(FCmpInst::FCMP_UEQ, LHS, RHS);
3289 return new FCmpInst(FCmpInst::FCMP_OGE, LHS, RHS);
3291 return new FCmpInst(FCmpInst::FCMP_UGE, LHS, RHS);
3294 return new FCmpInst(FCmpInst::FCMP_OLT, LHS, RHS);
3296 return new FCmpInst(FCmpInst::FCMP_ULT, LHS, RHS);
3299 return new FCmpInst(FCmpInst::FCMP_ONE, LHS, RHS);
3301 return new FCmpInst(FCmpInst::FCMP_UNE, LHS, RHS);
3304 return new FCmpInst(FCmpInst::FCMP_OLE, LHS, RHS);
3306 return new FCmpInst(FCmpInst::FCMP_ULE, LHS, RHS);
3307 case 7: return ConstantInt::getTrue();
3311 /// PredicatesFoldable - Return true if both predicates match sign or if at
3312 /// least one of them is an equality comparison (which is signless).
3313 static bool PredicatesFoldable(ICmpInst::Predicate p1, ICmpInst::Predicate p2) {
3314 return (ICmpInst::isSignedPredicate(p1) == ICmpInst::isSignedPredicate(p2)) ||
3315 (ICmpInst::isSignedPredicate(p1) && ICmpInst::isEquality(p2)) ||
3316 (ICmpInst::isSignedPredicate(p2) && ICmpInst::isEquality(p1));
3320 // FoldICmpLogical - Implements (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
3321 struct FoldICmpLogical {
3324 ICmpInst::Predicate pred;
3325 FoldICmpLogical(InstCombiner &ic, ICmpInst *ICI)
3326 : IC(ic), LHS(ICI->getOperand(0)), RHS(ICI->getOperand(1)),
3327 pred(ICI->getPredicate()) {}
3328 bool shouldApply(Value *V) const {
3329 if (ICmpInst *ICI = dyn_cast<ICmpInst>(V))
3330 if (PredicatesFoldable(pred, ICI->getPredicate()))
3331 return ((ICI->getOperand(0) == LHS && ICI->getOperand(1) == RHS) ||
3332 (ICI->getOperand(0) == RHS && ICI->getOperand(1) == LHS));
3335 Instruction *apply(Instruction &Log) const {
3336 ICmpInst *ICI = cast<ICmpInst>(Log.getOperand(0));
3337 if (ICI->getOperand(0) != LHS) {
3338 assert(ICI->getOperand(1) == LHS);
3339 ICI->swapOperands(); // Swap the LHS and RHS of the ICmp
3342 ICmpInst *RHSICI = cast<ICmpInst>(Log.getOperand(1));
3343 unsigned LHSCode = getICmpCode(ICI);
3344 unsigned RHSCode = getICmpCode(RHSICI);
3346 switch (Log.getOpcode()) {
3347 case Instruction::And: Code = LHSCode & RHSCode; break;
3348 case Instruction::Or: Code = LHSCode | RHSCode; break;
3349 case Instruction::Xor: Code = LHSCode ^ RHSCode; break;
3350 default: assert(0 && "Illegal logical opcode!"); return 0;
3353 bool isSigned = ICmpInst::isSignedPredicate(RHSICI->getPredicate()) ||
3354 ICmpInst::isSignedPredicate(ICI->getPredicate());
3356 Value *RV = getICmpValue(isSigned, Code, LHS, RHS);
3357 if (Instruction *I = dyn_cast<Instruction>(RV))
3359 // Otherwise, it's a constant boolean value...
3360 return IC.ReplaceInstUsesWith(Log, RV);
3363 } // end anonymous namespace
3365 // OptAndOp - This handles expressions of the form ((val OP C1) & C2). Where
3366 // the Op parameter is 'OP', OpRHS is 'C1', and AndRHS is 'C2'. Op is
3367 // guaranteed to be a binary operator.
3368 Instruction *InstCombiner::OptAndOp(Instruction *Op,
3370 ConstantInt *AndRHS,
3371 BinaryOperator &TheAnd) {
3372 Value *X = Op->getOperand(0);
3373 Constant *Together = 0;
3375 Together = And(AndRHS, OpRHS);
3377 switch (Op->getOpcode()) {
3378 case Instruction::Xor:
3379 if (Op->hasOneUse()) {
3380 // (X ^ C1) & C2 --> (X & C2) ^ (C1&C2)
3381 Instruction *And = BinaryOperator::CreateAnd(X, AndRHS);
3382 InsertNewInstBefore(And, TheAnd);
3384 return BinaryOperator::CreateXor(And, Together);
3387 case Instruction::Or:
3388 if (Together == AndRHS) // (X | C) & C --> C
3389 return ReplaceInstUsesWith(TheAnd, AndRHS);
3391 if (Op->hasOneUse() && Together != OpRHS) {
3392 // (X | C1) & C2 --> (X | (C1&C2)) & C2
3393 Instruction *Or = BinaryOperator::CreateOr(X, Together);
3394 InsertNewInstBefore(Or, TheAnd);
3396 return BinaryOperator::CreateAnd(Or, AndRHS);
3399 case Instruction::Add:
3400 if (Op->hasOneUse()) {
3401 // Adding a one to a single bit bit-field should be turned into an XOR
3402 // of the bit. First thing to check is to see if this AND is with a
3403 // single bit constant.
3404 const APInt& AndRHSV = cast<ConstantInt>(AndRHS)->getValue();
3406 // If there is only one bit set...
3407 if (isOneBitSet(cast<ConstantInt>(AndRHS))) {
3408 // Ok, at this point, we know that we are masking the result of the
3409 // ADD down to exactly one bit. If the constant we are adding has
3410 // no bits set below this bit, then we can eliminate the ADD.
3411 const APInt& AddRHS = cast<ConstantInt>(OpRHS)->getValue();
3413 // Check to see if any bits below the one bit set in AndRHSV are set.
3414 if ((AddRHS & (AndRHSV-1)) == 0) {
3415 // If not, the only thing that can effect the output of the AND is
3416 // the bit specified by AndRHSV. If that bit is set, the effect of
3417 // the XOR is to toggle the bit. If it is clear, then the ADD has
3419 if ((AddRHS & AndRHSV) == 0) { // Bit is not set, noop
3420 TheAnd.setOperand(0, X);
3423 // Pull the XOR out of the AND.
3424 Instruction *NewAnd = BinaryOperator::CreateAnd(X, AndRHS);
3425 InsertNewInstBefore(NewAnd, TheAnd);
3426 NewAnd->takeName(Op);
3427 return BinaryOperator::CreateXor(NewAnd, AndRHS);
3434 case Instruction::Shl: {
3435 // We know that the AND will not produce any of the bits shifted in, so if
3436 // the anded constant includes them, clear them now!
3438 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3439 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3440 APInt ShlMask(APInt::getHighBitsSet(BitWidth, BitWidth-OpRHSVal));
3441 ConstantInt *CI = ConstantInt::get(AndRHS->getValue() & ShlMask);
3443 if (CI->getValue() == ShlMask) {
3444 // Masking out bits that the shift already masks
3445 return ReplaceInstUsesWith(TheAnd, Op); // No need for the and.
3446 } else if (CI != AndRHS) { // Reducing bits set in and.
3447 TheAnd.setOperand(1, CI);
3452 case Instruction::LShr:
3454 // We know that the AND will not produce any of the bits shifted in, so if
3455 // the anded constant includes them, clear them now! This only applies to
3456 // unsigned shifts, because a signed shr may bring in set bits!
3458 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3459 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3460 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
3461 ConstantInt *CI = ConstantInt::get(AndRHS->getValue() & ShrMask);
3463 if (CI->getValue() == ShrMask) {
3464 // Masking out bits that the shift already masks.
3465 return ReplaceInstUsesWith(TheAnd, Op);
3466 } else if (CI != AndRHS) {
3467 TheAnd.setOperand(1, CI); // Reduce bits set in and cst.
3472 case Instruction::AShr:
3474 // See if this is shifting in some sign extension, then masking it out
3476 if (Op->hasOneUse()) {
3477 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3478 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3479 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
3480 Constant *C = ConstantInt::get(AndRHS->getValue() & ShrMask);
3481 if (C == AndRHS) { // Masking out bits shifted in.
3482 // (Val ashr C1) & C2 -> (Val lshr C1) & C2
3483 // Make the argument unsigned.
3484 Value *ShVal = Op->getOperand(0);
3485 ShVal = InsertNewInstBefore(
3486 BinaryOperator::CreateLShr(ShVal, OpRHS,
3487 Op->getName()), TheAnd);
3488 return BinaryOperator::CreateAnd(ShVal, AndRHS, TheAnd.getName());
3497 /// InsertRangeTest - Emit a computation of: (V >= Lo && V < Hi) if Inside is
3498 /// true, otherwise (V < Lo || V >= Hi). In pratice, we emit the more efficient
3499 /// (V-Lo) <u Hi-Lo. This method expects that Lo <= Hi. isSigned indicates
3500 /// whether to treat the V, Lo and HI as signed or not. IB is the location to
3501 /// insert new instructions.
3502 Instruction *InstCombiner::InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
3503 bool isSigned, bool Inside,
3505 assert(cast<ConstantInt>(ConstantExpr::getICmp((isSigned ?
3506 ICmpInst::ICMP_SLE:ICmpInst::ICMP_ULE), Lo, Hi))->getZExtValue() &&
3507 "Lo is not <= Hi in range emission code!");
3510 if (Lo == Hi) // Trivially false.
3511 return new ICmpInst(ICmpInst::ICMP_NE, V, V);
3513 // V >= Min && V < Hi --> V < Hi
3514 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
3515 ICmpInst::Predicate pred = (isSigned ?
3516 ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT);
3517 return new ICmpInst(pred, V, Hi);
3520 // Emit V-Lo <u Hi-Lo
3521 Constant *NegLo = ConstantExpr::getNeg(Lo);
3522 Instruction *Add = BinaryOperator::CreateAdd(V, NegLo, V->getName()+".off");
3523 InsertNewInstBefore(Add, IB);
3524 Constant *UpperBound = ConstantExpr::getAdd(NegLo, Hi);
3525 return new ICmpInst(ICmpInst::ICMP_ULT, Add, UpperBound);
3528 if (Lo == Hi) // Trivially true.
3529 return new ICmpInst(ICmpInst::ICMP_EQ, V, V);
3531 // V < Min || V >= Hi -> V > Hi-1
3532 Hi = SubOne(cast<ConstantInt>(Hi));
3533 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
3534 ICmpInst::Predicate pred = (isSigned ?
3535 ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT);
3536 return new ICmpInst(pred, V, Hi);
3539 // Emit V-Lo >u Hi-1-Lo
3540 // Note that Hi has already had one subtracted from it, above.
3541 ConstantInt *NegLo = cast<ConstantInt>(ConstantExpr::getNeg(Lo));
3542 Instruction *Add = BinaryOperator::CreateAdd(V, NegLo, V->getName()+".off");
3543 InsertNewInstBefore(Add, IB);
3544 Constant *LowerBound = ConstantExpr::getAdd(NegLo, Hi);
3545 return new ICmpInst(ICmpInst::ICMP_UGT, Add, LowerBound);
3548 // isRunOfOnes - Returns true iff Val consists of one contiguous run of 1s with
3549 // any number of 0s on either side. The 1s are allowed to wrap from LSB to
3550 // MSB, so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs. 0x0F0F0000 is
3551 // not, since all 1s are not contiguous.
3552 static bool isRunOfOnes(ConstantInt *Val, uint32_t &MB, uint32_t &ME) {
3553 const APInt& V = Val->getValue();
3554 uint32_t BitWidth = Val->getType()->getBitWidth();
3555 if (!APIntOps::isShiftedMask(BitWidth, V)) return false;
3557 // look for the first zero bit after the run of ones
3558 MB = BitWidth - ((V - 1) ^ V).countLeadingZeros();
3559 // look for the first non-zero bit
3560 ME = V.getActiveBits();
3564 /// FoldLogicalPlusAnd - This is part of an expression (LHS +/- RHS) & Mask,
3565 /// where isSub determines whether the operator is a sub. If we can fold one of
3566 /// the following xforms:
3568 /// ((A & N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == Mask
3569 /// ((A | N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3570 /// ((A ^ N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3572 /// return (A +/- B).
3574 Value *InstCombiner::FoldLogicalPlusAnd(Value *LHS, Value *RHS,
3575 ConstantInt *Mask, bool isSub,
3577 Instruction *LHSI = dyn_cast<Instruction>(LHS);
3578 if (!LHSI || LHSI->getNumOperands() != 2 ||
3579 !isa<ConstantInt>(LHSI->getOperand(1))) return 0;
3581 ConstantInt *N = cast<ConstantInt>(LHSI->getOperand(1));
3583 switch (LHSI->getOpcode()) {
3585 case Instruction::And:
3586 if (And(N, Mask) == Mask) {
3587 // If the AndRHS is a power of two minus one (0+1+), this is simple.
3588 if ((Mask->getValue().countLeadingZeros() +
3589 Mask->getValue().countPopulation()) ==
3590 Mask->getValue().getBitWidth())
3593 // Otherwise, if Mask is 0+1+0+, and if B is known to have the low 0+
3594 // part, we don't need any explicit masks to take them out of A. If that
3595 // is all N is, ignore it.
3596 uint32_t MB = 0, ME = 0;
3597 if (isRunOfOnes(Mask, MB, ME)) { // begin/end bit of run, inclusive
3598 uint32_t BitWidth = cast<IntegerType>(RHS->getType())->getBitWidth();
3599 APInt Mask(APInt::getLowBitsSet(BitWidth, MB-1));
3600 if (MaskedValueIsZero(RHS, Mask))
3605 case Instruction::Or:
3606 case Instruction::Xor:
3607 // If the AndRHS is a power of two minus one (0+1+), and N&Mask == 0
3608 if ((Mask->getValue().countLeadingZeros() +
3609 Mask->getValue().countPopulation()) == Mask->getValue().getBitWidth()
3610 && And(N, Mask)->isZero())
3617 New = BinaryOperator::CreateSub(LHSI->getOperand(0), RHS, "fold");
3619 New = BinaryOperator::CreateAdd(LHSI->getOperand(0), RHS, "fold");
3620 return InsertNewInstBefore(New, I);
3623 /// FoldAndOfICmps - Fold (icmp)&(icmp) if possible.
3624 Instruction *InstCombiner::FoldAndOfICmps(Instruction &I,
3625 ICmpInst *LHS, ICmpInst *RHS) {
3627 ConstantInt *LHSCst, *RHSCst;
3628 ICmpInst::Predicate LHSCC, RHSCC;
3630 // This only handles icmp of constants: (icmp1 A, C1) & (icmp2 B, C2).
3631 if (!match(LHS, m_ICmp(LHSCC, m_Value(Val), m_ConstantInt(LHSCst))) ||
3632 !match(RHS, m_ICmp(RHSCC, m_Value(Val2), m_ConstantInt(RHSCst))))
3635 // (icmp ult A, C) & (icmp ult B, C) --> (icmp ult (A|B), C)
3636 // where C is a power of 2
3637 if (LHSCst == RHSCst && LHSCC == RHSCC && LHSCC == ICmpInst::ICMP_ULT &&
3638 LHSCst->getValue().isPowerOf2()) {
3639 Instruction *NewOr = BinaryOperator::CreateOr(Val, Val2);
3640 InsertNewInstBefore(NewOr, I);
3641 return new ICmpInst(LHSCC, NewOr, LHSCst);
3644 // From here on, we only handle:
3645 // (icmp1 A, C1) & (icmp2 A, C2) --> something simpler.
3646 if (Val != Val2) return 0;
3648 // ICMP_[US][GL]E X, CST is folded to ICMP_[US][GL]T elsewhere.
3649 if (LHSCC == ICmpInst::ICMP_UGE || LHSCC == ICmpInst::ICMP_ULE ||
3650 RHSCC == ICmpInst::ICMP_UGE || RHSCC == ICmpInst::ICMP_ULE ||
3651 LHSCC == ICmpInst::ICMP_SGE || LHSCC == ICmpInst::ICMP_SLE ||
3652 RHSCC == ICmpInst::ICMP_SGE || RHSCC == ICmpInst::ICMP_SLE)
3655 // We can't fold (ugt x, C) & (sgt x, C2).
3656 if (!PredicatesFoldable(LHSCC, RHSCC))
3659 // Ensure that the larger constant is on the RHS.
3661 if (ICmpInst::isSignedPredicate(LHSCC) ||
3662 (ICmpInst::isEquality(LHSCC) &&
3663 ICmpInst::isSignedPredicate(RHSCC)))
3664 ShouldSwap = LHSCst->getValue().sgt(RHSCst->getValue());
3666 ShouldSwap = LHSCst->getValue().ugt(RHSCst->getValue());
3669 std::swap(LHS, RHS);
3670 std::swap(LHSCst, RHSCst);
3671 std::swap(LHSCC, RHSCC);
3674 // At this point, we know we have have two icmp instructions
3675 // comparing a value against two constants and and'ing the result
3676 // together. Because of the above check, we know that we only have
3677 // icmp eq, icmp ne, icmp [su]lt, and icmp [SU]gt here. We also know
3678 // (from the FoldICmpLogical check above), that the two constants
3679 // are not equal and that the larger constant is on the RHS
3680 assert(LHSCst != RHSCst && "Compares not folded above?");
3683 default: assert(0 && "Unknown integer condition code!");
3684 case ICmpInst::ICMP_EQ:
3686 default: assert(0 && "Unknown integer condition code!");
3687 case ICmpInst::ICMP_EQ: // (X == 13 & X == 15) -> false
3688 case ICmpInst::ICMP_UGT: // (X == 13 & X > 15) -> false
3689 case ICmpInst::ICMP_SGT: // (X == 13 & X > 15) -> false
3690 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3691 case ICmpInst::ICMP_NE: // (X == 13 & X != 15) -> X == 13
3692 case ICmpInst::ICMP_ULT: // (X == 13 & X < 15) -> X == 13
3693 case ICmpInst::ICMP_SLT: // (X == 13 & X < 15) -> X == 13
3694 return ReplaceInstUsesWith(I, LHS);
3696 case ICmpInst::ICMP_NE:
3698 default: assert(0 && "Unknown integer condition code!");
3699 case ICmpInst::ICMP_ULT:
3700 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X u< 14) -> X < 13
3701 return new ICmpInst(ICmpInst::ICMP_ULT, Val, LHSCst);
3702 break; // (X != 13 & X u< 15) -> no change
3703 case ICmpInst::ICMP_SLT:
3704 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X s< 14) -> X < 13
3705 return new ICmpInst(ICmpInst::ICMP_SLT, Val, LHSCst);
3706 break; // (X != 13 & X s< 15) -> no change
3707 case ICmpInst::ICMP_EQ: // (X != 13 & X == 15) -> X == 15
3708 case ICmpInst::ICMP_UGT: // (X != 13 & X u> 15) -> X u> 15
3709 case ICmpInst::ICMP_SGT: // (X != 13 & X s> 15) -> X s> 15
3710 return ReplaceInstUsesWith(I, RHS);
3711 case ICmpInst::ICMP_NE:
3712 if (LHSCst == SubOne(RHSCst)){// (X != 13 & X != 14) -> X-13 >u 1
3713 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
3714 Instruction *Add = BinaryOperator::CreateAdd(Val, AddCST,
3715 Val->getName()+".off");
3716 InsertNewInstBefore(Add, I);
3717 return new ICmpInst(ICmpInst::ICMP_UGT, Add,
3718 ConstantInt::get(Add->getType(), 1));
3720 break; // (X != 13 & X != 15) -> no change
3723 case ICmpInst::ICMP_ULT:
3725 default: assert(0 && "Unknown integer condition code!");
3726 case ICmpInst::ICMP_EQ: // (X u< 13 & X == 15) -> false
3727 case ICmpInst::ICMP_UGT: // (X u< 13 & X u> 15) -> false
3728 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3729 case ICmpInst::ICMP_SGT: // (X u< 13 & X s> 15) -> no change
3731 case ICmpInst::ICMP_NE: // (X u< 13 & X != 15) -> X u< 13
3732 case ICmpInst::ICMP_ULT: // (X u< 13 & X u< 15) -> X u< 13
3733 return ReplaceInstUsesWith(I, LHS);
3734 case ICmpInst::ICMP_SLT: // (X u< 13 & X s< 15) -> no change
3738 case ICmpInst::ICMP_SLT:
3740 default: assert(0 && "Unknown integer condition code!");
3741 case ICmpInst::ICMP_EQ: // (X s< 13 & X == 15) -> false
3742 case ICmpInst::ICMP_SGT: // (X s< 13 & X s> 15) -> false
3743 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3744 case ICmpInst::ICMP_UGT: // (X s< 13 & X u> 15) -> no change
3746 case ICmpInst::ICMP_NE: // (X s< 13 & X != 15) -> X < 13
3747 case ICmpInst::ICMP_SLT: // (X s< 13 & X s< 15) -> X < 13
3748 return ReplaceInstUsesWith(I, LHS);
3749 case ICmpInst::ICMP_ULT: // (X s< 13 & X u< 15) -> no change
3753 case ICmpInst::ICMP_UGT:
3755 default: assert(0 && "Unknown integer condition code!");
3756 case ICmpInst::ICMP_EQ: // (X u> 13 & X == 15) -> X == 15
3757 case ICmpInst::ICMP_UGT: // (X u> 13 & X u> 15) -> X u> 15
3758 return ReplaceInstUsesWith(I, RHS);
3759 case ICmpInst::ICMP_SGT: // (X u> 13 & X s> 15) -> no change
3761 case ICmpInst::ICMP_NE:
3762 if (RHSCst == AddOne(LHSCst)) // (X u> 13 & X != 14) -> X u> 14
3763 return new ICmpInst(LHSCC, Val, RHSCst);
3764 break; // (X u> 13 & X != 15) -> no change
3765 case ICmpInst::ICMP_ULT: // (X u> 13 & X u< 15) -> (X-14) <u 1
3766 return InsertRangeTest(Val, AddOne(LHSCst), RHSCst, false, true, I);
3767 case ICmpInst::ICMP_SLT: // (X u> 13 & X s< 15) -> no change
3771 case ICmpInst::ICMP_SGT:
3773 default: assert(0 && "Unknown integer condition code!");
3774 case ICmpInst::ICMP_EQ: // (X s> 13 & X == 15) -> X == 15
3775 case ICmpInst::ICMP_SGT: // (X s> 13 & X s> 15) -> X s> 15
3776 return ReplaceInstUsesWith(I, RHS);
3777 case ICmpInst::ICMP_UGT: // (X s> 13 & X u> 15) -> no change
3779 case ICmpInst::ICMP_NE:
3780 if (RHSCst == AddOne(LHSCst)) // (X s> 13 & X != 14) -> X s> 14
3781 return new ICmpInst(LHSCC, Val, RHSCst);
3782 break; // (X s> 13 & X != 15) -> no change
3783 case ICmpInst::ICMP_SLT: // (X s> 13 & X s< 15) -> (X-14) s< 1
3784 return InsertRangeTest(Val, AddOne(LHSCst), RHSCst, true, true, I);
3785 case ICmpInst::ICMP_ULT: // (X s> 13 & X u< 15) -> no change
3795 Instruction *InstCombiner::visitAnd(BinaryOperator &I) {
3796 bool Changed = SimplifyCommutative(I);
3797 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3799 if (isa<UndefValue>(Op1)) // X & undef -> 0
3800 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3804 return ReplaceInstUsesWith(I, Op1);
3806 // See if we can simplify any instructions used by the instruction whose sole
3807 // purpose is to compute bits we don't care about.
3808 if (!isa<VectorType>(I.getType())) {
3809 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
3810 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
3811 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
3812 KnownZero, KnownOne))
3815 if (ConstantVector *CP = dyn_cast<ConstantVector>(Op1)) {
3816 if (CP->isAllOnesValue()) // X & <-1,-1> -> X
3817 return ReplaceInstUsesWith(I, I.getOperand(0));
3818 } else if (isa<ConstantAggregateZero>(Op1)) {
3819 return ReplaceInstUsesWith(I, Op1); // X & <0,0> -> <0,0>
3823 if (ConstantInt *AndRHS = dyn_cast<ConstantInt>(Op1)) {
3824 const APInt& AndRHSMask = AndRHS->getValue();
3825 APInt NotAndRHS(~AndRHSMask);
3827 // Optimize a variety of ((val OP C1) & C2) combinations...
3828 if (isa<BinaryOperator>(Op0)) {
3829 Instruction *Op0I = cast<Instruction>(Op0);
3830 Value *Op0LHS = Op0I->getOperand(0);
3831 Value *Op0RHS = Op0I->getOperand(1);
3832 switch (Op0I->getOpcode()) {
3833 case Instruction::Xor:
3834 case Instruction::Or:
3835 // If the mask is only needed on one incoming arm, push it up.
3836 if (Op0I->hasOneUse()) {
3837 if (MaskedValueIsZero(Op0LHS, NotAndRHS)) {
3838 // Not masking anything out for the LHS, move to RHS.
3839 Instruction *NewRHS = BinaryOperator::CreateAnd(Op0RHS, AndRHS,
3840 Op0RHS->getName()+".masked");
3841 InsertNewInstBefore(NewRHS, I);
3842 return BinaryOperator::Create(
3843 cast<BinaryOperator>(Op0I)->getOpcode(), Op0LHS, NewRHS);
3845 if (!isa<Constant>(Op0RHS) &&
3846 MaskedValueIsZero(Op0RHS, NotAndRHS)) {
3847 // Not masking anything out for the RHS, move to LHS.
3848 Instruction *NewLHS = BinaryOperator::CreateAnd(Op0LHS, AndRHS,
3849 Op0LHS->getName()+".masked");
3850 InsertNewInstBefore(NewLHS, I);
3851 return BinaryOperator::Create(
3852 cast<BinaryOperator>(Op0I)->getOpcode(), NewLHS, Op0RHS);
3857 case Instruction::Add:
3858 // ((A & N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == AndRHS.
3859 // ((A | N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
3860 // ((A ^ N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
3861 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, false, I))
3862 return BinaryOperator::CreateAnd(V, AndRHS);
3863 if (Value *V = FoldLogicalPlusAnd(Op0RHS, Op0LHS, AndRHS, false, I))
3864 return BinaryOperator::CreateAnd(V, AndRHS); // Add commutes
3867 case Instruction::Sub:
3868 // ((A & N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == AndRHS.
3869 // ((A | N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
3870 // ((A ^ N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
3871 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, true, I))
3872 return BinaryOperator::CreateAnd(V, AndRHS);
3874 // (A - N) & AndRHS -> -N & AndRHS iff A&AndRHS==0 and AndRHS
3875 // has 1's for all bits that the subtraction with A might affect.
3876 if (Op0I->hasOneUse()) {
3877 uint32_t BitWidth = AndRHSMask.getBitWidth();
3878 uint32_t Zeros = AndRHSMask.countLeadingZeros();
3879 APInt Mask = APInt::getLowBitsSet(BitWidth, BitWidth - Zeros);
3881 ConstantInt *A = dyn_cast<ConstantInt>(Op0LHS);
3882 if (!(A && A->isZero()) && // avoid infinite recursion.
3883 MaskedValueIsZero(Op0LHS, Mask)) {
3884 Instruction *NewNeg = BinaryOperator::CreateNeg(Op0RHS);
3885 InsertNewInstBefore(NewNeg, I);
3886 return BinaryOperator::CreateAnd(NewNeg, AndRHS);
3891 case Instruction::Shl:
3892 case Instruction::LShr:
3893 // (1 << x) & 1 --> zext(x == 0)
3894 // (1 >> x) & 1 --> zext(x == 0)
3895 if (AndRHSMask == 1 && Op0LHS == AndRHS) {
3896 Instruction *NewICmp = new ICmpInst(ICmpInst::ICMP_EQ, Op0RHS,
3897 Constant::getNullValue(I.getType()));
3898 InsertNewInstBefore(NewICmp, I);
3899 return new ZExtInst(NewICmp, I.getType());
3904 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
3905 if (Instruction *Res = OptAndOp(Op0I, Op0CI, AndRHS, I))
3907 } else if (CastInst *CI = dyn_cast<CastInst>(Op0)) {
3908 // If this is an integer truncation or change from signed-to-unsigned, and
3909 // if the source is an and/or with immediate, transform it. This
3910 // frequently occurs for bitfield accesses.
3911 if (Instruction *CastOp = dyn_cast<Instruction>(CI->getOperand(0))) {
3912 if ((isa<TruncInst>(CI) || isa<BitCastInst>(CI)) &&
3913 CastOp->getNumOperands() == 2)
3914 if (ConstantInt *AndCI = dyn_cast<ConstantInt>(CastOp->getOperand(1))) {
3915 if (CastOp->getOpcode() == Instruction::And) {
3916 // Change: and (cast (and X, C1) to T), C2
3917 // into : and (cast X to T), trunc_or_bitcast(C1)&C2
3918 // This will fold the two constants together, which may allow
3919 // other simplifications.
3920 Instruction *NewCast = CastInst::CreateTruncOrBitCast(
3921 CastOp->getOperand(0), I.getType(),
3922 CastOp->getName()+".shrunk");
3923 NewCast = InsertNewInstBefore(NewCast, I);
3924 // trunc_or_bitcast(C1)&C2
3925 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
3926 C3 = ConstantExpr::getAnd(C3, AndRHS);
3927 return BinaryOperator::CreateAnd(NewCast, C3);
3928 } else if (CastOp->getOpcode() == Instruction::Or) {
3929 // Change: and (cast (or X, C1) to T), C2
3930 // into : trunc(C1)&C2 iff trunc(C1)&C2 == C2
3931 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
3932 if (ConstantExpr::getAnd(C3, AndRHS) == AndRHS) // trunc(C1)&C2
3933 return ReplaceInstUsesWith(I, AndRHS);
3939 // Try to fold constant and into select arguments.
3940 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3941 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3943 if (isa<PHINode>(Op0))
3944 if (Instruction *NV = FoldOpIntoPhi(I))
3948 Value *Op0NotVal = dyn_castNotVal(Op0);
3949 Value *Op1NotVal = dyn_castNotVal(Op1);
3951 if (Op0NotVal == Op1 || Op1NotVal == Op0) // A & ~A == ~A & A == 0
3952 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3954 // (~A & ~B) == (~(A | B)) - De Morgan's Law
3955 if (Op0NotVal && Op1NotVal && isOnlyUse(Op0) && isOnlyUse(Op1)) {
3956 Instruction *Or = BinaryOperator::CreateOr(Op0NotVal, Op1NotVal,
3957 I.getName()+".demorgan");
3958 InsertNewInstBefore(Or, I);
3959 return BinaryOperator::CreateNot(Or);
3963 Value *A = 0, *B = 0, *C = 0, *D = 0;
3964 if (match(Op0, m_Or(m_Value(A), m_Value(B)))) {
3965 if (A == Op1 || B == Op1) // (A | ?) & A --> A
3966 return ReplaceInstUsesWith(I, Op1);
3968 // (A|B) & ~(A&B) -> A^B
3969 if (match(Op1, m_Not(m_And(m_Value(C), m_Value(D))))) {
3970 if ((A == C && B == D) || (A == D && B == C))
3971 return BinaryOperator::CreateXor(A, B);
3975 if (match(Op1, m_Or(m_Value(A), m_Value(B)))) {
3976 if (A == Op0 || B == Op0) // A & (A | ?) --> A
3977 return ReplaceInstUsesWith(I, Op0);
3979 // ~(A&B) & (A|B) -> A^B
3980 if (match(Op0, m_Not(m_And(m_Value(C), m_Value(D))))) {
3981 if ((A == C && B == D) || (A == D && B == C))
3982 return BinaryOperator::CreateXor(A, B);
3986 if (Op0->hasOneUse() &&
3987 match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
3988 if (A == Op1) { // (A^B)&A -> A&(A^B)
3989 I.swapOperands(); // Simplify below
3990 std::swap(Op0, Op1);
3991 } else if (B == Op1) { // (A^B)&B -> B&(B^A)
3992 cast<BinaryOperator>(Op0)->swapOperands();
3993 I.swapOperands(); // Simplify below
3994 std::swap(Op0, Op1);
3997 if (Op1->hasOneUse() &&
3998 match(Op1, m_Xor(m_Value(A), m_Value(B)))) {
3999 if (B == Op0) { // B&(A^B) -> B&(B^A)
4000 cast<BinaryOperator>(Op1)->swapOperands();
4003 if (A == Op0) { // A&(A^B) -> A & ~B
4004 Instruction *NotB = BinaryOperator::CreateNot(B, "tmp");
4005 InsertNewInstBefore(NotB, I);
4006 return BinaryOperator::CreateAnd(A, NotB);
4011 if (ICmpInst *RHS = dyn_cast<ICmpInst>(Op1)) {
4012 // (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
4013 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
4016 if (ICmpInst *LHS = dyn_cast<ICmpInst>(Op0))
4017 if (Instruction *Res = FoldAndOfICmps(I, LHS, RHS))
4021 // fold (and (cast A), (cast B)) -> (cast (and A, B))
4022 if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
4023 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4024 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind ?
4025 const Type *SrcTy = Op0C->getOperand(0)->getType();
4026 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
4027 // Only do this if the casts both really cause code to be generated.
4028 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4030 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4032 Instruction *NewOp = BinaryOperator::CreateAnd(Op0C->getOperand(0),
4033 Op1C->getOperand(0),
4035 InsertNewInstBefore(NewOp, I);
4036 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
4040 // (X >> Z) & (Y >> Z) -> (X&Y) >> Z for all shifts.
4041 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
4042 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
4043 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
4044 SI0->getOperand(1) == SI1->getOperand(1) &&
4045 (SI0->hasOneUse() || SI1->hasOneUse())) {
4046 Instruction *NewOp =
4047 InsertNewInstBefore(BinaryOperator::CreateAnd(SI0->getOperand(0),
4049 SI0->getName()), I);
4050 return BinaryOperator::Create(SI1->getOpcode(), NewOp,
4051 SI1->getOperand(1));
4055 // If and'ing two fcmp, try combine them into one.
4056 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0))) {
4057 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1))) {
4058 if (LHS->getPredicate() == FCmpInst::FCMP_ORD &&
4059 RHS->getPredicate() == FCmpInst::FCMP_ORD) {
4060 // (fcmp ord x, c) & (fcmp ord y, c) -> (fcmp ord x, y)
4061 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
4062 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
4063 // If either of the constants are nans, then the whole thing returns
4065 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
4066 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4067 return new FCmpInst(FCmpInst::FCMP_ORD, LHS->getOperand(0),
4068 RHS->getOperand(0));
4071 Value *Op0LHS, *Op0RHS, *Op1LHS, *Op1RHS;
4072 FCmpInst::Predicate Op0CC, Op1CC;
4073 if (match(Op0, m_FCmp(Op0CC, m_Value(Op0LHS), m_Value(Op0RHS))) &&
4074 match(Op1, m_FCmp(Op1CC, m_Value(Op1LHS), m_Value(Op1RHS)))) {
4075 if (Op0LHS == Op1RHS && Op0RHS == Op1LHS) {
4076 // Swap RHS operands to match LHS.
4077 Op1CC = FCmpInst::getSwappedPredicate(Op1CC);
4078 std::swap(Op1LHS, Op1RHS);
4080 if (Op0LHS == Op1LHS && Op0RHS == Op1RHS) {
4081 // Simplify (fcmp cc0 x, y) & (fcmp cc1 x, y).
4083 return new FCmpInst((FCmpInst::Predicate)Op0CC, Op0LHS, Op0RHS);
4084 else if (Op0CC == FCmpInst::FCMP_FALSE ||
4085 Op1CC == FCmpInst::FCMP_FALSE)
4086 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4087 else if (Op0CC == FCmpInst::FCMP_TRUE)
4088 return ReplaceInstUsesWith(I, Op1);
4089 else if (Op1CC == FCmpInst::FCMP_TRUE)
4090 return ReplaceInstUsesWith(I, Op0);
4093 unsigned Op0Pred = getFCmpCode(Op0CC, Op0Ordered);
4094 unsigned Op1Pred = getFCmpCode(Op1CC, Op1Ordered);
4096 std::swap(Op0, Op1);
4097 std::swap(Op0Pred, Op1Pred);
4098 std::swap(Op0Ordered, Op1Ordered);
4101 // uno && ueq -> uno && (uno || eq) -> ueq
4102 // ord && olt -> ord && (ord && lt) -> olt
4103 if (Op0Ordered == Op1Ordered)
4104 return ReplaceInstUsesWith(I, Op1);
4105 // uno && oeq -> uno && (ord && eq) -> false
4106 // uno && ord -> false
4108 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4109 // ord && ueq -> ord && (uno || eq) -> oeq
4110 return cast<Instruction>(getFCmpValue(true, Op1Pred,
4119 return Changed ? &I : 0;
4122 /// CollectBSwapParts - Analyze the specified subexpression and see if it is
4123 /// capable of providing pieces of a bswap. The subexpression provides pieces
4124 /// of a bswap if it is proven that each of the non-zero bytes in the output of
4125 /// the expression came from the corresponding "byte swapped" byte in some other
4126 /// value. For example, if the current subexpression is "(shl i32 %X, 24)" then
4127 /// we know that the expression deposits the low byte of %X into the high byte
4128 /// of the bswap result and that all other bytes are zero. This expression is
4129 /// accepted, the high byte of ByteValues is set to X to indicate a correct
4132 /// This function returns true if the match was unsuccessful and false if so.
4133 /// On entry to the function the "OverallLeftShift" is a signed integer value
4134 /// indicating the number of bytes that the subexpression is later shifted. For
4135 /// example, if the expression is later right shifted by 16 bits, the
4136 /// OverallLeftShift value would be -2 on entry. This is used to specify which
4137 /// byte of ByteValues is actually being set.
4139 /// Similarly, ByteMask is a bitmask where a bit is clear if its corresponding
4140 /// byte is masked to zero by a user. For example, in (X & 255), X will be
4141 /// processed with a bytemask of 1. Because bytemask is 32-bits, this limits
4142 /// this function to working on up to 32-byte (256 bit) values. ByteMask is
4143 /// always in the local (OverallLeftShift) coordinate space.
4145 static bool CollectBSwapParts(Value *V, int OverallLeftShift, uint32_t ByteMask,
4146 SmallVector<Value*, 8> &ByteValues) {
4147 if (Instruction *I = dyn_cast<Instruction>(V)) {
4148 // If this is an or instruction, it may be an inner node of the bswap.
4149 if (I->getOpcode() == Instruction::Or) {
4150 return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
4152 CollectBSwapParts(I->getOperand(1), OverallLeftShift, ByteMask,
4156 // If this is a logical shift by a constant multiple of 8, recurse with
4157 // OverallLeftShift and ByteMask adjusted.
4158 if (I->isLogicalShift() && isa<ConstantInt>(I->getOperand(1))) {
4160 cast<ConstantInt>(I->getOperand(1))->getLimitedValue(~0U);
4161 // Ensure the shift amount is defined and of a byte value.
4162 if ((ShAmt & 7) || (ShAmt > 8*ByteValues.size()))
4165 unsigned ByteShift = ShAmt >> 3;
4166 if (I->getOpcode() == Instruction::Shl) {
4167 // X << 2 -> collect(X, +2)
4168 OverallLeftShift += ByteShift;
4169 ByteMask >>= ByteShift;
4171 // X >>u 2 -> collect(X, -2)
4172 OverallLeftShift -= ByteShift;
4173 ByteMask <<= ByteShift;
4174 ByteMask &= (~0U >> (32-ByteValues.size()));
4177 if (OverallLeftShift >= (int)ByteValues.size()) return true;
4178 if (OverallLeftShift <= -(int)ByteValues.size()) return true;
4180 return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
4184 // If this is a logical 'and' with a mask that clears bytes, clear the
4185 // corresponding bytes in ByteMask.
4186 if (I->getOpcode() == Instruction::And &&
4187 isa<ConstantInt>(I->getOperand(1))) {
4188 // Scan every byte of the and mask, seeing if the byte is either 0 or 255.
4189 unsigned NumBytes = ByteValues.size();
4190 APInt Byte(I->getType()->getPrimitiveSizeInBits(), 255);
4191 const APInt &AndMask = cast<ConstantInt>(I->getOperand(1))->getValue();
4193 for (unsigned i = 0; i != NumBytes; ++i, Byte <<= 8) {
4194 // If this byte is masked out by a later operation, we don't care what
4196 if ((ByteMask & (1 << i)) == 0)
4199 // If the AndMask is all zeros for this byte, clear the bit.
4200 APInt MaskB = AndMask & Byte;
4202 ByteMask &= ~(1U << i);
4206 // If the AndMask is not all ones for this byte, it's not a bytezap.
4210 // Otherwise, this byte is kept.
4213 return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
4218 // Okay, we got to something that isn't a shift, 'or' or 'and'. This must be
4219 // the input value to the bswap. Some observations: 1) if more than one byte
4220 // is demanded from this input, then it could not be successfully assembled
4221 // into a byteswap. At least one of the two bytes would not be aligned with
4222 // their ultimate destination.
4223 if (!isPowerOf2_32(ByteMask)) return true;
4224 unsigned InputByteNo = CountTrailingZeros_32(ByteMask);
4226 // 2) The input and ultimate destinations must line up: if byte 3 of an i32
4227 // is demanded, it needs to go into byte 0 of the result. This means that the
4228 // byte needs to be shifted until it lands in the right byte bucket. The
4229 // shift amount depends on the position: if the byte is coming from the high
4230 // part of the value (e.g. byte 3) then it must be shifted right. If from the
4231 // low part, it must be shifted left.
4232 unsigned DestByteNo = InputByteNo + OverallLeftShift;
4233 if (InputByteNo < ByteValues.size()/2) {
4234 if (ByteValues.size()-1-DestByteNo != InputByteNo)
4237 if (ByteValues.size()-1-DestByteNo != InputByteNo)
4241 // If the destination byte value is already defined, the values are or'd
4242 // together, which isn't a bswap (unless it's an or of the same bits).
4243 if (ByteValues[DestByteNo] && ByteValues[DestByteNo] != V)
4245 ByteValues[DestByteNo] = V;
4249 /// MatchBSwap - Given an OR instruction, check to see if this is a bswap idiom.
4250 /// If so, insert the new bswap intrinsic and return it.
4251 Instruction *InstCombiner::MatchBSwap(BinaryOperator &I) {
4252 const IntegerType *ITy = dyn_cast<IntegerType>(I.getType());
4253 if (!ITy || ITy->getBitWidth() % 16 ||
4254 // ByteMask only allows up to 32-byte values.
4255 ITy->getBitWidth() > 32*8)
4256 return 0; // Can only bswap pairs of bytes. Can't do vectors.
4258 /// ByteValues - For each byte of the result, we keep track of which value
4259 /// defines each byte.
4260 SmallVector<Value*, 8> ByteValues;
4261 ByteValues.resize(ITy->getBitWidth()/8);
4263 // Try to find all the pieces corresponding to the bswap.
4264 uint32_t ByteMask = ~0U >> (32-ByteValues.size());
4265 if (CollectBSwapParts(&I, 0, ByteMask, ByteValues))
4268 // Check to see if all of the bytes come from the same value.
4269 Value *V = ByteValues[0];
4270 if (V == 0) return 0; // Didn't find a byte? Must be zero.
4272 // Check to make sure that all of the bytes come from the same value.
4273 for (unsigned i = 1, e = ByteValues.size(); i != e; ++i)
4274 if (ByteValues[i] != V)
4276 const Type *Tys[] = { ITy };
4277 Module *M = I.getParent()->getParent()->getParent();
4278 Function *F = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 1);
4279 return CallInst::Create(F, V);
4282 /// MatchSelectFromAndOr - We have an expression of the form (A&C)|(B&D). Check
4283 /// If A is (cond?-1:0) and either B or D is ~(cond?-1,0) or (cond?0,-1), then
4284 /// we can simplify this expression to "cond ? C : D or B".
4285 static Instruction *MatchSelectFromAndOr(Value *A, Value *B,
4286 Value *C, Value *D) {
4287 // If A is not a select of -1/0, this cannot match.
4289 if (!match(A, m_SelectCst(m_Value(Cond), -1, 0)))
4292 // ((cond?-1:0)&C) | (B&(cond?0:-1)) -> cond ? C : B.
4293 if (match(D, m_SelectCst(m_Specific(Cond), 0, -1)))
4294 return SelectInst::Create(Cond, C, B);
4295 if (match(D, m_Not(m_SelectCst(m_Specific(Cond), -1, 0))))
4296 return SelectInst::Create(Cond, C, B);
4297 // ((cond?-1:0)&C) | ((cond?0:-1)&D) -> cond ? C : D.
4298 if (match(B, m_SelectCst(m_Specific(Cond), 0, -1)))
4299 return SelectInst::Create(Cond, C, D);
4300 if (match(B, m_Not(m_SelectCst(m_Specific(Cond), -1, 0))))
4301 return SelectInst::Create(Cond, C, D);
4305 /// FoldOrOfICmps - Fold (icmp)|(icmp) if possible.
4306 Instruction *InstCombiner::FoldOrOfICmps(Instruction &I,
4307 ICmpInst *LHS, ICmpInst *RHS) {
4309 ConstantInt *LHSCst, *RHSCst;
4310 ICmpInst::Predicate LHSCC, RHSCC;
4312 // This only handles icmp of constants: (icmp1 A, C1) | (icmp2 B, C2).
4313 if (!match(LHS, m_ICmp(LHSCC, m_Value(Val), m_ConstantInt(LHSCst))) ||
4314 !match(RHS, m_ICmp(RHSCC, m_Value(Val2), m_ConstantInt(RHSCst))))
4317 // From here on, we only handle:
4318 // (icmp1 A, C1) | (icmp2 A, C2) --> something simpler.
4319 if (Val != Val2) return 0;
4321 // ICMP_[US][GL]E X, CST is folded to ICMP_[US][GL]T elsewhere.
4322 if (LHSCC == ICmpInst::ICMP_UGE || LHSCC == ICmpInst::ICMP_ULE ||
4323 RHSCC == ICmpInst::ICMP_UGE || RHSCC == ICmpInst::ICMP_ULE ||
4324 LHSCC == ICmpInst::ICMP_SGE || LHSCC == ICmpInst::ICMP_SLE ||
4325 RHSCC == ICmpInst::ICMP_SGE || RHSCC == ICmpInst::ICMP_SLE)
4328 // We can't fold (ugt x, C) | (sgt x, C2).
4329 if (!PredicatesFoldable(LHSCC, RHSCC))
4332 // Ensure that the larger constant is on the RHS.
4334 if (ICmpInst::isSignedPredicate(LHSCC) ||
4335 (ICmpInst::isEquality(LHSCC) &&
4336 ICmpInst::isSignedPredicate(RHSCC)))
4337 ShouldSwap = LHSCst->getValue().sgt(RHSCst->getValue());
4339 ShouldSwap = LHSCst->getValue().ugt(RHSCst->getValue());
4342 std::swap(LHS, RHS);
4343 std::swap(LHSCst, RHSCst);
4344 std::swap(LHSCC, RHSCC);
4347 // At this point, we know we have have two icmp instructions
4348 // comparing a value against two constants and or'ing the result
4349 // together. Because of the above check, we know that we only have
4350 // ICMP_EQ, ICMP_NE, ICMP_LT, and ICMP_GT here. We also know (from the
4351 // FoldICmpLogical check above), that the two constants are not
4353 assert(LHSCst != RHSCst && "Compares not folded above?");
4356 default: assert(0 && "Unknown integer condition code!");
4357 case ICmpInst::ICMP_EQ:
4359 default: assert(0 && "Unknown integer condition code!");
4360 case ICmpInst::ICMP_EQ:
4361 if (LHSCst == SubOne(RHSCst)) { // (X == 13 | X == 14) -> X-13 <u 2
4362 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
4363 Instruction *Add = BinaryOperator::CreateAdd(Val, AddCST,
4364 Val->getName()+".off");
4365 InsertNewInstBefore(Add, I);
4366 AddCST = Subtract(AddOne(RHSCst), LHSCst);
4367 return new ICmpInst(ICmpInst::ICMP_ULT, Add, AddCST);
4369 break; // (X == 13 | X == 15) -> no change
4370 case ICmpInst::ICMP_UGT: // (X == 13 | X u> 14) -> no change
4371 case ICmpInst::ICMP_SGT: // (X == 13 | X s> 14) -> no change
4373 case ICmpInst::ICMP_NE: // (X == 13 | X != 15) -> X != 15
4374 case ICmpInst::ICMP_ULT: // (X == 13 | X u< 15) -> X u< 15
4375 case ICmpInst::ICMP_SLT: // (X == 13 | X s< 15) -> X s< 15
4376 return ReplaceInstUsesWith(I, RHS);
4379 case ICmpInst::ICMP_NE:
4381 default: assert(0 && "Unknown integer condition code!");
4382 case ICmpInst::ICMP_EQ: // (X != 13 | X == 15) -> X != 13
4383 case ICmpInst::ICMP_UGT: // (X != 13 | X u> 15) -> X != 13
4384 case ICmpInst::ICMP_SGT: // (X != 13 | X s> 15) -> X != 13
4385 return ReplaceInstUsesWith(I, LHS);
4386 case ICmpInst::ICMP_NE: // (X != 13 | X != 15) -> true
4387 case ICmpInst::ICMP_ULT: // (X != 13 | X u< 15) -> true
4388 case ICmpInst::ICMP_SLT: // (X != 13 | X s< 15) -> true
4389 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4392 case ICmpInst::ICMP_ULT:
4394 default: assert(0 && "Unknown integer condition code!");
4395 case ICmpInst::ICMP_EQ: // (X u< 13 | X == 14) -> no change
4397 case ICmpInst::ICMP_UGT: // (X u< 13 | X u> 15) -> (X-13) u> 2
4398 // If RHSCst is [us]MAXINT, it is always false. Not handling
4399 // this can cause overflow.
4400 if (RHSCst->isMaxValue(false))
4401 return ReplaceInstUsesWith(I, LHS);
4402 return InsertRangeTest(Val, LHSCst, AddOne(RHSCst), false, false, I);
4403 case ICmpInst::ICMP_SGT: // (X u< 13 | X s> 15) -> no change
4405 case ICmpInst::ICMP_NE: // (X u< 13 | X != 15) -> X != 15
4406 case ICmpInst::ICMP_ULT: // (X u< 13 | X u< 15) -> X u< 15
4407 return ReplaceInstUsesWith(I, RHS);
4408 case ICmpInst::ICMP_SLT: // (X u< 13 | X s< 15) -> no change
4412 case ICmpInst::ICMP_SLT:
4414 default: assert(0 && "Unknown integer condition code!");
4415 case ICmpInst::ICMP_EQ: // (X s< 13 | X == 14) -> no change
4417 case ICmpInst::ICMP_SGT: // (X s< 13 | X s> 15) -> (X-13) s> 2
4418 // If RHSCst is [us]MAXINT, it is always false. Not handling
4419 // this can cause overflow.
4420 if (RHSCst->isMaxValue(true))
4421 return ReplaceInstUsesWith(I, LHS);
4422 return InsertRangeTest(Val, LHSCst, AddOne(RHSCst), true, false, I);
4423 case ICmpInst::ICMP_UGT: // (X s< 13 | X u> 15) -> no change
4425 case ICmpInst::ICMP_NE: // (X s< 13 | X != 15) -> X != 15
4426 case ICmpInst::ICMP_SLT: // (X s< 13 | X s< 15) -> X s< 15
4427 return ReplaceInstUsesWith(I, RHS);
4428 case ICmpInst::ICMP_ULT: // (X s< 13 | X u< 15) -> no change
4432 case ICmpInst::ICMP_UGT:
4434 default: assert(0 && "Unknown integer condition code!");
4435 case ICmpInst::ICMP_EQ: // (X u> 13 | X == 15) -> X u> 13
4436 case ICmpInst::ICMP_UGT: // (X u> 13 | X u> 15) -> X u> 13
4437 return ReplaceInstUsesWith(I, LHS);
4438 case ICmpInst::ICMP_SGT: // (X u> 13 | X s> 15) -> no change
4440 case ICmpInst::ICMP_NE: // (X u> 13 | X != 15) -> true
4441 case ICmpInst::ICMP_ULT: // (X u> 13 | X u< 15) -> true
4442 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4443 case ICmpInst::ICMP_SLT: // (X u> 13 | X s< 15) -> no change
4447 case ICmpInst::ICMP_SGT:
4449 default: assert(0 && "Unknown integer condition code!");
4450 case ICmpInst::ICMP_EQ: // (X s> 13 | X == 15) -> X > 13
4451 case ICmpInst::ICMP_SGT: // (X s> 13 | X s> 15) -> X > 13
4452 return ReplaceInstUsesWith(I, LHS);
4453 case ICmpInst::ICMP_UGT: // (X s> 13 | X u> 15) -> no change
4455 case ICmpInst::ICMP_NE: // (X s> 13 | X != 15) -> true
4456 case ICmpInst::ICMP_SLT: // (X s> 13 | X s< 15) -> true
4457 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4458 case ICmpInst::ICMP_ULT: // (X s> 13 | X u< 15) -> no change
4466 Instruction *InstCombiner::visitOr(BinaryOperator &I) {
4467 bool Changed = SimplifyCommutative(I);
4468 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4470 if (isa<UndefValue>(Op1)) // X | undef -> -1
4471 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4475 return ReplaceInstUsesWith(I, Op0);
4477 // See if we can simplify any instructions used by the instruction whose sole
4478 // purpose is to compute bits we don't care about.
4479 if (!isa<VectorType>(I.getType())) {
4480 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
4481 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
4482 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
4483 KnownZero, KnownOne))
4485 } else if (isa<ConstantAggregateZero>(Op1)) {
4486 return ReplaceInstUsesWith(I, Op0); // X | <0,0> -> X
4487 } else if (ConstantVector *CP = dyn_cast<ConstantVector>(Op1)) {
4488 if (CP->isAllOnesValue()) // X | <-1,-1> -> <-1,-1>
4489 return ReplaceInstUsesWith(I, I.getOperand(1));
4495 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
4496 ConstantInt *C1 = 0; Value *X = 0;
4497 // (X & C1) | C2 --> (X | C2) & (C1|C2)
4498 if (match(Op0, m_And(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
4499 Instruction *Or = BinaryOperator::CreateOr(X, RHS);
4500 InsertNewInstBefore(Or, I);
4502 return BinaryOperator::CreateAnd(Or,
4503 ConstantInt::get(RHS->getValue() | C1->getValue()));
4506 // (X ^ C1) | C2 --> (X | C2) ^ (C1&~C2)
4507 if (match(Op0, m_Xor(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
4508 Instruction *Or = BinaryOperator::CreateOr(X, RHS);
4509 InsertNewInstBefore(Or, I);
4511 return BinaryOperator::CreateXor(Or,
4512 ConstantInt::get(C1->getValue() & ~RHS->getValue()));
4515 // Try to fold constant and into select arguments.
4516 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
4517 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
4519 if (isa<PHINode>(Op0))
4520 if (Instruction *NV = FoldOpIntoPhi(I))
4524 Value *A = 0, *B = 0;
4525 ConstantInt *C1 = 0, *C2 = 0;
4527 if (match(Op0, m_And(m_Value(A), m_Value(B))))
4528 if (A == Op1 || B == Op1) // (A & ?) | A --> A
4529 return ReplaceInstUsesWith(I, Op1);
4530 if (match(Op1, m_And(m_Value(A), m_Value(B))))
4531 if (A == Op0 || B == Op0) // A | (A & ?) --> A
4532 return ReplaceInstUsesWith(I, Op0);
4534 // (A | B) | C and A | (B | C) -> bswap if possible.
4535 // (A >> B) | (C << D) and (A << B) | (B >> C) -> bswap if possible.
4536 if (match(Op0, m_Or(m_Value(), m_Value())) ||
4537 match(Op1, m_Or(m_Value(), m_Value())) ||
4538 (match(Op0, m_Shift(m_Value(), m_Value())) &&
4539 match(Op1, m_Shift(m_Value(), m_Value())))) {
4540 if (Instruction *BSwap = MatchBSwap(I))
4544 // (X^C)|Y -> (X|Y)^C iff Y&C == 0
4545 if (Op0->hasOneUse() && match(Op0, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
4546 MaskedValueIsZero(Op1, C1->getValue())) {
4547 Instruction *NOr = BinaryOperator::CreateOr(A, Op1);
4548 InsertNewInstBefore(NOr, I);
4550 return BinaryOperator::CreateXor(NOr, C1);
4553 // Y|(X^C) -> (X|Y)^C iff Y&C == 0
4554 if (Op1->hasOneUse() && match(Op1, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
4555 MaskedValueIsZero(Op0, C1->getValue())) {
4556 Instruction *NOr = BinaryOperator::CreateOr(A, Op0);
4557 InsertNewInstBefore(NOr, I);
4559 return BinaryOperator::CreateXor(NOr, C1);
4563 Value *C = 0, *D = 0;
4564 if (match(Op0, m_And(m_Value(A), m_Value(C))) &&
4565 match(Op1, m_And(m_Value(B), m_Value(D)))) {
4566 Value *V1 = 0, *V2 = 0, *V3 = 0;
4567 C1 = dyn_cast<ConstantInt>(C);
4568 C2 = dyn_cast<ConstantInt>(D);
4569 if (C1 && C2) { // (A & C1)|(B & C2)
4570 // If we have: ((V + N) & C1) | (V & C2)
4571 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
4572 // replace with V+N.
4573 if (C1->getValue() == ~C2->getValue()) {
4574 if ((C2->getValue() & (C2->getValue()+1)) == 0 && // C2 == 0+1+
4575 match(A, m_Add(m_Value(V1), m_Value(V2)))) {
4576 // Add commutes, try both ways.
4577 if (V1 == B && MaskedValueIsZero(V2, C2->getValue()))
4578 return ReplaceInstUsesWith(I, A);
4579 if (V2 == B && MaskedValueIsZero(V1, C2->getValue()))
4580 return ReplaceInstUsesWith(I, A);
4582 // Or commutes, try both ways.
4583 if ((C1->getValue() & (C1->getValue()+1)) == 0 &&
4584 match(B, m_Add(m_Value(V1), m_Value(V2)))) {
4585 // Add commutes, try both ways.
4586 if (V1 == A && MaskedValueIsZero(V2, C1->getValue()))
4587 return ReplaceInstUsesWith(I, B);
4588 if (V2 == A && MaskedValueIsZero(V1, C1->getValue()))
4589 return ReplaceInstUsesWith(I, B);
4592 V1 = 0; V2 = 0; V3 = 0;
4595 // Check to see if we have any common things being and'ed. If so, find the
4596 // terms for V1 & (V2|V3).
4597 if (isOnlyUse(Op0) || isOnlyUse(Op1)) {
4598 if (A == B) // (A & C)|(A & D) == A & (C|D)
4599 V1 = A, V2 = C, V3 = D;
4600 else if (A == D) // (A & C)|(B & A) == A & (B|C)
4601 V1 = A, V2 = B, V3 = C;
4602 else if (C == B) // (A & C)|(C & D) == C & (A|D)
4603 V1 = C, V2 = A, V3 = D;
4604 else if (C == D) // (A & C)|(B & C) == C & (A|B)
4605 V1 = C, V2 = A, V3 = B;
4609 InsertNewInstBefore(BinaryOperator::CreateOr(V2, V3, "tmp"), I);
4610 return BinaryOperator::CreateAnd(V1, Or);
4614 // (A & (C0?-1:0)) | (B & ~(C0?-1:0)) -> C0 ? A : B, and commuted variants
4615 if (Instruction *Match = MatchSelectFromAndOr(A, B, C, D))
4617 if (Instruction *Match = MatchSelectFromAndOr(B, A, D, C))
4619 if (Instruction *Match = MatchSelectFromAndOr(C, B, A, D))
4621 if (Instruction *Match = MatchSelectFromAndOr(D, A, B, C))
4625 // (X >> Z) | (Y >> Z) -> (X|Y) >> Z for all shifts.
4626 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
4627 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
4628 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
4629 SI0->getOperand(1) == SI1->getOperand(1) &&
4630 (SI0->hasOneUse() || SI1->hasOneUse())) {
4631 Instruction *NewOp =
4632 InsertNewInstBefore(BinaryOperator::CreateOr(SI0->getOperand(0),
4634 SI0->getName()), I);
4635 return BinaryOperator::Create(SI1->getOpcode(), NewOp,
4636 SI1->getOperand(1));
4640 if (match(Op0, m_Not(m_Value(A)))) { // ~A | Op1
4641 if (A == Op1) // ~A | A == -1
4642 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4646 // Note, A is still live here!
4647 if (match(Op1, m_Not(m_Value(B)))) { // Op0 | ~B
4649 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4651 // (~A | ~B) == (~(A & B)) - De Morgan's Law
4652 if (A && isOnlyUse(Op0) && isOnlyUse(Op1)) {
4653 Value *And = InsertNewInstBefore(BinaryOperator::CreateAnd(A, B,
4654 I.getName()+".demorgan"), I);
4655 return BinaryOperator::CreateNot(And);
4659 // (icmp1 A, B) | (icmp2 A, B) --> (icmp3 A, B)
4660 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1))) {
4661 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
4664 if (ICmpInst *LHS = dyn_cast<ICmpInst>(I.getOperand(0)))
4665 if (Instruction *Res = FoldOrOfICmps(I, LHS, RHS))
4669 // fold (or (cast A), (cast B)) -> (cast (or A, B))
4670 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
4671 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4672 if (Op0C->getOpcode() == Op1C->getOpcode()) {// same cast kind ?
4673 if (!isa<ICmpInst>(Op0C->getOperand(0)) ||
4674 !isa<ICmpInst>(Op1C->getOperand(0))) {
4675 const Type *SrcTy = Op0C->getOperand(0)->getType();
4676 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
4677 // Only do this if the casts both really cause code to be
4679 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4681 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4683 Instruction *NewOp = BinaryOperator::CreateOr(Op0C->getOperand(0),
4684 Op1C->getOperand(0),
4686 InsertNewInstBefore(NewOp, I);
4687 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
4694 // (fcmp uno x, c) | (fcmp uno y, c) -> (fcmp uno x, y)
4695 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0))) {
4696 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1))) {
4697 if (LHS->getPredicate() == FCmpInst::FCMP_UNO &&
4698 RHS->getPredicate() == FCmpInst::FCMP_UNO &&
4699 LHS->getOperand(0)->getType() == RHS->getOperand(0)->getType()) {
4700 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
4701 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
4702 // If either of the constants are nans, then the whole thing returns
4704 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
4705 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4707 // Otherwise, no need to compare the two constants, compare the
4709 return new FCmpInst(FCmpInst::FCMP_UNO, LHS->getOperand(0),
4710 RHS->getOperand(0));
4713 Value *Op0LHS, *Op0RHS, *Op1LHS, *Op1RHS;
4714 FCmpInst::Predicate Op0CC, Op1CC;
4715 if (match(Op0, m_FCmp(Op0CC, m_Value(Op0LHS), m_Value(Op0RHS))) &&
4716 match(Op1, m_FCmp(Op1CC, m_Value(Op1LHS), m_Value(Op1RHS)))) {
4717 if (Op0LHS == Op1RHS && Op0RHS == Op1LHS) {
4718 // Swap RHS operands to match LHS.
4719 Op1CC = FCmpInst::getSwappedPredicate(Op1CC);
4720 std::swap(Op1LHS, Op1RHS);
4722 if (Op0LHS == Op1LHS && Op0RHS == Op1RHS) {
4723 // Simplify (fcmp cc0 x, y) | (fcmp cc1 x, y).
4725 return new FCmpInst((FCmpInst::Predicate)Op0CC, Op0LHS, Op0RHS);
4726 else if (Op0CC == FCmpInst::FCMP_TRUE ||
4727 Op1CC == FCmpInst::FCMP_TRUE)
4728 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4729 else if (Op0CC == FCmpInst::FCMP_FALSE)
4730 return ReplaceInstUsesWith(I, Op1);
4731 else if (Op1CC == FCmpInst::FCMP_FALSE)
4732 return ReplaceInstUsesWith(I, Op0);
4735 unsigned Op0Pred = getFCmpCode(Op0CC, Op0Ordered);
4736 unsigned Op1Pred = getFCmpCode(Op1CC, Op1Ordered);
4737 if (Op0Ordered == Op1Ordered) {
4738 // If both are ordered or unordered, return a new fcmp with
4739 // or'ed predicates.
4740 Value *RV = getFCmpValue(Op0Ordered, Op0Pred|Op1Pred,
4742 if (Instruction *I = dyn_cast<Instruction>(RV))
4744 // Otherwise, it's a constant boolean value...
4745 return ReplaceInstUsesWith(I, RV);
4753 return Changed ? &I : 0;
4758 // XorSelf - Implements: X ^ X --> 0
4761 XorSelf(Value *rhs) : RHS(rhs) {}
4762 bool shouldApply(Value *LHS) const { return LHS == RHS; }
4763 Instruction *apply(BinaryOperator &Xor) const {
4770 Instruction *InstCombiner::visitXor(BinaryOperator &I) {
4771 bool Changed = SimplifyCommutative(I);
4772 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4774 if (isa<UndefValue>(Op1)) {
4775 if (isa<UndefValue>(Op0))
4776 // Handle undef ^ undef -> 0 special case. This is a common
4778 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
4779 return ReplaceInstUsesWith(I, Op1); // X ^ undef -> undef
4782 // xor X, X = 0, even if X is nested in a sequence of Xor's.
4783 if (Instruction *Result = AssociativeOpt(I, XorSelf(Op1))) {
4784 assert(Result == &I && "AssociativeOpt didn't work?"); Result=Result;
4785 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
4788 // See if we can simplify any instructions used by the instruction whose sole
4789 // purpose is to compute bits we don't care about.
4790 if (!isa<VectorType>(I.getType())) {
4791 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
4792 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
4793 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
4794 KnownZero, KnownOne))
4796 } else if (isa<ConstantAggregateZero>(Op1)) {
4797 return ReplaceInstUsesWith(I, Op0); // X ^ <0,0> -> X
4800 // Is this a ~ operation?
4801 if (Value *NotOp = dyn_castNotVal(&I)) {
4802 // ~(~X & Y) --> (X | ~Y) - De Morgan's Law
4803 // ~(~X | Y) === (X & ~Y) - De Morgan's Law
4804 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(NotOp)) {
4805 if (Op0I->getOpcode() == Instruction::And ||
4806 Op0I->getOpcode() == Instruction::Or) {
4807 if (dyn_castNotVal(Op0I->getOperand(1))) Op0I->swapOperands();
4808 if (Value *Op0NotVal = dyn_castNotVal(Op0I->getOperand(0))) {
4810 BinaryOperator::CreateNot(Op0I->getOperand(1),
4811 Op0I->getOperand(1)->getName()+".not");
4812 InsertNewInstBefore(NotY, I);
4813 if (Op0I->getOpcode() == Instruction::And)
4814 return BinaryOperator::CreateOr(Op0NotVal, NotY);
4816 return BinaryOperator::CreateAnd(Op0NotVal, NotY);
4823 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
4824 // xor (cmp A, B), true = not (cmp A, B) = !cmp A, B
4825 if (RHS == ConstantInt::getTrue() && Op0->hasOneUse()) {
4826 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Op0))
4827 return new ICmpInst(ICI->getInversePredicate(),
4828 ICI->getOperand(0), ICI->getOperand(1));
4830 if (FCmpInst *FCI = dyn_cast<FCmpInst>(Op0))
4831 return new FCmpInst(FCI->getInversePredicate(),
4832 FCI->getOperand(0), FCI->getOperand(1));
4835 // fold (xor(zext(cmp)), 1) and (xor(sext(cmp)), -1) to ext(!cmp).
4836 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
4837 if (CmpInst *CI = dyn_cast<CmpInst>(Op0C->getOperand(0))) {
4838 if (CI->hasOneUse() && Op0C->hasOneUse()) {
4839 Instruction::CastOps Opcode = Op0C->getOpcode();
4840 if (Opcode == Instruction::ZExt || Opcode == Instruction::SExt) {
4841 if (RHS == ConstantExpr::getCast(Opcode, ConstantInt::getTrue(),
4842 Op0C->getDestTy())) {
4843 Instruction *NewCI = InsertNewInstBefore(CmpInst::Create(
4844 CI->getOpcode(), CI->getInversePredicate(),
4845 CI->getOperand(0), CI->getOperand(1)), I);
4846 NewCI->takeName(CI);
4847 return CastInst::Create(Opcode, NewCI, Op0C->getType());
4854 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
4855 // ~(c-X) == X-c-1 == X+(-c-1)
4856 if (Op0I->getOpcode() == Instruction::Sub && RHS->isAllOnesValue())
4857 if (Constant *Op0I0C = dyn_cast<Constant>(Op0I->getOperand(0))) {
4858 Constant *NegOp0I0C = ConstantExpr::getNeg(Op0I0C);
4859 Constant *ConstantRHS = ConstantExpr::getSub(NegOp0I0C,
4860 ConstantInt::get(I.getType(), 1));
4861 return BinaryOperator::CreateAdd(Op0I->getOperand(1), ConstantRHS);
4864 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) {
4865 if (Op0I->getOpcode() == Instruction::Add) {
4866 // ~(X-c) --> (-c-1)-X
4867 if (RHS->isAllOnesValue()) {
4868 Constant *NegOp0CI = ConstantExpr::getNeg(Op0CI);
4869 return BinaryOperator::CreateSub(
4870 ConstantExpr::getSub(NegOp0CI,
4871 ConstantInt::get(I.getType(), 1)),
4872 Op0I->getOperand(0));
4873 } else if (RHS->getValue().isSignBit()) {
4874 // (X + C) ^ signbit -> (X + C + signbit)
4875 Constant *C = ConstantInt::get(RHS->getValue() + Op0CI->getValue());
4876 return BinaryOperator::CreateAdd(Op0I->getOperand(0), C);
4879 } else if (Op0I->getOpcode() == Instruction::Or) {
4880 // (X|C1)^C2 -> X^(C1|C2) iff X&~C1 == 0
4881 if (MaskedValueIsZero(Op0I->getOperand(0), Op0CI->getValue())) {
4882 Constant *NewRHS = ConstantExpr::getOr(Op0CI, RHS);
4883 // Anything in both C1 and C2 is known to be zero, remove it from
4885 Constant *CommonBits = And(Op0CI, RHS);
4886 NewRHS = ConstantExpr::getAnd(NewRHS,
4887 ConstantExpr::getNot(CommonBits));
4888 AddToWorkList(Op0I);
4889 I.setOperand(0, Op0I->getOperand(0));
4890 I.setOperand(1, NewRHS);
4897 // Try to fold constant and into select arguments.
4898 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
4899 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
4901 if (isa<PHINode>(Op0))
4902 if (Instruction *NV = FoldOpIntoPhi(I))
4906 if (Value *X = dyn_castNotVal(Op0)) // ~A ^ A == -1
4908 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4910 if (Value *X = dyn_castNotVal(Op1)) // A ^ ~A == -1
4912 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4915 BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1);
4918 if (match(Op1I, m_Or(m_Value(A), m_Value(B)))) {
4919 if (A == Op0) { // B^(B|A) == (A|B)^B
4920 Op1I->swapOperands();
4922 std::swap(Op0, Op1);
4923 } else if (B == Op0) { // B^(A|B) == (A|B)^B
4924 I.swapOperands(); // Simplified below.
4925 std::swap(Op0, Op1);
4927 } else if (match(Op1I, m_Xor(m_Specific(Op0), m_Value(B)))) {
4928 return ReplaceInstUsesWith(I, B); // A^(A^B) == B
4929 } else if (match(Op1I, m_Xor(m_Value(A), m_Specific(Op0)))) {
4930 return ReplaceInstUsesWith(I, A); // A^(B^A) == B
4931 } else if (match(Op1I, m_And(m_Value(A), m_Value(B))) && Op1I->hasOneUse()){
4932 if (A == Op0) { // A^(A&B) -> A^(B&A)
4933 Op1I->swapOperands();
4936 if (B == Op0) { // A^(B&A) -> (B&A)^A
4937 I.swapOperands(); // Simplified below.
4938 std::swap(Op0, Op1);
4943 BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0);
4946 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) && Op0I->hasOneUse()) {
4947 if (A == Op1) // (B|A)^B == (A|B)^B
4949 if (B == Op1) { // (A|B)^B == A & ~B
4951 InsertNewInstBefore(BinaryOperator::CreateNot(Op1, "tmp"), I);
4952 return BinaryOperator::CreateAnd(A, NotB);
4954 } else if (match(Op0I, m_Xor(m_Specific(Op1), m_Value(B)))) {
4955 return ReplaceInstUsesWith(I, B); // (A^B)^A == B
4956 } else if (match(Op0I, m_Xor(m_Value(A), m_Specific(Op1)))) {
4957 return ReplaceInstUsesWith(I, A); // (B^A)^A == B
4958 } else if (match(Op0I, m_And(m_Value(A), m_Value(B))) && Op0I->hasOneUse()){
4959 if (A == Op1) // (A&B)^A -> (B&A)^A
4961 if (B == Op1 && // (B&A)^A == ~B & A
4962 !isa<ConstantInt>(Op1)) { // Canonical form is (B&C)^C
4964 InsertNewInstBefore(BinaryOperator::CreateNot(A, "tmp"), I);
4965 return BinaryOperator::CreateAnd(N, Op1);
4970 // (X >> Z) ^ (Y >> Z) -> (X^Y) >> Z for all shifts.
4971 if (Op0I && Op1I && Op0I->isShift() &&
4972 Op0I->getOpcode() == Op1I->getOpcode() &&
4973 Op0I->getOperand(1) == Op1I->getOperand(1) &&
4974 (Op1I->hasOneUse() || Op1I->hasOneUse())) {
4975 Instruction *NewOp =
4976 InsertNewInstBefore(BinaryOperator::CreateXor(Op0I->getOperand(0),
4977 Op1I->getOperand(0),
4978 Op0I->getName()), I);
4979 return BinaryOperator::Create(Op1I->getOpcode(), NewOp,
4980 Op1I->getOperand(1));
4984 Value *A, *B, *C, *D;
4985 // (A & B)^(A | B) -> A ^ B
4986 if (match(Op0I, m_And(m_Value(A), m_Value(B))) &&
4987 match(Op1I, m_Or(m_Value(C), m_Value(D)))) {
4988 if ((A == C && B == D) || (A == D && B == C))
4989 return BinaryOperator::CreateXor(A, B);
4991 // (A | B)^(A & B) -> A ^ B
4992 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) &&
4993 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
4994 if ((A == C && B == D) || (A == D && B == C))
4995 return BinaryOperator::CreateXor(A, B);
4999 if ((Op0I->hasOneUse() || Op1I->hasOneUse()) &&
5000 match(Op0I, m_And(m_Value(A), m_Value(B))) &&
5001 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
5002 // (X & Y)^(X & Y) -> (Y^Z) & X
5003 Value *X = 0, *Y = 0, *Z = 0;
5005 X = A, Y = B, Z = D;
5007 X = A, Y = B, Z = C;
5009 X = B, Y = A, Z = D;
5011 X = B, Y = A, Z = C;
5014 Instruction *NewOp =
5015 InsertNewInstBefore(BinaryOperator::CreateXor(Y, Z, Op0->getName()), I);
5016 return BinaryOperator::CreateAnd(NewOp, X);
5021 // (icmp1 A, B) ^ (icmp2 A, B) --> (icmp3 A, B)
5022 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1)))
5023 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
5026 // fold (xor (cast A), (cast B)) -> (cast (xor A, B))
5027 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
5028 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
5029 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind?
5030 const Type *SrcTy = Op0C->getOperand(0)->getType();
5031 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
5032 // Only do this if the casts both really cause code to be generated.
5033 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
5035 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
5037 Instruction *NewOp = BinaryOperator::CreateXor(Op0C->getOperand(0),
5038 Op1C->getOperand(0),
5040 InsertNewInstBefore(NewOp, I);
5041 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
5046 return Changed ? &I : 0;
5049 /// AddWithOverflow - Compute Result = In1+In2, returning true if the result
5050 /// overflowed for this type.
5051 static bool AddWithOverflow(ConstantInt *&Result, ConstantInt *In1,
5052 ConstantInt *In2, bool IsSigned = false) {
5053 Result = cast<ConstantInt>(Add(In1, In2));
5056 if (In2->getValue().isNegative())
5057 return Result->getValue().sgt(In1->getValue());
5059 return Result->getValue().slt(In1->getValue());
5061 return Result->getValue().ult(In1->getValue());
5064 /// SubWithOverflow - Compute Result = In1-In2, returning true if the result
5065 /// overflowed for this type.
5066 static bool SubWithOverflow(ConstantInt *&Result, ConstantInt *In1,
5067 ConstantInt *In2, bool IsSigned = false) {
5068 Result = cast<ConstantInt>(Subtract(In1, In2));
5071 if (In2->getValue().isNegative())
5072 return Result->getValue().slt(In1->getValue());
5074 return Result->getValue().sgt(In1->getValue());
5076 return Result->getValue().ugt(In1->getValue());
5079 /// EmitGEPOffset - Given a getelementptr instruction/constantexpr, emit the
5080 /// code necessary to compute the offset from the base pointer (without adding
5081 /// in the base pointer). Return the result as a signed integer of intptr size.
5082 static Value *EmitGEPOffset(User *GEP, Instruction &I, InstCombiner &IC) {
5083 TargetData &TD = IC.getTargetData();
5084 gep_type_iterator GTI = gep_type_begin(GEP);
5085 const Type *IntPtrTy = TD.getIntPtrType();
5086 Value *Result = Constant::getNullValue(IntPtrTy);
5088 // Build a mask for high order bits.
5089 unsigned IntPtrWidth = TD.getPointerSizeInBits();
5090 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth);
5092 for (User::op_iterator i = GEP->op_begin() + 1, e = GEP->op_end(); i != e;
5095 uint64_t Size = TD.getABITypeSize(GTI.getIndexedType()) & PtrSizeMask;
5096 if (ConstantInt *OpC = dyn_cast<ConstantInt>(Op)) {
5097 if (OpC->isZero()) continue;
5099 // Handle a struct index, which adds its field offset to the pointer.
5100 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
5101 Size = TD.getStructLayout(STy)->getElementOffset(OpC->getZExtValue());
5103 if (ConstantInt *RC = dyn_cast<ConstantInt>(Result))
5104 Result = ConstantInt::get(RC->getValue() + APInt(IntPtrWidth, Size));
5106 Result = IC.InsertNewInstBefore(
5107 BinaryOperator::CreateAdd(Result,
5108 ConstantInt::get(IntPtrTy, Size),
5109 GEP->getName()+".offs"), I);
5113 Constant *Scale = ConstantInt::get(IntPtrTy, Size);
5114 Constant *OC = ConstantExpr::getIntegerCast(OpC, IntPtrTy, true /*SExt*/);
5115 Scale = ConstantExpr::getMul(OC, Scale);
5116 if (Constant *RC = dyn_cast<Constant>(Result))
5117 Result = ConstantExpr::getAdd(RC, Scale);
5119 // Emit an add instruction.
5120 Result = IC.InsertNewInstBefore(
5121 BinaryOperator::CreateAdd(Result, Scale,
5122 GEP->getName()+".offs"), I);
5126 // Convert to correct type.
5127 if (Op->getType() != IntPtrTy) {
5128 if (Constant *OpC = dyn_cast<Constant>(Op))
5129 Op = ConstantExpr::getSExt(OpC, IntPtrTy);
5131 Op = IC.InsertNewInstBefore(new SExtInst(Op, IntPtrTy,
5132 Op->getName()+".c"), I);
5135 Constant *Scale = ConstantInt::get(IntPtrTy, Size);
5136 if (Constant *OpC = dyn_cast<Constant>(Op))
5137 Op = ConstantExpr::getMul(OpC, Scale);
5138 else // We'll let instcombine(mul) convert this to a shl if possible.
5139 Op = IC.InsertNewInstBefore(BinaryOperator::CreateMul(Op, Scale,
5140 GEP->getName()+".idx"), I);
5143 // Emit an add instruction.
5144 if (isa<Constant>(Op) && isa<Constant>(Result))
5145 Result = ConstantExpr::getAdd(cast<Constant>(Op),
5146 cast<Constant>(Result));
5148 Result = IC.InsertNewInstBefore(BinaryOperator::CreateAdd(Op, Result,
5149 GEP->getName()+".offs"), I);
5155 /// EvaluateGEPOffsetExpression - Return an value that can be used to compare of
5156 /// the *offset* implied by GEP to zero. For example, if we have &A[i], we want
5157 /// to return 'i' for "icmp ne i, 0". Note that, in general, indices can be
5158 /// complex, and scales are involved. The above expression would also be legal
5159 /// to codegen as "icmp ne (i*4), 0" (assuming A is a pointer to i32). This
5160 /// later form is less amenable to optimization though, and we are allowed to
5161 /// generate the first by knowing that pointer arithmetic doesn't overflow.
5163 /// If we can't emit an optimized form for this expression, this returns null.
5165 static Value *EvaluateGEPOffsetExpression(User *GEP, Instruction &I,
5167 TargetData &TD = IC.getTargetData();
5168 gep_type_iterator GTI = gep_type_begin(GEP);
5170 // Check to see if this gep only has a single variable index. If so, and if
5171 // any constant indices are a multiple of its scale, then we can compute this
5172 // in terms of the scale of the variable index. For example, if the GEP
5173 // implies an offset of "12 + i*4", then we can codegen this as "3 + i",
5174 // because the expression will cross zero at the same point.
5175 unsigned i, e = GEP->getNumOperands();
5177 for (i = 1; i != e; ++i, ++GTI) {
5178 if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i))) {
5179 // Compute the aggregate offset of constant indices.
5180 if (CI->isZero()) continue;
5182 // Handle a struct index, which adds its field offset to the pointer.
5183 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
5184 Offset += TD.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
5186 uint64_t Size = TD.getABITypeSize(GTI.getIndexedType());
5187 Offset += Size*CI->getSExtValue();
5190 // Found our variable index.
5195 // If there are no variable indices, we must have a constant offset, just
5196 // evaluate it the general way.
5197 if (i == e) return 0;
5199 Value *VariableIdx = GEP->getOperand(i);
5200 // Determine the scale factor of the variable element. For example, this is
5201 // 4 if the variable index is into an array of i32.
5202 uint64_t VariableScale = TD.getABITypeSize(GTI.getIndexedType());
5204 // Verify that there are no other variable indices. If so, emit the hard way.
5205 for (++i, ++GTI; i != e; ++i, ++GTI) {
5206 ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i));
5209 // Compute the aggregate offset of constant indices.
5210 if (CI->isZero()) continue;
5212 // Handle a struct index, which adds its field offset to the pointer.
5213 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
5214 Offset += TD.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
5216 uint64_t Size = TD.getABITypeSize(GTI.getIndexedType());
5217 Offset += Size*CI->getSExtValue();
5221 // Okay, we know we have a single variable index, which must be a
5222 // pointer/array/vector index. If there is no offset, life is simple, return
5224 unsigned IntPtrWidth = TD.getPointerSizeInBits();
5226 // Cast to intptrty in case a truncation occurs. If an extension is needed,
5227 // we don't need to bother extending: the extension won't affect where the
5228 // computation crosses zero.
5229 if (VariableIdx->getType()->getPrimitiveSizeInBits() > IntPtrWidth)
5230 VariableIdx = new TruncInst(VariableIdx, TD.getIntPtrType(),
5231 VariableIdx->getNameStart(), &I);
5235 // Otherwise, there is an index. The computation we will do will be modulo
5236 // the pointer size, so get it.
5237 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth);
5239 Offset &= PtrSizeMask;
5240 VariableScale &= PtrSizeMask;
5242 // To do this transformation, any constant index must be a multiple of the
5243 // variable scale factor. For example, we can evaluate "12 + 4*i" as "3 + i",
5244 // but we can't evaluate "10 + 3*i" in terms of i. Check that the offset is a
5245 // multiple of the variable scale.
5246 int64_t NewOffs = Offset / (int64_t)VariableScale;
5247 if (Offset != NewOffs*(int64_t)VariableScale)
5250 // Okay, we can do this evaluation. Start by converting the index to intptr.
5251 const Type *IntPtrTy = TD.getIntPtrType();
5252 if (VariableIdx->getType() != IntPtrTy)
5253 VariableIdx = CastInst::CreateIntegerCast(VariableIdx, IntPtrTy,
5255 VariableIdx->getNameStart(), &I);
5256 Constant *OffsetVal = ConstantInt::get(IntPtrTy, NewOffs);
5257 return BinaryOperator::CreateAdd(VariableIdx, OffsetVal, "offset", &I);
5261 /// FoldGEPICmp - Fold comparisons between a GEP instruction and something
5262 /// else. At this point we know that the GEP is on the LHS of the comparison.
5263 Instruction *InstCombiner::FoldGEPICmp(User *GEPLHS, Value *RHS,
5264 ICmpInst::Predicate Cond,
5266 assert(dyn_castGetElementPtr(GEPLHS) && "LHS is not a getelementptr!");
5268 // Look through bitcasts.
5269 if (BitCastInst *BCI = dyn_cast<BitCastInst>(RHS))
5270 RHS = BCI->getOperand(0);
5272 Value *PtrBase = GEPLHS->getOperand(0);
5273 if (PtrBase == RHS) {
5274 // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
5275 // This transformation (ignoring the base and scales) is valid because we
5276 // know pointers can't overflow. See if we can output an optimized form.
5277 Value *Offset = EvaluateGEPOffsetExpression(GEPLHS, I, *this);
5279 // If not, synthesize the offset the hard way.
5281 Offset = EmitGEPOffset(GEPLHS, I, *this);
5282 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Offset,
5283 Constant::getNullValue(Offset->getType()));
5284 } else if (User *GEPRHS = dyn_castGetElementPtr(RHS)) {
5285 // If the base pointers are different, but the indices are the same, just
5286 // compare the base pointer.
5287 if (PtrBase != GEPRHS->getOperand(0)) {
5288 bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands();
5289 IndicesTheSame &= GEPLHS->getOperand(0)->getType() ==
5290 GEPRHS->getOperand(0)->getType();
5292 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
5293 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
5294 IndicesTheSame = false;
5298 // If all indices are the same, just compare the base pointers.
5300 return new ICmpInst(ICmpInst::getSignedPredicate(Cond),
5301 GEPLHS->getOperand(0), GEPRHS->getOperand(0));
5303 // Otherwise, the base pointers are different and the indices are
5304 // different, bail out.
5308 // If one of the GEPs has all zero indices, recurse.
5309 bool AllZeros = true;
5310 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
5311 if (!isa<Constant>(GEPLHS->getOperand(i)) ||
5312 !cast<Constant>(GEPLHS->getOperand(i))->isNullValue()) {
5317 return FoldGEPICmp(GEPRHS, GEPLHS->getOperand(0),
5318 ICmpInst::getSwappedPredicate(Cond), I);
5320 // If the other GEP has all zero indices, recurse.
5322 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
5323 if (!isa<Constant>(GEPRHS->getOperand(i)) ||
5324 !cast<Constant>(GEPRHS->getOperand(i))->isNullValue()) {
5329 return FoldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I);
5331 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
5332 // If the GEPs only differ by one index, compare it.
5333 unsigned NumDifferences = 0; // Keep track of # differences.
5334 unsigned DiffOperand = 0; // The operand that differs.
5335 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
5336 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
5337 if (GEPLHS->getOperand(i)->getType()->getPrimitiveSizeInBits() !=
5338 GEPRHS->getOperand(i)->getType()->getPrimitiveSizeInBits()) {
5339 // Irreconcilable differences.
5343 if (NumDifferences++) break;
5348 if (NumDifferences == 0) // SAME GEP?
5349 return ReplaceInstUsesWith(I, // No comparison is needed here.
5350 ConstantInt::get(Type::Int1Ty,
5351 ICmpInst::isTrueWhenEqual(Cond)));
5353 else if (NumDifferences == 1) {
5354 Value *LHSV = GEPLHS->getOperand(DiffOperand);
5355 Value *RHSV = GEPRHS->getOperand(DiffOperand);
5356 // Make sure we do a signed comparison here.
5357 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), LHSV, RHSV);
5361 // Only lower this if the icmp is the only user of the GEP or if we expect
5362 // the result to fold to a constant!
5363 if ((isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
5364 (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
5365 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
5366 Value *L = EmitGEPOffset(GEPLHS, I, *this);
5367 Value *R = EmitGEPOffset(GEPRHS, I, *this);
5368 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), L, R);
5374 /// FoldFCmp_IntToFP_Cst - Fold fcmp ([us]itofp x, cst) if possible.
5376 Instruction *InstCombiner::FoldFCmp_IntToFP_Cst(FCmpInst &I,
5379 if (!isa<ConstantFP>(RHSC)) return 0;
5380 const APFloat &RHS = cast<ConstantFP>(RHSC)->getValueAPF();
5382 // Get the width of the mantissa. We don't want to hack on conversions that
5383 // might lose information from the integer, e.g. "i64 -> float"
5384 int MantissaWidth = LHSI->getType()->getFPMantissaWidth();
5385 if (MantissaWidth == -1) return 0; // Unknown.
5387 // Check to see that the input is converted from an integer type that is small
5388 // enough that preserves all bits. TODO: check here for "known" sign bits.
5389 // This would allow us to handle (fptosi (x >>s 62) to float) if x is i64 f.e.
5390 unsigned InputSize = LHSI->getOperand(0)->getType()->getPrimitiveSizeInBits();
5392 // If this is a uitofp instruction, we need an extra bit to hold the sign.
5393 bool LHSUnsigned = isa<UIToFPInst>(LHSI);
5397 // If the conversion would lose info, don't hack on this.
5398 if ((int)InputSize > MantissaWidth)
5401 // Otherwise, we can potentially simplify the comparison. We know that it
5402 // will always come through as an integer value and we know the constant is
5403 // not a NAN (it would have been previously simplified).
5404 assert(!RHS.isNaN() && "NaN comparison not already folded!");
5406 ICmpInst::Predicate Pred;
5407 switch (I.getPredicate()) {
5408 default: assert(0 && "Unexpected predicate!");
5409 case FCmpInst::FCMP_UEQ:
5410 case FCmpInst::FCMP_OEQ:
5411 Pred = ICmpInst::ICMP_EQ;
5413 case FCmpInst::FCMP_UGT:
5414 case FCmpInst::FCMP_OGT:
5415 Pred = LHSUnsigned ? ICmpInst::ICMP_UGT : ICmpInst::ICMP_SGT;
5417 case FCmpInst::FCMP_UGE:
5418 case FCmpInst::FCMP_OGE:
5419 Pred = LHSUnsigned ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_SGE;
5421 case FCmpInst::FCMP_ULT:
5422 case FCmpInst::FCMP_OLT:
5423 Pred = LHSUnsigned ? ICmpInst::ICMP_ULT : ICmpInst::ICMP_SLT;
5425 case FCmpInst::FCMP_ULE:
5426 case FCmpInst::FCMP_OLE:
5427 Pred = LHSUnsigned ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_SLE;
5429 case FCmpInst::FCMP_UNE:
5430 case FCmpInst::FCMP_ONE:
5431 Pred = ICmpInst::ICMP_NE;
5433 case FCmpInst::FCMP_ORD:
5434 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
5435 case FCmpInst::FCMP_UNO:
5436 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
5439 const IntegerType *IntTy = cast<IntegerType>(LHSI->getOperand(0)->getType());
5441 // Now we know that the APFloat is a normal number, zero or inf.
5443 // See if the FP constant is too large for the integer. For example,
5444 // comparing an i8 to 300.0.
5445 unsigned IntWidth = IntTy->getPrimitiveSizeInBits();
5448 // If the RHS value is > SignedMax, fold the comparison. This handles +INF
5449 // and large values.
5450 APFloat SMax(RHS.getSemantics(), APFloat::fcZero, false);
5451 SMax.convertFromAPInt(APInt::getSignedMaxValue(IntWidth), true,
5452 APFloat::rmNearestTiesToEven);
5453 if (SMax.compare(RHS) == APFloat::cmpLessThan) { // smax < 13123.0
5454 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SLT ||
5455 Pred == ICmpInst::ICMP_SLE)
5456 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
5457 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
5460 // If the RHS value is > UnsignedMax, fold the comparison. This handles
5461 // +INF and large values.
5462 APFloat UMax(RHS.getSemantics(), APFloat::fcZero, false);
5463 UMax.convertFromAPInt(APInt::getMaxValue(IntWidth), false,
5464 APFloat::rmNearestTiesToEven);
5465 if (UMax.compare(RHS) == APFloat::cmpLessThan) { // umax < 13123.0
5466 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_ULT ||
5467 Pred == ICmpInst::ICMP_ULE)
5468 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
5469 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
5474 // See if the RHS value is < SignedMin.
5475 APFloat SMin(RHS.getSemantics(), APFloat::fcZero, false);
5476 SMin.convertFromAPInt(APInt::getSignedMinValue(IntWidth), true,
5477 APFloat::rmNearestTiesToEven);
5478 if (SMin.compare(RHS) == APFloat::cmpGreaterThan) { // smin > 12312.0
5479 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SGT ||
5480 Pred == ICmpInst::ICMP_SGE)
5481 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
5482 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
5486 // Okay, now we know that the FP constant fits in the range [SMIN, SMAX] or
5487 // [0, UMAX], but it may still be fractional. See if it is fractional by
5488 // casting the FP value to the integer value and back, checking for equality.
5489 // Don't do this for zero, because -0.0 is not fractional.
5490 Constant *RHSInt = ConstantExpr::getFPToSI(RHSC, IntTy);
5491 if (!RHS.isZero() &&
5492 ConstantExpr::getSIToFP(RHSInt, RHSC->getType()) != RHSC) {
5493 // If we had a comparison against a fractional value, we have to adjust the
5494 // compare predicate and sometimes the value. RHSC is rounded towards zero
5497 default: assert(0 && "Unexpected integer comparison!");
5498 case ICmpInst::ICMP_NE: // (float)int != 4.4 --> true
5499 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
5500 case ICmpInst::ICMP_EQ: // (float)int == 4.4 --> false
5501 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
5502 case ICmpInst::ICMP_ULE:
5503 // (float)int <= 4.4 --> int <= 4
5504 // (float)int <= -4.4 --> false
5505 if (RHS.isNegative())
5506 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
5508 case ICmpInst::ICMP_SLE:
5509 // (float)int <= 4.4 --> int <= 4
5510 // (float)int <= -4.4 --> int < -4
5511 if (RHS.isNegative())
5512 Pred = ICmpInst::ICMP_SLT;
5514 case ICmpInst::ICMP_ULT:
5515 // (float)int < -4.4 --> false
5516 // (float)int < 4.4 --> int <= 4
5517 if (RHS.isNegative())
5518 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
5519 Pred = ICmpInst::ICMP_ULE;
5521 case ICmpInst::ICMP_SLT:
5522 // (float)int < -4.4 --> int < -4
5523 // (float)int < 4.4 --> int <= 4
5524 if (!RHS.isNegative())
5525 Pred = ICmpInst::ICMP_SLE;
5527 case ICmpInst::ICMP_UGT:
5528 // (float)int > 4.4 --> int > 4
5529 // (float)int > -4.4 --> true
5530 if (RHS.isNegative())
5531 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
5533 case ICmpInst::ICMP_SGT:
5534 // (float)int > 4.4 --> int > 4
5535 // (float)int > -4.4 --> int >= -4
5536 if (RHS.isNegative())
5537 Pred = ICmpInst::ICMP_SGE;
5539 case ICmpInst::ICMP_UGE:
5540 // (float)int >= -4.4 --> true
5541 // (float)int >= 4.4 --> int > 4
5542 if (!RHS.isNegative())
5543 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
5544 Pred = ICmpInst::ICMP_UGT;
5546 case ICmpInst::ICMP_SGE:
5547 // (float)int >= -4.4 --> int >= -4
5548 // (float)int >= 4.4 --> int > 4
5549 if (!RHS.isNegative())
5550 Pred = ICmpInst::ICMP_SGT;
5555 // Lower this FP comparison into an appropriate integer version of the
5557 return new ICmpInst(Pred, LHSI->getOperand(0), RHSInt);
5560 Instruction *InstCombiner::visitFCmpInst(FCmpInst &I) {
5561 bool Changed = SimplifyCompare(I);
5562 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5564 // Fold trivial predicates.
5565 if (I.getPredicate() == FCmpInst::FCMP_FALSE)
5566 return ReplaceInstUsesWith(I, Constant::getNullValue(Type::Int1Ty));
5567 if (I.getPredicate() == FCmpInst::FCMP_TRUE)
5568 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
5570 // Simplify 'fcmp pred X, X'
5572 switch (I.getPredicate()) {
5573 default: assert(0 && "Unknown predicate!");
5574 case FCmpInst::FCMP_UEQ: // True if unordered or equal
5575 case FCmpInst::FCMP_UGE: // True if unordered, greater than, or equal
5576 case FCmpInst::FCMP_ULE: // True if unordered, less than, or equal
5577 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
5578 case FCmpInst::FCMP_OGT: // True if ordered and greater than
5579 case FCmpInst::FCMP_OLT: // True if ordered and less than
5580 case FCmpInst::FCMP_ONE: // True if ordered and operands are unequal
5581 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
5583 case FCmpInst::FCMP_UNO: // True if unordered: isnan(X) | isnan(Y)
5584 case FCmpInst::FCMP_ULT: // True if unordered or less than
5585 case FCmpInst::FCMP_UGT: // True if unordered or greater than
5586 case FCmpInst::FCMP_UNE: // True if unordered or not equal
5587 // Canonicalize these to be 'fcmp uno %X, 0.0'.
5588 I.setPredicate(FCmpInst::FCMP_UNO);
5589 I.setOperand(1, Constant::getNullValue(Op0->getType()));
5592 case FCmpInst::FCMP_ORD: // True if ordered (no nans)
5593 case FCmpInst::FCMP_OEQ: // True if ordered and equal
5594 case FCmpInst::FCMP_OGE: // True if ordered and greater than or equal
5595 case FCmpInst::FCMP_OLE: // True if ordered and less than or equal
5596 // Canonicalize these to be 'fcmp ord %X, 0.0'.
5597 I.setPredicate(FCmpInst::FCMP_ORD);
5598 I.setOperand(1, Constant::getNullValue(Op0->getType()));
5603 if (isa<UndefValue>(Op1)) // fcmp pred X, undef -> undef
5604 return ReplaceInstUsesWith(I, UndefValue::get(Type::Int1Ty));
5606 // Handle fcmp with constant RHS
5607 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
5608 // If the constant is a nan, see if we can fold the comparison based on it.
5609 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHSC)) {
5610 if (CFP->getValueAPF().isNaN()) {
5611 if (FCmpInst::isOrdered(I.getPredicate())) // True if ordered and...
5612 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
5613 assert(FCmpInst::isUnordered(I.getPredicate()) &&
5614 "Comparison must be either ordered or unordered!");
5615 // True if unordered.
5616 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
5620 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
5621 switch (LHSI->getOpcode()) {
5622 case Instruction::PHI:
5623 // Only fold fcmp into the PHI if the phi and fcmp are in the same
5624 // block. If in the same block, we're encouraging jump threading. If
5625 // not, we are just pessimizing the code by making an i1 phi.
5626 if (LHSI->getParent() == I.getParent())
5627 if (Instruction *NV = FoldOpIntoPhi(I))
5630 case Instruction::SIToFP:
5631 case Instruction::UIToFP:
5632 if (Instruction *NV = FoldFCmp_IntToFP_Cst(I, LHSI, RHSC))
5635 case Instruction::Select:
5636 // If either operand of the select is a constant, we can fold the
5637 // comparison into the select arms, which will cause one to be
5638 // constant folded and the select turned into a bitwise or.
5639 Value *Op1 = 0, *Op2 = 0;
5640 if (LHSI->hasOneUse()) {
5641 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
5642 // Fold the known value into the constant operand.
5643 Op1 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
5644 // Insert a new FCmp of the other select operand.
5645 Op2 = InsertNewInstBefore(new FCmpInst(I.getPredicate(),
5646 LHSI->getOperand(2), RHSC,
5648 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
5649 // Fold the known value into the constant operand.
5650 Op2 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
5651 // Insert a new FCmp of the other select operand.
5652 Op1 = InsertNewInstBefore(new FCmpInst(I.getPredicate(),
5653 LHSI->getOperand(1), RHSC,
5659 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
5664 return Changed ? &I : 0;
5667 Instruction *InstCombiner::visitICmpInst(ICmpInst &I) {
5668 bool Changed = SimplifyCompare(I);
5669 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5670 const Type *Ty = Op0->getType();
5674 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
5675 I.isTrueWhenEqual()));
5677 if (isa<UndefValue>(Op1)) // X icmp undef -> undef
5678 return ReplaceInstUsesWith(I, UndefValue::get(Type::Int1Ty));
5680 // icmp <global/alloca*/null>, <global/alloca*/null> - Global/Stack value
5681 // addresses never equal each other! We already know that Op0 != Op1.
5682 if ((isa<GlobalValue>(Op0) || isa<AllocaInst>(Op0) ||
5683 isa<ConstantPointerNull>(Op0)) &&
5684 (isa<GlobalValue>(Op1) || isa<AllocaInst>(Op1) ||
5685 isa<ConstantPointerNull>(Op1)))
5686 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
5687 !I.isTrueWhenEqual()));
5689 // icmp's with boolean values can always be turned into bitwise operations
5690 if (Ty == Type::Int1Ty) {
5691 switch (I.getPredicate()) {
5692 default: assert(0 && "Invalid icmp instruction!");
5693 case ICmpInst::ICMP_EQ: { // icmp eq i1 A, B -> ~(A^B)
5694 Instruction *Xor = BinaryOperator::CreateXor(Op0, Op1, I.getName()+"tmp");
5695 InsertNewInstBefore(Xor, I);
5696 return BinaryOperator::CreateNot(Xor);
5698 case ICmpInst::ICMP_NE: // icmp eq i1 A, B -> A^B
5699 return BinaryOperator::CreateXor(Op0, Op1);
5701 case ICmpInst::ICMP_UGT:
5702 std::swap(Op0, Op1); // Change icmp ugt -> icmp ult
5704 case ICmpInst::ICMP_ULT:{ // icmp ult i1 A, B -> ~A & B
5705 Instruction *Not = BinaryOperator::CreateNot(Op0, I.getName()+"tmp");
5706 InsertNewInstBefore(Not, I);
5707 return BinaryOperator::CreateAnd(Not, Op1);
5709 case ICmpInst::ICMP_SGT:
5710 std::swap(Op0, Op1); // Change icmp sgt -> icmp slt
5712 case ICmpInst::ICMP_SLT: { // icmp slt i1 A, B -> A & ~B
5713 Instruction *Not = BinaryOperator::CreateNot(Op1, I.getName()+"tmp");
5714 InsertNewInstBefore(Not, I);
5715 return BinaryOperator::CreateAnd(Not, Op0);
5717 case ICmpInst::ICMP_UGE:
5718 std::swap(Op0, Op1); // Change icmp uge -> icmp ule
5720 case ICmpInst::ICMP_ULE: { // icmp ule i1 A, B -> ~A | B
5721 Instruction *Not = BinaryOperator::CreateNot(Op0, I.getName()+"tmp");
5722 InsertNewInstBefore(Not, I);
5723 return BinaryOperator::CreateOr(Not, Op1);
5725 case ICmpInst::ICMP_SGE:
5726 std::swap(Op0, Op1); // Change icmp sge -> icmp sle
5728 case ICmpInst::ICMP_SLE: { // icmp sle i1 A, B -> A | ~B
5729 Instruction *Not = BinaryOperator::CreateNot(Op1, I.getName()+"tmp");
5730 InsertNewInstBefore(Not, I);
5731 return BinaryOperator::CreateOr(Not, Op0);
5736 // See if we are doing a comparison with a constant.
5737 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
5740 // (icmp ne/eq (sub A B) 0) -> (icmp ne/eq A, B)
5741 if (I.isEquality() && CI->isNullValue() &&
5742 match(Op0, m_Sub(m_Value(A), m_Value(B)))) {
5743 // (icmp cond A B) if cond is equality
5744 return new ICmpInst(I.getPredicate(), A, B);
5747 // If we have an icmp le or icmp ge instruction, turn it into the
5748 // appropriate icmp lt or icmp gt instruction. This allows us to rely on
5749 // them being folded in the code below.
5750 switch (I.getPredicate()) {
5752 case ICmpInst::ICMP_ULE:
5753 if (CI->isMaxValue(false)) // A <=u MAX -> TRUE
5754 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5755 return new ICmpInst(ICmpInst::ICMP_ULT, Op0, AddOne(CI));
5756 case ICmpInst::ICMP_SLE:
5757 if (CI->isMaxValue(true)) // A <=s MAX -> TRUE
5758 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5759 return new ICmpInst(ICmpInst::ICMP_SLT, Op0, AddOne(CI));
5760 case ICmpInst::ICMP_UGE:
5761 if (CI->isMinValue(false)) // A >=u MIN -> TRUE
5762 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5763 return new ICmpInst( ICmpInst::ICMP_UGT, Op0, SubOne(CI));
5764 case ICmpInst::ICMP_SGE:
5765 if (CI->isMinValue(true)) // A >=s MIN -> TRUE
5766 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5767 return new ICmpInst(ICmpInst::ICMP_SGT, Op0, SubOne(CI));
5770 // See if we can fold the comparison based on range information we can get
5771 // by checking whether bits are known to be zero or one in the input.
5772 uint32_t BitWidth = cast<IntegerType>(Ty)->getBitWidth();
5773 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
5775 // If this comparison is a normal comparison, it demands all
5776 // bits, if it is a sign bit comparison, it only demands the sign bit.
5778 bool isSignBit = isSignBitCheck(I.getPredicate(), CI, UnusedBit);
5780 if (SimplifyDemandedBits(Op0,
5781 isSignBit ? APInt::getSignBit(BitWidth)
5782 : APInt::getAllOnesValue(BitWidth),
5783 KnownZero, KnownOne, 0))
5786 // Given the known and unknown bits, compute a range that the LHS could be
5787 // in. Compute the Min, Max and RHS values based on the known bits. For the
5788 // EQ and NE we use unsigned values.
5789 APInt Min(BitWidth, 0), Max(BitWidth, 0);
5790 if (ICmpInst::isSignedPredicate(I.getPredicate()))
5791 ComputeSignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne, Min, Max);
5793 ComputeUnsignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne,Min,Max);
5795 // If Min and Max are known to be the same, then SimplifyDemandedBits
5796 // figured out that the LHS is a constant. Just constant fold this now so
5797 // that code below can assume that Min != Max.
5799 return ReplaceInstUsesWith(I, ConstantExpr::getICmp(I.getPredicate(),
5800 ConstantInt::get(Min),
5803 // Based on the range information we know about the LHS, see if we can
5804 // simplify this comparison. For example, (x&4) < 8 is always true.
5805 const APInt &RHSVal = CI->getValue();
5806 switch (I.getPredicate()) { // LE/GE have been folded already.
5807 default: assert(0 && "Unknown icmp opcode!");
5808 case ICmpInst::ICMP_EQ:
5809 if (Max.ult(RHSVal) || Min.ugt(RHSVal))
5810 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5812 case ICmpInst::ICMP_NE:
5813 if (Max.ult(RHSVal) || Min.ugt(RHSVal))
5814 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5816 case ICmpInst::ICMP_ULT:
5817 if (Max.ult(RHSVal)) // A <u C -> true iff max(A) < C
5818 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5819 if (Min.uge(RHSVal)) // A <u C -> false iff min(A) >= C
5820 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5821 if (RHSVal == Max) // A <u MAX -> A != MAX
5822 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5823 if (RHSVal == Min+1) // A <u MIN+1 -> A == MIN
5824 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, SubOne(CI));
5826 // (x <u 2147483648) -> (x >s -1) -> true if sign bit clear
5827 if (CI->isMinValue(true))
5828 return new ICmpInst(ICmpInst::ICMP_SGT, Op0,
5829 ConstantInt::getAllOnesValue(Op0->getType()));
5831 case ICmpInst::ICMP_UGT:
5832 if (Min.ugt(RHSVal)) // A >u C -> true iff min(A) > C
5833 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5834 if (Max.ule(RHSVal)) // A >u C -> false iff max(A) <= C
5835 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5837 if (RHSVal == Min) // A >u MIN -> A != MIN
5838 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5839 if (RHSVal == Max-1) // A >u MAX-1 -> A == MAX
5840 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, AddOne(CI));
5842 // (x >u 2147483647) -> (x <s 0) -> true if sign bit set
5843 if (CI->isMaxValue(true))
5844 return new ICmpInst(ICmpInst::ICMP_SLT, Op0,
5845 ConstantInt::getNullValue(Op0->getType()));
5847 case ICmpInst::ICMP_SLT:
5848 if (Max.slt(RHSVal)) // A <s C -> true iff max(A) < C
5849 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5850 if (Min.sge(RHSVal)) // A <s C -> false iff min(A) >= C
5851 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5852 if (RHSVal == Max) // A <s MAX -> A != MAX
5853 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5854 if (RHSVal == Min+1) // A <s MIN+1 -> A == MIN
5855 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, SubOne(CI));
5857 case ICmpInst::ICMP_SGT:
5858 if (Min.sgt(RHSVal)) // A >s C -> true iff min(A) > C
5859 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5860 if (Max.sle(RHSVal)) // A >s C -> false iff max(A) <= C
5861 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5863 if (RHSVal == Min) // A >s MIN -> A != MIN
5864 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5865 if (RHSVal == Max-1) // A >s MAX-1 -> A == MAX
5866 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, AddOne(CI));
5871 // Test if the ICmpInst instruction is used exclusively by a select as
5872 // part of a minimum or maximum operation. If so, refrain from doing
5873 // any other folding. This helps out other analyses which understand
5874 // non-obfuscated minimum and maximum idioms, such as ScalarEvolution
5875 // and CodeGen. And in this case, at least one of the comparison
5876 // operands has at least one user besides the compare (the select),
5877 // which would often largely negate the benefit of folding anyway.
5879 if (SelectInst *SI = dyn_cast<SelectInst>(*I.use_begin()))
5880 if ((SI->getOperand(1) == Op0 && SI->getOperand(2) == Op1) ||
5881 (SI->getOperand(2) == Op0 && SI->getOperand(1) == Op1))
5884 // See if we are doing a comparison between a constant and an instruction that
5885 // can be folded into the comparison.
5886 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
5887 // Since the RHS is a ConstantInt (CI), if the left hand side is an
5888 // instruction, see if that instruction also has constants so that the
5889 // instruction can be folded into the icmp
5890 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
5891 if (Instruction *Res = visitICmpInstWithInstAndIntCst(I, LHSI, CI))
5895 // Handle icmp with constant (but not simple integer constant) RHS
5896 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
5897 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
5898 switch (LHSI->getOpcode()) {
5899 case Instruction::GetElementPtr:
5900 if (RHSC->isNullValue()) {
5901 // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
5902 bool isAllZeros = true;
5903 for (unsigned i = 1, e = LHSI->getNumOperands(); i != e; ++i)
5904 if (!isa<Constant>(LHSI->getOperand(i)) ||
5905 !cast<Constant>(LHSI->getOperand(i))->isNullValue()) {
5910 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
5911 Constant::getNullValue(LHSI->getOperand(0)->getType()));
5915 case Instruction::PHI:
5916 // Only fold icmp into the PHI if the phi and fcmp are in the same
5917 // block. If in the same block, we're encouraging jump threading. If
5918 // not, we are just pessimizing the code by making an i1 phi.
5919 if (LHSI->getParent() == I.getParent())
5920 if (Instruction *NV = FoldOpIntoPhi(I))
5923 case Instruction::Select: {
5924 // If either operand of the select is a constant, we can fold the
5925 // comparison into the select arms, which will cause one to be
5926 // constant folded and the select turned into a bitwise or.
5927 Value *Op1 = 0, *Op2 = 0;
5928 if (LHSI->hasOneUse()) {
5929 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
5930 // Fold the known value into the constant operand.
5931 Op1 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
5932 // Insert a new ICmp of the other select operand.
5933 Op2 = InsertNewInstBefore(new ICmpInst(I.getPredicate(),
5934 LHSI->getOperand(2), RHSC,
5936 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
5937 // Fold the known value into the constant operand.
5938 Op2 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
5939 // Insert a new ICmp of the other select operand.
5940 Op1 = InsertNewInstBefore(new ICmpInst(I.getPredicate(),
5941 LHSI->getOperand(1), RHSC,
5947 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
5950 case Instruction::Malloc:
5951 // If we have (malloc != null), and if the malloc has a single use, we
5952 // can assume it is successful and remove the malloc.
5953 if (LHSI->hasOneUse() && isa<ConstantPointerNull>(RHSC)) {
5954 AddToWorkList(LHSI);
5955 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
5956 !I.isTrueWhenEqual()));
5962 // If we can optimize a 'icmp GEP, P' or 'icmp P, GEP', do so now.
5963 if (User *GEP = dyn_castGetElementPtr(Op0))
5964 if (Instruction *NI = FoldGEPICmp(GEP, Op1, I.getPredicate(), I))
5966 if (User *GEP = dyn_castGetElementPtr(Op1))
5967 if (Instruction *NI = FoldGEPICmp(GEP, Op0,
5968 ICmpInst::getSwappedPredicate(I.getPredicate()), I))
5971 // Test to see if the operands of the icmp are casted versions of other
5972 // values. If the ptr->ptr cast can be stripped off both arguments, we do so
5974 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op0)) {
5975 if (isa<PointerType>(Op0->getType()) &&
5976 (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
5977 // We keep moving the cast from the left operand over to the right
5978 // operand, where it can often be eliminated completely.
5979 Op0 = CI->getOperand(0);
5981 // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast
5982 // so eliminate it as well.
5983 if (BitCastInst *CI2 = dyn_cast<BitCastInst>(Op1))
5984 Op1 = CI2->getOperand(0);
5986 // If Op1 is a constant, we can fold the cast into the constant.
5987 if (Op0->getType() != Op1->getType()) {
5988 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
5989 Op1 = ConstantExpr::getBitCast(Op1C, Op0->getType());
5991 // Otherwise, cast the RHS right before the icmp
5992 Op1 = InsertBitCastBefore(Op1, Op0->getType(), I);
5995 return new ICmpInst(I.getPredicate(), Op0, Op1);
5999 if (isa<CastInst>(Op0)) {
6000 // Handle the special case of: icmp (cast bool to X), <cst>
6001 // This comes up when you have code like
6004 // For generality, we handle any zero-extension of any operand comparison
6005 // with a constant or another cast from the same type.
6006 if (isa<ConstantInt>(Op1) || isa<CastInst>(Op1))
6007 if (Instruction *R = visitICmpInstWithCastAndCast(I))
6011 // See if it's the same type of instruction on the left and right.
6012 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
6013 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
6014 if (Op0I->getOpcode() == Op1I->getOpcode() && Op0I->hasOneUse() &&
6015 Op1I->hasOneUse() && Op0I->getOperand(1) == Op1I->getOperand(1) &&
6017 switch (Op0I->getOpcode()) {
6019 case Instruction::Add:
6020 case Instruction::Sub:
6021 case Instruction::Xor:
6022 // a+x icmp eq/ne b+x --> a icmp b
6023 return new ICmpInst(I.getPredicate(), Op0I->getOperand(0),
6024 Op1I->getOperand(0));
6026 case Instruction::Mul:
6027 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) {
6028 // a * Cst icmp eq/ne b * Cst --> a & Mask icmp b & Mask
6029 // Mask = -1 >> count-trailing-zeros(Cst).
6030 if (!CI->isZero() && !CI->isOne()) {
6031 const APInt &AP = CI->getValue();
6032 ConstantInt *Mask = ConstantInt::get(
6033 APInt::getLowBitsSet(AP.getBitWidth(),
6035 AP.countTrailingZeros()));
6036 Instruction *And1 = BinaryOperator::CreateAnd(Op0I->getOperand(0),
6038 Instruction *And2 = BinaryOperator::CreateAnd(Op1I->getOperand(0),
6040 InsertNewInstBefore(And1, I);
6041 InsertNewInstBefore(And2, I);
6042 return new ICmpInst(I.getPredicate(), And1, And2);
6051 // ~x < ~y --> y < x
6053 if (match(Op0, m_Not(m_Value(A))) &&
6054 match(Op1, m_Not(m_Value(B))))
6055 return new ICmpInst(I.getPredicate(), B, A);
6058 if (I.isEquality()) {
6059 Value *A, *B, *C, *D;
6061 // -x == -y --> x == y
6062 if (match(Op0, m_Neg(m_Value(A))) &&
6063 match(Op1, m_Neg(m_Value(B))))
6064 return new ICmpInst(I.getPredicate(), A, B);
6066 if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
6067 if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0
6068 Value *OtherVal = A == Op1 ? B : A;
6069 return new ICmpInst(I.getPredicate(), OtherVal,
6070 Constant::getNullValue(A->getType()));
6073 if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) {
6074 // A^c1 == C^c2 --> A == C^(c1^c2)
6075 ConstantInt *C1, *C2;
6076 if (match(B, m_ConstantInt(C1)) &&
6077 match(D, m_ConstantInt(C2)) && Op1->hasOneUse()) {
6078 Constant *NC = ConstantInt::get(C1->getValue() ^ C2->getValue());
6079 Instruction *Xor = BinaryOperator::CreateXor(C, NC, "tmp");
6080 return new ICmpInst(I.getPredicate(), A,
6081 InsertNewInstBefore(Xor, I));
6084 // A^B == A^D -> B == D
6085 if (A == C) return new ICmpInst(I.getPredicate(), B, D);
6086 if (A == D) return new ICmpInst(I.getPredicate(), B, C);
6087 if (B == C) return new ICmpInst(I.getPredicate(), A, D);
6088 if (B == D) return new ICmpInst(I.getPredicate(), A, C);
6092 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
6093 (A == Op0 || B == Op0)) {
6094 // A == (A^B) -> B == 0
6095 Value *OtherVal = A == Op0 ? B : A;
6096 return new ICmpInst(I.getPredicate(), OtherVal,
6097 Constant::getNullValue(A->getType()));
6100 // (A-B) == A -> B == 0
6101 if (match(Op0, m_Sub(m_Specific(Op1), m_Value(B))))
6102 return new ICmpInst(I.getPredicate(), B,
6103 Constant::getNullValue(B->getType()));
6105 // A == (A-B) -> B == 0
6106 if (match(Op1, m_Sub(m_Specific(Op0), m_Value(B))))
6107 return new ICmpInst(I.getPredicate(), B,
6108 Constant::getNullValue(B->getType()));
6110 // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
6111 if (Op0->hasOneUse() && Op1->hasOneUse() &&
6112 match(Op0, m_And(m_Value(A), m_Value(B))) &&
6113 match(Op1, m_And(m_Value(C), m_Value(D)))) {
6114 Value *X = 0, *Y = 0, *Z = 0;
6117 X = B; Y = D; Z = A;
6118 } else if (A == D) {
6119 X = B; Y = C; Z = A;
6120 } else if (B == C) {
6121 X = A; Y = D; Z = B;
6122 } else if (B == D) {
6123 X = A; Y = C; Z = B;
6126 if (X) { // Build (X^Y) & Z
6127 Op1 = InsertNewInstBefore(BinaryOperator::CreateXor(X, Y, "tmp"), I);
6128 Op1 = InsertNewInstBefore(BinaryOperator::CreateAnd(Op1, Z, "tmp"), I);
6129 I.setOperand(0, Op1);
6130 I.setOperand(1, Constant::getNullValue(Op1->getType()));
6135 return Changed ? &I : 0;
6139 /// FoldICmpDivCst - Fold "icmp pred, ([su]div X, DivRHS), CmpRHS" where DivRHS
6140 /// and CmpRHS are both known to be integer constants.
6141 Instruction *InstCombiner::FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
6142 ConstantInt *DivRHS) {
6143 ConstantInt *CmpRHS = cast<ConstantInt>(ICI.getOperand(1));
6144 const APInt &CmpRHSV = CmpRHS->getValue();
6146 // FIXME: If the operand types don't match the type of the divide
6147 // then don't attempt this transform. The code below doesn't have the
6148 // logic to deal with a signed divide and an unsigned compare (and
6149 // vice versa). This is because (x /s C1) <s C2 produces different
6150 // results than (x /s C1) <u C2 or (x /u C1) <s C2 or even
6151 // (x /u C1) <u C2. Simply casting the operands and result won't
6152 // work. :( The if statement below tests that condition and bails
6154 bool DivIsSigned = DivI->getOpcode() == Instruction::SDiv;
6155 if (!ICI.isEquality() && DivIsSigned != ICI.isSignedPredicate())
6157 if (DivRHS->isZero())
6158 return 0; // The ProdOV computation fails on divide by zero.
6159 if (DivIsSigned && DivRHS->isAllOnesValue())
6160 return 0; // The overflow computation also screws up here
6161 if (DivRHS->isOne())
6162 return 0; // Not worth bothering, and eliminates some funny cases
6165 // Compute Prod = CI * DivRHS. We are essentially solving an equation
6166 // of form X/C1=C2. We solve for X by multiplying C1 (DivRHS) and
6167 // C2 (CI). By solving for X we can turn this into a range check
6168 // instead of computing a divide.
6169 ConstantInt *Prod = Multiply(CmpRHS, DivRHS);
6171 // Determine if the product overflows by seeing if the product is
6172 // not equal to the divide. Make sure we do the same kind of divide
6173 // as in the LHS instruction that we're folding.
6174 bool ProdOV = (DivIsSigned ? ConstantExpr::getSDiv(Prod, DivRHS) :
6175 ConstantExpr::getUDiv(Prod, DivRHS)) != CmpRHS;
6177 // Get the ICmp opcode
6178 ICmpInst::Predicate Pred = ICI.getPredicate();
6180 // Figure out the interval that is being checked. For example, a comparison
6181 // like "X /u 5 == 0" is really checking that X is in the interval [0, 5).
6182 // Compute this interval based on the constants involved and the signedness of
6183 // the compare/divide. This computes a half-open interval, keeping track of
6184 // whether either value in the interval overflows. After analysis each
6185 // overflow variable is set to 0 if it's corresponding bound variable is valid
6186 // -1 if overflowed off the bottom end, or +1 if overflowed off the top end.
6187 int LoOverflow = 0, HiOverflow = 0;
6188 ConstantInt *LoBound = 0, *HiBound = 0;
6190 if (!DivIsSigned) { // udiv
6191 // e.g. X/5 op 3 --> [15, 20)
6193 HiOverflow = LoOverflow = ProdOV;
6195 HiOverflow = AddWithOverflow(HiBound, LoBound, DivRHS, false);
6196 } else if (DivRHS->getValue().isStrictlyPositive()) { // Divisor is > 0.
6197 if (CmpRHSV == 0) { // (X / pos) op 0
6198 // Can't overflow. e.g. X/2 op 0 --> [-1, 2)
6199 LoBound = cast<ConstantInt>(ConstantExpr::getNeg(SubOne(DivRHS)));
6201 } else if (CmpRHSV.isStrictlyPositive()) { // (X / pos) op pos
6202 LoBound = Prod; // e.g. X/5 op 3 --> [15, 20)
6203 HiOverflow = LoOverflow = ProdOV;
6205 HiOverflow = AddWithOverflow(HiBound, Prod, DivRHS, true);
6206 } else { // (X / pos) op neg
6207 // e.g. X/5 op -3 --> [-15-4, -15+1) --> [-19, -14)
6208 HiBound = AddOne(Prod);
6209 LoOverflow = HiOverflow = ProdOV ? -1 : 0;
6211 ConstantInt* DivNeg = cast<ConstantInt>(ConstantExpr::getNeg(DivRHS));
6212 LoOverflow = AddWithOverflow(LoBound, HiBound, DivNeg,
6216 } else if (DivRHS->getValue().isNegative()) { // Divisor is < 0.
6217 if (CmpRHSV == 0) { // (X / neg) op 0
6218 // e.g. X/-5 op 0 --> [-4, 5)
6219 LoBound = AddOne(DivRHS);
6220 HiBound = cast<ConstantInt>(ConstantExpr::getNeg(DivRHS));
6221 if (HiBound == DivRHS) { // -INTMIN = INTMIN
6222 HiOverflow = 1; // [INTMIN+1, overflow)
6223 HiBound = 0; // e.g. X/INTMIN = 0 --> X > INTMIN
6225 } else if (CmpRHSV.isStrictlyPositive()) { // (X / neg) op pos
6226 // e.g. X/-5 op 3 --> [-19, -14)
6227 HiBound = AddOne(Prod);
6228 HiOverflow = LoOverflow = ProdOV ? -1 : 0;
6230 LoOverflow = AddWithOverflow(LoBound, HiBound, DivRHS, true) ? -1 : 0;
6231 } else { // (X / neg) op neg
6232 LoBound = Prod; // e.g. X/-5 op -3 --> [15, 20)
6233 LoOverflow = HiOverflow = ProdOV;
6235 HiOverflow = SubWithOverflow(HiBound, Prod, DivRHS, true);
6238 // Dividing by a negative swaps the condition. LT <-> GT
6239 Pred = ICmpInst::getSwappedPredicate(Pred);
6242 Value *X = DivI->getOperand(0);
6244 default: assert(0 && "Unhandled icmp opcode!");
6245 case ICmpInst::ICMP_EQ:
6246 if (LoOverflow && HiOverflow)
6247 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
6248 else if (HiOverflow)
6249 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
6250 ICmpInst::ICMP_UGE, X, LoBound);
6251 else if (LoOverflow)
6252 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
6253 ICmpInst::ICMP_ULT, X, HiBound);
6255 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned, true, ICI);
6256 case ICmpInst::ICMP_NE:
6257 if (LoOverflow && HiOverflow)
6258 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
6259 else if (HiOverflow)
6260 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
6261 ICmpInst::ICMP_ULT, X, LoBound);
6262 else if (LoOverflow)
6263 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
6264 ICmpInst::ICMP_UGE, X, HiBound);
6266 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned, false, ICI);
6267 case ICmpInst::ICMP_ULT:
6268 case ICmpInst::ICMP_SLT:
6269 if (LoOverflow == +1) // Low bound is greater than input range.
6270 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
6271 if (LoOverflow == -1) // Low bound is less than input range.
6272 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
6273 return new ICmpInst(Pred, X, LoBound);
6274 case ICmpInst::ICMP_UGT:
6275 case ICmpInst::ICMP_SGT:
6276 if (HiOverflow == +1) // High bound greater than input range.
6277 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
6278 else if (HiOverflow == -1) // High bound less than input range.
6279 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
6280 if (Pred == ICmpInst::ICMP_UGT)
6281 return new ICmpInst(ICmpInst::ICMP_UGE, X, HiBound);
6283 return new ICmpInst(ICmpInst::ICMP_SGE, X, HiBound);
6288 /// visitICmpInstWithInstAndIntCst - Handle "icmp (instr, intcst)".
6290 Instruction *InstCombiner::visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
6293 const APInt &RHSV = RHS->getValue();
6295 switch (LHSI->getOpcode()) {
6296 case Instruction::Xor: // (icmp pred (xor X, XorCST), CI)
6297 if (ConstantInt *XorCST = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
6298 // If this is a comparison that tests the signbit (X < 0) or (x > -1),
6300 if ((ICI.getPredicate() == ICmpInst::ICMP_SLT && RHSV == 0) ||
6301 (ICI.getPredicate() == ICmpInst::ICMP_SGT && RHSV.isAllOnesValue())) {
6302 Value *CompareVal = LHSI->getOperand(0);
6304 // If the sign bit of the XorCST is not set, there is no change to
6305 // the operation, just stop using the Xor.
6306 if (!XorCST->getValue().isNegative()) {
6307 ICI.setOperand(0, CompareVal);
6308 AddToWorkList(LHSI);
6312 // Was the old condition true if the operand is positive?
6313 bool isTrueIfPositive = ICI.getPredicate() == ICmpInst::ICMP_SGT;
6315 // If so, the new one isn't.
6316 isTrueIfPositive ^= true;
6318 if (isTrueIfPositive)
6319 return new ICmpInst(ICmpInst::ICMP_SGT, CompareVal, SubOne(RHS));
6321 return new ICmpInst(ICmpInst::ICMP_SLT, CompareVal, AddOne(RHS));
6325 case Instruction::And: // (icmp pred (and X, AndCST), RHS)
6326 if (LHSI->hasOneUse() && isa<ConstantInt>(LHSI->getOperand(1)) &&
6327 LHSI->getOperand(0)->hasOneUse()) {
6328 ConstantInt *AndCST = cast<ConstantInt>(LHSI->getOperand(1));
6330 // If the LHS is an AND of a truncating cast, we can widen the
6331 // and/compare to be the input width without changing the value
6332 // produced, eliminating a cast.
6333 if (TruncInst *Cast = dyn_cast<TruncInst>(LHSI->getOperand(0))) {
6334 // We can do this transformation if either the AND constant does not
6335 // have its sign bit set or if it is an equality comparison.
6336 // Extending a relational comparison when we're checking the sign
6337 // bit would not work.
6338 if (Cast->hasOneUse() &&
6339 (ICI.isEquality() ||
6340 (AndCST->getValue().isNonNegative() && RHSV.isNonNegative()))) {
6342 cast<IntegerType>(Cast->getOperand(0)->getType())->getBitWidth();
6343 APInt NewCST = AndCST->getValue();
6344 NewCST.zext(BitWidth);
6346 NewCI.zext(BitWidth);
6347 Instruction *NewAnd =
6348 BinaryOperator::CreateAnd(Cast->getOperand(0),
6349 ConstantInt::get(NewCST),LHSI->getName());
6350 InsertNewInstBefore(NewAnd, ICI);
6351 return new ICmpInst(ICI.getPredicate(), NewAnd,
6352 ConstantInt::get(NewCI));
6356 // If this is: (X >> C1) & C2 != C3 (where any shift and any compare
6357 // could exist), turn it into (X & (C2 << C1)) != (C3 << C1). This
6358 // happens a LOT in code produced by the C front-end, for bitfield
6360 BinaryOperator *Shift = dyn_cast<BinaryOperator>(LHSI->getOperand(0));
6361 if (Shift && !Shift->isShift())
6365 ShAmt = Shift ? dyn_cast<ConstantInt>(Shift->getOperand(1)) : 0;
6366 const Type *Ty = Shift ? Shift->getType() : 0; // Type of the shift.
6367 const Type *AndTy = AndCST->getType(); // Type of the and.
6369 // We can fold this as long as we can't shift unknown bits
6370 // into the mask. This can only happen with signed shift
6371 // rights, as they sign-extend.
6373 bool CanFold = Shift->isLogicalShift();
6375 // To test for the bad case of the signed shr, see if any
6376 // of the bits shifted in could be tested after the mask.
6377 uint32_t TyBits = Ty->getPrimitiveSizeInBits();
6378 int ShAmtVal = TyBits - ShAmt->getLimitedValue(TyBits);
6380 uint32_t BitWidth = AndTy->getPrimitiveSizeInBits();
6381 if ((APInt::getHighBitsSet(BitWidth, BitWidth-ShAmtVal) &
6382 AndCST->getValue()) == 0)
6388 if (Shift->getOpcode() == Instruction::Shl)
6389 NewCst = ConstantExpr::getLShr(RHS, ShAmt);
6391 NewCst = ConstantExpr::getShl(RHS, ShAmt);
6393 // Check to see if we are shifting out any of the bits being
6395 if (ConstantExpr::get(Shift->getOpcode(), NewCst, ShAmt) != RHS) {
6396 // If we shifted bits out, the fold is not going to work out.
6397 // As a special case, check to see if this means that the
6398 // result is always true or false now.
6399 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
6400 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
6401 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
6402 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
6404 ICI.setOperand(1, NewCst);
6405 Constant *NewAndCST;
6406 if (Shift->getOpcode() == Instruction::Shl)
6407 NewAndCST = ConstantExpr::getLShr(AndCST, ShAmt);
6409 NewAndCST = ConstantExpr::getShl(AndCST, ShAmt);
6410 LHSI->setOperand(1, NewAndCST);
6411 LHSI->setOperand(0, Shift->getOperand(0));
6412 AddToWorkList(Shift); // Shift is dead.
6413 AddUsesToWorkList(ICI);
6419 // Turn ((X >> Y) & C) == 0 into (X & (C << Y)) == 0. The later is
6420 // preferable because it allows the C<<Y expression to be hoisted out
6421 // of a loop if Y is invariant and X is not.
6422 if (Shift && Shift->hasOneUse() && RHSV == 0 &&
6423 ICI.isEquality() && !Shift->isArithmeticShift() &&
6424 isa<Instruction>(Shift->getOperand(0))) {
6427 if (Shift->getOpcode() == Instruction::LShr) {
6428 NS = BinaryOperator::CreateShl(AndCST,
6429 Shift->getOperand(1), "tmp");
6431 // Insert a logical shift.
6432 NS = BinaryOperator::CreateLShr(AndCST,
6433 Shift->getOperand(1), "tmp");
6435 InsertNewInstBefore(cast<Instruction>(NS), ICI);
6437 // Compute X & (C << Y).
6438 Instruction *NewAnd =
6439 BinaryOperator::CreateAnd(Shift->getOperand(0), NS, LHSI->getName());
6440 InsertNewInstBefore(NewAnd, ICI);
6442 ICI.setOperand(0, NewAnd);
6448 case Instruction::Shl: { // (icmp pred (shl X, ShAmt), CI)
6449 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
6452 uint32_t TypeBits = RHSV.getBitWidth();
6454 // Check that the shift amount is in range. If not, don't perform
6455 // undefined shifts. When the shift is visited it will be
6457 if (ShAmt->uge(TypeBits))
6460 if (ICI.isEquality()) {
6461 // If we are comparing against bits always shifted out, the
6462 // comparison cannot succeed.
6464 ConstantExpr::getShl(ConstantExpr::getLShr(RHS, ShAmt), ShAmt);
6465 if (Comp != RHS) {// Comparing against a bit that we know is zero.
6466 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
6467 Constant *Cst = ConstantInt::get(Type::Int1Ty, IsICMP_NE);
6468 return ReplaceInstUsesWith(ICI, Cst);
6471 if (LHSI->hasOneUse()) {
6472 // Otherwise strength reduce the shift into an and.
6473 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
6475 ConstantInt::get(APInt::getLowBitsSet(TypeBits, TypeBits-ShAmtVal));
6478 BinaryOperator::CreateAnd(LHSI->getOperand(0),
6479 Mask, LHSI->getName()+".mask");
6480 Value *And = InsertNewInstBefore(AndI, ICI);
6481 return new ICmpInst(ICI.getPredicate(), And,
6482 ConstantInt::get(RHSV.lshr(ShAmtVal)));
6486 // Otherwise, if this is a comparison of the sign bit, simplify to and/test.
6487 bool TrueIfSigned = false;
6488 if (LHSI->hasOneUse() &&
6489 isSignBitCheck(ICI.getPredicate(), RHS, TrueIfSigned)) {
6490 // (X << 31) <s 0 --> (X&1) != 0
6491 Constant *Mask = ConstantInt::get(APInt(TypeBits, 1) <<
6492 (TypeBits-ShAmt->getZExtValue()-1));
6494 BinaryOperator::CreateAnd(LHSI->getOperand(0),
6495 Mask, LHSI->getName()+".mask");
6496 Value *And = InsertNewInstBefore(AndI, ICI);
6498 return new ICmpInst(TrueIfSigned ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ,
6499 And, Constant::getNullValue(And->getType()));
6504 case Instruction::LShr: // (icmp pred (shr X, ShAmt), CI)
6505 case Instruction::AShr: {
6506 // Only handle equality comparisons of shift-by-constant.
6507 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
6508 if (!ShAmt || !ICI.isEquality()) break;
6510 // Check that the shift amount is in range. If not, don't perform
6511 // undefined shifts. When the shift is visited it will be
6513 uint32_t TypeBits = RHSV.getBitWidth();
6514 if (ShAmt->uge(TypeBits))
6517 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
6519 // If we are comparing against bits always shifted out, the
6520 // comparison cannot succeed.
6521 APInt Comp = RHSV << ShAmtVal;
6522 if (LHSI->getOpcode() == Instruction::LShr)
6523 Comp = Comp.lshr(ShAmtVal);
6525 Comp = Comp.ashr(ShAmtVal);
6527 if (Comp != RHSV) { // Comparing against a bit that we know is zero.
6528 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
6529 Constant *Cst = ConstantInt::get(Type::Int1Ty, IsICMP_NE);
6530 return ReplaceInstUsesWith(ICI, Cst);
6533 // Otherwise, check to see if the bits shifted out are known to be zero.
6534 // If so, we can compare against the unshifted value:
6535 // (X & 4) >> 1 == 2 --> (X & 4) == 4.
6536 if (LHSI->hasOneUse() &&
6537 MaskedValueIsZero(LHSI->getOperand(0),
6538 APInt::getLowBitsSet(Comp.getBitWidth(), ShAmtVal))) {
6539 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
6540 ConstantExpr::getShl(RHS, ShAmt));
6543 if (LHSI->hasOneUse()) {
6544 // Otherwise strength reduce the shift into an and.
6545 APInt Val(APInt::getHighBitsSet(TypeBits, TypeBits - ShAmtVal));
6546 Constant *Mask = ConstantInt::get(Val);
6549 BinaryOperator::CreateAnd(LHSI->getOperand(0),
6550 Mask, LHSI->getName()+".mask");
6551 Value *And = InsertNewInstBefore(AndI, ICI);
6552 return new ICmpInst(ICI.getPredicate(), And,
6553 ConstantExpr::getShl(RHS, ShAmt));
6558 case Instruction::SDiv:
6559 case Instruction::UDiv:
6560 // Fold: icmp pred ([us]div X, C1), C2 -> range test
6561 // Fold this div into the comparison, producing a range check.
6562 // Determine, based on the divide type, what the range is being
6563 // checked. If there is an overflow on the low or high side, remember
6564 // it, otherwise compute the range [low, hi) bounding the new value.
6565 // See: InsertRangeTest above for the kinds of replacements possible.
6566 if (ConstantInt *DivRHS = dyn_cast<ConstantInt>(LHSI->getOperand(1)))
6567 if (Instruction *R = FoldICmpDivCst(ICI, cast<BinaryOperator>(LHSI),
6572 case Instruction::Add:
6573 // Fold: icmp pred (add, X, C1), C2
6575 if (!ICI.isEquality()) {
6576 ConstantInt *LHSC = dyn_cast<ConstantInt>(LHSI->getOperand(1));
6578 const APInt &LHSV = LHSC->getValue();
6580 ConstantRange CR = ICI.makeConstantRange(ICI.getPredicate(), RHSV)
6583 if (ICI.isSignedPredicate()) {
6584 if (CR.getLower().isSignBit()) {
6585 return new ICmpInst(ICmpInst::ICMP_SLT, LHSI->getOperand(0),
6586 ConstantInt::get(CR.getUpper()));
6587 } else if (CR.getUpper().isSignBit()) {
6588 return new ICmpInst(ICmpInst::ICMP_SGE, LHSI->getOperand(0),
6589 ConstantInt::get(CR.getLower()));
6592 if (CR.getLower().isMinValue()) {
6593 return new ICmpInst(ICmpInst::ICMP_ULT, LHSI->getOperand(0),
6594 ConstantInt::get(CR.getUpper()));
6595 } else if (CR.getUpper().isMinValue()) {
6596 return new ICmpInst(ICmpInst::ICMP_UGE, LHSI->getOperand(0),
6597 ConstantInt::get(CR.getLower()));
6604 // Simplify icmp_eq and icmp_ne instructions with integer constant RHS.
6605 if (ICI.isEquality()) {
6606 bool isICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
6608 // If the first operand is (add|sub|and|or|xor|rem) with a constant, and
6609 // the second operand is a constant, simplify a bit.
6610 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(LHSI)) {
6611 switch (BO->getOpcode()) {
6612 case Instruction::SRem:
6613 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
6614 if (RHSV == 0 && isa<ConstantInt>(BO->getOperand(1)) &&BO->hasOneUse()){
6615 const APInt &V = cast<ConstantInt>(BO->getOperand(1))->getValue();
6616 if (V.sgt(APInt(V.getBitWidth(), 1)) && V.isPowerOf2()) {
6617 Instruction *NewRem =
6618 BinaryOperator::CreateURem(BO->getOperand(0), BO->getOperand(1),
6620 InsertNewInstBefore(NewRem, ICI);
6621 return new ICmpInst(ICI.getPredicate(), NewRem,
6622 Constant::getNullValue(BO->getType()));
6626 case Instruction::Add:
6627 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
6628 if (ConstantInt *BOp1C = dyn_cast<ConstantInt>(BO->getOperand(1))) {
6629 if (BO->hasOneUse())
6630 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
6631 Subtract(RHS, BOp1C));
6632 } else if (RHSV == 0) {
6633 // Replace ((add A, B) != 0) with (A != -B) if A or B is
6634 // efficiently invertible, or if the add has just this one use.
6635 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
6637 if (Value *NegVal = dyn_castNegVal(BOp1))
6638 return new ICmpInst(ICI.getPredicate(), BOp0, NegVal);
6639 else if (Value *NegVal = dyn_castNegVal(BOp0))
6640 return new ICmpInst(ICI.getPredicate(), NegVal, BOp1);
6641 else if (BO->hasOneUse()) {
6642 Instruction *Neg = BinaryOperator::CreateNeg(BOp1);
6643 InsertNewInstBefore(Neg, ICI);
6645 return new ICmpInst(ICI.getPredicate(), BOp0, Neg);
6649 case Instruction::Xor:
6650 // For the xor case, we can xor two constants together, eliminating
6651 // the explicit xor.
6652 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1)))
6653 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
6654 ConstantExpr::getXor(RHS, BOC));
6657 case Instruction::Sub:
6658 // Replace (([sub|xor] A, B) != 0) with (A != B)
6660 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
6664 case Instruction::Or:
6665 // If bits are being or'd in that are not present in the constant we
6666 // are comparing against, then the comparison could never succeed!
6667 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1))) {
6668 Constant *NotCI = ConstantExpr::getNot(RHS);
6669 if (!ConstantExpr::getAnd(BOC, NotCI)->isNullValue())
6670 return ReplaceInstUsesWith(ICI, ConstantInt::get(Type::Int1Ty,
6675 case Instruction::And:
6676 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
6677 // If bits are being compared against that are and'd out, then the
6678 // comparison can never succeed!
6679 if ((RHSV & ~BOC->getValue()) != 0)
6680 return ReplaceInstUsesWith(ICI, ConstantInt::get(Type::Int1Ty,
6683 // If we have ((X & C) == C), turn it into ((X & C) != 0).
6684 if (RHS == BOC && RHSV.isPowerOf2())
6685 return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ :
6686 ICmpInst::ICMP_NE, LHSI,
6687 Constant::getNullValue(RHS->getType()));
6689 // Replace (and X, (1 << size(X)-1) != 0) with x s< 0
6690 if (BOC->getValue().isSignBit()) {
6691 Value *X = BO->getOperand(0);
6692 Constant *Zero = Constant::getNullValue(X->getType());
6693 ICmpInst::Predicate pred = isICMP_NE ?
6694 ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
6695 return new ICmpInst(pred, X, Zero);
6698 // ((X & ~7) == 0) --> X < 8
6699 if (RHSV == 0 && isHighOnes(BOC)) {
6700 Value *X = BO->getOperand(0);
6701 Constant *NegX = ConstantExpr::getNeg(BOC);
6702 ICmpInst::Predicate pred = isICMP_NE ?
6703 ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
6704 return new ICmpInst(pred, X, NegX);
6709 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(LHSI)) {
6710 // Handle icmp {eq|ne} <intrinsic>, intcst.
6711 if (II->getIntrinsicID() == Intrinsic::bswap) {
6713 ICI.setOperand(0, II->getOperand(1));
6714 ICI.setOperand(1, ConstantInt::get(RHSV.byteSwap()));
6718 } else { // Not a ICMP_EQ/ICMP_NE
6719 // If the LHS is a cast from an integral value of the same size,
6720 // then since we know the RHS is a constant, try to simlify.
6721 if (CastInst *Cast = dyn_cast<CastInst>(LHSI)) {
6722 Value *CastOp = Cast->getOperand(0);
6723 const Type *SrcTy = CastOp->getType();
6724 uint32_t SrcTySize = SrcTy->getPrimitiveSizeInBits();
6725 if (SrcTy->isInteger() &&
6726 SrcTySize == Cast->getType()->getPrimitiveSizeInBits()) {
6727 // If this is an unsigned comparison, try to make the comparison use
6728 // smaller constant values.
6729 if (ICI.getPredicate() == ICmpInst::ICMP_ULT && RHSV.isSignBit()) {
6730 // X u< 128 => X s> -1
6731 return new ICmpInst(ICmpInst::ICMP_SGT, CastOp,
6732 ConstantInt::get(APInt::getAllOnesValue(SrcTySize)));
6733 } else if (ICI.getPredicate() == ICmpInst::ICMP_UGT &&
6734 RHSV == APInt::getSignedMaxValue(SrcTySize)) {
6735 // X u> 127 => X s< 0
6736 return new ICmpInst(ICmpInst::ICMP_SLT, CastOp,
6737 Constant::getNullValue(SrcTy));
6745 /// visitICmpInstWithCastAndCast - Handle icmp (cast x to y), (cast/cst).
6746 /// We only handle extending casts so far.
6748 Instruction *InstCombiner::visitICmpInstWithCastAndCast(ICmpInst &ICI) {
6749 const CastInst *LHSCI = cast<CastInst>(ICI.getOperand(0));
6750 Value *LHSCIOp = LHSCI->getOperand(0);
6751 const Type *SrcTy = LHSCIOp->getType();
6752 const Type *DestTy = LHSCI->getType();
6755 // Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the
6756 // integer type is the same size as the pointer type.
6757 if (LHSCI->getOpcode() == Instruction::PtrToInt &&
6758 getTargetData().getPointerSizeInBits() ==
6759 cast<IntegerType>(DestTy)->getBitWidth()) {
6761 if (Constant *RHSC = dyn_cast<Constant>(ICI.getOperand(1))) {
6762 RHSOp = ConstantExpr::getIntToPtr(RHSC, SrcTy);
6763 } else if (PtrToIntInst *RHSC = dyn_cast<PtrToIntInst>(ICI.getOperand(1))) {
6764 RHSOp = RHSC->getOperand(0);
6765 // If the pointer types don't match, insert a bitcast.
6766 if (LHSCIOp->getType() != RHSOp->getType())
6767 RHSOp = InsertBitCastBefore(RHSOp, LHSCIOp->getType(), ICI);
6771 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSOp);
6774 // The code below only handles extension cast instructions, so far.
6776 if (LHSCI->getOpcode() != Instruction::ZExt &&
6777 LHSCI->getOpcode() != Instruction::SExt)
6780 bool isSignedExt = LHSCI->getOpcode() == Instruction::SExt;
6781 bool isSignedCmp = ICI.isSignedPredicate();
6783 if (CastInst *CI = dyn_cast<CastInst>(ICI.getOperand(1))) {
6784 // Not an extension from the same type?
6785 RHSCIOp = CI->getOperand(0);
6786 if (RHSCIOp->getType() != LHSCIOp->getType())
6789 // If the signedness of the two casts doesn't agree (i.e. one is a sext
6790 // and the other is a zext), then we can't handle this.
6791 if (CI->getOpcode() != LHSCI->getOpcode())
6794 // Deal with equality cases early.
6795 if (ICI.isEquality())
6796 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
6798 // A signed comparison of sign extended values simplifies into a
6799 // signed comparison.
6800 if (isSignedCmp && isSignedExt)
6801 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
6803 // The other three cases all fold into an unsigned comparison.
6804 return new ICmpInst(ICI.getUnsignedPredicate(), LHSCIOp, RHSCIOp);
6807 // If we aren't dealing with a constant on the RHS, exit early
6808 ConstantInt *CI = dyn_cast<ConstantInt>(ICI.getOperand(1));
6812 // Compute the constant that would happen if we truncated to SrcTy then
6813 // reextended to DestTy.
6814 Constant *Res1 = ConstantExpr::getTrunc(CI, SrcTy);
6815 Constant *Res2 = ConstantExpr::getCast(LHSCI->getOpcode(), Res1, DestTy);
6817 // If the re-extended constant didn't change...
6819 // Make sure that sign of the Cmp and the sign of the Cast are the same.
6820 // For example, we might have:
6821 // %A = sext short %X to uint
6822 // %B = icmp ugt uint %A, 1330
6823 // It is incorrect to transform this into
6824 // %B = icmp ugt short %X, 1330
6825 // because %A may have negative value.
6827 // However, we allow this when the compare is EQ/NE, because they are
6829 if (isSignedExt == isSignedCmp || ICI.isEquality())
6830 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
6834 // The re-extended constant changed so the constant cannot be represented
6835 // in the shorter type. Consequently, we cannot emit a simple comparison.
6837 // First, handle some easy cases. We know the result cannot be equal at this
6838 // point so handle the ICI.isEquality() cases
6839 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
6840 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
6841 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
6842 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
6844 // Evaluate the comparison for LT (we invert for GT below). LE and GE cases
6845 // should have been folded away previously and not enter in here.
6848 // We're performing a signed comparison.
6849 if (cast<ConstantInt>(CI)->getValue().isNegative())
6850 Result = ConstantInt::getFalse(); // X < (small) --> false
6852 Result = ConstantInt::getTrue(); // X < (large) --> true
6854 // We're performing an unsigned comparison.
6856 // We're performing an unsigned comp with a sign extended value.
6857 // This is true if the input is >= 0. [aka >s -1]
6858 Constant *NegOne = ConstantInt::getAllOnesValue(SrcTy);
6859 Result = InsertNewInstBefore(new ICmpInst(ICmpInst::ICMP_SGT, LHSCIOp,
6860 NegOne, ICI.getName()), ICI);
6862 // Unsigned extend & unsigned compare -> always true.
6863 Result = ConstantInt::getTrue();
6867 // Finally, return the value computed.
6868 if (ICI.getPredicate() == ICmpInst::ICMP_ULT ||
6869 ICI.getPredicate() == ICmpInst::ICMP_SLT)
6870 return ReplaceInstUsesWith(ICI, Result);
6872 assert((ICI.getPredicate()==ICmpInst::ICMP_UGT ||
6873 ICI.getPredicate()==ICmpInst::ICMP_SGT) &&
6874 "ICmp should be folded!");
6875 if (Constant *CI = dyn_cast<Constant>(Result))
6876 return ReplaceInstUsesWith(ICI, ConstantExpr::getNot(CI));
6877 return BinaryOperator::CreateNot(Result);
6880 Instruction *InstCombiner::visitShl(BinaryOperator &I) {
6881 return commonShiftTransforms(I);
6884 Instruction *InstCombiner::visitLShr(BinaryOperator &I) {
6885 return commonShiftTransforms(I);
6888 Instruction *InstCombiner::visitAShr(BinaryOperator &I) {
6889 if (Instruction *R = commonShiftTransforms(I))
6892 Value *Op0 = I.getOperand(0);
6894 // ashr int -1, X = -1 (for any arithmetic shift rights of ~0)
6895 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
6896 if (CSI->isAllOnesValue())
6897 return ReplaceInstUsesWith(I, CSI);
6899 // See if we can turn a signed shr into an unsigned shr.
6900 if (!isa<VectorType>(I.getType()) &&
6901 MaskedValueIsZero(Op0,
6902 APInt::getSignBit(I.getType()->getPrimitiveSizeInBits())))
6903 return BinaryOperator::CreateLShr(Op0, I.getOperand(1));
6908 Instruction *InstCombiner::commonShiftTransforms(BinaryOperator &I) {
6909 assert(I.getOperand(1)->getType() == I.getOperand(0)->getType());
6910 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
6912 // shl X, 0 == X and shr X, 0 == X
6913 // shl 0, X == 0 and shr 0, X == 0
6914 if (Op1 == Constant::getNullValue(Op1->getType()) ||
6915 Op0 == Constant::getNullValue(Op0->getType()))
6916 return ReplaceInstUsesWith(I, Op0);
6918 if (isa<UndefValue>(Op0)) {
6919 if (I.getOpcode() == Instruction::AShr) // undef >>s X -> undef
6920 return ReplaceInstUsesWith(I, Op0);
6921 else // undef << X -> 0, undef >>u X -> 0
6922 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
6924 if (isa<UndefValue>(Op1)) {
6925 if (I.getOpcode() == Instruction::AShr) // X >>s undef -> X
6926 return ReplaceInstUsesWith(I, Op0);
6927 else // X << undef, X >>u undef -> 0
6928 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
6931 // Try to fold constant and into select arguments.
6932 if (isa<Constant>(Op0))
6933 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
6934 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
6937 if (ConstantInt *CUI = dyn_cast<ConstantInt>(Op1))
6938 if (Instruction *Res = FoldShiftByConstant(Op0, CUI, I))
6943 Instruction *InstCombiner::FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
6944 BinaryOperator &I) {
6945 bool isLeftShift = I.getOpcode() == Instruction::Shl;
6947 // See if we can simplify any instructions used by the instruction whose sole
6948 // purpose is to compute bits we don't care about.
6949 uint32_t TypeBits = Op0->getType()->getPrimitiveSizeInBits();
6950 APInt KnownZero(TypeBits, 0), KnownOne(TypeBits, 0);
6951 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(TypeBits),
6952 KnownZero, KnownOne))
6955 // shl uint X, 32 = 0 and shr ubyte Y, 9 = 0, ... just don't eliminate shr
6956 // of a signed value.
6958 if (Op1->uge(TypeBits)) {
6959 if (I.getOpcode() != Instruction::AShr)
6960 return ReplaceInstUsesWith(I, Constant::getNullValue(Op0->getType()));
6962 I.setOperand(1, ConstantInt::get(I.getType(), TypeBits-1));
6967 // ((X*C1) << C2) == (X * (C1 << C2))
6968 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0))
6969 if (BO->getOpcode() == Instruction::Mul && isLeftShift)
6970 if (Constant *BOOp = dyn_cast<Constant>(BO->getOperand(1)))
6971 return BinaryOperator::CreateMul(BO->getOperand(0),
6972 ConstantExpr::getShl(BOOp, Op1));
6974 // Try to fold constant and into select arguments.
6975 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
6976 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
6978 if (isa<PHINode>(Op0))
6979 if (Instruction *NV = FoldOpIntoPhi(I))
6982 // Fold shift2(trunc(shift1(x,c1)), c2) -> trunc(shift2(shift1(x,c1),c2))
6983 if (TruncInst *TI = dyn_cast<TruncInst>(Op0)) {
6984 Instruction *TrOp = dyn_cast<Instruction>(TI->getOperand(0));
6985 // If 'shift2' is an ashr, we would have to get the sign bit into a funny
6986 // place. Don't try to do this transformation in this case. Also, we
6987 // require that the input operand is a shift-by-constant so that we have
6988 // confidence that the shifts will get folded together. We could do this
6989 // xform in more cases, but it is unlikely to be profitable.
6990 if (TrOp && I.isLogicalShift() && TrOp->isShift() &&
6991 isa<ConstantInt>(TrOp->getOperand(1))) {
6992 // Okay, we'll do this xform. Make the shift of shift.
6993 Constant *ShAmt = ConstantExpr::getZExt(Op1, TrOp->getType());
6994 Instruction *NSh = BinaryOperator::Create(I.getOpcode(), TrOp, ShAmt,
6996 InsertNewInstBefore(NSh, I); // (shift2 (shift1 & 0x00FF), c2)
6998 // For logical shifts, the truncation has the effect of making the high
6999 // part of the register be zeros. Emulate this by inserting an AND to
7000 // clear the top bits as needed. This 'and' will usually be zapped by
7001 // other xforms later if dead.
7002 unsigned SrcSize = TrOp->getType()->getPrimitiveSizeInBits();
7003 unsigned DstSize = TI->getType()->getPrimitiveSizeInBits();
7004 APInt MaskV(APInt::getLowBitsSet(SrcSize, DstSize));
7006 // The mask we constructed says what the trunc would do if occurring
7007 // between the shifts. We want to know the effect *after* the second
7008 // shift. We know that it is a logical shift by a constant, so adjust the
7009 // mask as appropriate.
7010 if (I.getOpcode() == Instruction::Shl)
7011 MaskV <<= Op1->getZExtValue();
7013 assert(I.getOpcode() == Instruction::LShr && "Unknown logical shift");
7014 MaskV = MaskV.lshr(Op1->getZExtValue());
7017 Instruction *And = BinaryOperator::CreateAnd(NSh, ConstantInt::get(MaskV),
7019 InsertNewInstBefore(And, I); // shift1 & 0x00FF
7021 // Return the value truncated to the interesting size.
7022 return new TruncInst(And, I.getType());
7026 if (Op0->hasOneUse()) {
7027 if (BinaryOperator *Op0BO = dyn_cast<BinaryOperator>(Op0)) {
7028 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
7031 switch (Op0BO->getOpcode()) {
7033 case Instruction::Add:
7034 case Instruction::And:
7035 case Instruction::Or:
7036 case Instruction::Xor: {
7037 // These operators commute.
7038 // Turn (Y + (X >> C)) << C -> (X + (Y << C)) & (~0 << C)
7039 if (isLeftShift && Op0BO->getOperand(1)->hasOneUse() &&
7040 match(Op0BO->getOperand(1), m_Shr(m_Value(V1), m_Specific(Op1)))){
7041 Instruction *YS = BinaryOperator::CreateShl(
7042 Op0BO->getOperand(0), Op1,
7044 InsertNewInstBefore(YS, I); // (Y << C)
7046 BinaryOperator::Create(Op0BO->getOpcode(), YS, V1,
7047 Op0BO->getOperand(1)->getName());
7048 InsertNewInstBefore(X, I); // (X + (Y << C))
7049 uint32_t Op1Val = Op1->getLimitedValue(TypeBits);
7050 return BinaryOperator::CreateAnd(X, ConstantInt::get(
7051 APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
7054 // Turn (Y + ((X >> C) & CC)) << C -> ((X & (CC << C)) + (Y << C))
7055 Value *Op0BOOp1 = Op0BO->getOperand(1);
7056 if (isLeftShift && Op0BOOp1->hasOneUse() &&
7058 m_And(m_Shr(m_Value(V1), m_Specific(Op1)),
7059 m_ConstantInt(CC))) &&
7060 cast<BinaryOperator>(Op0BOOp1)->getOperand(0)->hasOneUse()) {
7061 Instruction *YS = BinaryOperator::CreateShl(
7062 Op0BO->getOperand(0), Op1,
7064 InsertNewInstBefore(YS, I); // (Y << C)
7066 BinaryOperator::CreateAnd(V1, ConstantExpr::getShl(CC, Op1),
7067 V1->getName()+".mask");
7068 InsertNewInstBefore(XM, I); // X & (CC << C)
7070 return BinaryOperator::Create(Op0BO->getOpcode(), YS, XM);
7075 case Instruction::Sub: {
7076 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
7077 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
7078 match(Op0BO->getOperand(0), m_Shr(m_Value(V1), m_Specific(Op1)))){
7079 Instruction *YS = BinaryOperator::CreateShl(
7080 Op0BO->getOperand(1), Op1,
7082 InsertNewInstBefore(YS, I); // (Y << C)
7084 BinaryOperator::Create(Op0BO->getOpcode(), V1, YS,
7085 Op0BO->getOperand(0)->getName());
7086 InsertNewInstBefore(X, I); // (X + (Y << C))
7087 uint32_t Op1Val = Op1->getLimitedValue(TypeBits);
7088 return BinaryOperator::CreateAnd(X, ConstantInt::get(
7089 APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
7092 // Turn (((X >> C)&CC) + Y) << C -> (X + (Y << C)) & (CC << C)
7093 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
7094 match(Op0BO->getOperand(0),
7095 m_And(m_Shr(m_Value(V1), m_Value(V2)),
7096 m_ConstantInt(CC))) && V2 == Op1 &&
7097 cast<BinaryOperator>(Op0BO->getOperand(0))
7098 ->getOperand(0)->hasOneUse()) {
7099 Instruction *YS = BinaryOperator::CreateShl(
7100 Op0BO->getOperand(1), Op1,
7102 InsertNewInstBefore(YS, I); // (Y << C)
7104 BinaryOperator::CreateAnd(V1, ConstantExpr::getShl(CC, Op1),
7105 V1->getName()+".mask");
7106 InsertNewInstBefore(XM, I); // X & (CC << C)
7108 return BinaryOperator::Create(Op0BO->getOpcode(), XM, YS);
7116 // If the operand is an bitwise operator with a constant RHS, and the
7117 // shift is the only use, we can pull it out of the shift.
7118 if (ConstantInt *Op0C = dyn_cast<ConstantInt>(Op0BO->getOperand(1))) {
7119 bool isValid = true; // Valid only for And, Or, Xor
7120 bool highBitSet = false; // Transform if high bit of constant set?
7122 switch (Op0BO->getOpcode()) {
7123 default: isValid = false; break; // Do not perform transform!
7124 case Instruction::Add:
7125 isValid = isLeftShift;
7127 case Instruction::Or:
7128 case Instruction::Xor:
7131 case Instruction::And:
7136 // If this is a signed shift right, and the high bit is modified
7137 // by the logical operation, do not perform the transformation.
7138 // The highBitSet boolean indicates the value of the high bit of
7139 // the constant which would cause it to be modified for this
7142 if (isValid && I.getOpcode() == Instruction::AShr)
7143 isValid = Op0C->getValue()[TypeBits-1] == highBitSet;
7146 Constant *NewRHS = ConstantExpr::get(I.getOpcode(), Op0C, Op1);
7148 Instruction *NewShift =
7149 BinaryOperator::Create(I.getOpcode(), Op0BO->getOperand(0), Op1);
7150 InsertNewInstBefore(NewShift, I);
7151 NewShift->takeName(Op0BO);
7153 return BinaryOperator::Create(Op0BO->getOpcode(), NewShift,
7160 // Find out if this is a shift of a shift by a constant.
7161 BinaryOperator *ShiftOp = dyn_cast<BinaryOperator>(Op0);
7162 if (ShiftOp && !ShiftOp->isShift())
7165 if (ShiftOp && isa<ConstantInt>(ShiftOp->getOperand(1))) {
7166 ConstantInt *ShiftAmt1C = cast<ConstantInt>(ShiftOp->getOperand(1));
7167 uint32_t ShiftAmt1 = ShiftAmt1C->getLimitedValue(TypeBits);
7168 uint32_t ShiftAmt2 = Op1->getLimitedValue(TypeBits);
7169 assert(ShiftAmt2 != 0 && "Should have been simplified earlier");
7170 if (ShiftAmt1 == 0) return 0; // Will be simplified in the future.
7171 Value *X = ShiftOp->getOperand(0);
7173 uint32_t AmtSum = ShiftAmt1+ShiftAmt2; // Fold into one big shift.
7174 if (AmtSum > TypeBits)
7177 const IntegerType *Ty = cast<IntegerType>(I.getType());
7179 // Check for (X << c1) << c2 and (X >> c1) >> c2
7180 if (I.getOpcode() == ShiftOp->getOpcode()) {
7181 return BinaryOperator::Create(I.getOpcode(), X,
7182 ConstantInt::get(Ty, AmtSum));
7183 } else if (ShiftOp->getOpcode() == Instruction::LShr &&
7184 I.getOpcode() == Instruction::AShr) {
7185 // ((X >>u C1) >>s C2) -> (X >>u (C1+C2)) since C1 != 0.
7186 return BinaryOperator::CreateLShr(X, ConstantInt::get(Ty, AmtSum));
7187 } else if (ShiftOp->getOpcode() == Instruction::AShr &&
7188 I.getOpcode() == Instruction::LShr) {
7189 // ((X >>s C1) >>u C2) -> ((X >>s (C1+C2)) & mask) since C1 != 0.
7190 Instruction *Shift =
7191 BinaryOperator::CreateAShr(X, ConstantInt::get(Ty, AmtSum));
7192 InsertNewInstBefore(Shift, I);
7194 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
7195 return BinaryOperator::CreateAnd(Shift, ConstantInt::get(Mask));
7198 // Okay, if we get here, one shift must be left, and the other shift must be
7199 // right. See if the amounts are equal.
7200 if (ShiftAmt1 == ShiftAmt2) {
7201 // If we have ((X >>? C) << C), turn this into X & (-1 << C).
7202 if (I.getOpcode() == Instruction::Shl) {
7203 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt1));
7204 return BinaryOperator::CreateAnd(X, ConstantInt::get(Mask));
7206 // If we have ((X << C) >>u C), turn this into X & (-1 >>u C).
7207 if (I.getOpcode() == Instruction::LShr) {
7208 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt1));
7209 return BinaryOperator::CreateAnd(X, ConstantInt::get(Mask));
7211 // We can simplify ((X << C) >>s C) into a trunc + sext.
7212 // NOTE: we could do this for any C, but that would make 'unusual' integer
7213 // types. For now, just stick to ones well-supported by the code
7215 const Type *SExtType = 0;
7216 switch (Ty->getBitWidth() - ShiftAmt1) {
7223 SExtType = IntegerType::get(Ty->getBitWidth() - ShiftAmt1);
7228 Instruction *NewTrunc = new TruncInst(X, SExtType, "sext");
7229 InsertNewInstBefore(NewTrunc, I);
7230 return new SExtInst(NewTrunc, Ty);
7232 // Otherwise, we can't handle it yet.
7233 } else if (ShiftAmt1 < ShiftAmt2) {
7234 uint32_t ShiftDiff = ShiftAmt2-ShiftAmt1;
7236 // (X >>? C1) << C2 --> X << (C2-C1) & (-1 << C2)
7237 if (I.getOpcode() == Instruction::Shl) {
7238 assert(ShiftOp->getOpcode() == Instruction::LShr ||
7239 ShiftOp->getOpcode() == Instruction::AShr);
7240 Instruction *Shift =
7241 BinaryOperator::CreateShl(X, ConstantInt::get(Ty, ShiftDiff));
7242 InsertNewInstBefore(Shift, I);
7244 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
7245 return BinaryOperator::CreateAnd(Shift, ConstantInt::get(Mask));
7248 // (X << C1) >>u C2 --> X >>u (C2-C1) & (-1 >> C2)
7249 if (I.getOpcode() == Instruction::LShr) {
7250 assert(ShiftOp->getOpcode() == Instruction::Shl);
7251 Instruction *Shift =
7252 BinaryOperator::CreateLShr(X, ConstantInt::get(Ty, ShiftDiff));
7253 InsertNewInstBefore(Shift, I);
7255 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
7256 return BinaryOperator::CreateAnd(Shift, ConstantInt::get(Mask));
7259 // We can't handle (X << C1) >>s C2, it shifts arbitrary bits in.
7261 assert(ShiftAmt2 < ShiftAmt1);
7262 uint32_t ShiftDiff = ShiftAmt1-ShiftAmt2;
7264 // (X >>? C1) << C2 --> X >>? (C1-C2) & (-1 << C2)
7265 if (I.getOpcode() == Instruction::Shl) {
7266 assert(ShiftOp->getOpcode() == Instruction::LShr ||
7267 ShiftOp->getOpcode() == Instruction::AShr);
7268 Instruction *Shift =
7269 BinaryOperator::Create(ShiftOp->getOpcode(), X,
7270 ConstantInt::get(Ty, ShiftDiff));
7271 InsertNewInstBefore(Shift, I);
7273 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
7274 return BinaryOperator::CreateAnd(Shift, ConstantInt::get(Mask));
7277 // (X << C1) >>u C2 --> X << (C1-C2) & (-1 >> C2)
7278 if (I.getOpcode() == Instruction::LShr) {
7279 assert(ShiftOp->getOpcode() == Instruction::Shl);
7280 Instruction *Shift =
7281 BinaryOperator::CreateShl(X, ConstantInt::get(Ty, ShiftDiff));
7282 InsertNewInstBefore(Shift, I);
7284 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
7285 return BinaryOperator::CreateAnd(Shift, ConstantInt::get(Mask));
7288 // We can't handle (X << C1) >>a C2, it shifts arbitrary bits in.
7295 /// DecomposeSimpleLinearExpr - Analyze 'Val', seeing if it is a simple linear
7296 /// expression. If so, decompose it, returning some value X, such that Val is
7299 static Value *DecomposeSimpleLinearExpr(Value *Val, unsigned &Scale,
7301 assert(Val->getType() == Type::Int32Ty && "Unexpected allocation size type!");
7302 if (ConstantInt *CI = dyn_cast<ConstantInt>(Val)) {
7303 Offset = CI->getZExtValue();
7305 return ConstantInt::get(Type::Int32Ty, 0);
7306 } else if (BinaryOperator *I = dyn_cast<BinaryOperator>(Val)) {
7307 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
7308 if (I->getOpcode() == Instruction::Shl) {
7309 // This is a value scaled by '1 << the shift amt'.
7310 Scale = 1U << RHS->getZExtValue();
7312 return I->getOperand(0);
7313 } else if (I->getOpcode() == Instruction::Mul) {
7314 // This value is scaled by 'RHS'.
7315 Scale = RHS->getZExtValue();
7317 return I->getOperand(0);
7318 } else if (I->getOpcode() == Instruction::Add) {
7319 // We have X+C. Check to see if we really have (X*C2)+C1,
7320 // where C1 is divisible by C2.
7323 DecomposeSimpleLinearExpr(I->getOperand(0), SubScale, Offset);
7324 Offset += RHS->getZExtValue();
7331 // Otherwise, we can't look past this.
7338 /// PromoteCastOfAllocation - If we find a cast of an allocation instruction,
7339 /// try to eliminate the cast by moving the type information into the alloc.
7340 Instruction *InstCombiner::PromoteCastOfAllocation(BitCastInst &CI,
7341 AllocationInst &AI) {
7342 const PointerType *PTy = cast<PointerType>(CI.getType());
7344 // Remove any uses of AI that are dead.
7345 assert(!CI.use_empty() && "Dead instructions should be removed earlier!");
7347 for (Value::use_iterator UI = AI.use_begin(), E = AI.use_end(); UI != E; ) {
7348 Instruction *User = cast<Instruction>(*UI++);
7349 if (isInstructionTriviallyDead(User)) {
7350 while (UI != E && *UI == User)
7351 ++UI; // If this instruction uses AI more than once, don't break UI.
7354 DOUT << "IC: DCE: " << *User;
7355 EraseInstFromFunction(*User);
7359 // Get the type really allocated and the type casted to.
7360 const Type *AllocElTy = AI.getAllocatedType();
7361 const Type *CastElTy = PTy->getElementType();
7362 if (!AllocElTy->isSized() || !CastElTy->isSized()) return 0;
7364 unsigned AllocElTyAlign = TD->getABITypeAlignment(AllocElTy);
7365 unsigned CastElTyAlign = TD->getABITypeAlignment(CastElTy);
7366 if (CastElTyAlign < AllocElTyAlign) return 0;
7368 // If the allocation has multiple uses, only promote it if we are strictly
7369 // increasing the alignment of the resultant allocation. If we keep it the
7370 // same, we open the door to infinite loops of various kinds.
7371 if (!AI.hasOneUse() && CastElTyAlign == AllocElTyAlign) return 0;
7373 uint64_t AllocElTySize = TD->getABITypeSize(AllocElTy);
7374 uint64_t CastElTySize = TD->getABITypeSize(CastElTy);
7375 if (CastElTySize == 0 || AllocElTySize == 0) return 0;
7377 // See if we can satisfy the modulus by pulling a scale out of the array
7379 unsigned ArraySizeScale;
7381 Value *NumElements = // See if the array size is a decomposable linear expr.
7382 DecomposeSimpleLinearExpr(AI.getOperand(0), ArraySizeScale, ArrayOffset);
7384 // If we can now satisfy the modulus, by using a non-1 scale, we really can
7386 if ((AllocElTySize*ArraySizeScale) % CastElTySize != 0 ||
7387 (AllocElTySize*ArrayOffset ) % CastElTySize != 0) return 0;
7389 unsigned Scale = (AllocElTySize*ArraySizeScale)/CastElTySize;
7394 // If the allocation size is constant, form a constant mul expression
7395 Amt = ConstantInt::get(Type::Int32Ty, Scale);
7396 if (isa<ConstantInt>(NumElements))
7397 Amt = Multiply(cast<ConstantInt>(NumElements), cast<ConstantInt>(Amt));
7398 // otherwise multiply the amount and the number of elements
7399 else if (Scale != 1) {
7400 Instruction *Tmp = BinaryOperator::CreateMul(Amt, NumElements, "tmp");
7401 Amt = InsertNewInstBefore(Tmp, AI);
7405 if (int Offset = (AllocElTySize*ArrayOffset)/CastElTySize) {
7406 Value *Off = ConstantInt::get(Type::Int32Ty, Offset, true);
7407 Instruction *Tmp = BinaryOperator::CreateAdd(Amt, Off, "tmp");
7408 Amt = InsertNewInstBefore(Tmp, AI);
7411 AllocationInst *New;
7412 if (isa<MallocInst>(AI))
7413 New = new MallocInst(CastElTy, Amt, AI.getAlignment());
7415 New = new AllocaInst(CastElTy, Amt, AI.getAlignment());
7416 InsertNewInstBefore(New, AI);
7419 // If the allocation has multiple uses, insert a cast and change all things
7420 // that used it to use the new cast. This will also hack on CI, but it will
7422 if (!AI.hasOneUse()) {
7423 AddUsesToWorkList(AI);
7424 // New is the allocation instruction, pointer typed. AI is the original
7425 // allocation instruction, also pointer typed. Thus, cast to use is BitCast.
7426 CastInst *NewCast = new BitCastInst(New, AI.getType(), "tmpcast");
7427 InsertNewInstBefore(NewCast, AI);
7428 AI.replaceAllUsesWith(NewCast);
7430 return ReplaceInstUsesWith(CI, New);
7433 /// CanEvaluateInDifferentType - Return true if we can take the specified value
7434 /// and return it as type Ty without inserting any new casts and without
7435 /// changing the computed value. This is used by code that tries to decide
7436 /// whether promoting or shrinking integer operations to wider or smaller types
7437 /// will allow us to eliminate a truncate or extend.
7439 /// This is a truncation operation if Ty is smaller than V->getType(), or an
7440 /// extension operation if Ty is larger.
7442 /// If CastOpc is a truncation, then Ty will be a type smaller than V. We
7443 /// should return true if trunc(V) can be computed by computing V in the smaller
7444 /// type. If V is an instruction, then trunc(inst(x,y)) can be computed as
7445 /// inst(trunc(x),trunc(y)), which only makes sense if x and y can be
7446 /// efficiently truncated.
7448 /// If CastOpc is a sext or zext, we are asking if the low bits of the value can
7449 /// bit computed in a larger type, which is then and'd or sext_in_reg'd to get
7450 /// the final result.
7451 bool InstCombiner::CanEvaluateInDifferentType(Value *V, const IntegerType *Ty,
7453 int &NumCastsRemoved) {
7454 // We can always evaluate constants in another type.
7455 if (isa<ConstantInt>(V))
7458 Instruction *I = dyn_cast<Instruction>(V);
7459 if (!I) return false;
7461 const IntegerType *OrigTy = cast<IntegerType>(V->getType());
7463 // If this is an extension or truncate, we can often eliminate it.
7464 if (isa<TruncInst>(I) || isa<ZExtInst>(I) || isa<SExtInst>(I)) {
7465 // If this is a cast from the destination type, we can trivially eliminate
7466 // it, and this will remove a cast overall.
7467 if (I->getOperand(0)->getType() == Ty) {
7468 // If the first operand is itself a cast, and is eliminable, do not count
7469 // this as an eliminable cast. We would prefer to eliminate those two
7471 if (!isa<CastInst>(I->getOperand(0)) && I->hasOneUse())
7477 // We can't extend or shrink something that has multiple uses: doing so would
7478 // require duplicating the instruction in general, which isn't profitable.
7479 if (!I->hasOneUse()) return false;
7481 switch (I->getOpcode()) {
7482 case Instruction::Add:
7483 case Instruction::Sub:
7484 case Instruction::Mul:
7485 case Instruction::And:
7486 case Instruction::Or:
7487 case Instruction::Xor:
7488 // These operators can all arbitrarily be extended or truncated.
7489 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
7491 CanEvaluateInDifferentType(I->getOperand(1), Ty, CastOpc,
7494 case Instruction::Shl:
7495 // If we are truncating the result of this SHL, and if it's a shift of a
7496 // constant amount, we can always perform a SHL in a smaller type.
7497 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
7498 uint32_t BitWidth = Ty->getBitWidth();
7499 if (BitWidth < OrigTy->getBitWidth() &&
7500 CI->getLimitedValue(BitWidth) < BitWidth)
7501 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
7505 case Instruction::LShr:
7506 // If this is a truncate of a logical shr, we can truncate it to a smaller
7507 // lshr iff we know that the bits we would otherwise be shifting in are
7509 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
7510 uint32_t OrigBitWidth = OrigTy->getBitWidth();
7511 uint32_t BitWidth = Ty->getBitWidth();
7512 if (BitWidth < OrigBitWidth &&
7513 MaskedValueIsZero(I->getOperand(0),
7514 APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth)) &&
7515 CI->getLimitedValue(BitWidth) < BitWidth) {
7516 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
7521 case Instruction::ZExt:
7522 case Instruction::SExt:
7523 case Instruction::Trunc:
7524 // If this is the same kind of case as our original (e.g. zext+zext), we
7525 // can safely replace it. Note that replacing it does not reduce the number
7526 // of casts in the input.
7527 if (I->getOpcode() == CastOpc)
7530 case Instruction::Select: {
7531 SelectInst *SI = cast<SelectInst>(I);
7532 return CanEvaluateInDifferentType(SI->getTrueValue(), Ty, CastOpc,
7534 CanEvaluateInDifferentType(SI->getFalseValue(), Ty, CastOpc,
7537 case Instruction::PHI: {
7538 // We can change a phi if we can change all operands.
7539 PHINode *PN = cast<PHINode>(I);
7540 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
7541 if (!CanEvaluateInDifferentType(PN->getIncomingValue(i), Ty, CastOpc,
7547 // TODO: Can handle more cases here.
7554 /// EvaluateInDifferentType - Given an expression that
7555 /// CanEvaluateInDifferentType returns true for, actually insert the code to
7556 /// evaluate the expression.
7557 Value *InstCombiner::EvaluateInDifferentType(Value *V, const Type *Ty,
7559 if (Constant *C = dyn_cast<Constant>(V))
7560 return ConstantExpr::getIntegerCast(C, Ty, isSigned /*Sext or ZExt*/);
7562 // Otherwise, it must be an instruction.
7563 Instruction *I = cast<Instruction>(V);
7564 Instruction *Res = 0;
7565 switch (I->getOpcode()) {
7566 case Instruction::Add:
7567 case Instruction::Sub:
7568 case Instruction::Mul:
7569 case Instruction::And:
7570 case Instruction::Or:
7571 case Instruction::Xor:
7572 case Instruction::AShr:
7573 case Instruction::LShr:
7574 case Instruction::Shl: {
7575 Value *LHS = EvaluateInDifferentType(I->getOperand(0), Ty, isSigned);
7576 Value *RHS = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
7577 Res = BinaryOperator::Create((Instruction::BinaryOps)I->getOpcode(),
7581 case Instruction::Trunc:
7582 case Instruction::ZExt:
7583 case Instruction::SExt:
7584 // If the source type of the cast is the type we're trying for then we can
7585 // just return the source. There's no need to insert it because it is not
7587 if (I->getOperand(0)->getType() == Ty)
7588 return I->getOperand(0);
7590 // Otherwise, must be the same type of cast, so just reinsert a new one.
7591 Res = CastInst::Create(cast<CastInst>(I)->getOpcode(), I->getOperand(0),
7594 case Instruction::Select: {
7595 Value *True = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
7596 Value *False = EvaluateInDifferentType(I->getOperand(2), Ty, isSigned);
7597 Res = SelectInst::Create(I->getOperand(0), True, False);
7600 case Instruction::PHI: {
7601 PHINode *OPN = cast<PHINode>(I);
7602 PHINode *NPN = PHINode::Create(Ty);
7603 for (unsigned i = 0, e = OPN->getNumIncomingValues(); i != e; ++i) {
7604 Value *V =EvaluateInDifferentType(OPN->getIncomingValue(i), Ty, isSigned);
7605 NPN->addIncoming(V, OPN->getIncomingBlock(i));
7611 // TODO: Can handle more cases here.
7612 assert(0 && "Unreachable!");
7617 return InsertNewInstBefore(Res, *I);
7620 /// @brief Implement the transforms common to all CastInst visitors.
7621 Instruction *InstCombiner::commonCastTransforms(CastInst &CI) {
7622 Value *Src = CI.getOperand(0);
7624 // Many cases of "cast of a cast" are eliminable. If it's eliminable we just
7625 // eliminate it now.
7626 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
7627 if (Instruction::CastOps opc =
7628 isEliminableCastPair(CSrc, CI.getOpcode(), CI.getType(), TD)) {
7629 // The first cast (CSrc) is eliminable so we need to fix up or replace
7630 // the second cast (CI). CSrc will then have a good chance of being dead.
7631 return CastInst::Create(opc, CSrc->getOperand(0), CI.getType());
7635 // If we are casting a select then fold the cast into the select
7636 if (SelectInst *SI = dyn_cast<SelectInst>(Src))
7637 if (Instruction *NV = FoldOpIntoSelect(CI, SI, this))
7640 // If we are casting a PHI then fold the cast into the PHI
7641 if (isa<PHINode>(Src))
7642 if (Instruction *NV = FoldOpIntoPhi(CI))
7648 /// @brief Implement the transforms for cast of pointer (bitcast/ptrtoint)
7649 Instruction *InstCombiner::commonPointerCastTransforms(CastInst &CI) {
7650 Value *Src = CI.getOperand(0);
7652 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Src)) {
7653 // If casting the result of a getelementptr instruction with no offset, turn
7654 // this into a cast of the original pointer!
7655 if (GEP->hasAllZeroIndices()) {
7656 // Changing the cast operand is usually not a good idea but it is safe
7657 // here because the pointer operand is being replaced with another
7658 // pointer operand so the opcode doesn't need to change.
7660 CI.setOperand(0, GEP->getOperand(0));
7664 // If the GEP has a single use, and the base pointer is a bitcast, and the
7665 // GEP computes a constant offset, see if we can convert these three
7666 // instructions into fewer. This typically happens with unions and other
7667 // non-type-safe code.
7668 if (GEP->hasOneUse() && isa<BitCastInst>(GEP->getOperand(0))) {
7669 if (GEP->hasAllConstantIndices()) {
7670 // We are guaranteed to get a constant from EmitGEPOffset.
7671 ConstantInt *OffsetV = cast<ConstantInt>(EmitGEPOffset(GEP, CI, *this));
7672 int64_t Offset = OffsetV->getSExtValue();
7674 // Get the base pointer input of the bitcast, and the type it points to.
7675 Value *OrigBase = cast<BitCastInst>(GEP->getOperand(0))->getOperand(0);
7676 const Type *GEPIdxTy =
7677 cast<PointerType>(OrigBase->getType())->getElementType();
7678 if (GEPIdxTy->isSized()) {
7679 SmallVector<Value*, 8> NewIndices;
7681 // Start with the index over the outer type. Note that the type size
7682 // might be zero (even if the offset isn't zero) if the indexed type
7683 // is something like [0 x {int, int}]
7684 const Type *IntPtrTy = TD->getIntPtrType();
7685 int64_t FirstIdx = 0;
7686 if (int64_t TySize = TD->getABITypeSize(GEPIdxTy)) {
7687 FirstIdx = Offset/TySize;
7690 // Handle silly modulus not returning values values [0..TySize).
7694 assert(Offset >= 0);
7696 assert((uint64_t)Offset < (uint64_t)TySize &&"Out of range offset");
7699 NewIndices.push_back(ConstantInt::get(IntPtrTy, FirstIdx));
7701 // Index into the types. If we fail, set OrigBase to null.
7703 if (const StructType *STy = dyn_cast<StructType>(GEPIdxTy)) {
7704 const StructLayout *SL = TD->getStructLayout(STy);
7705 if (Offset < (int64_t)SL->getSizeInBytes()) {
7706 unsigned Elt = SL->getElementContainingOffset(Offset);
7707 NewIndices.push_back(ConstantInt::get(Type::Int32Ty, Elt));
7709 Offset -= SL->getElementOffset(Elt);
7710 GEPIdxTy = STy->getElementType(Elt);
7712 // Otherwise, we can't index into this, bail out.
7716 } else if (isa<ArrayType>(GEPIdxTy) || isa<VectorType>(GEPIdxTy)) {
7717 const SequentialType *STy = cast<SequentialType>(GEPIdxTy);
7718 if (uint64_t EltSize = TD->getABITypeSize(STy->getElementType())){
7719 NewIndices.push_back(ConstantInt::get(IntPtrTy,Offset/EltSize));
7722 NewIndices.push_back(ConstantInt::get(IntPtrTy, 0));
7724 GEPIdxTy = STy->getElementType();
7726 // Otherwise, we can't index into this, bail out.
7732 // If we were able to index down into an element, create the GEP
7733 // and bitcast the result. This eliminates one bitcast, potentially
7735 Instruction *NGEP = GetElementPtrInst::Create(OrigBase,
7737 NewIndices.end(), "");
7738 InsertNewInstBefore(NGEP, CI);
7739 NGEP->takeName(GEP);
7741 if (isa<BitCastInst>(CI))
7742 return new BitCastInst(NGEP, CI.getType());
7743 assert(isa<PtrToIntInst>(CI));
7744 return new PtrToIntInst(NGEP, CI.getType());
7751 return commonCastTransforms(CI);
7756 /// Only the TRUNC, ZEXT, SEXT, and BITCAST can both operand and result as
7757 /// integer types. This function implements the common transforms for all those
7759 /// @brief Implement the transforms common to CastInst with integer operands
7760 Instruction *InstCombiner::commonIntCastTransforms(CastInst &CI) {
7761 if (Instruction *Result = commonCastTransforms(CI))
7764 Value *Src = CI.getOperand(0);
7765 const Type *SrcTy = Src->getType();
7766 const Type *DestTy = CI.getType();
7767 uint32_t SrcBitSize = SrcTy->getPrimitiveSizeInBits();
7768 uint32_t DestBitSize = DestTy->getPrimitiveSizeInBits();
7770 // See if we can simplify any instructions used by the LHS whose sole
7771 // purpose is to compute bits we don't care about.
7772 APInt KnownZero(DestBitSize, 0), KnownOne(DestBitSize, 0);
7773 if (SimplifyDemandedBits(&CI, APInt::getAllOnesValue(DestBitSize),
7774 KnownZero, KnownOne))
7777 // If the source isn't an instruction or has more than one use then we
7778 // can't do anything more.
7779 Instruction *SrcI = dyn_cast<Instruction>(Src);
7780 if (!SrcI || !Src->hasOneUse())
7783 // Attempt to propagate the cast into the instruction for int->int casts.
7784 int NumCastsRemoved = 0;
7785 if (!isa<BitCastInst>(CI) &&
7786 CanEvaluateInDifferentType(SrcI, cast<IntegerType>(DestTy),
7787 CI.getOpcode(), NumCastsRemoved)) {
7788 // If this cast is a truncate, evaluting in a different type always
7789 // eliminates the cast, so it is always a win. If this is a zero-extension,
7790 // we need to do an AND to maintain the clear top-part of the computation,
7791 // so we require that the input have eliminated at least one cast. If this
7792 // is a sign extension, we insert two new casts (to do the extension) so we
7793 // require that two casts have been eliminated.
7795 switch (CI.getOpcode()) {
7797 // All the others use floating point so we shouldn't actually
7798 // get here because of the check above.
7799 assert(0 && "Unknown cast type");
7800 case Instruction::Trunc:
7803 case Instruction::ZExt:
7804 DoXForm = NumCastsRemoved >= 1;
7806 case Instruction::SExt:
7807 DoXForm = NumCastsRemoved >= 2;
7812 Value *Res = EvaluateInDifferentType(SrcI, DestTy,
7813 CI.getOpcode() == Instruction::SExt);
7814 assert(Res->getType() == DestTy);
7815 switch (CI.getOpcode()) {
7816 default: assert(0 && "Unknown cast type!");
7817 case Instruction::Trunc:
7818 case Instruction::BitCast:
7819 // Just replace this cast with the result.
7820 return ReplaceInstUsesWith(CI, Res);
7821 case Instruction::ZExt: {
7822 // We need to emit an AND to clear the high bits.
7823 assert(SrcBitSize < DestBitSize && "Not a zext?");
7824 Constant *C = ConstantInt::get(APInt::getLowBitsSet(DestBitSize,
7826 return BinaryOperator::CreateAnd(Res, C);
7828 case Instruction::SExt:
7829 // We need to emit a cast to truncate, then a cast to sext.
7830 return CastInst::Create(Instruction::SExt,
7831 InsertCastBefore(Instruction::Trunc, Res, Src->getType(),
7837 Value *Op0 = SrcI->getNumOperands() > 0 ? SrcI->getOperand(0) : 0;
7838 Value *Op1 = SrcI->getNumOperands() > 1 ? SrcI->getOperand(1) : 0;
7840 switch (SrcI->getOpcode()) {
7841 case Instruction::Add:
7842 case Instruction::Mul:
7843 case Instruction::And:
7844 case Instruction::Or:
7845 case Instruction::Xor:
7846 // If we are discarding information, rewrite.
7847 if (DestBitSize <= SrcBitSize && DestBitSize != 1) {
7848 // Don't insert two casts if they cannot be eliminated. We allow
7849 // two casts to be inserted if the sizes are the same. This could
7850 // only be converting signedness, which is a noop.
7851 if (DestBitSize == SrcBitSize ||
7852 !ValueRequiresCast(CI.getOpcode(), Op1, DestTy,TD) ||
7853 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
7854 Instruction::CastOps opcode = CI.getOpcode();
7855 Value *Op0c = InsertOperandCastBefore(opcode, Op0, DestTy, SrcI);
7856 Value *Op1c = InsertOperandCastBefore(opcode, Op1, DestTy, SrcI);
7857 return BinaryOperator::Create(
7858 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
7862 // cast (xor bool X, true) to int --> xor (cast bool X to int), 1
7863 if (isa<ZExtInst>(CI) && SrcBitSize == 1 &&
7864 SrcI->getOpcode() == Instruction::Xor &&
7865 Op1 == ConstantInt::getTrue() &&
7866 (!Op0->hasOneUse() || !isa<CmpInst>(Op0))) {
7867 Value *New = InsertOperandCastBefore(Instruction::ZExt, Op0, DestTy, &CI);
7868 return BinaryOperator::CreateXor(New, ConstantInt::get(CI.getType(), 1));
7871 case Instruction::SDiv:
7872 case Instruction::UDiv:
7873 case Instruction::SRem:
7874 case Instruction::URem:
7875 // If we are just changing the sign, rewrite.
7876 if (DestBitSize == SrcBitSize) {
7877 // Don't insert two casts if they cannot be eliminated. We allow
7878 // two casts to be inserted if the sizes are the same. This could
7879 // only be converting signedness, which is a noop.
7880 if (!ValueRequiresCast(CI.getOpcode(), Op1, DestTy, TD) ||
7881 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
7882 Value *Op0c = InsertOperandCastBefore(Instruction::BitCast,
7884 Value *Op1c = InsertOperandCastBefore(Instruction::BitCast,
7886 return BinaryOperator::Create(
7887 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
7892 case Instruction::Shl:
7893 // Allow changing the sign of the source operand. Do not allow
7894 // changing the size of the shift, UNLESS the shift amount is a
7895 // constant. We must not change variable sized shifts to a smaller
7896 // size, because it is undefined to shift more bits out than exist
7898 if (DestBitSize == SrcBitSize ||
7899 (DestBitSize < SrcBitSize && isa<Constant>(Op1))) {
7900 Instruction::CastOps opcode = (DestBitSize == SrcBitSize ?
7901 Instruction::BitCast : Instruction::Trunc);
7902 Value *Op0c = InsertOperandCastBefore(opcode, Op0, DestTy, SrcI);
7903 Value *Op1c = InsertOperandCastBefore(opcode, Op1, DestTy, SrcI);
7904 return BinaryOperator::CreateShl(Op0c, Op1c);
7907 case Instruction::AShr:
7908 // If this is a signed shr, and if all bits shifted in are about to be
7909 // truncated off, turn it into an unsigned shr to allow greater
7911 if (DestBitSize < SrcBitSize &&
7912 isa<ConstantInt>(Op1)) {
7913 uint32_t ShiftAmt = cast<ConstantInt>(Op1)->getLimitedValue(SrcBitSize);
7914 if (SrcBitSize > ShiftAmt && SrcBitSize-ShiftAmt >= DestBitSize) {
7915 // Insert the new logical shift right.
7916 return BinaryOperator::CreateLShr(Op0, Op1);
7924 Instruction *InstCombiner::visitTrunc(TruncInst &CI) {
7925 if (Instruction *Result = commonIntCastTransforms(CI))
7928 Value *Src = CI.getOperand(0);
7929 const Type *Ty = CI.getType();
7930 uint32_t DestBitWidth = Ty->getPrimitiveSizeInBits();
7931 uint32_t SrcBitWidth = cast<IntegerType>(Src->getType())->getBitWidth();
7933 if (Instruction *SrcI = dyn_cast<Instruction>(Src)) {
7934 switch (SrcI->getOpcode()) {
7936 case Instruction::LShr:
7937 // We can shrink lshr to something smaller if we know the bits shifted in
7938 // are already zeros.
7939 if (ConstantInt *ShAmtV = dyn_cast<ConstantInt>(SrcI->getOperand(1))) {
7940 uint32_t ShAmt = ShAmtV->getLimitedValue(SrcBitWidth);
7942 // Get a mask for the bits shifting in.
7943 APInt Mask(APInt::getLowBitsSet(SrcBitWidth, ShAmt).shl(DestBitWidth));
7944 Value* SrcIOp0 = SrcI->getOperand(0);
7945 if (SrcI->hasOneUse() && MaskedValueIsZero(SrcIOp0, Mask)) {
7946 if (ShAmt >= DestBitWidth) // All zeros.
7947 return ReplaceInstUsesWith(CI, Constant::getNullValue(Ty));
7949 // Okay, we can shrink this. Truncate the input, then return a new
7951 Value *V1 = InsertCastBefore(Instruction::Trunc, SrcIOp0, Ty, CI);
7952 Value *V2 = InsertCastBefore(Instruction::Trunc, SrcI->getOperand(1),
7954 return BinaryOperator::CreateLShr(V1, V2);
7956 } else { // This is a variable shr.
7958 // Turn 'trunc (lshr X, Y) to bool' into '(X & (1 << Y)) != 0'. This is
7959 // more LLVM instructions, but allows '1 << Y' to be hoisted if
7960 // loop-invariant and CSE'd.
7961 if (CI.getType() == Type::Int1Ty && SrcI->hasOneUse()) {
7962 Value *One = ConstantInt::get(SrcI->getType(), 1);
7964 Value *V = InsertNewInstBefore(
7965 BinaryOperator::CreateShl(One, SrcI->getOperand(1),
7967 V = InsertNewInstBefore(BinaryOperator::CreateAnd(V,
7968 SrcI->getOperand(0),
7970 Value *Zero = Constant::getNullValue(V->getType());
7971 return new ICmpInst(ICmpInst::ICMP_NE, V, Zero);
7981 /// transformZExtICmp - Transform (zext icmp) to bitwise / integer operations
7982 /// in order to eliminate the icmp.
7983 Instruction *InstCombiner::transformZExtICmp(ICmpInst *ICI, Instruction &CI,
7985 // If we are just checking for a icmp eq of a single bit and zext'ing it
7986 // to an integer, then shift the bit to the appropriate place and then
7987 // cast to integer to avoid the comparison.
7988 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(ICI->getOperand(1))) {
7989 const APInt &Op1CV = Op1C->getValue();
7991 // zext (x <s 0) to i32 --> x>>u31 true if signbit set.
7992 // zext (x >s -1) to i32 --> (x>>u31)^1 true if signbit clear.
7993 if ((ICI->getPredicate() == ICmpInst::ICMP_SLT && Op1CV == 0) ||
7994 (ICI->getPredicate() == ICmpInst::ICMP_SGT &&Op1CV.isAllOnesValue())) {
7995 if (!DoXform) return ICI;
7997 Value *In = ICI->getOperand(0);
7998 Value *Sh = ConstantInt::get(In->getType(),
7999 In->getType()->getPrimitiveSizeInBits()-1);
8000 In = InsertNewInstBefore(BinaryOperator::CreateLShr(In, Sh,
8001 In->getName()+".lobit"),
8003 if (In->getType() != CI.getType())
8004 In = CastInst::CreateIntegerCast(In, CI.getType(),
8005 false/*ZExt*/, "tmp", &CI);
8007 if (ICI->getPredicate() == ICmpInst::ICMP_SGT) {
8008 Constant *One = ConstantInt::get(In->getType(), 1);
8009 In = InsertNewInstBefore(BinaryOperator::CreateXor(In, One,
8010 In->getName()+".not"),
8014 return ReplaceInstUsesWith(CI, In);
8019 // zext (X == 0) to i32 --> X^1 iff X has only the low bit set.
8020 // zext (X == 0) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
8021 // zext (X == 1) to i32 --> X iff X has only the low bit set.
8022 // zext (X == 2) to i32 --> X>>1 iff X has only the 2nd bit set.
8023 // zext (X != 0) to i32 --> X iff X has only the low bit set.
8024 // zext (X != 0) to i32 --> X>>1 iff X has only the 2nd bit set.
8025 // zext (X != 1) to i32 --> X^1 iff X has only the low bit set.
8026 // zext (X != 2) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
8027 if ((Op1CV == 0 || Op1CV.isPowerOf2()) &&
8028 // This only works for EQ and NE
8029 ICI->isEquality()) {
8030 // If Op1C some other power of two, convert:
8031 uint32_t BitWidth = Op1C->getType()->getBitWidth();
8032 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
8033 APInt TypeMask(APInt::getAllOnesValue(BitWidth));
8034 ComputeMaskedBits(ICI->getOperand(0), TypeMask, KnownZero, KnownOne);
8036 APInt KnownZeroMask(~KnownZero);
8037 if (KnownZeroMask.isPowerOf2()) { // Exactly 1 possible 1?
8038 if (!DoXform) return ICI;
8040 bool isNE = ICI->getPredicate() == ICmpInst::ICMP_NE;
8041 if (Op1CV != 0 && (Op1CV != KnownZeroMask)) {
8042 // (X&4) == 2 --> false
8043 // (X&4) != 2 --> true
8044 Constant *Res = ConstantInt::get(Type::Int1Ty, isNE);
8045 Res = ConstantExpr::getZExt(Res, CI.getType());
8046 return ReplaceInstUsesWith(CI, Res);
8049 uint32_t ShiftAmt = KnownZeroMask.logBase2();
8050 Value *In = ICI->getOperand(0);
8052 // Perform a logical shr by shiftamt.
8053 // Insert the shift to put the result in the low bit.
8054 In = InsertNewInstBefore(BinaryOperator::CreateLShr(In,
8055 ConstantInt::get(In->getType(), ShiftAmt),
8056 In->getName()+".lobit"), CI);
8059 if ((Op1CV != 0) == isNE) { // Toggle the low bit.
8060 Constant *One = ConstantInt::get(In->getType(), 1);
8061 In = BinaryOperator::CreateXor(In, One, "tmp");
8062 InsertNewInstBefore(cast<Instruction>(In), CI);
8065 if (CI.getType() == In->getType())
8066 return ReplaceInstUsesWith(CI, In);
8068 return CastInst::CreateIntegerCast(In, CI.getType(), false/*ZExt*/);
8076 Instruction *InstCombiner::visitZExt(ZExtInst &CI) {
8077 // If one of the common conversion will work ..
8078 if (Instruction *Result = commonIntCastTransforms(CI))
8081 Value *Src = CI.getOperand(0);
8083 // If this is a cast of a cast
8084 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
8085 // If this is a TRUNC followed by a ZEXT then we are dealing with integral
8086 // types and if the sizes are just right we can convert this into a logical
8087 // 'and' which will be much cheaper than the pair of casts.
8088 if (isa<TruncInst>(CSrc)) {
8089 // Get the sizes of the types involved
8090 Value *A = CSrc->getOperand(0);
8091 uint32_t SrcSize = A->getType()->getPrimitiveSizeInBits();
8092 uint32_t MidSize = CSrc->getType()->getPrimitiveSizeInBits();
8093 uint32_t DstSize = CI.getType()->getPrimitiveSizeInBits();
8094 // If we're actually extending zero bits and the trunc is a no-op
8095 if (MidSize < DstSize && SrcSize == DstSize) {
8096 // Replace both of the casts with an And of the type mask.
8097 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
8098 Constant *AndConst = ConstantInt::get(AndValue);
8100 BinaryOperator::CreateAnd(CSrc->getOperand(0), AndConst);
8101 // Unfortunately, if the type changed, we need to cast it back.
8102 if (And->getType() != CI.getType()) {
8103 And->setName(CSrc->getName()+".mask");
8104 InsertNewInstBefore(And, CI);
8105 And = CastInst::CreateIntegerCast(And, CI.getType(), false/*ZExt*/);
8112 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src))
8113 return transformZExtICmp(ICI, CI);
8115 BinaryOperator *SrcI = dyn_cast<BinaryOperator>(Src);
8116 if (SrcI && SrcI->getOpcode() == Instruction::Or) {
8117 // zext (or icmp, icmp) --> or (zext icmp), (zext icmp) if at least one
8118 // of the (zext icmp) will be transformed.
8119 ICmpInst *LHS = dyn_cast<ICmpInst>(SrcI->getOperand(0));
8120 ICmpInst *RHS = dyn_cast<ICmpInst>(SrcI->getOperand(1));
8121 if (LHS && RHS && LHS->hasOneUse() && RHS->hasOneUse() &&
8122 (transformZExtICmp(LHS, CI, false) ||
8123 transformZExtICmp(RHS, CI, false))) {
8124 Value *LCast = InsertCastBefore(Instruction::ZExt, LHS, CI.getType(), CI);
8125 Value *RCast = InsertCastBefore(Instruction::ZExt, RHS, CI.getType(), CI);
8126 return BinaryOperator::Create(Instruction::Or, LCast, RCast);
8133 Instruction *InstCombiner::visitSExt(SExtInst &CI) {
8134 if (Instruction *I = commonIntCastTransforms(CI))
8137 Value *Src = CI.getOperand(0);
8139 // Canonicalize sign-extend from i1 to a select.
8140 if (Src->getType() == Type::Int1Ty)
8141 return SelectInst::Create(Src,
8142 ConstantInt::getAllOnesValue(CI.getType()),
8143 Constant::getNullValue(CI.getType()));
8145 // See if the value being truncated is already sign extended. If so, just
8146 // eliminate the trunc/sext pair.
8147 if (getOpcode(Src) == Instruction::Trunc) {
8148 Value *Op = cast<User>(Src)->getOperand(0);
8149 unsigned OpBits = cast<IntegerType>(Op->getType())->getBitWidth();
8150 unsigned MidBits = cast<IntegerType>(Src->getType())->getBitWidth();
8151 unsigned DestBits = cast<IntegerType>(CI.getType())->getBitWidth();
8152 unsigned NumSignBits = ComputeNumSignBits(Op);
8154 if (OpBits == DestBits) {
8155 // Op is i32, Mid is i8, and Dest is i32. If Op has more than 24 sign
8156 // bits, it is already ready.
8157 if (NumSignBits > DestBits-MidBits)
8158 return ReplaceInstUsesWith(CI, Op);
8159 } else if (OpBits < DestBits) {
8160 // Op is i32, Mid is i8, and Dest is i64. If Op has more than 24 sign
8161 // bits, just sext from i32.
8162 if (NumSignBits > OpBits-MidBits)
8163 return new SExtInst(Op, CI.getType(), "tmp");
8165 // Op is i64, Mid is i8, and Dest is i32. If Op has more than 56 sign
8166 // bits, just truncate to i32.
8167 if (NumSignBits > OpBits-MidBits)
8168 return new TruncInst(Op, CI.getType(), "tmp");
8172 // If the input is a shl/ashr pair of a same constant, then this is a sign
8173 // extension from a smaller value. If we could trust arbitrary bitwidth
8174 // integers, we could turn this into a truncate to the smaller bit and then
8175 // use a sext for the whole extension. Since we don't, look deeper and check
8176 // for a truncate. If the source and dest are the same type, eliminate the
8177 // trunc and extend and just do shifts. For example, turn:
8178 // %a = trunc i32 %i to i8
8179 // %b = shl i8 %a, 6
8180 // %c = ashr i8 %b, 6
8181 // %d = sext i8 %c to i32
8183 // %a = shl i32 %i, 30
8184 // %d = ashr i32 %a, 30
8186 ConstantInt *BA = 0, *CA = 0;
8187 if (match(Src, m_AShr(m_Shl(m_Value(A), m_ConstantInt(BA)),
8188 m_ConstantInt(CA))) &&
8189 BA == CA && isa<TruncInst>(A)) {
8190 Value *I = cast<TruncInst>(A)->getOperand(0);
8191 if (I->getType() == CI.getType()) {
8192 unsigned MidSize = Src->getType()->getPrimitiveSizeInBits();
8193 unsigned SrcDstSize = CI.getType()->getPrimitiveSizeInBits();
8194 unsigned ShAmt = CA->getZExtValue()+SrcDstSize-MidSize;
8195 Constant *ShAmtV = ConstantInt::get(CI.getType(), ShAmt);
8196 I = InsertNewInstBefore(BinaryOperator::CreateShl(I, ShAmtV,
8198 return BinaryOperator::CreateAShr(I, ShAmtV);
8205 /// FitsInFPType - Return a Constant* for the specified FP constant if it fits
8206 /// in the specified FP type without changing its value.
8207 static Constant *FitsInFPType(ConstantFP *CFP, const fltSemantics &Sem) {
8209 APFloat F = CFP->getValueAPF();
8210 (void)F.convert(Sem, APFloat::rmNearestTiesToEven, &losesInfo);
8212 return ConstantFP::get(F);
8216 /// LookThroughFPExtensions - If this is an fp extension instruction, look
8217 /// through it until we get the source value.
8218 static Value *LookThroughFPExtensions(Value *V) {
8219 if (Instruction *I = dyn_cast<Instruction>(V))
8220 if (I->getOpcode() == Instruction::FPExt)
8221 return LookThroughFPExtensions(I->getOperand(0));
8223 // If this value is a constant, return the constant in the smallest FP type
8224 // that can accurately represent it. This allows us to turn
8225 // (float)((double)X+2.0) into x+2.0f.
8226 if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
8227 if (CFP->getType() == Type::PPC_FP128Ty)
8228 return V; // No constant folding of this.
8229 // See if the value can be truncated to float and then reextended.
8230 if (Value *V = FitsInFPType(CFP, APFloat::IEEEsingle))
8232 if (CFP->getType() == Type::DoubleTy)
8233 return V; // Won't shrink.
8234 if (Value *V = FitsInFPType(CFP, APFloat::IEEEdouble))
8236 // Don't try to shrink to various long double types.
8242 Instruction *InstCombiner::visitFPTrunc(FPTruncInst &CI) {
8243 if (Instruction *I = commonCastTransforms(CI))
8246 // If we have fptrunc(add (fpextend x), (fpextend y)), where x and y are
8247 // smaller than the destination type, we can eliminate the truncate by doing
8248 // the add as the smaller type. This applies to add/sub/mul/div as well as
8249 // many builtins (sqrt, etc).
8250 BinaryOperator *OpI = dyn_cast<BinaryOperator>(CI.getOperand(0));
8251 if (OpI && OpI->hasOneUse()) {
8252 switch (OpI->getOpcode()) {
8254 case Instruction::Add:
8255 case Instruction::Sub:
8256 case Instruction::Mul:
8257 case Instruction::FDiv:
8258 case Instruction::FRem:
8259 const Type *SrcTy = OpI->getType();
8260 Value *LHSTrunc = LookThroughFPExtensions(OpI->getOperand(0));
8261 Value *RHSTrunc = LookThroughFPExtensions(OpI->getOperand(1));
8262 if (LHSTrunc->getType() != SrcTy &&
8263 RHSTrunc->getType() != SrcTy) {
8264 unsigned DstSize = CI.getType()->getPrimitiveSizeInBits();
8265 // If the source types were both smaller than the destination type of
8266 // the cast, do this xform.
8267 if (LHSTrunc->getType()->getPrimitiveSizeInBits() <= DstSize &&
8268 RHSTrunc->getType()->getPrimitiveSizeInBits() <= DstSize) {
8269 LHSTrunc = InsertCastBefore(Instruction::FPExt, LHSTrunc,
8271 RHSTrunc = InsertCastBefore(Instruction::FPExt, RHSTrunc,
8273 return BinaryOperator::Create(OpI->getOpcode(), LHSTrunc, RHSTrunc);
8282 Instruction *InstCombiner::visitFPExt(CastInst &CI) {
8283 return commonCastTransforms(CI);
8286 Instruction *InstCombiner::visitFPToUI(FPToUIInst &FI) {
8287 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0));
8289 return commonCastTransforms(FI);
8291 // fptoui(uitofp(X)) --> X
8292 // fptoui(sitofp(X)) --> X
8293 // This is safe if the intermediate type has enough bits in its mantissa to
8294 // accurately represent all values of X. For example, do not do this with
8295 // i64->float->i64. This is also safe for sitofp case, because any negative
8296 // 'X' value would cause an undefined result for the fptoui.
8297 if ((isa<UIToFPInst>(OpI) || isa<SIToFPInst>(OpI)) &&
8298 OpI->getOperand(0)->getType() == FI.getType() &&
8299 (int)FI.getType()->getPrimitiveSizeInBits() < /*extra bit for sign */
8300 OpI->getType()->getFPMantissaWidth())
8301 return ReplaceInstUsesWith(FI, OpI->getOperand(0));
8303 return commonCastTransforms(FI);
8306 Instruction *InstCombiner::visitFPToSI(FPToSIInst &FI) {
8307 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0));
8309 return commonCastTransforms(FI);
8311 // fptosi(sitofp(X)) --> X
8312 // fptosi(uitofp(X)) --> X
8313 // This is safe if the intermediate type has enough bits in its mantissa to
8314 // accurately represent all values of X. For example, do not do this with
8315 // i64->float->i64. This is also safe for sitofp case, because any negative
8316 // 'X' value would cause an undefined result for the fptoui.
8317 if ((isa<UIToFPInst>(OpI) || isa<SIToFPInst>(OpI)) &&
8318 OpI->getOperand(0)->getType() == FI.getType() &&
8319 (int)FI.getType()->getPrimitiveSizeInBits() <=
8320 OpI->getType()->getFPMantissaWidth())
8321 return ReplaceInstUsesWith(FI, OpI->getOperand(0));
8323 return commonCastTransforms(FI);
8326 Instruction *InstCombiner::visitUIToFP(CastInst &CI) {
8327 return commonCastTransforms(CI);
8330 Instruction *InstCombiner::visitSIToFP(CastInst &CI) {
8331 return commonCastTransforms(CI);
8334 Instruction *InstCombiner::visitPtrToInt(CastInst &CI) {
8335 return commonPointerCastTransforms(CI);
8338 Instruction *InstCombiner::visitIntToPtr(IntToPtrInst &CI) {
8339 if (Instruction *I = commonCastTransforms(CI))
8342 const Type *DestPointee = cast<PointerType>(CI.getType())->getElementType();
8343 if (!DestPointee->isSized()) return 0;
8345 // If this is inttoptr(add (ptrtoint x), cst), try to turn this into a GEP.
8348 if (match(CI.getOperand(0), m_Add(m_Cast<PtrToIntInst>(m_Value(X)),
8349 m_ConstantInt(Cst)))) {
8350 // If the source and destination operands have the same type, see if this
8351 // is a single-index GEP.
8352 if (X->getType() == CI.getType()) {
8353 // Get the size of the pointee type.
8354 uint64_t Size = TD->getABITypeSize(DestPointee);
8356 // Convert the constant to intptr type.
8357 APInt Offset = Cst->getValue();
8358 Offset.sextOrTrunc(TD->getPointerSizeInBits());
8360 // If Offset is evenly divisible by Size, we can do this xform.
8361 if (Size && !APIntOps::srem(Offset, APInt(Offset.getBitWidth(), Size))){
8362 Offset = APIntOps::sdiv(Offset, APInt(Offset.getBitWidth(), Size));
8363 return GetElementPtrInst::Create(X, ConstantInt::get(Offset));
8366 // TODO: Could handle other cases, e.g. where add is indexing into field of
8368 } else if (CI.getOperand(0)->hasOneUse() &&
8369 match(CI.getOperand(0), m_Add(m_Value(X), m_ConstantInt(Cst)))) {
8370 // Otherwise, if this is inttoptr(add x, cst), try to turn this into an
8371 // "inttoptr+GEP" instead of "add+intptr".
8373 // Get the size of the pointee type.
8374 uint64_t Size = TD->getABITypeSize(DestPointee);
8376 // Convert the constant to intptr type.
8377 APInt Offset = Cst->getValue();
8378 Offset.sextOrTrunc(TD->getPointerSizeInBits());
8380 // If Offset is evenly divisible by Size, we can do this xform.
8381 if (Size && !APIntOps::srem(Offset, APInt(Offset.getBitWidth(), Size))){
8382 Offset = APIntOps::sdiv(Offset, APInt(Offset.getBitWidth(), Size));
8384 Instruction *P = InsertNewInstBefore(new IntToPtrInst(X, CI.getType(),
8386 return GetElementPtrInst::Create(P, ConstantInt::get(Offset), "tmp");
8392 Instruction *InstCombiner::visitBitCast(BitCastInst &CI) {
8393 // If the operands are integer typed then apply the integer transforms,
8394 // otherwise just apply the common ones.
8395 Value *Src = CI.getOperand(0);
8396 const Type *SrcTy = Src->getType();
8397 const Type *DestTy = CI.getType();
8399 if (SrcTy->isInteger() && DestTy->isInteger()) {
8400 if (Instruction *Result = commonIntCastTransforms(CI))
8402 } else if (isa<PointerType>(SrcTy)) {
8403 if (Instruction *I = commonPointerCastTransforms(CI))
8406 if (Instruction *Result = commonCastTransforms(CI))
8411 // Get rid of casts from one type to the same type. These are useless and can
8412 // be replaced by the operand.
8413 if (DestTy == Src->getType())
8414 return ReplaceInstUsesWith(CI, Src);
8416 if (const PointerType *DstPTy = dyn_cast<PointerType>(DestTy)) {
8417 const PointerType *SrcPTy = cast<PointerType>(SrcTy);
8418 const Type *DstElTy = DstPTy->getElementType();
8419 const Type *SrcElTy = SrcPTy->getElementType();
8421 // If the address spaces don't match, don't eliminate the bitcast, which is
8422 // required for changing types.
8423 if (SrcPTy->getAddressSpace() != DstPTy->getAddressSpace())
8426 // If we are casting a malloc or alloca to a pointer to a type of the same
8427 // size, rewrite the allocation instruction to allocate the "right" type.
8428 if (AllocationInst *AI = dyn_cast<AllocationInst>(Src))
8429 if (Instruction *V = PromoteCastOfAllocation(CI, *AI))
8432 // If the source and destination are pointers, and this cast is equivalent
8433 // to a getelementptr X, 0, 0, 0... turn it into the appropriate gep.
8434 // This can enhance SROA and other transforms that want type-safe pointers.
8435 Constant *ZeroUInt = Constant::getNullValue(Type::Int32Ty);
8436 unsigned NumZeros = 0;
8437 while (SrcElTy != DstElTy &&
8438 isa<CompositeType>(SrcElTy) && !isa<PointerType>(SrcElTy) &&
8439 SrcElTy->getNumContainedTypes() /* not "{}" */) {
8440 SrcElTy = cast<CompositeType>(SrcElTy)->getTypeAtIndex(ZeroUInt);
8444 // If we found a path from the src to dest, create the getelementptr now.
8445 if (SrcElTy == DstElTy) {
8446 SmallVector<Value*, 8> Idxs(NumZeros+1, ZeroUInt);
8447 return GetElementPtrInst::Create(Src, Idxs.begin(), Idxs.end(), "",
8448 ((Instruction*) NULL));
8452 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(Src)) {
8453 if (SVI->hasOneUse()) {
8454 // Okay, we have (bitconvert (shuffle ..)). Check to see if this is
8455 // a bitconvert to a vector with the same # elts.
8456 if (isa<VectorType>(DestTy) &&
8457 cast<VectorType>(DestTy)->getNumElements() ==
8458 SVI->getType()->getNumElements() &&
8459 SVI->getType()->getNumElements() ==
8460 cast<VectorType>(SVI->getOperand(0)->getType())->getNumElements()) {
8462 // If either of the operands is a cast from CI.getType(), then
8463 // evaluating the shuffle in the casted destination's type will allow
8464 // us to eliminate at least one cast.
8465 if (((Tmp = dyn_cast<CastInst>(SVI->getOperand(0))) &&
8466 Tmp->getOperand(0)->getType() == DestTy) ||
8467 ((Tmp = dyn_cast<CastInst>(SVI->getOperand(1))) &&
8468 Tmp->getOperand(0)->getType() == DestTy)) {
8469 Value *LHS = InsertOperandCastBefore(Instruction::BitCast,
8470 SVI->getOperand(0), DestTy, &CI);
8471 Value *RHS = InsertOperandCastBefore(Instruction::BitCast,
8472 SVI->getOperand(1), DestTy, &CI);
8473 // Return a new shuffle vector. Use the same element ID's, as we
8474 // know the vector types match #elts.
8475 return new ShuffleVectorInst(LHS, RHS, SVI->getOperand(2));
8483 /// GetSelectFoldableOperands - We want to turn code that looks like this:
8485 /// %D = select %cond, %C, %A
8487 /// %C = select %cond, %B, 0
8490 /// Assuming that the specified instruction is an operand to the select, return
8491 /// a bitmask indicating which operands of this instruction are foldable if they
8492 /// equal the other incoming value of the select.
8494 static unsigned GetSelectFoldableOperands(Instruction *I) {
8495 switch (I->getOpcode()) {
8496 case Instruction::Add:
8497 case Instruction::Mul:
8498 case Instruction::And:
8499 case Instruction::Or:
8500 case Instruction::Xor:
8501 return 3; // Can fold through either operand.
8502 case Instruction::Sub: // Can only fold on the amount subtracted.
8503 case Instruction::Shl: // Can only fold on the shift amount.
8504 case Instruction::LShr:
8505 case Instruction::AShr:
8508 return 0; // Cannot fold
8512 /// GetSelectFoldableConstant - For the same transformation as the previous
8513 /// function, return the identity constant that goes into the select.
8514 static Constant *GetSelectFoldableConstant(Instruction *I) {
8515 switch (I->getOpcode()) {
8516 default: assert(0 && "This cannot happen!"); abort();
8517 case Instruction::Add:
8518 case Instruction::Sub:
8519 case Instruction::Or:
8520 case Instruction::Xor:
8521 case Instruction::Shl:
8522 case Instruction::LShr:
8523 case Instruction::AShr:
8524 return Constant::getNullValue(I->getType());
8525 case Instruction::And:
8526 return Constant::getAllOnesValue(I->getType());
8527 case Instruction::Mul:
8528 return ConstantInt::get(I->getType(), 1);
8532 /// FoldSelectOpOp - Here we have (select c, TI, FI), and we know that TI and FI
8533 /// have the same opcode and only one use each. Try to simplify this.
8534 Instruction *InstCombiner::FoldSelectOpOp(SelectInst &SI, Instruction *TI,
8536 if (TI->getNumOperands() == 1) {
8537 // If this is a non-volatile load or a cast from the same type,
8540 if (TI->getOperand(0)->getType() != FI->getOperand(0)->getType())
8543 return 0; // unknown unary op.
8546 // Fold this by inserting a select from the input values.
8547 SelectInst *NewSI = SelectInst::Create(SI.getCondition(), TI->getOperand(0),
8548 FI->getOperand(0), SI.getName()+".v");
8549 InsertNewInstBefore(NewSI, SI);
8550 return CastInst::Create(Instruction::CastOps(TI->getOpcode()), NewSI,
8554 // Only handle binary operators here.
8555 if (!isa<BinaryOperator>(TI))
8558 // Figure out if the operations have any operands in common.
8559 Value *MatchOp, *OtherOpT, *OtherOpF;
8561 if (TI->getOperand(0) == FI->getOperand(0)) {
8562 MatchOp = TI->getOperand(0);
8563 OtherOpT = TI->getOperand(1);
8564 OtherOpF = FI->getOperand(1);
8565 MatchIsOpZero = true;
8566 } else if (TI->getOperand(1) == FI->getOperand(1)) {
8567 MatchOp = TI->getOperand(1);
8568 OtherOpT = TI->getOperand(0);
8569 OtherOpF = FI->getOperand(0);
8570 MatchIsOpZero = false;
8571 } else if (!TI->isCommutative()) {
8573 } else if (TI->getOperand(0) == FI->getOperand(1)) {
8574 MatchOp = TI->getOperand(0);
8575 OtherOpT = TI->getOperand(1);
8576 OtherOpF = FI->getOperand(0);
8577 MatchIsOpZero = true;
8578 } else if (TI->getOperand(1) == FI->getOperand(0)) {
8579 MatchOp = TI->getOperand(1);
8580 OtherOpT = TI->getOperand(0);
8581 OtherOpF = FI->getOperand(1);
8582 MatchIsOpZero = true;
8587 // If we reach here, they do have operations in common.
8588 SelectInst *NewSI = SelectInst::Create(SI.getCondition(), OtherOpT,
8589 OtherOpF, SI.getName()+".v");
8590 InsertNewInstBefore(NewSI, SI);
8592 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TI)) {
8594 return BinaryOperator::Create(BO->getOpcode(), MatchOp, NewSI);
8596 return BinaryOperator::Create(BO->getOpcode(), NewSI, MatchOp);
8598 assert(0 && "Shouldn't get here");
8602 /// visitSelectInstWithICmp - Visit a SelectInst that has an
8603 /// ICmpInst as its first operand.
8605 Instruction *InstCombiner::visitSelectInstWithICmp(SelectInst &SI,
8607 bool Changed = false;
8608 ICmpInst::Predicate Pred = ICI->getPredicate();
8609 Value *CmpLHS = ICI->getOperand(0);
8610 Value *CmpRHS = ICI->getOperand(1);
8611 Value *TrueVal = SI.getTrueValue();
8612 Value *FalseVal = SI.getFalseValue();
8614 // Check cases where the comparison is with a constant that
8615 // can be adjusted to fit the min/max idiom. We may edit ICI in
8616 // place here, so make sure the select is the only user.
8617 if (ICI->hasOneUse())
8618 if (ConstantInt *CI = dyn_cast<ConstantInt>(CmpRHS)) {
8621 case ICmpInst::ICMP_ULT:
8622 case ICmpInst::ICMP_SLT: {
8623 // X < MIN ? T : F --> F
8624 if (CI->isMinValue(Pred == ICmpInst::ICMP_SLT))
8625 return ReplaceInstUsesWith(SI, FalseVal);
8626 // X < C ? X : C-1 --> X > C-1 ? C-1 : X
8627 Constant *AdjustedRHS = SubOne(CI);
8628 if ((CmpLHS == TrueVal && AdjustedRHS == FalseVal) ||
8629 (CmpLHS == FalseVal && AdjustedRHS == TrueVal)) {
8630 Pred = ICmpInst::getSwappedPredicate(Pred);
8631 CmpRHS = AdjustedRHS;
8632 std::swap(FalseVal, TrueVal);
8633 ICI->setPredicate(Pred);
8634 ICI->setOperand(1, CmpRHS);
8635 SI.setOperand(1, TrueVal);
8636 SI.setOperand(2, FalseVal);
8641 case ICmpInst::ICMP_UGT:
8642 case ICmpInst::ICMP_SGT: {
8643 // X > MAX ? T : F --> F
8644 if (CI->isMaxValue(Pred == ICmpInst::ICMP_SGT))
8645 return ReplaceInstUsesWith(SI, FalseVal);
8646 // X > C ? X : C+1 --> X < C+1 ? C+1 : X
8647 Constant *AdjustedRHS = AddOne(CI);
8648 if ((CmpLHS == TrueVal && AdjustedRHS == FalseVal) ||
8649 (CmpLHS == FalseVal && AdjustedRHS == TrueVal)) {
8650 Pred = ICmpInst::getSwappedPredicate(Pred);
8651 CmpRHS = AdjustedRHS;
8652 std::swap(FalseVal, TrueVal);
8653 ICI->setPredicate(Pred);
8654 ICI->setOperand(1, CmpRHS);
8655 SI.setOperand(1, TrueVal);
8656 SI.setOperand(2, FalseVal);
8663 // (x <s 0) ? -1 : 0 -> ashr x, 31 -> all ones if signed
8664 // (x >s -1) ? -1 : 0 -> ashr x, 31 -> all ones if not signed
8665 CmpInst::Predicate Pred = CmpInst::BAD_ICMP_PREDICATE;
8666 if (match(TrueVal, m_ConstantInt(-1)) &&
8667 match(FalseVal, m_ConstantInt(0)))
8668 Pred = ICI->getPredicate();
8669 else if (match(TrueVal, m_ConstantInt(0)) &&
8670 match(FalseVal, m_ConstantInt(-1)))
8671 Pred = CmpInst::getInversePredicate(ICI->getPredicate());
8673 if (Pred != CmpInst::BAD_ICMP_PREDICATE) {
8674 // If we are just checking for a icmp eq of a single bit and zext'ing it
8675 // to an integer, then shift the bit to the appropriate place and then
8676 // cast to integer to avoid the comparison.
8677 const APInt &Op1CV = CI->getValue();
8679 // sext (x <s 0) to i32 --> x>>s31 true if signbit set.
8680 // sext (x >s -1) to i32 --> (x>>s31)^-1 true if signbit clear.
8681 if ((Pred == ICmpInst::ICMP_SLT && Op1CV == 0) ||
8682 (Pred == ICmpInst::ICMP_SGT && Op1CV.isAllOnesValue())) {
8683 Value *In = ICI->getOperand(0);
8684 Value *Sh = ConstantInt::get(In->getType(),
8685 In->getType()->getPrimitiveSizeInBits()-1);
8686 In = InsertNewInstBefore(BinaryOperator::CreateAShr(In, Sh,
8687 In->getName()+".lobit"),
8689 if (In->getType() != SI.getType())
8690 In = CastInst::CreateIntegerCast(In, SI.getType(),
8691 true/*SExt*/, "tmp", ICI);
8693 if (Pred == ICmpInst::ICMP_SGT)
8694 In = InsertNewInstBefore(BinaryOperator::CreateNot(In,
8695 In->getName()+".not"), *ICI);
8697 return ReplaceInstUsesWith(SI, In);
8702 if (CmpLHS == TrueVal && CmpRHS == FalseVal) {
8703 // Transform (X == Y) ? X : Y -> Y
8704 if (Pred == ICmpInst::ICMP_EQ)
8705 return ReplaceInstUsesWith(SI, FalseVal);
8706 // Transform (X != Y) ? X : Y -> X
8707 if (Pred == ICmpInst::ICMP_NE)
8708 return ReplaceInstUsesWith(SI, TrueVal);
8709 /// NOTE: if we wanted to, this is where to detect integer MIN/MAX
8711 } else if (CmpLHS == FalseVal && CmpRHS == TrueVal) {
8712 // Transform (X == Y) ? Y : X -> X
8713 if (Pred == ICmpInst::ICMP_EQ)
8714 return ReplaceInstUsesWith(SI, FalseVal);
8715 // Transform (X != Y) ? Y : X -> Y
8716 if (Pred == ICmpInst::ICMP_NE)
8717 return ReplaceInstUsesWith(SI, TrueVal);
8718 /// NOTE: if we wanted to, this is where to detect integer MIN/MAX
8721 /// NOTE: if we wanted to, this is where to detect integer ABS
8723 return Changed ? &SI : 0;
8726 Instruction *InstCombiner::visitSelectInst(SelectInst &SI) {
8727 Value *CondVal = SI.getCondition();
8728 Value *TrueVal = SI.getTrueValue();
8729 Value *FalseVal = SI.getFalseValue();
8731 // select true, X, Y -> X
8732 // select false, X, Y -> Y
8733 if (ConstantInt *C = dyn_cast<ConstantInt>(CondVal))
8734 return ReplaceInstUsesWith(SI, C->getZExtValue() ? TrueVal : FalseVal);
8736 // select C, X, X -> X
8737 if (TrueVal == FalseVal)
8738 return ReplaceInstUsesWith(SI, TrueVal);
8740 if (isa<UndefValue>(TrueVal)) // select C, undef, X -> X
8741 return ReplaceInstUsesWith(SI, FalseVal);
8742 if (isa<UndefValue>(FalseVal)) // select C, X, undef -> X
8743 return ReplaceInstUsesWith(SI, TrueVal);
8744 if (isa<UndefValue>(CondVal)) { // select undef, X, Y -> X or Y
8745 if (isa<Constant>(TrueVal))
8746 return ReplaceInstUsesWith(SI, TrueVal);
8748 return ReplaceInstUsesWith(SI, FalseVal);
8751 if (SI.getType() == Type::Int1Ty) {
8752 if (ConstantInt *C = dyn_cast<ConstantInt>(TrueVal)) {
8753 if (C->getZExtValue()) {
8754 // Change: A = select B, true, C --> A = or B, C
8755 return BinaryOperator::CreateOr(CondVal, FalseVal);
8757 // Change: A = select B, false, C --> A = and !B, C
8759 InsertNewInstBefore(BinaryOperator::CreateNot(CondVal,
8760 "not."+CondVal->getName()), SI);
8761 return BinaryOperator::CreateAnd(NotCond, FalseVal);
8763 } else if (ConstantInt *C = dyn_cast<ConstantInt>(FalseVal)) {
8764 if (C->getZExtValue() == false) {
8765 // Change: A = select B, C, false --> A = and B, C
8766 return BinaryOperator::CreateAnd(CondVal, TrueVal);
8768 // Change: A = select B, C, true --> A = or !B, C
8770 InsertNewInstBefore(BinaryOperator::CreateNot(CondVal,
8771 "not."+CondVal->getName()), SI);
8772 return BinaryOperator::CreateOr(NotCond, TrueVal);
8776 // select a, b, a -> a&b
8777 // select a, a, b -> a|b
8778 if (CondVal == TrueVal)
8779 return BinaryOperator::CreateOr(CondVal, FalseVal);
8780 else if (CondVal == FalseVal)
8781 return BinaryOperator::CreateAnd(CondVal, TrueVal);
8784 // Selecting between two integer constants?
8785 if (ConstantInt *TrueValC = dyn_cast<ConstantInt>(TrueVal))
8786 if (ConstantInt *FalseValC = dyn_cast<ConstantInt>(FalseVal)) {
8787 // select C, 1, 0 -> zext C to int
8788 if (FalseValC->isZero() && TrueValC->getValue() == 1) {
8789 return CastInst::Create(Instruction::ZExt, CondVal, SI.getType());
8790 } else if (TrueValC->isZero() && FalseValC->getValue() == 1) {
8791 // select C, 0, 1 -> zext !C to int
8793 InsertNewInstBefore(BinaryOperator::CreateNot(CondVal,
8794 "not."+CondVal->getName()), SI);
8795 return CastInst::Create(Instruction::ZExt, NotCond, SI.getType());
8798 // FIXME: Turn select 0/-1 and -1/0 into sext from condition!
8800 if (ICmpInst *IC = dyn_cast<ICmpInst>(SI.getCondition())) {
8802 // (x <s 0) ? -1 : 0 -> ashr x, 31
8803 if (TrueValC->isAllOnesValue() && FalseValC->isZero())
8804 if (ConstantInt *CmpCst = dyn_cast<ConstantInt>(IC->getOperand(1))) {
8805 if (IC->getPredicate() == ICmpInst::ICMP_SLT && CmpCst->isZero()) {
8806 // The comparison constant and the result are not neccessarily the
8807 // same width. Make an all-ones value by inserting a AShr.
8808 Value *X = IC->getOperand(0);
8809 uint32_t Bits = X->getType()->getPrimitiveSizeInBits();
8810 Constant *ShAmt = ConstantInt::get(X->getType(), Bits-1);
8811 Instruction *SRA = BinaryOperator::Create(Instruction::AShr, X,
8813 InsertNewInstBefore(SRA, SI);
8815 // Finally, convert to the type of the select RHS. We figure out
8816 // if this requires a SExt, Trunc or BitCast based on the sizes.
8817 Instruction::CastOps opc = Instruction::BitCast;
8818 uint32_t SRASize = SRA->getType()->getPrimitiveSizeInBits();
8819 uint32_t SISize = SI.getType()->getPrimitiveSizeInBits();
8820 if (SRASize < SISize)
8821 opc = Instruction::SExt;
8822 else if (SRASize > SISize)
8823 opc = Instruction::Trunc;
8824 return CastInst::Create(opc, SRA, SI.getType());
8829 // If one of the constants is zero (we know they can't both be) and we
8830 // have an icmp instruction with zero, and we have an 'and' with the
8831 // non-constant value, eliminate this whole mess. This corresponds to
8832 // cases like this: ((X & 27) ? 27 : 0)
8833 if (TrueValC->isZero() || FalseValC->isZero())
8834 if (IC->isEquality() && isa<ConstantInt>(IC->getOperand(1)) &&
8835 cast<Constant>(IC->getOperand(1))->isNullValue())
8836 if (Instruction *ICA = dyn_cast<Instruction>(IC->getOperand(0)))
8837 if (ICA->getOpcode() == Instruction::And &&
8838 isa<ConstantInt>(ICA->getOperand(1)) &&
8839 (ICA->getOperand(1) == TrueValC ||
8840 ICA->getOperand(1) == FalseValC) &&
8841 isOneBitSet(cast<ConstantInt>(ICA->getOperand(1)))) {
8842 // Okay, now we know that everything is set up, we just don't
8843 // know whether we have a icmp_ne or icmp_eq and whether the
8844 // true or false val is the zero.
8845 bool ShouldNotVal = !TrueValC->isZero();
8846 ShouldNotVal ^= IC->getPredicate() == ICmpInst::ICMP_NE;
8849 V = InsertNewInstBefore(BinaryOperator::Create(
8850 Instruction::Xor, V, ICA->getOperand(1)), SI);
8851 return ReplaceInstUsesWith(SI, V);
8856 // See if we are selecting two values based on a comparison of the two values.
8857 if (FCmpInst *FCI = dyn_cast<FCmpInst>(CondVal)) {
8858 if (FCI->getOperand(0) == TrueVal && FCI->getOperand(1) == FalseVal) {
8859 // Transform (X == Y) ? X : Y -> Y
8860 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ) {
8861 // This is not safe in general for floating point:
8862 // consider X== -0, Y== +0.
8863 // It becomes safe if either operand is a nonzero constant.
8864 ConstantFP *CFPt, *CFPf;
8865 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) &&
8866 !CFPt->getValueAPF().isZero()) ||
8867 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) &&
8868 !CFPf->getValueAPF().isZero()))
8869 return ReplaceInstUsesWith(SI, FalseVal);
8871 // Transform (X != Y) ? X : Y -> X
8872 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
8873 return ReplaceInstUsesWith(SI, TrueVal);
8874 // NOTE: if we wanted to, this is where to detect MIN/MAX
8876 } else if (FCI->getOperand(0) == FalseVal && FCI->getOperand(1) == TrueVal){
8877 // Transform (X == Y) ? Y : X -> X
8878 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ) {
8879 // This is not safe in general for floating point:
8880 // consider X== -0, Y== +0.
8881 // It becomes safe if either operand is a nonzero constant.
8882 ConstantFP *CFPt, *CFPf;
8883 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) &&
8884 !CFPt->getValueAPF().isZero()) ||
8885 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) &&
8886 !CFPf->getValueAPF().isZero()))
8887 return ReplaceInstUsesWith(SI, FalseVal);
8889 // Transform (X != Y) ? Y : X -> Y
8890 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
8891 return ReplaceInstUsesWith(SI, TrueVal);
8892 // NOTE: if we wanted to, this is where to detect MIN/MAX
8894 // NOTE: if we wanted to, this is where to detect ABS
8897 // See if we are selecting two values based on a comparison of the two values.
8898 if (ICmpInst *ICI = dyn_cast<ICmpInst>(CondVal))
8899 if (Instruction *Result = visitSelectInstWithICmp(SI, ICI))
8902 if (Instruction *TI = dyn_cast<Instruction>(TrueVal))
8903 if (Instruction *FI = dyn_cast<Instruction>(FalseVal))
8904 if (TI->hasOneUse() && FI->hasOneUse()) {
8905 Instruction *AddOp = 0, *SubOp = 0;
8907 // Turn (select C, (op X, Y), (op X, Z)) -> (op X, (select C, Y, Z))
8908 if (TI->getOpcode() == FI->getOpcode())
8909 if (Instruction *IV = FoldSelectOpOp(SI, TI, FI))
8912 // Turn select C, (X+Y), (X-Y) --> (X+(select C, Y, (-Y))). This is
8913 // even legal for FP.
8914 if (TI->getOpcode() == Instruction::Sub &&
8915 FI->getOpcode() == Instruction::Add) {
8916 AddOp = FI; SubOp = TI;
8917 } else if (FI->getOpcode() == Instruction::Sub &&
8918 TI->getOpcode() == Instruction::Add) {
8919 AddOp = TI; SubOp = FI;
8923 Value *OtherAddOp = 0;
8924 if (SubOp->getOperand(0) == AddOp->getOperand(0)) {
8925 OtherAddOp = AddOp->getOperand(1);
8926 } else if (SubOp->getOperand(0) == AddOp->getOperand(1)) {
8927 OtherAddOp = AddOp->getOperand(0);
8931 // So at this point we know we have (Y -> OtherAddOp):
8932 // select C, (add X, Y), (sub X, Z)
8933 Value *NegVal; // Compute -Z
8934 if (Constant *C = dyn_cast<Constant>(SubOp->getOperand(1))) {
8935 NegVal = ConstantExpr::getNeg(C);
8937 NegVal = InsertNewInstBefore(
8938 BinaryOperator::CreateNeg(SubOp->getOperand(1), "tmp"), SI);
8941 Value *NewTrueOp = OtherAddOp;
8942 Value *NewFalseOp = NegVal;
8944 std::swap(NewTrueOp, NewFalseOp);
8945 Instruction *NewSel =
8946 SelectInst::Create(CondVal, NewTrueOp,
8947 NewFalseOp, SI.getName() + ".p");
8949 NewSel = InsertNewInstBefore(NewSel, SI);
8950 return BinaryOperator::CreateAdd(SubOp->getOperand(0), NewSel);
8955 // See if we can fold the select into one of our operands.
8956 if (SI.getType()->isInteger()) {
8957 // See the comment above GetSelectFoldableOperands for a description of the
8958 // transformation we are doing here.
8959 if (Instruction *TVI = dyn_cast<Instruction>(TrueVal))
8960 if (TVI->hasOneUse() && TVI->getNumOperands() == 2 &&
8961 !isa<Constant>(FalseVal))
8962 if (unsigned SFO = GetSelectFoldableOperands(TVI)) {
8963 unsigned OpToFold = 0;
8964 if ((SFO & 1) && FalseVal == TVI->getOperand(0)) {
8966 } else if ((SFO & 2) && FalseVal == TVI->getOperand(1)) {
8971 Constant *C = GetSelectFoldableConstant(TVI);
8972 Instruction *NewSel =
8973 SelectInst::Create(SI.getCondition(),
8974 TVI->getOperand(2-OpToFold), C);
8975 InsertNewInstBefore(NewSel, SI);
8976 NewSel->takeName(TVI);
8977 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TVI))
8978 return BinaryOperator::Create(BO->getOpcode(), FalseVal, NewSel);
8980 assert(0 && "Unknown instruction!!");
8985 if (Instruction *FVI = dyn_cast<Instruction>(FalseVal))
8986 if (FVI->hasOneUse() && FVI->getNumOperands() == 2 &&
8987 !isa<Constant>(TrueVal))
8988 if (unsigned SFO = GetSelectFoldableOperands(FVI)) {
8989 unsigned OpToFold = 0;
8990 if ((SFO & 1) && TrueVal == FVI->getOperand(0)) {
8992 } else if ((SFO & 2) && TrueVal == FVI->getOperand(1)) {
8997 Constant *C = GetSelectFoldableConstant(FVI);
8998 Instruction *NewSel =
8999 SelectInst::Create(SI.getCondition(), C,
9000 FVI->getOperand(2-OpToFold));
9001 InsertNewInstBefore(NewSel, SI);
9002 NewSel->takeName(FVI);
9003 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(FVI))
9004 return BinaryOperator::Create(BO->getOpcode(), TrueVal, NewSel);
9006 assert(0 && "Unknown instruction!!");
9011 if (BinaryOperator::isNot(CondVal)) {
9012 SI.setOperand(0, BinaryOperator::getNotArgument(CondVal));
9013 SI.setOperand(1, FalseVal);
9014 SI.setOperand(2, TrueVal);
9021 /// EnforceKnownAlignment - If the specified pointer points to an object that
9022 /// we control, modify the object's alignment to PrefAlign. This isn't
9023 /// often possible though. If alignment is important, a more reliable approach
9024 /// is to simply align all global variables and allocation instructions to
9025 /// their preferred alignment from the beginning.
9027 static unsigned EnforceKnownAlignment(Value *V,
9028 unsigned Align, unsigned PrefAlign) {
9030 User *U = dyn_cast<User>(V);
9031 if (!U) return Align;
9033 switch (getOpcode(U)) {
9035 case Instruction::BitCast:
9036 return EnforceKnownAlignment(U->getOperand(0), Align, PrefAlign);
9037 case Instruction::GetElementPtr: {
9038 // If all indexes are zero, it is just the alignment of the base pointer.
9039 bool AllZeroOperands = true;
9040 for (User::op_iterator i = U->op_begin() + 1, e = U->op_end(); i != e; ++i)
9041 if (!isa<Constant>(*i) ||
9042 !cast<Constant>(*i)->isNullValue()) {
9043 AllZeroOperands = false;
9047 if (AllZeroOperands) {
9048 // Treat this like a bitcast.
9049 return EnforceKnownAlignment(U->getOperand(0), Align, PrefAlign);
9055 if (GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
9056 // If there is a large requested alignment and we can, bump up the alignment
9058 if (!GV->isDeclaration()) {
9059 GV->setAlignment(PrefAlign);
9062 } else if (AllocationInst *AI = dyn_cast<AllocationInst>(V)) {
9063 // If there is a requested alignment and if this is an alloca, round up. We
9064 // don't do this for malloc, because some systems can't respect the request.
9065 if (isa<AllocaInst>(AI)) {
9066 AI->setAlignment(PrefAlign);
9074 /// GetOrEnforceKnownAlignment - If the specified pointer has an alignment that
9075 /// we can determine, return it, otherwise return 0. If PrefAlign is specified,
9076 /// and it is more than the alignment of the ultimate object, see if we can
9077 /// increase the alignment of the ultimate object, making this check succeed.
9078 unsigned InstCombiner::GetOrEnforceKnownAlignment(Value *V,
9079 unsigned PrefAlign) {
9080 unsigned BitWidth = TD ? TD->getTypeSizeInBits(V->getType()) :
9081 sizeof(PrefAlign) * CHAR_BIT;
9082 APInt Mask = APInt::getAllOnesValue(BitWidth);
9083 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
9084 ComputeMaskedBits(V, Mask, KnownZero, KnownOne);
9085 unsigned TrailZ = KnownZero.countTrailingOnes();
9086 unsigned Align = 1u << std::min(BitWidth - 1, TrailZ);
9088 if (PrefAlign > Align)
9089 Align = EnforceKnownAlignment(V, Align, PrefAlign);
9091 // We don't need to make any adjustment.
9095 Instruction *InstCombiner::SimplifyMemTransfer(MemIntrinsic *MI) {
9096 unsigned DstAlign = GetOrEnforceKnownAlignment(MI->getOperand(1));
9097 unsigned SrcAlign = GetOrEnforceKnownAlignment(MI->getOperand(2));
9098 unsigned MinAlign = std::min(DstAlign, SrcAlign);
9099 unsigned CopyAlign = MI->getAlignment()->getZExtValue();
9101 if (CopyAlign < MinAlign) {
9102 MI->setAlignment(ConstantInt::get(Type::Int32Ty, MinAlign));
9106 // If MemCpyInst length is 1/2/4/8 bytes then replace memcpy with
9108 ConstantInt *MemOpLength = dyn_cast<ConstantInt>(MI->getOperand(3));
9109 if (MemOpLength == 0) return 0;
9111 // Source and destination pointer types are always "i8*" for intrinsic. See
9112 // if the size is something we can handle with a single primitive load/store.
9113 // A single load+store correctly handles overlapping memory in the memmove
9115 unsigned Size = MemOpLength->getZExtValue();
9116 if (Size == 0) return MI; // Delete this mem transfer.
9118 if (Size > 8 || (Size&(Size-1)))
9119 return 0; // If not 1/2/4/8 bytes, exit.
9121 // Use an integer load+store unless we can find something better.
9122 Type *NewPtrTy = PointerType::getUnqual(IntegerType::get(Size<<3));
9124 // Memcpy forces the use of i8* for the source and destination. That means
9125 // that if you're using memcpy to move one double around, you'll get a cast
9126 // from double* to i8*. We'd much rather use a double load+store rather than
9127 // an i64 load+store, here because this improves the odds that the source or
9128 // dest address will be promotable. See if we can find a better type than the
9129 // integer datatype.
9130 if (Value *Op = getBitCastOperand(MI->getOperand(1))) {
9131 const Type *SrcETy = cast<PointerType>(Op->getType())->getElementType();
9132 if (SrcETy->isSized() && TD->getTypeStoreSize(SrcETy) == Size) {
9133 // The SrcETy might be something like {{{double}}} or [1 x double]. Rip
9134 // down through these levels if so.
9135 while (!SrcETy->isSingleValueType()) {
9136 if (const StructType *STy = dyn_cast<StructType>(SrcETy)) {
9137 if (STy->getNumElements() == 1)
9138 SrcETy = STy->getElementType(0);
9141 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(SrcETy)) {
9142 if (ATy->getNumElements() == 1)
9143 SrcETy = ATy->getElementType();
9150 if (SrcETy->isSingleValueType())
9151 NewPtrTy = PointerType::getUnqual(SrcETy);
9156 // If the memcpy/memmove provides better alignment info than we can
9158 SrcAlign = std::max(SrcAlign, CopyAlign);
9159 DstAlign = std::max(DstAlign, CopyAlign);
9161 Value *Src = InsertBitCastBefore(MI->getOperand(2), NewPtrTy, *MI);
9162 Value *Dest = InsertBitCastBefore(MI->getOperand(1), NewPtrTy, *MI);
9163 Instruction *L = new LoadInst(Src, "tmp", false, SrcAlign);
9164 InsertNewInstBefore(L, *MI);
9165 InsertNewInstBefore(new StoreInst(L, Dest, false, DstAlign), *MI);
9167 // Set the size of the copy to 0, it will be deleted on the next iteration.
9168 MI->setOperand(3, Constant::getNullValue(MemOpLength->getType()));
9172 Instruction *InstCombiner::SimplifyMemSet(MemSetInst *MI) {
9173 unsigned Alignment = GetOrEnforceKnownAlignment(MI->getDest());
9174 if (MI->getAlignment()->getZExtValue() < Alignment) {
9175 MI->setAlignment(ConstantInt::get(Type::Int32Ty, Alignment));
9179 // Extract the length and alignment and fill if they are constant.
9180 ConstantInt *LenC = dyn_cast<ConstantInt>(MI->getLength());
9181 ConstantInt *FillC = dyn_cast<ConstantInt>(MI->getValue());
9182 if (!LenC || !FillC || FillC->getType() != Type::Int8Ty)
9184 uint64_t Len = LenC->getZExtValue();
9185 Alignment = MI->getAlignment()->getZExtValue();
9187 // If the length is zero, this is a no-op
9188 if (Len == 0) return MI; // memset(d,c,0,a) -> noop
9190 // memset(s,c,n) -> store s, c (for n=1,2,4,8)
9191 if (Len <= 8 && isPowerOf2_32((uint32_t)Len)) {
9192 const Type *ITy = IntegerType::get(Len*8); // n=1 -> i8.
9194 Value *Dest = MI->getDest();
9195 Dest = InsertBitCastBefore(Dest, PointerType::getUnqual(ITy), *MI);
9197 // Alignment 0 is identity for alignment 1 for memset, but not store.
9198 if (Alignment == 0) Alignment = 1;
9200 // Extract the fill value and store.
9201 uint64_t Fill = FillC->getZExtValue()*0x0101010101010101ULL;
9202 InsertNewInstBefore(new StoreInst(ConstantInt::get(ITy, Fill), Dest, false,
9205 // Set the size of the copy to 0, it will be deleted on the next iteration.
9206 MI->setLength(Constant::getNullValue(LenC->getType()));
9214 /// visitCallInst - CallInst simplification. This mostly only handles folding
9215 /// of intrinsic instructions. For normal calls, it allows visitCallSite to do
9216 /// the heavy lifting.
9218 Instruction *InstCombiner::visitCallInst(CallInst &CI) {
9219 IntrinsicInst *II = dyn_cast<IntrinsicInst>(&CI);
9220 if (!II) return visitCallSite(&CI);
9222 // Intrinsics cannot occur in an invoke, so handle them here instead of in
9224 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(II)) {
9225 bool Changed = false;
9227 // memmove/cpy/set of zero bytes is a noop.
9228 if (Constant *NumBytes = dyn_cast<Constant>(MI->getLength())) {
9229 if (NumBytes->isNullValue()) return EraseInstFromFunction(CI);
9231 if (ConstantInt *CI = dyn_cast<ConstantInt>(NumBytes))
9232 if (CI->getZExtValue() == 1) {
9233 // Replace the instruction with just byte operations. We would
9234 // transform other cases to loads/stores, but we don't know if
9235 // alignment is sufficient.
9239 // If we have a memmove and the source operation is a constant global,
9240 // then the source and dest pointers can't alias, so we can change this
9241 // into a call to memcpy.
9242 if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(MI)) {
9243 if (GlobalVariable *GVSrc = dyn_cast<GlobalVariable>(MMI->getSource()))
9244 if (GVSrc->isConstant()) {
9245 Module *M = CI.getParent()->getParent()->getParent();
9246 Intrinsic::ID MemCpyID = Intrinsic::memcpy;
9248 Tys[0] = CI.getOperand(3)->getType();
9250 Intrinsic::getDeclaration(M, MemCpyID, Tys, 1));
9254 // memmove(x,x,size) -> noop.
9255 if (MMI->getSource() == MMI->getDest())
9256 return EraseInstFromFunction(CI);
9259 // If we can determine a pointer alignment that is bigger than currently
9260 // set, update the alignment.
9261 if (isa<MemCpyInst>(MI) || isa<MemMoveInst>(MI)) {
9262 if (Instruction *I = SimplifyMemTransfer(MI))
9264 } else if (MemSetInst *MSI = dyn_cast<MemSetInst>(MI)) {
9265 if (Instruction *I = SimplifyMemSet(MSI))
9269 if (Changed) return II;
9272 switch (II->getIntrinsicID()) {
9274 case Intrinsic::bswap:
9275 // bswap(bswap(x)) -> x
9276 if (IntrinsicInst *Operand = dyn_cast<IntrinsicInst>(II->getOperand(1)))
9277 if (Operand->getIntrinsicID() == Intrinsic::bswap)
9278 return ReplaceInstUsesWith(CI, Operand->getOperand(1));
9280 case Intrinsic::ppc_altivec_lvx:
9281 case Intrinsic::ppc_altivec_lvxl:
9282 case Intrinsic::x86_sse_loadu_ps:
9283 case Intrinsic::x86_sse2_loadu_pd:
9284 case Intrinsic::x86_sse2_loadu_dq:
9285 // Turn PPC lvx -> load if the pointer is known aligned.
9286 // Turn X86 loadups -> load if the pointer is known aligned.
9287 if (GetOrEnforceKnownAlignment(II->getOperand(1), 16) >= 16) {
9288 Value *Ptr = InsertBitCastBefore(II->getOperand(1),
9289 PointerType::getUnqual(II->getType()),
9291 return new LoadInst(Ptr);
9294 case Intrinsic::ppc_altivec_stvx:
9295 case Intrinsic::ppc_altivec_stvxl:
9296 // Turn stvx -> store if the pointer is known aligned.
9297 if (GetOrEnforceKnownAlignment(II->getOperand(2), 16) >= 16) {
9298 const Type *OpPtrTy =
9299 PointerType::getUnqual(II->getOperand(1)->getType());
9300 Value *Ptr = InsertBitCastBefore(II->getOperand(2), OpPtrTy, CI);
9301 return new StoreInst(II->getOperand(1), Ptr);
9304 case Intrinsic::x86_sse_storeu_ps:
9305 case Intrinsic::x86_sse2_storeu_pd:
9306 case Intrinsic::x86_sse2_storeu_dq:
9307 // Turn X86 storeu -> store if the pointer is known aligned.
9308 if (GetOrEnforceKnownAlignment(II->getOperand(1), 16) >= 16) {
9309 const Type *OpPtrTy =
9310 PointerType::getUnqual(II->getOperand(2)->getType());
9311 Value *Ptr = InsertBitCastBefore(II->getOperand(1), OpPtrTy, CI);
9312 return new StoreInst(II->getOperand(2), Ptr);
9316 case Intrinsic::x86_sse_cvttss2si: {
9317 // These intrinsics only demands the 0th element of its input vector. If
9318 // we can simplify the input based on that, do so now.
9320 if (Value *V = SimplifyDemandedVectorElts(II->getOperand(1), 1,
9322 II->setOperand(1, V);
9328 case Intrinsic::ppc_altivec_vperm:
9329 // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant.
9330 if (ConstantVector *Mask = dyn_cast<ConstantVector>(II->getOperand(3))) {
9331 assert(Mask->getNumOperands() == 16 && "Bad type for intrinsic!");
9333 // Check that all of the elements are integer constants or undefs.
9334 bool AllEltsOk = true;
9335 for (unsigned i = 0; i != 16; ++i) {
9336 if (!isa<ConstantInt>(Mask->getOperand(i)) &&
9337 !isa<UndefValue>(Mask->getOperand(i))) {
9344 // Cast the input vectors to byte vectors.
9345 Value *Op0 =InsertBitCastBefore(II->getOperand(1),Mask->getType(),CI);
9346 Value *Op1 =InsertBitCastBefore(II->getOperand(2),Mask->getType(),CI);
9347 Value *Result = UndefValue::get(Op0->getType());
9349 // Only extract each element once.
9350 Value *ExtractedElts[32];
9351 memset(ExtractedElts, 0, sizeof(ExtractedElts));
9353 for (unsigned i = 0; i != 16; ++i) {
9354 if (isa<UndefValue>(Mask->getOperand(i)))
9356 unsigned Idx=cast<ConstantInt>(Mask->getOperand(i))->getZExtValue();
9357 Idx &= 31; // Match the hardware behavior.
9359 if (ExtractedElts[Idx] == 0) {
9361 new ExtractElementInst(Idx < 16 ? Op0 : Op1, Idx&15, "tmp");
9362 InsertNewInstBefore(Elt, CI);
9363 ExtractedElts[Idx] = Elt;
9366 // Insert this value into the result vector.
9367 Result = InsertElementInst::Create(Result, ExtractedElts[Idx],
9369 InsertNewInstBefore(cast<Instruction>(Result), CI);
9371 return CastInst::Create(Instruction::BitCast, Result, CI.getType());
9376 case Intrinsic::stackrestore: {
9377 // If the save is right next to the restore, remove the restore. This can
9378 // happen when variable allocas are DCE'd.
9379 if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getOperand(1))) {
9380 if (SS->getIntrinsicID() == Intrinsic::stacksave) {
9381 BasicBlock::iterator BI = SS;
9383 return EraseInstFromFunction(CI);
9387 // Scan down this block to see if there is another stack restore in the
9388 // same block without an intervening call/alloca.
9389 BasicBlock::iterator BI = II;
9390 TerminatorInst *TI = II->getParent()->getTerminator();
9391 bool CannotRemove = false;
9392 for (++BI; &*BI != TI; ++BI) {
9393 if (isa<AllocaInst>(BI)) {
9394 CannotRemove = true;
9397 if (CallInst *BCI = dyn_cast<CallInst>(BI)) {
9398 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(BCI)) {
9399 // If there is a stackrestore below this one, remove this one.
9400 if (II->getIntrinsicID() == Intrinsic::stackrestore)
9401 return EraseInstFromFunction(CI);
9402 // Otherwise, ignore the intrinsic.
9404 // If we found a non-intrinsic call, we can't remove the stack
9406 CannotRemove = true;
9412 // If the stack restore is in a return/unwind block and if there are no
9413 // allocas or calls between the restore and the return, nuke the restore.
9414 if (!CannotRemove && (isa<ReturnInst>(TI) || isa<UnwindInst>(TI)))
9415 return EraseInstFromFunction(CI);
9420 return visitCallSite(II);
9423 // InvokeInst simplification
9425 Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) {
9426 return visitCallSite(&II);
9429 /// isSafeToEliminateVarargsCast - If this cast does not affect the value
9430 /// passed through the varargs area, we can eliminate the use of the cast.
9431 static bool isSafeToEliminateVarargsCast(const CallSite CS,
9432 const CastInst * const CI,
9433 const TargetData * const TD,
9435 if (!CI->isLosslessCast())
9438 // The size of ByVal arguments is derived from the type, so we
9439 // can't change to a type with a different size. If the size were
9440 // passed explicitly we could avoid this check.
9441 if (!CS.paramHasAttr(ix, Attribute::ByVal))
9445 cast<PointerType>(CI->getOperand(0)->getType())->getElementType();
9446 const Type* DstTy = cast<PointerType>(CI->getType())->getElementType();
9447 if (!SrcTy->isSized() || !DstTy->isSized())
9449 if (TD->getABITypeSize(SrcTy) != TD->getABITypeSize(DstTy))
9454 // visitCallSite - Improvements for call and invoke instructions.
9456 Instruction *InstCombiner::visitCallSite(CallSite CS) {
9457 bool Changed = false;
9459 // If the callee is a constexpr cast of a function, attempt to move the cast
9460 // to the arguments of the call/invoke.
9461 if (transformConstExprCastCall(CS)) return 0;
9463 Value *Callee = CS.getCalledValue();
9465 if (Function *CalleeF = dyn_cast<Function>(Callee))
9466 if (CalleeF->getCallingConv() != CS.getCallingConv()) {
9467 Instruction *OldCall = CS.getInstruction();
9468 // If the call and callee calling conventions don't match, this call must
9469 // be unreachable, as the call is undefined.
9470 new StoreInst(ConstantInt::getTrue(),
9471 UndefValue::get(PointerType::getUnqual(Type::Int1Ty)),
9473 if (!OldCall->use_empty())
9474 OldCall->replaceAllUsesWith(UndefValue::get(OldCall->getType()));
9475 if (isa<CallInst>(OldCall)) // Not worth removing an invoke here.
9476 return EraseInstFromFunction(*OldCall);
9480 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) {
9481 // This instruction is not reachable, just remove it. We insert a store to
9482 // undef so that we know that this code is not reachable, despite the fact
9483 // that we can't modify the CFG here.
9484 new StoreInst(ConstantInt::getTrue(),
9485 UndefValue::get(PointerType::getUnqual(Type::Int1Ty)),
9486 CS.getInstruction());
9488 if (!CS.getInstruction()->use_empty())
9489 CS.getInstruction()->
9490 replaceAllUsesWith(UndefValue::get(CS.getInstruction()->getType()));
9492 if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) {
9493 // Don't break the CFG, insert a dummy cond branch.
9494 BranchInst::Create(II->getNormalDest(), II->getUnwindDest(),
9495 ConstantInt::getTrue(), II);
9497 return EraseInstFromFunction(*CS.getInstruction());
9500 if (BitCastInst *BC = dyn_cast<BitCastInst>(Callee))
9501 if (IntrinsicInst *In = dyn_cast<IntrinsicInst>(BC->getOperand(0)))
9502 if (In->getIntrinsicID() == Intrinsic::init_trampoline)
9503 return transformCallThroughTrampoline(CS);
9505 const PointerType *PTy = cast<PointerType>(Callee->getType());
9506 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
9507 if (FTy->isVarArg()) {
9508 int ix = FTy->getNumParams() + (isa<InvokeInst>(Callee) ? 3 : 1);
9509 // See if we can optimize any arguments passed through the varargs area of
9511 for (CallSite::arg_iterator I = CS.arg_begin()+FTy->getNumParams(),
9512 E = CS.arg_end(); I != E; ++I, ++ix) {
9513 CastInst *CI = dyn_cast<CastInst>(*I);
9514 if (CI && isSafeToEliminateVarargsCast(CS, CI, TD, ix)) {
9515 *I = CI->getOperand(0);
9521 if (isa<InlineAsm>(Callee) && !CS.doesNotThrow()) {
9522 // Inline asm calls cannot throw - mark them 'nounwind'.
9523 CS.setDoesNotThrow();
9527 return Changed ? CS.getInstruction() : 0;
9530 // transformConstExprCastCall - If the callee is a constexpr cast of a function,
9531 // attempt to move the cast to the arguments of the call/invoke.
9533 bool InstCombiner::transformConstExprCastCall(CallSite CS) {
9534 if (!isa<ConstantExpr>(CS.getCalledValue())) return false;
9535 ConstantExpr *CE = cast<ConstantExpr>(CS.getCalledValue());
9536 if (CE->getOpcode() != Instruction::BitCast ||
9537 !isa<Function>(CE->getOperand(0)))
9539 Function *Callee = cast<Function>(CE->getOperand(0));
9540 Instruction *Caller = CS.getInstruction();
9541 const AttrListPtr &CallerPAL = CS.getAttributes();
9543 // Okay, this is a cast from a function to a different type. Unless doing so
9544 // would cause a type conversion of one of our arguments, change this call to
9545 // be a direct call with arguments casted to the appropriate types.
9547 const FunctionType *FT = Callee->getFunctionType();
9548 const Type *OldRetTy = Caller->getType();
9549 const Type *NewRetTy = FT->getReturnType();
9551 if (isa<StructType>(NewRetTy))
9552 return false; // TODO: Handle multiple return values.
9554 // Check to see if we are changing the return type...
9555 if (OldRetTy != NewRetTy) {
9556 if (Callee->isDeclaration() &&
9557 // Conversion is ok if changing from one pointer type to another or from
9558 // a pointer to an integer of the same size.
9559 !((isa<PointerType>(OldRetTy) || OldRetTy == TD->getIntPtrType()) &&
9560 (isa<PointerType>(NewRetTy) || NewRetTy == TD->getIntPtrType())))
9561 return false; // Cannot transform this return value.
9563 if (!Caller->use_empty() &&
9564 // void -> non-void is handled specially
9565 NewRetTy != Type::VoidTy && !CastInst::isCastable(NewRetTy, OldRetTy))
9566 return false; // Cannot transform this return value.
9568 if (!CallerPAL.isEmpty() && !Caller->use_empty()) {
9569 Attributes RAttrs = CallerPAL.getRetAttributes();
9570 if (RAttrs & Attribute::typeIncompatible(NewRetTy))
9571 return false; // Attribute not compatible with transformed value.
9574 // If the callsite is an invoke instruction, and the return value is used by
9575 // a PHI node in a successor, we cannot change the return type of the call
9576 // because there is no place to put the cast instruction (without breaking
9577 // the critical edge). Bail out in this case.
9578 if (!Caller->use_empty())
9579 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller))
9580 for (Value::use_iterator UI = II->use_begin(), E = II->use_end();
9582 if (PHINode *PN = dyn_cast<PHINode>(*UI))
9583 if (PN->getParent() == II->getNormalDest() ||
9584 PN->getParent() == II->getUnwindDest())
9588 unsigned NumActualArgs = unsigned(CS.arg_end()-CS.arg_begin());
9589 unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs);
9591 CallSite::arg_iterator AI = CS.arg_begin();
9592 for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) {
9593 const Type *ParamTy = FT->getParamType(i);
9594 const Type *ActTy = (*AI)->getType();
9596 if (!CastInst::isCastable(ActTy, ParamTy))
9597 return false; // Cannot transform this parameter value.
9599 if (CallerPAL.getParamAttributes(i + 1)
9600 & Attribute::typeIncompatible(ParamTy))
9601 return false; // Attribute not compatible with transformed value.
9603 // Converting from one pointer type to another or between a pointer and an
9604 // integer of the same size is safe even if we do not have a body.
9605 bool isConvertible = ActTy == ParamTy ||
9606 ((isa<PointerType>(ParamTy) || ParamTy == TD->getIntPtrType()) &&
9607 (isa<PointerType>(ActTy) || ActTy == TD->getIntPtrType()));
9608 if (Callee->isDeclaration() && !isConvertible) return false;
9611 if (FT->getNumParams() < NumActualArgs && !FT->isVarArg() &&
9612 Callee->isDeclaration())
9613 return false; // Do not delete arguments unless we have a function body.
9615 if (FT->getNumParams() < NumActualArgs && FT->isVarArg() &&
9616 !CallerPAL.isEmpty())
9617 // In this case we have more arguments than the new function type, but we
9618 // won't be dropping them. Check that these extra arguments have attributes
9619 // that are compatible with being a vararg call argument.
9620 for (unsigned i = CallerPAL.getNumSlots(); i; --i) {
9621 if (CallerPAL.getSlot(i - 1).Index <= FT->getNumParams())
9623 Attributes PAttrs = CallerPAL.getSlot(i - 1).Attrs;
9624 if (PAttrs & Attribute::VarArgsIncompatible)
9628 // Okay, we decided that this is a safe thing to do: go ahead and start
9629 // inserting cast instructions as necessary...
9630 std::vector<Value*> Args;
9631 Args.reserve(NumActualArgs);
9632 SmallVector<AttributeWithIndex, 8> attrVec;
9633 attrVec.reserve(NumCommonArgs);
9635 // Get any return attributes.
9636 Attributes RAttrs = CallerPAL.getRetAttributes();
9638 // If the return value is not being used, the type may not be compatible
9639 // with the existing attributes. Wipe out any problematic attributes.
9640 RAttrs &= ~Attribute::typeIncompatible(NewRetTy);
9642 // Add the new return attributes.
9644 attrVec.push_back(AttributeWithIndex::get(0, RAttrs));
9646 AI = CS.arg_begin();
9647 for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) {
9648 const Type *ParamTy = FT->getParamType(i);
9649 if ((*AI)->getType() == ParamTy) {
9650 Args.push_back(*AI);
9652 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI,
9653 false, ParamTy, false);
9654 CastInst *NewCast = CastInst::Create(opcode, *AI, ParamTy, "tmp");
9655 Args.push_back(InsertNewInstBefore(NewCast, *Caller));
9658 // Add any parameter attributes.
9659 if (Attributes PAttrs = CallerPAL.getParamAttributes(i + 1))
9660 attrVec.push_back(AttributeWithIndex::get(i + 1, PAttrs));
9663 // If the function takes more arguments than the call was taking, add them
9665 for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i)
9666 Args.push_back(Constant::getNullValue(FT->getParamType(i)));
9668 // If we are removing arguments to the function, emit an obnoxious warning...
9669 if (FT->getNumParams() < NumActualArgs) {
9670 if (!FT->isVarArg()) {
9671 cerr << "WARNING: While resolving call to function '"
9672 << Callee->getName() << "' arguments were dropped!\n";
9674 // Add all of the arguments in their promoted form to the arg list...
9675 for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) {
9676 const Type *PTy = getPromotedType((*AI)->getType());
9677 if (PTy != (*AI)->getType()) {
9678 // Must promote to pass through va_arg area!
9679 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI, false,
9681 Instruction *Cast = CastInst::Create(opcode, *AI, PTy, "tmp");
9682 InsertNewInstBefore(Cast, *Caller);
9683 Args.push_back(Cast);
9685 Args.push_back(*AI);
9688 // Add any parameter attributes.
9689 if (Attributes PAttrs = CallerPAL.getParamAttributes(i + 1))
9690 attrVec.push_back(AttributeWithIndex::get(i + 1, PAttrs));
9695 if (Attributes FnAttrs = CallerPAL.getFnAttributes())
9696 attrVec.push_back(AttributeWithIndex::get(~0, FnAttrs));
9698 if (NewRetTy == Type::VoidTy)
9699 Caller->setName(""); // Void type should not have a name.
9701 const AttrListPtr &NewCallerPAL = AttrListPtr::get(attrVec.begin(),attrVec.end());
9704 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
9705 NC = InvokeInst::Create(Callee, II->getNormalDest(), II->getUnwindDest(),
9706 Args.begin(), Args.end(),
9707 Caller->getName(), Caller);
9708 cast<InvokeInst>(NC)->setCallingConv(II->getCallingConv());
9709 cast<InvokeInst>(NC)->setAttributes(NewCallerPAL);
9711 NC = CallInst::Create(Callee, Args.begin(), Args.end(),
9712 Caller->getName(), Caller);
9713 CallInst *CI = cast<CallInst>(Caller);
9714 if (CI->isTailCall())
9715 cast<CallInst>(NC)->setTailCall();
9716 cast<CallInst>(NC)->setCallingConv(CI->getCallingConv());
9717 cast<CallInst>(NC)->setAttributes(NewCallerPAL);
9720 // Insert a cast of the return type as necessary.
9722 if (OldRetTy != NV->getType() && !Caller->use_empty()) {
9723 if (NV->getType() != Type::VoidTy) {
9724 Instruction::CastOps opcode = CastInst::getCastOpcode(NC, false,
9726 NV = NC = CastInst::Create(opcode, NC, OldRetTy, "tmp");
9728 // If this is an invoke instruction, we should insert it after the first
9729 // non-phi, instruction in the normal successor block.
9730 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
9731 BasicBlock::iterator I = II->getNormalDest()->getFirstNonPHI();
9732 InsertNewInstBefore(NC, *I);
9734 // Otherwise, it's a call, just insert cast right after the call instr
9735 InsertNewInstBefore(NC, *Caller);
9737 AddUsersToWorkList(*Caller);
9739 NV = UndefValue::get(Caller->getType());
9743 if (Caller->getType() != Type::VoidTy && !Caller->use_empty())
9744 Caller->replaceAllUsesWith(NV);
9745 Caller->eraseFromParent();
9746 RemoveFromWorkList(Caller);
9750 // transformCallThroughTrampoline - Turn a call to a function created by the
9751 // init_trampoline intrinsic into a direct call to the underlying function.
9753 Instruction *InstCombiner::transformCallThroughTrampoline(CallSite CS) {
9754 Value *Callee = CS.getCalledValue();
9755 const PointerType *PTy = cast<PointerType>(Callee->getType());
9756 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
9757 const AttrListPtr &Attrs = CS.getAttributes();
9759 // If the call already has the 'nest' attribute somewhere then give up -
9760 // otherwise 'nest' would occur twice after splicing in the chain.
9761 if (Attrs.hasAttrSomewhere(Attribute::Nest))
9764 IntrinsicInst *Tramp =
9765 cast<IntrinsicInst>(cast<BitCastInst>(Callee)->getOperand(0));
9767 Function *NestF = cast<Function>(Tramp->getOperand(2)->stripPointerCasts());
9768 const PointerType *NestFPTy = cast<PointerType>(NestF->getType());
9769 const FunctionType *NestFTy = cast<FunctionType>(NestFPTy->getElementType());
9771 const AttrListPtr &NestAttrs = NestF->getAttributes();
9772 if (!NestAttrs.isEmpty()) {
9773 unsigned NestIdx = 1;
9774 const Type *NestTy = 0;
9775 Attributes NestAttr = Attribute::None;
9777 // Look for a parameter marked with the 'nest' attribute.
9778 for (FunctionType::param_iterator I = NestFTy->param_begin(),
9779 E = NestFTy->param_end(); I != E; ++NestIdx, ++I)
9780 if (NestAttrs.paramHasAttr(NestIdx, Attribute::Nest)) {
9781 // Record the parameter type and any other attributes.
9783 NestAttr = NestAttrs.getParamAttributes(NestIdx);
9788 Instruction *Caller = CS.getInstruction();
9789 std::vector<Value*> NewArgs;
9790 NewArgs.reserve(unsigned(CS.arg_end()-CS.arg_begin())+1);
9792 SmallVector<AttributeWithIndex, 8> NewAttrs;
9793 NewAttrs.reserve(Attrs.getNumSlots() + 1);
9795 // Insert the nest argument into the call argument list, which may
9796 // mean appending it. Likewise for attributes.
9798 // Add any result attributes.
9799 if (Attributes Attr = Attrs.getRetAttributes())
9800 NewAttrs.push_back(AttributeWithIndex::get(0, Attr));
9804 CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
9806 if (Idx == NestIdx) {
9807 // Add the chain argument and attributes.
9808 Value *NestVal = Tramp->getOperand(3);
9809 if (NestVal->getType() != NestTy)
9810 NestVal = new BitCastInst(NestVal, NestTy, "nest", Caller);
9811 NewArgs.push_back(NestVal);
9812 NewAttrs.push_back(AttributeWithIndex::get(NestIdx, NestAttr));
9818 // Add the original argument and attributes.
9819 NewArgs.push_back(*I);
9820 if (Attributes Attr = Attrs.getParamAttributes(Idx))
9822 (AttributeWithIndex::get(Idx + (Idx >= NestIdx), Attr));
9828 // Add any function attributes.
9829 if (Attributes Attr = Attrs.getFnAttributes())
9830 NewAttrs.push_back(AttributeWithIndex::get(~0, Attr));
9832 // The trampoline may have been bitcast to a bogus type (FTy).
9833 // Handle this by synthesizing a new function type, equal to FTy
9834 // with the chain parameter inserted.
9836 std::vector<const Type*> NewTypes;
9837 NewTypes.reserve(FTy->getNumParams()+1);
9839 // Insert the chain's type into the list of parameter types, which may
9840 // mean appending it.
9843 FunctionType::param_iterator I = FTy->param_begin(),
9844 E = FTy->param_end();
9848 // Add the chain's type.
9849 NewTypes.push_back(NestTy);
9854 // Add the original type.
9855 NewTypes.push_back(*I);
9861 // Replace the trampoline call with a direct call. Let the generic
9862 // code sort out any function type mismatches.
9863 FunctionType *NewFTy =
9864 FunctionType::get(FTy->getReturnType(), NewTypes, FTy->isVarArg());
9865 Constant *NewCallee = NestF->getType() == PointerType::getUnqual(NewFTy) ?
9866 NestF : ConstantExpr::getBitCast(NestF, PointerType::getUnqual(NewFTy));
9867 const AttrListPtr &NewPAL = AttrListPtr::get(NewAttrs.begin(),NewAttrs.end());
9869 Instruction *NewCaller;
9870 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
9871 NewCaller = InvokeInst::Create(NewCallee,
9872 II->getNormalDest(), II->getUnwindDest(),
9873 NewArgs.begin(), NewArgs.end(),
9874 Caller->getName(), Caller);
9875 cast<InvokeInst>(NewCaller)->setCallingConv(II->getCallingConv());
9876 cast<InvokeInst>(NewCaller)->setAttributes(NewPAL);
9878 NewCaller = CallInst::Create(NewCallee, NewArgs.begin(), NewArgs.end(),
9879 Caller->getName(), Caller);
9880 if (cast<CallInst>(Caller)->isTailCall())
9881 cast<CallInst>(NewCaller)->setTailCall();
9882 cast<CallInst>(NewCaller)->
9883 setCallingConv(cast<CallInst>(Caller)->getCallingConv());
9884 cast<CallInst>(NewCaller)->setAttributes(NewPAL);
9886 if (Caller->getType() != Type::VoidTy && !Caller->use_empty())
9887 Caller->replaceAllUsesWith(NewCaller);
9888 Caller->eraseFromParent();
9889 RemoveFromWorkList(Caller);
9894 // Replace the trampoline call with a direct call. Since there is no 'nest'
9895 // parameter, there is no need to adjust the argument list. Let the generic
9896 // code sort out any function type mismatches.
9897 Constant *NewCallee =
9898 NestF->getType() == PTy ? NestF : ConstantExpr::getBitCast(NestF, PTy);
9899 CS.setCalledFunction(NewCallee);
9900 return CS.getInstruction();
9903 /// FoldPHIArgBinOpIntoPHI - If we have something like phi [add (a,b), add(c,d)]
9904 /// and if a/b/c/d and the add's all have a single use, turn this into two phi's
9905 /// and a single binop.
9906 Instruction *InstCombiner::FoldPHIArgBinOpIntoPHI(PHINode &PN) {
9907 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
9908 assert(isa<BinaryOperator>(FirstInst) || isa<GetElementPtrInst>(FirstInst) ||
9909 isa<CmpInst>(FirstInst));
9910 unsigned Opc = FirstInst->getOpcode();
9911 Value *LHSVal = FirstInst->getOperand(0);
9912 Value *RHSVal = FirstInst->getOperand(1);
9914 const Type *LHSType = LHSVal->getType();
9915 const Type *RHSType = RHSVal->getType();
9917 // Scan to see if all operands are the same opcode, all have one use, and all
9918 // kill their operands (i.e. the operands have one use).
9919 for (unsigned i = 0; i != PN.getNumIncomingValues(); ++i) {
9920 Instruction *I = dyn_cast<Instruction>(PN.getIncomingValue(i));
9921 if (!I || I->getOpcode() != Opc || !I->hasOneUse() ||
9922 // Verify type of the LHS matches so we don't fold cmp's of different
9923 // types or GEP's with different index types.
9924 I->getOperand(0)->getType() != LHSType ||
9925 I->getOperand(1)->getType() != RHSType)
9928 // If they are CmpInst instructions, check their predicates
9929 if (Opc == Instruction::ICmp || Opc == Instruction::FCmp)
9930 if (cast<CmpInst>(I)->getPredicate() !=
9931 cast<CmpInst>(FirstInst)->getPredicate())
9934 // Keep track of which operand needs a phi node.
9935 if (I->getOperand(0) != LHSVal) LHSVal = 0;
9936 if (I->getOperand(1) != RHSVal) RHSVal = 0;
9939 // Otherwise, this is safe to transform, determine if it is profitable.
9941 // If this is a GEP, and if the index (not the pointer) needs a PHI, bail out.
9942 // Indexes are often folded into load/store instructions, so we don't want to
9943 // hide them behind a phi.
9944 if (isa<GetElementPtrInst>(FirstInst) && RHSVal == 0)
9947 Value *InLHS = FirstInst->getOperand(0);
9948 Value *InRHS = FirstInst->getOperand(1);
9949 PHINode *NewLHS = 0, *NewRHS = 0;
9951 NewLHS = PHINode::Create(LHSType,
9952 FirstInst->getOperand(0)->getName() + ".pn");
9953 NewLHS->reserveOperandSpace(PN.getNumOperands()/2);
9954 NewLHS->addIncoming(InLHS, PN.getIncomingBlock(0));
9955 InsertNewInstBefore(NewLHS, PN);
9960 NewRHS = PHINode::Create(RHSType,
9961 FirstInst->getOperand(1)->getName() + ".pn");
9962 NewRHS->reserveOperandSpace(PN.getNumOperands()/2);
9963 NewRHS->addIncoming(InRHS, PN.getIncomingBlock(0));
9964 InsertNewInstBefore(NewRHS, PN);
9968 // Add all operands to the new PHIs.
9969 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
9971 Value *NewInLHS =cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
9972 NewLHS->addIncoming(NewInLHS, PN.getIncomingBlock(i));
9975 Value *NewInRHS =cast<Instruction>(PN.getIncomingValue(i))->getOperand(1);
9976 NewRHS->addIncoming(NewInRHS, PN.getIncomingBlock(i));
9980 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
9981 return BinaryOperator::Create(BinOp->getOpcode(), LHSVal, RHSVal);
9982 else if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst))
9983 return CmpInst::Create(CIOp->getOpcode(), CIOp->getPredicate(), LHSVal,
9986 assert(isa<GetElementPtrInst>(FirstInst));
9987 return GetElementPtrInst::Create(LHSVal, RHSVal);
9991 /// isSafeToSinkLoad - Return true if we know that it is safe sink the load out
9992 /// of the block that defines it. This means that it must be obvious the value
9993 /// of the load is not changed from the point of the load to the end of the
9996 /// Finally, it is safe, but not profitable, to sink a load targetting a
9997 /// non-address-taken alloca. Doing so will cause us to not promote the alloca
9999 static bool isSafeToSinkLoad(LoadInst *L) {
10000 BasicBlock::iterator BBI = L, E = L->getParent()->end();
10002 for (++BBI; BBI != E; ++BBI)
10003 if (BBI->mayWriteToMemory())
10006 // Check for non-address taken alloca. If not address-taken already, it isn't
10007 // profitable to do this xform.
10008 if (AllocaInst *AI = dyn_cast<AllocaInst>(L->getOperand(0))) {
10009 bool isAddressTaken = false;
10010 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
10012 if (isa<LoadInst>(UI)) continue;
10013 if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
10014 // If storing TO the alloca, then the address isn't taken.
10015 if (SI->getOperand(1) == AI) continue;
10017 isAddressTaken = true;
10021 if (!isAddressTaken)
10029 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
10030 // operator and they all are only used by the PHI, PHI together their
10031 // inputs, and do the operation once, to the result of the PHI.
10032 Instruction *InstCombiner::FoldPHIArgOpIntoPHI(PHINode &PN) {
10033 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
10035 // Scan the instruction, looking for input operations that can be folded away.
10036 // If all input operands to the phi are the same instruction (e.g. a cast from
10037 // the same type or "+42") we can pull the operation through the PHI, reducing
10038 // code size and simplifying code.
10039 Constant *ConstantOp = 0;
10040 const Type *CastSrcTy = 0;
10041 bool isVolatile = false;
10042 if (isa<CastInst>(FirstInst)) {
10043 CastSrcTy = FirstInst->getOperand(0)->getType();
10044 } else if (isa<BinaryOperator>(FirstInst) || isa<CmpInst>(FirstInst)) {
10045 // Can fold binop, compare or shift here if the RHS is a constant,
10046 // otherwise call FoldPHIArgBinOpIntoPHI.
10047 ConstantOp = dyn_cast<Constant>(FirstInst->getOperand(1));
10048 if (ConstantOp == 0)
10049 return FoldPHIArgBinOpIntoPHI(PN);
10050 } else if (LoadInst *LI = dyn_cast<LoadInst>(FirstInst)) {
10051 isVolatile = LI->isVolatile();
10052 // We can't sink the load if the loaded value could be modified between the
10053 // load and the PHI.
10054 if (LI->getParent() != PN.getIncomingBlock(0) ||
10055 !isSafeToSinkLoad(LI))
10058 // If the PHI is of volatile loads and the load block has multiple
10059 // successors, sinking it would remove a load of the volatile value from
10060 // the path through the other successor.
10062 LI->getParent()->getTerminator()->getNumSuccessors() != 1)
10065 } else if (isa<GetElementPtrInst>(FirstInst)) {
10066 if (FirstInst->getNumOperands() == 2)
10067 return FoldPHIArgBinOpIntoPHI(PN);
10068 // Can't handle general GEPs yet.
10071 return 0; // Cannot fold this operation.
10074 // Check to see if all arguments are the same operation.
10075 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
10076 if (!isa<Instruction>(PN.getIncomingValue(i))) return 0;
10077 Instruction *I = cast<Instruction>(PN.getIncomingValue(i));
10078 if (!I->hasOneUse() || !I->isSameOperationAs(FirstInst))
10081 if (I->getOperand(0)->getType() != CastSrcTy)
10082 return 0; // Cast operation must match.
10083 } else if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
10084 // We can't sink the load if the loaded value could be modified between
10085 // the load and the PHI.
10086 if (LI->isVolatile() != isVolatile ||
10087 LI->getParent() != PN.getIncomingBlock(i) ||
10088 !isSafeToSinkLoad(LI))
10091 // If the PHI is of volatile loads and the load block has multiple
10092 // successors, sinking it would remove a load of the volatile value from
10093 // the path through the other successor.
10095 LI->getParent()->getTerminator()->getNumSuccessors() != 1)
10099 } else if (I->getOperand(1) != ConstantOp) {
10104 // Okay, they are all the same operation. Create a new PHI node of the
10105 // correct type, and PHI together all of the LHS's of the instructions.
10106 PHINode *NewPN = PHINode::Create(FirstInst->getOperand(0)->getType(),
10107 PN.getName()+".in");
10108 NewPN->reserveOperandSpace(PN.getNumOperands()/2);
10110 Value *InVal = FirstInst->getOperand(0);
10111 NewPN->addIncoming(InVal, PN.getIncomingBlock(0));
10113 // Add all operands to the new PHI.
10114 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
10115 Value *NewInVal = cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
10116 if (NewInVal != InVal)
10118 NewPN->addIncoming(NewInVal, PN.getIncomingBlock(i));
10123 // The new PHI unions all of the same values together. This is really
10124 // common, so we handle it intelligently here for compile-time speed.
10128 InsertNewInstBefore(NewPN, PN);
10132 // Insert and return the new operation.
10133 if (CastInst* FirstCI = dyn_cast<CastInst>(FirstInst))
10134 return CastInst::Create(FirstCI->getOpcode(), PhiVal, PN.getType());
10135 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
10136 return BinaryOperator::Create(BinOp->getOpcode(), PhiVal, ConstantOp);
10137 if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst))
10138 return CmpInst::Create(CIOp->getOpcode(), CIOp->getPredicate(),
10139 PhiVal, ConstantOp);
10140 assert(isa<LoadInst>(FirstInst) && "Unknown operation");
10142 // If this was a volatile load that we are merging, make sure to loop through
10143 // and mark all the input loads as non-volatile. If we don't do this, we will
10144 // insert a new volatile load and the old ones will not be deletable.
10146 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i)
10147 cast<LoadInst>(PN.getIncomingValue(i))->setVolatile(false);
10149 return new LoadInst(PhiVal, "", isVolatile);
10152 /// DeadPHICycle - Return true if this PHI node is only used by a PHI node cycle
10154 static bool DeadPHICycle(PHINode *PN,
10155 SmallPtrSet<PHINode*, 16> &PotentiallyDeadPHIs) {
10156 if (PN->use_empty()) return true;
10157 if (!PN->hasOneUse()) return false;
10159 // Remember this node, and if we find the cycle, return.
10160 if (!PotentiallyDeadPHIs.insert(PN))
10163 // Don't scan crazily complex things.
10164 if (PotentiallyDeadPHIs.size() == 16)
10167 if (PHINode *PU = dyn_cast<PHINode>(PN->use_back()))
10168 return DeadPHICycle(PU, PotentiallyDeadPHIs);
10173 /// PHIsEqualValue - Return true if this phi node is always equal to
10174 /// NonPhiInVal. This happens with mutually cyclic phi nodes like:
10175 /// z = some value; x = phi (y, z); y = phi (x, z)
10176 static bool PHIsEqualValue(PHINode *PN, Value *NonPhiInVal,
10177 SmallPtrSet<PHINode*, 16> &ValueEqualPHIs) {
10178 // See if we already saw this PHI node.
10179 if (!ValueEqualPHIs.insert(PN))
10182 // Don't scan crazily complex things.
10183 if (ValueEqualPHIs.size() == 16)
10186 // Scan the operands to see if they are either phi nodes or are equal to
10188 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
10189 Value *Op = PN->getIncomingValue(i);
10190 if (PHINode *OpPN = dyn_cast<PHINode>(Op)) {
10191 if (!PHIsEqualValue(OpPN, NonPhiInVal, ValueEqualPHIs))
10193 } else if (Op != NonPhiInVal)
10201 // PHINode simplification
10203 Instruction *InstCombiner::visitPHINode(PHINode &PN) {
10204 // If LCSSA is around, don't mess with Phi nodes
10205 if (MustPreserveLCSSA) return 0;
10207 if (Value *V = PN.hasConstantValue())
10208 return ReplaceInstUsesWith(PN, V);
10210 // If all PHI operands are the same operation, pull them through the PHI,
10211 // reducing code size.
10212 if (isa<Instruction>(PN.getIncomingValue(0)) &&
10213 PN.getIncomingValue(0)->hasOneUse())
10214 if (Instruction *Result = FoldPHIArgOpIntoPHI(PN))
10217 // If this is a trivial cycle in the PHI node graph, remove it. Basically, if
10218 // this PHI only has a single use (a PHI), and if that PHI only has one use (a
10219 // PHI)... break the cycle.
10220 if (PN.hasOneUse()) {
10221 Instruction *PHIUser = cast<Instruction>(PN.use_back());
10222 if (PHINode *PU = dyn_cast<PHINode>(PHIUser)) {
10223 SmallPtrSet<PHINode*, 16> PotentiallyDeadPHIs;
10224 PotentiallyDeadPHIs.insert(&PN);
10225 if (DeadPHICycle(PU, PotentiallyDeadPHIs))
10226 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
10229 // If this phi has a single use, and if that use just computes a value for
10230 // the next iteration of a loop, delete the phi. This occurs with unused
10231 // induction variables, e.g. "for (int j = 0; ; ++j);". Detecting this
10232 // common case here is good because the only other things that catch this
10233 // are induction variable analysis (sometimes) and ADCE, which is only run
10235 if (PHIUser->hasOneUse() &&
10236 (isa<BinaryOperator>(PHIUser) || isa<GetElementPtrInst>(PHIUser)) &&
10237 PHIUser->use_back() == &PN) {
10238 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
10242 // We sometimes end up with phi cycles that non-obviously end up being the
10243 // same value, for example:
10244 // z = some value; x = phi (y, z); y = phi (x, z)
10245 // where the phi nodes don't necessarily need to be in the same block. Do a
10246 // quick check to see if the PHI node only contains a single non-phi value, if
10247 // so, scan to see if the phi cycle is actually equal to that value.
10249 unsigned InValNo = 0, NumOperandVals = PN.getNumIncomingValues();
10250 // Scan for the first non-phi operand.
10251 while (InValNo != NumOperandVals &&
10252 isa<PHINode>(PN.getIncomingValue(InValNo)))
10255 if (InValNo != NumOperandVals) {
10256 Value *NonPhiInVal = PN.getOperand(InValNo);
10258 // Scan the rest of the operands to see if there are any conflicts, if so
10259 // there is no need to recursively scan other phis.
10260 for (++InValNo; InValNo != NumOperandVals; ++InValNo) {
10261 Value *OpVal = PN.getIncomingValue(InValNo);
10262 if (OpVal != NonPhiInVal && !isa<PHINode>(OpVal))
10266 // If we scanned over all operands, then we have one unique value plus
10267 // phi values. Scan PHI nodes to see if they all merge in each other or
10269 if (InValNo == NumOperandVals) {
10270 SmallPtrSet<PHINode*, 16> ValueEqualPHIs;
10271 if (PHIsEqualValue(&PN, NonPhiInVal, ValueEqualPHIs))
10272 return ReplaceInstUsesWith(PN, NonPhiInVal);
10279 static Value *InsertCastToIntPtrTy(Value *V, const Type *DTy,
10280 Instruction *InsertPoint,
10281 InstCombiner *IC) {
10282 unsigned PtrSize = DTy->getPrimitiveSizeInBits();
10283 unsigned VTySize = V->getType()->getPrimitiveSizeInBits();
10284 // We must cast correctly to the pointer type. Ensure that we
10285 // sign extend the integer value if it is smaller as this is
10286 // used for address computation.
10287 Instruction::CastOps opcode =
10288 (VTySize < PtrSize ? Instruction::SExt :
10289 (VTySize == PtrSize ? Instruction::BitCast : Instruction::Trunc));
10290 return IC->InsertCastBefore(opcode, V, DTy, *InsertPoint);
10294 Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) {
10295 Value *PtrOp = GEP.getOperand(0);
10296 // Is it 'getelementptr %P, i32 0' or 'getelementptr %P'
10297 // If so, eliminate the noop.
10298 if (GEP.getNumOperands() == 1)
10299 return ReplaceInstUsesWith(GEP, PtrOp);
10301 if (isa<UndefValue>(GEP.getOperand(0)))
10302 return ReplaceInstUsesWith(GEP, UndefValue::get(GEP.getType()));
10304 bool HasZeroPointerIndex = false;
10305 if (Constant *C = dyn_cast<Constant>(GEP.getOperand(1)))
10306 HasZeroPointerIndex = C->isNullValue();
10308 if (GEP.getNumOperands() == 2 && HasZeroPointerIndex)
10309 return ReplaceInstUsesWith(GEP, PtrOp);
10311 // Eliminate unneeded casts for indices.
10312 bool MadeChange = false;
10314 gep_type_iterator GTI = gep_type_begin(GEP);
10315 for (User::op_iterator i = GEP.op_begin() + 1, e = GEP.op_end();
10316 i != e; ++i, ++GTI) {
10317 if (isa<SequentialType>(*GTI)) {
10318 if (CastInst *CI = dyn_cast<CastInst>(*i)) {
10319 if (CI->getOpcode() == Instruction::ZExt ||
10320 CI->getOpcode() == Instruction::SExt) {
10321 const Type *SrcTy = CI->getOperand(0)->getType();
10322 // We can eliminate a cast from i32 to i64 iff the target
10323 // is a 32-bit pointer target.
10324 if (SrcTy->getPrimitiveSizeInBits() >= TD->getPointerSizeInBits()) {
10326 *i = CI->getOperand(0);
10330 // If we are using a wider index than needed for this platform, shrink it
10331 // to what we need. If narrower, sign-extend it to what we need.
10332 // If the incoming value needs a cast instruction,
10333 // insert it. This explicit cast can make subsequent optimizations more
10336 if (TD->getTypeSizeInBits(Op->getType()) > TD->getPointerSizeInBits()) {
10337 if (Constant *C = dyn_cast<Constant>(Op)) {
10338 *i = ConstantExpr::getTrunc(C, TD->getIntPtrType());
10341 Op = InsertCastBefore(Instruction::Trunc, Op, TD->getIntPtrType(),
10346 } else if (TD->getTypeSizeInBits(Op->getType()) < TD->getPointerSizeInBits()) {
10347 if (Constant *C = dyn_cast<Constant>(Op)) {
10348 *i = ConstantExpr::getSExt(C, TD->getIntPtrType());
10351 Op = InsertCastBefore(Instruction::SExt, Op, TD->getIntPtrType(),
10359 if (MadeChange) return &GEP;
10361 // If this GEP instruction doesn't move the pointer, and if the input operand
10362 // is a bitcast of another pointer, just replace the GEP with a bitcast of the
10363 // real input to the dest type.
10364 if (GEP.hasAllZeroIndices()) {
10365 if (BitCastInst *BCI = dyn_cast<BitCastInst>(GEP.getOperand(0))) {
10366 // If the bitcast is of an allocation, and the allocation will be
10367 // converted to match the type of the cast, don't touch this.
10368 if (isa<AllocationInst>(BCI->getOperand(0))) {
10369 // See if the bitcast simplifies, if so, don't nuke this GEP yet.
10370 if (Instruction *I = visitBitCast(*BCI)) {
10373 BCI->getParent()->getInstList().insert(BCI, I);
10374 ReplaceInstUsesWith(*BCI, I);
10379 return new BitCastInst(BCI->getOperand(0), GEP.getType());
10383 // Combine Indices - If the source pointer to this getelementptr instruction
10384 // is a getelementptr instruction, combine the indices of the two
10385 // getelementptr instructions into a single instruction.
10387 SmallVector<Value*, 8> SrcGEPOperands;
10388 if (User *Src = dyn_castGetElementPtr(PtrOp))
10389 SrcGEPOperands.append(Src->op_begin(), Src->op_end());
10391 if (!SrcGEPOperands.empty()) {
10392 // Note that if our source is a gep chain itself that we wait for that
10393 // chain to be resolved before we perform this transformation. This
10394 // avoids us creating a TON of code in some cases.
10396 if (isa<GetElementPtrInst>(SrcGEPOperands[0]) &&
10397 cast<Instruction>(SrcGEPOperands[0])->getNumOperands() == 2)
10398 return 0; // Wait until our source is folded to completion.
10400 SmallVector<Value*, 8> Indices;
10402 // Find out whether the last index in the source GEP is a sequential idx.
10403 bool EndsWithSequential = false;
10404 for (gep_type_iterator I = gep_type_begin(*cast<User>(PtrOp)),
10405 E = gep_type_end(*cast<User>(PtrOp)); I != E; ++I)
10406 EndsWithSequential = !isa<StructType>(*I);
10408 // Can we combine the two pointer arithmetics offsets?
10409 if (EndsWithSequential) {
10410 // Replace: gep (gep %P, long B), long A, ...
10411 // With: T = long A+B; gep %P, T, ...
10413 Value *Sum, *SO1 = SrcGEPOperands.back(), *GO1 = GEP.getOperand(1);
10414 if (SO1 == Constant::getNullValue(SO1->getType())) {
10416 } else if (GO1 == Constant::getNullValue(GO1->getType())) {
10419 // If they aren't the same type, convert both to an integer of the
10420 // target's pointer size.
10421 if (SO1->getType() != GO1->getType()) {
10422 if (Constant *SO1C = dyn_cast<Constant>(SO1)) {
10423 SO1 = ConstantExpr::getIntegerCast(SO1C, GO1->getType(), true);
10424 } else if (Constant *GO1C = dyn_cast<Constant>(GO1)) {
10425 GO1 = ConstantExpr::getIntegerCast(GO1C, SO1->getType(), true);
10427 unsigned PS = TD->getPointerSizeInBits();
10428 if (TD->getTypeSizeInBits(SO1->getType()) == PS) {
10429 // Convert GO1 to SO1's type.
10430 GO1 = InsertCastToIntPtrTy(GO1, SO1->getType(), &GEP, this);
10432 } else if (TD->getTypeSizeInBits(GO1->getType()) == PS) {
10433 // Convert SO1 to GO1's type.
10434 SO1 = InsertCastToIntPtrTy(SO1, GO1->getType(), &GEP, this);
10436 const Type *PT = TD->getIntPtrType();
10437 SO1 = InsertCastToIntPtrTy(SO1, PT, &GEP, this);
10438 GO1 = InsertCastToIntPtrTy(GO1, PT, &GEP, this);
10442 if (isa<Constant>(SO1) && isa<Constant>(GO1))
10443 Sum = ConstantExpr::getAdd(cast<Constant>(SO1), cast<Constant>(GO1));
10445 Sum = BinaryOperator::CreateAdd(SO1, GO1, PtrOp->getName()+".sum");
10446 InsertNewInstBefore(cast<Instruction>(Sum), GEP);
10450 // Recycle the GEP we already have if possible.
10451 if (SrcGEPOperands.size() == 2) {
10452 GEP.setOperand(0, SrcGEPOperands[0]);
10453 GEP.setOperand(1, Sum);
10456 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
10457 SrcGEPOperands.end()-1);
10458 Indices.push_back(Sum);
10459 Indices.insert(Indices.end(), GEP.op_begin()+2, GEP.op_end());
10461 } else if (isa<Constant>(*GEP.idx_begin()) &&
10462 cast<Constant>(*GEP.idx_begin())->isNullValue() &&
10463 SrcGEPOperands.size() != 1) {
10464 // Otherwise we can do the fold if the first index of the GEP is a zero
10465 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
10466 SrcGEPOperands.end());
10467 Indices.insert(Indices.end(), GEP.idx_begin()+1, GEP.idx_end());
10470 if (!Indices.empty())
10471 return GetElementPtrInst::Create(SrcGEPOperands[0], Indices.begin(),
10472 Indices.end(), GEP.getName());
10474 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(PtrOp)) {
10475 // GEP of global variable. If all of the indices for this GEP are
10476 // constants, we can promote this to a constexpr instead of an instruction.
10478 // Scan for nonconstants...
10479 SmallVector<Constant*, 8> Indices;
10480 User::op_iterator I = GEP.idx_begin(), E = GEP.idx_end();
10481 for (; I != E && isa<Constant>(*I); ++I)
10482 Indices.push_back(cast<Constant>(*I));
10484 if (I == E) { // If they are all constants...
10485 Constant *CE = ConstantExpr::getGetElementPtr(GV,
10486 &Indices[0],Indices.size());
10488 // Replace all uses of the GEP with the new constexpr...
10489 return ReplaceInstUsesWith(GEP, CE);
10491 } else if (Value *X = getBitCastOperand(PtrOp)) { // Is the operand a cast?
10492 if (!isa<PointerType>(X->getType())) {
10493 // Not interesting. Source pointer must be a cast from pointer.
10494 } else if (HasZeroPointerIndex) {
10495 // transform: GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ...
10496 // into : GEP [10 x i8]* X, i32 0, ...
10498 // This occurs when the program declares an array extern like "int X[];"
10500 const PointerType *CPTy = cast<PointerType>(PtrOp->getType());
10501 const PointerType *XTy = cast<PointerType>(X->getType());
10502 if (const ArrayType *XATy =
10503 dyn_cast<ArrayType>(XTy->getElementType()))
10504 if (const ArrayType *CATy =
10505 dyn_cast<ArrayType>(CPTy->getElementType()))
10506 if (CATy->getElementType() == XATy->getElementType()) {
10507 // At this point, we know that the cast source type is a pointer
10508 // to an array of the same type as the destination pointer
10509 // array. Because the array type is never stepped over (there
10510 // is a leading zero) we can fold the cast into this GEP.
10511 GEP.setOperand(0, X);
10514 } else if (GEP.getNumOperands() == 2) {
10515 // Transform things like:
10516 // %t = getelementptr i32* bitcast ([2 x i32]* %str to i32*), i32 %V
10517 // into: %t1 = getelementptr [2 x i32]* %str, i32 0, i32 %V; bitcast
10518 const Type *SrcElTy = cast<PointerType>(X->getType())->getElementType();
10519 const Type *ResElTy=cast<PointerType>(PtrOp->getType())->getElementType();
10520 if (isa<ArrayType>(SrcElTy) &&
10521 TD->getABITypeSize(cast<ArrayType>(SrcElTy)->getElementType()) ==
10522 TD->getABITypeSize(ResElTy)) {
10524 Idx[0] = Constant::getNullValue(Type::Int32Ty);
10525 Idx[1] = GEP.getOperand(1);
10526 Value *V = InsertNewInstBefore(
10527 GetElementPtrInst::Create(X, Idx, Idx + 2, GEP.getName()), GEP);
10528 // V and GEP are both pointer types --> BitCast
10529 return new BitCastInst(V, GEP.getType());
10532 // Transform things like:
10533 // getelementptr i8* bitcast ([100 x double]* X to i8*), i32 %tmp
10534 // (where tmp = 8*tmp2) into:
10535 // getelementptr [100 x double]* %arr, i32 0, i32 %tmp2; bitcast
10537 if (isa<ArrayType>(SrcElTy) && ResElTy == Type::Int8Ty) {
10538 uint64_t ArrayEltSize =
10539 TD->getABITypeSize(cast<ArrayType>(SrcElTy)->getElementType());
10541 // Check to see if "tmp" is a scale by a multiple of ArrayEltSize. We
10542 // allow either a mul, shift, or constant here.
10544 ConstantInt *Scale = 0;
10545 if (ArrayEltSize == 1) {
10546 NewIdx = GEP.getOperand(1);
10547 Scale = ConstantInt::get(NewIdx->getType(), 1);
10548 } else if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP.getOperand(1))) {
10549 NewIdx = ConstantInt::get(CI->getType(), 1);
10551 } else if (Instruction *Inst =dyn_cast<Instruction>(GEP.getOperand(1))){
10552 if (Inst->getOpcode() == Instruction::Shl &&
10553 isa<ConstantInt>(Inst->getOperand(1))) {
10554 ConstantInt *ShAmt = cast<ConstantInt>(Inst->getOperand(1));
10555 uint32_t ShAmtVal = ShAmt->getLimitedValue(64);
10556 Scale = ConstantInt::get(Inst->getType(), 1ULL << ShAmtVal);
10557 NewIdx = Inst->getOperand(0);
10558 } else if (Inst->getOpcode() == Instruction::Mul &&
10559 isa<ConstantInt>(Inst->getOperand(1))) {
10560 Scale = cast<ConstantInt>(Inst->getOperand(1));
10561 NewIdx = Inst->getOperand(0);
10565 // If the index will be to exactly the right offset with the scale taken
10566 // out, perform the transformation. Note, we don't know whether Scale is
10567 // signed or not. We'll use unsigned version of division/modulo
10568 // operation after making sure Scale doesn't have the sign bit set.
10569 if (Scale && Scale->getSExtValue() >= 0LL &&
10570 Scale->getZExtValue() % ArrayEltSize == 0) {
10571 Scale = ConstantInt::get(Scale->getType(),
10572 Scale->getZExtValue() / ArrayEltSize);
10573 if (Scale->getZExtValue() != 1) {
10574 Constant *C = ConstantExpr::getIntegerCast(Scale, NewIdx->getType(),
10576 Instruction *Sc = BinaryOperator::CreateMul(NewIdx, C, "idxscale");
10577 NewIdx = InsertNewInstBefore(Sc, GEP);
10580 // Insert the new GEP instruction.
10582 Idx[0] = Constant::getNullValue(Type::Int32Ty);
10584 Instruction *NewGEP =
10585 GetElementPtrInst::Create(X, Idx, Idx + 2, GEP.getName());
10586 NewGEP = InsertNewInstBefore(NewGEP, GEP);
10587 // The NewGEP must be pointer typed, so must the old one -> BitCast
10588 return new BitCastInst(NewGEP, GEP.getType());
10597 Instruction *InstCombiner::visitAllocationInst(AllocationInst &AI) {
10598 // Convert: malloc Ty, C - where C is a constant != 1 into: malloc [C x Ty], 1
10599 if (AI.isArrayAllocation()) { // Check C != 1
10600 if (const ConstantInt *C = dyn_cast<ConstantInt>(AI.getArraySize())) {
10601 const Type *NewTy =
10602 ArrayType::get(AI.getAllocatedType(), C->getZExtValue());
10603 AllocationInst *New = 0;
10605 // Create and insert the replacement instruction...
10606 if (isa<MallocInst>(AI))
10607 New = new MallocInst(NewTy, 0, AI.getAlignment(), AI.getName());
10609 assert(isa<AllocaInst>(AI) && "Unknown type of allocation inst!");
10610 New = new AllocaInst(NewTy, 0, AI.getAlignment(), AI.getName());
10613 InsertNewInstBefore(New, AI);
10615 // Scan to the end of the allocation instructions, to skip over a block of
10616 // allocas if possible...
10618 BasicBlock::iterator It = New;
10619 while (isa<AllocationInst>(*It)) ++It;
10621 // Now that I is pointing to the first non-allocation-inst in the block,
10622 // insert our getelementptr instruction...
10624 Value *NullIdx = Constant::getNullValue(Type::Int32Ty);
10628 Value *V = GetElementPtrInst::Create(New, Idx, Idx + 2,
10629 New->getName()+".sub", It);
10631 // Now make everything use the getelementptr instead of the original
10633 return ReplaceInstUsesWith(AI, V);
10634 } else if (isa<UndefValue>(AI.getArraySize())) {
10635 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
10639 // If alloca'ing a zero byte object, replace the alloca with a null pointer.
10640 // Note that we only do this for alloca's, because malloc should allocate and
10641 // return a unique pointer, even for a zero byte allocation.
10642 if (isa<AllocaInst>(AI) && AI.getAllocatedType()->isSized() &&
10643 TD->getABITypeSize(AI.getAllocatedType()) == 0)
10644 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
10649 Instruction *InstCombiner::visitFreeInst(FreeInst &FI) {
10650 Value *Op = FI.getOperand(0);
10652 // free undef -> unreachable.
10653 if (isa<UndefValue>(Op)) {
10654 // Insert a new store to null because we cannot modify the CFG here.
10655 new StoreInst(ConstantInt::getTrue(),
10656 UndefValue::get(PointerType::getUnqual(Type::Int1Ty)), &FI);
10657 return EraseInstFromFunction(FI);
10660 // If we have 'free null' delete the instruction. This can happen in stl code
10661 // when lots of inlining happens.
10662 if (isa<ConstantPointerNull>(Op))
10663 return EraseInstFromFunction(FI);
10665 // Change free <ty>* (cast <ty2>* X to <ty>*) into free <ty2>* X
10666 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op)) {
10667 FI.setOperand(0, CI->getOperand(0));
10671 // Change free (gep X, 0,0,0,0) into free(X)
10672 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op)) {
10673 if (GEPI->hasAllZeroIndices()) {
10674 AddToWorkList(GEPI);
10675 FI.setOperand(0, GEPI->getOperand(0));
10680 // Change free(malloc) into nothing, if the malloc has a single use.
10681 if (MallocInst *MI = dyn_cast<MallocInst>(Op))
10682 if (MI->hasOneUse()) {
10683 EraseInstFromFunction(FI);
10684 return EraseInstFromFunction(*MI);
10691 /// InstCombineLoadCast - Fold 'load (cast P)' -> cast (load P)' when possible.
10692 static Instruction *InstCombineLoadCast(InstCombiner &IC, LoadInst &LI,
10693 const TargetData *TD) {
10694 User *CI = cast<User>(LI.getOperand(0));
10695 Value *CastOp = CI->getOperand(0);
10697 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(CI)) {
10698 // Instead of loading constant c string, use corresponding integer value
10699 // directly if string length is small enough.
10701 if (GetConstantStringInfo(CE->getOperand(0), Str) && !Str.empty()) {
10702 unsigned len = Str.length();
10703 const Type *Ty = cast<PointerType>(CE->getType())->getElementType();
10704 unsigned numBits = Ty->getPrimitiveSizeInBits();
10705 // Replace LI with immediate integer store.
10706 if ((numBits >> 3) == len + 1) {
10707 APInt StrVal(numBits, 0);
10708 APInt SingleChar(numBits, 0);
10709 if (TD->isLittleEndian()) {
10710 for (signed i = len-1; i >= 0; i--) {
10711 SingleChar = (uint64_t) Str[i];
10712 StrVal = (StrVal << 8) | SingleChar;
10715 for (unsigned i = 0; i < len; i++) {
10716 SingleChar = (uint64_t) Str[i];
10717 StrVal = (StrVal << 8) | SingleChar;
10719 // Append NULL at the end.
10721 StrVal = (StrVal << 8) | SingleChar;
10723 Value *NL = ConstantInt::get(StrVal);
10724 return IC.ReplaceInstUsesWith(LI, NL);
10729 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
10730 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
10731 const Type *SrcPTy = SrcTy->getElementType();
10733 if (DestPTy->isInteger() || isa<PointerType>(DestPTy) ||
10734 isa<VectorType>(DestPTy)) {
10735 // If the source is an array, the code below will not succeed. Check to
10736 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
10738 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
10739 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
10740 if (ASrcTy->getNumElements() != 0) {
10742 Idxs[0] = Idxs[1] = Constant::getNullValue(Type::Int32Ty);
10743 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs, 2);
10744 SrcTy = cast<PointerType>(CastOp->getType());
10745 SrcPTy = SrcTy->getElementType();
10748 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy) ||
10749 isa<VectorType>(SrcPTy)) &&
10750 // Do not allow turning this into a load of an integer, which is then
10751 // casted to a pointer, this pessimizes pointer analysis a lot.
10752 (isa<PointerType>(SrcPTy) == isa<PointerType>(LI.getType())) &&
10753 IC.getTargetData().getTypeSizeInBits(SrcPTy) ==
10754 IC.getTargetData().getTypeSizeInBits(DestPTy)) {
10756 // Okay, we are casting from one integer or pointer type to another of
10757 // the same size. Instead of casting the pointer before the load, cast
10758 // the result of the loaded value.
10759 Value *NewLoad = IC.InsertNewInstBefore(new LoadInst(CastOp,
10761 LI.isVolatile()),LI);
10762 // Now cast the result of the load.
10763 return new BitCastInst(NewLoad, LI.getType());
10770 /// isSafeToLoadUnconditionally - Return true if we know that executing a load
10771 /// from this value cannot trap. If it is not obviously safe to load from the
10772 /// specified pointer, we do a quick local scan of the basic block containing
10773 /// ScanFrom, to determine if the address is already accessed.
10774 static bool isSafeToLoadUnconditionally(Value *V, Instruction *ScanFrom) {
10775 // If it is an alloca it is always safe to load from.
10776 if (isa<AllocaInst>(V)) return true;
10778 // If it is a global variable it is mostly safe to load from.
10779 if (const GlobalValue *GV = dyn_cast<GlobalVariable>(V))
10780 // Don't try to evaluate aliases. External weak GV can be null.
10781 return !isa<GlobalAlias>(GV) && !GV->hasExternalWeakLinkage();
10783 // Otherwise, be a little bit agressive by scanning the local block where we
10784 // want to check to see if the pointer is already being loaded or stored
10785 // from/to. If so, the previous load or store would have already trapped,
10786 // so there is no harm doing an extra load (also, CSE will later eliminate
10787 // the load entirely).
10788 BasicBlock::iterator BBI = ScanFrom, E = ScanFrom->getParent()->begin();
10793 // If we see a free or a call (which might do a free) the pointer could be
10795 if (isa<FreeInst>(BBI) || isa<CallInst>(BBI))
10798 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
10799 if (LI->getOperand(0) == V) return true;
10800 } else if (StoreInst *SI = dyn_cast<StoreInst>(BBI)) {
10801 if (SI->getOperand(1) == V) return true;
10808 Instruction *InstCombiner::visitLoadInst(LoadInst &LI) {
10809 Value *Op = LI.getOperand(0);
10811 // Attempt to improve the alignment.
10812 unsigned KnownAlign = GetOrEnforceKnownAlignment(Op);
10814 (LI.getAlignment() == 0 ? TD->getABITypeAlignment(LI.getType()) :
10815 LI.getAlignment()))
10816 LI.setAlignment(KnownAlign);
10818 // load (cast X) --> cast (load X) iff safe
10819 if (isa<CastInst>(Op))
10820 if (Instruction *Res = InstCombineLoadCast(*this, LI, TD))
10823 // None of the following transforms are legal for volatile loads.
10824 if (LI.isVolatile()) return 0;
10826 // Do really simple store-to-load forwarding and load CSE, to catch cases
10827 // where there are several consequtive memory accesses to the same location,
10828 // separated by a few arithmetic operations.
10829 BasicBlock::iterator BBI = &LI;
10830 if (Value *AvailableVal = FindAvailableLoadedValue(Op, LI.getParent(), BBI,6))
10831 return ReplaceInstUsesWith(LI, AvailableVal);
10833 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op)) {
10834 const Value *GEPI0 = GEPI->getOperand(0);
10835 // TODO: Consider a target hook for valid address spaces for this xform.
10836 if (isa<ConstantPointerNull>(GEPI0) &&
10837 cast<PointerType>(GEPI0->getType())->getAddressSpace() == 0) {
10838 // Insert a new store to null instruction before the load to indicate
10839 // that this code is not reachable. We do this instead of inserting
10840 // an unreachable instruction directly because we cannot modify the
10842 new StoreInst(UndefValue::get(LI.getType()),
10843 Constant::getNullValue(Op->getType()), &LI);
10844 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
10848 if (Constant *C = dyn_cast<Constant>(Op)) {
10849 // load null/undef -> undef
10850 // TODO: Consider a target hook for valid address spaces for this xform.
10851 if (isa<UndefValue>(C) || (C->isNullValue() &&
10852 cast<PointerType>(Op->getType())->getAddressSpace() == 0)) {
10853 // Insert a new store to null instruction before the load to indicate that
10854 // this code is not reachable. We do this instead of inserting an
10855 // unreachable instruction directly because we cannot modify the CFG.
10856 new StoreInst(UndefValue::get(LI.getType()),
10857 Constant::getNullValue(Op->getType()), &LI);
10858 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
10861 // Instcombine load (constant global) into the value loaded.
10862 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Op))
10863 if (GV->isConstant() && !GV->isDeclaration())
10864 return ReplaceInstUsesWith(LI, GV->getInitializer());
10866 // Instcombine load (constantexpr_GEP global, 0, ...) into the value loaded.
10867 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Op)) {
10868 if (CE->getOpcode() == Instruction::GetElementPtr) {
10869 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(CE->getOperand(0)))
10870 if (GV->isConstant() && !GV->isDeclaration())
10872 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE))
10873 return ReplaceInstUsesWith(LI, V);
10874 if (CE->getOperand(0)->isNullValue()) {
10875 // Insert a new store to null instruction before the load to indicate
10876 // that this code is not reachable. We do this instead of inserting
10877 // an unreachable instruction directly because we cannot modify the
10879 new StoreInst(UndefValue::get(LI.getType()),
10880 Constant::getNullValue(Op->getType()), &LI);
10881 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
10884 } else if (CE->isCast()) {
10885 if (Instruction *Res = InstCombineLoadCast(*this, LI, TD))
10891 // If this load comes from anywhere in a constant global, and if the global
10892 // is all undef or zero, we know what it loads.
10893 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Op->getUnderlyingObject())){
10894 if (GV->isConstant() && GV->hasInitializer()) {
10895 if (GV->getInitializer()->isNullValue())
10896 return ReplaceInstUsesWith(LI, Constant::getNullValue(LI.getType()));
10897 else if (isa<UndefValue>(GV->getInitializer()))
10898 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
10902 if (Op->hasOneUse()) {
10903 // Change select and PHI nodes to select values instead of addresses: this
10904 // helps alias analysis out a lot, allows many others simplifications, and
10905 // exposes redundancy in the code.
10907 // Note that we cannot do the transformation unless we know that the
10908 // introduced loads cannot trap! Something like this is valid as long as
10909 // the condition is always false: load (select bool %C, int* null, int* %G),
10910 // but it would not be valid if we transformed it to load from null
10911 // unconditionally.
10913 if (SelectInst *SI = dyn_cast<SelectInst>(Op)) {
10914 // load (select (Cond, &V1, &V2)) --> select(Cond, load &V1, load &V2).
10915 if (isSafeToLoadUnconditionally(SI->getOperand(1), SI) &&
10916 isSafeToLoadUnconditionally(SI->getOperand(2), SI)) {
10917 Value *V1 = InsertNewInstBefore(new LoadInst(SI->getOperand(1),
10918 SI->getOperand(1)->getName()+".val"), LI);
10919 Value *V2 = InsertNewInstBefore(new LoadInst(SI->getOperand(2),
10920 SI->getOperand(2)->getName()+".val"), LI);
10921 return SelectInst::Create(SI->getCondition(), V1, V2);
10924 // load (select (cond, null, P)) -> load P
10925 if (Constant *C = dyn_cast<Constant>(SI->getOperand(1)))
10926 if (C->isNullValue()) {
10927 LI.setOperand(0, SI->getOperand(2));
10931 // load (select (cond, P, null)) -> load P
10932 if (Constant *C = dyn_cast<Constant>(SI->getOperand(2)))
10933 if (C->isNullValue()) {
10934 LI.setOperand(0, SI->getOperand(1));
10942 /// InstCombineStoreToCast - Fold store V, (cast P) -> store (cast V), P
10944 static Instruction *InstCombineStoreToCast(InstCombiner &IC, StoreInst &SI) {
10945 User *CI = cast<User>(SI.getOperand(1));
10946 Value *CastOp = CI->getOperand(0);
10948 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
10949 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
10950 const Type *SrcPTy = SrcTy->getElementType();
10952 if (DestPTy->isInteger() || isa<PointerType>(DestPTy)) {
10953 // If the source is an array, the code below will not succeed. Check to
10954 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
10956 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
10957 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
10958 if (ASrcTy->getNumElements() != 0) {
10960 Idxs[0] = Idxs[1] = Constant::getNullValue(Type::Int32Ty);
10961 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs, 2);
10962 SrcTy = cast<PointerType>(CastOp->getType());
10963 SrcPTy = SrcTy->getElementType();
10966 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy)) &&
10967 IC.getTargetData().getTypeSizeInBits(SrcPTy) ==
10968 IC.getTargetData().getTypeSizeInBits(DestPTy)) {
10970 // Okay, we are casting from one integer or pointer type to another of
10971 // the same size. Instead of casting the pointer before
10972 // the store, cast the value to be stored.
10974 Value *SIOp0 = SI.getOperand(0);
10975 Instruction::CastOps opcode = Instruction::BitCast;
10976 const Type* CastSrcTy = SIOp0->getType();
10977 const Type* CastDstTy = SrcPTy;
10978 if (isa<PointerType>(CastDstTy)) {
10979 if (CastSrcTy->isInteger())
10980 opcode = Instruction::IntToPtr;
10981 } else if (isa<IntegerType>(CastDstTy)) {
10982 if (isa<PointerType>(SIOp0->getType()))
10983 opcode = Instruction::PtrToInt;
10985 if (Constant *C = dyn_cast<Constant>(SIOp0))
10986 NewCast = ConstantExpr::getCast(opcode, C, CastDstTy);
10988 NewCast = IC.InsertNewInstBefore(
10989 CastInst::Create(opcode, SIOp0, CastDstTy, SIOp0->getName()+".c"),
10991 return new StoreInst(NewCast, CastOp);
10998 /// equivalentAddressValues - Test if A and B will obviously have the same
10999 /// value. This includes recognizing that %t0 and %t1 will have the same
11000 /// value in code like this:
11001 /// %t0 = getelementptr @a, 0, 3
11002 /// store i32 0, i32* %t0
11003 /// %t1 = getelementptr @a, 0, 3
11004 /// %t2 = load i32* %t1
11006 static bool equivalentAddressValues(Value *A, Value *B) {
11007 // Test if the values are trivially equivalent.
11008 if (A == B) return true;
11010 // Test if the values come form identical arithmetic instructions.
11011 if (isa<BinaryOperator>(A) ||
11012 isa<CastInst>(A) ||
11014 isa<GetElementPtrInst>(A))
11015 if (Instruction *BI = dyn_cast<Instruction>(B))
11016 if (cast<Instruction>(A)->isIdenticalTo(BI))
11019 // Otherwise they may not be equivalent.
11023 Instruction *InstCombiner::visitStoreInst(StoreInst &SI) {
11024 Value *Val = SI.getOperand(0);
11025 Value *Ptr = SI.getOperand(1);
11027 if (isa<UndefValue>(Ptr)) { // store X, undef -> noop (even if volatile)
11028 EraseInstFromFunction(SI);
11033 // If the RHS is an alloca with a single use, zapify the store, making the
11035 if (Ptr->hasOneUse() && !SI.isVolatile()) {
11036 if (isa<AllocaInst>(Ptr)) {
11037 EraseInstFromFunction(SI);
11042 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr))
11043 if (isa<AllocaInst>(GEP->getOperand(0)) &&
11044 GEP->getOperand(0)->hasOneUse()) {
11045 EraseInstFromFunction(SI);
11051 // Attempt to improve the alignment.
11052 unsigned KnownAlign = GetOrEnforceKnownAlignment(Ptr);
11054 (SI.getAlignment() == 0 ? TD->getABITypeAlignment(Val->getType()) :
11055 SI.getAlignment()))
11056 SI.setAlignment(KnownAlign);
11058 // Do really simple DSE, to catch cases where there are several consequtive
11059 // stores to the same location, separated by a few arithmetic operations. This
11060 // situation often occurs with bitfield accesses.
11061 BasicBlock::iterator BBI = &SI;
11062 for (unsigned ScanInsts = 6; BBI != SI.getParent()->begin() && ScanInsts;
11066 if (StoreInst *PrevSI = dyn_cast<StoreInst>(BBI)) {
11067 // Prev store isn't volatile, and stores to the same location?
11068 if (!PrevSI->isVolatile() &&equivalentAddressValues(PrevSI->getOperand(1),
11069 SI.getOperand(1))) {
11072 EraseInstFromFunction(*PrevSI);
11078 // If this is a load, we have to stop. However, if the loaded value is from
11079 // the pointer we're loading and is producing the pointer we're storing,
11080 // then *this* store is dead (X = load P; store X -> P).
11081 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
11082 if (LI == Val && equivalentAddressValues(LI->getOperand(0), Ptr) &&
11083 !SI.isVolatile()) {
11084 EraseInstFromFunction(SI);
11088 // Otherwise, this is a load from some other location. Stores before it
11089 // may not be dead.
11093 // Don't skip over loads or things that can modify memory.
11094 if (BBI->mayWriteToMemory() || BBI->mayReadFromMemory())
11099 if (SI.isVolatile()) return 0; // Don't hack volatile stores.
11101 // store X, null -> turns into 'unreachable' in SimplifyCFG
11102 if (isa<ConstantPointerNull>(Ptr)) {
11103 if (!isa<UndefValue>(Val)) {
11104 SI.setOperand(0, UndefValue::get(Val->getType()));
11105 if (Instruction *U = dyn_cast<Instruction>(Val))
11106 AddToWorkList(U); // Dropped a use.
11109 return 0; // Do not modify these!
11112 // store undef, Ptr -> noop
11113 if (isa<UndefValue>(Val)) {
11114 EraseInstFromFunction(SI);
11119 // If the pointer destination is a cast, see if we can fold the cast into the
11121 if (isa<CastInst>(Ptr))
11122 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
11124 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
11126 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
11130 // If this store is the last instruction in the basic block, and if the block
11131 // ends with an unconditional branch, try to move it to the successor block.
11133 if (BranchInst *BI = dyn_cast<BranchInst>(BBI))
11134 if (BI->isUnconditional())
11135 if (SimplifyStoreAtEndOfBlock(SI))
11136 return 0; // xform done!
11141 /// SimplifyStoreAtEndOfBlock - Turn things like:
11142 /// if () { *P = v1; } else { *P = v2 }
11143 /// into a phi node with a store in the successor.
11145 /// Simplify things like:
11146 /// *P = v1; if () { *P = v2; }
11147 /// into a phi node with a store in the successor.
11149 bool InstCombiner::SimplifyStoreAtEndOfBlock(StoreInst &SI) {
11150 BasicBlock *StoreBB = SI.getParent();
11152 // Check to see if the successor block has exactly two incoming edges. If
11153 // so, see if the other predecessor contains a store to the same location.
11154 // if so, insert a PHI node (if needed) and move the stores down.
11155 BasicBlock *DestBB = StoreBB->getTerminator()->getSuccessor(0);
11157 // Determine whether Dest has exactly two predecessors and, if so, compute
11158 // the other predecessor.
11159 pred_iterator PI = pred_begin(DestBB);
11160 BasicBlock *OtherBB = 0;
11161 if (*PI != StoreBB)
11164 if (PI == pred_end(DestBB))
11167 if (*PI != StoreBB) {
11172 if (++PI != pred_end(DestBB))
11175 // Bail out if all the relevant blocks aren't distinct (this can happen,
11176 // for example, if SI is in an infinite loop)
11177 if (StoreBB == DestBB || OtherBB == DestBB)
11180 // Verify that the other block ends in a branch and is not otherwise empty.
11181 BasicBlock::iterator BBI = OtherBB->getTerminator();
11182 BranchInst *OtherBr = dyn_cast<BranchInst>(BBI);
11183 if (!OtherBr || BBI == OtherBB->begin())
11186 // If the other block ends in an unconditional branch, check for the 'if then
11187 // else' case. there is an instruction before the branch.
11188 StoreInst *OtherStore = 0;
11189 if (OtherBr->isUnconditional()) {
11190 // If this isn't a store, or isn't a store to the same location, bail out.
11192 OtherStore = dyn_cast<StoreInst>(BBI);
11193 if (!OtherStore || OtherStore->getOperand(1) != SI.getOperand(1))
11196 // Otherwise, the other block ended with a conditional branch. If one of the
11197 // destinations is StoreBB, then we have the if/then case.
11198 if (OtherBr->getSuccessor(0) != StoreBB &&
11199 OtherBr->getSuccessor(1) != StoreBB)
11202 // Okay, we know that OtherBr now goes to Dest and StoreBB, so this is an
11203 // if/then triangle. See if there is a store to the same ptr as SI that
11204 // lives in OtherBB.
11206 // Check to see if we find the matching store.
11207 if ((OtherStore = dyn_cast<StoreInst>(BBI))) {
11208 if (OtherStore->getOperand(1) != SI.getOperand(1))
11212 // If we find something that may be using or overwriting the stored
11213 // value, or if we run out of instructions, we can't do the xform.
11214 if (BBI->mayReadFromMemory() || BBI->mayWriteToMemory() ||
11215 BBI == OtherBB->begin())
11219 // In order to eliminate the store in OtherBr, we have to
11220 // make sure nothing reads or overwrites the stored value in
11222 for (BasicBlock::iterator I = StoreBB->begin(); &*I != &SI; ++I) {
11223 // FIXME: This should really be AA driven.
11224 if (I->mayReadFromMemory() || I->mayWriteToMemory())
11229 // Insert a PHI node now if we need it.
11230 Value *MergedVal = OtherStore->getOperand(0);
11231 if (MergedVal != SI.getOperand(0)) {
11232 PHINode *PN = PHINode::Create(MergedVal->getType(), "storemerge");
11233 PN->reserveOperandSpace(2);
11234 PN->addIncoming(SI.getOperand(0), SI.getParent());
11235 PN->addIncoming(OtherStore->getOperand(0), OtherBB);
11236 MergedVal = InsertNewInstBefore(PN, DestBB->front());
11239 // Advance to a place where it is safe to insert the new store and
11241 BBI = DestBB->getFirstNonPHI();
11242 InsertNewInstBefore(new StoreInst(MergedVal, SI.getOperand(1),
11243 OtherStore->isVolatile()), *BBI);
11245 // Nuke the old stores.
11246 EraseInstFromFunction(SI);
11247 EraseInstFromFunction(*OtherStore);
11253 Instruction *InstCombiner::visitBranchInst(BranchInst &BI) {
11254 // Change br (not X), label True, label False to: br X, label False, True
11256 BasicBlock *TrueDest;
11257 BasicBlock *FalseDest;
11258 if (match(&BI, m_Br(m_Not(m_Value(X)), TrueDest, FalseDest)) &&
11259 !isa<Constant>(X)) {
11260 // Swap Destinations and condition...
11261 BI.setCondition(X);
11262 BI.setSuccessor(0, FalseDest);
11263 BI.setSuccessor(1, TrueDest);
11267 // Cannonicalize fcmp_one -> fcmp_oeq
11268 FCmpInst::Predicate FPred; Value *Y;
11269 if (match(&BI, m_Br(m_FCmp(FPred, m_Value(X), m_Value(Y)),
11270 TrueDest, FalseDest)))
11271 if ((FPred == FCmpInst::FCMP_ONE || FPred == FCmpInst::FCMP_OLE ||
11272 FPred == FCmpInst::FCMP_OGE) && BI.getCondition()->hasOneUse()) {
11273 FCmpInst *I = cast<FCmpInst>(BI.getCondition());
11274 FCmpInst::Predicate NewPred = FCmpInst::getInversePredicate(FPred);
11275 Instruction *NewSCC = new FCmpInst(NewPred, X, Y, "", I);
11276 NewSCC->takeName(I);
11277 // Swap Destinations and condition...
11278 BI.setCondition(NewSCC);
11279 BI.setSuccessor(0, FalseDest);
11280 BI.setSuccessor(1, TrueDest);
11281 RemoveFromWorkList(I);
11282 I->eraseFromParent();
11283 AddToWorkList(NewSCC);
11287 // Cannonicalize icmp_ne -> icmp_eq
11288 ICmpInst::Predicate IPred;
11289 if (match(&BI, m_Br(m_ICmp(IPred, m_Value(X), m_Value(Y)),
11290 TrueDest, FalseDest)))
11291 if ((IPred == ICmpInst::ICMP_NE || IPred == ICmpInst::ICMP_ULE ||
11292 IPred == ICmpInst::ICMP_SLE || IPred == ICmpInst::ICMP_UGE ||
11293 IPred == ICmpInst::ICMP_SGE) && BI.getCondition()->hasOneUse()) {
11294 ICmpInst *I = cast<ICmpInst>(BI.getCondition());
11295 ICmpInst::Predicate NewPred = ICmpInst::getInversePredicate(IPred);
11296 Instruction *NewSCC = new ICmpInst(NewPred, X, Y, "", I);
11297 NewSCC->takeName(I);
11298 // Swap Destinations and condition...
11299 BI.setCondition(NewSCC);
11300 BI.setSuccessor(0, FalseDest);
11301 BI.setSuccessor(1, TrueDest);
11302 RemoveFromWorkList(I);
11303 I->eraseFromParent();;
11304 AddToWorkList(NewSCC);
11311 Instruction *InstCombiner::visitSwitchInst(SwitchInst &SI) {
11312 Value *Cond = SI.getCondition();
11313 if (Instruction *I = dyn_cast<Instruction>(Cond)) {
11314 if (I->getOpcode() == Instruction::Add)
11315 if (ConstantInt *AddRHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
11316 // change 'switch (X+4) case 1:' into 'switch (X) case -3'
11317 for (unsigned i = 2, e = SI.getNumOperands(); i != e; i += 2)
11318 SI.setOperand(i,ConstantExpr::getSub(cast<Constant>(SI.getOperand(i)),
11320 SI.setOperand(0, I->getOperand(0));
11328 Instruction *InstCombiner::visitExtractValueInst(ExtractValueInst &EV) {
11329 Value *Agg = EV.getAggregateOperand();
11331 if (!EV.hasIndices())
11332 return ReplaceInstUsesWith(EV, Agg);
11334 if (Constant *C = dyn_cast<Constant>(Agg)) {
11335 if (isa<UndefValue>(C))
11336 return ReplaceInstUsesWith(EV, UndefValue::get(EV.getType()));
11338 if (isa<ConstantAggregateZero>(C))
11339 return ReplaceInstUsesWith(EV, Constant::getNullValue(EV.getType()));
11341 if (isa<ConstantArray>(C) || isa<ConstantStruct>(C)) {
11342 // Extract the element indexed by the first index out of the constant
11343 Value *V = C->getOperand(*EV.idx_begin());
11344 if (EV.getNumIndices() > 1)
11345 // Extract the remaining indices out of the constant indexed by the
11347 return ExtractValueInst::Create(V, EV.idx_begin() + 1, EV.idx_end());
11349 return ReplaceInstUsesWith(EV, V);
11351 return 0; // Can't handle other constants
11353 if (InsertValueInst *IV = dyn_cast<InsertValueInst>(Agg)) {
11354 // We're extracting from an insertvalue instruction, compare the indices
11355 const unsigned *exti, *exte, *insi, *inse;
11356 for (exti = EV.idx_begin(), insi = IV->idx_begin(),
11357 exte = EV.idx_end(), inse = IV->idx_end();
11358 exti != exte && insi != inse;
11360 if (*insi != *exti)
11361 // The insert and extract both reference distinctly different elements.
11362 // This means the extract is not influenced by the insert, and we can
11363 // replace the aggregate operand of the extract with the aggregate
11364 // operand of the insert. i.e., replace
11365 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1
11366 // %E = extractvalue { i32, { i32 } } %I, 0
11368 // %E = extractvalue { i32, { i32 } } %A, 0
11369 return ExtractValueInst::Create(IV->getAggregateOperand(),
11370 EV.idx_begin(), EV.idx_end());
11372 if (exti == exte && insi == inse)
11373 // Both iterators are at the end: Index lists are identical. Replace
11374 // %B = insertvalue { i32, { i32 } } %A, i32 42, 1, 0
11375 // %C = extractvalue { i32, { i32 } } %B, 1, 0
11377 return ReplaceInstUsesWith(EV, IV->getInsertedValueOperand());
11378 if (exti == exte) {
11379 // The extract list is a prefix of the insert list. i.e. replace
11380 // %I = insertvalue { i32, { i32 } } %A, i32 42, 1, 0
11381 // %E = extractvalue { i32, { i32 } } %I, 1
11383 // %X = extractvalue { i32, { i32 } } %A, 1
11384 // %E = insertvalue { i32 } %X, i32 42, 0
11385 // by switching the order of the insert and extract (though the
11386 // insertvalue should be left in, since it may have other uses).
11387 Value *NewEV = InsertNewInstBefore(
11388 ExtractValueInst::Create(IV->getAggregateOperand(),
11389 EV.idx_begin(), EV.idx_end()),
11391 return InsertValueInst::Create(NewEV, IV->getInsertedValueOperand(),
11395 // The insert list is a prefix of the extract list
11396 // We can simply remove the common indices from the extract and make it
11397 // operate on the inserted value instead of the insertvalue result.
11399 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1
11400 // %E = extractvalue { i32, { i32 } } %I, 1, 0
11402 // %E extractvalue { i32 } { i32 42 }, 0
11403 return ExtractValueInst::Create(IV->getInsertedValueOperand(),
11406 // Can't simplify extracts from other values. Note that nested extracts are
11407 // already simplified implicitely by the above (extract ( extract (insert) )
11408 // will be translated into extract ( insert ( extract ) ) first and then just
11409 // the value inserted, if appropriate).
11413 /// CheapToScalarize - Return true if the value is cheaper to scalarize than it
11414 /// is to leave as a vector operation.
11415 static bool CheapToScalarize(Value *V, bool isConstant) {
11416 if (isa<ConstantAggregateZero>(V))
11418 if (ConstantVector *C = dyn_cast<ConstantVector>(V)) {
11419 if (isConstant) return true;
11420 // If all elts are the same, we can extract.
11421 Constant *Op0 = C->getOperand(0);
11422 for (unsigned i = 1; i < C->getNumOperands(); ++i)
11423 if (C->getOperand(i) != Op0)
11427 Instruction *I = dyn_cast<Instruction>(V);
11428 if (!I) return false;
11430 // Insert element gets simplified to the inserted element or is deleted if
11431 // this is constant idx extract element and its a constant idx insertelt.
11432 if (I->getOpcode() == Instruction::InsertElement && isConstant &&
11433 isa<ConstantInt>(I->getOperand(2)))
11435 if (I->getOpcode() == Instruction::Load && I->hasOneUse())
11437 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I))
11438 if (BO->hasOneUse() &&
11439 (CheapToScalarize(BO->getOperand(0), isConstant) ||
11440 CheapToScalarize(BO->getOperand(1), isConstant)))
11442 if (CmpInst *CI = dyn_cast<CmpInst>(I))
11443 if (CI->hasOneUse() &&
11444 (CheapToScalarize(CI->getOperand(0), isConstant) ||
11445 CheapToScalarize(CI->getOperand(1), isConstant)))
11451 /// Read and decode a shufflevector mask.
11453 /// It turns undef elements into values that are larger than the number of
11454 /// elements in the input.
11455 static std::vector<unsigned> getShuffleMask(const ShuffleVectorInst *SVI) {
11456 unsigned NElts = SVI->getType()->getNumElements();
11457 if (isa<ConstantAggregateZero>(SVI->getOperand(2)))
11458 return std::vector<unsigned>(NElts, 0);
11459 if (isa<UndefValue>(SVI->getOperand(2)))
11460 return std::vector<unsigned>(NElts, 2*NElts);
11462 std::vector<unsigned> Result;
11463 const ConstantVector *CP = cast<ConstantVector>(SVI->getOperand(2));
11464 for (User::const_op_iterator i = CP->op_begin(), e = CP->op_end(); i!=e; ++i)
11465 if (isa<UndefValue>(*i))
11466 Result.push_back(NElts*2); // undef -> 8
11468 Result.push_back(cast<ConstantInt>(*i)->getZExtValue());
11472 /// FindScalarElement - Given a vector and an element number, see if the scalar
11473 /// value is already around as a register, for example if it were inserted then
11474 /// extracted from the vector.
11475 static Value *FindScalarElement(Value *V, unsigned EltNo) {
11476 assert(isa<VectorType>(V->getType()) && "Not looking at a vector?");
11477 const VectorType *PTy = cast<VectorType>(V->getType());
11478 unsigned Width = PTy->getNumElements();
11479 if (EltNo >= Width) // Out of range access.
11480 return UndefValue::get(PTy->getElementType());
11482 if (isa<UndefValue>(V))
11483 return UndefValue::get(PTy->getElementType());
11484 else if (isa<ConstantAggregateZero>(V))
11485 return Constant::getNullValue(PTy->getElementType());
11486 else if (ConstantVector *CP = dyn_cast<ConstantVector>(V))
11487 return CP->getOperand(EltNo);
11488 else if (InsertElementInst *III = dyn_cast<InsertElementInst>(V)) {
11489 // If this is an insert to a variable element, we don't know what it is.
11490 if (!isa<ConstantInt>(III->getOperand(2)))
11492 unsigned IIElt = cast<ConstantInt>(III->getOperand(2))->getZExtValue();
11494 // If this is an insert to the element we are looking for, return the
11496 if (EltNo == IIElt)
11497 return III->getOperand(1);
11499 // Otherwise, the insertelement doesn't modify the value, recurse on its
11501 return FindScalarElement(III->getOperand(0), EltNo);
11502 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(V)) {
11503 unsigned LHSWidth =
11504 cast<VectorType>(SVI->getOperand(0)->getType())->getNumElements();
11505 unsigned InEl = getShuffleMask(SVI)[EltNo];
11506 if (InEl < LHSWidth)
11507 return FindScalarElement(SVI->getOperand(0), InEl);
11508 else if (InEl < LHSWidth*2)
11509 return FindScalarElement(SVI->getOperand(1), InEl - LHSWidth);
11511 return UndefValue::get(PTy->getElementType());
11514 // Otherwise, we don't know.
11518 Instruction *InstCombiner::visitExtractElementInst(ExtractElementInst &EI) {
11519 // If vector val is undef, replace extract with scalar undef.
11520 if (isa<UndefValue>(EI.getOperand(0)))
11521 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
11523 // If vector val is constant 0, replace extract with scalar 0.
11524 if (isa<ConstantAggregateZero>(EI.getOperand(0)))
11525 return ReplaceInstUsesWith(EI, Constant::getNullValue(EI.getType()));
11527 if (ConstantVector *C = dyn_cast<ConstantVector>(EI.getOperand(0))) {
11528 // If vector val is constant with all elements the same, replace EI with
11529 // that element. When the elements are not identical, we cannot replace yet
11530 // (we do that below, but only when the index is constant).
11531 Constant *op0 = C->getOperand(0);
11532 for (unsigned i = 1; i < C->getNumOperands(); ++i)
11533 if (C->getOperand(i) != op0) {
11538 return ReplaceInstUsesWith(EI, op0);
11541 // If extracting a specified index from the vector, see if we can recursively
11542 // find a previously computed scalar that was inserted into the vector.
11543 if (ConstantInt *IdxC = dyn_cast<ConstantInt>(EI.getOperand(1))) {
11544 unsigned IndexVal = IdxC->getZExtValue();
11545 unsigned VectorWidth =
11546 cast<VectorType>(EI.getOperand(0)->getType())->getNumElements();
11548 // If this is extracting an invalid index, turn this into undef, to avoid
11549 // crashing the code below.
11550 if (IndexVal >= VectorWidth)
11551 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
11553 // This instruction only demands the single element from the input vector.
11554 // If the input vector has a single use, simplify it based on this use
11556 if (EI.getOperand(0)->hasOneUse() && VectorWidth != 1) {
11557 uint64_t UndefElts;
11558 if (Value *V = SimplifyDemandedVectorElts(EI.getOperand(0),
11561 EI.setOperand(0, V);
11566 if (Value *Elt = FindScalarElement(EI.getOperand(0), IndexVal))
11567 return ReplaceInstUsesWith(EI, Elt);
11569 // If the this extractelement is directly using a bitcast from a vector of
11570 // the same number of elements, see if we can find the source element from
11571 // it. In this case, we will end up needing to bitcast the scalars.
11572 if (BitCastInst *BCI = dyn_cast<BitCastInst>(EI.getOperand(0))) {
11573 if (const VectorType *VT =
11574 dyn_cast<VectorType>(BCI->getOperand(0)->getType()))
11575 if (VT->getNumElements() == VectorWidth)
11576 if (Value *Elt = FindScalarElement(BCI->getOperand(0), IndexVal))
11577 return new BitCastInst(Elt, EI.getType());
11581 if (Instruction *I = dyn_cast<Instruction>(EI.getOperand(0))) {
11582 if (I->hasOneUse()) {
11583 // Push extractelement into predecessor operation if legal and
11584 // profitable to do so
11585 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
11586 bool isConstantElt = isa<ConstantInt>(EI.getOperand(1));
11587 if (CheapToScalarize(BO, isConstantElt)) {
11588 ExtractElementInst *newEI0 =
11589 new ExtractElementInst(BO->getOperand(0), EI.getOperand(1),
11590 EI.getName()+".lhs");
11591 ExtractElementInst *newEI1 =
11592 new ExtractElementInst(BO->getOperand(1), EI.getOperand(1),
11593 EI.getName()+".rhs");
11594 InsertNewInstBefore(newEI0, EI);
11595 InsertNewInstBefore(newEI1, EI);
11596 return BinaryOperator::Create(BO->getOpcode(), newEI0, newEI1);
11598 } else if (isa<LoadInst>(I)) {
11600 cast<PointerType>(I->getOperand(0)->getType())->getAddressSpace();
11601 Value *Ptr = InsertBitCastBefore(I->getOperand(0),
11602 PointerType::get(EI.getType(), AS),EI);
11603 GetElementPtrInst *GEP =
11604 GetElementPtrInst::Create(Ptr, EI.getOperand(1), I->getName()+".gep");
11605 InsertNewInstBefore(GEP, EI);
11606 return new LoadInst(GEP);
11609 if (InsertElementInst *IE = dyn_cast<InsertElementInst>(I)) {
11610 // Extracting the inserted element?
11611 if (IE->getOperand(2) == EI.getOperand(1))
11612 return ReplaceInstUsesWith(EI, IE->getOperand(1));
11613 // If the inserted and extracted elements are constants, they must not
11614 // be the same value, extract from the pre-inserted value instead.
11615 if (isa<Constant>(IE->getOperand(2)) &&
11616 isa<Constant>(EI.getOperand(1))) {
11617 AddUsesToWorkList(EI);
11618 EI.setOperand(0, IE->getOperand(0));
11621 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I)) {
11622 // If this is extracting an element from a shufflevector, figure out where
11623 // it came from and extract from the appropriate input element instead.
11624 if (ConstantInt *Elt = dyn_cast<ConstantInt>(EI.getOperand(1))) {
11625 unsigned SrcIdx = getShuffleMask(SVI)[Elt->getZExtValue()];
11627 unsigned LHSWidth =
11628 cast<VectorType>(SVI->getOperand(0)->getType())->getNumElements();
11630 if (SrcIdx < LHSWidth)
11631 Src = SVI->getOperand(0);
11632 else if (SrcIdx < LHSWidth*2) {
11633 SrcIdx -= LHSWidth;
11634 Src = SVI->getOperand(1);
11636 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
11638 return new ExtractElementInst(Src, SrcIdx);
11645 /// CollectSingleShuffleElements - If V is a shuffle of values that ONLY returns
11646 /// elements from either LHS or RHS, return the shuffle mask and true.
11647 /// Otherwise, return false.
11648 static bool CollectSingleShuffleElements(Value *V, Value *LHS, Value *RHS,
11649 std::vector<Constant*> &Mask) {
11650 assert(V->getType() == LHS->getType() && V->getType() == RHS->getType() &&
11651 "Invalid CollectSingleShuffleElements");
11652 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
11654 if (isa<UndefValue>(V)) {
11655 Mask.assign(NumElts, UndefValue::get(Type::Int32Ty));
11657 } else if (V == LHS) {
11658 for (unsigned i = 0; i != NumElts; ++i)
11659 Mask.push_back(ConstantInt::get(Type::Int32Ty, i));
11661 } else if (V == RHS) {
11662 for (unsigned i = 0; i != NumElts; ++i)
11663 Mask.push_back(ConstantInt::get(Type::Int32Ty, i+NumElts));
11665 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
11666 // If this is an insert of an extract from some other vector, include it.
11667 Value *VecOp = IEI->getOperand(0);
11668 Value *ScalarOp = IEI->getOperand(1);
11669 Value *IdxOp = IEI->getOperand(2);
11671 if (!isa<ConstantInt>(IdxOp))
11673 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
11675 if (isa<UndefValue>(ScalarOp)) { // inserting undef into vector.
11676 // Okay, we can handle this if the vector we are insertinting into is
11677 // transitively ok.
11678 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
11679 // If so, update the mask to reflect the inserted undef.
11680 Mask[InsertedIdx] = UndefValue::get(Type::Int32Ty);
11683 } else if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)){
11684 if (isa<ConstantInt>(EI->getOperand(1)) &&
11685 EI->getOperand(0)->getType() == V->getType()) {
11686 unsigned ExtractedIdx =
11687 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
11689 // This must be extracting from either LHS or RHS.
11690 if (EI->getOperand(0) == LHS || EI->getOperand(0) == RHS) {
11691 // Okay, we can handle this if the vector we are insertinting into is
11692 // transitively ok.
11693 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
11694 // If so, update the mask to reflect the inserted value.
11695 if (EI->getOperand(0) == LHS) {
11696 Mask[InsertedIdx % NumElts] =
11697 ConstantInt::get(Type::Int32Ty, ExtractedIdx);
11699 assert(EI->getOperand(0) == RHS);
11700 Mask[InsertedIdx % NumElts] =
11701 ConstantInt::get(Type::Int32Ty, ExtractedIdx+NumElts);
11710 // TODO: Handle shufflevector here!
11715 /// CollectShuffleElements - We are building a shuffle of V, using RHS as the
11716 /// RHS of the shuffle instruction, if it is not null. Return a shuffle mask
11717 /// that computes V and the LHS value of the shuffle.
11718 static Value *CollectShuffleElements(Value *V, std::vector<Constant*> &Mask,
11720 assert(isa<VectorType>(V->getType()) &&
11721 (RHS == 0 || V->getType() == RHS->getType()) &&
11722 "Invalid shuffle!");
11723 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
11725 if (isa<UndefValue>(V)) {
11726 Mask.assign(NumElts, UndefValue::get(Type::Int32Ty));
11728 } else if (isa<ConstantAggregateZero>(V)) {
11729 Mask.assign(NumElts, ConstantInt::get(Type::Int32Ty, 0));
11731 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
11732 // If this is an insert of an extract from some other vector, include it.
11733 Value *VecOp = IEI->getOperand(0);
11734 Value *ScalarOp = IEI->getOperand(1);
11735 Value *IdxOp = IEI->getOperand(2);
11737 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
11738 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
11739 EI->getOperand(0)->getType() == V->getType()) {
11740 unsigned ExtractedIdx =
11741 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
11742 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
11744 // Either the extracted from or inserted into vector must be RHSVec,
11745 // otherwise we'd end up with a shuffle of three inputs.
11746 if (EI->getOperand(0) == RHS || RHS == 0) {
11747 RHS = EI->getOperand(0);
11748 Value *V = CollectShuffleElements(VecOp, Mask, RHS);
11749 Mask[InsertedIdx % NumElts] =
11750 ConstantInt::get(Type::Int32Ty, NumElts+ExtractedIdx);
11754 if (VecOp == RHS) {
11755 Value *V = CollectShuffleElements(EI->getOperand(0), Mask, RHS);
11756 // Everything but the extracted element is replaced with the RHS.
11757 for (unsigned i = 0; i != NumElts; ++i) {
11758 if (i != InsertedIdx)
11759 Mask[i] = ConstantInt::get(Type::Int32Ty, NumElts+i);
11764 // If this insertelement is a chain that comes from exactly these two
11765 // vectors, return the vector and the effective shuffle.
11766 if (CollectSingleShuffleElements(IEI, EI->getOperand(0), RHS, Mask))
11767 return EI->getOperand(0);
11772 // TODO: Handle shufflevector here!
11774 // Otherwise, can't do anything fancy. Return an identity vector.
11775 for (unsigned i = 0; i != NumElts; ++i)
11776 Mask.push_back(ConstantInt::get(Type::Int32Ty, i));
11780 Instruction *InstCombiner::visitInsertElementInst(InsertElementInst &IE) {
11781 Value *VecOp = IE.getOperand(0);
11782 Value *ScalarOp = IE.getOperand(1);
11783 Value *IdxOp = IE.getOperand(2);
11785 // Inserting an undef or into an undefined place, remove this.
11786 if (isa<UndefValue>(ScalarOp) || isa<UndefValue>(IdxOp))
11787 ReplaceInstUsesWith(IE, VecOp);
11789 // If the inserted element was extracted from some other vector, and if the
11790 // indexes are constant, try to turn this into a shufflevector operation.
11791 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
11792 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
11793 EI->getOperand(0)->getType() == IE.getType()) {
11794 unsigned NumVectorElts = IE.getType()->getNumElements();
11795 unsigned ExtractedIdx =
11796 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
11797 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
11799 if (ExtractedIdx >= NumVectorElts) // Out of range extract.
11800 return ReplaceInstUsesWith(IE, VecOp);
11802 if (InsertedIdx >= NumVectorElts) // Out of range insert.
11803 return ReplaceInstUsesWith(IE, UndefValue::get(IE.getType()));
11805 // If we are extracting a value from a vector, then inserting it right
11806 // back into the same place, just use the input vector.
11807 if (EI->getOperand(0) == VecOp && ExtractedIdx == InsertedIdx)
11808 return ReplaceInstUsesWith(IE, VecOp);
11810 // We could theoretically do this for ANY input. However, doing so could
11811 // turn chains of insertelement instructions into a chain of shufflevector
11812 // instructions, and right now we do not merge shufflevectors. As such,
11813 // only do this in a situation where it is clear that there is benefit.
11814 if (isa<UndefValue>(VecOp) || isa<ConstantAggregateZero>(VecOp)) {
11815 // Turn this into shuffle(EIOp0, VecOp, Mask). The result has all of
11816 // the values of VecOp, except then one read from EIOp0.
11817 // Build a new shuffle mask.
11818 std::vector<Constant*> Mask;
11819 if (isa<UndefValue>(VecOp))
11820 Mask.assign(NumVectorElts, UndefValue::get(Type::Int32Ty));
11822 assert(isa<ConstantAggregateZero>(VecOp) && "Unknown thing");
11823 Mask.assign(NumVectorElts, ConstantInt::get(Type::Int32Ty,
11826 Mask[InsertedIdx] = ConstantInt::get(Type::Int32Ty, ExtractedIdx);
11827 return new ShuffleVectorInst(EI->getOperand(0), VecOp,
11828 ConstantVector::get(Mask));
11831 // If this insertelement isn't used by some other insertelement, turn it
11832 // (and any insertelements it points to), into one big shuffle.
11833 if (!IE.hasOneUse() || !isa<InsertElementInst>(IE.use_back())) {
11834 std::vector<Constant*> Mask;
11836 Value *LHS = CollectShuffleElements(&IE, Mask, RHS);
11837 if (RHS == 0) RHS = UndefValue::get(LHS->getType());
11838 // We now have a shuffle of LHS, RHS, Mask.
11839 return new ShuffleVectorInst(LHS, RHS, ConstantVector::get(Mask));
11848 Instruction *InstCombiner::visitShuffleVectorInst(ShuffleVectorInst &SVI) {
11849 Value *LHS = SVI.getOperand(0);
11850 Value *RHS = SVI.getOperand(1);
11851 std::vector<unsigned> Mask = getShuffleMask(&SVI);
11853 bool MadeChange = false;
11855 // Undefined shuffle mask -> undefined value.
11856 if (isa<UndefValue>(SVI.getOperand(2)))
11857 return ReplaceInstUsesWith(SVI, UndefValue::get(SVI.getType()));
11859 uint64_t UndefElts;
11860 unsigned VWidth = cast<VectorType>(SVI.getType())->getNumElements();
11862 if (VWidth != cast<VectorType>(LHS->getType())->getNumElements())
11865 uint64_t AllOnesEltMask = ~0ULL >> (64-VWidth);
11866 if (VWidth <= 64 &&
11867 SimplifyDemandedVectorElts(&SVI, AllOnesEltMask, UndefElts)) {
11868 LHS = SVI.getOperand(0);
11869 RHS = SVI.getOperand(1);
11873 // Canonicalize shuffle(x ,x,mask) -> shuffle(x, undef,mask')
11874 // Canonicalize shuffle(undef,x,mask) -> shuffle(x, undef,mask').
11875 if (LHS == RHS || isa<UndefValue>(LHS)) {
11876 if (isa<UndefValue>(LHS) && LHS == RHS) {
11877 // shuffle(undef,undef,mask) -> undef.
11878 return ReplaceInstUsesWith(SVI, LHS);
11881 // Remap any references to RHS to use LHS.
11882 std::vector<Constant*> Elts;
11883 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
11884 if (Mask[i] >= 2*e)
11885 Elts.push_back(UndefValue::get(Type::Int32Ty));
11887 if ((Mask[i] >= e && isa<UndefValue>(RHS)) ||
11888 (Mask[i] < e && isa<UndefValue>(LHS))) {
11889 Mask[i] = 2*e; // Turn into undef.
11890 Elts.push_back(UndefValue::get(Type::Int32Ty));
11892 Mask[i] = Mask[i] % e; // Force to LHS.
11893 Elts.push_back(ConstantInt::get(Type::Int32Ty, Mask[i]));
11897 SVI.setOperand(0, SVI.getOperand(1));
11898 SVI.setOperand(1, UndefValue::get(RHS->getType()));
11899 SVI.setOperand(2, ConstantVector::get(Elts));
11900 LHS = SVI.getOperand(0);
11901 RHS = SVI.getOperand(1);
11905 // Analyze the shuffle, are the LHS or RHS and identity shuffles?
11906 bool isLHSID = true, isRHSID = true;
11908 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
11909 if (Mask[i] >= e*2) continue; // Ignore undef values.
11910 // Is this an identity shuffle of the LHS value?
11911 isLHSID &= (Mask[i] == i);
11913 // Is this an identity shuffle of the RHS value?
11914 isRHSID &= (Mask[i]-e == i);
11917 // Eliminate identity shuffles.
11918 if (isLHSID) return ReplaceInstUsesWith(SVI, LHS);
11919 if (isRHSID) return ReplaceInstUsesWith(SVI, RHS);
11921 // If the LHS is a shufflevector itself, see if we can combine it with this
11922 // one without producing an unusual shuffle. Here we are really conservative:
11923 // we are absolutely afraid of producing a shuffle mask not in the input
11924 // program, because the code gen may not be smart enough to turn a merged
11925 // shuffle into two specific shuffles: it may produce worse code. As such,
11926 // we only merge two shuffles if the result is one of the two input shuffle
11927 // masks. In this case, merging the shuffles just removes one instruction,
11928 // which we know is safe. This is good for things like turning:
11929 // (splat(splat)) -> splat.
11930 if (ShuffleVectorInst *LHSSVI = dyn_cast<ShuffleVectorInst>(LHS)) {
11931 if (isa<UndefValue>(RHS)) {
11932 std::vector<unsigned> LHSMask = getShuffleMask(LHSSVI);
11934 std::vector<unsigned> NewMask;
11935 for (unsigned i = 0, e = Mask.size(); i != e; ++i)
11936 if (Mask[i] >= 2*e)
11937 NewMask.push_back(2*e);
11939 NewMask.push_back(LHSMask[Mask[i]]);
11941 // If the result mask is equal to the src shuffle or this shuffle mask, do
11942 // the replacement.
11943 if (NewMask == LHSMask || NewMask == Mask) {
11944 std::vector<Constant*> Elts;
11945 for (unsigned i = 0, e = NewMask.size(); i != e; ++i) {
11946 if (NewMask[i] >= e*2) {
11947 Elts.push_back(UndefValue::get(Type::Int32Ty));
11949 Elts.push_back(ConstantInt::get(Type::Int32Ty, NewMask[i]));
11952 return new ShuffleVectorInst(LHSSVI->getOperand(0),
11953 LHSSVI->getOperand(1),
11954 ConstantVector::get(Elts));
11959 return MadeChange ? &SVI : 0;
11965 /// TryToSinkInstruction - Try to move the specified instruction from its
11966 /// current block into the beginning of DestBlock, which can only happen if it's
11967 /// safe to move the instruction past all of the instructions between it and the
11968 /// end of its block.
11969 static bool TryToSinkInstruction(Instruction *I, BasicBlock *DestBlock) {
11970 assert(I->hasOneUse() && "Invariants didn't hold!");
11972 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
11973 if (isa<PHINode>(I) || I->mayWriteToMemory() || isa<TerminatorInst>(I))
11976 // Do not sink alloca instructions out of the entry block.
11977 if (isa<AllocaInst>(I) && I->getParent() ==
11978 &DestBlock->getParent()->getEntryBlock())
11981 // We can only sink load instructions if there is nothing between the load and
11982 // the end of block that could change the value.
11983 if (I->mayReadFromMemory()) {
11984 for (BasicBlock::iterator Scan = I, E = I->getParent()->end();
11986 if (Scan->mayWriteToMemory())
11990 BasicBlock::iterator InsertPos = DestBlock->getFirstNonPHI();
11992 I->moveBefore(InsertPos);
11998 /// AddReachableCodeToWorklist - Walk the function in depth-first order, adding
11999 /// all reachable code to the worklist.
12001 /// This has a couple of tricks to make the code faster and more powerful. In
12002 /// particular, we constant fold and DCE instructions as we go, to avoid adding
12003 /// them to the worklist (this significantly speeds up instcombine on code where
12004 /// many instructions are dead or constant). Additionally, if we find a branch
12005 /// whose condition is a known constant, we only visit the reachable successors.
12007 static void AddReachableCodeToWorklist(BasicBlock *BB,
12008 SmallPtrSet<BasicBlock*, 64> &Visited,
12010 const TargetData *TD) {
12011 SmallVector<BasicBlock*, 256> Worklist;
12012 Worklist.push_back(BB);
12014 while (!Worklist.empty()) {
12015 BB = Worklist.back();
12016 Worklist.pop_back();
12018 // We have now visited this block! If we've already been here, ignore it.
12019 if (!Visited.insert(BB)) continue;
12021 DbgInfoIntrinsic *DBI_Prev = NULL;
12022 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) {
12023 Instruction *Inst = BBI++;
12025 // DCE instruction if trivially dead.
12026 if (isInstructionTriviallyDead(Inst)) {
12028 DOUT << "IC: DCE: " << *Inst;
12029 Inst->eraseFromParent();
12033 // ConstantProp instruction if trivially constant.
12034 if (Constant *C = ConstantFoldInstruction(Inst, TD)) {
12035 DOUT << "IC: ConstFold to: " << *C << " from: " << *Inst;
12036 Inst->replaceAllUsesWith(C);
12038 Inst->eraseFromParent();
12042 // If there are two consecutive llvm.dbg.stoppoint calls then
12043 // it is likely that the optimizer deleted code in between these
12045 DbgInfoIntrinsic *DBI_Next = dyn_cast<DbgInfoIntrinsic>(Inst);
12048 && DBI_Prev->getIntrinsicID() == llvm::Intrinsic::dbg_stoppoint
12049 && DBI_Next->getIntrinsicID() == llvm::Intrinsic::dbg_stoppoint) {
12050 IC.RemoveFromWorkList(DBI_Prev);
12051 DBI_Prev->eraseFromParent();
12053 DBI_Prev = DBI_Next;
12056 IC.AddToWorkList(Inst);
12059 // Recursively visit successors. If this is a branch or switch on a
12060 // constant, only visit the reachable successor.
12061 TerminatorInst *TI = BB->getTerminator();
12062 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
12063 if (BI->isConditional() && isa<ConstantInt>(BI->getCondition())) {
12064 bool CondVal = cast<ConstantInt>(BI->getCondition())->getZExtValue();
12065 BasicBlock *ReachableBB = BI->getSuccessor(!CondVal);
12066 Worklist.push_back(ReachableBB);
12069 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
12070 if (ConstantInt *Cond = dyn_cast<ConstantInt>(SI->getCondition())) {
12071 // See if this is an explicit destination.
12072 for (unsigned i = 1, e = SI->getNumSuccessors(); i != e; ++i)
12073 if (SI->getCaseValue(i) == Cond) {
12074 BasicBlock *ReachableBB = SI->getSuccessor(i);
12075 Worklist.push_back(ReachableBB);
12079 // Otherwise it is the default destination.
12080 Worklist.push_back(SI->getSuccessor(0));
12085 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
12086 Worklist.push_back(TI->getSuccessor(i));
12090 bool InstCombiner::DoOneIteration(Function &F, unsigned Iteration) {
12091 bool Changed = false;
12092 TD = &getAnalysis<TargetData>();
12094 DEBUG(DOUT << "\n\nINSTCOMBINE ITERATION #" << Iteration << " on "
12095 << F.getNameStr() << "\n");
12098 // Do a depth-first traversal of the function, populate the worklist with
12099 // the reachable instructions. Ignore blocks that are not reachable. Keep
12100 // track of which blocks we visit.
12101 SmallPtrSet<BasicBlock*, 64> Visited;
12102 AddReachableCodeToWorklist(F.begin(), Visited, *this, TD);
12104 // Do a quick scan over the function. If we find any blocks that are
12105 // unreachable, remove any instructions inside of them. This prevents
12106 // the instcombine code from having to deal with some bad special cases.
12107 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
12108 if (!Visited.count(BB)) {
12109 Instruction *Term = BB->getTerminator();
12110 while (Term != BB->begin()) { // Remove instrs bottom-up
12111 BasicBlock::iterator I = Term; --I;
12113 DOUT << "IC: DCE: " << *I;
12116 if (!I->use_empty())
12117 I->replaceAllUsesWith(UndefValue::get(I->getType()));
12118 I->eraseFromParent();
12123 while (!Worklist.empty()) {
12124 Instruction *I = RemoveOneFromWorkList();
12125 if (I == 0) continue; // skip null values.
12127 // Check to see if we can DCE the instruction.
12128 if (isInstructionTriviallyDead(I)) {
12129 // Add operands to the worklist.
12130 if (I->getNumOperands() < 4)
12131 AddUsesToWorkList(*I);
12134 DOUT << "IC: DCE: " << *I;
12136 I->eraseFromParent();
12137 RemoveFromWorkList(I);
12141 // Instruction isn't dead, see if we can constant propagate it.
12142 if (Constant *C = ConstantFoldInstruction(I, TD)) {
12143 DOUT << "IC: ConstFold to: " << *C << " from: " << *I;
12145 // Add operands to the worklist.
12146 AddUsesToWorkList(*I);
12147 ReplaceInstUsesWith(*I, C);
12150 I->eraseFromParent();
12151 RemoveFromWorkList(I);
12155 if (TD && I->getType()->getTypeID() == Type::VoidTyID) {
12156 // See if we can constant fold its operands.
12157 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i) {
12158 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(i)) {
12159 if (Constant *NewC = ConstantFoldConstantExpression(CE, TD))
12165 // See if we can trivially sink this instruction to a successor basic block.
12166 if (I->hasOneUse()) {
12167 BasicBlock *BB = I->getParent();
12168 BasicBlock *UserParent = cast<Instruction>(I->use_back())->getParent();
12169 if (UserParent != BB) {
12170 bool UserIsSuccessor = false;
12171 // See if the user is one of our successors.
12172 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI)
12173 if (*SI == UserParent) {
12174 UserIsSuccessor = true;
12178 // If the user is one of our immediate successors, and if that successor
12179 // only has us as a predecessors (we'd have to split the critical edge
12180 // otherwise), we can keep going.
12181 if (UserIsSuccessor && !isa<PHINode>(I->use_back()) &&
12182 next(pred_begin(UserParent)) == pred_end(UserParent))
12183 // Okay, the CFG is simple enough, try to sink this instruction.
12184 Changed |= TryToSinkInstruction(I, UserParent);
12188 // Now that we have an instruction, try combining it to simplify it...
12192 DEBUG(std::ostringstream SS; I->print(SS); OrigI = SS.str(););
12193 if (Instruction *Result = visit(*I)) {
12195 // Should we replace the old instruction with a new one?
12197 DOUT << "IC: Old = " << *I
12198 << " New = " << *Result;
12200 // Everything uses the new instruction now.
12201 I->replaceAllUsesWith(Result);
12203 // Push the new instruction and any users onto the worklist.
12204 AddToWorkList(Result);
12205 AddUsersToWorkList(*Result);
12207 // Move the name to the new instruction first.
12208 Result->takeName(I);
12210 // Insert the new instruction into the basic block...
12211 BasicBlock *InstParent = I->getParent();
12212 BasicBlock::iterator InsertPos = I;
12214 if (!isa<PHINode>(Result)) // If combining a PHI, don't insert
12215 while (isa<PHINode>(InsertPos)) // middle of a block of PHIs.
12218 InstParent->getInstList().insert(InsertPos, Result);
12220 // Make sure that we reprocess all operands now that we reduced their
12222 AddUsesToWorkList(*I);
12224 // Instructions can end up on the worklist more than once. Make sure
12225 // we do not process an instruction that has been deleted.
12226 RemoveFromWorkList(I);
12228 // Erase the old instruction.
12229 InstParent->getInstList().erase(I);
12232 DOUT << "IC: Mod = " << OrigI
12233 << " New = " << *I;
12236 // If the instruction was modified, it's possible that it is now dead.
12237 // if so, remove it.
12238 if (isInstructionTriviallyDead(I)) {
12239 // Make sure we process all operands now that we are reducing their
12241 AddUsesToWorkList(*I);
12243 // Instructions may end up in the worklist more than once. Erase all
12244 // occurrences of this instruction.
12245 RemoveFromWorkList(I);
12246 I->eraseFromParent();
12249 AddUsersToWorkList(*I);
12256 assert(WorklistMap.empty() && "Worklist empty, but map not?");
12258 // Do an explicit clear, this shrinks the map if needed.
12259 WorklistMap.clear();
12264 bool InstCombiner::runOnFunction(Function &F) {
12265 MustPreserveLCSSA = mustPreserveAnalysisID(LCSSAID);
12267 bool EverMadeChange = false;
12269 // Iterate while there is work to do.
12270 unsigned Iteration = 0;
12271 while (DoOneIteration(F, Iteration++))
12272 EverMadeChange = true;
12273 return EverMadeChange;
12276 FunctionPass *llvm::createInstructionCombiningPass() {
12277 return new InstCombiner();