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->getValue());
2962 APInt NegOne = -APInt(RHSNeg->getBitWidth(), 1, true);
2963 APInt TwoToExp(RHSNeg->getBitWidth(), 1 << (RHSNeg->getBitWidth() - 1));
2965 // -X/C -> X/-C, if and only if negation doesn't overflow.
2966 if ((RHS->getValue().isNegative() && RHSNegAPI.slt(TwoToExp - 1)) ||
2967 (RHS->getValue().isNonNegative() && RHSNegAPI.sgt(TwoToExp * NegOne))) {
2968 if (Value *LHSNeg = dyn_castNegVal(Op0)) {
2969 if (ConstantInt *CI = dyn_cast<ConstantInt>(LHSNeg)) {
2970 ConstantInt *CINeg = cast<ConstantInt>(ConstantExpr::getNeg(CI));
2971 APInt CINegAPI(CINeg->getValue());
2973 if ((CI->getValue().isNegative() && CINegAPI.slt(TwoToExp - 1)) ||
2974 (CI->getValue().isNonNegative() && CINegAPI.sgt(TwoToExp*NegOne)))
2975 return BinaryOperator::CreateSDiv(LHSNeg,
2976 ConstantExpr::getNeg(RHS));
2982 // If the sign bits of both operands are zero (i.e. we can prove they are
2983 // unsigned inputs), turn this into a udiv.
2984 if (I.getType()->isInteger()) {
2985 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()));
2986 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
2987 // X sdiv Y -> X udiv Y, iff X and Y don't have sign bit set
2988 return BinaryOperator::CreateUDiv(Op0, Op1, I.getName());
2995 Instruction *InstCombiner::visitFDiv(BinaryOperator &I) {
2996 return commonDivTransforms(I);
2999 /// This function implements the transforms on rem instructions that work
3000 /// regardless of the kind of rem instruction it is (urem, srem, or frem). It
3001 /// is used by the visitors to those instructions.
3002 /// @brief Transforms common to all three rem instructions
3003 Instruction *InstCombiner::commonRemTransforms(BinaryOperator &I) {
3004 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3006 // 0 % X == 0 for integer, we don't need to preserve faults!
3007 if (Constant *LHS = dyn_cast<Constant>(Op0))
3008 if (LHS->isNullValue())
3009 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3011 if (isa<UndefValue>(Op0)) { // undef % X -> 0
3012 if (I.getType()->isFPOrFPVector())
3013 return ReplaceInstUsesWith(I, Op0); // X % undef -> undef (could be SNaN)
3014 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3016 if (isa<UndefValue>(Op1))
3017 return ReplaceInstUsesWith(I, Op1); // X % undef -> undef
3019 // Handle cases involving: rem X, (select Cond, Y, Z)
3020 if (isa<SelectInst>(Op1) && SimplifyDivRemOfSelect(I))
3026 /// This function implements the transforms common to both integer remainder
3027 /// instructions (urem and srem). It is called by the visitors to those integer
3028 /// remainder instructions.
3029 /// @brief Common integer remainder transforms
3030 Instruction *InstCombiner::commonIRemTransforms(BinaryOperator &I) {
3031 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3033 if (Instruction *common = commonRemTransforms(I))
3036 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3037 // X % 0 == undef, we don't need to preserve faults!
3038 if (RHS->equalsInt(0))
3039 return ReplaceInstUsesWith(I, UndefValue::get(I.getType()));
3041 if (RHS->equalsInt(1)) // X % 1 == 0
3042 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3044 if (Instruction *Op0I = dyn_cast<Instruction>(Op0)) {
3045 if (SelectInst *SI = dyn_cast<SelectInst>(Op0I)) {
3046 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3048 } else if (isa<PHINode>(Op0I)) {
3049 if (Instruction *NV = FoldOpIntoPhi(I))
3053 // See if we can fold away this rem instruction.
3054 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
3055 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
3056 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
3057 KnownZero, KnownOne))
3065 Instruction *InstCombiner::visitURem(BinaryOperator &I) {
3066 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3068 if (Instruction *common = commonIRemTransforms(I))
3071 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3072 // X urem C^2 -> X and C
3073 // Check to see if this is an unsigned remainder with an exact power of 2,
3074 // if so, convert to a bitwise and.
3075 if (ConstantInt *C = dyn_cast<ConstantInt>(RHS))
3076 if (C->getValue().isPowerOf2())
3077 return BinaryOperator::CreateAnd(Op0, SubOne(C));
3080 if (Instruction *RHSI = dyn_cast<Instruction>(I.getOperand(1))) {
3081 // Turn A % (C << N), where C is 2^k, into A & ((C << N)-1)
3082 if (RHSI->getOpcode() == Instruction::Shl &&
3083 isa<ConstantInt>(RHSI->getOperand(0))) {
3084 if (cast<ConstantInt>(RHSI->getOperand(0))->getValue().isPowerOf2()) {
3085 Constant *N1 = ConstantInt::getAllOnesValue(I.getType());
3086 Value *Add = InsertNewInstBefore(BinaryOperator::CreateAdd(RHSI, N1,
3088 return BinaryOperator::CreateAnd(Op0, Add);
3093 // urem X, (select Cond, 2^C1, 2^C2) --> select Cond, (and X, C1), (and X, C2)
3094 // where C1&C2 are powers of two.
3095 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
3096 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
3097 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
3098 // STO == 0 and SFO == 0 handled above.
3099 if ((STO->getValue().isPowerOf2()) &&
3100 (SFO->getValue().isPowerOf2())) {
3101 Value *TrueAnd = InsertNewInstBefore(
3102 BinaryOperator::CreateAnd(Op0, SubOne(STO), SI->getName()+".t"), I);
3103 Value *FalseAnd = InsertNewInstBefore(
3104 BinaryOperator::CreateAnd(Op0, SubOne(SFO), SI->getName()+".f"), I);
3105 return SelectInst::Create(SI->getOperand(0), TrueAnd, FalseAnd);
3113 Instruction *InstCombiner::visitSRem(BinaryOperator &I) {
3114 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3116 // Handle the integer rem common cases
3117 if (Instruction *common = commonIRemTransforms(I))
3120 if (Value *RHSNeg = dyn_castNegVal(Op1))
3121 if (!isa<Constant>(RHSNeg) ||
3122 (isa<ConstantInt>(RHSNeg) &&
3123 cast<ConstantInt>(RHSNeg)->getValue().isStrictlyPositive())) {
3125 AddUsesToWorkList(I);
3126 I.setOperand(1, RHSNeg);
3130 // If the sign bits of both operands are zero (i.e. we can prove they are
3131 // unsigned inputs), turn this into a urem.
3132 if (I.getType()->isInteger()) {
3133 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()));
3134 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
3135 // X srem Y -> X urem Y, iff X and Y don't have sign bit set
3136 return BinaryOperator::CreateURem(Op0, Op1, I.getName());
3143 Instruction *InstCombiner::visitFRem(BinaryOperator &I) {
3144 return commonRemTransforms(I);
3147 // isOneBitSet - Return true if there is exactly one bit set in the specified
3149 static bool isOneBitSet(const ConstantInt *CI) {
3150 return CI->getValue().isPowerOf2();
3153 // isHighOnes - Return true if the constant is of the form 1+0+.
3154 // This is the same as lowones(~X).
3155 static bool isHighOnes(const ConstantInt *CI) {
3156 return (~CI->getValue() + 1).isPowerOf2();
3159 /// getICmpCode - Encode a icmp predicate into a three bit mask. These bits
3160 /// are carefully arranged to allow folding of expressions such as:
3162 /// (A < B) | (A > B) --> (A != B)
3164 /// Note that this is only valid if the first and second predicates have the
3165 /// same sign. Is illegal to do: (A u< B) | (A s> B)
3167 /// Three bits are used to represent the condition, as follows:
3172 /// <=> Value Definition
3173 /// 000 0 Always false
3180 /// 111 7 Always true
3182 static unsigned getICmpCode(const ICmpInst *ICI) {
3183 switch (ICI->getPredicate()) {
3185 case ICmpInst::ICMP_UGT: return 1; // 001
3186 case ICmpInst::ICMP_SGT: return 1; // 001
3187 case ICmpInst::ICMP_EQ: return 2; // 010
3188 case ICmpInst::ICMP_UGE: return 3; // 011
3189 case ICmpInst::ICMP_SGE: return 3; // 011
3190 case ICmpInst::ICMP_ULT: return 4; // 100
3191 case ICmpInst::ICMP_SLT: return 4; // 100
3192 case ICmpInst::ICMP_NE: return 5; // 101
3193 case ICmpInst::ICMP_ULE: return 6; // 110
3194 case ICmpInst::ICMP_SLE: return 6; // 110
3197 assert(0 && "Invalid ICmp predicate!");
3202 /// getFCmpCode - Similar to getICmpCode but for FCmpInst. This encodes a fcmp
3203 /// predicate into a three bit mask. It also returns whether it is an ordered
3204 /// predicate by reference.
3205 static unsigned getFCmpCode(FCmpInst::Predicate CC, bool &isOrdered) {
3208 case FCmpInst::FCMP_ORD: isOrdered = true; return 0; // 000
3209 case FCmpInst::FCMP_UNO: return 0; // 000
3210 case FCmpInst::FCMP_OGT: isOrdered = true; return 1; // 001
3211 case FCmpInst::FCMP_UGT: return 1; // 001
3212 case FCmpInst::FCMP_OEQ: isOrdered = true; return 2; // 010
3213 case FCmpInst::FCMP_UEQ: return 2; // 010
3214 case FCmpInst::FCMP_OGE: isOrdered = true; return 3; // 011
3215 case FCmpInst::FCMP_UGE: return 3; // 011
3216 case FCmpInst::FCMP_OLT: isOrdered = true; return 4; // 100
3217 case FCmpInst::FCMP_ULT: return 4; // 100
3218 case FCmpInst::FCMP_ONE: isOrdered = true; return 5; // 101
3219 case FCmpInst::FCMP_UNE: return 5; // 101
3220 case FCmpInst::FCMP_OLE: isOrdered = true; return 6; // 110
3221 case FCmpInst::FCMP_ULE: return 6; // 110
3224 // Not expecting FCMP_FALSE and FCMP_TRUE;
3225 assert(0 && "Unexpected FCmp predicate!");
3230 /// getICmpValue - This is the complement of getICmpCode, which turns an
3231 /// opcode and two operands into either a constant true or false, or a brand
3232 /// new ICmp instruction. The sign is passed in to determine which kind
3233 /// of predicate to use in the new icmp instruction.
3234 static Value *getICmpValue(bool sign, unsigned code, Value *LHS, Value *RHS) {
3236 default: assert(0 && "Illegal ICmp code!");
3237 case 0: return ConstantInt::getFalse();
3240 return new ICmpInst(ICmpInst::ICMP_SGT, LHS, RHS);
3242 return new ICmpInst(ICmpInst::ICMP_UGT, LHS, RHS);
3243 case 2: return new ICmpInst(ICmpInst::ICMP_EQ, LHS, RHS);
3246 return new ICmpInst(ICmpInst::ICMP_SGE, LHS, RHS);
3248 return new ICmpInst(ICmpInst::ICMP_UGE, LHS, RHS);
3251 return new ICmpInst(ICmpInst::ICMP_SLT, LHS, RHS);
3253 return new ICmpInst(ICmpInst::ICMP_ULT, LHS, RHS);
3254 case 5: return new ICmpInst(ICmpInst::ICMP_NE, LHS, RHS);
3257 return new ICmpInst(ICmpInst::ICMP_SLE, LHS, RHS);
3259 return new ICmpInst(ICmpInst::ICMP_ULE, LHS, RHS);
3260 case 7: return ConstantInt::getTrue();
3264 /// getFCmpValue - This is the complement of getFCmpCode, which turns an
3265 /// opcode and two operands into either a FCmp instruction. isordered is passed
3266 /// in to determine which kind of predicate to use in the new fcmp instruction.
3267 static Value *getFCmpValue(bool isordered, unsigned code,
3268 Value *LHS, Value *RHS) {
3270 default: assert(0 && "Illegal FCmp code!");
3273 return new FCmpInst(FCmpInst::FCMP_ORD, LHS, RHS);
3275 return new FCmpInst(FCmpInst::FCMP_UNO, LHS, RHS);
3278 return new FCmpInst(FCmpInst::FCMP_OGT, LHS, RHS);
3280 return new FCmpInst(FCmpInst::FCMP_UGT, LHS, RHS);
3283 return new FCmpInst(FCmpInst::FCMP_OEQ, LHS, RHS);
3285 return new FCmpInst(FCmpInst::FCMP_UEQ, LHS, RHS);
3288 return new FCmpInst(FCmpInst::FCMP_OGE, LHS, RHS);
3290 return new FCmpInst(FCmpInst::FCMP_UGE, LHS, RHS);
3293 return new FCmpInst(FCmpInst::FCMP_OLT, LHS, RHS);
3295 return new FCmpInst(FCmpInst::FCMP_ULT, LHS, RHS);
3298 return new FCmpInst(FCmpInst::FCMP_ONE, LHS, RHS);
3300 return new FCmpInst(FCmpInst::FCMP_UNE, LHS, RHS);
3303 return new FCmpInst(FCmpInst::FCMP_OLE, LHS, RHS);
3305 return new FCmpInst(FCmpInst::FCMP_ULE, LHS, RHS);
3306 case 7: return ConstantInt::getTrue();
3310 /// PredicatesFoldable - Return true if both predicates match sign or if at
3311 /// least one of them is an equality comparison (which is signless).
3312 static bool PredicatesFoldable(ICmpInst::Predicate p1, ICmpInst::Predicate p2) {
3313 return (ICmpInst::isSignedPredicate(p1) == ICmpInst::isSignedPredicate(p2)) ||
3314 (ICmpInst::isSignedPredicate(p1) && ICmpInst::isEquality(p2)) ||
3315 (ICmpInst::isSignedPredicate(p2) && ICmpInst::isEquality(p1));
3319 // FoldICmpLogical - Implements (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
3320 struct FoldICmpLogical {
3323 ICmpInst::Predicate pred;
3324 FoldICmpLogical(InstCombiner &ic, ICmpInst *ICI)
3325 : IC(ic), LHS(ICI->getOperand(0)), RHS(ICI->getOperand(1)),
3326 pred(ICI->getPredicate()) {}
3327 bool shouldApply(Value *V) const {
3328 if (ICmpInst *ICI = dyn_cast<ICmpInst>(V))
3329 if (PredicatesFoldable(pred, ICI->getPredicate()))
3330 return ((ICI->getOperand(0) == LHS && ICI->getOperand(1) == RHS) ||
3331 (ICI->getOperand(0) == RHS && ICI->getOperand(1) == LHS));
3334 Instruction *apply(Instruction &Log) const {
3335 ICmpInst *ICI = cast<ICmpInst>(Log.getOperand(0));
3336 if (ICI->getOperand(0) != LHS) {
3337 assert(ICI->getOperand(1) == LHS);
3338 ICI->swapOperands(); // Swap the LHS and RHS of the ICmp
3341 ICmpInst *RHSICI = cast<ICmpInst>(Log.getOperand(1));
3342 unsigned LHSCode = getICmpCode(ICI);
3343 unsigned RHSCode = getICmpCode(RHSICI);
3345 switch (Log.getOpcode()) {
3346 case Instruction::And: Code = LHSCode & RHSCode; break;
3347 case Instruction::Or: Code = LHSCode | RHSCode; break;
3348 case Instruction::Xor: Code = LHSCode ^ RHSCode; break;
3349 default: assert(0 && "Illegal logical opcode!"); return 0;
3352 bool isSigned = ICmpInst::isSignedPredicate(RHSICI->getPredicate()) ||
3353 ICmpInst::isSignedPredicate(ICI->getPredicate());
3355 Value *RV = getICmpValue(isSigned, Code, LHS, RHS);
3356 if (Instruction *I = dyn_cast<Instruction>(RV))
3358 // Otherwise, it's a constant boolean value...
3359 return IC.ReplaceInstUsesWith(Log, RV);
3362 } // end anonymous namespace
3364 // OptAndOp - This handles expressions of the form ((val OP C1) & C2). Where
3365 // the Op parameter is 'OP', OpRHS is 'C1', and AndRHS is 'C2'. Op is
3366 // guaranteed to be a binary operator.
3367 Instruction *InstCombiner::OptAndOp(Instruction *Op,
3369 ConstantInt *AndRHS,
3370 BinaryOperator &TheAnd) {
3371 Value *X = Op->getOperand(0);
3372 Constant *Together = 0;
3374 Together = And(AndRHS, OpRHS);
3376 switch (Op->getOpcode()) {
3377 case Instruction::Xor:
3378 if (Op->hasOneUse()) {
3379 // (X ^ C1) & C2 --> (X & C2) ^ (C1&C2)
3380 Instruction *And = BinaryOperator::CreateAnd(X, AndRHS);
3381 InsertNewInstBefore(And, TheAnd);
3383 return BinaryOperator::CreateXor(And, Together);
3386 case Instruction::Or:
3387 if (Together == AndRHS) // (X | C) & C --> C
3388 return ReplaceInstUsesWith(TheAnd, AndRHS);
3390 if (Op->hasOneUse() && Together != OpRHS) {
3391 // (X | C1) & C2 --> (X | (C1&C2)) & C2
3392 Instruction *Or = BinaryOperator::CreateOr(X, Together);
3393 InsertNewInstBefore(Or, TheAnd);
3395 return BinaryOperator::CreateAnd(Or, AndRHS);
3398 case Instruction::Add:
3399 if (Op->hasOneUse()) {
3400 // Adding a one to a single bit bit-field should be turned into an XOR
3401 // of the bit. First thing to check is to see if this AND is with a
3402 // single bit constant.
3403 const APInt& AndRHSV = cast<ConstantInt>(AndRHS)->getValue();
3405 // If there is only one bit set...
3406 if (isOneBitSet(cast<ConstantInt>(AndRHS))) {
3407 // Ok, at this point, we know that we are masking the result of the
3408 // ADD down to exactly one bit. If the constant we are adding has
3409 // no bits set below this bit, then we can eliminate the ADD.
3410 const APInt& AddRHS = cast<ConstantInt>(OpRHS)->getValue();
3412 // Check to see if any bits below the one bit set in AndRHSV are set.
3413 if ((AddRHS & (AndRHSV-1)) == 0) {
3414 // If not, the only thing that can effect the output of the AND is
3415 // the bit specified by AndRHSV. If that bit is set, the effect of
3416 // the XOR is to toggle the bit. If it is clear, then the ADD has
3418 if ((AddRHS & AndRHSV) == 0) { // Bit is not set, noop
3419 TheAnd.setOperand(0, X);
3422 // Pull the XOR out of the AND.
3423 Instruction *NewAnd = BinaryOperator::CreateAnd(X, AndRHS);
3424 InsertNewInstBefore(NewAnd, TheAnd);
3425 NewAnd->takeName(Op);
3426 return BinaryOperator::CreateXor(NewAnd, AndRHS);
3433 case Instruction::Shl: {
3434 // We know that the AND will not produce any of the bits shifted in, so if
3435 // the anded constant includes them, clear them now!
3437 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3438 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3439 APInt ShlMask(APInt::getHighBitsSet(BitWidth, BitWidth-OpRHSVal));
3440 ConstantInt *CI = ConstantInt::get(AndRHS->getValue() & ShlMask);
3442 if (CI->getValue() == ShlMask) {
3443 // Masking out bits that the shift already masks
3444 return ReplaceInstUsesWith(TheAnd, Op); // No need for the and.
3445 } else if (CI != AndRHS) { // Reducing bits set in and.
3446 TheAnd.setOperand(1, CI);
3451 case Instruction::LShr:
3453 // We know that the AND will not produce any of the bits shifted in, so if
3454 // the anded constant includes them, clear them now! This only applies to
3455 // unsigned shifts, because a signed shr may bring in set bits!
3457 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3458 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3459 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
3460 ConstantInt *CI = ConstantInt::get(AndRHS->getValue() & ShrMask);
3462 if (CI->getValue() == ShrMask) {
3463 // Masking out bits that the shift already masks.
3464 return ReplaceInstUsesWith(TheAnd, Op);
3465 } else if (CI != AndRHS) {
3466 TheAnd.setOperand(1, CI); // Reduce bits set in and cst.
3471 case Instruction::AShr:
3473 // See if this is shifting in some sign extension, then masking it out
3475 if (Op->hasOneUse()) {
3476 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3477 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3478 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
3479 Constant *C = ConstantInt::get(AndRHS->getValue() & ShrMask);
3480 if (C == AndRHS) { // Masking out bits shifted in.
3481 // (Val ashr C1) & C2 -> (Val lshr C1) & C2
3482 // Make the argument unsigned.
3483 Value *ShVal = Op->getOperand(0);
3484 ShVal = InsertNewInstBefore(
3485 BinaryOperator::CreateLShr(ShVal, OpRHS,
3486 Op->getName()), TheAnd);
3487 return BinaryOperator::CreateAnd(ShVal, AndRHS, TheAnd.getName());
3496 /// InsertRangeTest - Emit a computation of: (V >= Lo && V < Hi) if Inside is
3497 /// true, otherwise (V < Lo || V >= Hi). In pratice, we emit the more efficient
3498 /// (V-Lo) <u Hi-Lo. This method expects that Lo <= Hi. isSigned indicates
3499 /// whether to treat the V, Lo and HI as signed or not. IB is the location to
3500 /// insert new instructions.
3501 Instruction *InstCombiner::InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
3502 bool isSigned, bool Inside,
3504 assert(cast<ConstantInt>(ConstantExpr::getICmp((isSigned ?
3505 ICmpInst::ICMP_SLE:ICmpInst::ICMP_ULE), Lo, Hi))->getZExtValue() &&
3506 "Lo is not <= Hi in range emission code!");
3509 if (Lo == Hi) // Trivially false.
3510 return new ICmpInst(ICmpInst::ICMP_NE, V, V);
3512 // V >= Min && V < Hi --> V < Hi
3513 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
3514 ICmpInst::Predicate pred = (isSigned ?
3515 ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT);
3516 return new ICmpInst(pred, V, Hi);
3519 // Emit V-Lo <u Hi-Lo
3520 Constant *NegLo = ConstantExpr::getNeg(Lo);
3521 Instruction *Add = BinaryOperator::CreateAdd(V, NegLo, V->getName()+".off");
3522 InsertNewInstBefore(Add, IB);
3523 Constant *UpperBound = ConstantExpr::getAdd(NegLo, Hi);
3524 return new ICmpInst(ICmpInst::ICMP_ULT, Add, UpperBound);
3527 if (Lo == Hi) // Trivially true.
3528 return new ICmpInst(ICmpInst::ICMP_EQ, V, V);
3530 // V < Min || V >= Hi -> V > Hi-1
3531 Hi = SubOne(cast<ConstantInt>(Hi));
3532 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
3533 ICmpInst::Predicate pred = (isSigned ?
3534 ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT);
3535 return new ICmpInst(pred, V, Hi);
3538 // Emit V-Lo >u Hi-1-Lo
3539 // Note that Hi has already had one subtracted from it, above.
3540 ConstantInt *NegLo = cast<ConstantInt>(ConstantExpr::getNeg(Lo));
3541 Instruction *Add = BinaryOperator::CreateAdd(V, NegLo, V->getName()+".off");
3542 InsertNewInstBefore(Add, IB);
3543 Constant *LowerBound = ConstantExpr::getAdd(NegLo, Hi);
3544 return new ICmpInst(ICmpInst::ICMP_UGT, Add, LowerBound);
3547 // isRunOfOnes - Returns true iff Val consists of one contiguous run of 1s with
3548 // any number of 0s on either side. The 1s are allowed to wrap from LSB to
3549 // MSB, so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs. 0x0F0F0000 is
3550 // not, since all 1s are not contiguous.
3551 static bool isRunOfOnes(ConstantInt *Val, uint32_t &MB, uint32_t &ME) {
3552 const APInt& V = Val->getValue();
3553 uint32_t BitWidth = Val->getType()->getBitWidth();
3554 if (!APIntOps::isShiftedMask(BitWidth, V)) return false;
3556 // look for the first zero bit after the run of ones
3557 MB = BitWidth - ((V - 1) ^ V).countLeadingZeros();
3558 // look for the first non-zero bit
3559 ME = V.getActiveBits();
3563 /// FoldLogicalPlusAnd - This is part of an expression (LHS +/- RHS) & Mask,
3564 /// where isSub determines whether the operator is a sub. If we can fold one of
3565 /// the following xforms:
3567 /// ((A & N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == Mask
3568 /// ((A | N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3569 /// ((A ^ N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3571 /// return (A +/- B).
3573 Value *InstCombiner::FoldLogicalPlusAnd(Value *LHS, Value *RHS,
3574 ConstantInt *Mask, bool isSub,
3576 Instruction *LHSI = dyn_cast<Instruction>(LHS);
3577 if (!LHSI || LHSI->getNumOperands() != 2 ||
3578 !isa<ConstantInt>(LHSI->getOperand(1))) return 0;
3580 ConstantInt *N = cast<ConstantInt>(LHSI->getOperand(1));
3582 switch (LHSI->getOpcode()) {
3584 case Instruction::And:
3585 if (And(N, Mask) == Mask) {
3586 // If the AndRHS is a power of two minus one (0+1+), this is simple.
3587 if ((Mask->getValue().countLeadingZeros() +
3588 Mask->getValue().countPopulation()) ==
3589 Mask->getValue().getBitWidth())
3592 // Otherwise, if Mask is 0+1+0+, and if B is known to have the low 0+
3593 // part, we don't need any explicit masks to take them out of A. If that
3594 // is all N is, ignore it.
3595 uint32_t MB = 0, ME = 0;
3596 if (isRunOfOnes(Mask, MB, ME)) { // begin/end bit of run, inclusive
3597 uint32_t BitWidth = cast<IntegerType>(RHS->getType())->getBitWidth();
3598 APInt Mask(APInt::getLowBitsSet(BitWidth, MB-1));
3599 if (MaskedValueIsZero(RHS, Mask))
3604 case Instruction::Or:
3605 case Instruction::Xor:
3606 // If the AndRHS is a power of two minus one (0+1+), and N&Mask == 0
3607 if ((Mask->getValue().countLeadingZeros() +
3608 Mask->getValue().countPopulation()) == Mask->getValue().getBitWidth()
3609 && And(N, Mask)->isZero())
3616 New = BinaryOperator::CreateSub(LHSI->getOperand(0), RHS, "fold");
3618 New = BinaryOperator::CreateAdd(LHSI->getOperand(0), RHS, "fold");
3619 return InsertNewInstBefore(New, I);
3622 /// FoldAndOfICmps - Fold (icmp)&(icmp) if possible.
3623 Instruction *InstCombiner::FoldAndOfICmps(Instruction &I,
3624 ICmpInst *LHS, ICmpInst *RHS) {
3626 ConstantInt *LHSCst, *RHSCst;
3627 ICmpInst::Predicate LHSCC, RHSCC;
3629 // This only handles icmp of constants: (icmp1 A, C1) & (icmp2 B, C2).
3630 if (!match(LHS, m_ICmp(LHSCC, m_Value(Val), m_ConstantInt(LHSCst))) ||
3631 !match(RHS, m_ICmp(RHSCC, m_Value(Val2), m_ConstantInt(RHSCst))))
3634 // (icmp ult A, C) & (icmp ult B, C) --> (icmp ult (A|B), C)
3635 // where C is a power of 2
3636 if (LHSCst == RHSCst && LHSCC == RHSCC && LHSCC == ICmpInst::ICMP_ULT &&
3637 LHSCst->getValue().isPowerOf2()) {
3638 Instruction *NewOr = BinaryOperator::CreateOr(Val, Val2);
3639 InsertNewInstBefore(NewOr, I);
3640 return new ICmpInst(LHSCC, NewOr, LHSCst);
3643 // From here on, we only handle:
3644 // (icmp1 A, C1) & (icmp2 A, C2) --> something simpler.
3645 if (Val != Val2) return 0;
3647 // ICMP_[US][GL]E X, CST is folded to ICMP_[US][GL]T elsewhere.
3648 if (LHSCC == ICmpInst::ICMP_UGE || LHSCC == ICmpInst::ICMP_ULE ||
3649 RHSCC == ICmpInst::ICMP_UGE || RHSCC == ICmpInst::ICMP_ULE ||
3650 LHSCC == ICmpInst::ICMP_SGE || LHSCC == ICmpInst::ICMP_SLE ||
3651 RHSCC == ICmpInst::ICMP_SGE || RHSCC == ICmpInst::ICMP_SLE)
3654 // We can't fold (ugt x, C) & (sgt x, C2).
3655 if (!PredicatesFoldable(LHSCC, RHSCC))
3658 // Ensure that the larger constant is on the RHS.
3660 if (ICmpInst::isSignedPredicate(LHSCC) ||
3661 (ICmpInst::isEquality(LHSCC) &&
3662 ICmpInst::isSignedPredicate(RHSCC)))
3663 ShouldSwap = LHSCst->getValue().sgt(RHSCst->getValue());
3665 ShouldSwap = LHSCst->getValue().ugt(RHSCst->getValue());
3668 std::swap(LHS, RHS);
3669 std::swap(LHSCst, RHSCst);
3670 std::swap(LHSCC, RHSCC);
3673 // At this point, we know we have have two icmp instructions
3674 // comparing a value against two constants and and'ing the result
3675 // together. Because of the above check, we know that we only have
3676 // icmp eq, icmp ne, icmp [su]lt, and icmp [SU]gt here. We also know
3677 // (from the FoldICmpLogical check above), that the two constants
3678 // are not equal and that the larger constant is on the RHS
3679 assert(LHSCst != RHSCst && "Compares not folded above?");
3682 default: assert(0 && "Unknown integer condition code!");
3683 case ICmpInst::ICMP_EQ:
3685 default: assert(0 && "Unknown integer condition code!");
3686 case ICmpInst::ICMP_EQ: // (X == 13 & X == 15) -> false
3687 case ICmpInst::ICMP_UGT: // (X == 13 & X > 15) -> false
3688 case ICmpInst::ICMP_SGT: // (X == 13 & X > 15) -> false
3689 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3690 case ICmpInst::ICMP_NE: // (X == 13 & X != 15) -> X == 13
3691 case ICmpInst::ICMP_ULT: // (X == 13 & X < 15) -> X == 13
3692 case ICmpInst::ICMP_SLT: // (X == 13 & X < 15) -> X == 13
3693 return ReplaceInstUsesWith(I, LHS);
3695 case ICmpInst::ICMP_NE:
3697 default: assert(0 && "Unknown integer condition code!");
3698 case ICmpInst::ICMP_ULT:
3699 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X u< 14) -> X < 13
3700 return new ICmpInst(ICmpInst::ICMP_ULT, Val, LHSCst);
3701 break; // (X != 13 & X u< 15) -> no change
3702 case ICmpInst::ICMP_SLT:
3703 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X s< 14) -> X < 13
3704 return new ICmpInst(ICmpInst::ICMP_SLT, Val, LHSCst);
3705 break; // (X != 13 & X s< 15) -> no change
3706 case ICmpInst::ICMP_EQ: // (X != 13 & X == 15) -> X == 15
3707 case ICmpInst::ICMP_UGT: // (X != 13 & X u> 15) -> X u> 15
3708 case ICmpInst::ICMP_SGT: // (X != 13 & X s> 15) -> X s> 15
3709 return ReplaceInstUsesWith(I, RHS);
3710 case ICmpInst::ICMP_NE:
3711 if (LHSCst == SubOne(RHSCst)){// (X != 13 & X != 14) -> X-13 >u 1
3712 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
3713 Instruction *Add = BinaryOperator::CreateAdd(Val, AddCST,
3714 Val->getName()+".off");
3715 InsertNewInstBefore(Add, I);
3716 return new ICmpInst(ICmpInst::ICMP_UGT, Add,
3717 ConstantInt::get(Add->getType(), 1));
3719 break; // (X != 13 & X != 15) -> no change
3722 case ICmpInst::ICMP_ULT:
3724 default: assert(0 && "Unknown integer condition code!");
3725 case ICmpInst::ICMP_EQ: // (X u< 13 & X == 15) -> false
3726 case ICmpInst::ICMP_UGT: // (X u< 13 & X u> 15) -> false
3727 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3728 case ICmpInst::ICMP_SGT: // (X u< 13 & X s> 15) -> no change
3730 case ICmpInst::ICMP_NE: // (X u< 13 & X != 15) -> X u< 13
3731 case ICmpInst::ICMP_ULT: // (X u< 13 & X u< 15) -> X u< 13
3732 return ReplaceInstUsesWith(I, LHS);
3733 case ICmpInst::ICMP_SLT: // (X u< 13 & X s< 15) -> no change
3737 case ICmpInst::ICMP_SLT:
3739 default: assert(0 && "Unknown integer condition code!");
3740 case ICmpInst::ICMP_EQ: // (X s< 13 & X == 15) -> false
3741 case ICmpInst::ICMP_SGT: // (X s< 13 & X s> 15) -> false
3742 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3743 case ICmpInst::ICMP_UGT: // (X s< 13 & X u> 15) -> no change
3745 case ICmpInst::ICMP_NE: // (X s< 13 & X != 15) -> X < 13
3746 case ICmpInst::ICMP_SLT: // (X s< 13 & X s< 15) -> X < 13
3747 return ReplaceInstUsesWith(I, LHS);
3748 case ICmpInst::ICMP_ULT: // (X s< 13 & X u< 15) -> no change
3752 case ICmpInst::ICMP_UGT:
3754 default: assert(0 && "Unknown integer condition code!");
3755 case ICmpInst::ICMP_EQ: // (X u> 13 & X == 15) -> X == 15
3756 case ICmpInst::ICMP_UGT: // (X u> 13 & X u> 15) -> X u> 15
3757 return ReplaceInstUsesWith(I, RHS);
3758 case ICmpInst::ICMP_SGT: // (X u> 13 & X s> 15) -> no change
3760 case ICmpInst::ICMP_NE:
3761 if (RHSCst == AddOne(LHSCst)) // (X u> 13 & X != 14) -> X u> 14
3762 return new ICmpInst(LHSCC, Val, RHSCst);
3763 break; // (X u> 13 & X != 15) -> no change
3764 case ICmpInst::ICMP_ULT: // (X u> 13 & X u< 15) -> (X-14) <u 1
3765 return InsertRangeTest(Val, AddOne(LHSCst), RHSCst, false, true, I);
3766 case ICmpInst::ICMP_SLT: // (X u> 13 & X s< 15) -> no change
3770 case ICmpInst::ICMP_SGT:
3772 default: assert(0 && "Unknown integer condition code!");
3773 case ICmpInst::ICMP_EQ: // (X s> 13 & X == 15) -> X == 15
3774 case ICmpInst::ICMP_SGT: // (X s> 13 & X s> 15) -> X s> 15
3775 return ReplaceInstUsesWith(I, RHS);
3776 case ICmpInst::ICMP_UGT: // (X s> 13 & X u> 15) -> no change
3778 case ICmpInst::ICMP_NE:
3779 if (RHSCst == AddOne(LHSCst)) // (X s> 13 & X != 14) -> X s> 14
3780 return new ICmpInst(LHSCC, Val, RHSCst);
3781 break; // (X s> 13 & X != 15) -> no change
3782 case ICmpInst::ICMP_SLT: // (X s> 13 & X s< 15) -> (X-14) s< 1
3783 return InsertRangeTest(Val, AddOne(LHSCst), RHSCst, true, true, I);
3784 case ICmpInst::ICMP_ULT: // (X s> 13 & X u< 15) -> no change
3794 Instruction *InstCombiner::visitAnd(BinaryOperator &I) {
3795 bool Changed = SimplifyCommutative(I);
3796 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3798 if (isa<UndefValue>(Op1)) // X & undef -> 0
3799 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3803 return ReplaceInstUsesWith(I, Op1);
3805 // See if we can simplify any instructions used by the instruction whose sole
3806 // purpose is to compute bits we don't care about.
3807 if (!isa<VectorType>(I.getType())) {
3808 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
3809 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
3810 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
3811 KnownZero, KnownOne))
3814 if (ConstantVector *CP = dyn_cast<ConstantVector>(Op1)) {
3815 if (CP->isAllOnesValue()) // X & <-1,-1> -> X
3816 return ReplaceInstUsesWith(I, I.getOperand(0));
3817 } else if (isa<ConstantAggregateZero>(Op1)) {
3818 return ReplaceInstUsesWith(I, Op1); // X & <0,0> -> <0,0>
3822 if (ConstantInt *AndRHS = dyn_cast<ConstantInt>(Op1)) {
3823 const APInt& AndRHSMask = AndRHS->getValue();
3824 APInt NotAndRHS(~AndRHSMask);
3826 // Optimize a variety of ((val OP C1) & C2) combinations...
3827 if (isa<BinaryOperator>(Op0)) {
3828 Instruction *Op0I = cast<Instruction>(Op0);
3829 Value *Op0LHS = Op0I->getOperand(0);
3830 Value *Op0RHS = Op0I->getOperand(1);
3831 switch (Op0I->getOpcode()) {
3832 case Instruction::Xor:
3833 case Instruction::Or:
3834 // If the mask is only needed on one incoming arm, push it up.
3835 if (Op0I->hasOneUse()) {
3836 if (MaskedValueIsZero(Op0LHS, NotAndRHS)) {
3837 // Not masking anything out for the LHS, move to RHS.
3838 Instruction *NewRHS = BinaryOperator::CreateAnd(Op0RHS, AndRHS,
3839 Op0RHS->getName()+".masked");
3840 InsertNewInstBefore(NewRHS, I);
3841 return BinaryOperator::Create(
3842 cast<BinaryOperator>(Op0I)->getOpcode(), Op0LHS, NewRHS);
3844 if (!isa<Constant>(Op0RHS) &&
3845 MaskedValueIsZero(Op0RHS, NotAndRHS)) {
3846 // Not masking anything out for the RHS, move to LHS.
3847 Instruction *NewLHS = BinaryOperator::CreateAnd(Op0LHS, AndRHS,
3848 Op0LHS->getName()+".masked");
3849 InsertNewInstBefore(NewLHS, I);
3850 return BinaryOperator::Create(
3851 cast<BinaryOperator>(Op0I)->getOpcode(), NewLHS, Op0RHS);
3856 case Instruction::Add:
3857 // ((A & N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == AndRHS.
3858 // ((A | N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
3859 // ((A ^ N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
3860 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, false, I))
3861 return BinaryOperator::CreateAnd(V, AndRHS);
3862 if (Value *V = FoldLogicalPlusAnd(Op0RHS, Op0LHS, AndRHS, false, I))
3863 return BinaryOperator::CreateAnd(V, AndRHS); // Add commutes
3866 case Instruction::Sub:
3867 // ((A & N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == AndRHS.
3868 // ((A | N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
3869 // ((A ^ N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
3870 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, true, I))
3871 return BinaryOperator::CreateAnd(V, AndRHS);
3873 // (A - N) & AndRHS -> -N & AndRHS iff A&AndRHS==0 and AndRHS
3874 // has 1's for all bits that the subtraction with A might affect.
3875 if (Op0I->hasOneUse()) {
3876 uint32_t BitWidth = AndRHSMask.getBitWidth();
3877 uint32_t Zeros = AndRHSMask.countLeadingZeros();
3878 APInt Mask = APInt::getLowBitsSet(BitWidth, BitWidth - Zeros);
3880 ConstantInt *A = dyn_cast<ConstantInt>(Op0LHS);
3881 if (!(A && A->isZero()) && // avoid infinite recursion.
3882 MaskedValueIsZero(Op0LHS, Mask)) {
3883 Instruction *NewNeg = BinaryOperator::CreateNeg(Op0RHS);
3884 InsertNewInstBefore(NewNeg, I);
3885 return BinaryOperator::CreateAnd(NewNeg, AndRHS);
3890 case Instruction::Shl:
3891 case Instruction::LShr:
3892 // (1 << x) & 1 --> zext(x == 0)
3893 // (1 >> x) & 1 --> zext(x == 0)
3894 if (AndRHSMask == 1 && Op0LHS == AndRHS) {
3895 Instruction *NewICmp = new ICmpInst(ICmpInst::ICMP_EQ, Op0RHS,
3896 Constant::getNullValue(I.getType()));
3897 InsertNewInstBefore(NewICmp, I);
3898 return new ZExtInst(NewICmp, I.getType());
3903 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
3904 if (Instruction *Res = OptAndOp(Op0I, Op0CI, AndRHS, I))
3906 } else if (CastInst *CI = dyn_cast<CastInst>(Op0)) {
3907 // If this is an integer truncation or change from signed-to-unsigned, and
3908 // if the source is an and/or with immediate, transform it. This
3909 // frequently occurs for bitfield accesses.
3910 if (Instruction *CastOp = dyn_cast<Instruction>(CI->getOperand(0))) {
3911 if ((isa<TruncInst>(CI) || isa<BitCastInst>(CI)) &&
3912 CastOp->getNumOperands() == 2)
3913 if (ConstantInt *AndCI = dyn_cast<ConstantInt>(CastOp->getOperand(1))) {
3914 if (CastOp->getOpcode() == Instruction::And) {
3915 // Change: and (cast (and X, C1) to T), C2
3916 // into : and (cast X to T), trunc_or_bitcast(C1)&C2
3917 // This will fold the two constants together, which may allow
3918 // other simplifications.
3919 Instruction *NewCast = CastInst::CreateTruncOrBitCast(
3920 CastOp->getOperand(0), I.getType(),
3921 CastOp->getName()+".shrunk");
3922 NewCast = InsertNewInstBefore(NewCast, I);
3923 // trunc_or_bitcast(C1)&C2
3924 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
3925 C3 = ConstantExpr::getAnd(C3, AndRHS);
3926 return BinaryOperator::CreateAnd(NewCast, C3);
3927 } else if (CastOp->getOpcode() == Instruction::Or) {
3928 // Change: and (cast (or X, C1) to T), C2
3929 // into : trunc(C1)&C2 iff trunc(C1)&C2 == C2
3930 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
3931 if (ConstantExpr::getAnd(C3, AndRHS) == AndRHS) // trunc(C1)&C2
3932 return ReplaceInstUsesWith(I, AndRHS);
3938 // Try to fold constant and into select arguments.
3939 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3940 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3942 if (isa<PHINode>(Op0))
3943 if (Instruction *NV = FoldOpIntoPhi(I))
3947 Value *Op0NotVal = dyn_castNotVal(Op0);
3948 Value *Op1NotVal = dyn_castNotVal(Op1);
3950 if (Op0NotVal == Op1 || Op1NotVal == Op0) // A & ~A == ~A & A == 0
3951 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3953 // (~A & ~B) == (~(A | B)) - De Morgan's Law
3954 if (Op0NotVal && Op1NotVal && isOnlyUse(Op0) && isOnlyUse(Op1)) {
3955 Instruction *Or = BinaryOperator::CreateOr(Op0NotVal, Op1NotVal,
3956 I.getName()+".demorgan");
3957 InsertNewInstBefore(Or, I);
3958 return BinaryOperator::CreateNot(Or);
3962 Value *A = 0, *B = 0, *C = 0, *D = 0;
3963 if (match(Op0, m_Or(m_Value(A), m_Value(B)))) {
3964 if (A == Op1 || B == Op1) // (A | ?) & A --> A
3965 return ReplaceInstUsesWith(I, Op1);
3967 // (A|B) & ~(A&B) -> A^B
3968 if (match(Op1, m_Not(m_And(m_Value(C), m_Value(D))))) {
3969 if ((A == C && B == D) || (A == D && B == C))
3970 return BinaryOperator::CreateXor(A, B);
3974 if (match(Op1, m_Or(m_Value(A), m_Value(B)))) {
3975 if (A == Op0 || B == Op0) // A & (A | ?) --> A
3976 return ReplaceInstUsesWith(I, Op0);
3978 // ~(A&B) & (A|B) -> A^B
3979 if (match(Op0, m_Not(m_And(m_Value(C), m_Value(D))))) {
3980 if ((A == C && B == D) || (A == D && B == C))
3981 return BinaryOperator::CreateXor(A, B);
3985 if (Op0->hasOneUse() &&
3986 match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
3987 if (A == Op1) { // (A^B)&A -> A&(A^B)
3988 I.swapOperands(); // Simplify below
3989 std::swap(Op0, Op1);
3990 } else if (B == Op1) { // (A^B)&B -> B&(B^A)
3991 cast<BinaryOperator>(Op0)->swapOperands();
3992 I.swapOperands(); // Simplify below
3993 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);
4010 // (A&((~A)|B)) -> A&B
4011 if (match(Op0, m_Or(m_Not(m_Value(A)), m_Value(B)))) {
4013 return BinaryOperator::CreateAnd(A, B);
4015 if (match(Op0, m_Or(m_Value(A), m_Not(m_Value(B))))) {
4017 return BinaryOperator::CreateAnd(A, B);
4019 if (match(Op1, m_Or(m_Not(m_Value(A)), m_Value(B)))) {
4021 return BinaryOperator::CreateAnd(A, B);
4023 if (match(Op1, m_Or(m_Value(A), m_Not(m_Value(B))))) {
4025 return BinaryOperator::CreateAnd(A, B);
4029 if (ICmpInst *RHS = dyn_cast<ICmpInst>(Op1)) {
4030 // (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
4031 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
4034 if (ICmpInst *LHS = dyn_cast<ICmpInst>(Op0))
4035 if (Instruction *Res = FoldAndOfICmps(I, LHS, RHS))
4039 // fold (and (cast A), (cast B)) -> (cast (and A, B))
4040 if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
4041 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4042 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind ?
4043 const Type *SrcTy = Op0C->getOperand(0)->getType();
4044 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
4045 // Only do this if the casts both really cause code to be generated.
4046 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4048 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4050 Instruction *NewOp = BinaryOperator::CreateAnd(Op0C->getOperand(0),
4051 Op1C->getOperand(0),
4053 InsertNewInstBefore(NewOp, I);
4054 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
4058 // (X >> Z) & (Y >> Z) -> (X&Y) >> Z for all shifts.
4059 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
4060 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
4061 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
4062 SI0->getOperand(1) == SI1->getOperand(1) &&
4063 (SI0->hasOneUse() || SI1->hasOneUse())) {
4064 Instruction *NewOp =
4065 InsertNewInstBefore(BinaryOperator::CreateAnd(SI0->getOperand(0),
4067 SI0->getName()), I);
4068 return BinaryOperator::Create(SI1->getOpcode(), NewOp,
4069 SI1->getOperand(1));
4073 // If and'ing two fcmp, try combine them into one.
4074 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0))) {
4075 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1))) {
4076 if (LHS->getPredicate() == FCmpInst::FCMP_ORD &&
4077 RHS->getPredicate() == FCmpInst::FCMP_ORD) {
4078 // (fcmp ord x, c) & (fcmp ord y, c) -> (fcmp ord x, y)
4079 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
4080 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
4081 // If either of the constants are nans, then the whole thing returns
4083 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
4084 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4085 return new FCmpInst(FCmpInst::FCMP_ORD, LHS->getOperand(0),
4086 RHS->getOperand(0));
4089 Value *Op0LHS, *Op0RHS, *Op1LHS, *Op1RHS;
4090 FCmpInst::Predicate Op0CC, Op1CC;
4091 if (match(Op0, m_FCmp(Op0CC, m_Value(Op0LHS), m_Value(Op0RHS))) &&
4092 match(Op1, m_FCmp(Op1CC, m_Value(Op1LHS), m_Value(Op1RHS)))) {
4093 if (Op0LHS == Op1RHS && Op0RHS == Op1LHS) {
4094 // Swap RHS operands to match LHS.
4095 Op1CC = FCmpInst::getSwappedPredicate(Op1CC);
4096 std::swap(Op1LHS, Op1RHS);
4098 if (Op0LHS == Op1LHS && Op0RHS == Op1RHS) {
4099 // Simplify (fcmp cc0 x, y) & (fcmp cc1 x, y).
4101 return new FCmpInst((FCmpInst::Predicate)Op0CC, Op0LHS, Op0RHS);
4102 else if (Op0CC == FCmpInst::FCMP_FALSE ||
4103 Op1CC == FCmpInst::FCMP_FALSE)
4104 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4105 else if (Op0CC == FCmpInst::FCMP_TRUE)
4106 return ReplaceInstUsesWith(I, Op1);
4107 else if (Op1CC == FCmpInst::FCMP_TRUE)
4108 return ReplaceInstUsesWith(I, Op0);
4111 unsigned Op0Pred = getFCmpCode(Op0CC, Op0Ordered);
4112 unsigned Op1Pred = getFCmpCode(Op1CC, Op1Ordered);
4114 std::swap(Op0, Op1);
4115 std::swap(Op0Pred, Op1Pred);
4116 std::swap(Op0Ordered, Op1Ordered);
4119 // uno && ueq -> uno && (uno || eq) -> ueq
4120 // ord && olt -> ord && (ord && lt) -> olt
4121 if (Op0Ordered == Op1Ordered)
4122 return ReplaceInstUsesWith(I, Op1);
4123 // uno && oeq -> uno && (ord && eq) -> false
4124 // uno && ord -> false
4126 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4127 // ord && ueq -> ord && (uno || eq) -> oeq
4128 return cast<Instruction>(getFCmpValue(true, Op1Pred,
4137 return Changed ? &I : 0;
4140 /// CollectBSwapParts - Analyze the specified subexpression and see if it is
4141 /// capable of providing pieces of a bswap. The subexpression provides pieces
4142 /// of a bswap if it is proven that each of the non-zero bytes in the output of
4143 /// the expression came from the corresponding "byte swapped" byte in some other
4144 /// value. For example, if the current subexpression is "(shl i32 %X, 24)" then
4145 /// we know that the expression deposits the low byte of %X into the high byte
4146 /// of the bswap result and that all other bytes are zero. This expression is
4147 /// accepted, the high byte of ByteValues is set to X to indicate a correct
4150 /// This function returns true if the match was unsuccessful and false if so.
4151 /// On entry to the function the "OverallLeftShift" is a signed integer value
4152 /// indicating the number of bytes that the subexpression is later shifted. For
4153 /// example, if the expression is later right shifted by 16 bits, the
4154 /// OverallLeftShift value would be -2 on entry. This is used to specify which
4155 /// byte of ByteValues is actually being set.
4157 /// Similarly, ByteMask is a bitmask where a bit is clear if its corresponding
4158 /// byte is masked to zero by a user. For example, in (X & 255), X will be
4159 /// processed with a bytemask of 1. Because bytemask is 32-bits, this limits
4160 /// this function to working on up to 32-byte (256 bit) values. ByteMask is
4161 /// always in the local (OverallLeftShift) coordinate space.
4163 static bool CollectBSwapParts(Value *V, int OverallLeftShift, uint32_t ByteMask,
4164 SmallVector<Value*, 8> &ByteValues) {
4165 if (Instruction *I = dyn_cast<Instruction>(V)) {
4166 // If this is an or instruction, it may be an inner node of the bswap.
4167 if (I->getOpcode() == Instruction::Or) {
4168 return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
4170 CollectBSwapParts(I->getOperand(1), OverallLeftShift, ByteMask,
4174 // If this is a logical shift by a constant multiple of 8, recurse with
4175 // OverallLeftShift and ByteMask adjusted.
4176 if (I->isLogicalShift() && isa<ConstantInt>(I->getOperand(1))) {
4178 cast<ConstantInt>(I->getOperand(1))->getLimitedValue(~0U);
4179 // Ensure the shift amount is defined and of a byte value.
4180 if ((ShAmt & 7) || (ShAmt > 8*ByteValues.size()))
4183 unsigned ByteShift = ShAmt >> 3;
4184 if (I->getOpcode() == Instruction::Shl) {
4185 // X << 2 -> collect(X, +2)
4186 OverallLeftShift += ByteShift;
4187 ByteMask >>= ByteShift;
4189 // X >>u 2 -> collect(X, -2)
4190 OverallLeftShift -= ByteShift;
4191 ByteMask <<= ByteShift;
4192 ByteMask &= (~0U >> (32-ByteValues.size()));
4195 if (OverallLeftShift >= (int)ByteValues.size()) return true;
4196 if (OverallLeftShift <= -(int)ByteValues.size()) return true;
4198 return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
4202 // If this is a logical 'and' with a mask that clears bytes, clear the
4203 // corresponding bytes in ByteMask.
4204 if (I->getOpcode() == Instruction::And &&
4205 isa<ConstantInt>(I->getOperand(1))) {
4206 // Scan every byte of the and mask, seeing if the byte is either 0 or 255.
4207 unsigned NumBytes = ByteValues.size();
4208 APInt Byte(I->getType()->getPrimitiveSizeInBits(), 255);
4209 const APInt &AndMask = cast<ConstantInt>(I->getOperand(1))->getValue();
4211 for (unsigned i = 0; i != NumBytes; ++i, Byte <<= 8) {
4212 // If this byte is masked out by a later operation, we don't care what
4214 if ((ByteMask & (1 << i)) == 0)
4217 // If the AndMask is all zeros for this byte, clear the bit.
4218 APInt MaskB = AndMask & Byte;
4220 ByteMask &= ~(1U << i);
4224 // If the AndMask is not all ones for this byte, it's not a bytezap.
4228 // Otherwise, this byte is kept.
4231 return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
4236 // Okay, we got to something that isn't a shift, 'or' or 'and'. This must be
4237 // the input value to the bswap. Some observations: 1) if more than one byte
4238 // is demanded from this input, then it could not be successfully assembled
4239 // into a byteswap. At least one of the two bytes would not be aligned with
4240 // their ultimate destination.
4241 if (!isPowerOf2_32(ByteMask)) return true;
4242 unsigned InputByteNo = CountTrailingZeros_32(ByteMask);
4244 // 2) The input and ultimate destinations must line up: if byte 3 of an i32
4245 // is demanded, it needs to go into byte 0 of the result. This means that the
4246 // byte needs to be shifted until it lands in the right byte bucket. The
4247 // shift amount depends on the position: if the byte is coming from the high
4248 // part of the value (e.g. byte 3) then it must be shifted right. If from the
4249 // low part, it must be shifted left.
4250 unsigned DestByteNo = InputByteNo + OverallLeftShift;
4251 if (InputByteNo < ByteValues.size()/2) {
4252 if (ByteValues.size()-1-DestByteNo != InputByteNo)
4255 if (ByteValues.size()-1-DestByteNo != InputByteNo)
4259 // If the destination byte value is already defined, the values are or'd
4260 // together, which isn't a bswap (unless it's an or of the same bits).
4261 if (ByteValues[DestByteNo] && ByteValues[DestByteNo] != V)
4263 ByteValues[DestByteNo] = V;
4267 /// MatchBSwap - Given an OR instruction, check to see if this is a bswap idiom.
4268 /// If so, insert the new bswap intrinsic and return it.
4269 Instruction *InstCombiner::MatchBSwap(BinaryOperator &I) {
4270 const IntegerType *ITy = dyn_cast<IntegerType>(I.getType());
4271 if (!ITy || ITy->getBitWidth() % 16 ||
4272 // ByteMask only allows up to 32-byte values.
4273 ITy->getBitWidth() > 32*8)
4274 return 0; // Can only bswap pairs of bytes. Can't do vectors.
4276 /// ByteValues - For each byte of the result, we keep track of which value
4277 /// defines each byte.
4278 SmallVector<Value*, 8> ByteValues;
4279 ByteValues.resize(ITy->getBitWidth()/8);
4281 // Try to find all the pieces corresponding to the bswap.
4282 uint32_t ByteMask = ~0U >> (32-ByteValues.size());
4283 if (CollectBSwapParts(&I, 0, ByteMask, ByteValues))
4286 // Check to see if all of the bytes come from the same value.
4287 Value *V = ByteValues[0];
4288 if (V == 0) return 0; // Didn't find a byte? Must be zero.
4290 // Check to make sure that all of the bytes come from the same value.
4291 for (unsigned i = 1, e = ByteValues.size(); i != e; ++i)
4292 if (ByteValues[i] != V)
4294 const Type *Tys[] = { ITy };
4295 Module *M = I.getParent()->getParent()->getParent();
4296 Function *F = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 1);
4297 return CallInst::Create(F, V);
4300 /// MatchSelectFromAndOr - We have an expression of the form (A&C)|(B&D). Check
4301 /// If A is (cond?-1:0) and either B or D is ~(cond?-1,0) or (cond?0,-1), then
4302 /// we can simplify this expression to "cond ? C : D or B".
4303 static Instruction *MatchSelectFromAndOr(Value *A, Value *B,
4304 Value *C, Value *D) {
4305 // If A is not a select of -1/0, this cannot match.
4307 if (!match(A, m_SelectCst(m_Value(Cond), -1, 0)))
4310 // ((cond?-1:0)&C) | (B&(cond?0:-1)) -> cond ? C : B.
4311 if (match(D, m_SelectCst(m_Specific(Cond), 0, -1)))
4312 return SelectInst::Create(Cond, C, B);
4313 if (match(D, m_Not(m_SelectCst(m_Specific(Cond), -1, 0))))
4314 return SelectInst::Create(Cond, C, B);
4315 // ((cond?-1:0)&C) | ((cond?0:-1)&D) -> cond ? C : D.
4316 if (match(B, m_SelectCst(m_Specific(Cond), 0, -1)))
4317 return SelectInst::Create(Cond, C, D);
4318 if (match(B, m_Not(m_SelectCst(m_Specific(Cond), -1, 0))))
4319 return SelectInst::Create(Cond, C, D);
4323 /// FoldOrOfICmps - Fold (icmp)|(icmp) if possible.
4324 Instruction *InstCombiner::FoldOrOfICmps(Instruction &I,
4325 ICmpInst *LHS, ICmpInst *RHS) {
4327 ConstantInt *LHSCst, *RHSCst;
4328 ICmpInst::Predicate LHSCC, RHSCC;
4330 // This only handles icmp of constants: (icmp1 A, C1) | (icmp2 B, C2).
4331 if (!match(LHS, m_ICmp(LHSCC, m_Value(Val), m_ConstantInt(LHSCst))) ||
4332 !match(RHS, m_ICmp(RHSCC, m_Value(Val2), m_ConstantInt(RHSCst))))
4335 // From here on, we only handle:
4336 // (icmp1 A, C1) | (icmp2 A, C2) --> something simpler.
4337 if (Val != Val2) return 0;
4339 // ICMP_[US][GL]E X, CST is folded to ICMP_[US][GL]T elsewhere.
4340 if (LHSCC == ICmpInst::ICMP_UGE || LHSCC == ICmpInst::ICMP_ULE ||
4341 RHSCC == ICmpInst::ICMP_UGE || RHSCC == ICmpInst::ICMP_ULE ||
4342 LHSCC == ICmpInst::ICMP_SGE || LHSCC == ICmpInst::ICMP_SLE ||
4343 RHSCC == ICmpInst::ICMP_SGE || RHSCC == ICmpInst::ICMP_SLE)
4346 // We can't fold (ugt x, C) | (sgt x, C2).
4347 if (!PredicatesFoldable(LHSCC, RHSCC))
4350 // Ensure that the larger constant is on the RHS.
4352 if (ICmpInst::isSignedPredicate(LHSCC) ||
4353 (ICmpInst::isEquality(LHSCC) &&
4354 ICmpInst::isSignedPredicate(RHSCC)))
4355 ShouldSwap = LHSCst->getValue().sgt(RHSCst->getValue());
4357 ShouldSwap = LHSCst->getValue().ugt(RHSCst->getValue());
4360 std::swap(LHS, RHS);
4361 std::swap(LHSCst, RHSCst);
4362 std::swap(LHSCC, RHSCC);
4365 // At this point, we know we have have two icmp instructions
4366 // comparing a value against two constants and or'ing the result
4367 // together. Because of the above check, we know that we only have
4368 // ICMP_EQ, ICMP_NE, ICMP_LT, and ICMP_GT here. We also know (from the
4369 // FoldICmpLogical check above), that the two constants are not
4371 assert(LHSCst != RHSCst && "Compares not folded above?");
4374 default: assert(0 && "Unknown integer condition code!");
4375 case ICmpInst::ICMP_EQ:
4377 default: assert(0 && "Unknown integer condition code!");
4378 case ICmpInst::ICMP_EQ:
4379 if (LHSCst == SubOne(RHSCst)) { // (X == 13 | X == 14) -> X-13 <u 2
4380 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
4381 Instruction *Add = BinaryOperator::CreateAdd(Val, AddCST,
4382 Val->getName()+".off");
4383 InsertNewInstBefore(Add, I);
4384 AddCST = Subtract(AddOne(RHSCst), LHSCst);
4385 return new ICmpInst(ICmpInst::ICMP_ULT, Add, AddCST);
4387 break; // (X == 13 | X == 15) -> no change
4388 case ICmpInst::ICMP_UGT: // (X == 13 | X u> 14) -> no change
4389 case ICmpInst::ICMP_SGT: // (X == 13 | X s> 14) -> no change
4391 case ICmpInst::ICMP_NE: // (X == 13 | X != 15) -> X != 15
4392 case ICmpInst::ICMP_ULT: // (X == 13 | X u< 15) -> X u< 15
4393 case ICmpInst::ICMP_SLT: // (X == 13 | X s< 15) -> X s< 15
4394 return ReplaceInstUsesWith(I, RHS);
4397 case ICmpInst::ICMP_NE:
4399 default: assert(0 && "Unknown integer condition code!");
4400 case ICmpInst::ICMP_EQ: // (X != 13 | X == 15) -> X != 13
4401 case ICmpInst::ICMP_UGT: // (X != 13 | X u> 15) -> X != 13
4402 case ICmpInst::ICMP_SGT: // (X != 13 | X s> 15) -> X != 13
4403 return ReplaceInstUsesWith(I, LHS);
4404 case ICmpInst::ICMP_NE: // (X != 13 | X != 15) -> true
4405 case ICmpInst::ICMP_ULT: // (X != 13 | X u< 15) -> true
4406 case ICmpInst::ICMP_SLT: // (X != 13 | X s< 15) -> true
4407 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4410 case ICmpInst::ICMP_ULT:
4412 default: assert(0 && "Unknown integer condition code!");
4413 case ICmpInst::ICMP_EQ: // (X u< 13 | X == 14) -> no change
4415 case ICmpInst::ICMP_UGT: // (X u< 13 | X u> 15) -> (X-13) u> 2
4416 // If RHSCst is [us]MAXINT, it is always false. Not handling
4417 // this can cause overflow.
4418 if (RHSCst->isMaxValue(false))
4419 return ReplaceInstUsesWith(I, LHS);
4420 return InsertRangeTest(Val, LHSCst, AddOne(RHSCst), false, false, I);
4421 case ICmpInst::ICMP_SGT: // (X u< 13 | X s> 15) -> no change
4423 case ICmpInst::ICMP_NE: // (X u< 13 | X != 15) -> X != 15
4424 case ICmpInst::ICMP_ULT: // (X u< 13 | X u< 15) -> X u< 15
4425 return ReplaceInstUsesWith(I, RHS);
4426 case ICmpInst::ICMP_SLT: // (X u< 13 | X s< 15) -> no change
4430 case ICmpInst::ICMP_SLT:
4432 default: assert(0 && "Unknown integer condition code!");
4433 case ICmpInst::ICMP_EQ: // (X s< 13 | X == 14) -> no change
4435 case ICmpInst::ICMP_SGT: // (X s< 13 | X s> 15) -> (X-13) s> 2
4436 // If RHSCst is [us]MAXINT, it is always false. Not handling
4437 // this can cause overflow.
4438 if (RHSCst->isMaxValue(true))
4439 return ReplaceInstUsesWith(I, LHS);
4440 return InsertRangeTest(Val, LHSCst, AddOne(RHSCst), true, false, I);
4441 case ICmpInst::ICMP_UGT: // (X s< 13 | X u> 15) -> no change
4443 case ICmpInst::ICMP_NE: // (X s< 13 | X != 15) -> X != 15
4444 case ICmpInst::ICMP_SLT: // (X s< 13 | X s< 15) -> X s< 15
4445 return ReplaceInstUsesWith(I, RHS);
4446 case ICmpInst::ICMP_ULT: // (X s< 13 | X u< 15) -> no change
4450 case ICmpInst::ICMP_UGT:
4452 default: assert(0 && "Unknown integer condition code!");
4453 case ICmpInst::ICMP_EQ: // (X u> 13 | X == 15) -> X u> 13
4454 case ICmpInst::ICMP_UGT: // (X u> 13 | X u> 15) -> X u> 13
4455 return ReplaceInstUsesWith(I, LHS);
4456 case ICmpInst::ICMP_SGT: // (X u> 13 | X s> 15) -> no change
4458 case ICmpInst::ICMP_NE: // (X u> 13 | X != 15) -> true
4459 case ICmpInst::ICMP_ULT: // (X u> 13 | X u< 15) -> true
4460 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4461 case ICmpInst::ICMP_SLT: // (X u> 13 | X s< 15) -> no change
4465 case ICmpInst::ICMP_SGT:
4467 default: assert(0 && "Unknown integer condition code!");
4468 case ICmpInst::ICMP_EQ: // (X s> 13 | X == 15) -> X > 13
4469 case ICmpInst::ICMP_SGT: // (X s> 13 | X s> 15) -> X > 13
4470 return ReplaceInstUsesWith(I, LHS);
4471 case ICmpInst::ICMP_UGT: // (X s> 13 | X u> 15) -> no change
4473 case ICmpInst::ICMP_NE: // (X s> 13 | X != 15) -> true
4474 case ICmpInst::ICMP_SLT: // (X s> 13 | X s< 15) -> true
4475 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4476 case ICmpInst::ICMP_ULT: // (X s> 13 | X u< 15) -> no change
4484 Instruction *InstCombiner::visitOr(BinaryOperator &I) {
4485 bool Changed = SimplifyCommutative(I);
4486 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4488 if (isa<UndefValue>(Op1)) // X | undef -> -1
4489 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4493 return ReplaceInstUsesWith(I, Op0);
4495 // See if we can simplify any instructions used by the instruction whose sole
4496 // purpose is to compute bits we don't care about.
4497 if (!isa<VectorType>(I.getType())) {
4498 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
4499 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
4500 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
4501 KnownZero, KnownOne))
4503 } else if (isa<ConstantAggregateZero>(Op1)) {
4504 return ReplaceInstUsesWith(I, Op0); // X | <0,0> -> X
4505 } else if (ConstantVector *CP = dyn_cast<ConstantVector>(Op1)) {
4506 if (CP->isAllOnesValue()) // X | <-1,-1> -> <-1,-1>
4507 return ReplaceInstUsesWith(I, I.getOperand(1));
4513 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
4514 ConstantInt *C1 = 0; Value *X = 0;
4515 // (X & C1) | C2 --> (X | C2) & (C1|C2)
4516 if (match(Op0, m_And(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
4517 Instruction *Or = BinaryOperator::CreateOr(X, RHS);
4518 InsertNewInstBefore(Or, I);
4520 return BinaryOperator::CreateAnd(Or,
4521 ConstantInt::get(RHS->getValue() | C1->getValue()));
4524 // (X ^ C1) | C2 --> (X | C2) ^ (C1&~C2)
4525 if (match(Op0, m_Xor(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
4526 Instruction *Or = BinaryOperator::CreateOr(X, RHS);
4527 InsertNewInstBefore(Or, I);
4529 return BinaryOperator::CreateXor(Or,
4530 ConstantInt::get(C1->getValue() & ~RHS->getValue()));
4533 // Try to fold constant and into select arguments.
4534 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
4535 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
4537 if (isa<PHINode>(Op0))
4538 if (Instruction *NV = FoldOpIntoPhi(I))
4542 Value *A = 0, *B = 0;
4543 ConstantInt *C1 = 0, *C2 = 0;
4545 if (match(Op0, m_And(m_Value(A), m_Value(B))))
4546 if (A == Op1 || B == Op1) // (A & ?) | A --> A
4547 return ReplaceInstUsesWith(I, Op1);
4548 if (match(Op1, m_And(m_Value(A), m_Value(B))))
4549 if (A == Op0 || B == Op0) // A | (A & ?) --> A
4550 return ReplaceInstUsesWith(I, Op0);
4552 // (A | B) | C and A | (B | C) -> bswap if possible.
4553 // (A >> B) | (C << D) and (A << B) | (B >> C) -> bswap if possible.
4554 if (match(Op0, m_Or(m_Value(), m_Value())) ||
4555 match(Op1, m_Or(m_Value(), m_Value())) ||
4556 (match(Op0, m_Shift(m_Value(), m_Value())) &&
4557 match(Op1, m_Shift(m_Value(), m_Value())))) {
4558 if (Instruction *BSwap = MatchBSwap(I))
4562 // (X^C)|Y -> (X|Y)^C iff Y&C == 0
4563 if (Op0->hasOneUse() && match(Op0, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
4564 MaskedValueIsZero(Op1, C1->getValue())) {
4565 Instruction *NOr = BinaryOperator::CreateOr(A, Op1);
4566 InsertNewInstBefore(NOr, I);
4568 return BinaryOperator::CreateXor(NOr, C1);
4571 // Y|(X^C) -> (X|Y)^C iff Y&C == 0
4572 if (Op1->hasOneUse() && match(Op1, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
4573 MaskedValueIsZero(Op0, C1->getValue())) {
4574 Instruction *NOr = BinaryOperator::CreateOr(A, Op0);
4575 InsertNewInstBefore(NOr, I);
4577 return BinaryOperator::CreateXor(NOr, C1);
4581 Value *C = 0, *D = 0;
4582 if (match(Op0, m_And(m_Value(A), m_Value(C))) &&
4583 match(Op1, m_And(m_Value(B), m_Value(D)))) {
4584 Value *V1 = 0, *V2 = 0, *V3 = 0;
4585 C1 = dyn_cast<ConstantInt>(C);
4586 C2 = dyn_cast<ConstantInt>(D);
4587 if (C1 && C2) { // (A & C1)|(B & C2)
4588 // If we have: ((V + N) & C1) | (V & C2)
4589 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
4590 // replace with V+N.
4591 if (C1->getValue() == ~C2->getValue()) {
4592 if ((C2->getValue() & (C2->getValue()+1)) == 0 && // C2 == 0+1+
4593 match(A, m_Add(m_Value(V1), m_Value(V2)))) {
4594 // Add commutes, try both ways.
4595 if (V1 == B && MaskedValueIsZero(V2, C2->getValue()))
4596 return ReplaceInstUsesWith(I, A);
4597 if (V2 == B && MaskedValueIsZero(V1, C2->getValue()))
4598 return ReplaceInstUsesWith(I, A);
4600 // Or commutes, try both ways.
4601 if ((C1->getValue() & (C1->getValue()+1)) == 0 &&
4602 match(B, m_Add(m_Value(V1), m_Value(V2)))) {
4603 // Add commutes, try both ways.
4604 if (V1 == A && MaskedValueIsZero(V2, C1->getValue()))
4605 return ReplaceInstUsesWith(I, B);
4606 if (V2 == A && MaskedValueIsZero(V1, C1->getValue()))
4607 return ReplaceInstUsesWith(I, B);
4610 V1 = 0; V2 = 0; V3 = 0;
4613 // Check to see if we have any common things being and'ed. If so, find the
4614 // terms for V1 & (V2|V3).
4615 if (isOnlyUse(Op0) || isOnlyUse(Op1)) {
4616 if (A == B) // (A & C)|(A & D) == A & (C|D)
4617 V1 = A, V2 = C, V3 = D;
4618 else if (A == D) // (A & C)|(B & A) == A & (B|C)
4619 V1 = A, V2 = B, V3 = C;
4620 else if (C == B) // (A & C)|(C & D) == C & (A|D)
4621 V1 = C, V2 = A, V3 = D;
4622 else if (C == D) // (A & C)|(B & C) == C & (A|B)
4623 V1 = C, V2 = A, V3 = B;
4627 InsertNewInstBefore(BinaryOperator::CreateOr(V2, V3, "tmp"), I);
4628 return BinaryOperator::CreateAnd(V1, Or);
4632 // (A & (C0?-1:0)) | (B & ~(C0?-1:0)) -> C0 ? A : B, and commuted variants
4633 if (Instruction *Match = MatchSelectFromAndOr(A, B, C, D))
4635 if (Instruction *Match = MatchSelectFromAndOr(B, A, D, C))
4637 if (Instruction *Match = MatchSelectFromAndOr(C, B, A, D))
4639 if (Instruction *Match = MatchSelectFromAndOr(D, A, B, C))
4643 // (X >> Z) | (Y >> Z) -> (X|Y) >> Z for all shifts.
4644 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
4645 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
4646 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
4647 SI0->getOperand(1) == SI1->getOperand(1) &&
4648 (SI0->hasOneUse() || SI1->hasOneUse())) {
4649 Instruction *NewOp =
4650 InsertNewInstBefore(BinaryOperator::CreateOr(SI0->getOperand(0),
4652 SI0->getName()), I);
4653 return BinaryOperator::Create(SI1->getOpcode(), NewOp,
4654 SI1->getOperand(1));
4658 if (match(Op0, m_Not(m_Value(A)))) { // ~A | Op1
4659 if (A == Op1) // ~A | A == -1
4660 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4664 // Note, A is still live here!
4665 if (match(Op1, m_Not(m_Value(B)))) { // Op0 | ~B
4667 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4669 // (~A | ~B) == (~(A & B)) - De Morgan's Law
4670 if (A && isOnlyUse(Op0) && isOnlyUse(Op1)) {
4671 Value *And = InsertNewInstBefore(BinaryOperator::CreateAnd(A, B,
4672 I.getName()+".demorgan"), I);
4673 return BinaryOperator::CreateNot(And);
4677 // (icmp1 A, B) | (icmp2 A, B) --> (icmp3 A, B)
4678 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1))) {
4679 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
4682 if (ICmpInst *LHS = dyn_cast<ICmpInst>(I.getOperand(0)))
4683 if (Instruction *Res = FoldOrOfICmps(I, LHS, RHS))
4687 // fold (or (cast A), (cast B)) -> (cast (or A, B))
4688 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
4689 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4690 if (Op0C->getOpcode() == Op1C->getOpcode()) {// same cast kind ?
4691 if (!isa<ICmpInst>(Op0C->getOperand(0)) ||
4692 !isa<ICmpInst>(Op1C->getOperand(0))) {
4693 const Type *SrcTy = Op0C->getOperand(0)->getType();
4694 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
4695 // Only do this if the casts both really cause code to be
4697 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4699 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4701 Instruction *NewOp = BinaryOperator::CreateOr(Op0C->getOperand(0),
4702 Op1C->getOperand(0),
4704 InsertNewInstBefore(NewOp, I);
4705 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
4712 // (fcmp uno x, c) | (fcmp uno y, c) -> (fcmp uno x, y)
4713 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0))) {
4714 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1))) {
4715 if (LHS->getPredicate() == FCmpInst::FCMP_UNO &&
4716 RHS->getPredicate() == FCmpInst::FCMP_UNO &&
4717 LHS->getOperand(0)->getType() == RHS->getOperand(0)->getType()) {
4718 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
4719 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
4720 // If either of the constants are nans, then the whole thing returns
4722 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
4723 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4725 // Otherwise, no need to compare the two constants, compare the
4727 return new FCmpInst(FCmpInst::FCMP_UNO, LHS->getOperand(0),
4728 RHS->getOperand(0));
4731 Value *Op0LHS, *Op0RHS, *Op1LHS, *Op1RHS;
4732 FCmpInst::Predicate Op0CC, Op1CC;
4733 if (match(Op0, m_FCmp(Op0CC, m_Value(Op0LHS), m_Value(Op0RHS))) &&
4734 match(Op1, m_FCmp(Op1CC, m_Value(Op1LHS), m_Value(Op1RHS)))) {
4735 if (Op0LHS == Op1RHS && Op0RHS == Op1LHS) {
4736 // Swap RHS operands to match LHS.
4737 Op1CC = FCmpInst::getSwappedPredicate(Op1CC);
4738 std::swap(Op1LHS, Op1RHS);
4740 if (Op0LHS == Op1LHS && Op0RHS == Op1RHS) {
4741 // Simplify (fcmp cc0 x, y) | (fcmp cc1 x, y).
4743 return new FCmpInst((FCmpInst::Predicate)Op0CC, Op0LHS, Op0RHS);
4744 else if (Op0CC == FCmpInst::FCMP_TRUE ||
4745 Op1CC == FCmpInst::FCMP_TRUE)
4746 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4747 else if (Op0CC == FCmpInst::FCMP_FALSE)
4748 return ReplaceInstUsesWith(I, Op1);
4749 else if (Op1CC == FCmpInst::FCMP_FALSE)
4750 return ReplaceInstUsesWith(I, Op0);
4753 unsigned Op0Pred = getFCmpCode(Op0CC, Op0Ordered);
4754 unsigned Op1Pred = getFCmpCode(Op1CC, Op1Ordered);
4755 if (Op0Ordered == Op1Ordered) {
4756 // If both are ordered or unordered, return a new fcmp with
4757 // or'ed predicates.
4758 Value *RV = getFCmpValue(Op0Ordered, Op0Pred|Op1Pred,
4760 if (Instruction *I = dyn_cast<Instruction>(RV))
4762 // Otherwise, it's a constant boolean value...
4763 return ReplaceInstUsesWith(I, RV);
4771 return Changed ? &I : 0;
4776 // XorSelf - Implements: X ^ X --> 0
4779 XorSelf(Value *rhs) : RHS(rhs) {}
4780 bool shouldApply(Value *LHS) const { return LHS == RHS; }
4781 Instruction *apply(BinaryOperator &Xor) const {
4788 Instruction *InstCombiner::visitXor(BinaryOperator &I) {
4789 bool Changed = SimplifyCommutative(I);
4790 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4792 if (isa<UndefValue>(Op1)) {
4793 if (isa<UndefValue>(Op0))
4794 // Handle undef ^ undef -> 0 special case. This is a common
4796 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
4797 return ReplaceInstUsesWith(I, Op1); // X ^ undef -> undef
4800 // xor X, X = 0, even if X is nested in a sequence of Xor's.
4801 if (Instruction *Result = AssociativeOpt(I, XorSelf(Op1))) {
4802 assert(Result == &I && "AssociativeOpt didn't work?"); Result=Result;
4803 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
4806 // See if we can simplify any instructions used by the instruction whose sole
4807 // purpose is to compute bits we don't care about.
4808 if (!isa<VectorType>(I.getType())) {
4809 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
4810 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
4811 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
4812 KnownZero, KnownOne))
4814 } else if (isa<ConstantAggregateZero>(Op1)) {
4815 return ReplaceInstUsesWith(I, Op0); // X ^ <0,0> -> X
4818 // Is this a ~ operation?
4819 if (Value *NotOp = dyn_castNotVal(&I)) {
4820 // ~(~X & Y) --> (X | ~Y) - De Morgan's Law
4821 // ~(~X | Y) === (X & ~Y) - De Morgan's Law
4822 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(NotOp)) {
4823 if (Op0I->getOpcode() == Instruction::And ||
4824 Op0I->getOpcode() == Instruction::Or) {
4825 if (dyn_castNotVal(Op0I->getOperand(1))) Op0I->swapOperands();
4826 if (Value *Op0NotVal = dyn_castNotVal(Op0I->getOperand(0))) {
4828 BinaryOperator::CreateNot(Op0I->getOperand(1),
4829 Op0I->getOperand(1)->getName()+".not");
4830 InsertNewInstBefore(NotY, I);
4831 if (Op0I->getOpcode() == Instruction::And)
4832 return BinaryOperator::CreateOr(Op0NotVal, NotY);
4834 return BinaryOperator::CreateAnd(Op0NotVal, NotY);
4841 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
4842 // xor (cmp A, B), true = not (cmp A, B) = !cmp A, B
4843 if (RHS == ConstantInt::getTrue() && Op0->hasOneUse()) {
4844 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Op0))
4845 return new ICmpInst(ICI->getInversePredicate(),
4846 ICI->getOperand(0), ICI->getOperand(1));
4848 if (FCmpInst *FCI = dyn_cast<FCmpInst>(Op0))
4849 return new FCmpInst(FCI->getInversePredicate(),
4850 FCI->getOperand(0), FCI->getOperand(1));
4853 // fold (xor(zext(cmp)), 1) and (xor(sext(cmp)), -1) to ext(!cmp).
4854 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
4855 if (CmpInst *CI = dyn_cast<CmpInst>(Op0C->getOperand(0))) {
4856 if (CI->hasOneUse() && Op0C->hasOneUse()) {
4857 Instruction::CastOps Opcode = Op0C->getOpcode();
4858 if (Opcode == Instruction::ZExt || Opcode == Instruction::SExt) {
4859 if (RHS == ConstantExpr::getCast(Opcode, ConstantInt::getTrue(),
4860 Op0C->getDestTy())) {
4861 Instruction *NewCI = InsertNewInstBefore(CmpInst::Create(
4862 CI->getOpcode(), CI->getInversePredicate(),
4863 CI->getOperand(0), CI->getOperand(1)), I);
4864 NewCI->takeName(CI);
4865 return CastInst::Create(Opcode, NewCI, Op0C->getType());
4872 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
4873 // ~(c-X) == X-c-1 == X+(-c-1)
4874 if (Op0I->getOpcode() == Instruction::Sub && RHS->isAllOnesValue())
4875 if (Constant *Op0I0C = dyn_cast<Constant>(Op0I->getOperand(0))) {
4876 Constant *NegOp0I0C = ConstantExpr::getNeg(Op0I0C);
4877 Constant *ConstantRHS = ConstantExpr::getSub(NegOp0I0C,
4878 ConstantInt::get(I.getType(), 1));
4879 return BinaryOperator::CreateAdd(Op0I->getOperand(1), ConstantRHS);
4882 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) {
4883 if (Op0I->getOpcode() == Instruction::Add) {
4884 // ~(X-c) --> (-c-1)-X
4885 if (RHS->isAllOnesValue()) {
4886 Constant *NegOp0CI = ConstantExpr::getNeg(Op0CI);
4887 return BinaryOperator::CreateSub(
4888 ConstantExpr::getSub(NegOp0CI,
4889 ConstantInt::get(I.getType(), 1)),
4890 Op0I->getOperand(0));
4891 } else if (RHS->getValue().isSignBit()) {
4892 // (X + C) ^ signbit -> (X + C + signbit)
4893 Constant *C = ConstantInt::get(RHS->getValue() + Op0CI->getValue());
4894 return BinaryOperator::CreateAdd(Op0I->getOperand(0), C);
4897 } else if (Op0I->getOpcode() == Instruction::Or) {
4898 // (X|C1)^C2 -> X^(C1|C2) iff X&~C1 == 0
4899 if (MaskedValueIsZero(Op0I->getOperand(0), Op0CI->getValue())) {
4900 Constant *NewRHS = ConstantExpr::getOr(Op0CI, RHS);
4901 // Anything in both C1 and C2 is known to be zero, remove it from
4903 Constant *CommonBits = And(Op0CI, RHS);
4904 NewRHS = ConstantExpr::getAnd(NewRHS,
4905 ConstantExpr::getNot(CommonBits));
4906 AddToWorkList(Op0I);
4907 I.setOperand(0, Op0I->getOperand(0));
4908 I.setOperand(1, NewRHS);
4915 // Try to fold constant and into select arguments.
4916 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
4917 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
4919 if (isa<PHINode>(Op0))
4920 if (Instruction *NV = FoldOpIntoPhi(I))
4924 if (Value *X = dyn_castNotVal(Op0)) // ~A ^ A == -1
4926 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4928 if (Value *X = dyn_castNotVal(Op1)) // A ^ ~A == -1
4930 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4933 BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1);
4936 if (match(Op1I, m_Or(m_Value(A), m_Value(B)))) {
4937 if (A == Op0) { // B^(B|A) == (A|B)^B
4938 Op1I->swapOperands();
4940 std::swap(Op0, Op1);
4941 } else if (B == Op0) { // B^(A|B) == (A|B)^B
4942 I.swapOperands(); // Simplified below.
4943 std::swap(Op0, Op1);
4945 } else if (match(Op1I, m_Xor(m_Specific(Op0), m_Value(B)))) {
4946 return ReplaceInstUsesWith(I, B); // A^(A^B) == B
4947 } else if (match(Op1I, m_Xor(m_Value(A), m_Specific(Op0)))) {
4948 return ReplaceInstUsesWith(I, A); // A^(B^A) == B
4949 } else if (match(Op1I, m_And(m_Value(A), m_Value(B))) && Op1I->hasOneUse()){
4950 if (A == Op0) { // A^(A&B) -> A^(B&A)
4951 Op1I->swapOperands();
4954 if (B == Op0) { // A^(B&A) -> (B&A)^A
4955 I.swapOperands(); // Simplified below.
4956 std::swap(Op0, Op1);
4961 BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0);
4964 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) && Op0I->hasOneUse()) {
4965 if (A == Op1) // (B|A)^B == (A|B)^B
4967 if (B == Op1) { // (A|B)^B == A & ~B
4969 InsertNewInstBefore(BinaryOperator::CreateNot(Op1, "tmp"), I);
4970 return BinaryOperator::CreateAnd(A, NotB);
4972 } else if (match(Op0I, m_Xor(m_Specific(Op1), m_Value(B)))) {
4973 return ReplaceInstUsesWith(I, B); // (A^B)^A == B
4974 } else if (match(Op0I, m_Xor(m_Value(A), m_Specific(Op1)))) {
4975 return ReplaceInstUsesWith(I, A); // (B^A)^A == B
4976 } else if (match(Op0I, m_And(m_Value(A), m_Value(B))) && Op0I->hasOneUse()){
4977 if (A == Op1) // (A&B)^A -> (B&A)^A
4979 if (B == Op1 && // (B&A)^A == ~B & A
4980 !isa<ConstantInt>(Op1)) { // Canonical form is (B&C)^C
4982 InsertNewInstBefore(BinaryOperator::CreateNot(A, "tmp"), I);
4983 return BinaryOperator::CreateAnd(N, Op1);
4988 // (X >> Z) ^ (Y >> Z) -> (X^Y) >> Z for all shifts.
4989 if (Op0I && Op1I && Op0I->isShift() &&
4990 Op0I->getOpcode() == Op1I->getOpcode() &&
4991 Op0I->getOperand(1) == Op1I->getOperand(1) &&
4992 (Op1I->hasOneUse() || Op1I->hasOneUse())) {
4993 Instruction *NewOp =
4994 InsertNewInstBefore(BinaryOperator::CreateXor(Op0I->getOperand(0),
4995 Op1I->getOperand(0),
4996 Op0I->getName()), I);
4997 return BinaryOperator::Create(Op1I->getOpcode(), NewOp,
4998 Op1I->getOperand(1));
5002 Value *A, *B, *C, *D;
5003 // (A & B)^(A | B) -> A ^ B
5004 if (match(Op0I, m_And(m_Value(A), m_Value(B))) &&
5005 match(Op1I, m_Or(m_Value(C), m_Value(D)))) {
5006 if ((A == C && B == D) || (A == D && B == C))
5007 return BinaryOperator::CreateXor(A, B);
5009 // (A | B)^(A & B) -> A ^ B
5010 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) &&
5011 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
5012 if ((A == C && B == D) || (A == D && B == C))
5013 return BinaryOperator::CreateXor(A, B);
5017 if ((Op0I->hasOneUse() || Op1I->hasOneUse()) &&
5018 match(Op0I, m_And(m_Value(A), m_Value(B))) &&
5019 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
5020 // (X & Y)^(X & Y) -> (Y^Z) & X
5021 Value *X = 0, *Y = 0, *Z = 0;
5023 X = A, Y = B, Z = D;
5025 X = A, Y = B, Z = C;
5027 X = B, Y = A, Z = D;
5029 X = B, Y = A, Z = C;
5032 Instruction *NewOp =
5033 InsertNewInstBefore(BinaryOperator::CreateXor(Y, Z, Op0->getName()), I);
5034 return BinaryOperator::CreateAnd(NewOp, X);
5039 // (icmp1 A, B) ^ (icmp2 A, B) --> (icmp3 A, B)
5040 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1)))
5041 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
5044 // fold (xor (cast A), (cast B)) -> (cast (xor A, B))
5045 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
5046 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
5047 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind?
5048 const Type *SrcTy = Op0C->getOperand(0)->getType();
5049 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
5050 // Only do this if the casts both really cause code to be generated.
5051 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
5053 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
5055 Instruction *NewOp = BinaryOperator::CreateXor(Op0C->getOperand(0),
5056 Op1C->getOperand(0),
5058 InsertNewInstBefore(NewOp, I);
5059 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
5064 return Changed ? &I : 0;
5067 /// AddWithOverflow - Compute Result = In1+In2, returning true if the result
5068 /// overflowed for this type.
5069 static bool AddWithOverflow(ConstantInt *&Result, ConstantInt *In1,
5070 ConstantInt *In2, bool IsSigned = false) {
5071 Result = cast<ConstantInt>(Add(In1, In2));
5074 if (In2->getValue().isNegative())
5075 return Result->getValue().sgt(In1->getValue());
5077 return Result->getValue().slt(In1->getValue());
5079 return Result->getValue().ult(In1->getValue());
5082 /// SubWithOverflow - Compute Result = In1-In2, returning true if the result
5083 /// overflowed for this type.
5084 static bool SubWithOverflow(ConstantInt *&Result, ConstantInt *In1,
5085 ConstantInt *In2, bool IsSigned = false) {
5086 Result = cast<ConstantInt>(Subtract(In1, In2));
5089 if (In2->getValue().isNegative())
5090 return Result->getValue().slt(In1->getValue());
5092 return Result->getValue().sgt(In1->getValue());
5094 return Result->getValue().ugt(In1->getValue());
5097 /// EmitGEPOffset - Given a getelementptr instruction/constantexpr, emit the
5098 /// code necessary to compute the offset from the base pointer (without adding
5099 /// in the base pointer). Return the result as a signed integer of intptr size.
5100 static Value *EmitGEPOffset(User *GEP, Instruction &I, InstCombiner &IC) {
5101 TargetData &TD = IC.getTargetData();
5102 gep_type_iterator GTI = gep_type_begin(GEP);
5103 const Type *IntPtrTy = TD.getIntPtrType();
5104 Value *Result = Constant::getNullValue(IntPtrTy);
5106 // Build a mask for high order bits.
5107 unsigned IntPtrWidth = TD.getPointerSizeInBits();
5108 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth);
5110 for (User::op_iterator i = GEP->op_begin() + 1, e = GEP->op_end(); i != e;
5113 uint64_t Size = TD.getABITypeSize(GTI.getIndexedType()) & PtrSizeMask;
5114 if (ConstantInt *OpC = dyn_cast<ConstantInt>(Op)) {
5115 if (OpC->isZero()) continue;
5117 // Handle a struct index, which adds its field offset to the pointer.
5118 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
5119 Size = TD.getStructLayout(STy)->getElementOffset(OpC->getZExtValue());
5121 if (ConstantInt *RC = dyn_cast<ConstantInt>(Result))
5122 Result = ConstantInt::get(RC->getValue() + APInt(IntPtrWidth, Size));
5124 Result = IC.InsertNewInstBefore(
5125 BinaryOperator::CreateAdd(Result,
5126 ConstantInt::get(IntPtrTy, Size),
5127 GEP->getName()+".offs"), I);
5131 Constant *Scale = ConstantInt::get(IntPtrTy, Size);
5132 Constant *OC = ConstantExpr::getIntegerCast(OpC, IntPtrTy, true /*SExt*/);
5133 Scale = ConstantExpr::getMul(OC, Scale);
5134 if (Constant *RC = dyn_cast<Constant>(Result))
5135 Result = ConstantExpr::getAdd(RC, Scale);
5137 // Emit an add instruction.
5138 Result = IC.InsertNewInstBefore(
5139 BinaryOperator::CreateAdd(Result, Scale,
5140 GEP->getName()+".offs"), I);
5144 // Convert to correct type.
5145 if (Op->getType() != IntPtrTy) {
5146 if (Constant *OpC = dyn_cast<Constant>(Op))
5147 Op = ConstantExpr::getSExt(OpC, IntPtrTy);
5149 Op = IC.InsertNewInstBefore(new SExtInst(Op, IntPtrTy,
5150 Op->getName()+".c"), I);
5153 Constant *Scale = ConstantInt::get(IntPtrTy, Size);
5154 if (Constant *OpC = dyn_cast<Constant>(Op))
5155 Op = ConstantExpr::getMul(OpC, Scale);
5156 else // We'll let instcombine(mul) convert this to a shl if possible.
5157 Op = IC.InsertNewInstBefore(BinaryOperator::CreateMul(Op, Scale,
5158 GEP->getName()+".idx"), I);
5161 // Emit an add instruction.
5162 if (isa<Constant>(Op) && isa<Constant>(Result))
5163 Result = ConstantExpr::getAdd(cast<Constant>(Op),
5164 cast<Constant>(Result));
5166 Result = IC.InsertNewInstBefore(BinaryOperator::CreateAdd(Op, Result,
5167 GEP->getName()+".offs"), I);
5173 /// EvaluateGEPOffsetExpression - Return an value that can be used to compare of
5174 /// the *offset* implied by GEP to zero. For example, if we have &A[i], we want
5175 /// to return 'i' for "icmp ne i, 0". Note that, in general, indices can be
5176 /// complex, and scales are involved. The above expression would also be legal
5177 /// to codegen as "icmp ne (i*4), 0" (assuming A is a pointer to i32). This
5178 /// later form is less amenable to optimization though, and we are allowed to
5179 /// generate the first by knowing that pointer arithmetic doesn't overflow.
5181 /// If we can't emit an optimized form for this expression, this returns null.
5183 static Value *EvaluateGEPOffsetExpression(User *GEP, Instruction &I,
5185 TargetData &TD = IC.getTargetData();
5186 gep_type_iterator GTI = gep_type_begin(GEP);
5188 // Check to see if this gep only has a single variable index. If so, and if
5189 // any constant indices are a multiple of its scale, then we can compute this
5190 // in terms of the scale of the variable index. For example, if the GEP
5191 // implies an offset of "12 + i*4", then we can codegen this as "3 + i",
5192 // because the expression will cross zero at the same point.
5193 unsigned i, e = GEP->getNumOperands();
5195 for (i = 1; i != e; ++i, ++GTI) {
5196 if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i))) {
5197 // Compute the aggregate offset of constant indices.
5198 if (CI->isZero()) continue;
5200 // Handle a struct index, which adds its field offset to the pointer.
5201 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
5202 Offset += TD.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
5204 uint64_t Size = TD.getABITypeSize(GTI.getIndexedType());
5205 Offset += Size*CI->getSExtValue();
5208 // Found our variable index.
5213 // If there are no variable indices, we must have a constant offset, just
5214 // evaluate it the general way.
5215 if (i == e) return 0;
5217 Value *VariableIdx = GEP->getOperand(i);
5218 // Determine the scale factor of the variable element. For example, this is
5219 // 4 if the variable index is into an array of i32.
5220 uint64_t VariableScale = TD.getABITypeSize(GTI.getIndexedType());
5222 // Verify that there are no other variable indices. If so, emit the hard way.
5223 for (++i, ++GTI; i != e; ++i, ++GTI) {
5224 ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i));
5227 // Compute the aggregate offset of constant indices.
5228 if (CI->isZero()) continue;
5230 // Handle a struct index, which adds its field offset to the pointer.
5231 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
5232 Offset += TD.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
5234 uint64_t Size = TD.getABITypeSize(GTI.getIndexedType());
5235 Offset += Size*CI->getSExtValue();
5239 // Okay, we know we have a single variable index, which must be a
5240 // pointer/array/vector index. If there is no offset, life is simple, return
5242 unsigned IntPtrWidth = TD.getPointerSizeInBits();
5244 // Cast to intptrty in case a truncation occurs. If an extension is needed,
5245 // we don't need to bother extending: the extension won't affect where the
5246 // computation crosses zero.
5247 if (VariableIdx->getType()->getPrimitiveSizeInBits() > IntPtrWidth)
5248 VariableIdx = new TruncInst(VariableIdx, TD.getIntPtrType(),
5249 VariableIdx->getNameStart(), &I);
5253 // Otherwise, there is an index. The computation we will do will be modulo
5254 // the pointer size, so get it.
5255 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth);
5257 Offset &= PtrSizeMask;
5258 VariableScale &= PtrSizeMask;
5260 // To do this transformation, any constant index must be a multiple of the
5261 // variable scale factor. For example, we can evaluate "12 + 4*i" as "3 + i",
5262 // but we can't evaluate "10 + 3*i" in terms of i. Check that the offset is a
5263 // multiple of the variable scale.
5264 int64_t NewOffs = Offset / (int64_t)VariableScale;
5265 if (Offset != NewOffs*(int64_t)VariableScale)
5268 // Okay, we can do this evaluation. Start by converting the index to intptr.
5269 const Type *IntPtrTy = TD.getIntPtrType();
5270 if (VariableIdx->getType() != IntPtrTy)
5271 VariableIdx = CastInst::CreateIntegerCast(VariableIdx, IntPtrTy,
5273 VariableIdx->getNameStart(), &I);
5274 Constant *OffsetVal = ConstantInt::get(IntPtrTy, NewOffs);
5275 return BinaryOperator::CreateAdd(VariableIdx, OffsetVal, "offset", &I);
5279 /// FoldGEPICmp - Fold comparisons between a GEP instruction and something
5280 /// else. At this point we know that the GEP is on the LHS of the comparison.
5281 Instruction *InstCombiner::FoldGEPICmp(User *GEPLHS, Value *RHS,
5282 ICmpInst::Predicate Cond,
5284 assert(dyn_castGetElementPtr(GEPLHS) && "LHS is not a getelementptr!");
5286 // Look through bitcasts.
5287 if (BitCastInst *BCI = dyn_cast<BitCastInst>(RHS))
5288 RHS = BCI->getOperand(0);
5290 Value *PtrBase = GEPLHS->getOperand(0);
5291 if (PtrBase == RHS) {
5292 // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
5293 // This transformation (ignoring the base and scales) is valid because we
5294 // know pointers can't overflow. See if we can output an optimized form.
5295 Value *Offset = EvaluateGEPOffsetExpression(GEPLHS, I, *this);
5297 // If not, synthesize the offset the hard way.
5299 Offset = EmitGEPOffset(GEPLHS, I, *this);
5300 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Offset,
5301 Constant::getNullValue(Offset->getType()));
5302 } else if (User *GEPRHS = dyn_castGetElementPtr(RHS)) {
5303 // If the base pointers are different, but the indices are the same, just
5304 // compare the base pointer.
5305 if (PtrBase != GEPRHS->getOperand(0)) {
5306 bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands();
5307 IndicesTheSame &= GEPLHS->getOperand(0)->getType() ==
5308 GEPRHS->getOperand(0)->getType();
5310 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
5311 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
5312 IndicesTheSame = false;
5316 // If all indices are the same, just compare the base pointers.
5318 return new ICmpInst(ICmpInst::getSignedPredicate(Cond),
5319 GEPLHS->getOperand(0), GEPRHS->getOperand(0));
5321 // Otherwise, the base pointers are different and the indices are
5322 // different, bail out.
5326 // If one of the GEPs has all zero indices, recurse.
5327 bool AllZeros = true;
5328 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
5329 if (!isa<Constant>(GEPLHS->getOperand(i)) ||
5330 !cast<Constant>(GEPLHS->getOperand(i))->isNullValue()) {
5335 return FoldGEPICmp(GEPRHS, GEPLHS->getOperand(0),
5336 ICmpInst::getSwappedPredicate(Cond), I);
5338 // If the other GEP has all zero indices, recurse.
5340 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
5341 if (!isa<Constant>(GEPRHS->getOperand(i)) ||
5342 !cast<Constant>(GEPRHS->getOperand(i))->isNullValue()) {
5347 return FoldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I);
5349 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
5350 // If the GEPs only differ by one index, compare it.
5351 unsigned NumDifferences = 0; // Keep track of # differences.
5352 unsigned DiffOperand = 0; // The operand that differs.
5353 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
5354 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
5355 if (GEPLHS->getOperand(i)->getType()->getPrimitiveSizeInBits() !=
5356 GEPRHS->getOperand(i)->getType()->getPrimitiveSizeInBits()) {
5357 // Irreconcilable differences.
5361 if (NumDifferences++) break;
5366 if (NumDifferences == 0) // SAME GEP?
5367 return ReplaceInstUsesWith(I, // No comparison is needed here.
5368 ConstantInt::get(Type::Int1Ty,
5369 ICmpInst::isTrueWhenEqual(Cond)));
5371 else if (NumDifferences == 1) {
5372 Value *LHSV = GEPLHS->getOperand(DiffOperand);
5373 Value *RHSV = GEPRHS->getOperand(DiffOperand);
5374 // Make sure we do a signed comparison here.
5375 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), LHSV, RHSV);
5379 // Only lower this if the icmp is the only user of the GEP or if we expect
5380 // the result to fold to a constant!
5381 if ((isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
5382 (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
5383 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
5384 Value *L = EmitGEPOffset(GEPLHS, I, *this);
5385 Value *R = EmitGEPOffset(GEPRHS, I, *this);
5386 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), L, R);
5392 /// FoldFCmp_IntToFP_Cst - Fold fcmp ([us]itofp x, cst) if possible.
5394 Instruction *InstCombiner::FoldFCmp_IntToFP_Cst(FCmpInst &I,
5397 if (!isa<ConstantFP>(RHSC)) return 0;
5398 const APFloat &RHS = cast<ConstantFP>(RHSC)->getValueAPF();
5400 // Get the width of the mantissa. We don't want to hack on conversions that
5401 // might lose information from the integer, e.g. "i64 -> float"
5402 int MantissaWidth = LHSI->getType()->getFPMantissaWidth();
5403 if (MantissaWidth == -1) return 0; // Unknown.
5405 // Check to see that the input is converted from an integer type that is small
5406 // enough that preserves all bits. TODO: check here for "known" sign bits.
5407 // This would allow us to handle (fptosi (x >>s 62) to float) if x is i64 f.e.
5408 unsigned InputSize = LHSI->getOperand(0)->getType()->getPrimitiveSizeInBits();
5410 // If this is a uitofp instruction, we need an extra bit to hold the sign.
5411 bool LHSUnsigned = isa<UIToFPInst>(LHSI);
5415 // If the conversion would lose info, don't hack on this.
5416 if ((int)InputSize > MantissaWidth)
5419 // Otherwise, we can potentially simplify the comparison. We know that it
5420 // will always come through as an integer value and we know the constant is
5421 // not a NAN (it would have been previously simplified).
5422 assert(!RHS.isNaN() && "NaN comparison not already folded!");
5424 ICmpInst::Predicate Pred;
5425 switch (I.getPredicate()) {
5426 default: assert(0 && "Unexpected predicate!");
5427 case FCmpInst::FCMP_UEQ:
5428 case FCmpInst::FCMP_OEQ:
5429 Pred = ICmpInst::ICMP_EQ;
5431 case FCmpInst::FCMP_UGT:
5432 case FCmpInst::FCMP_OGT:
5433 Pred = LHSUnsigned ? ICmpInst::ICMP_UGT : ICmpInst::ICMP_SGT;
5435 case FCmpInst::FCMP_UGE:
5436 case FCmpInst::FCMP_OGE:
5437 Pred = LHSUnsigned ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_SGE;
5439 case FCmpInst::FCMP_ULT:
5440 case FCmpInst::FCMP_OLT:
5441 Pred = LHSUnsigned ? ICmpInst::ICMP_ULT : ICmpInst::ICMP_SLT;
5443 case FCmpInst::FCMP_ULE:
5444 case FCmpInst::FCMP_OLE:
5445 Pred = LHSUnsigned ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_SLE;
5447 case FCmpInst::FCMP_UNE:
5448 case FCmpInst::FCMP_ONE:
5449 Pred = ICmpInst::ICMP_NE;
5451 case FCmpInst::FCMP_ORD:
5452 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
5453 case FCmpInst::FCMP_UNO:
5454 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
5457 const IntegerType *IntTy = cast<IntegerType>(LHSI->getOperand(0)->getType());
5459 // Now we know that the APFloat is a normal number, zero or inf.
5461 // See if the FP constant is too large for the integer. For example,
5462 // comparing an i8 to 300.0.
5463 unsigned IntWidth = IntTy->getPrimitiveSizeInBits();
5466 // If the RHS value is > SignedMax, fold the comparison. This handles +INF
5467 // and large values.
5468 APFloat SMax(RHS.getSemantics(), APFloat::fcZero, false);
5469 SMax.convertFromAPInt(APInt::getSignedMaxValue(IntWidth), true,
5470 APFloat::rmNearestTiesToEven);
5471 if (SMax.compare(RHS) == APFloat::cmpLessThan) { // smax < 13123.0
5472 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SLT ||
5473 Pred == ICmpInst::ICMP_SLE)
5474 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
5475 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
5478 // If the RHS value is > UnsignedMax, fold the comparison. This handles
5479 // +INF and large values.
5480 APFloat UMax(RHS.getSemantics(), APFloat::fcZero, false);
5481 UMax.convertFromAPInt(APInt::getMaxValue(IntWidth), false,
5482 APFloat::rmNearestTiesToEven);
5483 if (UMax.compare(RHS) == APFloat::cmpLessThan) { // umax < 13123.0
5484 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_ULT ||
5485 Pred == ICmpInst::ICMP_ULE)
5486 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
5487 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
5492 // See if the RHS value is < SignedMin.
5493 APFloat SMin(RHS.getSemantics(), APFloat::fcZero, false);
5494 SMin.convertFromAPInt(APInt::getSignedMinValue(IntWidth), true,
5495 APFloat::rmNearestTiesToEven);
5496 if (SMin.compare(RHS) == APFloat::cmpGreaterThan) { // smin > 12312.0
5497 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SGT ||
5498 Pred == ICmpInst::ICMP_SGE)
5499 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
5500 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
5504 // Okay, now we know that the FP constant fits in the range [SMIN, SMAX] or
5505 // [0, UMAX], but it may still be fractional. See if it is fractional by
5506 // casting the FP value to the integer value and back, checking for equality.
5507 // Don't do this for zero, because -0.0 is not fractional.
5508 Constant *RHSInt = ConstantExpr::getFPToSI(RHSC, IntTy);
5509 if (!RHS.isZero() &&
5510 ConstantExpr::getSIToFP(RHSInt, RHSC->getType()) != RHSC) {
5511 // If we had a comparison against a fractional value, we have to adjust the
5512 // compare predicate and sometimes the value. RHSC is rounded towards zero
5515 default: assert(0 && "Unexpected integer comparison!");
5516 case ICmpInst::ICMP_NE: // (float)int != 4.4 --> true
5517 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
5518 case ICmpInst::ICMP_EQ: // (float)int == 4.4 --> false
5519 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
5520 case ICmpInst::ICMP_ULE:
5521 // (float)int <= 4.4 --> int <= 4
5522 // (float)int <= -4.4 --> false
5523 if (RHS.isNegative())
5524 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
5526 case ICmpInst::ICMP_SLE:
5527 // (float)int <= 4.4 --> int <= 4
5528 // (float)int <= -4.4 --> int < -4
5529 if (RHS.isNegative())
5530 Pred = ICmpInst::ICMP_SLT;
5532 case ICmpInst::ICMP_ULT:
5533 // (float)int < -4.4 --> false
5534 // (float)int < 4.4 --> int <= 4
5535 if (RHS.isNegative())
5536 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
5537 Pred = ICmpInst::ICMP_ULE;
5539 case ICmpInst::ICMP_SLT:
5540 // (float)int < -4.4 --> int < -4
5541 // (float)int < 4.4 --> int <= 4
5542 if (!RHS.isNegative())
5543 Pred = ICmpInst::ICMP_SLE;
5545 case ICmpInst::ICMP_UGT:
5546 // (float)int > 4.4 --> int > 4
5547 // (float)int > -4.4 --> true
5548 if (RHS.isNegative())
5549 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
5551 case ICmpInst::ICMP_SGT:
5552 // (float)int > 4.4 --> int > 4
5553 // (float)int > -4.4 --> int >= -4
5554 if (RHS.isNegative())
5555 Pred = ICmpInst::ICMP_SGE;
5557 case ICmpInst::ICMP_UGE:
5558 // (float)int >= -4.4 --> true
5559 // (float)int >= 4.4 --> int > 4
5560 if (!RHS.isNegative())
5561 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
5562 Pred = ICmpInst::ICMP_UGT;
5564 case ICmpInst::ICMP_SGE:
5565 // (float)int >= -4.4 --> int >= -4
5566 // (float)int >= 4.4 --> int > 4
5567 if (!RHS.isNegative())
5568 Pred = ICmpInst::ICMP_SGT;
5573 // Lower this FP comparison into an appropriate integer version of the
5575 return new ICmpInst(Pred, LHSI->getOperand(0), RHSInt);
5578 Instruction *InstCombiner::visitFCmpInst(FCmpInst &I) {
5579 bool Changed = SimplifyCompare(I);
5580 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5582 // Fold trivial predicates.
5583 if (I.getPredicate() == FCmpInst::FCMP_FALSE)
5584 return ReplaceInstUsesWith(I, Constant::getNullValue(Type::Int1Ty));
5585 if (I.getPredicate() == FCmpInst::FCMP_TRUE)
5586 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
5588 // Simplify 'fcmp pred X, X'
5590 switch (I.getPredicate()) {
5591 default: assert(0 && "Unknown predicate!");
5592 case FCmpInst::FCMP_UEQ: // True if unordered or equal
5593 case FCmpInst::FCMP_UGE: // True if unordered, greater than, or equal
5594 case FCmpInst::FCMP_ULE: // True if unordered, less than, or equal
5595 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
5596 case FCmpInst::FCMP_OGT: // True if ordered and greater than
5597 case FCmpInst::FCMP_OLT: // True if ordered and less than
5598 case FCmpInst::FCMP_ONE: // True if ordered and operands are unequal
5599 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
5601 case FCmpInst::FCMP_UNO: // True if unordered: isnan(X) | isnan(Y)
5602 case FCmpInst::FCMP_ULT: // True if unordered or less than
5603 case FCmpInst::FCMP_UGT: // True if unordered or greater than
5604 case FCmpInst::FCMP_UNE: // True if unordered or not equal
5605 // Canonicalize these to be 'fcmp uno %X, 0.0'.
5606 I.setPredicate(FCmpInst::FCMP_UNO);
5607 I.setOperand(1, Constant::getNullValue(Op0->getType()));
5610 case FCmpInst::FCMP_ORD: // True if ordered (no nans)
5611 case FCmpInst::FCMP_OEQ: // True if ordered and equal
5612 case FCmpInst::FCMP_OGE: // True if ordered and greater than or equal
5613 case FCmpInst::FCMP_OLE: // True if ordered and less than or equal
5614 // Canonicalize these to be 'fcmp ord %X, 0.0'.
5615 I.setPredicate(FCmpInst::FCMP_ORD);
5616 I.setOperand(1, Constant::getNullValue(Op0->getType()));
5621 if (isa<UndefValue>(Op1)) // fcmp pred X, undef -> undef
5622 return ReplaceInstUsesWith(I, UndefValue::get(Type::Int1Ty));
5624 // Handle fcmp with constant RHS
5625 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
5626 // If the constant is a nan, see if we can fold the comparison based on it.
5627 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHSC)) {
5628 if (CFP->getValueAPF().isNaN()) {
5629 if (FCmpInst::isOrdered(I.getPredicate())) // True if ordered and...
5630 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
5631 assert(FCmpInst::isUnordered(I.getPredicate()) &&
5632 "Comparison must be either ordered or unordered!");
5633 // True if unordered.
5634 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
5638 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
5639 switch (LHSI->getOpcode()) {
5640 case Instruction::PHI:
5641 // Only fold fcmp into the PHI if the phi and fcmp are in the same
5642 // block. If in the same block, we're encouraging jump threading. If
5643 // not, we are just pessimizing the code by making an i1 phi.
5644 if (LHSI->getParent() == I.getParent())
5645 if (Instruction *NV = FoldOpIntoPhi(I))
5648 case Instruction::SIToFP:
5649 case Instruction::UIToFP:
5650 if (Instruction *NV = FoldFCmp_IntToFP_Cst(I, LHSI, RHSC))
5653 case Instruction::Select:
5654 // If either operand of the select is a constant, we can fold the
5655 // comparison into the select arms, which will cause one to be
5656 // constant folded and the select turned into a bitwise or.
5657 Value *Op1 = 0, *Op2 = 0;
5658 if (LHSI->hasOneUse()) {
5659 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
5660 // Fold the known value into the constant operand.
5661 Op1 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
5662 // Insert a new FCmp of the other select operand.
5663 Op2 = InsertNewInstBefore(new FCmpInst(I.getPredicate(),
5664 LHSI->getOperand(2), RHSC,
5666 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
5667 // Fold the known value into the constant operand.
5668 Op2 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
5669 // Insert a new FCmp of the other select operand.
5670 Op1 = InsertNewInstBefore(new FCmpInst(I.getPredicate(),
5671 LHSI->getOperand(1), RHSC,
5677 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
5682 return Changed ? &I : 0;
5685 Instruction *InstCombiner::visitICmpInst(ICmpInst &I) {
5686 bool Changed = SimplifyCompare(I);
5687 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5688 const Type *Ty = Op0->getType();
5692 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
5693 I.isTrueWhenEqual()));
5695 if (isa<UndefValue>(Op1)) // X icmp undef -> undef
5696 return ReplaceInstUsesWith(I, UndefValue::get(Type::Int1Ty));
5698 // icmp <global/alloca*/null>, <global/alloca*/null> - Global/Stack value
5699 // addresses never equal each other! We already know that Op0 != Op1.
5700 if ((isa<GlobalValue>(Op0) || isa<AllocaInst>(Op0) ||
5701 isa<ConstantPointerNull>(Op0)) &&
5702 (isa<GlobalValue>(Op1) || isa<AllocaInst>(Op1) ||
5703 isa<ConstantPointerNull>(Op1)))
5704 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
5705 !I.isTrueWhenEqual()));
5707 // icmp's with boolean values can always be turned into bitwise operations
5708 if (Ty == Type::Int1Ty) {
5709 switch (I.getPredicate()) {
5710 default: assert(0 && "Invalid icmp instruction!");
5711 case ICmpInst::ICMP_EQ: { // icmp eq i1 A, B -> ~(A^B)
5712 Instruction *Xor = BinaryOperator::CreateXor(Op0, Op1, I.getName()+"tmp");
5713 InsertNewInstBefore(Xor, I);
5714 return BinaryOperator::CreateNot(Xor);
5716 case ICmpInst::ICMP_NE: // icmp eq i1 A, B -> A^B
5717 return BinaryOperator::CreateXor(Op0, Op1);
5719 case ICmpInst::ICMP_UGT:
5720 std::swap(Op0, Op1); // Change icmp ugt -> icmp ult
5722 case ICmpInst::ICMP_ULT:{ // icmp ult i1 A, B -> ~A & B
5723 Instruction *Not = BinaryOperator::CreateNot(Op0, I.getName()+"tmp");
5724 InsertNewInstBefore(Not, I);
5725 return BinaryOperator::CreateAnd(Not, Op1);
5727 case ICmpInst::ICMP_SGT:
5728 std::swap(Op0, Op1); // Change icmp sgt -> icmp slt
5730 case ICmpInst::ICMP_SLT: { // icmp slt i1 A, B -> A & ~B
5731 Instruction *Not = BinaryOperator::CreateNot(Op1, I.getName()+"tmp");
5732 InsertNewInstBefore(Not, I);
5733 return BinaryOperator::CreateAnd(Not, Op0);
5735 case ICmpInst::ICMP_UGE:
5736 std::swap(Op0, Op1); // Change icmp uge -> icmp ule
5738 case ICmpInst::ICMP_ULE: { // icmp ule i1 A, B -> ~A | B
5739 Instruction *Not = BinaryOperator::CreateNot(Op0, I.getName()+"tmp");
5740 InsertNewInstBefore(Not, I);
5741 return BinaryOperator::CreateOr(Not, Op1);
5743 case ICmpInst::ICMP_SGE:
5744 std::swap(Op0, Op1); // Change icmp sge -> icmp sle
5746 case ICmpInst::ICMP_SLE: { // icmp sle i1 A, B -> A | ~B
5747 Instruction *Not = BinaryOperator::CreateNot(Op1, I.getName()+"tmp");
5748 InsertNewInstBefore(Not, I);
5749 return BinaryOperator::CreateOr(Not, Op0);
5754 // See if we are doing a comparison with a constant.
5755 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
5758 // (icmp ne/eq (sub A B) 0) -> (icmp ne/eq A, B)
5759 if (I.isEquality() && CI->isNullValue() &&
5760 match(Op0, m_Sub(m_Value(A), m_Value(B)))) {
5761 // (icmp cond A B) if cond is equality
5762 return new ICmpInst(I.getPredicate(), A, B);
5765 // If we have an icmp le or icmp ge instruction, turn it into the
5766 // appropriate icmp lt or icmp gt instruction. This allows us to rely on
5767 // them being folded in the code below.
5768 switch (I.getPredicate()) {
5770 case ICmpInst::ICMP_ULE:
5771 if (CI->isMaxValue(false)) // A <=u MAX -> TRUE
5772 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5773 return new ICmpInst(ICmpInst::ICMP_ULT, Op0, AddOne(CI));
5774 case ICmpInst::ICMP_SLE:
5775 if (CI->isMaxValue(true)) // A <=s MAX -> TRUE
5776 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5777 return new ICmpInst(ICmpInst::ICMP_SLT, Op0, AddOne(CI));
5778 case ICmpInst::ICMP_UGE:
5779 if (CI->isMinValue(false)) // A >=u MIN -> TRUE
5780 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5781 return new ICmpInst( ICmpInst::ICMP_UGT, Op0, SubOne(CI));
5782 case ICmpInst::ICMP_SGE:
5783 if (CI->isMinValue(true)) // A >=s MIN -> TRUE
5784 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5785 return new ICmpInst(ICmpInst::ICMP_SGT, Op0, SubOne(CI));
5788 // See if we can fold the comparison based on range information we can get
5789 // by checking whether bits are known to be zero or one in the input.
5790 uint32_t BitWidth = cast<IntegerType>(Ty)->getBitWidth();
5791 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
5793 // If this comparison is a normal comparison, it demands all
5794 // bits, if it is a sign bit comparison, it only demands the sign bit.
5796 bool isSignBit = isSignBitCheck(I.getPredicate(), CI, UnusedBit);
5798 if (SimplifyDemandedBits(Op0,
5799 isSignBit ? APInt::getSignBit(BitWidth)
5800 : APInt::getAllOnesValue(BitWidth),
5801 KnownZero, KnownOne, 0))
5804 // Given the known and unknown bits, compute a range that the LHS could be
5805 // in. Compute the Min, Max and RHS values based on the known bits. For the
5806 // EQ and NE we use unsigned values.
5807 APInt Min(BitWidth, 0), Max(BitWidth, 0);
5808 if (ICmpInst::isSignedPredicate(I.getPredicate()))
5809 ComputeSignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne, Min, Max);
5811 ComputeUnsignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne,Min,Max);
5813 // If Min and Max are known to be the same, then SimplifyDemandedBits
5814 // figured out that the LHS is a constant. Just constant fold this now so
5815 // that code below can assume that Min != Max.
5817 return ReplaceInstUsesWith(I, ConstantExpr::getICmp(I.getPredicate(),
5818 ConstantInt::get(Min),
5821 // Based on the range information we know about the LHS, see if we can
5822 // simplify this comparison. For example, (x&4) < 8 is always true.
5823 const APInt &RHSVal = CI->getValue();
5824 switch (I.getPredicate()) { // LE/GE have been folded already.
5825 default: assert(0 && "Unknown icmp opcode!");
5826 case ICmpInst::ICMP_EQ:
5827 if (Max.ult(RHSVal) || Min.ugt(RHSVal))
5828 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5830 case ICmpInst::ICMP_NE:
5831 if (Max.ult(RHSVal) || Min.ugt(RHSVal))
5832 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5834 case ICmpInst::ICMP_ULT:
5835 if (Max.ult(RHSVal)) // A <u C -> true iff max(A) < C
5836 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5837 if (Min.uge(RHSVal)) // A <u C -> false iff min(A) >= C
5838 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5839 if (RHSVal == Max) // A <u MAX -> A != MAX
5840 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5841 if (RHSVal == Min+1) // A <u MIN+1 -> A == MIN
5842 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, SubOne(CI));
5844 // (x <u 2147483648) -> (x >s -1) -> true if sign bit clear
5845 if (CI->isMinValue(true))
5846 return new ICmpInst(ICmpInst::ICMP_SGT, Op0,
5847 ConstantInt::getAllOnesValue(Op0->getType()));
5849 case ICmpInst::ICMP_UGT:
5850 if (Min.ugt(RHSVal)) // A >u C -> true iff min(A) > C
5851 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5852 if (Max.ule(RHSVal)) // A >u C -> false iff max(A) <= C
5853 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5855 if (RHSVal == Min) // A >u MIN -> A != MIN
5856 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5857 if (RHSVal == Max-1) // A >u MAX-1 -> A == MAX
5858 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, AddOne(CI));
5860 // (x >u 2147483647) -> (x <s 0) -> true if sign bit set
5861 if (CI->isMaxValue(true))
5862 return new ICmpInst(ICmpInst::ICMP_SLT, Op0,
5863 ConstantInt::getNullValue(Op0->getType()));
5865 case ICmpInst::ICMP_SLT:
5866 if (Max.slt(RHSVal)) // A <s C -> true iff max(A) < C
5867 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5868 if (Min.sge(RHSVal)) // A <s C -> false iff min(A) >= C
5869 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5870 if (RHSVal == Max) // A <s MAX -> A != MAX
5871 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5872 if (RHSVal == Min+1) // A <s MIN+1 -> A == MIN
5873 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, SubOne(CI));
5875 case ICmpInst::ICMP_SGT:
5876 if (Min.sgt(RHSVal)) // A >s C -> true iff min(A) > C
5877 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5878 if (Max.sle(RHSVal)) // A >s C -> false iff max(A) <= C
5879 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5881 if (RHSVal == Min) // A >s MIN -> A != MIN
5882 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5883 if (RHSVal == Max-1) // A >s MAX-1 -> A == MAX
5884 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, AddOne(CI));
5889 // Test if the ICmpInst instruction is used exclusively by a select as
5890 // part of a minimum or maximum operation. If so, refrain from doing
5891 // any other folding. This helps out other analyses which understand
5892 // non-obfuscated minimum and maximum idioms, such as ScalarEvolution
5893 // and CodeGen. And in this case, at least one of the comparison
5894 // operands has at least one user besides the compare (the select),
5895 // which would often largely negate the benefit of folding anyway.
5897 if (SelectInst *SI = dyn_cast<SelectInst>(*I.use_begin()))
5898 if ((SI->getOperand(1) == Op0 && SI->getOperand(2) == Op1) ||
5899 (SI->getOperand(2) == Op0 && SI->getOperand(1) == Op1))
5902 // See if we are doing a comparison between a constant and an instruction that
5903 // can be folded into the comparison.
5904 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
5905 // Since the RHS is a ConstantInt (CI), if the left hand side is an
5906 // instruction, see if that instruction also has constants so that the
5907 // instruction can be folded into the icmp
5908 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
5909 if (Instruction *Res = visitICmpInstWithInstAndIntCst(I, LHSI, CI))
5913 // Handle icmp with constant (but not simple integer constant) RHS
5914 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
5915 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
5916 switch (LHSI->getOpcode()) {
5917 case Instruction::GetElementPtr:
5918 if (RHSC->isNullValue()) {
5919 // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
5920 bool isAllZeros = true;
5921 for (unsigned i = 1, e = LHSI->getNumOperands(); i != e; ++i)
5922 if (!isa<Constant>(LHSI->getOperand(i)) ||
5923 !cast<Constant>(LHSI->getOperand(i))->isNullValue()) {
5928 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
5929 Constant::getNullValue(LHSI->getOperand(0)->getType()));
5933 case Instruction::PHI:
5934 // Only fold icmp into the PHI if the phi and fcmp are in the same
5935 // block. If in the same block, we're encouraging jump threading. If
5936 // not, we are just pessimizing the code by making an i1 phi.
5937 if (LHSI->getParent() == I.getParent())
5938 if (Instruction *NV = FoldOpIntoPhi(I))
5941 case Instruction::Select: {
5942 // If either operand of the select is a constant, we can fold the
5943 // comparison into the select arms, which will cause one to be
5944 // constant folded and the select turned into a bitwise or.
5945 Value *Op1 = 0, *Op2 = 0;
5946 if (LHSI->hasOneUse()) {
5947 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
5948 // Fold the known value into the constant operand.
5949 Op1 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
5950 // Insert a new ICmp of the other select operand.
5951 Op2 = InsertNewInstBefore(new ICmpInst(I.getPredicate(),
5952 LHSI->getOperand(2), RHSC,
5954 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
5955 // Fold the known value into the constant operand.
5956 Op2 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
5957 // Insert a new ICmp of the other select operand.
5958 Op1 = InsertNewInstBefore(new ICmpInst(I.getPredicate(),
5959 LHSI->getOperand(1), RHSC,
5965 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
5968 case Instruction::Malloc:
5969 // If we have (malloc != null), and if the malloc has a single use, we
5970 // can assume it is successful and remove the malloc.
5971 if (LHSI->hasOneUse() && isa<ConstantPointerNull>(RHSC)) {
5972 AddToWorkList(LHSI);
5973 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
5974 !I.isTrueWhenEqual()));
5980 // If we can optimize a 'icmp GEP, P' or 'icmp P, GEP', do so now.
5981 if (User *GEP = dyn_castGetElementPtr(Op0))
5982 if (Instruction *NI = FoldGEPICmp(GEP, Op1, I.getPredicate(), I))
5984 if (User *GEP = dyn_castGetElementPtr(Op1))
5985 if (Instruction *NI = FoldGEPICmp(GEP, Op0,
5986 ICmpInst::getSwappedPredicate(I.getPredicate()), I))
5989 // Test to see if the operands of the icmp are casted versions of other
5990 // values. If the ptr->ptr cast can be stripped off both arguments, we do so
5992 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op0)) {
5993 if (isa<PointerType>(Op0->getType()) &&
5994 (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
5995 // We keep moving the cast from the left operand over to the right
5996 // operand, where it can often be eliminated completely.
5997 Op0 = CI->getOperand(0);
5999 // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast
6000 // so eliminate it as well.
6001 if (BitCastInst *CI2 = dyn_cast<BitCastInst>(Op1))
6002 Op1 = CI2->getOperand(0);
6004 // If Op1 is a constant, we can fold the cast into the constant.
6005 if (Op0->getType() != Op1->getType()) {
6006 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
6007 Op1 = ConstantExpr::getBitCast(Op1C, Op0->getType());
6009 // Otherwise, cast the RHS right before the icmp
6010 Op1 = InsertBitCastBefore(Op1, Op0->getType(), I);
6013 return new ICmpInst(I.getPredicate(), Op0, Op1);
6017 if (isa<CastInst>(Op0)) {
6018 // Handle the special case of: icmp (cast bool to X), <cst>
6019 // This comes up when you have code like
6022 // For generality, we handle any zero-extension of any operand comparison
6023 // with a constant or another cast from the same type.
6024 if (isa<ConstantInt>(Op1) || isa<CastInst>(Op1))
6025 if (Instruction *R = visitICmpInstWithCastAndCast(I))
6029 // See if it's the same type of instruction on the left and right.
6030 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
6031 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
6032 if (Op0I->getOpcode() == Op1I->getOpcode() && Op0I->hasOneUse() &&
6033 Op1I->hasOneUse() && Op0I->getOperand(1) == Op1I->getOperand(1) &&
6035 switch (Op0I->getOpcode()) {
6037 case Instruction::Add:
6038 case Instruction::Sub:
6039 case Instruction::Xor:
6040 // a+x icmp eq/ne b+x --> a icmp b
6041 return new ICmpInst(I.getPredicate(), Op0I->getOperand(0),
6042 Op1I->getOperand(0));
6044 case Instruction::Mul:
6045 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) {
6046 // a * Cst icmp eq/ne b * Cst --> a & Mask icmp b & Mask
6047 // Mask = -1 >> count-trailing-zeros(Cst).
6048 if (!CI->isZero() && !CI->isOne()) {
6049 const APInt &AP = CI->getValue();
6050 ConstantInt *Mask = ConstantInt::get(
6051 APInt::getLowBitsSet(AP.getBitWidth(),
6053 AP.countTrailingZeros()));
6054 Instruction *And1 = BinaryOperator::CreateAnd(Op0I->getOperand(0),
6056 Instruction *And2 = BinaryOperator::CreateAnd(Op1I->getOperand(0),
6058 InsertNewInstBefore(And1, I);
6059 InsertNewInstBefore(And2, I);
6060 return new ICmpInst(I.getPredicate(), And1, And2);
6069 // ~x < ~y --> y < x
6071 if (match(Op0, m_Not(m_Value(A))) &&
6072 match(Op1, m_Not(m_Value(B))))
6073 return new ICmpInst(I.getPredicate(), B, A);
6076 if (I.isEquality()) {
6077 Value *A, *B, *C, *D;
6079 // -x == -y --> x == y
6080 if (match(Op0, m_Neg(m_Value(A))) &&
6081 match(Op1, m_Neg(m_Value(B))))
6082 return new ICmpInst(I.getPredicate(), A, B);
6084 if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
6085 if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0
6086 Value *OtherVal = A == Op1 ? B : A;
6087 return new ICmpInst(I.getPredicate(), OtherVal,
6088 Constant::getNullValue(A->getType()));
6091 if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) {
6092 // A^c1 == C^c2 --> A == C^(c1^c2)
6093 ConstantInt *C1, *C2;
6094 if (match(B, m_ConstantInt(C1)) &&
6095 match(D, m_ConstantInt(C2)) && Op1->hasOneUse()) {
6096 Constant *NC = ConstantInt::get(C1->getValue() ^ C2->getValue());
6097 Instruction *Xor = BinaryOperator::CreateXor(C, NC, "tmp");
6098 return new ICmpInst(I.getPredicate(), A,
6099 InsertNewInstBefore(Xor, I));
6102 // A^B == A^D -> B == D
6103 if (A == C) return new ICmpInst(I.getPredicate(), B, D);
6104 if (A == D) return new ICmpInst(I.getPredicate(), B, C);
6105 if (B == C) return new ICmpInst(I.getPredicate(), A, D);
6106 if (B == D) return new ICmpInst(I.getPredicate(), A, C);
6110 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
6111 (A == Op0 || B == Op0)) {
6112 // A == (A^B) -> B == 0
6113 Value *OtherVal = A == Op0 ? B : A;
6114 return new ICmpInst(I.getPredicate(), OtherVal,
6115 Constant::getNullValue(A->getType()));
6118 // (A-B) == A -> B == 0
6119 if (match(Op0, m_Sub(m_Specific(Op1), m_Value(B))))
6120 return new ICmpInst(I.getPredicate(), B,
6121 Constant::getNullValue(B->getType()));
6123 // A == (A-B) -> B == 0
6124 if (match(Op1, m_Sub(m_Specific(Op0), m_Value(B))))
6125 return new ICmpInst(I.getPredicate(), B,
6126 Constant::getNullValue(B->getType()));
6128 // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
6129 if (Op0->hasOneUse() && Op1->hasOneUse() &&
6130 match(Op0, m_And(m_Value(A), m_Value(B))) &&
6131 match(Op1, m_And(m_Value(C), m_Value(D)))) {
6132 Value *X = 0, *Y = 0, *Z = 0;
6135 X = B; Y = D; Z = A;
6136 } else if (A == D) {
6137 X = B; Y = C; Z = A;
6138 } else if (B == C) {
6139 X = A; Y = D; Z = B;
6140 } else if (B == D) {
6141 X = A; Y = C; Z = B;
6144 if (X) { // Build (X^Y) & Z
6145 Op1 = InsertNewInstBefore(BinaryOperator::CreateXor(X, Y, "tmp"), I);
6146 Op1 = InsertNewInstBefore(BinaryOperator::CreateAnd(Op1, Z, "tmp"), I);
6147 I.setOperand(0, Op1);
6148 I.setOperand(1, Constant::getNullValue(Op1->getType()));
6153 return Changed ? &I : 0;
6157 /// FoldICmpDivCst - Fold "icmp pred, ([su]div X, DivRHS), CmpRHS" where DivRHS
6158 /// and CmpRHS are both known to be integer constants.
6159 Instruction *InstCombiner::FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
6160 ConstantInt *DivRHS) {
6161 ConstantInt *CmpRHS = cast<ConstantInt>(ICI.getOperand(1));
6162 const APInt &CmpRHSV = CmpRHS->getValue();
6164 // FIXME: If the operand types don't match the type of the divide
6165 // then don't attempt this transform. The code below doesn't have the
6166 // logic to deal with a signed divide and an unsigned compare (and
6167 // vice versa). This is because (x /s C1) <s C2 produces different
6168 // results than (x /s C1) <u C2 or (x /u C1) <s C2 or even
6169 // (x /u C1) <u C2. Simply casting the operands and result won't
6170 // work. :( The if statement below tests that condition and bails
6172 bool DivIsSigned = DivI->getOpcode() == Instruction::SDiv;
6173 if (!ICI.isEquality() && DivIsSigned != ICI.isSignedPredicate())
6175 if (DivRHS->isZero())
6176 return 0; // The ProdOV computation fails on divide by zero.
6177 if (DivIsSigned && DivRHS->isAllOnesValue())
6178 return 0; // The overflow computation also screws up here
6179 if (DivRHS->isOne())
6180 return 0; // Not worth bothering, and eliminates some funny cases
6183 // Compute Prod = CI * DivRHS. We are essentially solving an equation
6184 // of form X/C1=C2. We solve for X by multiplying C1 (DivRHS) and
6185 // C2 (CI). By solving for X we can turn this into a range check
6186 // instead of computing a divide.
6187 ConstantInt *Prod = Multiply(CmpRHS, DivRHS);
6189 // Determine if the product overflows by seeing if the product is
6190 // not equal to the divide. Make sure we do the same kind of divide
6191 // as in the LHS instruction that we're folding.
6192 bool ProdOV = (DivIsSigned ? ConstantExpr::getSDiv(Prod, DivRHS) :
6193 ConstantExpr::getUDiv(Prod, DivRHS)) != CmpRHS;
6195 // Get the ICmp opcode
6196 ICmpInst::Predicate Pred = ICI.getPredicate();
6198 // Figure out the interval that is being checked. For example, a comparison
6199 // like "X /u 5 == 0" is really checking that X is in the interval [0, 5).
6200 // Compute this interval based on the constants involved and the signedness of
6201 // the compare/divide. This computes a half-open interval, keeping track of
6202 // whether either value in the interval overflows. After analysis each
6203 // overflow variable is set to 0 if it's corresponding bound variable is valid
6204 // -1 if overflowed off the bottom end, or +1 if overflowed off the top end.
6205 int LoOverflow = 0, HiOverflow = 0;
6206 ConstantInt *LoBound = 0, *HiBound = 0;
6208 if (!DivIsSigned) { // udiv
6209 // e.g. X/5 op 3 --> [15, 20)
6211 HiOverflow = LoOverflow = ProdOV;
6213 HiOverflow = AddWithOverflow(HiBound, LoBound, DivRHS, false);
6214 } else if (DivRHS->getValue().isStrictlyPositive()) { // Divisor is > 0.
6215 if (CmpRHSV == 0) { // (X / pos) op 0
6216 // Can't overflow. e.g. X/2 op 0 --> [-1, 2)
6217 LoBound = cast<ConstantInt>(ConstantExpr::getNeg(SubOne(DivRHS)));
6219 } else if (CmpRHSV.isStrictlyPositive()) { // (X / pos) op pos
6220 LoBound = Prod; // e.g. X/5 op 3 --> [15, 20)
6221 HiOverflow = LoOverflow = ProdOV;
6223 HiOverflow = AddWithOverflow(HiBound, Prod, DivRHS, true);
6224 } else { // (X / pos) op neg
6225 // e.g. X/5 op -3 --> [-15-4, -15+1) --> [-19, -14)
6226 HiBound = AddOne(Prod);
6227 LoOverflow = HiOverflow = ProdOV ? -1 : 0;
6229 ConstantInt* DivNeg = cast<ConstantInt>(ConstantExpr::getNeg(DivRHS));
6230 LoOverflow = AddWithOverflow(LoBound, HiBound, DivNeg,
6234 } else if (DivRHS->getValue().isNegative()) { // Divisor is < 0.
6235 if (CmpRHSV == 0) { // (X / neg) op 0
6236 // e.g. X/-5 op 0 --> [-4, 5)
6237 LoBound = AddOne(DivRHS);
6238 HiBound = cast<ConstantInt>(ConstantExpr::getNeg(DivRHS));
6239 if (HiBound == DivRHS) { // -INTMIN = INTMIN
6240 HiOverflow = 1; // [INTMIN+1, overflow)
6241 HiBound = 0; // e.g. X/INTMIN = 0 --> X > INTMIN
6243 } else if (CmpRHSV.isStrictlyPositive()) { // (X / neg) op pos
6244 // e.g. X/-5 op 3 --> [-19, -14)
6245 HiBound = AddOne(Prod);
6246 HiOverflow = LoOverflow = ProdOV ? -1 : 0;
6248 LoOverflow = AddWithOverflow(LoBound, HiBound, DivRHS, true) ? -1 : 0;
6249 } else { // (X / neg) op neg
6250 LoBound = Prod; // e.g. X/-5 op -3 --> [15, 20)
6251 LoOverflow = HiOverflow = ProdOV;
6253 HiOverflow = SubWithOverflow(HiBound, Prod, DivRHS, true);
6256 // Dividing by a negative swaps the condition. LT <-> GT
6257 Pred = ICmpInst::getSwappedPredicate(Pred);
6260 Value *X = DivI->getOperand(0);
6262 default: assert(0 && "Unhandled icmp opcode!");
6263 case ICmpInst::ICMP_EQ:
6264 if (LoOverflow && HiOverflow)
6265 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
6266 else if (HiOverflow)
6267 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
6268 ICmpInst::ICMP_UGE, X, LoBound);
6269 else if (LoOverflow)
6270 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
6271 ICmpInst::ICMP_ULT, X, HiBound);
6273 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned, true, ICI);
6274 case ICmpInst::ICMP_NE:
6275 if (LoOverflow && HiOverflow)
6276 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
6277 else if (HiOverflow)
6278 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
6279 ICmpInst::ICMP_ULT, X, LoBound);
6280 else if (LoOverflow)
6281 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
6282 ICmpInst::ICMP_UGE, X, HiBound);
6284 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned, false, ICI);
6285 case ICmpInst::ICMP_ULT:
6286 case ICmpInst::ICMP_SLT:
6287 if (LoOverflow == +1) // Low bound is greater than input range.
6288 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
6289 if (LoOverflow == -1) // Low bound is less than input range.
6290 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
6291 return new ICmpInst(Pred, X, LoBound);
6292 case ICmpInst::ICMP_UGT:
6293 case ICmpInst::ICMP_SGT:
6294 if (HiOverflow == +1) // High bound greater than input range.
6295 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
6296 else if (HiOverflow == -1) // High bound less than input range.
6297 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
6298 if (Pred == ICmpInst::ICMP_UGT)
6299 return new ICmpInst(ICmpInst::ICMP_UGE, X, HiBound);
6301 return new ICmpInst(ICmpInst::ICMP_SGE, X, HiBound);
6306 /// visitICmpInstWithInstAndIntCst - Handle "icmp (instr, intcst)".
6308 Instruction *InstCombiner::visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
6311 const APInt &RHSV = RHS->getValue();
6313 switch (LHSI->getOpcode()) {
6314 case Instruction::Xor: // (icmp pred (xor X, XorCST), CI)
6315 if (ConstantInt *XorCST = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
6316 // If this is a comparison that tests the signbit (X < 0) or (x > -1),
6318 if ((ICI.getPredicate() == ICmpInst::ICMP_SLT && RHSV == 0) ||
6319 (ICI.getPredicate() == ICmpInst::ICMP_SGT && RHSV.isAllOnesValue())) {
6320 Value *CompareVal = LHSI->getOperand(0);
6322 // If the sign bit of the XorCST is not set, there is no change to
6323 // the operation, just stop using the Xor.
6324 if (!XorCST->getValue().isNegative()) {
6325 ICI.setOperand(0, CompareVal);
6326 AddToWorkList(LHSI);
6330 // Was the old condition true if the operand is positive?
6331 bool isTrueIfPositive = ICI.getPredicate() == ICmpInst::ICMP_SGT;
6333 // If so, the new one isn't.
6334 isTrueIfPositive ^= true;
6336 if (isTrueIfPositive)
6337 return new ICmpInst(ICmpInst::ICMP_SGT, CompareVal, SubOne(RHS));
6339 return new ICmpInst(ICmpInst::ICMP_SLT, CompareVal, AddOne(RHS));
6343 case Instruction::And: // (icmp pred (and X, AndCST), RHS)
6344 if (LHSI->hasOneUse() && isa<ConstantInt>(LHSI->getOperand(1)) &&
6345 LHSI->getOperand(0)->hasOneUse()) {
6346 ConstantInt *AndCST = cast<ConstantInt>(LHSI->getOperand(1));
6348 // If the LHS is an AND of a truncating cast, we can widen the
6349 // and/compare to be the input width without changing the value
6350 // produced, eliminating a cast.
6351 if (TruncInst *Cast = dyn_cast<TruncInst>(LHSI->getOperand(0))) {
6352 // We can do this transformation if either the AND constant does not
6353 // have its sign bit set or if it is an equality comparison.
6354 // Extending a relational comparison when we're checking the sign
6355 // bit would not work.
6356 if (Cast->hasOneUse() &&
6357 (ICI.isEquality() ||
6358 (AndCST->getValue().isNonNegative() && RHSV.isNonNegative()))) {
6360 cast<IntegerType>(Cast->getOperand(0)->getType())->getBitWidth();
6361 APInt NewCST = AndCST->getValue();
6362 NewCST.zext(BitWidth);
6364 NewCI.zext(BitWidth);
6365 Instruction *NewAnd =
6366 BinaryOperator::CreateAnd(Cast->getOperand(0),
6367 ConstantInt::get(NewCST),LHSI->getName());
6368 InsertNewInstBefore(NewAnd, ICI);
6369 return new ICmpInst(ICI.getPredicate(), NewAnd,
6370 ConstantInt::get(NewCI));
6374 // If this is: (X >> C1) & C2 != C3 (where any shift and any compare
6375 // could exist), turn it into (X & (C2 << C1)) != (C3 << C1). This
6376 // happens a LOT in code produced by the C front-end, for bitfield
6378 BinaryOperator *Shift = dyn_cast<BinaryOperator>(LHSI->getOperand(0));
6379 if (Shift && !Shift->isShift())
6383 ShAmt = Shift ? dyn_cast<ConstantInt>(Shift->getOperand(1)) : 0;
6384 const Type *Ty = Shift ? Shift->getType() : 0; // Type of the shift.
6385 const Type *AndTy = AndCST->getType(); // Type of the and.
6387 // We can fold this as long as we can't shift unknown bits
6388 // into the mask. This can only happen with signed shift
6389 // rights, as they sign-extend.
6391 bool CanFold = Shift->isLogicalShift();
6393 // To test for the bad case of the signed shr, see if any
6394 // of the bits shifted in could be tested after the mask.
6395 uint32_t TyBits = Ty->getPrimitiveSizeInBits();
6396 int ShAmtVal = TyBits - ShAmt->getLimitedValue(TyBits);
6398 uint32_t BitWidth = AndTy->getPrimitiveSizeInBits();
6399 if ((APInt::getHighBitsSet(BitWidth, BitWidth-ShAmtVal) &
6400 AndCST->getValue()) == 0)
6406 if (Shift->getOpcode() == Instruction::Shl)
6407 NewCst = ConstantExpr::getLShr(RHS, ShAmt);
6409 NewCst = ConstantExpr::getShl(RHS, ShAmt);
6411 // Check to see if we are shifting out any of the bits being
6413 if (ConstantExpr::get(Shift->getOpcode(), NewCst, ShAmt) != RHS) {
6414 // If we shifted bits out, the fold is not going to work out.
6415 // As a special case, check to see if this means that the
6416 // result is always true or false now.
6417 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
6418 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
6419 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
6420 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
6422 ICI.setOperand(1, NewCst);
6423 Constant *NewAndCST;
6424 if (Shift->getOpcode() == Instruction::Shl)
6425 NewAndCST = ConstantExpr::getLShr(AndCST, ShAmt);
6427 NewAndCST = ConstantExpr::getShl(AndCST, ShAmt);
6428 LHSI->setOperand(1, NewAndCST);
6429 LHSI->setOperand(0, Shift->getOperand(0));
6430 AddToWorkList(Shift); // Shift is dead.
6431 AddUsesToWorkList(ICI);
6437 // Turn ((X >> Y) & C) == 0 into (X & (C << Y)) == 0. The later is
6438 // preferable because it allows the C<<Y expression to be hoisted out
6439 // of a loop if Y is invariant and X is not.
6440 if (Shift && Shift->hasOneUse() && RHSV == 0 &&
6441 ICI.isEquality() && !Shift->isArithmeticShift() &&
6442 isa<Instruction>(Shift->getOperand(0))) {
6445 if (Shift->getOpcode() == Instruction::LShr) {
6446 NS = BinaryOperator::CreateShl(AndCST,
6447 Shift->getOperand(1), "tmp");
6449 // Insert a logical shift.
6450 NS = BinaryOperator::CreateLShr(AndCST,
6451 Shift->getOperand(1), "tmp");
6453 InsertNewInstBefore(cast<Instruction>(NS), ICI);
6455 // Compute X & (C << Y).
6456 Instruction *NewAnd =
6457 BinaryOperator::CreateAnd(Shift->getOperand(0), NS, LHSI->getName());
6458 InsertNewInstBefore(NewAnd, ICI);
6460 ICI.setOperand(0, NewAnd);
6466 case Instruction::Shl: { // (icmp pred (shl X, ShAmt), CI)
6467 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
6470 uint32_t TypeBits = RHSV.getBitWidth();
6472 // Check that the shift amount is in range. If not, don't perform
6473 // undefined shifts. When the shift is visited it will be
6475 if (ShAmt->uge(TypeBits))
6478 if (ICI.isEquality()) {
6479 // If we are comparing against bits always shifted out, the
6480 // comparison cannot succeed.
6482 ConstantExpr::getShl(ConstantExpr::getLShr(RHS, ShAmt), ShAmt);
6483 if (Comp != RHS) {// Comparing against a bit that we know is zero.
6484 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
6485 Constant *Cst = ConstantInt::get(Type::Int1Ty, IsICMP_NE);
6486 return ReplaceInstUsesWith(ICI, Cst);
6489 if (LHSI->hasOneUse()) {
6490 // Otherwise strength reduce the shift into an and.
6491 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
6493 ConstantInt::get(APInt::getLowBitsSet(TypeBits, TypeBits-ShAmtVal));
6496 BinaryOperator::CreateAnd(LHSI->getOperand(0),
6497 Mask, LHSI->getName()+".mask");
6498 Value *And = InsertNewInstBefore(AndI, ICI);
6499 return new ICmpInst(ICI.getPredicate(), And,
6500 ConstantInt::get(RHSV.lshr(ShAmtVal)));
6504 // Otherwise, if this is a comparison of the sign bit, simplify to and/test.
6505 bool TrueIfSigned = false;
6506 if (LHSI->hasOneUse() &&
6507 isSignBitCheck(ICI.getPredicate(), RHS, TrueIfSigned)) {
6508 // (X << 31) <s 0 --> (X&1) != 0
6509 Constant *Mask = ConstantInt::get(APInt(TypeBits, 1) <<
6510 (TypeBits-ShAmt->getZExtValue()-1));
6512 BinaryOperator::CreateAnd(LHSI->getOperand(0),
6513 Mask, LHSI->getName()+".mask");
6514 Value *And = InsertNewInstBefore(AndI, ICI);
6516 return new ICmpInst(TrueIfSigned ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ,
6517 And, Constant::getNullValue(And->getType()));
6522 case Instruction::LShr: // (icmp pred (shr X, ShAmt), CI)
6523 case Instruction::AShr: {
6524 // Only handle equality comparisons of shift-by-constant.
6525 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
6526 if (!ShAmt || !ICI.isEquality()) break;
6528 // Check that the shift amount is in range. If not, don't perform
6529 // undefined shifts. When the shift is visited it will be
6531 uint32_t TypeBits = RHSV.getBitWidth();
6532 if (ShAmt->uge(TypeBits))
6535 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
6537 // If we are comparing against bits always shifted out, the
6538 // comparison cannot succeed.
6539 APInt Comp = RHSV << ShAmtVal;
6540 if (LHSI->getOpcode() == Instruction::LShr)
6541 Comp = Comp.lshr(ShAmtVal);
6543 Comp = Comp.ashr(ShAmtVal);
6545 if (Comp != RHSV) { // Comparing against a bit that we know is zero.
6546 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
6547 Constant *Cst = ConstantInt::get(Type::Int1Ty, IsICMP_NE);
6548 return ReplaceInstUsesWith(ICI, Cst);
6551 // Otherwise, check to see if the bits shifted out are known to be zero.
6552 // If so, we can compare against the unshifted value:
6553 // (X & 4) >> 1 == 2 --> (X & 4) == 4.
6554 if (LHSI->hasOneUse() &&
6555 MaskedValueIsZero(LHSI->getOperand(0),
6556 APInt::getLowBitsSet(Comp.getBitWidth(), ShAmtVal))) {
6557 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
6558 ConstantExpr::getShl(RHS, ShAmt));
6561 if (LHSI->hasOneUse()) {
6562 // Otherwise strength reduce the shift into an and.
6563 APInt Val(APInt::getHighBitsSet(TypeBits, TypeBits - ShAmtVal));
6564 Constant *Mask = ConstantInt::get(Val);
6567 BinaryOperator::CreateAnd(LHSI->getOperand(0),
6568 Mask, LHSI->getName()+".mask");
6569 Value *And = InsertNewInstBefore(AndI, ICI);
6570 return new ICmpInst(ICI.getPredicate(), And,
6571 ConstantExpr::getShl(RHS, ShAmt));
6576 case Instruction::SDiv:
6577 case Instruction::UDiv:
6578 // Fold: icmp pred ([us]div X, C1), C2 -> range test
6579 // Fold this div into the comparison, producing a range check.
6580 // Determine, based on the divide type, what the range is being
6581 // checked. If there is an overflow on the low or high side, remember
6582 // it, otherwise compute the range [low, hi) bounding the new value.
6583 // See: InsertRangeTest above for the kinds of replacements possible.
6584 if (ConstantInt *DivRHS = dyn_cast<ConstantInt>(LHSI->getOperand(1)))
6585 if (Instruction *R = FoldICmpDivCst(ICI, cast<BinaryOperator>(LHSI),
6590 case Instruction::Add:
6591 // Fold: icmp pred (add, X, C1), C2
6593 if (!ICI.isEquality()) {
6594 ConstantInt *LHSC = dyn_cast<ConstantInt>(LHSI->getOperand(1));
6596 const APInt &LHSV = LHSC->getValue();
6598 ConstantRange CR = ICI.makeConstantRange(ICI.getPredicate(), RHSV)
6601 if (ICI.isSignedPredicate()) {
6602 if (CR.getLower().isSignBit()) {
6603 return new ICmpInst(ICmpInst::ICMP_SLT, LHSI->getOperand(0),
6604 ConstantInt::get(CR.getUpper()));
6605 } else if (CR.getUpper().isSignBit()) {
6606 return new ICmpInst(ICmpInst::ICMP_SGE, LHSI->getOperand(0),
6607 ConstantInt::get(CR.getLower()));
6610 if (CR.getLower().isMinValue()) {
6611 return new ICmpInst(ICmpInst::ICMP_ULT, LHSI->getOperand(0),
6612 ConstantInt::get(CR.getUpper()));
6613 } else if (CR.getUpper().isMinValue()) {
6614 return new ICmpInst(ICmpInst::ICMP_UGE, LHSI->getOperand(0),
6615 ConstantInt::get(CR.getLower()));
6622 // Simplify icmp_eq and icmp_ne instructions with integer constant RHS.
6623 if (ICI.isEquality()) {
6624 bool isICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
6626 // If the first operand is (add|sub|and|or|xor|rem) with a constant, and
6627 // the second operand is a constant, simplify a bit.
6628 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(LHSI)) {
6629 switch (BO->getOpcode()) {
6630 case Instruction::SRem:
6631 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
6632 if (RHSV == 0 && isa<ConstantInt>(BO->getOperand(1)) &&BO->hasOneUse()){
6633 const APInt &V = cast<ConstantInt>(BO->getOperand(1))->getValue();
6634 if (V.sgt(APInt(V.getBitWidth(), 1)) && V.isPowerOf2()) {
6635 Instruction *NewRem =
6636 BinaryOperator::CreateURem(BO->getOperand(0), BO->getOperand(1),
6638 InsertNewInstBefore(NewRem, ICI);
6639 return new ICmpInst(ICI.getPredicate(), NewRem,
6640 Constant::getNullValue(BO->getType()));
6644 case Instruction::Add:
6645 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
6646 if (ConstantInt *BOp1C = dyn_cast<ConstantInt>(BO->getOperand(1))) {
6647 if (BO->hasOneUse())
6648 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
6649 Subtract(RHS, BOp1C));
6650 } else if (RHSV == 0) {
6651 // Replace ((add A, B) != 0) with (A != -B) if A or B is
6652 // efficiently invertible, or if the add has just this one use.
6653 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
6655 if (Value *NegVal = dyn_castNegVal(BOp1))
6656 return new ICmpInst(ICI.getPredicate(), BOp0, NegVal);
6657 else if (Value *NegVal = dyn_castNegVal(BOp0))
6658 return new ICmpInst(ICI.getPredicate(), NegVal, BOp1);
6659 else if (BO->hasOneUse()) {
6660 Instruction *Neg = BinaryOperator::CreateNeg(BOp1);
6661 InsertNewInstBefore(Neg, ICI);
6663 return new ICmpInst(ICI.getPredicate(), BOp0, Neg);
6667 case Instruction::Xor:
6668 // For the xor case, we can xor two constants together, eliminating
6669 // the explicit xor.
6670 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1)))
6671 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
6672 ConstantExpr::getXor(RHS, BOC));
6675 case Instruction::Sub:
6676 // Replace (([sub|xor] A, B) != 0) with (A != B)
6678 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
6682 case Instruction::Or:
6683 // If bits are being or'd in that are not present in the constant we
6684 // are comparing against, then the comparison could never succeed!
6685 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1))) {
6686 Constant *NotCI = ConstantExpr::getNot(RHS);
6687 if (!ConstantExpr::getAnd(BOC, NotCI)->isNullValue())
6688 return ReplaceInstUsesWith(ICI, ConstantInt::get(Type::Int1Ty,
6693 case Instruction::And:
6694 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
6695 // If bits are being compared against that are and'd out, then the
6696 // comparison can never succeed!
6697 if ((RHSV & ~BOC->getValue()) != 0)
6698 return ReplaceInstUsesWith(ICI, ConstantInt::get(Type::Int1Ty,
6701 // If we have ((X & C) == C), turn it into ((X & C) != 0).
6702 if (RHS == BOC && RHSV.isPowerOf2())
6703 return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ :
6704 ICmpInst::ICMP_NE, LHSI,
6705 Constant::getNullValue(RHS->getType()));
6707 // Replace (and X, (1 << size(X)-1) != 0) with x s< 0
6708 if (BOC->getValue().isSignBit()) {
6709 Value *X = BO->getOperand(0);
6710 Constant *Zero = Constant::getNullValue(X->getType());
6711 ICmpInst::Predicate pred = isICMP_NE ?
6712 ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
6713 return new ICmpInst(pred, X, Zero);
6716 // ((X & ~7) == 0) --> X < 8
6717 if (RHSV == 0 && isHighOnes(BOC)) {
6718 Value *X = BO->getOperand(0);
6719 Constant *NegX = ConstantExpr::getNeg(BOC);
6720 ICmpInst::Predicate pred = isICMP_NE ?
6721 ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
6722 return new ICmpInst(pred, X, NegX);
6727 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(LHSI)) {
6728 // Handle icmp {eq|ne} <intrinsic>, intcst.
6729 if (II->getIntrinsicID() == Intrinsic::bswap) {
6731 ICI.setOperand(0, II->getOperand(1));
6732 ICI.setOperand(1, ConstantInt::get(RHSV.byteSwap()));
6736 } else { // Not a ICMP_EQ/ICMP_NE
6737 // If the LHS is a cast from an integral value of the same size,
6738 // then since we know the RHS is a constant, try to simlify.
6739 if (CastInst *Cast = dyn_cast<CastInst>(LHSI)) {
6740 Value *CastOp = Cast->getOperand(0);
6741 const Type *SrcTy = CastOp->getType();
6742 uint32_t SrcTySize = SrcTy->getPrimitiveSizeInBits();
6743 if (SrcTy->isInteger() &&
6744 SrcTySize == Cast->getType()->getPrimitiveSizeInBits()) {
6745 // If this is an unsigned comparison, try to make the comparison use
6746 // smaller constant values.
6747 if (ICI.getPredicate() == ICmpInst::ICMP_ULT && RHSV.isSignBit()) {
6748 // X u< 128 => X s> -1
6749 return new ICmpInst(ICmpInst::ICMP_SGT, CastOp,
6750 ConstantInt::get(APInt::getAllOnesValue(SrcTySize)));
6751 } else if (ICI.getPredicate() == ICmpInst::ICMP_UGT &&
6752 RHSV == APInt::getSignedMaxValue(SrcTySize)) {
6753 // X u> 127 => X s< 0
6754 return new ICmpInst(ICmpInst::ICMP_SLT, CastOp,
6755 Constant::getNullValue(SrcTy));
6763 /// visitICmpInstWithCastAndCast - Handle icmp (cast x to y), (cast/cst).
6764 /// We only handle extending casts so far.
6766 Instruction *InstCombiner::visitICmpInstWithCastAndCast(ICmpInst &ICI) {
6767 const CastInst *LHSCI = cast<CastInst>(ICI.getOperand(0));
6768 Value *LHSCIOp = LHSCI->getOperand(0);
6769 const Type *SrcTy = LHSCIOp->getType();
6770 const Type *DestTy = LHSCI->getType();
6773 // Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the
6774 // integer type is the same size as the pointer type.
6775 if (LHSCI->getOpcode() == Instruction::PtrToInt &&
6776 getTargetData().getPointerSizeInBits() ==
6777 cast<IntegerType>(DestTy)->getBitWidth()) {
6779 if (Constant *RHSC = dyn_cast<Constant>(ICI.getOperand(1))) {
6780 RHSOp = ConstantExpr::getIntToPtr(RHSC, SrcTy);
6781 } else if (PtrToIntInst *RHSC = dyn_cast<PtrToIntInst>(ICI.getOperand(1))) {
6782 RHSOp = RHSC->getOperand(0);
6783 // If the pointer types don't match, insert a bitcast.
6784 if (LHSCIOp->getType() != RHSOp->getType())
6785 RHSOp = InsertBitCastBefore(RHSOp, LHSCIOp->getType(), ICI);
6789 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSOp);
6792 // The code below only handles extension cast instructions, so far.
6794 if (LHSCI->getOpcode() != Instruction::ZExt &&
6795 LHSCI->getOpcode() != Instruction::SExt)
6798 bool isSignedExt = LHSCI->getOpcode() == Instruction::SExt;
6799 bool isSignedCmp = ICI.isSignedPredicate();
6801 if (CastInst *CI = dyn_cast<CastInst>(ICI.getOperand(1))) {
6802 // Not an extension from the same type?
6803 RHSCIOp = CI->getOperand(0);
6804 if (RHSCIOp->getType() != LHSCIOp->getType())
6807 // If the signedness of the two casts doesn't agree (i.e. one is a sext
6808 // and the other is a zext), then we can't handle this.
6809 if (CI->getOpcode() != LHSCI->getOpcode())
6812 // Deal with equality cases early.
6813 if (ICI.isEquality())
6814 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
6816 // A signed comparison of sign extended values simplifies into a
6817 // signed comparison.
6818 if (isSignedCmp && isSignedExt)
6819 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
6821 // The other three cases all fold into an unsigned comparison.
6822 return new ICmpInst(ICI.getUnsignedPredicate(), LHSCIOp, RHSCIOp);
6825 // If we aren't dealing with a constant on the RHS, exit early
6826 ConstantInt *CI = dyn_cast<ConstantInt>(ICI.getOperand(1));
6830 // Compute the constant that would happen if we truncated to SrcTy then
6831 // reextended to DestTy.
6832 Constant *Res1 = ConstantExpr::getTrunc(CI, SrcTy);
6833 Constant *Res2 = ConstantExpr::getCast(LHSCI->getOpcode(), Res1, DestTy);
6835 // If the re-extended constant didn't change...
6837 // Make sure that sign of the Cmp and the sign of the Cast are the same.
6838 // For example, we might have:
6839 // %A = sext short %X to uint
6840 // %B = icmp ugt uint %A, 1330
6841 // It is incorrect to transform this into
6842 // %B = icmp ugt short %X, 1330
6843 // because %A may have negative value.
6845 // However, we allow this when the compare is EQ/NE, because they are
6847 if (isSignedExt == isSignedCmp || ICI.isEquality())
6848 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
6852 // The re-extended constant changed so the constant cannot be represented
6853 // in the shorter type. Consequently, we cannot emit a simple comparison.
6855 // First, handle some easy cases. We know the result cannot be equal at this
6856 // point so handle the ICI.isEquality() cases
6857 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
6858 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
6859 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
6860 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
6862 // Evaluate the comparison for LT (we invert for GT below). LE and GE cases
6863 // should have been folded away previously and not enter in here.
6866 // We're performing a signed comparison.
6867 if (cast<ConstantInt>(CI)->getValue().isNegative())
6868 Result = ConstantInt::getFalse(); // X < (small) --> false
6870 Result = ConstantInt::getTrue(); // X < (large) --> true
6872 // We're performing an unsigned comparison.
6874 // We're performing an unsigned comp with a sign extended value.
6875 // This is true if the input is >= 0. [aka >s -1]
6876 Constant *NegOne = ConstantInt::getAllOnesValue(SrcTy);
6877 Result = InsertNewInstBefore(new ICmpInst(ICmpInst::ICMP_SGT, LHSCIOp,
6878 NegOne, ICI.getName()), ICI);
6880 // Unsigned extend & unsigned compare -> always true.
6881 Result = ConstantInt::getTrue();
6885 // Finally, return the value computed.
6886 if (ICI.getPredicate() == ICmpInst::ICMP_ULT ||
6887 ICI.getPredicate() == ICmpInst::ICMP_SLT)
6888 return ReplaceInstUsesWith(ICI, Result);
6890 assert((ICI.getPredicate()==ICmpInst::ICMP_UGT ||
6891 ICI.getPredicate()==ICmpInst::ICMP_SGT) &&
6892 "ICmp should be folded!");
6893 if (Constant *CI = dyn_cast<Constant>(Result))
6894 return ReplaceInstUsesWith(ICI, ConstantExpr::getNot(CI));
6895 return BinaryOperator::CreateNot(Result);
6898 Instruction *InstCombiner::visitShl(BinaryOperator &I) {
6899 return commonShiftTransforms(I);
6902 Instruction *InstCombiner::visitLShr(BinaryOperator &I) {
6903 return commonShiftTransforms(I);
6906 Instruction *InstCombiner::visitAShr(BinaryOperator &I) {
6907 if (Instruction *R = commonShiftTransforms(I))
6910 Value *Op0 = I.getOperand(0);
6912 // ashr int -1, X = -1 (for any arithmetic shift rights of ~0)
6913 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
6914 if (CSI->isAllOnesValue())
6915 return ReplaceInstUsesWith(I, CSI);
6917 // See if we can turn a signed shr into an unsigned shr.
6918 if (!isa<VectorType>(I.getType()) &&
6919 MaskedValueIsZero(Op0,
6920 APInt::getSignBit(I.getType()->getPrimitiveSizeInBits())))
6921 return BinaryOperator::CreateLShr(Op0, I.getOperand(1));
6926 Instruction *InstCombiner::commonShiftTransforms(BinaryOperator &I) {
6927 assert(I.getOperand(1)->getType() == I.getOperand(0)->getType());
6928 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
6930 // shl X, 0 == X and shr X, 0 == X
6931 // shl 0, X == 0 and shr 0, X == 0
6932 if (Op1 == Constant::getNullValue(Op1->getType()) ||
6933 Op0 == Constant::getNullValue(Op0->getType()))
6934 return ReplaceInstUsesWith(I, Op0);
6936 if (isa<UndefValue>(Op0)) {
6937 if (I.getOpcode() == Instruction::AShr) // undef >>s X -> undef
6938 return ReplaceInstUsesWith(I, Op0);
6939 else // undef << X -> 0, undef >>u X -> 0
6940 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
6942 if (isa<UndefValue>(Op1)) {
6943 if (I.getOpcode() == Instruction::AShr) // X >>s undef -> X
6944 return ReplaceInstUsesWith(I, Op0);
6945 else // X << undef, X >>u undef -> 0
6946 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
6949 // Try to fold constant and into select arguments.
6950 if (isa<Constant>(Op0))
6951 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
6952 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
6955 if (ConstantInt *CUI = dyn_cast<ConstantInt>(Op1))
6956 if (Instruction *Res = FoldShiftByConstant(Op0, CUI, I))
6961 Instruction *InstCombiner::FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
6962 BinaryOperator &I) {
6963 bool isLeftShift = I.getOpcode() == Instruction::Shl;
6965 // See if we can simplify any instructions used by the instruction whose sole
6966 // purpose is to compute bits we don't care about.
6967 uint32_t TypeBits = Op0->getType()->getPrimitiveSizeInBits();
6968 APInt KnownZero(TypeBits, 0), KnownOne(TypeBits, 0);
6969 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(TypeBits),
6970 KnownZero, KnownOne))
6973 // shl uint X, 32 = 0 and shr ubyte Y, 9 = 0, ... just don't eliminate shr
6974 // of a signed value.
6976 if (Op1->uge(TypeBits)) {
6977 if (I.getOpcode() != Instruction::AShr)
6978 return ReplaceInstUsesWith(I, Constant::getNullValue(Op0->getType()));
6980 I.setOperand(1, ConstantInt::get(I.getType(), TypeBits-1));
6985 // ((X*C1) << C2) == (X * (C1 << C2))
6986 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0))
6987 if (BO->getOpcode() == Instruction::Mul && isLeftShift)
6988 if (Constant *BOOp = dyn_cast<Constant>(BO->getOperand(1)))
6989 return BinaryOperator::CreateMul(BO->getOperand(0),
6990 ConstantExpr::getShl(BOOp, Op1));
6992 // Try to fold constant and into select arguments.
6993 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
6994 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
6996 if (isa<PHINode>(Op0))
6997 if (Instruction *NV = FoldOpIntoPhi(I))
7000 // Fold shift2(trunc(shift1(x,c1)), c2) -> trunc(shift2(shift1(x,c1),c2))
7001 if (TruncInst *TI = dyn_cast<TruncInst>(Op0)) {
7002 Instruction *TrOp = dyn_cast<Instruction>(TI->getOperand(0));
7003 // If 'shift2' is an ashr, we would have to get the sign bit into a funny
7004 // place. Don't try to do this transformation in this case. Also, we
7005 // require that the input operand is a shift-by-constant so that we have
7006 // confidence that the shifts will get folded together. We could do this
7007 // xform in more cases, but it is unlikely to be profitable.
7008 if (TrOp && I.isLogicalShift() && TrOp->isShift() &&
7009 isa<ConstantInt>(TrOp->getOperand(1))) {
7010 // Okay, we'll do this xform. Make the shift of shift.
7011 Constant *ShAmt = ConstantExpr::getZExt(Op1, TrOp->getType());
7012 Instruction *NSh = BinaryOperator::Create(I.getOpcode(), TrOp, ShAmt,
7014 InsertNewInstBefore(NSh, I); // (shift2 (shift1 & 0x00FF), c2)
7016 // For logical shifts, the truncation has the effect of making the high
7017 // part of the register be zeros. Emulate this by inserting an AND to
7018 // clear the top bits as needed. This 'and' will usually be zapped by
7019 // other xforms later if dead.
7020 unsigned SrcSize = TrOp->getType()->getPrimitiveSizeInBits();
7021 unsigned DstSize = TI->getType()->getPrimitiveSizeInBits();
7022 APInt MaskV(APInt::getLowBitsSet(SrcSize, DstSize));
7024 // The mask we constructed says what the trunc would do if occurring
7025 // between the shifts. We want to know the effect *after* the second
7026 // shift. We know that it is a logical shift by a constant, so adjust the
7027 // mask as appropriate.
7028 if (I.getOpcode() == Instruction::Shl)
7029 MaskV <<= Op1->getZExtValue();
7031 assert(I.getOpcode() == Instruction::LShr && "Unknown logical shift");
7032 MaskV = MaskV.lshr(Op1->getZExtValue());
7035 Instruction *And = BinaryOperator::CreateAnd(NSh, ConstantInt::get(MaskV),
7037 InsertNewInstBefore(And, I); // shift1 & 0x00FF
7039 // Return the value truncated to the interesting size.
7040 return new TruncInst(And, I.getType());
7044 if (Op0->hasOneUse()) {
7045 if (BinaryOperator *Op0BO = dyn_cast<BinaryOperator>(Op0)) {
7046 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
7049 switch (Op0BO->getOpcode()) {
7051 case Instruction::Add:
7052 case Instruction::And:
7053 case Instruction::Or:
7054 case Instruction::Xor: {
7055 // These operators commute.
7056 // Turn (Y + (X >> C)) << C -> (X + (Y << C)) & (~0 << C)
7057 if (isLeftShift && Op0BO->getOperand(1)->hasOneUse() &&
7058 match(Op0BO->getOperand(1), m_Shr(m_Value(V1), m_Specific(Op1)))){
7059 Instruction *YS = BinaryOperator::CreateShl(
7060 Op0BO->getOperand(0), Op1,
7062 InsertNewInstBefore(YS, I); // (Y << C)
7064 BinaryOperator::Create(Op0BO->getOpcode(), YS, V1,
7065 Op0BO->getOperand(1)->getName());
7066 InsertNewInstBefore(X, I); // (X + (Y << C))
7067 uint32_t Op1Val = Op1->getLimitedValue(TypeBits);
7068 return BinaryOperator::CreateAnd(X, ConstantInt::get(
7069 APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
7072 // Turn (Y + ((X >> C) & CC)) << C -> ((X & (CC << C)) + (Y << C))
7073 Value *Op0BOOp1 = Op0BO->getOperand(1);
7074 if (isLeftShift && Op0BOOp1->hasOneUse() &&
7076 m_And(m_Shr(m_Value(V1), m_Specific(Op1)),
7077 m_ConstantInt(CC))) &&
7078 cast<BinaryOperator>(Op0BOOp1)->getOperand(0)->hasOneUse()) {
7079 Instruction *YS = BinaryOperator::CreateShl(
7080 Op0BO->getOperand(0), Op1,
7082 InsertNewInstBefore(YS, I); // (Y << C)
7084 BinaryOperator::CreateAnd(V1, ConstantExpr::getShl(CC, Op1),
7085 V1->getName()+".mask");
7086 InsertNewInstBefore(XM, I); // X & (CC << C)
7088 return BinaryOperator::Create(Op0BO->getOpcode(), YS, XM);
7093 case Instruction::Sub: {
7094 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
7095 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
7096 match(Op0BO->getOperand(0), m_Shr(m_Value(V1), m_Specific(Op1)))){
7097 Instruction *YS = BinaryOperator::CreateShl(
7098 Op0BO->getOperand(1), Op1,
7100 InsertNewInstBefore(YS, I); // (Y << C)
7102 BinaryOperator::Create(Op0BO->getOpcode(), V1, YS,
7103 Op0BO->getOperand(0)->getName());
7104 InsertNewInstBefore(X, I); // (X + (Y << C))
7105 uint32_t Op1Val = Op1->getLimitedValue(TypeBits);
7106 return BinaryOperator::CreateAnd(X, ConstantInt::get(
7107 APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
7110 // Turn (((X >> C)&CC) + Y) << C -> (X + (Y << C)) & (CC << C)
7111 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
7112 match(Op0BO->getOperand(0),
7113 m_And(m_Shr(m_Value(V1), m_Value(V2)),
7114 m_ConstantInt(CC))) && V2 == Op1 &&
7115 cast<BinaryOperator>(Op0BO->getOperand(0))
7116 ->getOperand(0)->hasOneUse()) {
7117 Instruction *YS = BinaryOperator::CreateShl(
7118 Op0BO->getOperand(1), Op1,
7120 InsertNewInstBefore(YS, I); // (Y << C)
7122 BinaryOperator::CreateAnd(V1, ConstantExpr::getShl(CC, Op1),
7123 V1->getName()+".mask");
7124 InsertNewInstBefore(XM, I); // X & (CC << C)
7126 return BinaryOperator::Create(Op0BO->getOpcode(), XM, YS);
7134 // If the operand is an bitwise operator with a constant RHS, and the
7135 // shift is the only use, we can pull it out of the shift.
7136 if (ConstantInt *Op0C = dyn_cast<ConstantInt>(Op0BO->getOperand(1))) {
7137 bool isValid = true; // Valid only for And, Or, Xor
7138 bool highBitSet = false; // Transform if high bit of constant set?
7140 switch (Op0BO->getOpcode()) {
7141 default: isValid = false; break; // Do not perform transform!
7142 case Instruction::Add:
7143 isValid = isLeftShift;
7145 case Instruction::Or:
7146 case Instruction::Xor:
7149 case Instruction::And:
7154 // If this is a signed shift right, and the high bit is modified
7155 // by the logical operation, do not perform the transformation.
7156 // The highBitSet boolean indicates the value of the high bit of
7157 // the constant which would cause it to be modified for this
7160 if (isValid && I.getOpcode() == Instruction::AShr)
7161 isValid = Op0C->getValue()[TypeBits-1] == highBitSet;
7164 Constant *NewRHS = ConstantExpr::get(I.getOpcode(), Op0C, Op1);
7166 Instruction *NewShift =
7167 BinaryOperator::Create(I.getOpcode(), Op0BO->getOperand(0), Op1);
7168 InsertNewInstBefore(NewShift, I);
7169 NewShift->takeName(Op0BO);
7171 return BinaryOperator::Create(Op0BO->getOpcode(), NewShift,
7178 // Find out if this is a shift of a shift by a constant.
7179 BinaryOperator *ShiftOp = dyn_cast<BinaryOperator>(Op0);
7180 if (ShiftOp && !ShiftOp->isShift())
7183 if (ShiftOp && isa<ConstantInt>(ShiftOp->getOperand(1))) {
7184 ConstantInt *ShiftAmt1C = cast<ConstantInt>(ShiftOp->getOperand(1));
7185 uint32_t ShiftAmt1 = ShiftAmt1C->getLimitedValue(TypeBits);
7186 uint32_t ShiftAmt2 = Op1->getLimitedValue(TypeBits);
7187 assert(ShiftAmt2 != 0 && "Should have been simplified earlier");
7188 if (ShiftAmt1 == 0) return 0; // Will be simplified in the future.
7189 Value *X = ShiftOp->getOperand(0);
7191 uint32_t AmtSum = ShiftAmt1+ShiftAmt2; // Fold into one big shift.
7192 if (AmtSum > TypeBits)
7195 const IntegerType *Ty = cast<IntegerType>(I.getType());
7197 // Check for (X << c1) << c2 and (X >> c1) >> c2
7198 if (I.getOpcode() == ShiftOp->getOpcode()) {
7199 return BinaryOperator::Create(I.getOpcode(), X,
7200 ConstantInt::get(Ty, AmtSum));
7201 } else if (ShiftOp->getOpcode() == Instruction::LShr &&
7202 I.getOpcode() == Instruction::AShr) {
7203 // ((X >>u C1) >>s C2) -> (X >>u (C1+C2)) since C1 != 0.
7204 return BinaryOperator::CreateLShr(X, ConstantInt::get(Ty, AmtSum));
7205 } else if (ShiftOp->getOpcode() == Instruction::AShr &&
7206 I.getOpcode() == Instruction::LShr) {
7207 // ((X >>s C1) >>u C2) -> ((X >>s (C1+C2)) & mask) since C1 != 0.
7208 Instruction *Shift =
7209 BinaryOperator::CreateAShr(X, ConstantInt::get(Ty, AmtSum));
7210 InsertNewInstBefore(Shift, I);
7212 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
7213 return BinaryOperator::CreateAnd(Shift, ConstantInt::get(Mask));
7216 // Okay, if we get here, one shift must be left, and the other shift must be
7217 // right. See if the amounts are equal.
7218 if (ShiftAmt1 == ShiftAmt2) {
7219 // If we have ((X >>? C) << C), turn this into X & (-1 << C).
7220 if (I.getOpcode() == Instruction::Shl) {
7221 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt1));
7222 return BinaryOperator::CreateAnd(X, ConstantInt::get(Mask));
7224 // If we have ((X << C) >>u C), turn this into X & (-1 >>u C).
7225 if (I.getOpcode() == Instruction::LShr) {
7226 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt1));
7227 return BinaryOperator::CreateAnd(X, ConstantInt::get(Mask));
7229 // We can simplify ((X << C) >>s C) into a trunc + sext.
7230 // NOTE: we could do this for any C, but that would make 'unusual' integer
7231 // types. For now, just stick to ones well-supported by the code
7233 const Type *SExtType = 0;
7234 switch (Ty->getBitWidth() - ShiftAmt1) {
7241 SExtType = IntegerType::get(Ty->getBitWidth() - ShiftAmt1);
7246 Instruction *NewTrunc = new TruncInst(X, SExtType, "sext");
7247 InsertNewInstBefore(NewTrunc, I);
7248 return new SExtInst(NewTrunc, Ty);
7250 // Otherwise, we can't handle it yet.
7251 } else if (ShiftAmt1 < ShiftAmt2) {
7252 uint32_t ShiftDiff = ShiftAmt2-ShiftAmt1;
7254 // (X >>? C1) << C2 --> X << (C2-C1) & (-1 << C2)
7255 if (I.getOpcode() == Instruction::Shl) {
7256 assert(ShiftOp->getOpcode() == Instruction::LShr ||
7257 ShiftOp->getOpcode() == Instruction::AShr);
7258 Instruction *Shift =
7259 BinaryOperator::CreateShl(X, ConstantInt::get(Ty, ShiftDiff));
7260 InsertNewInstBefore(Shift, I);
7262 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
7263 return BinaryOperator::CreateAnd(Shift, ConstantInt::get(Mask));
7266 // (X << C1) >>u C2 --> X >>u (C2-C1) & (-1 >> C2)
7267 if (I.getOpcode() == Instruction::LShr) {
7268 assert(ShiftOp->getOpcode() == Instruction::Shl);
7269 Instruction *Shift =
7270 BinaryOperator::CreateLShr(X, ConstantInt::get(Ty, ShiftDiff));
7271 InsertNewInstBefore(Shift, I);
7273 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
7274 return BinaryOperator::CreateAnd(Shift, ConstantInt::get(Mask));
7277 // We can't handle (X << C1) >>s C2, it shifts arbitrary bits in.
7279 assert(ShiftAmt2 < ShiftAmt1);
7280 uint32_t ShiftDiff = ShiftAmt1-ShiftAmt2;
7282 // (X >>? C1) << C2 --> X >>? (C1-C2) & (-1 << C2)
7283 if (I.getOpcode() == Instruction::Shl) {
7284 assert(ShiftOp->getOpcode() == Instruction::LShr ||
7285 ShiftOp->getOpcode() == Instruction::AShr);
7286 Instruction *Shift =
7287 BinaryOperator::Create(ShiftOp->getOpcode(), X,
7288 ConstantInt::get(Ty, ShiftDiff));
7289 InsertNewInstBefore(Shift, I);
7291 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
7292 return BinaryOperator::CreateAnd(Shift, ConstantInt::get(Mask));
7295 // (X << C1) >>u C2 --> X << (C1-C2) & (-1 >> C2)
7296 if (I.getOpcode() == Instruction::LShr) {
7297 assert(ShiftOp->getOpcode() == Instruction::Shl);
7298 Instruction *Shift =
7299 BinaryOperator::CreateShl(X, ConstantInt::get(Ty, ShiftDiff));
7300 InsertNewInstBefore(Shift, I);
7302 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
7303 return BinaryOperator::CreateAnd(Shift, ConstantInt::get(Mask));
7306 // We can't handle (X << C1) >>a C2, it shifts arbitrary bits in.
7313 /// DecomposeSimpleLinearExpr - Analyze 'Val', seeing if it is a simple linear
7314 /// expression. If so, decompose it, returning some value X, such that Val is
7317 static Value *DecomposeSimpleLinearExpr(Value *Val, unsigned &Scale,
7319 assert(Val->getType() == Type::Int32Ty && "Unexpected allocation size type!");
7320 if (ConstantInt *CI = dyn_cast<ConstantInt>(Val)) {
7321 Offset = CI->getZExtValue();
7323 return ConstantInt::get(Type::Int32Ty, 0);
7324 } else if (BinaryOperator *I = dyn_cast<BinaryOperator>(Val)) {
7325 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
7326 if (I->getOpcode() == Instruction::Shl) {
7327 // This is a value scaled by '1 << the shift amt'.
7328 Scale = 1U << RHS->getZExtValue();
7330 return I->getOperand(0);
7331 } else if (I->getOpcode() == Instruction::Mul) {
7332 // This value is scaled by 'RHS'.
7333 Scale = RHS->getZExtValue();
7335 return I->getOperand(0);
7336 } else if (I->getOpcode() == Instruction::Add) {
7337 // We have X+C. Check to see if we really have (X*C2)+C1,
7338 // where C1 is divisible by C2.
7341 DecomposeSimpleLinearExpr(I->getOperand(0), SubScale, Offset);
7342 Offset += RHS->getZExtValue();
7349 // Otherwise, we can't look past this.
7356 /// PromoteCastOfAllocation - If we find a cast of an allocation instruction,
7357 /// try to eliminate the cast by moving the type information into the alloc.
7358 Instruction *InstCombiner::PromoteCastOfAllocation(BitCastInst &CI,
7359 AllocationInst &AI) {
7360 const PointerType *PTy = cast<PointerType>(CI.getType());
7362 // Remove any uses of AI that are dead.
7363 assert(!CI.use_empty() && "Dead instructions should be removed earlier!");
7365 for (Value::use_iterator UI = AI.use_begin(), E = AI.use_end(); UI != E; ) {
7366 Instruction *User = cast<Instruction>(*UI++);
7367 if (isInstructionTriviallyDead(User)) {
7368 while (UI != E && *UI == User)
7369 ++UI; // If this instruction uses AI more than once, don't break UI.
7372 DOUT << "IC: DCE: " << *User;
7373 EraseInstFromFunction(*User);
7377 // Get the type really allocated and the type casted to.
7378 const Type *AllocElTy = AI.getAllocatedType();
7379 const Type *CastElTy = PTy->getElementType();
7380 if (!AllocElTy->isSized() || !CastElTy->isSized()) return 0;
7382 unsigned AllocElTyAlign = TD->getABITypeAlignment(AllocElTy);
7383 unsigned CastElTyAlign = TD->getABITypeAlignment(CastElTy);
7384 if (CastElTyAlign < AllocElTyAlign) return 0;
7386 // If the allocation has multiple uses, only promote it if we are strictly
7387 // increasing the alignment of the resultant allocation. If we keep it the
7388 // same, we open the door to infinite loops of various kinds.
7389 if (!AI.hasOneUse() && CastElTyAlign == AllocElTyAlign) return 0;
7391 uint64_t AllocElTySize = TD->getABITypeSize(AllocElTy);
7392 uint64_t CastElTySize = TD->getABITypeSize(CastElTy);
7393 if (CastElTySize == 0 || AllocElTySize == 0) return 0;
7395 // See if we can satisfy the modulus by pulling a scale out of the array
7397 unsigned ArraySizeScale;
7399 Value *NumElements = // See if the array size is a decomposable linear expr.
7400 DecomposeSimpleLinearExpr(AI.getOperand(0), ArraySizeScale, ArrayOffset);
7402 // If we can now satisfy the modulus, by using a non-1 scale, we really can
7404 if ((AllocElTySize*ArraySizeScale) % CastElTySize != 0 ||
7405 (AllocElTySize*ArrayOffset ) % CastElTySize != 0) return 0;
7407 unsigned Scale = (AllocElTySize*ArraySizeScale)/CastElTySize;
7412 // If the allocation size is constant, form a constant mul expression
7413 Amt = ConstantInt::get(Type::Int32Ty, Scale);
7414 if (isa<ConstantInt>(NumElements))
7415 Amt = Multiply(cast<ConstantInt>(NumElements), cast<ConstantInt>(Amt));
7416 // otherwise multiply the amount and the number of elements
7417 else if (Scale != 1) {
7418 Instruction *Tmp = BinaryOperator::CreateMul(Amt, NumElements, "tmp");
7419 Amt = InsertNewInstBefore(Tmp, AI);
7423 if (int Offset = (AllocElTySize*ArrayOffset)/CastElTySize) {
7424 Value *Off = ConstantInt::get(Type::Int32Ty, Offset, true);
7425 Instruction *Tmp = BinaryOperator::CreateAdd(Amt, Off, "tmp");
7426 Amt = InsertNewInstBefore(Tmp, AI);
7429 AllocationInst *New;
7430 if (isa<MallocInst>(AI))
7431 New = new MallocInst(CastElTy, Amt, AI.getAlignment());
7433 New = new AllocaInst(CastElTy, Amt, AI.getAlignment());
7434 InsertNewInstBefore(New, AI);
7437 // If the allocation has multiple uses, insert a cast and change all things
7438 // that used it to use the new cast. This will also hack on CI, but it will
7440 if (!AI.hasOneUse()) {
7441 AddUsesToWorkList(AI);
7442 // New is the allocation instruction, pointer typed. AI is the original
7443 // allocation instruction, also pointer typed. Thus, cast to use is BitCast.
7444 CastInst *NewCast = new BitCastInst(New, AI.getType(), "tmpcast");
7445 InsertNewInstBefore(NewCast, AI);
7446 AI.replaceAllUsesWith(NewCast);
7448 return ReplaceInstUsesWith(CI, New);
7451 /// CanEvaluateInDifferentType - Return true if we can take the specified value
7452 /// and return it as type Ty without inserting any new casts and without
7453 /// changing the computed value. This is used by code that tries to decide
7454 /// whether promoting or shrinking integer operations to wider or smaller types
7455 /// will allow us to eliminate a truncate or extend.
7457 /// This is a truncation operation if Ty is smaller than V->getType(), or an
7458 /// extension operation if Ty is larger.
7460 /// If CastOpc is a truncation, then Ty will be a type smaller than V. We
7461 /// should return true if trunc(V) can be computed by computing V in the smaller
7462 /// type. If V is an instruction, then trunc(inst(x,y)) can be computed as
7463 /// inst(trunc(x),trunc(y)), which only makes sense if x and y can be
7464 /// efficiently truncated.
7466 /// If CastOpc is a sext or zext, we are asking if the low bits of the value can
7467 /// bit computed in a larger type, which is then and'd or sext_in_reg'd to get
7468 /// the final result.
7469 bool InstCombiner::CanEvaluateInDifferentType(Value *V, const IntegerType *Ty,
7471 int &NumCastsRemoved) {
7472 // We can always evaluate constants in another type.
7473 if (isa<ConstantInt>(V))
7476 Instruction *I = dyn_cast<Instruction>(V);
7477 if (!I) return false;
7479 const IntegerType *OrigTy = cast<IntegerType>(V->getType());
7481 // If this is an extension or truncate, we can often eliminate it.
7482 if (isa<TruncInst>(I) || isa<ZExtInst>(I) || isa<SExtInst>(I)) {
7483 // If this is a cast from the destination type, we can trivially eliminate
7484 // it, and this will remove a cast overall.
7485 if (I->getOperand(0)->getType() == Ty) {
7486 // If the first operand is itself a cast, and is eliminable, do not count
7487 // this as an eliminable cast. We would prefer to eliminate those two
7489 if (!isa<CastInst>(I->getOperand(0)) && I->hasOneUse())
7495 // We can't extend or shrink something that has multiple uses: doing so would
7496 // require duplicating the instruction in general, which isn't profitable.
7497 if (!I->hasOneUse()) return false;
7499 switch (I->getOpcode()) {
7500 case Instruction::Add:
7501 case Instruction::Sub:
7502 case Instruction::Mul:
7503 case Instruction::And:
7504 case Instruction::Or:
7505 case Instruction::Xor:
7506 // These operators can all arbitrarily be extended or truncated.
7507 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
7509 CanEvaluateInDifferentType(I->getOperand(1), Ty, CastOpc,
7512 case Instruction::Shl:
7513 // If we are truncating the result of this SHL, and if it's a shift of a
7514 // constant amount, we can always perform a SHL in a smaller type.
7515 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
7516 uint32_t BitWidth = Ty->getBitWidth();
7517 if (BitWidth < OrigTy->getBitWidth() &&
7518 CI->getLimitedValue(BitWidth) < BitWidth)
7519 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
7523 case Instruction::LShr:
7524 // If this is a truncate of a logical shr, we can truncate it to a smaller
7525 // lshr iff we know that the bits we would otherwise be shifting in are
7527 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
7528 uint32_t OrigBitWidth = OrigTy->getBitWidth();
7529 uint32_t BitWidth = Ty->getBitWidth();
7530 if (BitWidth < OrigBitWidth &&
7531 MaskedValueIsZero(I->getOperand(0),
7532 APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth)) &&
7533 CI->getLimitedValue(BitWidth) < BitWidth) {
7534 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
7539 case Instruction::ZExt:
7540 case Instruction::SExt:
7541 case Instruction::Trunc:
7542 // If this is the same kind of case as our original (e.g. zext+zext), we
7543 // can safely replace it. Note that replacing it does not reduce the number
7544 // of casts in the input.
7545 if (I->getOpcode() == CastOpc)
7548 case Instruction::Select: {
7549 SelectInst *SI = cast<SelectInst>(I);
7550 return CanEvaluateInDifferentType(SI->getTrueValue(), Ty, CastOpc,
7552 CanEvaluateInDifferentType(SI->getFalseValue(), Ty, CastOpc,
7555 case Instruction::PHI: {
7556 // We can change a phi if we can change all operands.
7557 PHINode *PN = cast<PHINode>(I);
7558 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
7559 if (!CanEvaluateInDifferentType(PN->getIncomingValue(i), Ty, CastOpc,
7565 // TODO: Can handle more cases here.
7572 /// EvaluateInDifferentType - Given an expression that
7573 /// CanEvaluateInDifferentType returns true for, actually insert the code to
7574 /// evaluate the expression.
7575 Value *InstCombiner::EvaluateInDifferentType(Value *V, const Type *Ty,
7577 if (Constant *C = dyn_cast<Constant>(V))
7578 return ConstantExpr::getIntegerCast(C, Ty, isSigned /*Sext or ZExt*/);
7580 // Otherwise, it must be an instruction.
7581 Instruction *I = cast<Instruction>(V);
7582 Instruction *Res = 0;
7583 switch (I->getOpcode()) {
7584 case Instruction::Add:
7585 case Instruction::Sub:
7586 case Instruction::Mul:
7587 case Instruction::And:
7588 case Instruction::Or:
7589 case Instruction::Xor:
7590 case Instruction::AShr:
7591 case Instruction::LShr:
7592 case Instruction::Shl: {
7593 Value *LHS = EvaluateInDifferentType(I->getOperand(0), Ty, isSigned);
7594 Value *RHS = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
7595 Res = BinaryOperator::Create((Instruction::BinaryOps)I->getOpcode(),
7599 case Instruction::Trunc:
7600 case Instruction::ZExt:
7601 case Instruction::SExt:
7602 // If the source type of the cast is the type we're trying for then we can
7603 // just return the source. There's no need to insert it because it is not
7605 if (I->getOperand(0)->getType() == Ty)
7606 return I->getOperand(0);
7608 // Otherwise, must be the same type of cast, so just reinsert a new one.
7609 Res = CastInst::Create(cast<CastInst>(I)->getOpcode(), I->getOperand(0),
7612 case Instruction::Select: {
7613 Value *True = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
7614 Value *False = EvaluateInDifferentType(I->getOperand(2), Ty, isSigned);
7615 Res = SelectInst::Create(I->getOperand(0), True, False);
7618 case Instruction::PHI: {
7619 PHINode *OPN = cast<PHINode>(I);
7620 PHINode *NPN = PHINode::Create(Ty);
7621 for (unsigned i = 0, e = OPN->getNumIncomingValues(); i != e; ++i) {
7622 Value *V =EvaluateInDifferentType(OPN->getIncomingValue(i), Ty, isSigned);
7623 NPN->addIncoming(V, OPN->getIncomingBlock(i));
7629 // TODO: Can handle more cases here.
7630 assert(0 && "Unreachable!");
7635 return InsertNewInstBefore(Res, *I);
7638 /// @brief Implement the transforms common to all CastInst visitors.
7639 Instruction *InstCombiner::commonCastTransforms(CastInst &CI) {
7640 Value *Src = CI.getOperand(0);
7642 // Many cases of "cast of a cast" are eliminable. If it's eliminable we just
7643 // eliminate it now.
7644 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
7645 if (Instruction::CastOps opc =
7646 isEliminableCastPair(CSrc, CI.getOpcode(), CI.getType(), TD)) {
7647 // The first cast (CSrc) is eliminable so we need to fix up or replace
7648 // the second cast (CI). CSrc will then have a good chance of being dead.
7649 return CastInst::Create(opc, CSrc->getOperand(0), CI.getType());
7653 // If we are casting a select then fold the cast into the select
7654 if (SelectInst *SI = dyn_cast<SelectInst>(Src))
7655 if (Instruction *NV = FoldOpIntoSelect(CI, SI, this))
7658 // If we are casting a PHI then fold the cast into the PHI
7659 if (isa<PHINode>(Src))
7660 if (Instruction *NV = FoldOpIntoPhi(CI))
7666 /// @brief Implement the transforms for cast of pointer (bitcast/ptrtoint)
7667 Instruction *InstCombiner::commonPointerCastTransforms(CastInst &CI) {
7668 Value *Src = CI.getOperand(0);
7670 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Src)) {
7671 // If casting the result of a getelementptr instruction with no offset, turn
7672 // this into a cast of the original pointer!
7673 if (GEP->hasAllZeroIndices()) {
7674 // Changing the cast operand is usually not a good idea but it is safe
7675 // here because the pointer operand is being replaced with another
7676 // pointer operand so the opcode doesn't need to change.
7678 CI.setOperand(0, GEP->getOperand(0));
7682 // If the GEP has a single use, and the base pointer is a bitcast, and the
7683 // GEP computes a constant offset, see if we can convert these three
7684 // instructions into fewer. This typically happens with unions and other
7685 // non-type-safe code.
7686 if (GEP->hasOneUse() && isa<BitCastInst>(GEP->getOperand(0))) {
7687 if (GEP->hasAllConstantIndices()) {
7688 // We are guaranteed to get a constant from EmitGEPOffset.
7689 ConstantInt *OffsetV = cast<ConstantInt>(EmitGEPOffset(GEP, CI, *this));
7690 int64_t Offset = OffsetV->getSExtValue();
7692 // Get the base pointer input of the bitcast, and the type it points to.
7693 Value *OrigBase = cast<BitCastInst>(GEP->getOperand(0))->getOperand(0);
7694 const Type *GEPIdxTy =
7695 cast<PointerType>(OrigBase->getType())->getElementType();
7696 if (GEPIdxTy->isSized()) {
7697 SmallVector<Value*, 8> NewIndices;
7699 // Start with the index over the outer type. Note that the type size
7700 // might be zero (even if the offset isn't zero) if the indexed type
7701 // is something like [0 x {int, int}]
7702 const Type *IntPtrTy = TD->getIntPtrType();
7703 int64_t FirstIdx = 0;
7704 if (int64_t TySize = TD->getABITypeSize(GEPIdxTy)) {
7705 FirstIdx = Offset/TySize;
7708 // Handle silly modulus not returning values values [0..TySize).
7712 assert(Offset >= 0);
7714 assert((uint64_t)Offset < (uint64_t)TySize &&"Out of range offset");
7717 NewIndices.push_back(ConstantInt::get(IntPtrTy, FirstIdx));
7719 // Index into the types. If we fail, set OrigBase to null.
7721 if (const StructType *STy = dyn_cast<StructType>(GEPIdxTy)) {
7722 const StructLayout *SL = TD->getStructLayout(STy);
7723 if (Offset < (int64_t)SL->getSizeInBytes()) {
7724 unsigned Elt = SL->getElementContainingOffset(Offset);
7725 NewIndices.push_back(ConstantInt::get(Type::Int32Ty, Elt));
7727 Offset -= SL->getElementOffset(Elt);
7728 GEPIdxTy = STy->getElementType(Elt);
7730 // Otherwise, we can't index into this, bail out.
7734 } else if (isa<ArrayType>(GEPIdxTy) || isa<VectorType>(GEPIdxTy)) {
7735 const SequentialType *STy = cast<SequentialType>(GEPIdxTy);
7736 if (uint64_t EltSize = TD->getABITypeSize(STy->getElementType())){
7737 NewIndices.push_back(ConstantInt::get(IntPtrTy,Offset/EltSize));
7740 NewIndices.push_back(ConstantInt::get(IntPtrTy, 0));
7742 GEPIdxTy = STy->getElementType();
7744 // Otherwise, we can't index into this, bail out.
7750 // If we were able to index down into an element, create the GEP
7751 // and bitcast the result. This eliminates one bitcast, potentially
7753 Instruction *NGEP = GetElementPtrInst::Create(OrigBase,
7755 NewIndices.end(), "");
7756 InsertNewInstBefore(NGEP, CI);
7757 NGEP->takeName(GEP);
7759 if (isa<BitCastInst>(CI))
7760 return new BitCastInst(NGEP, CI.getType());
7761 assert(isa<PtrToIntInst>(CI));
7762 return new PtrToIntInst(NGEP, CI.getType());
7769 return commonCastTransforms(CI);
7774 /// Only the TRUNC, ZEXT, SEXT, and BITCAST can both operand and result as
7775 /// integer types. This function implements the common transforms for all those
7777 /// @brief Implement the transforms common to CastInst with integer operands
7778 Instruction *InstCombiner::commonIntCastTransforms(CastInst &CI) {
7779 if (Instruction *Result = commonCastTransforms(CI))
7782 Value *Src = CI.getOperand(0);
7783 const Type *SrcTy = Src->getType();
7784 const Type *DestTy = CI.getType();
7785 uint32_t SrcBitSize = SrcTy->getPrimitiveSizeInBits();
7786 uint32_t DestBitSize = DestTy->getPrimitiveSizeInBits();
7788 // See if we can simplify any instructions used by the LHS whose sole
7789 // purpose is to compute bits we don't care about.
7790 APInt KnownZero(DestBitSize, 0), KnownOne(DestBitSize, 0);
7791 if (SimplifyDemandedBits(&CI, APInt::getAllOnesValue(DestBitSize),
7792 KnownZero, KnownOne))
7795 // If the source isn't an instruction or has more than one use then we
7796 // can't do anything more.
7797 Instruction *SrcI = dyn_cast<Instruction>(Src);
7798 if (!SrcI || !Src->hasOneUse())
7801 // Attempt to propagate the cast into the instruction for int->int casts.
7802 int NumCastsRemoved = 0;
7803 if (!isa<BitCastInst>(CI) &&
7804 CanEvaluateInDifferentType(SrcI, cast<IntegerType>(DestTy),
7805 CI.getOpcode(), NumCastsRemoved)) {
7806 // If this cast is a truncate, evaluting in a different type always
7807 // eliminates the cast, so it is always a win. If this is a zero-extension,
7808 // we need to do an AND to maintain the clear top-part of the computation,
7809 // so we require that the input have eliminated at least one cast. If this
7810 // is a sign extension, we insert two new casts (to do the extension) so we
7811 // require that two casts have been eliminated.
7813 switch (CI.getOpcode()) {
7815 // All the others use floating point so we shouldn't actually
7816 // get here because of the check above.
7817 assert(0 && "Unknown cast type");
7818 case Instruction::Trunc:
7821 case Instruction::ZExt:
7822 DoXForm = NumCastsRemoved >= 1;
7824 case Instruction::SExt:
7825 DoXForm = NumCastsRemoved >= 2;
7830 Value *Res = EvaluateInDifferentType(SrcI, DestTy,
7831 CI.getOpcode() == Instruction::SExt);
7832 assert(Res->getType() == DestTy);
7833 switch (CI.getOpcode()) {
7834 default: assert(0 && "Unknown cast type!");
7835 case Instruction::Trunc:
7836 case Instruction::BitCast:
7837 // Just replace this cast with the result.
7838 return ReplaceInstUsesWith(CI, Res);
7839 case Instruction::ZExt: {
7840 // We need to emit an AND to clear the high bits.
7841 assert(SrcBitSize < DestBitSize && "Not a zext?");
7842 Constant *C = ConstantInt::get(APInt::getLowBitsSet(DestBitSize,
7844 return BinaryOperator::CreateAnd(Res, C);
7846 case Instruction::SExt:
7847 // We need to emit a cast to truncate, then a cast to sext.
7848 return CastInst::Create(Instruction::SExt,
7849 InsertCastBefore(Instruction::Trunc, Res, Src->getType(),
7855 Value *Op0 = SrcI->getNumOperands() > 0 ? SrcI->getOperand(0) : 0;
7856 Value *Op1 = SrcI->getNumOperands() > 1 ? SrcI->getOperand(1) : 0;
7858 switch (SrcI->getOpcode()) {
7859 case Instruction::Add:
7860 case Instruction::Mul:
7861 case Instruction::And:
7862 case Instruction::Or:
7863 case Instruction::Xor:
7864 // If we are discarding information, rewrite.
7865 if (DestBitSize <= SrcBitSize && DestBitSize != 1) {
7866 // Don't insert two casts if they cannot be eliminated. We allow
7867 // two casts to be inserted if the sizes are the same. This could
7868 // only be converting signedness, which is a noop.
7869 if (DestBitSize == SrcBitSize ||
7870 !ValueRequiresCast(CI.getOpcode(), Op1, DestTy,TD) ||
7871 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
7872 Instruction::CastOps opcode = CI.getOpcode();
7873 Value *Op0c = InsertOperandCastBefore(opcode, Op0, DestTy, SrcI);
7874 Value *Op1c = InsertOperandCastBefore(opcode, Op1, DestTy, SrcI);
7875 return BinaryOperator::Create(
7876 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
7880 // cast (xor bool X, true) to int --> xor (cast bool X to int), 1
7881 if (isa<ZExtInst>(CI) && SrcBitSize == 1 &&
7882 SrcI->getOpcode() == Instruction::Xor &&
7883 Op1 == ConstantInt::getTrue() &&
7884 (!Op0->hasOneUse() || !isa<CmpInst>(Op0))) {
7885 Value *New = InsertOperandCastBefore(Instruction::ZExt, Op0, DestTy, &CI);
7886 return BinaryOperator::CreateXor(New, ConstantInt::get(CI.getType(), 1));
7889 case Instruction::SDiv:
7890 case Instruction::UDiv:
7891 case Instruction::SRem:
7892 case Instruction::URem:
7893 // If we are just changing the sign, rewrite.
7894 if (DestBitSize == SrcBitSize) {
7895 // Don't insert two casts if they cannot be eliminated. We allow
7896 // two casts to be inserted if the sizes are the same. This could
7897 // only be converting signedness, which is a noop.
7898 if (!ValueRequiresCast(CI.getOpcode(), Op1, DestTy, TD) ||
7899 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
7900 Value *Op0c = InsertOperandCastBefore(Instruction::BitCast,
7902 Value *Op1c = InsertOperandCastBefore(Instruction::BitCast,
7904 return BinaryOperator::Create(
7905 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
7910 case Instruction::Shl:
7911 // Allow changing the sign of the source operand. Do not allow
7912 // changing the size of the shift, UNLESS the shift amount is a
7913 // constant. We must not change variable sized shifts to a smaller
7914 // size, because it is undefined to shift more bits out than exist
7916 if (DestBitSize == SrcBitSize ||
7917 (DestBitSize < SrcBitSize && isa<Constant>(Op1))) {
7918 Instruction::CastOps opcode = (DestBitSize == SrcBitSize ?
7919 Instruction::BitCast : Instruction::Trunc);
7920 Value *Op0c = InsertOperandCastBefore(opcode, Op0, DestTy, SrcI);
7921 Value *Op1c = InsertOperandCastBefore(opcode, Op1, DestTy, SrcI);
7922 return BinaryOperator::CreateShl(Op0c, Op1c);
7925 case Instruction::AShr:
7926 // If this is a signed shr, and if all bits shifted in are about to be
7927 // truncated off, turn it into an unsigned shr to allow greater
7929 if (DestBitSize < SrcBitSize &&
7930 isa<ConstantInt>(Op1)) {
7931 uint32_t ShiftAmt = cast<ConstantInt>(Op1)->getLimitedValue(SrcBitSize);
7932 if (SrcBitSize > ShiftAmt && SrcBitSize-ShiftAmt >= DestBitSize) {
7933 // Insert the new logical shift right.
7934 return BinaryOperator::CreateLShr(Op0, Op1);
7942 Instruction *InstCombiner::visitTrunc(TruncInst &CI) {
7943 if (Instruction *Result = commonIntCastTransforms(CI))
7946 Value *Src = CI.getOperand(0);
7947 const Type *Ty = CI.getType();
7948 uint32_t DestBitWidth = Ty->getPrimitiveSizeInBits();
7949 uint32_t SrcBitWidth = cast<IntegerType>(Src->getType())->getBitWidth();
7951 if (Instruction *SrcI = dyn_cast<Instruction>(Src)) {
7952 switch (SrcI->getOpcode()) {
7954 case Instruction::LShr:
7955 // We can shrink lshr to something smaller if we know the bits shifted in
7956 // are already zeros.
7957 if (ConstantInt *ShAmtV = dyn_cast<ConstantInt>(SrcI->getOperand(1))) {
7958 uint32_t ShAmt = ShAmtV->getLimitedValue(SrcBitWidth);
7960 // Get a mask for the bits shifting in.
7961 APInt Mask(APInt::getLowBitsSet(SrcBitWidth, ShAmt).shl(DestBitWidth));
7962 Value* SrcIOp0 = SrcI->getOperand(0);
7963 if (SrcI->hasOneUse() && MaskedValueIsZero(SrcIOp0, Mask)) {
7964 if (ShAmt >= DestBitWidth) // All zeros.
7965 return ReplaceInstUsesWith(CI, Constant::getNullValue(Ty));
7967 // Okay, we can shrink this. Truncate the input, then return a new
7969 Value *V1 = InsertCastBefore(Instruction::Trunc, SrcIOp0, Ty, CI);
7970 Value *V2 = InsertCastBefore(Instruction::Trunc, SrcI->getOperand(1),
7972 return BinaryOperator::CreateLShr(V1, V2);
7974 } else { // This is a variable shr.
7976 // Turn 'trunc (lshr X, Y) to bool' into '(X & (1 << Y)) != 0'. This is
7977 // more LLVM instructions, but allows '1 << Y' to be hoisted if
7978 // loop-invariant and CSE'd.
7979 if (CI.getType() == Type::Int1Ty && SrcI->hasOneUse()) {
7980 Value *One = ConstantInt::get(SrcI->getType(), 1);
7982 Value *V = InsertNewInstBefore(
7983 BinaryOperator::CreateShl(One, SrcI->getOperand(1),
7985 V = InsertNewInstBefore(BinaryOperator::CreateAnd(V,
7986 SrcI->getOperand(0),
7988 Value *Zero = Constant::getNullValue(V->getType());
7989 return new ICmpInst(ICmpInst::ICMP_NE, V, Zero);
7999 /// transformZExtICmp - Transform (zext icmp) to bitwise / integer operations
8000 /// in order to eliminate the icmp.
8001 Instruction *InstCombiner::transformZExtICmp(ICmpInst *ICI, Instruction &CI,
8003 // If we are just checking for a icmp eq of a single bit and zext'ing it
8004 // to an integer, then shift the bit to the appropriate place and then
8005 // cast to integer to avoid the comparison.
8006 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(ICI->getOperand(1))) {
8007 const APInt &Op1CV = Op1C->getValue();
8009 // zext (x <s 0) to i32 --> x>>u31 true if signbit set.
8010 // zext (x >s -1) to i32 --> (x>>u31)^1 true if signbit clear.
8011 if ((ICI->getPredicate() == ICmpInst::ICMP_SLT && Op1CV == 0) ||
8012 (ICI->getPredicate() == ICmpInst::ICMP_SGT &&Op1CV.isAllOnesValue())) {
8013 if (!DoXform) return ICI;
8015 Value *In = ICI->getOperand(0);
8016 Value *Sh = ConstantInt::get(In->getType(),
8017 In->getType()->getPrimitiveSizeInBits()-1);
8018 In = InsertNewInstBefore(BinaryOperator::CreateLShr(In, Sh,
8019 In->getName()+".lobit"),
8021 if (In->getType() != CI.getType())
8022 In = CastInst::CreateIntegerCast(In, CI.getType(),
8023 false/*ZExt*/, "tmp", &CI);
8025 if (ICI->getPredicate() == ICmpInst::ICMP_SGT) {
8026 Constant *One = ConstantInt::get(In->getType(), 1);
8027 In = InsertNewInstBefore(BinaryOperator::CreateXor(In, One,
8028 In->getName()+".not"),
8032 return ReplaceInstUsesWith(CI, In);
8037 // zext (X == 0) to i32 --> X^1 iff X has only the low bit set.
8038 // zext (X == 0) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
8039 // zext (X == 1) to i32 --> X iff X has only the low bit set.
8040 // zext (X == 2) to i32 --> X>>1 iff X has only the 2nd bit set.
8041 // zext (X != 0) to i32 --> X iff X has only the low bit set.
8042 // zext (X != 0) to i32 --> X>>1 iff X has only the 2nd bit set.
8043 // zext (X != 1) to i32 --> X^1 iff X has only the low bit set.
8044 // zext (X != 2) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
8045 if ((Op1CV == 0 || Op1CV.isPowerOf2()) &&
8046 // This only works for EQ and NE
8047 ICI->isEquality()) {
8048 // If Op1C some other power of two, convert:
8049 uint32_t BitWidth = Op1C->getType()->getBitWidth();
8050 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
8051 APInt TypeMask(APInt::getAllOnesValue(BitWidth));
8052 ComputeMaskedBits(ICI->getOperand(0), TypeMask, KnownZero, KnownOne);
8054 APInt KnownZeroMask(~KnownZero);
8055 if (KnownZeroMask.isPowerOf2()) { // Exactly 1 possible 1?
8056 if (!DoXform) return ICI;
8058 bool isNE = ICI->getPredicate() == ICmpInst::ICMP_NE;
8059 if (Op1CV != 0 && (Op1CV != KnownZeroMask)) {
8060 // (X&4) == 2 --> false
8061 // (X&4) != 2 --> true
8062 Constant *Res = ConstantInt::get(Type::Int1Ty, isNE);
8063 Res = ConstantExpr::getZExt(Res, CI.getType());
8064 return ReplaceInstUsesWith(CI, Res);
8067 uint32_t ShiftAmt = KnownZeroMask.logBase2();
8068 Value *In = ICI->getOperand(0);
8070 // Perform a logical shr by shiftamt.
8071 // Insert the shift to put the result in the low bit.
8072 In = InsertNewInstBefore(BinaryOperator::CreateLShr(In,
8073 ConstantInt::get(In->getType(), ShiftAmt),
8074 In->getName()+".lobit"), CI);
8077 if ((Op1CV != 0) == isNE) { // Toggle the low bit.
8078 Constant *One = ConstantInt::get(In->getType(), 1);
8079 In = BinaryOperator::CreateXor(In, One, "tmp");
8080 InsertNewInstBefore(cast<Instruction>(In), CI);
8083 if (CI.getType() == In->getType())
8084 return ReplaceInstUsesWith(CI, In);
8086 return CastInst::CreateIntegerCast(In, CI.getType(), false/*ZExt*/);
8094 Instruction *InstCombiner::visitZExt(ZExtInst &CI) {
8095 // If one of the common conversion will work ..
8096 if (Instruction *Result = commonIntCastTransforms(CI))
8099 Value *Src = CI.getOperand(0);
8101 // If this is a cast of a cast
8102 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
8103 // If this is a TRUNC followed by a ZEXT then we are dealing with integral
8104 // types and if the sizes are just right we can convert this into a logical
8105 // 'and' which will be much cheaper than the pair of casts.
8106 if (isa<TruncInst>(CSrc)) {
8107 // Get the sizes of the types involved
8108 Value *A = CSrc->getOperand(0);
8109 uint32_t SrcSize = A->getType()->getPrimitiveSizeInBits();
8110 uint32_t MidSize = CSrc->getType()->getPrimitiveSizeInBits();
8111 uint32_t DstSize = CI.getType()->getPrimitiveSizeInBits();
8112 // If we're actually extending zero bits and the trunc is a no-op
8113 if (MidSize < DstSize && SrcSize == DstSize) {
8114 // Replace both of the casts with an And of the type mask.
8115 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
8116 Constant *AndConst = ConstantInt::get(AndValue);
8118 BinaryOperator::CreateAnd(CSrc->getOperand(0), AndConst);
8119 // Unfortunately, if the type changed, we need to cast it back.
8120 if (And->getType() != CI.getType()) {
8121 And->setName(CSrc->getName()+".mask");
8122 InsertNewInstBefore(And, CI);
8123 And = CastInst::CreateIntegerCast(And, CI.getType(), false/*ZExt*/);
8130 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src))
8131 return transformZExtICmp(ICI, CI);
8133 BinaryOperator *SrcI = dyn_cast<BinaryOperator>(Src);
8134 if (SrcI && SrcI->getOpcode() == Instruction::Or) {
8135 // zext (or icmp, icmp) --> or (zext icmp), (zext icmp) if at least one
8136 // of the (zext icmp) will be transformed.
8137 ICmpInst *LHS = dyn_cast<ICmpInst>(SrcI->getOperand(0));
8138 ICmpInst *RHS = dyn_cast<ICmpInst>(SrcI->getOperand(1));
8139 if (LHS && RHS && LHS->hasOneUse() && RHS->hasOneUse() &&
8140 (transformZExtICmp(LHS, CI, false) ||
8141 transformZExtICmp(RHS, CI, false))) {
8142 Value *LCast = InsertCastBefore(Instruction::ZExt, LHS, CI.getType(), CI);
8143 Value *RCast = InsertCastBefore(Instruction::ZExt, RHS, CI.getType(), CI);
8144 return BinaryOperator::Create(Instruction::Or, LCast, RCast);
8151 Instruction *InstCombiner::visitSExt(SExtInst &CI) {
8152 if (Instruction *I = commonIntCastTransforms(CI))
8155 Value *Src = CI.getOperand(0);
8157 // Canonicalize sign-extend from i1 to a select.
8158 if (Src->getType() == Type::Int1Ty)
8159 return SelectInst::Create(Src,
8160 ConstantInt::getAllOnesValue(CI.getType()),
8161 Constant::getNullValue(CI.getType()));
8163 // See if the value being truncated is already sign extended. If so, just
8164 // eliminate the trunc/sext pair.
8165 if (getOpcode(Src) == Instruction::Trunc) {
8166 Value *Op = cast<User>(Src)->getOperand(0);
8167 unsigned OpBits = cast<IntegerType>(Op->getType())->getBitWidth();
8168 unsigned MidBits = cast<IntegerType>(Src->getType())->getBitWidth();
8169 unsigned DestBits = cast<IntegerType>(CI.getType())->getBitWidth();
8170 unsigned NumSignBits = ComputeNumSignBits(Op);
8172 if (OpBits == DestBits) {
8173 // Op is i32, Mid is i8, and Dest is i32. If Op has more than 24 sign
8174 // bits, it is already ready.
8175 if (NumSignBits > DestBits-MidBits)
8176 return ReplaceInstUsesWith(CI, Op);
8177 } else if (OpBits < DestBits) {
8178 // Op is i32, Mid is i8, and Dest is i64. If Op has more than 24 sign
8179 // bits, just sext from i32.
8180 if (NumSignBits > OpBits-MidBits)
8181 return new SExtInst(Op, CI.getType(), "tmp");
8183 // Op is i64, Mid is i8, and Dest is i32. If Op has more than 56 sign
8184 // bits, just truncate to i32.
8185 if (NumSignBits > OpBits-MidBits)
8186 return new TruncInst(Op, CI.getType(), "tmp");
8190 // If the input is a shl/ashr pair of a same constant, then this is a sign
8191 // extension from a smaller value. If we could trust arbitrary bitwidth
8192 // integers, we could turn this into a truncate to the smaller bit and then
8193 // use a sext for the whole extension. Since we don't, look deeper and check
8194 // for a truncate. If the source and dest are the same type, eliminate the
8195 // trunc and extend and just do shifts. For example, turn:
8196 // %a = trunc i32 %i to i8
8197 // %b = shl i8 %a, 6
8198 // %c = ashr i8 %b, 6
8199 // %d = sext i8 %c to i32
8201 // %a = shl i32 %i, 30
8202 // %d = ashr i32 %a, 30
8204 ConstantInt *BA = 0, *CA = 0;
8205 if (match(Src, m_AShr(m_Shl(m_Value(A), m_ConstantInt(BA)),
8206 m_ConstantInt(CA))) &&
8207 BA == CA && isa<TruncInst>(A)) {
8208 Value *I = cast<TruncInst>(A)->getOperand(0);
8209 if (I->getType() == CI.getType()) {
8210 unsigned MidSize = Src->getType()->getPrimitiveSizeInBits();
8211 unsigned SrcDstSize = CI.getType()->getPrimitiveSizeInBits();
8212 unsigned ShAmt = CA->getZExtValue()+SrcDstSize-MidSize;
8213 Constant *ShAmtV = ConstantInt::get(CI.getType(), ShAmt);
8214 I = InsertNewInstBefore(BinaryOperator::CreateShl(I, ShAmtV,
8216 return BinaryOperator::CreateAShr(I, ShAmtV);
8223 /// FitsInFPType - Return a Constant* for the specified FP constant if it fits
8224 /// in the specified FP type without changing its value.
8225 static Constant *FitsInFPType(ConstantFP *CFP, const fltSemantics &Sem) {
8227 APFloat F = CFP->getValueAPF();
8228 (void)F.convert(Sem, APFloat::rmNearestTiesToEven, &losesInfo);
8230 return ConstantFP::get(F);
8234 /// LookThroughFPExtensions - If this is an fp extension instruction, look
8235 /// through it until we get the source value.
8236 static Value *LookThroughFPExtensions(Value *V) {
8237 if (Instruction *I = dyn_cast<Instruction>(V))
8238 if (I->getOpcode() == Instruction::FPExt)
8239 return LookThroughFPExtensions(I->getOperand(0));
8241 // If this value is a constant, return the constant in the smallest FP type
8242 // that can accurately represent it. This allows us to turn
8243 // (float)((double)X+2.0) into x+2.0f.
8244 if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
8245 if (CFP->getType() == Type::PPC_FP128Ty)
8246 return V; // No constant folding of this.
8247 // See if the value can be truncated to float and then reextended.
8248 if (Value *V = FitsInFPType(CFP, APFloat::IEEEsingle))
8250 if (CFP->getType() == Type::DoubleTy)
8251 return V; // Won't shrink.
8252 if (Value *V = FitsInFPType(CFP, APFloat::IEEEdouble))
8254 // Don't try to shrink to various long double types.
8260 Instruction *InstCombiner::visitFPTrunc(FPTruncInst &CI) {
8261 if (Instruction *I = commonCastTransforms(CI))
8264 // If we have fptrunc(add (fpextend x), (fpextend y)), where x and y are
8265 // smaller than the destination type, we can eliminate the truncate by doing
8266 // the add as the smaller type. This applies to add/sub/mul/div as well as
8267 // many builtins (sqrt, etc).
8268 BinaryOperator *OpI = dyn_cast<BinaryOperator>(CI.getOperand(0));
8269 if (OpI && OpI->hasOneUse()) {
8270 switch (OpI->getOpcode()) {
8272 case Instruction::Add:
8273 case Instruction::Sub:
8274 case Instruction::Mul:
8275 case Instruction::FDiv:
8276 case Instruction::FRem:
8277 const Type *SrcTy = OpI->getType();
8278 Value *LHSTrunc = LookThroughFPExtensions(OpI->getOperand(0));
8279 Value *RHSTrunc = LookThroughFPExtensions(OpI->getOperand(1));
8280 if (LHSTrunc->getType() != SrcTy &&
8281 RHSTrunc->getType() != SrcTy) {
8282 unsigned DstSize = CI.getType()->getPrimitiveSizeInBits();
8283 // If the source types were both smaller than the destination type of
8284 // the cast, do this xform.
8285 if (LHSTrunc->getType()->getPrimitiveSizeInBits() <= DstSize &&
8286 RHSTrunc->getType()->getPrimitiveSizeInBits() <= DstSize) {
8287 LHSTrunc = InsertCastBefore(Instruction::FPExt, LHSTrunc,
8289 RHSTrunc = InsertCastBefore(Instruction::FPExt, RHSTrunc,
8291 return BinaryOperator::Create(OpI->getOpcode(), LHSTrunc, RHSTrunc);
8300 Instruction *InstCombiner::visitFPExt(CastInst &CI) {
8301 return commonCastTransforms(CI);
8304 Instruction *InstCombiner::visitFPToUI(FPToUIInst &FI) {
8305 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0));
8307 return commonCastTransforms(FI);
8309 // fptoui(uitofp(X)) --> X
8310 // fptoui(sitofp(X)) --> X
8311 // This is safe if the intermediate type has enough bits in its mantissa to
8312 // accurately represent all values of X. For example, do not do this with
8313 // i64->float->i64. This is also safe for sitofp case, because any negative
8314 // 'X' value would cause an undefined result for the fptoui.
8315 if ((isa<UIToFPInst>(OpI) || isa<SIToFPInst>(OpI)) &&
8316 OpI->getOperand(0)->getType() == FI.getType() &&
8317 (int)FI.getType()->getPrimitiveSizeInBits() < /*extra bit for sign */
8318 OpI->getType()->getFPMantissaWidth())
8319 return ReplaceInstUsesWith(FI, OpI->getOperand(0));
8321 return commonCastTransforms(FI);
8324 Instruction *InstCombiner::visitFPToSI(FPToSIInst &FI) {
8325 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0));
8327 return commonCastTransforms(FI);
8329 // fptosi(sitofp(X)) --> X
8330 // fptosi(uitofp(X)) --> X
8331 // This is safe if the intermediate type has enough bits in its mantissa to
8332 // accurately represent all values of X. For example, do not do this with
8333 // i64->float->i64. This is also safe for sitofp case, because any negative
8334 // 'X' value would cause an undefined result for the fptoui.
8335 if ((isa<UIToFPInst>(OpI) || isa<SIToFPInst>(OpI)) &&
8336 OpI->getOperand(0)->getType() == FI.getType() &&
8337 (int)FI.getType()->getPrimitiveSizeInBits() <=
8338 OpI->getType()->getFPMantissaWidth())
8339 return ReplaceInstUsesWith(FI, OpI->getOperand(0));
8341 return commonCastTransforms(FI);
8344 Instruction *InstCombiner::visitUIToFP(CastInst &CI) {
8345 return commonCastTransforms(CI);
8348 Instruction *InstCombiner::visitSIToFP(CastInst &CI) {
8349 return commonCastTransforms(CI);
8352 Instruction *InstCombiner::visitPtrToInt(CastInst &CI) {
8353 return commonPointerCastTransforms(CI);
8356 Instruction *InstCombiner::visitIntToPtr(IntToPtrInst &CI) {
8357 if (Instruction *I = commonCastTransforms(CI))
8360 const Type *DestPointee = cast<PointerType>(CI.getType())->getElementType();
8361 if (!DestPointee->isSized()) return 0;
8363 // If this is inttoptr(add (ptrtoint x), cst), try to turn this into a GEP.
8366 if (match(CI.getOperand(0), m_Add(m_Cast<PtrToIntInst>(m_Value(X)),
8367 m_ConstantInt(Cst)))) {
8368 // If the source and destination operands have the same type, see if this
8369 // is a single-index GEP.
8370 if (X->getType() == CI.getType()) {
8371 // Get the size of the pointee type.
8372 uint64_t Size = TD->getABITypeSize(DestPointee);
8374 // Convert the constant to intptr type.
8375 APInt Offset = Cst->getValue();
8376 Offset.sextOrTrunc(TD->getPointerSizeInBits());
8378 // If Offset is evenly divisible by Size, we can do this xform.
8379 if (Size && !APIntOps::srem(Offset, APInt(Offset.getBitWidth(), Size))){
8380 Offset = APIntOps::sdiv(Offset, APInt(Offset.getBitWidth(), Size));
8381 return GetElementPtrInst::Create(X, ConstantInt::get(Offset));
8384 // TODO: Could handle other cases, e.g. where add is indexing into field of
8386 } else if (CI.getOperand(0)->hasOneUse() &&
8387 match(CI.getOperand(0), m_Add(m_Value(X), m_ConstantInt(Cst)))) {
8388 // Otherwise, if this is inttoptr(add x, cst), try to turn this into an
8389 // "inttoptr+GEP" instead of "add+intptr".
8391 // Get the size of the pointee type.
8392 uint64_t Size = TD->getABITypeSize(DestPointee);
8394 // Convert the constant to intptr type.
8395 APInt Offset = Cst->getValue();
8396 Offset.sextOrTrunc(TD->getPointerSizeInBits());
8398 // If Offset is evenly divisible by Size, we can do this xform.
8399 if (Size && !APIntOps::srem(Offset, APInt(Offset.getBitWidth(), Size))){
8400 Offset = APIntOps::sdiv(Offset, APInt(Offset.getBitWidth(), Size));
8402 Instruction *P = InsertNewInstBefore(new IntToPtrInst(X, CI.getType(),
8404 return GetElementPtrInst::Create(P, ConstantInt::get(Offset), "tmp");
8410 Instruction *InstCombiner::visitBitCast(BitCastInst &CI) {
8411 // If the operands are integer typed then apply the integer transforms,
8412 // otherwise just apply the common ones.
8413 Value *Src = CI.getOperand(0);
8414 const Type *SrcTy = Src->getType();
8415 const Type *DestTy = CI.getType();
8417 if (SrcTy->isInteger() && DestTy->isInteger()) {
8418 if (Instruction *Result = commonIntCastTransforms(CI))
8420 } else if (isa<PointerType>(SrcTy)) {
8421 if (Instruction *I = commonPointerCastTransforms(CI))
8424 if (Instruction *Result = commonCastTransforms(CI))
8429 // Get rid of casts from one type to the same type. These are useless and can
8430 // be replaced by the operand.
8431 if (DestTy == Src->getType())
8432 return ReplaceInstUsesWith(CI, Src);
8434 if (const PointerType *DstPTy = dyn_cast<PointerType>(DestTy)) {
8435 const PointerType *SrcPTy = cast<PointerType>(SrcTy);
8436 const Type *DstElTy = DstPTy->getElementType();
8437 const Type *SrcElTy = SrcPTy->getElementType();
8439 // If the address spaces don't match, don't eliminate the bitcast, which is
8440 // required for changing types.
8441 if (SrcPTy->getAddressSpace() != DstPTy->getAddressSpace())
8444 // If we are casting a malloc or alloca to a pointer to a type of the same
8445 // size, rewrite the allocation instruction to allocate the "right" type.
8446 if (AllocationInst *AI = dyn_cast<AllocationInst>(Src))
8447 if (Instruction *V = PromoteCastOfAllocation(CI, *AI))
8450 // If the source and destination are pointers, and this cast is equivalent
8451 // to a getelementptr X, 0, 0, 0... turn it into the appropriate gep.
8452 // This can enhance SROA and other transforms that want type-safe pointers.
8453 Constant *ZeroUInt = Constant::getNullValue(Type::Int32Ty);
8454 unsigned NumZeros = 0;
8455 while (SrcElTy != DstElTy &&
8456 isa<CompositeType>(SrcElTy) && !isa<PointerType>(SrcElTy) &&
8457 SrcElTy->getNumContainedTypes() /* not "{}" */) {
8458 SrcElTy = cast<CompositeType>(SrcElTy)->getTypeAtIndex(ZeroUInt);
8462 // If we found a path from the src to dest, create the getelementptr now.
8463 if (SrcElTy == DstElTy) {
8464 SmallVector<Value*, 8> Idxs(NumZeros+1, ZeroUInt);
8465 return GetElementPtrInst::Create(Src, Idxs.begin(), Idxs.end(), "",
8466 ((Instruction*) NULL));
8470 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(Src)) {
8471 if (SVI->hasOneUse()) {
8472 // Okay, we have (bitconvert (shuffle ..)). Check to see if this is
8473 // a bitconvert to a vector with the same # elts.
8474 if (isa<VectorType>(DestTy) &&
8475 cast<VectorType>(DestTy)->getNumElements() ==
8476 SVI->getType()->getNumElements() &&
8477 SVI->getType()->getNumElements() ==
8478 cast<VectorType>(SVI->getOperand(0)->getType())->getNumElements()) {
8480 // If either of the operands is a cast from CI.getType(), then
8481 // evaluating the shuffle in the casted destination's type will allow
8482 // us to eliminate at least one cast.
8483 if (((Tmp = dyn_cast<CastInst>(SVI->getOperand(0))) &&
8484 Tmp->getOperand(0)->getType() == DestTy) ||
8485 ((Tmp = dyn_cast<CastInst>(SVI->getOperand(1))) &&
8486 Tmp->getOperand(0)->getType() == DestTy)) {
8487 Value *LHS = InsertOperandCastBefore(Instruction::BitCast,
8488 SVI->getOperand(0), DestTy, &CI);
8489 Value *RHS = InsertOperandCastBefore(Instruction::BitCast,
8490 SVI->getOperand(1), DestTy, &CI);
8491 // Return a new shuffle vector. Use the same element ID's, as we
8492 // know the vector types match #elts.
8493 return new ShuffleVectorInst(LHS, RHS, SVI->getOperand(2));
8501 /// GetSelectFoldableOperands - We want to turn code that looks like this:
8503 /// %D = select %cond, %C, %A
8505 /// %C = select %cond, %B, 0
8508 /// Assuming that the specified instruction is an operand to the select, return
8509 /// a bitmask indicating which operands of this instruction are foldable if they
8510 /// equal the other incoming value of the select.
8512 static unsigned GetSelectFoldableOperands(Instruction *I) {
8513 switch (I->getOpcode()) {
8514 case Instruction::Add:
8515 case Instruction::Mul:
8516 case Instruction::And:
8517 case Instruction::Or:
8518 case Instruction::Xor:
8519 return 3; // Can fold through either operand.
8520 case Instruction::Sub: // Can only fold on the amount subtracted.
8521 case Instruction::Shl: // Can only fold on the shift amount.
8522 case Instruction::LShr:
8523 case Instruction::AShr:
8526 return 0; // Cannot fold
8530 /// GetSelectFoldableConstant - For the same transformation as the previous
8531 /// function, return the identity constant that goes into the select.
8532 static Constant *GetSelectFoldableConstant(Instruction *I) {
8533 switch (I->getOpcode()) {
8534 default: assert(0 && "This cannot happen!"); abort();
8535 case Instruction::Add:
8536 case Instruction::Sub:
8537 case Instruction::Or:
8538 case Instruction::Xor:
8539 case Instruction::Shl:
8540 case Instruction::LShr:
8541 case Instruction::AShr:
8542 return Constant::getNullValue(I->getType());
8543 case Instruction::And:
8544 return Constant::getAllOnesValue(I->getType());
8545 case Instruction::Mul:
8546 return ConstantInt::get(I->getType(), 1);
8550 /// FoldSelectOpOp - Here we have (select c, TI, FI), and we know that TI and FI
8551 /// have the same opcode and only one use each. Try to simplify this.
8552 Instruction *InstCombiner::FoldSelectOpOp(SelectInst &SI, Instruction *TI,
8554 if (TI->getNumOperands() == 1) {
8555 // If this is a non-volatile load or a cast from the same type,
8558 if (TI->getOperand(0)->getType() != FI->getOperand(0)->getType())
8561 return 0; // unknown unary op.
8564 // Fold this by inserting a select from the input values.
8565 SelectInst *NewSI = SelectInst::Create(SI.getCondition(), TI->getOperand(0),
8566 FI->getOperand(0), SI.getName()+".v");
8567 InsertNewInstBefore(NewSI, SI);
8568 return CastInst::Create(Instruction::CastOps(TI->getOpcode()), NewSI,
8572 // Only handle binary operators here.
8573 if (!isa<BinaryOperator>(TI))
8576 // Figure out if the operations have any operands in common.
8577 Value *MatchOp, *OtherOpT, *OtherOpF;
8579 if (TI->getOperand(0) == FI->getOperand(0)) {
8580 MatchOp = TI->getOperand(0);
8581 OtherOpT = TI->getOperand(1);
8582 OtherOpF = FI->getOperand(1);
8583 MatchIsOpZero = true;
8584 } else if (TI->getOperand(1) == FI->getOperand(1)) {
8585 MatchOp = TI->getOperand(1);
8586 OtherOpT = TI->getOperand(0);
8587 OtherOpF = FI->getOperand(0);
8588 MatchIsOpZero = false;
8589 } else if (!TI->isCommutative()) {
8591 } else if (TI->getOperand(0) == FI->getOperand(1)) {
8592 MatchOp = TI->getOperand(0);
8593 OtherOpT = TI->getOperand(1);
8594 OtherOpF = FI->getOperand(0);
8595 MatchIsOpZero = true;
8596 } else if (TI->getOperand(1) == FI->getOperand(0)) {
8597 MatchOp = TI->getOperand(1);
8598 OtherOpT = TI->getOperand(0);
8599 OtherOpF = FI->getOperand(1);
8600 MatchIsOpZero = true;
8605 // If we reach here, they do have operations in common.
8606 SelectInst *NewSI = SelectInst::Create(SI.getCondition(), OtherOpT,
8607 OtherOpF, SI.getName()+".v");
8608 InsertNewInstBefore(NewSI, SI);
8610 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TI)) {
8612 return BinaryOperator::Create(BO->getOpcode(), MatchOp, NewSI);
8614 return BinaryOperator::Create(BO->getOpcode(), NewSI, MatchOp);
8616 assert(0 && "Shouldn't get here");
8620 /// visitSelectInstWithICmp - Visit a SelectInst that has an
8621 /// ICmpInst as its first operand.
8623 Instruction *InstCombiner::visitSelectInstWithICmp(SelectInst &SI,
8625 bool Changed = false;
8626 ICmpInst::Predicate Pred = ICI->getPredicate();
8627 Value *CmpLHS = ICI->getOperand(0);
8628 Value *CmpRHS = ICI->getOperand(1);
8629 Value *TrueVal = SI.getTrueValue();
8630 Value *FalseVal = SI.getFalseValue();
8632 // Check cases where the comparison is with a constant that
8633 // can be adjusted to fit the min/max idiom. We may edit ICI in
8634 // place here, so make sure the select is the only user.
8635 if (ICI->hasOneUse())
8636 if (ConstantInt *CI = dyn_cast<ConstantInt>(CmpRHS)) {
8639 case ICmpInst::ICMP_ULT:
8640 case ICmpInst::ICMP_SLT: {
8641 // X < MIN ? T : F --> F
8642 if (CI->isMinValue(Pred == ICmpInst::ICMP_SLT))
8643 return ReplaceInstUsesWith(SI, FalseVal);
8644 // X < C ? X : C-1 --> X > C-1 ? C-1 : X
8645 Constant *AdjustedRHS = SubOne(CI);
8646 if ((CmpLHS == TrueVal && AdjustedRHS == FalseVal) ||
8647 (CmpLHS == FalseVal && AdjustedRHS == TrueVal)) {
8648 Pred = ICmpInst::getSwappedPredicate(Pred);
8649 CmpRHS = AdjustedRHS;
8650 std::swap(FalseVal, TrueVal);
8651 ICI->setPredicate(Pred);
8652 ICI->setOperand(1, CmpRHS);
8653 SI.setOperand(1, TrueVal);
8654 SI.setOperand(2, FalseVal);
8659 case ICmpInst::ICMP_UGT:
8660 case ICmpInst::ICMP_SGT: {
8661 // X > MAX ? T : F --> F
8662 if (CI->isMaxValue(Pred == ICmpInst::ICMP_SGT))
8663 return ReplaceInstUsesWith(SI, FalseVal);
8664 // X > C ? X : C+1 --> X < C+1 ? C+1 : X
8665 Constant *AdjustedRHS = AddOne(CI);
8666 if ((CmpLHS == TrueVal && AdjustedRHS == FalseVal) ||
8667 (CmpLHS == FalseVal && AdjustedRHS == TrueVal)) {
8668 Pred = ICmpInst::getSwappedPredicate(Pred);
8669 CmpRHS = AdjustedRHS;
8670 std::swap(FalseVal, TrueVal);
8671 ICI->setPredicate(Pred);
8672 ICI->setOperand(1, CmpRHS);
8673 SI.setOperand(1, TrueVal);
8674 SI.setOperand(2, FalseVal);
8681 // (x <s 0) ? -1 : 0 -> ashr x, 31 -> all ones if signed
8682 // (x >s -1) ? -1 : 0 -> ashr x, 31 -> all ones if not signed
8683 CmpInst::Predicate Pred = CmpInst::BAD_ICMP_PREDICATE;
8684 if (match(TrueVal, m_ConstantInt(-1)) &&
8685 match(FalseVal, m_ConstantInt(0)))
8686 Pred = ICI->getPredicate();
8687 else if (match(TrueVal, m_ConstantInt(0)) &&
8688 match(FalseVal, m_ConstantInt(-1)))
8689 Pred = CmpInst::getInversePredicate(ICI->getPredicate());
8691 if (Pred != CmpInst::BAD_ICMP_PREDICATE) {
8692 // If we are just checking for a icmp eq of a single bit and zext'ing it
8693 // to an integer, then shift the bit to the appropriate place and then
8694 // cast to integer to avoid the comparison.
8695 const APInt &Op1CV = CI->getValue();
8697 // sext (x <s 0) to i32 --> x>>s31 true if signbit set.
8698 // sext (x >s -1) to i32 --> (x>>s31)^-1 true if signbit clear.
8699 if ((Pred == ICmpInst::ICMP_SLT && Op1CV == 0) ||
8700 (Pred == ICmpInst::ICMP_SGT && Op1CV.isAllOnesValue())) {
8701 Value *In = ICI->getOperand(0);
8702 Value *Sh = ConstantInt::get(In->getType(),
8703 In->getType()->getPrimitiveSizeInBits()-1);
8704 In = InsertNewInstBefore(BinaryOperator::CreateAShr(In, Sh,
8705 In->getName()+".lobit"),
8707 if (In->getType() != SI.getType())
8708 In = CastInst::CreateIntegerCast(In, SI.getType(),
8709 true/*SExt*/, "tmp", ICI);
8711 if (Pred == ICmpInst::ICMP_SGT)
8712 In = InsertNewInstBefore(BinaryOperator::CreateNot(In,
8713 In->getName()+".not"), *ICI);
8715 return ReplaceInstUsesWith(SI, In);
8720 if (CmpLHS == TrueVal && CmpRHS == FalseVal) {
8721 // Transform (X == Y) ? X : Y -> Y
8722 if (Pred == ICmpInst::ICMP_EQ)
8723 return ReplaceInstUsesWith(SI, FalseVal);
8724 // Transform (X != Y) ? X : Y -> X
8725 if (Pred == ICmpInst::ICMP_NE)
8726 return ReplaceInstUsesWith(SI, TrueVal);
8727 /// NOTE: if we wanted to, this is where to detect integer MIN/MAX
8729 } else if (CmpLHS == FalseVal && CmpRHS == TrueVal) {
8730 // Transform (X == Y) ? Y : X -> X
8731 if (Pred == ICmpInst::ICMP_EQ)
8732 return ReplaceInstUsesWith(SI, FalseVal);
8733 // Transform (X != Y) ? Y : X -> Y
8734 if (Pred == ICmpInst::ICMP_NE)
8735 return ReplaceInstUsesWith(SI, TrueVal);
8736 /// NOTE: if we wanted to, this is where to detect integer MIN/MAX
8739 /// NOTE: if we wanted to, this is where to detect integer ABS
8741 return Changed ? &SI : 0;
8744 Instruction *InstCombiner::visitSelectInst(SelectInst &SI) {
8745 Value *CondVal = SI.getCondition();
8746 Value *TrueVal = SI.getTrueValue();
8747 Value *FalseVal = SI.getFalseValue();
8749 // select true, X, Y -> X
8750 // select false, X, Y -> Y
8751 if (ConstantInt *C = dyn_cast<ConstantInt>(CondVal))
8752 return ReplaceInstUsesWith(SI, C->getZExtValue() ? TrueVal : FalseVal);
8754 // select C, X, X -> X
8755 if (TrueVal == FalseVal)
8756 return ReplaceInstUsesWith(SI, TrueVal);
8758 if (isa<UndefValue>(TrueVal)) // select C, undef, X -> X
8759 return ReplaceInstUsesWith(SI, FalseVal);
8760 if (isa<UndefValue>(FalseVal)) // select C, X, undef -> X
8761 return ReplaceInstUsesWith(SI, TrueVal);
8762 if (isa<UndefValue>(CondVal)) { // select undef, X, Y -> X or Y
8763 if (isa<Constant>(TrueVal))
8764 return ReplaceInstUsesWith(SI, TrueVal);
8766 return ReplaceInstUsesWith(SI, FalseVal);
8769 if (SI.getType() == Type::Int1Ty) {
8770 if (ConstantInt *C = dyn_cast<ConstantInt>(TrueVal)) {
8771 if (C->getZExtValue()) {
8772 // Change: A = select B, true, C --> A = or B, C
8773 return BinaryOperator::CreateOr(CondVal, FalseVal);
8775 // Change: A = select B, false, C --> A = and !B, C
8777 InsertNewInstBefore(BinaryOperator::CreateNot(CondVal,
8778 "not."+CondVal->getName()), SI);
8779 return BinaryOperator::CreateAnd(NotCond, FalseVal);
8781 } else if (ConstantInt *C = dyn_cast<ConstantInt>(FalseVal)) {
8782 if (C->getZExtValue() == false) {
8783 // Change: A = select B, C, false --> A = and B, C
8784 return BinaryOperator::CreateAnd(CondVal, TrueVal);
8786 // Change: A = select B, C, true --> A = or !B, C
8788 InsertNewInstBefore(BinaryOperator::CreateNot(CondVal,
8789 "not."+CondVal->getName()), SI);
8790 return BinaryOperator::CreateOr(NotCond, TrueVal);
8794 // select a, b, a -> a&b
8795 // select a, a, b -> a|b
8796 if (CondVal == TrueVal)
8797 return BinaryOperator::CreateOr(CondVal, FalseVal);
8798 else if (CondVal == FalseVal)
8799 return BinaryOperator::CreateAnd(CondVal, TrueVal);
8802 // Selecting between two integer constants?
8803 if (ConstantInt *TrueValC = dyn_cast<ConstantInt>(TrueVal))
8804 if (ConstantInt *FalseValC = dyn_cast<ConstantInt>(FalseVal)) {
8805 // select C, 1, 0 -> zext C to int
8806 if (FalseValC->isZero() && TrueValC->getValue() == 1) {
8807 return CastInst::Create(Instruction::ZExt, CondVal, SI.getType());
8808 } else if (TrueValC->isZero() && FalseValC->getValue() == 1) {
8809 // select C, 0, 1 -> zext !C to int
8811 InsertNewInstBefore(BinaryOperator::CreateNot(CondVal,
8812 "not."+CondVal->getName()), SI);
8813 return CastInst::Create(Instruction::ZExt, NotCond, SI.getType());
8816 // FIXME: Turn select 0/-1 and -1/0 into sext from condition!
8818 if (ICmpInst *IC = dyn_cast<ICmpInst>(SI.getCondition())) {
8820 // (x <s 0) ? -1 : 0 -> ashr x, 31
8821 if (TrueValC->isAllOnesValue() && FalseValC->isZero())
8822 if (ConstantInt *CmpCst = dyn_cast<ConstantInt>(IC->getOperand(1))) {
8823 if (IC->getPredicate() == ICmpInst::ICMP_SLT && CmpCst->isZero()) {
8824 // The comparison constant and the result are not neccessarily the
8825 // same width. Make an all-ones value by inserting a AShr.
8826 Value *X = IC->getOperand(0);
8827 uint32_t Bits = X->getType()->getPrimitiveSizeInBits();
8828 Constant *ShAmt = ConstantInt::get(X->getType(), Bits-1);
8829 Instruction *SRA = BinaryOperator::Create(Instruction::AShr, X,
8831 InsertNewInstBefore(SRA, SI);
8833 // Finally, convert to the type of the select RHS. We figure out
8834 // if this requires a SExt, Trunc or BitCast based on the sizes.
8835 Instruction::CastOps opc = Instruction::BitCast;
8836 uint32_t SRASize = SRA->getType()->getPrimitiveSizeInBits();
8837 uint32_t SISize = SI.getType()->getPrimitiveSizeInBits();
8838 if (SRASize < SISize)
8839 opc = Instruction::SExt;
8840 else if (SRASize > SISize)
8841 opc = Instruction::Trunc;
8842 return CastInst::Create(opc, SRA, SI.getType());
8847 // If one of the constants is zero (we know they can't both be) and we
8848 // have an icmp instruction with zero, and we have an 'and' with the
8849 // non-constant value, eliminate this whole mess. This corresponds to
8850 // cases like this: ((X & 27) ? 27 : 0)
8851 if (TrueValC->isZero() || FalseValC->isZero())
8852 if (IC->isEquality() && isa<ConstantInt>(IC->getOperand(1)) &&
8853 cast<Constant>(IC->getOperand(1))->isNullValue())
8854 if (Instruction *ICA = dyn_cast<Instruction>(IC->getOperand(0)))
8855 if (ICA->getOpcode() == Instruction::And &&
8856 isa<ConstantInt>(ICA->getOperand(1)) &&
8857 (ICA->getOperand(1) == TrueValC ||
8858 ICA->getOperand(1) == FalseValC) &&
8859 isOneBitSet(cast<ConstantInt>(ICA->getOperand(1)))) {
8860 // Okay, now we know that everything is set up, we just don't
8861 // know whether we have a icmp_ne or icmp_eq and whether the
8862 // true or false val is the zero.
8863 bool ShouldNotVal = !TrueValC->isZero();
8864 ShouldNotVal ^= IC->getPredicate() == ICmpInst::ICMP_NE;
8867 V = InsertNewInstBefore(BinaryOperator::Create(
8868 Instruction::Xor, V, ICA->getOperand(1)), SI);
8869 return ReplaceInstUsesWith(SI, V);
8874 // See if we are selecting two values based on a comparison of the two values.
8875 if (FCmpInst *FCI = dyn_cast<FCmpInst>(CondVal)) {
8876 if (FCI->getOperand(0) == TrueVal && FCI->getOperand(1) == FalseVal) {
8877 // Transform (X == Y) ? X : Y -> Y
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) ? X : Y -> X
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 } else if (FCI->getOperand(0) == FalseVal && FCI->getOperand(1) == TrueVal){
8895 // Transform (X == Y) ? Y : X -> X
8896 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ) {
8897 // This is not safe in general for floating point:
8898 // consider X== -0, Y== +0.
8899 // It becomes safe if either operand is a nonzero constant.
8900 ConstantFP *CFPt, *CFPf;
8901 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) &&
8902 !CFPt->getValueAPF().isZero()) ||
8903 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) &&
8904 !CFPf->getValueAPF().isZero()))
8905 return ReplaceInstUsesWith(SI, FalseVal);
8907 // Transform (X != Y) ? Y : X -> Y
8908 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
8909 return ReplaceInstUsesWith(SI, TrueVal);
8910 // NOTE: if we wanted to, this is where to detect MIN/MAX
8912 // NOTE: if we wanted to, this is where to detect ABS
8915 // See if we are selecting two values based on a comparison of the two values.
8916 if (ICmpInst *ICI = dyn_cast<ICmpInst>(CondVal))
8917 if (Instruction *Result = visitSelectInstWithICmp(SI, ICI))
8920 if (Instruction *TI = dyn_cast<Instruction>(TrueVal))
8921 if (Instruction *FI = dyn_cast<Instruction>(FalseVal))
8922 if (TI->hasOneUse() && FI->hasOneUse()) {
8923 Instruction *AddOp = 0, *SubOp = 0;
8925 // Turn (select C, (op X, Y), (op X, Z)) -> (op X, (select C, Y, Z))
8926 if (TI->getOpcode() == FI->getOpcode())
8927 if (Instruction *IV = FoldSelectOpOp(SI, TI, FI))
8930 // Turn select C, (X+Y), (X-Y) --> (X+(select C, Y, (-Y))). This is
8931 // even legal for FP.
8932 if (TI->getOpcode() == Instruction::Sub &&
8933 FI->getOpcode() == Instruction::Add) {
8934 AddOp = FI; SubOp = TI;
8935 } else if (FI->getOpcode() == Instruction::Sub &&
8936 TI->getOpcode() == Instruction::Add) {
8937 AddOp = TI; SubOp = FI;
8941 Value *OtherAddOp = 0;
8942 if (SubOp->getOperand(0) == AddOp->getOperand(0)) {
8943 OtherAddOp = AddOp->getOperand(1);
8944 } else if (SubOp->getOperand(0) == AddOp->getOperand(1)) {
8945 OtherAddOp = AddOp->getOperand(0);
8949 // So at this point we know we have (Y -> OtherAddOp):
8950 // select C, (add X, Y), (sub X, Z)
8951 Value *NegVal; // Compute -Z
8952 if (Constant *C = dyn_cast<Constant>(SubOp->getOperand(1))) {
8953 NegVal = ConstantExpr::getNeg(C);
8955 NegVal = InsertNewInstBefore(
8956 BinaryOperator::CreateNeg(SubOp->getOperand(1), "tmp"), SI);
8959 Value *NewTrueOp = OtherAddOp;
8960 Value *NewFalseOp = NegVal;
8962 std::swap(NewTrueOp, NewFalseOp);
8963 Instruction *NewSel =
8964 SelectInst::Create(CondVal, NewTrueOp,
8965 NewFalseOp, SI.getName() + ".p");
8967 NewSel = InsertNewInstBefore(NewSel, SI);
8968 return BinaryOperator::CreateAdd(SubOp->getOperand(0), NewSel);
8973 // See if we can fold the select into one of our operands.
8974 if (SI.getType()->isInteger()) {
8975 // See the comment above GetSelectFoldableOperands for a description of the
8976 // transformation we are doing here.
8977 if (Instruction *TVI = dyn_cast<Instruction>(TrueVal))
8978 if (TVI->hasOneUse() && TVI->getNumOperands() == 2 &&
8979 !isa<Constant>(FalseVal))
8980 if (unsigned SFO = GetSelectFoldableOperands(TVI)) {
8981 unsigned OpToFold = 0;
8982 if ((SFO & 1) && FalseVal == TVI->getOperand(0)) {
8984 } else if ((SFO & 2) && FalseVal == TVI->getOperand(1)) {
8989 Constant *C = GetSelectFoldableConstant(TVI);
8990 Instruction *NewSel =
8991 SelectInst::Create(SI.getCondition(),
8992 TVI->getOperand(2-OpToFold), C);
8993 InsertNewInstBefore(NewSel, SI);
8994 NewSel->takeName(TVI);
8995 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TVI))
8996 return BinaryOperator::Create(BO->getOpcode(), FalseVal, NewSel);
8998 assert(0 && "Unknown instruction!!");
9003 if (Instruction *FVI = dyn_cast<Instruction>(FalseVal))
9004 if (FVI->hasOneUse() && FVI->getNumOperands() == 2 &&
9005 !isa<Constant>(TrueVal))
9006 if (unsigned SFO = GetSelectFoldableOperands(FVI)) {
9007 unsigned OpToFold = 0;
9008 if ((SFO & 1) && TrueVal == FVI->getOperand(0)) {
9010 } else if ((SFO & 2) && TrueVal == FVI->getOperand(1)) {
9015 Constant *C = GetSelectFoldableConstant(FVI);
9016 Instruction *NewSel =
9017 SelectInst::Create(SI.getCondition(), C,
9018 FVI->getOperand(2-OpToFold));
9019 InsertNewInstBefore(NewSel, SI);
9020 NewSel->takeName(FVI);
9021 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(FVI))
9022 return BinaryOperator::Create(BO->getOpcode(), TrueVal, NewSel);
9024 assert(0 && "Unknown instruction!!");
9029 if (BinaryOperator::isNot(CondVal)) {
9030 SI.setOperand(0, BinaryOperator::getNotArgument(CondVal));
9031 SI.setOperand(1, FalseVal);
9032 SI.setOperand(2, TrueVal);
9039 /// EnforceKnownAlignment - If the specified pointer points to an object that
9040 /// we control, modify the object's alignment to PrefAlign. This isn't
9041 /// often possible though. If alignment is important, a more reliable approach
9042 /// is to simply align all global variables and allocation instructions to
9043 /// their preferred alignment from the beginning.
9045 static unsigned EnforceKnownAlignment(Value *V,
9046 unsigned Align, unsigned PrefAlign) {
9048 User *U = dyn_cast<User>(V);
9049 if (!U) return Align;
9051 switch (getOpcode(U)) {
9053 case Instruction::BitCast:
9054 return EnforceKnownAlignment(U->getOperand(0), Align, PrefAlign);
9055 case Instruction::GetElementPtr: {
9056 // If all indexes are zero, it is just the alignment of the base pointer.
9057 bool AllZeroOperands = true;
9058 for (User::op_iterator i = U->op_begin() + 1, e = U->op_end(); i != e; ++i)
9059 if (!isa<Constant>(*i) ||
9060 !cast<Constant>(*i)->isNullValue()) {
9061 AllZeroOperands = false;
9065 if (AllZeroOperands) {
9066 // Treat this like a bitcast.
9067 return EnforceKnownAlignment(U->getOperand(0), Align, PrefAlign);
9073 if (GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
9074 // If there is a large requested alignment and we can, bump up the alignment
9076 if (!GV->isDeclaration()) {
9077 GV->setAlignment(PrefAlign);
9080 } else if (AllocationInst *AI = dyn_cast<AllocationInst>(V)) {
9081 // If there is a requested alignment and if this is an alloca, round up. We
9082 // don't do this for malloc, because some systems can't respect the request.
9083 if (isa<AllocaInst>(AI)) {
9084 AI->setAlignment(PrefAlign);
9092 /// GetOrEnforceKnownAlignment - If the specified pointer has an alignment that
9093 /// we can determine, return it, otherwise return 0. If PrefAlign is specified,
9094 /// and it is more than the alignment of the ultimate object, see if we can
9095 /// increase the alignment of the ultimate object, making this check succeed.
9096 unsigned InstCombiner::GetOrEnforceKnownAlignment(Value *V,
9097 unsigned PrefAlign) {
9098 unsigned BitWidth = TD ? TD->getTypeSizeInBits(V->getType()) :
9099 sizeof(PrefAlign) * CHAR_BIT;
9100 APInt Mask = APInt::getAllOnesValue(BitWidth);
9101 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
9102 ComputeMaskedBits(V, Mask, KnownZero, KnownOne);
9103 unsigned TrailZ = KnownZero.countTrailingOnes();
9104 unsigned Align = 1u << std::min(BitWidth - 1, TrailZ);
9106 if (PrefAlign > Align)
9107 Align = EnforceKnownAlignment(V, Align, PrefAlign);
9109 // We don't need to make any adjustment.
9113 Instruction *InstCombiner::SimplifyMemTransfer(MemIntrinsic *MI) {
9114 unsigned DstAlign = GetOrEnforceKnownAlignment(MI->getOperand(1));
9115 unsigned SrcAlign = GetOrEnforceKnownAlignment(MI->getOperand(2));
9116 unsigned MinAlign = std::min(DstAlign, SrcAlign);
9117 unsigned CopyAlign = MI->getAlignment()->getZExtValue();
9119 if (CopyAlign < MinAlign) {
9120 MI->setAlignment(ConstantInt::get(Type::Int32Ty, MinAlign));
9124 // If MemCpyInst length is 1/2/4/8 bytes then replace memcpy with
9126 ConstantInt *MemOpLength = dyn_cast<ConstantInt>(MI->getOperand(3));
9127 if (MemOpLength == 0) return 0;
9129 // Source and destination pointer types are always "i8*" for intrinsic. See
9130 // if the size is something we can handle with a single primitive load/store.
9131 // A single load+store correctly handles overlapping memory in the memmove
9133 unsigned Size = MemOpLength->getZExtValue();
9134 if (Size == 0) return MI; // Delete this mem transfer.
9136 if (Size > 8 || (Size&(Size-1)))
9137 return 0; // If not 1/2/4/8 bytes, exit.
9139 // Use an integer load+store unless we can find something better.
9140 Type *NewPtrTy = PointerType::getUnqual(IntegerType::get(Size<<3));
9142 // Memcpy forces the use of i8* for the source and destination. That means
9143 // that if you're using memcpy to move one double around, you'll get a cast
9144 // from double* to i8*. We'd much rather use a double load+store rather than
9145 // an i64 load+store, here because this improves the odds that the source or
9146 // dest address will be promotable. See if we can find a better type than the
9147 // integer datatype.
9148 if (Value *Op = getBitCastOperand(MI->getOperand(1))) {
9149 const Type *SrcETy = cast<PointerType>(Op->getType())->getElementType();
9150 if (SrcETy->isSized() && TD->getTypeStoreSize(SrcETy) == Size) {
9151 // The SrcETy might be something like {{{double}}} or [1 x double]. Rip
9152 // down through these levels if so.
9153 while (!SrcETy->isSingleValueType()) {
9154 if (const StructType *STy = dyn_cast<StructType>(SrcETy)) {
9155 if (STy->getNumElements() == 1)
9156 SrcETy = STy->getElementType(0);
9159 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(SrcETy)) {
9160 if (ATy->getNumElements() == 1)
9161 SrcETy = ATy->getElementType();
9168 if (SrcETy->isSingleValueType())
9169 NewPtrTy = PointerType::getUnqual(SrcETy);
9174 // If the memcpy/memmove provides better alignment info than we can
9176 SrcAlign = std::max(SrcAlign, CopyAlign);
9177 DstAlign = std::max(DstAlign, CopyAlign);
9179 Value *Src = InsertBitCastBefore(MI->getOperand(2), NewPtrTy, *MI);
9180 Value *Dest = InsertBitCastBefore(MI->getOperand(1), NewPtrTy, *MI);
9181 Instruction *L = new LoadInst(Src, "tmp", false, SrcAlign);
9182 InsertNewInstBefore(L, *MI);
9183 InsertNewInstBefore(new StoreInst(L, Dest, false, DstAlign), *MI);
9185 // Set the size of the copy to 0, it will be deleted on the next iteration.
9186 MI->setOperand(3, Constant::getNullValue(MemOpLength->getType()));
9190 Instruction *InstCombiner::SimplifyMemSet(MemSetInst *MI) {
9191 unsigned Alignment = GetOrEnforceKnownAlignment(MI->getDest());
9192 if (MI->getAlignment()->getZExtValue() < Alignment) {
9193 MI->setAlignment(ConstantInt::get(Type::Int32Ty, Alignment));
9197 // Extract the length and alignment and fill if they are constant.
9198 ConstantInt *LenC = dyn_cast<ConstantInt>(MI->getLength());
9199 ConstantInt *FillC = dyn_cast<ConstantInt>(MI->getValue());
9200 if (!LenC || !FillC || FillC->getType() != Type::Int8Ty)
9202 uint64_t Len = LenC->getZExtValue();
9203 Alignment = MI->getAlignment()->getZExtValue();
9205 // If the length is zero, this is a no-op
9206 if (Len == 0) return MI; // memset(d,c,0,a) -> noop
9208 // memset(s,c,n) -> store s, c (for n=1,2,4,8)
9209 if (Len <= 8 && isPowerOf2_32((uint32_t)Len)) {
9210 const Type *ITy = IntegerType::get(Len*8); // n=1 -> i8.
9212 Value *Dest = MI->getDest();
9213 Dest = InsertBitCastBefore(Dest, PointerType::getUnqual(ITy), *MI);
9215 // Alignment 0 is identity for alignment 1 for memset, but not store.
9216 if (Alignment == 0) Alignment = 1;
9218 // Extract the fill value and store.
9219 uint64_t Fill = FillC->getZExtValue()*0x0101010101010101ULL;
9220 InsertNewInstBefore(new StoreInst(ConstantInt::get(ITy, Fill), Dest, false,
9223 // Set the size of the copy to 0, it will be deleted on the next iteration.
9224 MI->setLength(Constant::getNullValue(LenC->getType()));
9232 /// visitCallInst - CallInst simplification. This mostly only handles folding
9233 /// of intrinsic instructions. For normal calls, it allows visitCallSite to do
9234 /// the heavy lifting.
9236 Instruction *InstCombiner::visitCallInst(CallInst &CI) {
9237 IntrinsicInst *II = dyn_cast<IntrinsicInst>(&CI);
9238 if (!II) return visitCallSite(&CI);
9240 // Intrinsics cannot occur in an invoke, so handle them here instead of in
9242 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(II)) {
9243 bool Changed = false;
9245 // memmove/cpy/set of zero bytes is a noop.
9246 if (Constant *NumBytes = dyn_cast<Constant>(MI->getLength())) {
9247 if (NumBytes->isNullValue()) return EraseInstFromFunction(CI);
9249 if (ConstantInt *CI = dyn_cast<ConstantInt>(NumBytes))
9250 if (CI->getZExtValue() == 1) {
9251 // Replace the instruction with just byte operations. We would
9252 // transform other cases to loads/stores, but we don't know if
9253 // alignment is sufficient.
9257 // If we have a memmove and the source operation is a constant global,
9258 // then the source and dest pointers can't alias, so we can change this
9259 // into a call to memcpy.
9260 if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(MI)) {
9261 if (GlobalVariable *GVSrc = dyn_cast<GlobalVariable>(MMI->getSource()))
9262 if (GVSrc->isConstant()) {
9263 Module *M = CI.getParent()->getParent()->getParent();
9264 Intrinsic::ID MemCpyID = Intrinsic::memcpy;
9266 Tys[0] = CI.getOperand(3)->getType();
9268 Intrinsic::getDeclaration(M, MemCpyID, Tys, 1));
9272 // memmove(x,x,size) -> noop.
9273 if (MMI->getSource() == MMI->getDest())
9274 return EraseInstFromFunction(CI);
9277 // If we can determine a pointer alignment that is bigger than currently
9278 // set, update the alignment.
9279 if (isa<MemCpyInst>(MI) || isa<MemMoveInst>(MI)) {
9280 if (Instruction *I = SimplifyMemTransfer(MI))
9282 } else if (MemSetInst *MSI = dyn_cast<MemSetInst>(MI)) {
9283 if (Instruction *I = SimplifyMemSet(MSI))
9287 if (Changed) return II;
9290 switch (II->getIntrinsicID()) {
9292 case Intrinsic::bswap:
9293 // bswap(bswap(x)) -> x
9294 if (IntrinsicInst *Operand = dyn_cast<IntrinsicInst>(II->getOperand(1)))
9295 if (Operand->getIntrinsicID() == Intrinsic::bswap)
9296 return ReplaceInstUsesWith(CI, Operand->getOperand(1));
9298 case Intrinsic::ppc_altivec_lvx:
9299 case Intrinsic::ppc_altivec_lvxl:
9300 case Intrinsic::x86_sse_loadu_ps:
9301 case Intrinsic::x86_sse2_loadu_pd:
9302 case Intrinsic::x86_sse2_loadu_dq:
9303 // Turn PPC lvx -> load if the pointer is known aligned.
9304 // Turn X86 loadups -> load if the pointer is known aligned.
9305 if (GetOrEnforceKnownAlignment(II->getOperand(1), 16) >= 16) {
9306 Value *Ptr = InsertBitCastBefore(II->getOperand(1),
9307 PointerType::getUnqual(II->getType()),
9309 return new LoadInst(Ptr);
9312 case Intrinsic::ppc_altivec_stvx:
9313 case Intrinsic::ppc_altivec_stvxl:
9314 // Turn stvx -> store if the pointer is known aligned.
9315 if (GetOrEnforceKnownAlignment(II->getOperand(2), 16) >= 16) {
9316 const Type *OpPtrTy =
9317 PointerType::getUnqual(II->getOperand(1)->getType());
9318 Value *Ptr = InsertBitCastBefore(II->getOperand(2), OpPtrTy, CI);
9319 return new StoreInst(II->getOperand(1), Ptr);
9322 case Intrinsic::x86_sse_storeu_ps:
9323 case Intrinsic::x86_sse2_storeu_pd:
9324 case Intrinsic::x86_sse2_storeu_dq:
9325 // Turn X86 storeu -> store if the pointer is known aligned.
9326 if (GetOrEnforceKnownAlignment(II->getOperand(1), 16) >= 16) {
9327 const Type *OpPtrTy =
9328 PointerType::getUnqual(II->getOperand(2)->getType());
9329 Value *Ptr = InsertBitCastBefore(II->getOperand(1), OpPtrTy, CI);
9330 return new StoreInst(II->getOperand(2), Ptr);
9334 case Intrinsic::x86_sse_cvttss2si: {
9335 // These intrinsics only demands the 0th element of its input vector. If
9336 // we can simplify the input based on that, do so now.
9338 if (Value *V = SimplifyDemandedVectorElts(II->getOperand(1), 1,
9340 II->setOperand(1, V);
9346 case Intrinsic::ppc_altivec_vperm:
9347 // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant.
9348 if (ConstantVector *Mask = dyn_cast<ConstantVector>(II->getOperand(3))) {
9349 assert(Mask->getNumOperands() == 16 && "Bad type for intrinsic!");
9351 // Check that all of the elements are integer constants or undefs.
9352 bool AllEltsOk = true;
9353 for (unsigned i = 0; i != 16; ++i) {
9354 if (!isa<ConstantInt>(Mask->getOperand(i)) &&
9355 !isa<UndefValue>(Mask->getOperand(i))) {
9362 // Cast the input vectors to byte vectors.
9363 Value *Op0 =InsertBitCastBefore(II->getOperand(1),Mask->getType(),CI);
9364 Value *Op1 =InsertBitCastBefore(II->getOperand(2),Mask->getType(),CI);
9365 Value *Result = UndefValue::get(Op0->getType());
9367 // Only extract each element once.
9368 Value *ExtractedElts[32];
9369 memset(ExtractedElts, 0, sizeof(ExtractedElts));
9371 for (unsigned i = 0; i != 16; ++i) {
9372 if (isa<UndefValue>(Mask->getOperand(i)))
9374 unsigned Idx=cast<ConstantInt>(Mask->getOperand(i))->getZExtValue();
9375 Idx &= 31; // Match the hardware behavior.
9377 if (ExtractedElts[Idx] == 0) {
9379 new ExtractElementInst(Idx < 16 ? Op0 : Op1, Idx&15, "tmp");
9380 InsertNewInstBefore(Elt, CI);
9381 ExtractedElts[Idx] = Elt;
9384 // Insert this value into the result vector.
9385 Result = InsertElementInst::Create(Result, ExtractedElts[Idx],
9387 InsertNewInstBefore(cast<Instruction>(Result), CI);
9389 return CastInst::Create(Instruction::BitCast, Result, CI.getType());
9394 case Intrinsic::stackrestore: {
9395 // If the save is right next to the restore, remove the restore. This can
9396 // happen when variable allocas are DCE'd.
9397 if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getOperand(1))) {
9398 if (SS->getIntrinsicID() == Intrinsic::stacksave) {
9399 BasicBlock::iterator BI = SS;
9401 return EraseInstFromFunction(CI);
9405 // Scan down this block to see if there is another stack restore in the
9406 // same block without an intervening call/alloca.
9407 BasicBlock::iterator BI = II;
9408 TerminatorInst *TI = II->getParent()->getTerminator();
9409 bool CannotRemove = false;
9410 for (++BI; &*BI != TI; ++BI) {
9411 if (isa<AllocaInst>(BI)) {
9412 CannotRemove = true;
9415 if (CallInst *BCI = dyn_cast<CallInst>(BI)) {
9416 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(BCI)) {
9417 // If there is a stackrestore below this one, remove this one.
9418 if (II->getIntrinsicID() == Intrinsic::stackrestore)
9419 return EraseInstFromFunction(CI);
9420 // Otherwise, ignore the intrinsic.
9422 // If we found a non-intrinsic call, we can't remove the stack
9424 CannotRemove = true;
9430 // If the stack restore is in a return/unwind block and if there are no
9431 // allocas or calls between the restore and the return, nuke the restore.
9432 if (!CannotRemove && (isa<ReturnInst>(TI) || isa<UnwindInst>(TI)))
9433 return EraseInstFromFunction(CI);
9438 return visitCallSite(II);
9441 // InvokeInst simplification
9443 Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) {
9444 return visitCallSite(&II);
9447 /// isSafeToEliminateVarargsCast - If this cast does not affect the value
9448 /// passed through the varargs area, we can eliminate the use of the cast.
9449 static bool isSafeToEliminateVarargsCast(const CallSite CS,
9450 const CastInst * const CI,
9451 const TargetData * const TD,
9453 if (!CI->isLosslessCast())
9456 // The size of ByVal arguments is derived from the type, so we
9457 // can't change to a type with a different size. If the size were
9458 // passed explicitly we could avoid this check.
9459 if (!CS.paramHasAttr(ix, Attribute::ByVal))
9463 cast<PointerType>(CI->getOperand(0)->getType())->getElementType();
9464 const Type* DstTy = cast<PointerType>(CI->getType())->getElementType();
9465 if (!SrcTy->isSized() || !DstTy->isSized())
9467 if (TD->getABITypeSize(SrcTy) != TD->getABITypeSize(DstTy))
9472 // visitCallSite - Improvements for call and invoke instructions.
9474 Instruction *InstCombiner::visitCallSite(CallSite CS) {
9475 bool Changed = false;
9477 // If the callee is a constexpr cast of a function, attempt to move the cast
9478 // to the arguments of the call/invoke.
9479 if (transformConstExprCastCall(CS)) return 0;
9481 Value *Callee = CS.getCalledValue();
9483 if (Function *CalleeF = dyn_cast<Function>(Callee))
9484 if (CalleeF->getCallingConv() != CS.getCallingConv()) {
9485 Instruction *OldCall = CS.getInstruction();
9486 // If the call and callee calling conventions don't match, this call must
9487 // be unreachable, as the call is undefined.
9488 new StoreInst(ConstantInt::getTrue(),
9489 UndefValue::get(PointerType::getUnqual(Type::Int1Ty)),
9491 if (!OldCall->use_empty())
9492 OldCall->replaceAllUsesWith(UndefValue::get(OldCall->getType()));
9493 if (isa<CallInst>(OldCall)) // Not worth removing an invoke here.
9494 return EraseInstFromFunction(*OldCall);
9498 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) {
9499 // This instruction is not reachable, just remove it. We insert a store to
9500 // undef so that we know that this code is not reachable, despite the fact
9501 // that we can't modify the CFG here.
9502 new StoreInst(ConstantInt::getTrue(),
9503 UndefValue::get(PointerType::getUnqual(Type::Int1Ty)),
9504 CS.getInstruction());
9506 if (!CS.getInstruction()->use_empty())
9507 CS.getInstruction()->
9508 replaceAllUsesWith(UndefValue::get(CS.getInstruction()->getType()));
9510 if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) {
9511 // Don't break the CFG, insert a dummy cond branch.
9512 BranchInst::Create(II->getNormalDest(), II->getUnwindDest(),
9513 ConstantInt::getTrue(), II);
9515 return EraseInstFromFunction(*CS.getInstruction());
9518 if (BitCastInst *BC = dyn_cast<BitCastInst>(Callee))
9519 if (IntrinsicInst *In = dyn_cast<IntrinsicInst>(BC->getOperand(0)))
9520 if (In->getIntrinsicID() == Intrinsic::init_trampoline)
9521 return transformCallThroughTrampoline(CS);
9523 const PointerType *PTy = cast<PointerType>(Callee->getType());
9524 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
9525 if (FTy->isVarArg()) {
9526 int ix = FTy->getNumParams() + (isa<InvokeInst>(Callee) ? 3 : 1);
9527 // See if we can optimize any arguments passed through the varargs area of
9529 for (CallSite::arg_iterator I = CS.arg_begin()+FTy->getNumParams(),
9530 E = CS.arg_end(); I != E; ++I, ++ix) {
9531 CastInst *CI = dyn_cast<CastInst>(*I);
9532 if (CI && isSafeToEliminateVarargsCast(CS, CI, TD, ix)) {
9533 *I = CI->getOperand(0);
9539 if (isa<InlineAsm>(Callee) && !CS.doesNotThrow()) {
9540 // Inline asm calls cannot throw - mark them 'nounwind'.
9541 CS.setDoesNotThrow();
9545 return Changed ? CS.getInstruction() : 0;
9548 // transformConstExprCastCall - If the callee is a constexpr cast of a function,
9549 // attempt to move the cast to the arguments of the call/invoke.
9551 bool InstCombiner::transformConstExprCastCall(CallSite CS) {
9552 if (!isa<ConstantExpr>(CS.getCalledValue())) return false;
9553 ConstantExpr *CE = cast<ConstantExpr>(CS.getCalledValue());
9554 if (CE->getOpcode() != Instruction::BitCast ||
9555 !isa<Function>(CE->getOperand(0)))
9557 Function *Callee = cast<Function>(CE->getOperand(0));
9558 Instruction *Caller = CS.getInstruction();
9559 const AttrListPtr &CallerPAL = CS.getAttributes();
9561 // Okay, this is a cast from a function to a different type. Unless doing so
9562 // would cause a type conversion of one of our arguments, change this call to
9563 // be a direct call with arguments casted to the appropriate types.
9565 const FunctionType *FT = Callee->getFunctionType();
9566 const Type *OldRetTy = Caller->getType();
9567 const Type *NewRetTy = FT->getReturnType();
9569 if (isa<StructType>(NewRetTy))
9570 return false; // TODO: Handle multiple return values.
9572 // Check to see if we are changing the return type...
9573 if (OldRetTy != NewRetTy) {
9574 if (Callee->isDeclaration() &&
9575 // Conversion is ok if changing from one pointer type to another or from
9576 // a pointer to an integer of the same size.
9577 !((isa<PointerType>(OldRetTy) || OldRetTy == TD->getIntPtrType()) &&
9578 (isa<PointerType>(NewRetTy) || NewRetTy == TD->getIntPtrType())))
9579 return false; // Cannot transform this return value.
9581 if (!Caller->use_empty() &&
9582 // void -> non-void is handled specially
9583 NewRetTy != Type::VoidTy && !CastInst::isCastable(NewRetTy, OldRetTy))
9584 return false; // Cannot transform this return value.
9586 if (!CallerPAL.isEmpty() && !Caller->use_empty()) {
9587 Attributes RAttrs = CallerPAL.getRetAttributes();
9588 if (RAttrs & Attribute::typeIncompatible(NewRetTy))
9589 return false; // Attribute not compatible with transformed value.
9592 // If the callsite is an invoke instruction, and the return value is used by
9593 // a PHI node in a successor, we cannot change the return type of the call
9594 // because there is no place to put the cast instruction (without breaking
9595 // the critical edge). Bail out in this case.
9596 if (!Caller->use_empty())
9597 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller))
9598 for (Value::use_iterator UI = II->use_begin(), E = II->use_end();
9600 if (PHINode *PN = dyn_cast<PHINode>(*UI))
9601 if (PN->getParent() == II->getNormalDest() ||
9602 PN->getParent() == II->getUnwindDest())
9606 unsigned NumActualArgs = unsigned(CS.arg_end()-CS.arg_begin());
9607 unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs);
9609 CallSite::arg_iterator AI = CS.arg_begin();
9610 for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) {
9611 const Type *ParamTy = FT->getParamType(i);
9612 const Type *ActTy = (*AI)->getType();
9614 if (!CastInst::isCastable(ActTy, ParamTy))
9615 return false; // Cannot transform this parameter value.
9617 if (CallerPAL.getParamAttributes(i + 1)
9618 & Attribute::typeIncompatible(ParamTy))
9619 return false; // Attribute not compatible with transformed value.
9621 // Converting from one pointer type to another or between a pointer and an
9622 // integer of the same size is safe even if we do not have a body.
9623 bool isConvertible = ActTy == ParamTy ||
9624 ((isa<PointerType>(ParamTy) || ParamTy == TD->getIntPtrType()) &&
9625 (isa<PointerType>(ActTy) || ActTy == TD->getIntPtrType()));
9626 if (Callee->isDeclaration() && !isConvertible) return false;
9629 if (FT->getNumParams() < NumActualArgs && !FT->isVarArg() &&
9630 Callee->isDeclaration())
9631 return false; // Do not delete arguments unless we have a function body.
9633 if (FT->getNumParams() < NumActualArgs && FT->isVarArg() &&
9634 !CallerPAL.isEmpty())
9635 // In this case we have more arguments than the new function type, but we
9636 // won't be dropping them. Check that these extra arguments have attributes
9637 // that are compatible with being a vararg call argument.
9638 for (unsigned i = CallerPAL.getNumSlots(); i; --i) {
9639 if (CallerPAL.getSlot(i - 1).Index <= FT->getNumParams())
9641 Attributes PAttrs = CallerPAL.getSlot(i - 1).Attrs;
9642 if (PAttrs & Attribute::VarArgsIncompatible)
9646 // Okay, we decided that this is a safe thing to do: go ahead and start
9647 // inserting cast instructions as necessary...
9648 std::vector<Value*> Args;
9649 Args.reserve(NumActualArgs);
9650 SmallVector<AttributeWithIndex, 8> attrVec;
9651 attrVec.reserve(NumCommonArgs);
9653 // Get any return attributes.
9654 Attributes RAttrs = CallerPAL.getRetAttributes();
9656 // If the return value is not being used, the type may not be compatible
9657 // with the existing attributes. Wipe out any problematic attributes.
9658 RAttrs &= ~Attribute::typeIncompatible(NewRetTy);
9660 // Add the new return attributes.
9662 attrVec.push_back(AttributeWithIndex::get(0, RAttrs));
9664 AI = CS.arg_begin();
9665 for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) {
9666 const Type *ParamTy = FT->getParamType(i);
9667 if ((*AI)->getType() == ParamTy) {
9668 Args.push_back(*AI);
9670 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI,
9671 false, ParamTy, false);
9672 CastInst *NewCast = CastInst::Create(opcode, *AI, ParamTy, "tmp");
9673 Args.push_back(InsertNewInstBefore(NewCast, *Caller));
9676 // Add any parameter attributes.
9677 if (Attributes PAttrs = CallerPAL.getParamAttributes(i + 1))
9678 attrVec.push_back(AttributeWithIndex::get(i + 1, PAttrs));
9681 // If the function takes more arguments than the call was taking, add them
9683 for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i)
9684 Args.push_back(Constant::getNullValue(FT->getParamType(i)));
9686 // If we are removing arguments to the function, emit an obnoxious warning...
9687 if (FT->getNumParams() < NumActualArgs) {
9688 if (!FT->isVarArg()) {
9689 cerr << "WARNING: While resolving call to function '"
9690 << Callee->getName() << "' arguments were dropped!\n";
9692 // Add all of the arguments in their promoted form to the arg list...
9693 for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) {
9694 const Type *PTy = getPromotedType((*AI)->getType());
9695 if (PTy != (*AI)->getType()) {
9696 // Must promote to pass through va_arg area!
9697 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI, false,
9699 Instruction *Cast = CastInst::Create(opcode, *AI, PTy, "tmp");
9700 InsertNewInstBefore(Cast, *Caller);
9701 Args.push_back(Cast);
9703 Args.push_back(*AI);
9706 // Add any parameter attributes.
9707 if (Attributes PAttrs = CallerPAL.getParamAttributes(i + 1))
9708 attrVec.push_back(AttributeWithIndex::get(i + 1, PAttrs));
9713 if (Attributes FnAttrs = CallerPAL.getFnAttributes())
9714 attrVec.push_back(AttributeWithIndex::get(~0, FnAttrs));
9716 if (NewRetTy == Type::VoidTy)
9717 Caller->setName(""); // Void type should not have a name.
9719 const AttrListPtr &NewCallerPAL = AttrListPtr::get(attrVec.begin(),attrVec.end());
9722 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
9723 NC = InvokeInst::Create(Callee, II->getNormalDest(), II->getUnwindDest(),
9724 Args.begin(), Args.end(),
9725 Caller->getName(), Caller);
9726 cast<InvokeInst>(NC)->setCallingConv(II->getCallingConv());
9727 cast<InvokeInst>(NC)->setAttributes(NewCallerPAL);
9729 NC = CallInst::Create(Callee, Args.begin(), Args.end(),
9730 Caller->getName(), Caller);
9731 CallInst *CI = cast<CallInst>(Caller);
9732 if (CI->isTailCall())
9733 cast<CallInst>(NC)->setTailCall();
9734 cast<CallInst>(NC)->setCallingConv(CI->getCallingConv());
9735 cast<CallInst>(NC)->setAttributes(NewCallerPAL);
9738 // Insert a cast of the return type as necessary.
9740 if (OldRetTy != NV->getType() && !Caller->use_empty()) {
9741 if (NV->getType() != Type::VoidTy) {
9742 Instruction::CastOps opcode = CastInst::getCastOpcode(NC, false,
9744 NV = NC = CastInst::Create(opcode, NC, OldRetTy, "tmp");
9746 // If this is an invoke instruction, we should insert it after the first
9747 // non-phi, instruction in the normal successor block.
9748 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
9749 BasicBlock::iterator I = II->getNormalDest()->getFirstNonPHI();
9750 InsertNewInstBefore(NC, *I);
9752 // Otherwise, it's a call, just insert cast right after the call instr
9753 InsertNewInstBefore(NC, *Caller);
9755 AddUsersToWorkList(*Caller);
9757 NV = UndefValue::get(Caller->getType());
9761 if (Caller->getType() != Type::VoidTy && !Caller->use_empty())
9762 Caller->replaceAllUsesWith(NV);
9763 Caller->eraseFromParent();
9764 RemoveFromWorkList(Caller);
9768 // transformCallThroughTrampoline - Turn a call to a function created by the
9769 // init_trampoline intrinsic into a direct call to the underlying function.
9771 Instruction *InstCombiner::transformCallThroughTrampoline(CallSite CS) {
9772 Value *Callee = CS.getCalledValue();
9773 const PointerType *PTy = cast<PointerType>(Callee->getType());
9774 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
9775 const AttrListPtr &Attrs = CS.getAttributes();
9777 // If the call already has the 'nest' attribute somewhere then give up -
9778 // otherwise 'nest' would occur twice after splicing in the chain.
9779 if (Attrs.hasAttrSomewhere(Attribute::Nest))
9782 IntrinsicInst *Tramp =
9783 cast<IntrinsicInst>(cast<BitCastInst>(Callee)->getOperand(0));
9785 Function *NestF = cast<Function>(Tramp->getOperand(2)->stripPointerCasts());
9786 const PointerType *NestFPTy = cast<PointerType>(NestF->getType());
9787 const FunctionType *NestFTy = cast<FunctionType>(NestFPTy->getElementType());
9789 const AttrListPtr &NestAttrs = NestF->getAttributes();
9790 if (!NestAttrs.isEmpty()) {
9791 unsigned NestIdx = 1;
9792 const Type *NestTy = 0;
9793 Attributes NestAttr = Attribute::None;
9795 // Look for a parameter marked with the 'nest' attribute.
9796 for (FunctionType::param_iterator I = NestFTy->param_begin(),
9797 E = NestFTy->param_end(); I != E; ++NestIdx, ++I)
9798 if (NestAttrs.paramHasAttr(NestIdx, Attribute::Nest)) {
9799 // Record the parameter type and any other attributes.
9801 NestAttr = NestAttrs.getParamAttributes(NestIdx);
9806 Instruction *Caller = CS.getInstruction();
9807 std::vector<Value*> NewArgs;
9808 NewArgs.reserve(unsigned(CS.arg_end()-CS.arg_begin())+1);
9810 SmallVector<AttributeWithIndex, 8> NewAttrs;
9811 NewAttrs.reserve(Attrs.getNumSlots() + 1);
9813 // Insert the nest argument into the call argument list, which may
9814 // mean appending it. Likewise for attributes.
9816 // Add any result attributes.
9817 if (Attributes Attr = Attrs.getRetAttributes())
9818 NewAttrs.push_back(AttributeWithIndex::get(0, Attr));
9822 CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
9824 if (Idx == NestIdx) {
9825 // Add the chain argument and attributes.
9826 Value *NestVal = Tramp->getOperand(3);
9827 if (NestVal->getType() != NestTy)
9828 NestVal = new BitCastInst(NestVal, NestTy, "nest", Caller);
9829 NewArgs.push_back(NestVal);
9830 NewAttrs.push_back(AttributeWithIndex::get(NestIdx, NestAttr));
9836 // Add the original argument and attributes.
9837 NewArgs.push_back(*I);
9838 if (Attributes Attr = Attrs.getParamAttributes(Idx))
9840 (AttributeWithIndex::get(Idx + (Idx >= NestIdx), Attr));
9846 // Add any function attributes.
9847 if (Attributes Attr = Attrs.getFnAttributes())
9848 NewAttrs.push_back(AttributeWithIndex::get(~0, Attr));
9850 // The trampoline may have been bitcast to a bogus type (FTy).
9851 // Handle this by synthesizing a new function type, equal to FTy
9852 // with the chain parameter inserted.
9854 std::vector<const Type*> NewTypes;
9855 NewTypes.reserve(FTy->getNumParams()+1);
9857 // Insert the chain's type into the list of parameter types, which may
9858 // mean appending it.
9861 FunctionType::param_iterator I = FTy->param_begin(),
9862 E = FTy->param_end();
9866 // Add the chain's type.
9867 NewTypes.push_back(NestTy);
9872 // Add the original type.
9873 NewTypes.push_back(*I);
9879 // Replace the trampoline call with a direct call. Let the generic
9880 // code sort out any function type mismatches.
9881 FunctionType *NewFTy =
9882 FunctionType::get(FTy->getReturnType(), NewTypes, FTy->isVarArg());
9883 Constant *NewCallee = NestF->getType() == PointerType::getUnqual(NewFTy) ?
9884 NestF : ConstantExpr::getBitCast(NestF, PointerType::getUnqual(NewFTy));
9885 const AttrListPtr &NewPAL = AttrListPtr::get(NewAttrs.begin(),NewAttrs.end());
9887 Instruction *NewCaller;
9888 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
9889 NewCaller = InvokeInst::Create(NewCallee,
9890 II->getNormalDest(), II->getUnwindDest(),
9891 NewArgs.begin(), NewArgs.end(),
9892 Caller->getName(), Caller);
9893 cast<InvokeInst>(NewCaller)->setCallingConv(II->getCallingConv());
9894 cast<InvokeInst>(NewCaller)->setAttributes(NewPAL);
9896 NewCaller = CallInst::Create(NewCallee, NewArgs.begin(), NewArgs.end(),
9897 Caller->getName(), Caller);
9898 if (cast<CallInst>(Caller)->isTailCall())
9899 cast<CallInst>(NewCaller)->setTailCall();
9900 cast<CallInst>(NewCaller)->
9901 setCallingConv(cast<CallInst>(Caller)->getCallingConv());
9902 cast<CallInst>(NewCaller)->setAttributes(NewPAL);
9904 if (Caller->getType() != Type::VoidTy && !Caller->use_empty())
9905 Caller->replaceAllUsesWith(NewCaller);
9906 Caller->eraseFromParent();
9907 RemoveFromWorkList(Caller);
9912 // Replace the trampoline call with a direct call. Since there is no 'nest'
9913 // parameter, there is no need to adjust the argument list. Let the generic
9914 // code sort out any function type mismatches.
9915 Constant *NewCallee =
9916 NestF->getType() == PTy ? NestF : ConstantExpr::getBitCast(NestF, PTy);
9917 CS.setCalledFunction(NewCallee);
9918 return CS.getInstruction();
9921 /// FoldPHIArgBinOpIntoPHI - If we have something like phi [add (a,b), add(c,d)]
9922 /// and if a/b/c/d and the add's all have a single use, turn this into two phi's
9923 /// and a single binop.
9924 Instruction *InstCombiner::FoldPHIArgBinOpIntoPHI(PHINode &PN) {
9925 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
9926 assert(isa<BinaryOperator>(FirstInst) || isa<GetElementPtrInst>(FirstInst) ||
9927 isa<CmpInst>(FirstInst));
9928 unsigned Opc = FirstInst->getOpcode();
9929 Value *LHSVal = FirstInst->getOperand(0);
9930 Value *RHSVal = FirstInst->getOperand(1);
9932 const Type *LHSType = LHSVal->getType();
9933 const Type *RHSType = RHSVal->getType();
9935 // Scan to see if all operands are the same opcode, all have one use, and all
9936 // kill their operands (i.e. the operands have one use).
9937 for (unsigned i = 0; i != PN.getNumIncomingValues(); ++i) {
9938 Instruction *I = dyn_cast<Instruction>(PN.getIncomingValue(i));
9939 if (!I || I->getOpcode() != Opc || !I->hasOneUse() ||
9940 // Verify type of the LHS matches so we don't fold cmp's of different
9941 // types or GEP's with different index types.
9942 I->getOperand(0)->getType() != LHSType ||
9943 I->getOperand(1)->getType() != RHSType)
9946 // If they are CmpInst instructions, check their predicates
9947 if (Opc == Instruction::ICmp || Opc == Instruction::FCmp)
9948 if (cast<CmpInst>(I)->getPredicate() !=
9949 cast<CmpInst>(FirstInst)->getPredicate())
9952 // Keep track of which operand needs a phi node.
9953 if (I->getOperand(0) != LHSVal) LHSVal = 0;
9954 if (I->getOperand(1) != RHSVal) RHSVal = 0;
9957 // Otherwise, this is safe to transform, determine if it is profitable.
9959 // If this is a GEP, and if the index (not the pointer) needs a PHI, bail out.
9960 // Indexes are often folded into load/store instructions, so we don't want to
9961 // hide them behind a phi.
9962 if (isa<GetElementPtrInst>(FirstInst) && RHSVal == 0)
9965 Value *InLHS = FirstInst->getOperand(0);
9966 Value *InRHS = FirstInst->getOperand(1);
9967 PHINode *NewLHS = 0, *NewRHS = 0;
9969 NewLHS = PHINode::Create(LHSType,
9970 FirstInst->getOperand(0)->getName() + ".pn");
9971 NewLHS->reserveOperandSpace(PN.getNumOperands()/2);
9972 NewLHS->addIncoming(InLHS, PN.getIncomingBlock(0));
9973 InsertNewInstBefore(NewLHS, PN);
9978 NewRHS = PHINode::Create(RHSType,
9979 FirstInst->getOperand(1)->getName() + ".pn");
9980 NewRHS->reserveOperandSpace(PN.getNumOperands()/2);
9981 NewRHS->addIncoming(InRHS, PN.getIncomingBlock(0));
9982 InsertNewInstBefore(NewRHS, PN);
9986 // Add all operands to the new PHIs.
9987 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
9989 Value *NewInLHS =cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
9990 NewLHS->addIncoming(NewInLHS, PN.getIncomingBlock(i));
9993 Value *NewInRHS =cast<Instruction>(PN.getIncomingValue(i))->getOperand(1);
9994 NewRHS->addIncoming(NewInRHS, PN.getIncomingBlock(i));
9998 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
9999 return BinaryOperator::Create(BinOp->getOpcode(), LHSVal, RHSVal);
10000 else if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst))
10001 return CmpInst::Create(CIOp->getOpcode(), CIOp->getPredicate(), LHSVal,
10004 assert(isa<GetElementPtrInst>(FirstInst));
10005 return GetElementPtrInst::Create(LHSVal, RHSVal);
10009 /// isSafeToSinkLoad - Return true if we know that it is safe sink the load out
10010 /// of the block that defines it. This means that it must be obvious the value
10011 /// of the load is not changed from the point of the load to the end of the
10012 /// block it is in.
10014 /// Finally, it is safe, but not profitable, to sink a load targetting a
10015 /// non-address-taken alloca. Doing so will cause us to not promote the alloca
10017 static bool isSafeToSinkLoad(LoadInst *L) {
10018 BasicBlock::iterator BBI = L, E = L->getParent()->end();
10020 for (++BBI; BBI != E; ++BBI)
10021 if (BBI->mayWriteToMemory())
10024 // Check for non-address taken alloca. If not address-taken already, it isn't
10025 // profitable to do this xform.
10026 if (AllocaInst *AI = dyn_cast<AllocaInst>(L->getOperand(0))) {
10027 bool isAddressTaken = false;
10028 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
10030 if (isa<LoadInst>(UI)) continue;
10031 if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
10032 // If storing TO the alloca, then the address isn't taken.
10033 if (SI->getOperand(1) == AI) continue;
10035 isAddressTaken = true;
10039 if (!isAddressTaken)
10047 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
10048 // operator and they all are only used by the PHI, PHI together their
10049 // inputs, and do the operation once, to the result of the PHI.
10050 Instruction *InstCombiner::FoldPHIArgOpIntoPHI(PHINode &PN) {
10051 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
10053 // Scan the instruction, looking for input operations that can be folded away.
10054 // If all input operands to the phi are the same instruction (e.g. a cast from
10055 // the same type or "+42") we can pull the operation through the PHI, reducing
10056 // code size and simplifying code.
10057 Constant *ConstantOp = 0;
10058 const Type *CastSrcTy = 0;
10059 bool isVolatile = false;
10060 if (isa<CastInst>(FirstInst)) {
10061 CastSrcTy = FirstInst->getOperand(0)->getType();
10062 } else if (isa<BinaryOperator>(FirstInst) || isa<CmpInst>(FirstInst)) {
10063 // Can fold binop, compare or shift here if the RHS is a constant,
10064 // otherwise call FoldPHIArgBinOpIntoPHI.
10065 ConstantOp = dyn_cast<Constant>(FirstInst->getOperand(1));
10066 if (ConstantOp == 0)
10067 return FoldPHIArgBinOpIntoPHI(PN);
10068 } else if (LoadInst *LI = dyn_cast<LoadInst>(FirstInst)) {
10069 isVolatile = LI->isVolatile();
10070 // We can't sink the load if the loaded value could be modified between the
10071 // load and the PHI.
10072 if (LI->getParent() != PN.getIncomingBlock(0) ||
10073 !isSafeToSinkLoad(LI))
10076 // If the PHI is of volatile loads and the load block has multiple
10077 // successors, sinking it would remove a load of the volatile value from
10078 // the path through the other successor.
10080 LI->getParent()->getTerminator()->getNumSuccessors() != 1)
10083 } else if (isa<GetElementPtrInst>(FirstInst)) {
10084 if (FirstInst->getNumOperands() == 2)
10085 return FoldPHIArgBinOpIntoPHI(PN);
10086 // Can't handle general GEPs yet.
10089 return 0; // Cannot fold this operation.
10092 // Check to see if all arguments are the same operation.
10093 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
10094 if (!isa<Instruction>(PN.getIncomingValue(i))) return 0;
10095 Instruction *I = cast<Instruction>(PN.getIncomingValue(i));
10096 if (!I->hasOneUse() || !I->isSameOperationAs(FirstInst))
10099 if (I->getOperand(0)->getType() != CastSrcTy)
10100 return 0; // Cast operation must match.
10101 } else if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
10102 // We can't sink the load if the loaded value could be modified between
10103 // the load and the PHI.
10104 if (LI->isVolatile() != isVolatile ||
10105 LI->getParent() != PN.getIncomingBlock(i) ||
10106 !isSafeToSinkLoad(LI))
10109 // If the PHI is of volatile loads and the load block has multiple
10110 // successors, sinking it would remove a load of the volatile value from
10111 // the path through the other successor.
10113 LI->getParent()->getTerminator()->getNumSuccessors() != 1)
10117 } else if (I->getOperand(1) != ConstantOp) {
10122 // Okay, they are all the same operation. Create a new PHI node of the
10123 // correct type, and PHI together all of the LHS's of the instructions.
10124 PHINode *NewPN = PHINode::Create(FirstInst->getOperand(0)->getType(),
10125 PN.getName()+".in");
10126 NewPN->reserveOperandSpace(PN.getNumOperands()/2);
10128 Value *InVal = FirstInst->getOperand(0);
10129 NewPN->addIncoming(InVal, PN.getIncomingBlock(0));
10131 // Add all operands to the new PHI.
10132 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
10133 Value *NewInVal = cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
10134 if (NewInVal != InVal)
10136 NewPN->addIncoming(NewInVal, PN.getIncomingBlock(i));
10141 // The new PHI unions all of the same values together. This is really
10142 // common, so we handle it intelligently here for compile-time speed.
10146 InsertNewInstBefore(NewPN, PN);
10150 // Insert and return the new operation.
10151 if (CastInst* FirstCI = dyn_cast<CastInst>(FirstInst))
10152 return CastInst::Create(FirstCI->getOpcode(), PhiVal, PN.getType());
10153 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
10154 return BinaryOperator::Create(BinOp->getOpcode(), PhiVal, ConstantOp);
10155 if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst))
10156 return CmpInst::Create(CIOp->getOpcode(), CIOp->getPredicate(),
10157 PhiVal, ConstantOp);
10158 assert(isa<LoadInst>(FirstInst) && "Unknown operation");
10160 // If this was a volatile load that we are merging, make sure to loop through
10161 // and mark all the input loads as non-volatile. If we don't do this, we will
10162 // insert a new volatile load and the old ones will not be deletable.
10164 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i)
10165 cast<LoadInst>(PN.getIncomingValue(i))->setVolatile(false);
10167 return new LoadInst(PhiVal, "", isVolatile);
10170 /// DeadPHICycle - Return true if this PHI node is only used by a PHI node cycle
10172 static bool DeadPHICycle(PHINode *PN,
10173 SmallPtrSet<PHINode*, 16> &PotentiallyDeadPHIs) {
10174 if (PN->use_empty()) return true;
10175 if (!PN->hasOneUse()) return false;
10177 // Remember this node, and if we find the cycle, return.
10178 if (!PotentiallyDeadPHIs.insert(PN))
10181 // Don't scan crazily complex things.
10182 if (PotentiallyDeadPHIs.size() == 16)
10185 if (PHINode *PU = dyn_cast<PHINode>(PN->use_back()))
10186 return DeadPHICycle(PU, PotentiallyDeadPHIs);
10191 /// PHIsEqualValue - Return true if this phi node is always equal to
10192 /// NonPhiInVal. This happens with mutually cyclic phi nodes like:
10193 /// z = some value; x = phi (y, z); y = phi (x, z)
10194 static bool PHIsEqualValue(PHINode *PN, Value *NonPhiInVal,
10195 SmallPtrSet<PHINode*, 16> &ValueEqualPHIs) {
10196 // See if we already saw this PHI node.
10197 if (!ValueEqualPHIs.insert(PN))
10200 // Don't scan crazily complex things.
10201 if (ValueEqualPHIs.size() == 16)
10204 // Scan the operands to see if they are either phi nodes or are equal to
10206 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
10207 Value *Op = PN->getIncomingValue(i);
10208 if (PHINode *OpPN = dyn_cast<PHINode>(Op)) {
10209 if (!PHIsEqualValue(OpPN, NonPhiInVal, ValueEqualPHIs))
10211 } else if (Op != NonPhiInVal)
10219 // PHINode simplification
10221 Instruction *InstCombiner::visitPHINode(PHINode &PN) {
10222 // If LCSSA is around, don't mess with Phi nodes
10223 if (MustPreserveLCSSA) return 0;
10225 if (Value *V = PN.hasConstantValue())
10226 return ReplaceInstUsesWith(PN, V);
10228 // If all PHI operands are the same operation, pull them through the PHI,
10229 // reducing code size.
10230 if (isa<Instruction>(PN.getIncomingValue(0)) &&
10231 PN.getIncomingValue(0)->hasOneUse())
10232 if (Instruction *Result = FoldPHIArgOpIntoPHI(PN))
10235 // If this is a trivial cycle in the PHI node graph, remove it. Basically, if
10236 // this PHI only has a single use (a PHI), and if that PHI only has one use (a
10237 // PHI)... break the cycle.
10238 if (PN.hasOneUse()) {
10239 Instruction *PHIUser = cast<Instruction>(PN.use_back());
10240 if (PHINode *PU = dyn_cast<PHINode>(PHIUser)) {
10241 SmallPtrSet<PHINode*, 16> PotentiallyDeadPHIs;
10242 PotentiallyDeadPHIs.insert(&PN);
10243 if (DeadPHICycle(PU, PotentiallyDeadPHIs))
10244 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
10247 // If this phi has a single use, and if that use just computes a value for
10248 // the next iteration of a loop, delete the phi. This occurs with unused
10249 // induction variables, e.g. "for (int j = 0; ; ++j);". Detecting this
10250 // common case here is good because the only other things that catch this
10251 // are induction variable analysis (sometimes) and ADCE, which is only run
10253 if (PHIUser->hasOneUse() &&
10254 (isa<BinaryOperator>(PHIUser) || isa<GetElementPtrInst>(PHIUser)) &&
10255 PHIUser->use_back() == &PN) {
10256 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
10260 // We sometimes end up with phi cycles that non-obviously end up being the
10261 // same value, for example:
10262 // z = some value; x = phi (y, z); y = phi (x, z)
10263 // where the phi nodes don't necessarily need to be in the same block. Do a
10264 // quick check to see if the PHI node only contains a single non-phi value, if
10265 // so, scan to see if the phi cycle is actually equal to that value.
10267 unsigned InValNo = 0, NumOperandVals = PN.getNumIncomingValues();
10268 // Scan for the first non-phi operand.
10269 while (InValNo != NumOperandVals &&
10270 isa<PHINode>(PN.getIncomingValue(InValNo)))
10273 if (InValNo != NumOperandVals) {
10274 Value *NonPhiInVal = PN.getOperand(InValNo);
10276 // Scan the rest of the operands to see if there are any conflicts, if so
10277 // there is no need to recursively scan other phis.
10278 for (++InValNo; InValNo != NumOperandVals; ++InValNo) {
10279 Value *OpVal = PN.getIncomingValue(InValNo);
10280 if (OpVal != NonPhiInVal && !isa<PHINode>(OpVal))
10284 // If we scanned over all operands, then we have one unique value plus
10285 // phi values. Scan PHI nodes to see if they all merge in each other or
10287 if (InValNo == NumOperandVals) {
10288 SmallPtrSet<PHINode*, 16> ValueEqualPHIs;
10289 if (PHIsEqualValue(&PN, NonPhiInVal, ValueEqualPHIs))
10290 return ReplaceInstUsesWith(PN, NonPhiInVal);
10297 static Value *InsertCastToIntPtrTy(Value *V, const Type *DTy,
10298 Instruction *InsertPoint,
10299 InstCombiner *IC) {
10300 unsigned PtrSize = DTy->getPrimitiveSizeInBits();
10301 unsigned VTySize = V->getType()->getPrimitiveSizeInBits();
10302 // We must cast correctly to the pointer type. Ensure that we
10303 // sign extend the integer value if it is smaller as this is
10304 // used for address computation.
10305 Instruction::CastOps opcode =
10306 (VTySize < PtrSize ? Instruction::SExt :
10307 (VTySize == PtrSize ? Instruction::BitCast : Instruction::Trunc));
10308 return IC->InsertCastBefore(opcode, V, DTy, *InsertPoint);
10312 Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) {
10313 Value *PtrOp = GEP.getOperand(0);
10314 // Is it 'getelementptr %P, i32 0' or 'getelementptr %P'
10315 // If so, eliminate the noop.
10316 if (GEP.getNumOperands() == 1)
10317 return ReplaceInstUsesWith(GEP, PtrOp);
10319 if (isa<UndefValue>(GEP.getOperand(0)))
10320 return ReplaceInstUsesWith(GEP, UndefValue::get(GEP.getType()));
10322 bool HasZeroPointerIndex = false;
10323 if (Constant *C = dyn_cast<Constant>(GEP.getOperand(1)))
10324 HasZeroPointerIndex = C->isNullValue();
10326 if (GEP.getNumOperands() == 2 && HasZeroPointerIndex)
10327 return ReplaceInstUsesWith(GEP, PtrOp);
10329 // Eliminate unneeded casts for indices.
10330 bool MadeChange = false;
10332 gep_type_iterator GTI = gep_type_begin(GEP);
10333 for (User::op_iterator i = GEP.op_begin() + 1, e = GEP.op_end();
10334 i != e; ++i, ++GTI) {
10335 if (isa<SequentialType>(*GTI)) {
10336 if (CastInst *CI = dyn_cast<CastInst>(*i)) {
10337 if (CI->getOpcode() == Instruction::ZExt ||
10338 CI->getOpcode() == Instruction::SExt) {
10339 const Type *SrcTy = CI->getOperand(0)->getType();
10340 // We can eliminate a cast from i32 to i64 iff the target
10341 // is a 32-bit pointer target.
10342 if (SrcTy->getPrimitiveSizeInBits() >= TD->getPointerSizeInBits()) {
10344 *i = CI->getOperand(0);
10348 // If we are using a wider index than needed for this platform, shrink it
10349 // to what we need. If narrower, sign-extend it to what we need.
10350 // If the incoming value needs a cast instruction,
10351 // insert it. This explicit cast can make subsequent optimizations more
10354 if (TD->getTypeSizeInBits(Op->getType()) > TD->getPointerSizeInBits()) {
10355 if (Constant *C = dyn_cast<Constant>(Op)) {
10356 *i = ConstantExpr::getTrunc(C, TD->getIntPtrType());
10359 Op = InsertCastBefore(Instruction::Trunc, Op, TD->getIntPtrType(),
10364 } else if (TD->getTypeSizeInBits(Op->getType()) < TD->getPointerSizeInBits()) {
10365 if (Constant *C = dyn_cast<Constant>(Op)) {
10366 *i = ConstantExpr::getSExt(C, TD->getIntPtrType());
10369 Op = InsertCastBefore(Instruction::SExt, Op, TD->getIntPtrType(),
10377 if (MadeChange) return &GEP;
10379 // If this GEP instruction doesn't move the pointer, and if the input operand
10380 // is a bitcast of another pointer, just replace the GEP with a bitcast of the
10381 // real input to the dest type.
10382 if (GEP.hasAllZeroIndices()) {
10383 if (BitCastInst *BCI = dyn_cast<BitCastInst>(GEP.getOperand(0))) {
10384 // If the bitcast is of an allocation, and the allocation will be
10385 // converted to match the type of the cast, don't touch this.
10386 if (isa<AllocationInst>(BCI->getOperand(0))) {
10387 // See if the bitcast simplifies, if so, don't nuke this GEP yet.
10388 if (Instruction *I = visitBitCast(*BCI)) {
10391 BCI->getParent()->getInstList().insert(BCI, I);
10392 ReplaceInstUsesWith(*BCI, I);
10397 return new BitCastInst(BCI->getOperand(0), GEP.getType());
10401 // Combine Indices - If the source pointer to this getelementptr instruction
10402 // is a getelementptr instruction, combine the indices of the two
10403 // getelementptr instructions into a single instruction.
10405 SmallVector<Value*, 8> SrcGEPOperands;
10406 if (User *Src = dyn_castGetElementPtr(PtrOp))
10407 SrcGEPOperands.append(Src->op_begin(), Src->op_end());
10409 if (!SrcGEPOperands.empty()) {
10410 // Note that if our source is a gep chain itself that we wait for that
10411 // chain to be resolved before we perform this transformation. This
10412 // avoids us creating a TON of code in some cases.
10414 if (isa<GetElementPtrInst>(SrcGEPOperands[0]) &&
10415 cast<Instruction>(SrcGEPOperands[0])->getNumOperands() == 2)
10416 return 0; // Wait until our source is folded to completion.
10418 SmallVector<Value*, 8> Indices;
10420 // Find out whether the last index in the source GEP is a sequential idx.
10421 bool EndsWithSequential = false;
10422 for (gep_type_iterator I = gep_type_begin(*cast<User>(PtrOp)),
10423 E = gep_type_end(*cast<User>(PtrOp)); I != E; ++I)
10424 EndsWithSequential = !isa<StructType>(*I);
10426 // Can we combine the two pointer arithmetics offsets?
10427 if (EndsWithSequential) {
10428 // Replace: gep (gep %P, long B), long A, ...
10429 // With: T = long A+B; gep %P, T, ...
10431 Value *Sum, *SO1 = SrcGEPOperands.back(), *GO1 = GEP.getOperand(1);
10432 if (SO1 == Constant::getNullValue(SO1->getType())) {
10434 } else if (GO1 == Constant::getNullValue(GO1->getType())) {
10437 // If they aren't the same type, convert both to an integer of the
10438 // target's pointer size.
10439 if (SO1->getType() != GO1->getType()) {
10440 if (Constant *SO1C = dyn_cast<Constant>(SO1)) {
10441 SO1 = ConstantExpr::getIntegerCast(SO1C, GO1->getType(), true);
10442 } else if (Constant *GO1C = dyn_cast<Constant>(GO1)) {
10443 GO1 = ConstantExpr::getIntegerCast(GO1C, SO1->getType(), true);
10445 unsigned PS = TD->getPointerSizeInBits();
10446 if (TD->getTypeSizeInBits(SO1->getType()) == PS) {
10447 // Convert GO1 to SO1's type.
10448 GO1 = InsertCastToIntPtrTy(GO1, SO1->getType(), &GEP, this);
10450 } else if (TD->getTypeSizeInBits(GO1->getType()) == PS) {
10451 // Convert SO1 to GO1's type.
10452 SO1 = InsertCastToIntPtrTy(SO1, GO1->getType(), &GEP, this);
10454 const Type *PT = TD->getIntPtrType();
10455 SO1 = InsertCastToIntPtrTy(SO1, PT, &GEP, this);
10456 GO1 = InsertCastToIntPtrTy(GO1, PT, &GEP, this);
10460 if (isa<Constant>(SO1) && isa<Constant>(GO1))
10461 Sum = ConstantExpr::getAdd(cast<Constant>(SO1), cast<Constant>(GO1));
10463 Sum = BinaryOperator::CreateAdd(SO1, GO1, PtrOp->getName()+".sum");
10464 InsertNewInstBefore(cast<Instruction>(Sum), GEP);
10468 // Recycle the GEP we already have if possible.
10469 if (SrcGEPOperands.size() == 2) {
10470 GEP.setOperand(0, SrcGEPOperands[0]);
10471 GEP.setOperand(1, Sum);
10474 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
10475 SrcGEPOperands.end()-1);
10476 Indices.push_back(Sum);
10477 Indices.insert(Indices.end(), GEP.op_begin()+2, GEP.op_end());
10479 } else if (isa<Constant>(*GEP.idx_begin()) &&
10480 cast<Constant>(*GEP.idx_begin())->isNullValue() &&
10481 SrcGEPOperands.size() != 1) {
10482 // Otherwise we can do the fold if the first index of the GEP is a zero
10483 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
10484 SrcGEPOperands.end());
10485 Indices.insert(Indices.end(), GEP.idx_begin()+1, GEP.idx_end());
10488 if (!Indices.empty())
10489 return GetElementPtrInst::Create(SrcGEPOperands[0], Indices.begin(),
10490 Indices.end(), GEP.getName());
10492 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(PtrOp)) {
10493 // GEP of global variable. If all of the indices for this GEP are
10494 // constants, we can promote this to a constexpr instead of an instruction.
10496 // Scan for nonconstants...
10497 SmallVector<Constant*, 8> Indices;
10498 User::op_iterator I = GEP.idx_begin(), E = GEP.idx_end();
10499 for (; I != E && isa<Constant>(*I); ++I)
10500 Indices.push_back(cast<Constant>(*I));
10502 if (I == E) { // If they are all constants...
10503 Constant *CE = ConstantExpr::getGetElementPtr(GV,
10504 &Indices[0],Indices.size());
10506 // Replace all uses of the GEP with the new constexpr...
10507 return ReplaceInstUsesWith(GEP, CE);
10509 } else if (Value *X = getBitCastOperand(PtrOp)) { // Is the operand a cast?
10510 if (!isa<PointerType>(X->getType())) {
10511 // Not interesting. Source pointer must be a cast from pointer.
10512 } else if (HasZeroPointerIndex) {
10513 // transform: GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ...
10514 // into : GEP [10 x i8]* X, i32 0, ...
10516 // This occurs when the program declares an array extern like "int X[];"
10518 const PointerType *CPTy = cast<PointerType>(PtrOp->getType());
10519 const PointerType *XTy = cast<PointerType>(X->getType());
10520 if (const ArrayType *XATy =
10521 dyn_cast<ArrayType>(XTy->getElementType()))
10522 if (const ArrayType *CATy =
10523 dyn_cast<ArrayType>(CPTy->getElementType()))
10524 if (CATy->getElementType() == XATy->getElementType()) {
10525 // At this point, we know that the cast source type is a pointer
10526 // to an array of the same type as the destination pointer
10527 // array. Because the array type is never stepped over (there
10528 // is a leading zero) we can fold the cast into this GEP.
10529 GEP.setOperand(0, X);
10532 } else if (GEP.getNumOperands() == 2) {
10533 // Transform things like:
10534 // %t = getelementptr i32* bitcast ([2 x i32]* %str to i32*), i32 %V
10535 // into: %t1 = getelementptr [2 x i32]* %str, i32 0, i32 %V; bitcast
10536 const Type *SrcElTy = cast<PointerType>(X->getType())->getElementType();
10537 const Type *ResElTy=cast<PointerType>(PtrOp->getType())->getElementType();
10538 if (isa<ArrayType>(SrcElTy) &&
10539 TD->getABITypeSize(cast<ArrayType>(SrcElTy)->getElementType()) ==
10540 TD->getABITypeSize(ResElTy)) {
10542 Idx[0] = Constant::getNullValue(Type::Int32Ty);
10543 Idx[1] = GEP.getOperand(1);
10544 Value *V = InsertNewInstBefore(
10545 GetElementPtrInst::Create(X, Idx, Idx + 2, GEP.getName()), GEP);
10546 // V and GEP are both pointer types --> BitCast
10547 return new BitCastInst(V, GEP.getType());
10550 // Transform things like:
10551 // getelementptr i8* bitcast ([100 x double]* X to i8*), i32 %tmp
10552 // (where tmp = 8*tmp2) into:
10553 // getelementptr [100 x double]* %arr, i32 0, i32 %tmp2; bitcast
10555 if (isa<ArrayType>(SrcElTy) && ResElTy == Type::Int8Ty) {
10556 uint64_t ArrayEltSize =
10557 TD->getABITypeSize(cast<ArrayType>(SrcElTy)->getElementType());
10559 // Check to see if "tmp" is a scale by a multiple of ArrayEltSize. We
10560 // allow either a mul, shift, or constant here.
10562 ConstantInt *Scale = 0;
10563 if (ArrayEltSize == 1) {
10564 NewIdx = GEP.getOperand(1);
10565 Scale = ConstantInt::get(NewIdx->getType(), 1);
10566 } else if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP.getOperand(1))) {
10567 NewIdx = ConstantInt::get(CI->getType(), 1);
10569 } else if (Instruction *Inst =dyn_cast<Instruction>(GEP.getOperand(1))){
10570 if (Inst->getOpcode() == Instruction::Shl &&
10571 isa<ConstantInt>(Inst->getOperand(1))) {
10572 ConstantInt *ShAmt = cast<ConstantInt>(Inst->getOperand(1));
10573 uint32_t ShAmtVal = ShAmt->getLimitedValue(64);
10574 Scale = ConstantInt::get(Inst->getType(), 1ULL << ShAmtVal);
10575 NewIdx = Inst->getOperand(0);
10576 } else if (Inst->getOpcode() == Instruction::Mul &&
10577 isa<ConstantInt>(Inst->getOperand(1))) {
10578 Scale = cast<ConstantInt>(Inst->getOperand(1));
10579 NewIdx = Inst->getOperand(0);
10583 // If the index will be to exactly the right offset with the scale taken
10584 // out, perform the transformation. Note, we don't know whether Scale is
10585 // signed or not. We'll use unsigned version of division/modulo
10586 // operation after making sure Scale doesn't have the sign bit set.
10587 if (Scale && Scale->getSExtValue() >= 0LL &&
10588 Scale->getZExtValue() % ArrayEltSize == 0) {
10589 Scale = ConstantInt::get(Scale->getType(),
10590 Scale->getZExtValue() / ArrayEltSize);
10591 if (Scale->getZExtValue() != 1) {
10592 Constant *C = ConstantExpr::getIntegerCast(Scale, NewIdx->getType(),
10594 Instruction *Sc = BinaryOperator::CreateMul(NewIdx, C, "idxscale");
10595 NewIdx = InsertNewInstBefore(Sc, GEP);
10598 // Insert the new GEP instruction.
10600 Idx[0] = Constant::getNullValue(Type::Int32Ty);
10602 Instruction *NewGEP =
10603 GetElementPtrInst::Create(X, Idx, Idx + 2, GEP.getName());
10604 NewGEP = InsertNewInstBefore(NewGEP, GEP);
10605 // The NewGEP must be pointer typed, so must the old one -> BitCast
10606 return new BitCastInst(NewGEP, GEP.getType());
10615 Instruction *InstCombiner::visitAllocationInst(AllocationInst &AI) {
10616 // Convert: malloc Ty, C - where C is a constant != 1 into: malloc [C x Ty], 1
10617 if (AI.isArrayAllocation()) { // Check C != 1
10618 if (const ConstantInt *C = dyn_cast<ConstantInt>(AI.getArraySize())) {
10619 const Type *NewTy =
10620 ArrayType::get(AI.getAllocatedType(), C->getZExtValue());
10621 AllocationInst *New = 0;
10623 // Create and insert the replacement instruction...
10624 if (isa<MallocInst>(AI))
10625 New = new MallocInst(NewTy, 0, AI.getAlignment(), AI.getName());
10627 assert(isa<AllocaInst>(AI) && "Unknown type of allocation inst!");
10628 New = new AllocaInst(NewTy, 0, AI.getAlignment(), AI.getName());
10631 InsertNewInstBefore(New, AI);
10633 // Scan to the end of the allocation instructions, to skip over a block of
10634 // allocas if possible...
10636 BasicBlock::iterator It = New;
10637 while (isa<AllocationInst>(*It)) ++It;
10639 // Now that I is pointing to the first non-allocation-inst in the block,
10640 // insert our getelementptr instruction...
10642 Value *NullIdx = Constant::getNullValue(Type::Int32Ty);
10646 Value *V = GetElementPtrInst::Create(New, Idx, Idx + 2,
10647 New->getName()+".sub", It);
10649 // Now make everything use the getelementptr instead of the original
10651 return ReplaceInstUsesWith(AI, V);
10652 } else if (isa<UndefValue>(AI.getArraySize())) {
10653 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
10657 // If alloca'ing a zero byte object, replace the alloca with a null pointer.
10658 // Note that we only do this for alloca's, because malloc should allocate and
10659 // return a unique pointer, even for a zero byte allocation.
10660 if (isa<AllocaInst>(AI) && AI.getAllocatedType()->isSized() &&
10661 TD->getABITypeSize(AI.getAllocatedType()) == 0)
10662 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
10667 Instruction *InstCombiner::visitFreeInst(FreeInst &FI) {
10668 Value *Op = FI.getOperand(0);
10670 // free undef -> unreachable.
10671 if (isa<UndefValue>(Op)) {
10672 // Insert a new store to null because we cannot modify the CFG here.
10673 new StoreInst(ConstantInt::getTrue(),
10674 UndefValue::get(PointerType::getUnqual(Type::Int1Ty)), &FI);
10675 return EraseInstFromFunction(FI);
10678 // If we have 'free null' delete the instruction. This can happen in stl code
10679 // when lots of inlining happens.
10680 if (isa<ConstantPointerNull>(Op))
10681 return EraseInstFromFunction(FI);
10683 // Change free <ty>* (cast <ty2>* X to <ty>*) into free <ty2>* X
10684 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op)) {
10685 FI.setOperand(0, CI->getOperand(0));
10689 // Change free (gep X, 0,0,0,0) into free(X)
10690 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op)) {
10691 if (GEPI->hasAllZeroIndices()) {
10692 AddToWorkList(GEPI);
10693 FI.setOperand(0, GEPI->getOperand(0));
10698 // Change free(malloc) into nothing, if the malloc has a single use.
10699 if (MallocInst *MI = dyn_cast<MallocInst>(Op))
10700 if (MI->hasOneUse()) {
10701 EraseInstFromFunction(FI);
10702 return EraseInstFromFunction(*MI);
10709 /// InstCombineLoadCast - Fold 'load (cast P)' -> cast (load P)' when possible.
10710 static Instruction *InstCombineLoadCast(InstCombiner &IC, LoadInst &LI,
10711 const TargetData *TD) {
10712 User *CI = cast<User>(LI.getOperand(0));
10713 Value *CastOp = CI->getOperand(0);
10715 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(CI)) {
10716 // Instead of loading constant c string, use corresponding integer value
10717 // directly if string length is small enough.
10719 if (GetConstantStringInfo(CE->getOperand(0), Str) && !Str.empty()) {
10720 unsigned len = Str.length();
10721 const Type *Ty = cast<PointerType>(CE->getType())->getElementType();
10722 unsigned numBits = Ty->getPrimitiveSizeInBits();
10723 // Replace LI with immediate integer store.
10724 if ((numBits >> 3) == len + 1) {
10725 APInt StrVal(numBits, 0);
10726 APInt SingleChar(numBits, 0);
10727 if (TD->isLittleEndian()) {
10728 for (signed i = len-1; i >= 0; i--) {
10729 SingleChar = (uint64_t) Str[i];
10730 StrVal = (StrVal << 8) | SingleChar;
10733 for (unsigned i = 0; i < len; i++) {
10734 SingleChar = (uint64_t) Str[i];
10735 StrVal = (StrVal << 8) | SingleChar;
10737 // Append NULL at the end.
10739 StrVal = (StrVal << 8) | SingleChar;
10741 Value *NL = ConstantInt::get(StrVal);
10742 return IC.ReplaceInstUsesWith(LI, NL);
10747 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
10748 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
10749 const Type *SrcPTy = SrcTy->getElementType();
10751 if (DestPTy->isInteger() || isa<PointerType>(DestPTy) ||
10752 isa<VectorType>(DestPTy)) {
10753 // If the source is an array, the code below will not succeed. Check to
10754 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
10756 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
10757 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
10758 if (ASrcTy->getNumElements() != 0) {
10760 Idxs[0] = Idxs[1] = Constant::getNullValue(Type::Int32Ty);
10761 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs, 2);
10762 SrcTy = cast<PointerType>(CastOp->getType());
10763 SrcPTy = SrcTy->getElementType();
10766 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy) ||
10767 isa<VectorType>(SrcPTy)) &&
10768 // Do not allow turning this into a load of an integer, which is then
10769 // casted to a pointer, this pessimizes pointer analysis a lot.
10770 (isa<PointerType>(SrcPTy) == isa<PointerType>(LI.getType())) &&
10771 IC.getTargetData().getTypeSizeInBits(SrcPTy) ==
10772 IC.getTargetData().getTypeSizeInBits(DestPTy)) {
10774 // Okay, we are casting from one integer or pointer type to another of
10775 // the same size. Instead of casting the pointer before the load, cast
10776 // the result of the loaded value.
10777 Value *NewLoad = IC.InsertNewInstBefore(new LoadInst(CastOp,
10779 LI.isVolatile()),LI);
10780 // Now cast the result of the load.
10781 return new BitCastInst(NewLoad, LI.getType());
10788 /// isSafeToLoadUnconditionally - Return true if we know that executing a load
10789 /// from this value cannot trap. If it is not obviously safe to load from the
10790 /// specified pointer, we do a quick local scan of the basic block containing
10791 /// ScanFrom, to determine if the address is already accessed.
10792 static bool isSafeToLoadUnconditionally(Value *V, Instruction *ScanFrom) {
10793 // If it is an alloca it is always safe to load from.
10794 if (isa<AllocaInst>(V)) return true;
10796 // If it is a global variable it is mostly safe to load from.
10797 if (const GlobalValue *GV = dyn_cast<GlobalVariable>(V))
10798 // Don't try to evaluate aliases. External weak GV can be null.
10799 return !isa<GlobalAlias>(GV) && !GV->hasExternalWeakLinkage();
10801 // Otherwise, be a little bit agressive by scanning the local block where we
10802 // want to check to see if the pointer is already being loaded or stored
10803 // from/to. If so, the previous load or store would have already trapped,
10804 // so there is no harm doing an extra load (also, CSE will later eliminate
10805 // the load entirely).
10806 BasicBlock::iterator BBI = ScanFrom, E = ScanFrom->getParent()->begin();
10811 // If we see a free or a call (which might do a free) the pointer could be
10813 if (isa<FreeInst>(BBI) || isa<CallInst>(BBI))
10816 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
10817 if (LI->getOperand(0) == V) return true;
10818 } else if (StoreInst *SI = dyn_cast<StoreInst>(BBI)) {
10819 if (SI->getOperand(1) == V) return true;
10826 Instruction *InstCombiner::visitLoadInst(LoadInst &LI) {
10827 Value *Op = LI.getOperand(0);
10829 // Attempt to improve the alignment.
10830 unsigned KnownAlign = GetOrEnforceKnownAlignment(Op);
10832 (LI.getAlignment() == 0 ? TD->getABITypeAlignment(LI.getType()) :
10833 LI.getAlignment()))
10834 LI.setAlignment(KnownAlign);
10836 // load (cast X) --> cast (load X) iff safe
10837 if (isa<CastInst>(Op))
10838 if (Instruction *Res = InstCombineLoadCast(*this, LI, TD))
10841 // None of the following transforms are legal for volatile loads.
10842 if (LI.isVolatile()) return 0;
10844 // Do really simple store-to-load forwarding and load CSE, to catch cases
10845 // where there are several consequtive memory accesses to the same location,
10846 // separated by a few arithmetic operations.
10847 BasicBlock::iterator BBI = &LI;
10848 if (Value *AvailableVal = FindAvailableLoadedValue(Op, LI.getParent(), BBI,6))
10849 return ReplaceInstUsesWith(LI, AvailableVal);
10851 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op)) {
10852 const Value *GEPI0 = GEPI->getOperand(0);
10853 // TODO: Consider a target hook for valid address spaces for this xform.
10854 if (isa<ConstantPointerNull>(GEPI0) &&
10855 cast<PointerType>(GEPI0->getType())->getAddressSpace() == 0) {
10856 // Insert a new store to null instruction before the load to indicate
10857 // that this code is not reachable. We do this instead of inserting
10858 // an unreachable instruction directly because we cannot modify the
10860 new StoreInst(UndefValue::get(LI.getType()),
10861 Constant::getNullValue(Op->getType()), &LI);
10862 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
10866 if (Constant *C = dyn_cast<Constant>(Op)) {
10867 // load null/undef -> undef
10868 // TODO: Consider a target hook for valid address spaces for this xform.
10869 if (isa<UndefValue>(C) || (C->isNullValue() &&
10870 cast<PointerType>(Op->getType())->getAddressSpace() == 0)) {
10871 // Insert a new store to null instruction before the load to indicate that
10872 // this code is not reachable. We do this instead of inserting an
10873 // unreachable instruction directly because we cannot modify the CFG.
10874 new StoreInst(UndefValue::get(LI.getType()),
10875 Constant::getNullValue(Op->getType()), &LI);
10876 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
10879 // Instcombine load (constant global) into the value loaded.
10880 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Op))
10881 if (GV->isConstant() && !GV->isDeclaration())
10882 return ReplaceInstUsesWith(LI, GV->getInitializer());
10884 // Instcombine load (constantexpr_GEP global, 0, ...) into the value loaded.
10885 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Op)) {
10886 if (CE->getOpcode() == Instruction::GetElementPtr) {
10887 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(CE->getOperand(0)))
10888 if (GV->isConstant() && !GV->isDeclaration())
10890 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE))
10891 return ReplaceInstUsesWith(LI, V);
10892 if (CE->getOperand(0)->isNullValue()) {
10893 // Insert a new store to null instruction before the load to indicate
10894 // that this code is not reachable. We do this instead of inserting
10895 // an unreachable instruction directly because we cannot modify the
10897 new StoreInst(UndefValue::get(LI.getType()),
10898 Constant::getNullValue(Op->getType()), &LI);
10899 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
10902 } else if (CE->isCast()) {
10903 if (Instruction *Res = InstCombineLoadCast(*this, LI, TD))
10909 // If this load comes from anywhere in a constant global, and if the global
10910 // is all undef or zero, we know what it loads.
10911 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Op->getUnderlyingObject())){
10912 if (GV->isConstant() && GV->hasInitializer()) {
10913 if (GV->getInitializer()->isNullValue())
10914 return ReplaceInstUsesWith(LI, Constant::getNullValue(LI.getType()));
10915 else if (isa<UndefValue>(GV->getInitializer()))
10916 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
10920 if (Op->hasOneUse()) {
10921 // Change select and PHI nodes to select values instead of addresses: this
10922 // helps alias analysis out a lot, allows many others simplifications, and
10923 // exposes redundancy in the code.
10925 // Note that we cannot do the transformation unless we know that the
10926 // introduced loads cannot trap! Something like this is valid as long as
10927 // the condition is always false: load (select bool %C, int* null, int* %G),
10928 // but it would not be valid if we transformed it to load from null
10929 // unconditionally.
10931 if (SelectInst *SI = dyn_cast<SelectInst>(Op)) {
10932 // load (select (Cond, &V1, &V2)) --> select(Cond, load &V1, load &V2).
10933 if (isSafeToLoadUnconditionally(SI->getOperand(1), SI) &&
10934 isSafeToLoadUnconditionally(SI->getOperand(2), SI)) {
10935 Value *V1 = InsertNewInstBefore(new LoadInst(SI->getOperand(1),
10936 SI->getOperand(1)->getName()+".val"), LI);
10937 Value *V2 = InsertNewInstBefore(new LoadInst(SI->getOperand(2),
10938 SI->getOperand(2)->getName()+".val"), LI);
10939 return SelectInst::Create(SI->getCondition(), V1, V2);
10942 // load (select (cond, null, P)) -> load P
10943 if (Constant *C = dyn_cast<Constant>(SI->getOperand(1)))
10944 if (C->isNullValue()) {
10945 LI.setOperand(0, SI->getOperand(2));
10949 // load (select (cond, P, null)) -> load P
10950 if (Constant *C = dyn_cast<Constant>(SI->getOperand(2)))
10951 if (C->isNullValue()) {
10952 LI.setOperand(0, SI->getOperand(1));
10960 /// InstCombineStoreToCast - Fold store V, (cast P) -> store (cast V), P
10962 static Instruction *InstCombineStoreToCast(InstCombiner &IC, StoreInst &SI) {
10963 User *CI = cast<User>(SI.getOperand(1));
10964 Value *CastOp = CI->getOperand(0);
10966 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
10967 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
10968 const Type *SrcPTy = SrcTy->getElementType();
10970 if (DestPTy->isInteger() || isa<PointerType>(DestPTy)) {
10971 // If the source is an array, the code below will not succeed. Check to
10972 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
10974 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
10975 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
10976 if (ASrcTy->getNumElements() != 0) {
10978 Idxs[0] = Idxs[1] = Constant::getNullValue(Type::Int32Ty);
10979 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs, 2);
10980 SrcTy = cast<PointerType>(CastOp->getType());
10981 SrcPTy = SrcTy->getElementType();
10984 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy)) &&
10985 IC.getTargetData().getTypeSizeInBits(SrcPTy) ==
10986 IC.getTargetData().getTypeSizeInBits(DestPTy)) {
10988 // Okay, we are casting from one integer or pointer type to another of
10989 // the same size. Instead of casting the pointer before
10990 // the store, cast the value to be stored.
10992 Value *SIOp0 = SI.getOperand(0);
10993 Instruction::CastOps opcode = Instruction::BitCast;
10994 const Type* CastSrcTy = SIOp0->getType();
10995 const Type* CastDstTy = SrcPTy;
10996 if (isa<PointerType>(CastDstTy)) {
10997 if (CastSrcTy->isInteger())
10998 opcode = Instruction::IntToPtr;
10999 } else if (isa<IntegerType>(CastDstTy)) {
11000 if (isa<PointerType>(SIOp0->getType()))
11001 opcode = Instruction::PtrToInt;
11003 if (Constant *C = dyn_cast<Constant>(SIOp0))
11004 NewCast = ConstantExpr::getCast(opcode, C, CastDstTy);
11006 NewCast = IC.InsertNewInstBefore(
11007 CastInst::Create(opcode, SIOp0, CastDstTy, SIOp0->getName()+".c"),
11009 return new StoreInst(NewCast, CastOp);
11016 /// equivalentAddressValues - Test if A and B will obviously have the same
11017 /// value. This includes recognizing that %t0 and %t1 will have the same
11018 /// value in code like this:
11019 /// %t0 = getelementptr @a, 0, 3
11020 /// store i32 0, i32* %t0
11021 /// %t1 = getelementptr @a, 0, 3
11022 /// %t2 = load i32* %t1
11024 static bool equivalentAddressValues(Value *A, Value *B) {
11025 // Test if the values are trivially equivalent.
11026 if (A == B) return true;
11028 // Test if the values come form identical arithmetic instructions.
11029 if (isa<BinaryOperator>(A) ||
11030 isa<CastInst>(A) ||
11032 isa<GetElementPtrInst>(A))
11033 if (Instruction *BI = dyn_cast<Instruction>(B))
11034 if (cast<Instruction>(A)->isIdenticalTo(BI))
11037 // Otherwise they may not be equivalent.
11041 Instruction *InstCombiner::visitStoreInst(StoreInst &SI) {
11042 Value *Val = SI.getOperand(0);
11043 Value *Ptr = SI.getOperand(1);
11045 if (isa<UndefValue>(Ptr)) { // store X, undef -> noop (even if volatile)
11046 EraseInstFromFunction(SI);
11051 // If the RHS is an alloca with a single use, zapify the store, making the
11053 if (Ptr->hasOneUse() && !SI.isVolatile()) {
11054 if (isa<AllocaInst>(Ptr)) {
11055 EraseInstFromFunction(SI);
11060 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr))
11061 if (isa<AllocaInst>(GEP->getOperand(0)) &&
11062 GEP->getOperand(0)->hasOneUse()) {
11063 EraseInstFromFunction(SI);
11069 // Attempt to improve the alignment.
11070 unsigned KnownAlign = GetOrEnforceKnownAlignment(Ptr);
11072 (SI.getAlignment() == 0 ? TD->getABITypeAlignment(Val->getType()) :
11073 SI.getAlignment()))
11074 SI.setAlignment(KnownAlign);
11076 // Do really simple DSE, to catch cases where there are several consequtive
11077 // stores to the same location, separated by a few arithmetic operations. This
11078 // situation often occurs with bitfield accesses.
11079 BasicBlock::iterator BBI = &SI;
11080 for (unsigned ScanInsts = 6; BBI != SI.getParent()->begin() && ScanInsts;
11084 if (StoreInst *PrevSI = dyn_cast<StoreInst>(BBI)) {
11085 // Prev store isn't volatile, and stores to the same location?
11086 if (!PrevSI->isVolatile() &&equivalentAddressValues(PrevSI->getOperand(1),
11087 SI.getOperand(1))) {
11090 EraseInstFromFunction(*PrevSI);
11096 // If this is a load, we have to stop. However, if the loaded value is from
11097 // the pointer we're loading and is producing the pointer we're storing,
11098 // then *this* store is dead (X = load P; store X -> P).
11099 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
11100 if (LI == Val && equivalentAddressValues(LI->getOperand(0), Ptr) &&
11101 !SI.isVolatile()) {
11102 EraseInstFromFunction(SI);
11106 // Otherwise, this is a load from some other location. Stores before it
11107 // may not be dead.
11111 // Don't skip over loads or things that can modify memory.
11112 if (BBI->mayWriteToMemory() || BBI->mayReadFromMemory())
11117 if (SI.isVolatile()) return 0; // Don't hack volatile stores.
11119 // store X, null -> turns into 'unreachable' in SimplifyCFG
11120 if (isa<ConstantPointerNull>(Ptr)) {
11121 if (!isa<UndefValue>(Val)) {
11122 SI.setOperand(0, UndefValue::get(Val->getType()));
11123 if (Instruction *U = dyn_cast<Instruction>(Val))
11124 AddToWorkList(U); // Dropped a use.
11127 return 0; // Do not modify these!
11130 // store undef, Ptr -> noop
11131 if (isa<UndefValue>(Val)) {
11132 EraseInstFromFunction(SI);
11137 // If the pointer destination is a cast, see if we can fold the cast into the
11139 if (isa<CastInst>(Ptr))
11140 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
11142 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
11144 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
11148 // If this store is the last instruction in the basic block, and if the block
11149 // ends with an unconditional branch, try to move it to the successor block.
11151 if (BranchInst *BI = dyn_cast<BranchInst>(BBI))
11152 if (BI->isUnconditional())
11153 if (SimplifyStoreAtEndOfBlock(SI))
11154 return 0; // xform done!
11159 /// SimplifyStoreAtEndOfBlock - Turn things like:
11160 /// if () { *P = v1; } else { *P = v2 }
11161 /// into a phi node with a store in the successor.
11163 /// Simplify things like:
11164 /// *P = v1; if () { *P = v2; }
11165 /// into a phi node with a store in the successor.
11167 bool InstCombiner::SimplifyStoreAtEndOfBlock(StoreInst &SI) {
11168 BasicBlock *StoreBB = SI.getParent();
11170 // Check to see if the successor block has exactly two incoming edges. If
11171 // so, see if the other predecessor contains a store to the same location.
11172 // if so, insert a PHI node (if needed) and move the stores down.
11173 BasicBlock *DestBB = StoreBB->getTerminator()->getSuccessor(0);
11175 // Determine whether Dest has exactly two predecessors and, if so, compute
11176 // the other predecessor.
11177 pred_iterator PI = pred_begin(DestBB);
11178 BasicBlock *OtherBB = 0;
11179 if (*PI != StoreBB)
11182 if (PI == pred_end(DestBB))
11185 if (*PI != StoreBB) {
11190 if (++PI != pred_end(DestBB))
11193 // Bail out if all the relevant blocks aren't distinct (this can happen,
11194 // for example, if SI is in an infinite loop)
11195 if (StoreBB == DestBB || OtherBB == DestBB)
11198 // Verify that the other block ends in a branch and is not otherwise empty.
11199 BasicBlock::iterator BBI = OtherBB->getTerminator();
11200 BranchInst *OtherBr = dyn_cast<BranchInst>(BBI);
11201 if (!OtherBr || BBI == OtherBB->begin())
11204 // If the other block ends in an unconditional branch, check for the 'if then
11205 // else' case. there is an instruction before the branch.
11206 StoreInst *OtherStore = 0;
11207 if (OtherBr->isUnconditional()) {
11208 // If this isn't a store, or isn't a store to the same location, bail out.
11210 OtherStore = dyn_cast<StoreInst>(BBI);
11211 if (!OtherStore || OtherStore->getOperand(1) != SI.getOperand(1))
11214 // Otherwise, the other block ended with a conditional branch. If one of the
11215 // destinations is StoreBB, then we have the if/then case.
11216 if (OtherBr->getSuccessor(0) != StoreBB &&
11217 OtherBr->getSuccessor(1) != StoreBB)
11220 // Okay, we know that OtherBr now goes to Dest and StoreBB, so this is an
11221 // if/then triangle. See if there is a store to the same ptr as SI that
11222 // lives in OtherBB.
11224 // Check to see if we find the matching store.
11225 if ((OtherStore = dyn_cast<StoreInst>(BBI))) {
11226 if (OtherStore->getOperand(1) != SI.getOperand(1))
11230 // If we find something that may be using or overwriting the stored
11231 // value, or if we run out of instructions, we can't do the xform.
11232 if (BBI->mayReadFromMemory() || BBI->mayWriteToMemory() ||
11233 BBI == OtherBB->begin())
11237 // In order to eliminate the store in OtherBr, we have to
11238 // make sure nothing reads or overwrites the stored value in
11240 for (BasicBlock::iterator I = StoreBB->begin(); &*I != &SI; ++I) {
11241 // FIXME: This should really be AA driven.
11242 if (I->mayReadFromMemory() || I->mayWriteToMemory())
11247 // Insert a PHI node now if we need it.
11248 Value *MergedVal = OtherStore->getOperand(0);
11249 if (MergedVal != SI.getOperand(0)) {
11250 PHINode *PN = PHINode::Create(MergedVal->getType(), "storemerge");
11251 PN->reserveOperandSpace(2);
11252 PN->addIncoming(SI.getOperand(0), SI.getParent());
11253 PN->addIncoming(OtherStore->getOperand(0), OtherBB);
11254 MergedVal = InsertNewInstBefore(PN, DestBB->front());
11257 // Advance to a place where it is safe to insert the new store and
11259 BBI = DestBB->getFirstNonPHI();
11260 InsertNewInstBefore(new StoreInst(MergedVal, SI.getOperand(1),
11261 OtherStore->isVolatile()), *BBI);
11263 // Nuke the old stores.
11264 EraseInstFromFunction(SI);
11265 EraseInstFromFunction(*OtherStore);
11271 Instruction *InstCombiner::visitBranchInst(BranchInst &BI) {
11272 // Change br (not X), label True, label False to: br X, label False, True
11274 BasicBlock *TrueDest;
11275 BasicBlock *FalseDest;
11276 if (match(&BI, m_Br(m_Not(m_Value(X)), TrueDest, FalseDest)) &&
11277 !isa<Constant>(X)) {
11278 // Swap Destinations and condition...
11279 BI.setCondition(X);
11280 BI.setSuccessor(0, FalseDest);
11281 BI.setSuccessor(1, TrueDest);
11285 // Cannonicalize fcmp_one -> fcmp_oeq
11286 FCmpInst::Predicate FPred; Value *Y;
11287 if (match(&BI, m_Br(m_FCmp(FPred, m_Value(X), m_Value(Y)),
11288 TrueDest, FalseDest)))
11289 if ((FPred == FCmpInst::FCMP_ONE || FPred == FCmpInst::FCMP_OLE ||
11290 FPred == FCmpInst::FCMP_OGE) && BI.getCondition()->hasOneUse()) {
11291 FCmpInst *I = cast<FCmpInst>(BI.getCondition());
11292 FCmpInst::Predicate NewPred = FCmpInst::getInversePredicate(FPred);
11293 Instruction *NewSCC = new FCmpInst(NewPred, X, Y, "", I);
11294 NewSCC->takeName(I);
11295 // Swap Destinations and condition...
11296 BI.setCondition(NewSCC);
11297 BI.setSuccessor(0, FalseDest);
11298 BI.setSuccessor(1, TrueDest);
11299 RemoveFromWorkList(I);
11300 I->eraseFromParent();
11301 AddToWorkList(NewSCC);
11305 // Cannonicalize icmp_ne -> icmp_eq
11306 ICmpInst::Predicate IPred;
11307 if (match(&BI, m_Br(m_ICmp(IPred, m_Value(X), m_Value(Y)),
11308 TrueDest, FalseDest)))
11309 if ((IPred == ICmpInst::ICMP_NE || IPred == ICmpInst::ICMP_ULE ||
11310 IPred == ICmpInst::ICMP_SLE || IPred == ICmpInst::ICMP_UGE ||
11311 IPred == ICmpInst::ICMP_SGE) && BI.getCondition()->hasOneUse()) {
11312 ICmpInst *I = cast<ICmpInst>(BI.getCondition());
11313 ICmpInst::Predicate NewPred = ICmpInst::getInversePredicate(IPred);
11314 Instruction *NewSCC = new ICmpInst(NewPred, X, Y, "", I);
11315 NewSCC->takeName(I);
11316 // Swap Destinations and condition...
11317 BI.setCondition(NewSCC);
11318 BI.setSuccessor(0, FalseDest);
11319 BI.setSuccessor(1, TrueDest);
11320 RemoveFromWorkList(I);
11321 I->eraseFromParent();;
11322 AddToWorkList(NewSCC);
11329 Instruction *InstCombiner::visitSwitchInst(SwitchInst &SI) {
11330 Value *Cond = SI.getCondition();
11331 if (Instruction *I = dyn_cast<Instruction>(Cond)) {
11332 if (I->getOpcode() == Instruction::Add)
11333 if (ConstantInt *AddRHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
11334 // change 'switch (X+4) case 1:' into 'switch (X) case -3'
11335 for (unsigned i = 2, e = SI.getNumOperands(); i != e; i += 2)
11336 SI.setOperand(i,ConstantExpr::getSub(cast<Constant>(SI.getOperand(i)),
11338 SI.setOperand(0, I->getOperand(0));
11346 Instruction *InstCombiner::visitExtractValueInst(ExtractValueInst &EV) {
11347 Value *Agg = EV.getAggregateOperand();
11349 if (!EV.hasIndices())
11350 return ReplaceInstUsesWith(EV, Agg);
11352 if (Constant *C = dyn_cast<Constant>(Agg)) {
11353 if (isa<UndefValue>(C))
11354 return ReplaceInstUsesWith(EV, UndefValue::get(EV.getType()));
11356 if (isa<ConstantAggregateZero>(C))
11357 return ReplaceInstUsesWith(EV, Constant::getNullValue(EV.getType()));
11359 if (isa<ConstantArray>(C) || isa<ConstantStruct>(C)) {
11360 // Extract the element indexed by the first index out of the constant
11361 Value *V = C->getOperand(*EV.idx_begin());
11362 if (EV.getNumIndices() > 1)
11363 // Extract the remaining indices out of the constant indexed by the
11365 return ExtractValueInst::Create(V, EV.idx_begin() + 1, EV.idx_end());
11367 return ReplaceInstUsesWith(EV, V);
11369 return 0; // Can't handle other constants
11371 if (InsertValueInst *IV = dyn_cast<InsertValueInst>(Agg)) {
11372 // We're extracting from an insertvalue instruction, compare the indices
11373 const unsigned *exti, *exte, *insi, *inse;
11374 for (exti = EV.idx_begin(), insi = IV->idx_begin(),
11375 exte = EV.idx_end(), inse = IV->idx_end();
11376 exti != exte && insi != inse;
11378 if (*insi != *exti)
11379 // The insert and extract both reference distinctly different elements.
11380 // This means the extract is not influenced by the insert, and we can
11381 // replace the aggregate operand of the extract with the aggregate
11382 // operand of the insert. i.e., replace
11383 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1
11384 // %E = extractvalue { i32, { i32 } } %I, 0
11386 // %E = extractvalue { i32, { i32 } } %A, 0
11387 return ExtractValueInst::Create(IV->getAggregateOperand(),
11388 EV.idx_begin(), EV.idx_end());
11390 if (exti == exte && insi == inse)
11391 // Both iterators are at the end: Index lists are identical. Replace
11392 // %B = insertvalue { i32, { i32 } } %A, i32 42, 1, 0
11393 // %C = extractvalue { i32, { i32 } } %B, 1, 0
11395 return ReplaceInstUsesWith(EV, IV->getInsertedValueOperand());
11396 if (exti == exte) {
11397 // The extract list is a prefix of the insert list. i.e. replace
11398 // %I = insertvalue { i32, { i32 } } %A, i32 42, 1, 0
11399 // %E = extractvalue { i32, { i32 } } %I, 1
11401 // %X = extractvalue { i32, { i32 } } %A, 1
11402 // %E = insertvalue { i32 } %X, i32 42, 0
11403 // by switching the order of the insert and extract (though the
11404 // insertvalue should be left in, since it may have other uses).
11405 Value *NewEV = InsertNewInstBefore(
11406 ExtractValueInst::Create(IV->getAggregateOperand(),
11407 EV.idx_begin(), EV.idx_end()),
11409 return InsertValueInst::Create(NewEV, IV->getInsertedValueOperand(),
11413 // The insert list is a prefix of the extract list
11414 // We can simply remove the common indices from the extract and make it
11415 // operate on the inserted value instead of the insertvalue result.
11417 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1
11418 // %E = extractvalue { i32, { i32 } } %I, 1, 0
11420 // %E extractvalue { i32 } { i32 42 }, 0
11421 return ExtractValueInst::Create(IV->getInsertedValueOperand(),
11424 // Can't simplify extracts from other values. Note that nested extracts are
11425 // already simplified implicitely by the above (extract ( extract (insert) )
11426 // will be translated into extract ( insert ( extract ) ) first and then just
11427 // the value inserted, if appropriate).
11431 /// CheapToScalarize - Return true if the value is cheaper to scalarize than it
11432 /// is to leave as a vector operation.
11433 static bool CheapToScalarize(Value *V, bool isConstant) {
11434 if (isa<ConstantAggregateZero>(V))
11436 if (ConstantVector *C = dyn_cast<ConstantVector>(V)) {
11437 if (isConstant) return true;
11438 // If all elts are the same, we can extract.
11439 Constant *Op0 = C->getOperand(0);
11440 for (unsigned i = 1; i < C->getNumOperands(); ++i)
11441 if (C->getOperand(i) != Op0)
11445 Instruction *I = dyn_cast<Instruction>(V);
11446 if (!I) return false;
11448 // Insert element gets simplified to the inserted element or is deleted if
11449 // this is constant idx extract element and its a constant idx insertelt.
11450 if (I->getOpcode() == Instruction::InsertElement && isConstant &&
11451 isa<ConstantInt>(I->getOperand(2)))
11453 if (I->getOpcode() == Instruction::Load && I->hasOneUse())
11455 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I))
11456 if (BO->hasOneUse() &&
11457 (CheapToScalarize(BO->getOperand(0), isConstant) ||
11458 CheapToScalarize(BO->getOperand(1), isConstant)))
11460 if (CmpInst *CI = dyn_cast<CmpInst>(I))
11461 if (CI->hasOneUse() &&
11462 (CheapToScalarize(CI->getOperand(0), isConstant) ||
11463 CheapToScalarize(CI->getOperand(1), isConstant)))
11469 /// Read and decode a shufflevector mask.
11471 /// It turns undef elements into values that are larger than the number of
11472 /// elements in the input.
11473 static std::vector<unsigned> getShuffleMask(const ShuffleVectorInst *SVI) {
11474 unsigned NElts = SVI->getType()->getNumElements();
11475 if (isa<ConstantAggregateZero>(SVI->getOperand(2)))
11476 return std::vector<unsigned>(NElts, 0);
11477 if (isa<UndefValue>(SVI->getOperand(2)))
11478 return std::vector<unsigned>(NElts, 2*NElts);
11480 std::vector<unsigned> Result;
11481 const ConstantVector *CP = cast<ConstantVector>(SVI->getOperand(2));
11482 for (User::const_op_iterator i = CP->op_begin(), e = CP->op_end(); i!=e; ++i)
11483 if (isa<UndefValue>(*i))
11484 Result.push_back(NElts*2); // undef -> 8
11486 Result.push_back(cast<ConstantInt>(*i)->getZExtValue());
11490 /// FindScalarElement - Given a vector and an element number, see if the scalar
11491 /// value is already around as a register, for example if it were inserted then
11492 /// extracted from the vector.
11493 static Value *FindScalarElement(Value *V, unsigned EltNo) {
11494 assert(isa<VectorType>(V->getType()) && "Not looking at a vector?");
11495 const VectorType *PTy = cast<VectorType>(V->getType());
11496 unsigned Width = PTy->getNumElements();
11497 if (EltNo >= Width) // Out of range access.
11498 return UndefValue::get(PTy->getElementType());
11500 if (isa<UndefValue>(V))
11501 return UndefValue::get(PTy->getElementType());
11502 else if (isa<ConstantAggregateZero>(V))
11503 return Constant::getNullValue(PTy->getElementType());
11504 else if (ConstantVector *CP = dyn_cast<ConstantVector>(V))
11505 return CP->getOperand(EltNo);
11506 else if (InsertElementInst *III = dyn_cast<InsertElementInst>(V)) {
11507 // If this is an insert to a variable element, we don't know what it is.
11508 if (!isa<ConstantInt>(III->getOperand(2)))
11510 unsigned IIElt = cast<ConstantInt>(III->getOperand(2))->getZExtValue();
11512 // If this is an insert to the element we are looking for, return the
11514 if (EltNo == IIElt)
11515 return III->getOperand(1);
11517 // Otherwise, the insertelement doesn't modify the value, recurse on its
11519 return FindScalarElement(III->getOperand(0), EltNo);
11520 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(V)) {
11521 unsigned LHSWidth =
11522 cast<VectorType>(SVI->getOperand(0)->getType())->getNumElements();
11523 unsigned InEl = getShuffleMask(SVI)[EltNo];
11524 if (InEl < LHSWidth)
11525 return FindScalarElement(SVI->getOperand(0), InEl);
11526 else if (InEl < LHSWidth*2)
11527 return FindScalarElement(SVI->getOperand(1), InEl - LHSWidth);
11529 return UndefValue::get(PTy->getElementType());
11532 // Otherwise, we don't know.
11536 Instruction *InstCombiner::visitExtractElementInst(ExtractElementInst &EI) {
11537 // If vector val is undef, replace extract with scalar undef.
11538 if (isa<UndefValue>(EI.getOperand(0)))
11539 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
11541 // If vector val is constant 0, replace extract with scalar 0.
11542 if (isa<ConstantAggregateZero>(EI.getOperand(0)))
11543 return ReplaceInstUsesWith(EI, Constant::getNullValue(EI.getType()));
11545 if (ConstantVector *C = dyn_cast<ConstantVector>(EI.getOperand(0))) {
11546 // If vector val is constant with all elements the same, replace EI with
11547 // that element. When the elements are not identical, we cannot replace yet
11548 // (we do that below, but only when the index is constant).
11549 Constant *op0 = C->getOperand(0);
11550 for (unsigned i = 1; i < C->getNumOperands(); ++i)
11551 if (C->getOperand(i) != op0) {
11556 return ReplaceInstUsesWith(EI, op0);
11559 // If extracting a specified index from the vector, see if we can recursively
11560 // find a previously computed scalar that was inserted into the vector.
11561 if (ConstantInt *IdxC = dyn_cast<ConstantInt>(EI.getOperand(1))) {
11562 unsigned IndexVal = IdxC->getZExtValue();
11563 unsigned VectorWidth =
11564 cast<VectorType>(EI.getOperand(0)->getType())->getNumElements();
11566 // If this is extracting an invalid index, turn this into undef, to avoid
11567 // crashing the code below.
11568 if (IndexVal >= VectorWidth)
11569 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
11571 // This instruction only demands the single element from the input vector.
11572 // If the input vector has a single use, simplify it based on this use
11574 if (EI.getOperand(0)->hasOneUse() && VectorWidth != 1) {
11575 uint64_t UndefElts;
11576 if (Value *V = SimplifyDemandedVectorElts(EI.getOperand(0),
11579 EI.setOperand(0, V);
11584 if (Value *Elt = FindScalarElement(EI.getOperand(0), IndexVal))
11585 return ReplaceInstUsesWith(EI, Elt);
11587 // If the this extractelement is directly using a bitcast from a vector of
11588 // the same number of elements, see if we can find the source element from
11589 // it. In this case, we will end up needing to bitcast the scalars.
11590 if (BitCastInst *BCI = dyn_cast<BitCastInst>(EI.getOperand(0))) {
11591 if (const VectorType *VT =
11592 dyn_cast<VectorType>(BCI->getOperand(0)->getType()))
11593 if (VT->getNumElements() == VectorWidth)
11594 if (Value *Elt = FindScalarElement(BCI->getOperand(0), IndexVal))
11595 return new BitCastInst(Elt, EI.getType());
11599 if (Instruction *I = dyn_cast<Instruction>(EI.getOperand(0))) {
11600 if (I->hasOneUse()) {
11601 // Push extractelement into predecessor operation if legal and
11602 // profitable to do so
11603 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
11604 bool isConstantElt = isa<ConstantInt>(EI.getOperand(1));
11605 if (CheapToScalarize(BO, isConstantElt)) {
11606 ExtractElementInst *newEI0 =
11607 new ExtractElementInst(BO->getOperand(0), EI.getOperand(1),
11608 EI.getName()+".lhs");
11609 ExtractElementInst *newEI1 =
11610 new ExtractElementInst(BO->getOperand(1), EI.getOperand(1),
11611 EI.getName()+".rhs");
11612 InsertNewInstBefore(newEI0, EI);
11613 InsertNewInstBefore(newEI1, EI);
11614 return BinaryOperator::Create(BO->getOpcode(), newEI0, newEI1);
11616 } else if (isa<LoadInst>(I)) {
11618 cast<PointerType>(I->getOperand(0)->getType())->getAddressSpace();
11619 Value *Ptr = InsertBitCastBefore(I->getOperand(0),
11620 PointerType::get(EI.getType(), AS),EI);
11621 GetElementPtrInst *GEP =
11622 GetElementPtrInst::Create(Ptr, EI.getOperand(1), I->getName()+".gep");
11623 InsertNewInstBefore(GEP, EI);
11624 return new LoadInst(GEP);
11627 if (InsertElementInst *IE = dyn_cast<InsertElementInst>(I)) {
11628 // Extracting the inserted element?
11629 if (IE->getOperand(2) == EI.getOperand(1))
11630 return ReplaceInstUsesWith(EI, IE->getOperand(1));
11631 // If the inserted and extracted elements are constants, they must not
11632 // be the same value, extract from the pre-inserted value instead.
11633 if (isa<Constant>(IE->getOperand(2)) &&
11634 isa<Constant>(EI.getOperand(1))) {
11635 AddUsesToWorkList(EI);
11636 EI.setOperand(0, IE->getOperand(0));
11639 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I)) {
11640 // If this is extracting an element from a shufflevector, figure out where
11641 // it came from and extract from the appropriate input element instead.
11642 if (ConstantInt *Elt = dyn_cast<ConstantInt>(EI.getOperand(1))) {
11643 unsigned SrcIdx = getShuffleMask(SVI)[Elt->getZExtValue()];
11645 unsigned LHSWidth =
11646 cast<VectorType>(SVI->getOperand(0)->getType())->getNumElements();
11648 if (SrcIdx < LHSWidth)
11649 Src = SVI->getOperand(0);
11650 else if (SrcIdx < LHSWidth*2) {
11651 SrcIdx -= LHSWidth;
11652 Src = SVI->getOperand(1);
11654 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
11656 return new ExtractElementInst(Src, SrcIdx);
11663 /// CollectSingleShuffleElements - If V is a shuffle of values that ONLY returns
11664 /// elements from either LHS or RHS, return the shuffle mask and true.
11665 /// Otherwise, return false.
11666 static bool CollectSingleShuffleElements(Value *V, Value *LHS, Value *RHS,
11667 std::vector<Constant*> &Mask) {
11668 assert(V->getType() == LHS->getType() && V->getType() == RHS->getType() &&
11669 "Invalid CollectSingleShuffleElements");
11670 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
11672 if (isa<UndefValue>(V)) {
11673 Mask.assign(NumElts, UndefValue::get(Type::Int32Ty));
11675 } else if (V == LHS) {
11676 for (unsigned i = 0; i != NumElts; ++i)
11677 Mask.push_back(ConstantInt::get(Type::Int32Ty, i));
11679 } else if (V == RHS) {
11680 for (unsigned i = 0; i != NumElts; ++i)
11681 Mask.push_back(ConstantInt::get(Type::Int32Ty, i+NumElts));
11683 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
11684 // If this is an insert of an extract from some other vector, include it.
11685 Value *VecOp = IEI->getOperand(0);
11686 Value *ScalarOp = IEI->getOperand(1);
11687 Value *IdxOp = IEI->getOperand(2);
11689 if (!isa<ConstantInt>(IdxOp))
11691 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
11693 if (isa<UndefValue>(ScalarOp)) { // inserting undef into vector.
11694 // Okay, we can handle this if the vector we are insertinting into is
11695 // transitively ok.
11696 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
11697 // If so, update the mask to reflect the inserted undef.
11698 Mask[InsertedIdx] = UndefValue::get(Type::Int32Ty);
11701 } else if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)){
11702 if (isa<ConstantInt>(EI->getOperand(1)) &&
11703 EI->getOperand(0)->getType() == V->getType()) {
11704 unsigned ExtractedIdx =
11705 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
11707 // This must be extracting from either LHS or RHS.
11708 if (EI->getOperand(0) == LHS || EI->getOperand(0) == RHS) {
11709 // Okay, we can handle this if the vector we are insertinting into is
11710 // transitively ok.
11711 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
11712 // If so, update the mask to reflect the inserted value.
11713 if (EI->getOperand(0) == LHS) {
11714 Mask[InsertedIdx % NumElts] =
11715 ConstantInt::get(Type::Int32Ty, ExtractedIdx);
11717 assert(EI->getOperand(0) == RHS);
11718 Mask[InsertedIdx % NumElts] =
11719 ConstantInt::get(Type::Int32Ty, ExtractedIdx+NumElts);
11728 // TODO: Handle shufflevector here!
11733 /// CollectShuffleElements - We are building a shuffle of V, using RHS as the
11734 /// RHS of the shuffle instruction, if it is not null. Return a shuffle mask
11735 /// that computes V and the LHS value of the shuffle.
11736 static Value *CollectShuffleElements(Value *V, std::vector<Constant*> &Mask,
11738 assert(isa<VectorType>(V->getType()) &&
11739 (RHS == 0 || V->getType() == RHS->getType()) &&
11740 "Invalid shuffle!");
11741 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
11743 if (isa<UndefValue>(V)) {
11744 Mask.assign(NumElts, UndefValue::get(Type::Int32Ty));
11746 } else if (isa<ConstantAggregateZero>(V)) {
11747 Mask.assign(NumElts, ConstantInt::get(Type::Int32Ty, 0));
11749 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
11750 // If this is an insert of an extract from some other vector, include it.
11751 Value *VecOp = IEI->getOperand(0);
11752 Value *ScalarOp = IEI->getOperand(1);
11753 Value *IdxOp = IEI->getOperand(2);
11755 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
11756 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
11757 EI->getOperand(0)->getType() == V->getType()) {
11758 unsigned ExtractedIdx =
11759 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
11760 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
11762 // Either the extracted from or inserted into vector must be RHSVec,
11763 // otherwise we'd end up with a shuffle of three inputs.
11764 if (EI->getOperand(0) == RHS || RHS == 0) {
11765 RHS = EI->getOperand(0);
11766 Value *V = CollectShuffleElements(VecOp, Mask, RHS);
11767 Mask[InsertedIdx % NumElts] =
11768 ConstantInt::get(Type::Int32Ty, NumElts+ExtractedIdx);
11772 if (VecOp == RHS) {
11773 Value *V = CollectShuffleElements(EI->getOperand(0), Mask, RHS);
11774 // Everything but the extracted element is replaced with the RHS.
11775 for (unsigned i = 0; i != NumElts; ++i) {
11776 if (i != InsertedIdx)
11777 Mask[i] = ConstantInt::get(Type::Int32Ty, NumElts+i);
11782 // If this insertelement is a chain that comes from exactly these two
11783 // vectors, return the vector and the effective shuffle.
11784 if (CollectSingleShuffleElements(IEI, EI->getOperand(0), RHS, Mask))
11785 return EI->getOperand(0);
11790 // TODO: Handle shufflevector here!
11792 // Otherwise, can't do anything fancy. Return an identity vector.
11793 for (unsigned i = 0; i != NumElts; ++i)
11794 Mask.push_back(ConstantInt::get(Type::Int32Ty, i));
11798 Instruction *InstCombiner::visitInsertElementInst(InsertElementInst &IE) {
11799 Value *VecOp = IE.getOperand(0);
11800 Value *ScalarOp = IE.getOperand(1);
11801 Value *IdxOp = IE.getOperand(2);
11803 // Inserting an undef or into an undefined place, remove this.
11804 if (isa<UndefValue>(ScalarOp) || isa<UndefValue>(IdxOp))
11805 ReplaceInstUsesWith(IE, VecOp);
11807 // If the inserted element was extracted from some other vector, and if the
11808 // indexes are constant, try to turn this into a shufflevector operation.
11809 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
11810 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
11811 EI->getOperand(0)->getType() == IE.getType()) {
11812 unsigned NumVectorElts = IE.getType()->getNumElements();
11813 unsigned ExtractedIdx =
11814 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
11815 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
11817 if (ExtractedIdx >= NumVectorElts) // Out of range extract.
11818 return ReplaceInstUsesWith(IE, VecOp);
11820 if (InsertedIdx >= NumVectorElts) // Out of range insert.
11821 return ReplaceInstUsesWith(IE, UndefValue::get(IE.getType()));
11823 // If we are extracting a value from a vector, then inserting it right
11824 // back into the same place, just use the input vector.
11825 if (EI->getOperand(0) == VecOp && ExtractedIdx == InsertedIdx)
11826 return ReplaceInstUsesWith(IE, VecOp);
11828 // We could theoretically do this for ANY input. However, doing so could
11829 // turn chains of insertelement instructions into a chain of shufflevector
11830 // instructions, and right now we do not merge shufflevectors. As such,
11831 // only do this in a situation where it is clear that there is benefit.
11832 if (isa<UndefValue>(VecOp) || isa<ConstantAggregateZero>(VecOp)) {
11833 // Turn this into shuffle(EIOp0, VecOp, Mask). The result has all of
11834 // the values of VecOp, except then one read from EIOp0.
11835 // Build a new shuffle mask.
11836 std::vector<Constant*> Mask;
11837 if (isa<UndefValue>(VecOp))
11838 Mask.assign(NumVectorElts, UndefValue::get(Type::Int32Ty));
11840 assert(isa<ConstantAggregateZero>(VecOp) && "Unknown thing");
11841 Mask.assign(NumVectorElts, ConstantInt::get(Type::Int32Ty,
11844 Mask[InsertedIdx] = ConstantInt::get(Type::Int32Ty, ExtractedIdx);
11845 return new ShuffleVectorInst(EI->getOperand(0), VecOp,
11846 ConstantVector::get(Mask));
11849 // If this insertelement isn't used by some other insertelement, turn it
11850 // (and any insertelements it points to), into one big shuffle.
11851 if (!IE.hasOneUse() || !isa<InsertElementInst>(IE.use_back())) {
11852 std::vector<Constant*> Mask;
11854 Value *LHS = CollectShuffleElements(&IE, Mask, RHS);
11855 if (RHS == 0) RHS = UndefValue::get(LHS->getType());
11856 // We now have a shuffle of LHS, RHS, Mask.
11857 return new ShuffleVectorInst(LHS, RHS, ConstantVector::get(Mask));
11866 Instruction *InstCombiner::visitShuffleVectorInst(ShuffleVectorInst &SVI) {
11867 Value *LHS = SVI.getOperand(0);
11868 Value *RHS = SVI.getOperand(1);
11869 std::vector<unsigned> Mask = getShuffleMask(&SVI);
11871 bool MadeChange = false;
11873 // Undefined shuffle mask -> undefined value.
11874 if (isa<UndefValue>(SVI.getOperand(2)))
11875 return ReplaceInstUsesWith(SVI, UndefValue::get(SVI.getType()));
11877 uint64_t UndefElts;
11878 unsigned VWidth = cast<VectorType>(SVI.getType())->getNumElements();
11880 if (VWidth != cast<VectorType>(LHS->getType())->getNumElements())
11883 uint64_t AllOnesEltMask = ~0ULL >> (64-VWidth);
11884 if (VWidth <= 64 &&
11885 SimplifyDemandedVectorElts(&SVI, AllOnesEltMask, UndefElts)) {
11886 LHS = SVI.getOperand(0);
11887 RHS = SVI.getOperand(1);
11891 // Canonicalize shuffle(x ,x,mask) -> shuffle(x, undef,mask')
11892 // Canonicalize shuffle(undef,x,mask) -> shuffle(x, undef,mask').
11893 if (LHS == RHS || isa<UndefValue>(LHS)) {
11894 if (isa<UndefValue>(LHS) && LHS == RHS) {
11895 // shuffle(undef,undef,mask) -> undef.
11896 return ReplaceInstUsesWith(SVI, LHS);
11899 // Remap any references to RHS to use LHS.
11900 std::vector<Constant*> Elts;
11901 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
11902 if (Mask[i] >= 2*e)
11903 Elts.push_back(UndefValue::get(Type::Int32Ty));
11905 if ((Mask[i] >= e && isa<UndefValue>(RHS)) ||
11906 (Mask[i] < e && isa<UndefValue>(LHS))) {
11907 Mask[i] = 2*e; // Turn into undef.
11908 Elts.push_back(UndefValue::get(Type::Int32Ty));
11910 Mask[i] = Mask[i] % e; // Force to LHS.
11911 Elts.push_back(ConstantInt::get(Type::Int32Ty, Mask[i]));
11915 SVI.setOperand(0, SVI.getOperand(1));
11916 SVI.setOperand(1, UndefValue::get(RHS->getType()));
11917 SVI.setOperand(2, ConstantVector::get(Elts));
11918 LHS = SVI.getOperand(0);
11919 RHS = SVI.getOperand(1);
11923 // Analyze the shuffle, are the LHS or RHS and identity shuffles?
11924 bool isLHSID = true, isRHSID = true;
11926 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
11927 if (Mask[i] >= e*2) continue; // Ignore undef values.
11928 // Is this an identity shuffle of the LHS value?
11929 isLHSID &= (Mask[i] == i);
11931 // Is this an identity shuffle of the RHS value?
11932 isRHSID &= (Mask[i]-e == i);
11935 // Eliminate identity shuffles.
11936 if (isLHSID) return ReplaceInstUsesWith(SVI, LHS);
11937 if (isRHSID) return ReplaceInstUsesWith(SVI, RHS);
11939 // If the LHS is a shufflevector itself, see if we can combine it with this
11940 // one without producing an unusual shuffle. Here we are really conservative:
11941 // we are absolutely afraid of producing a shuffle mask not in the input
11942 // program, because the code gen may not be smart enough to turn a merged
11943 // shuffle into two specific shuffles: it may produce worse code. As such,
11944 // we only merge two shuffles if the result is one of the two input shuffle
11945 // masks. In this case, merging the shuffles just removes one instruction,
11946 // which we know is safe. This is good for things like turning:
11947 // (splat(splat)) -> splat.
11948 if (ShuffleVectorInst *LHSSVI = dyn_cast<ShuffleVectorInst>(LHS)) {
11949 if (isa<UndefValue>(RHS)) {
11950 std::vector<unsigned> LHSMask = getShuffleMask(LHSSVI);
11952 std::vector<unsigned> NewMask;
11953 for (unsigned i = 0, e = Mask.size(); i != e; ++i)
11954 if (Mask[i] >= 2*e)
11955 NewMask.push_back(2*e);
11957 NewMask.push_back(LHSMask[Mask[i]]);
11959 // If the result mask is equal to the src shuffle or this shuffle mask, do
11960 // the replacement.
11961 if (NewMask == LHSMask || NewMask == Mask) {
11962 std::vector<Constant*> Elts;
11963 for (unsigned i = 0, e = NewMask.size(); i != e; ++i) {
11964 if (NewMask[i] >= e*2) {
11965 Elts.push_back(UndefValue::get(Type::Int32Ty));
11967 Elts.push_back(ConstantInt::get(Type::Int32Ty, NewMask[i]));
11970 return new ShuffleVectorInst(LHSSVI->getOperand(0),
11971 LHSSVI->getOperand(1),
11972 ConstantVector::get(Elts));
11977 return MadeChange ? &SVI : 0;
11983 /// TryToSinkInstruction - Try to move the specified instruction from its
11984 /// current block into the beginning of DestBlock, which can only happen if it's
11985 /// safe to move the instruction past all of the instructions between it and the
11986 /// end of its block.
11987 static bool TryToSinkInstruction(Instruction *I, BasicBlock *DestBlock) {
11988 assert(I->hasOneUse() && "Invariants didn't hold!");
11990 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
11991 if (isa<PHINode>(I) || I->mayWriteToMemory() || isa<TerminatorInst>(I))
11994 // Do not sink alloca instructions out of the entry block.
11995 if (isa<AllocaInst>(I) && I->getParent() ==
11996 &DestBlock->getParent()->getEntryBlock())
11999 // We can only sink load instructions if there is nothing between the load and
12000 // the end of block that could change the value.
12001 if (I->mayReadFromMemory()) {
12002 for (BasicBlock::iterator Scan = I, E = I->getParent()->end();
12004 if (Scan->mayWriteToMemory())
12008 BasicBlock::iterator InsertPos = DestBlock->getFirstNonPHI();
12010 I->moveBefore(InsertPos);
12016 /// AddReachableCodeToWorklist - Walk the function in depth-first order, adding
12017 /// all reachable code to the worklist.
12019 /// This has a couple of tricks to make the code faster and more powerful. In
12020 /// particular, we constant fold and DCE instructions as we go, to avoid adding
12021 /// them to the worklist (this significantly speeds up instcombine on code where
12022 /// many instructions are dead or constant). Additionally, if we find a branch
12023 /// whose condition is a known constant, we only visit the reachable successors.
12025 static void AddReachableCodeToWorklist(BasicBlock *BB,
12026 SmallPtrSet<BasicBlock*, 64> &Visited,
12028 const TargetData *TD) {
12029 SmallVector<BasicBlock*, 256> Worklist;
12030 Worklist.push_back(BB);
12032 while (!Worklist.empty()) {
12033 BB = Worklist.back();
12034 Worklist.pop_back();
12036 // We have now visited this block! If we've already been here, ignore it.
12037 if (!Visited.insert(BB)) continue;
12039 DbgInfoIntrinsic *DBI_Prev = NULL;
12040 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) {
12041 Instruction *Inst = BBI++;
12043 // DCE instruction if trivially dead.
12044 if (isInstructionTriviallyDead(Inst)) {
12046 DOUT << "IC: DCE: " << *Inst;
12047 Inst->eraseFromParent();
12051 // ConstantProp instruction if trivially constant.
12052 if (Constant *C = ConstantFoldInstruction(Inst, TD)) {
12053 DOUT << "IC: ConstFold to: " << *C << " from: " << *Inst;
12054 Inst->replaceAllUsesWith(C);
12056 Inst->eraseFromParent();
12060 // If there are two consecutive llvm.dbg.stoppoint calls then
12061 // it is likely that the optimizer deleted code in between these
12063 DbgInfoIntrinsic *DBI_Next = dyn_cast<DbgInfoIntrinsic>(Inst);
12066 && DBI_Prev->getIntrinsicID() == llvm::Intrinsic::dbg_stoppoint
12067 && DBI_Next->getIntrinsicID() == llvm::Intrinsic::dbg_stoppoint) {
12068 IC.RemoveFromWorkList(DBI_Prev);
12069 DBI_Prev->eraseFromParent();
12071 DBI_Prev = DBI_Next;
12074 IC.AddToWorkList(Inst);
12077 // Recursively visit successors. If this is a branch or switch on a
12078 // constant, only visit the reachable successor.
12079 TerminatorInst *TI = BB->getTerminator();
12080 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
12081 if (BI->isConditional() && isa<ConstantInt>(BI->getCondition())) {
12082 bool CondVal = cast<ConstantInt>(BI->getCondition())->getZExtValue();
12083 BasicBlock *ReachableBB = BI->getSuccessor(!CondVal);
12084 Worklist.push_back(ReachableBB);
12087 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
12088 if (ConstantInt *Cond = dyn_cast<ConstantInt>(SI->getCondition())) {
12089 // See if this is an explicit destination.
12090 for (unsigned i = 1, e = SI->getNumSuccessors(); i != e; ++i)
12091 if (SI->getCaseValue(i) == Cond) {
12092 BasicBlock *ReachableBB = SI->getSuccessor(i);
12093 Worklist.push_back(ReachableBB);
12097 // Otherwise it is the default destination.
12098 Worklist.push_back(SI->getSuccessor(0));
12103 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
12104 Worklist.push_back(TI->getSuccessor(i));
12108 bool InstCombiner::DoOneIteration(Function &F, unsigned Iteration) {
12109 bool Changed = false;
12110 TD = &getAnalysis<TargetData>();
12112 DEBUG(DOUT << "\n\nINSTCOMBINE ITERATION #" << Iteration << " on "
12113 << F.getNameStr() << "\n");
12116 // Do a depth-first traversal of the function, populate the worklist with
12117 // the reachable instructions. Ignore blocks that are not reachable. Keep
12118 // track of which blocks we visit.
12119 SmallPtrSet<BasicBlock*, 64> Visited;
12120 AddReachableCodeToWorklist(F.begin(), Visited, *this, TD);
12122 // Do a quick scan over the function. If we find any blocks that are
12123 // unreachable, remove any instructions inside of them. This prevents
12124 // the instcombine code from having to deal with some bad special cases.
12125 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
12126 if (!Visited.count(BB)) {
12127 Instruction *Term = BB->getTerminator();
12128 while (Term != BB->begin()) { // Remove instrs bottom-up
12129 BasicBlock::iterator I = Term; --I;
12131 DOUT << "IC: DCE: " << *I;
12134 if (!I->use_empty())
12135 I->replaceAllUsesWith(UndefValue::get(I->getType()));
12136 I->eraseFromParent();
12141 while (!Worklist.empty()) {
12142 Instruction *I = RemoveOneFromWorkList();
12143 if (I == 0) continue; // skip null values.
12145 // Check to see if we can DCE the instruction.
12146 if (isInstructionTriviallyDead(I)) {
12147 // Add operands to the worklist.
12148 if (I->getNumOperands() < 4)
12149 AddUsesToWorkList(*I);
12152 DOUT << "IC: DCE: " << *I;
12154 I->eraseFromParent();
12155 RemoveFromWorkList(I);
12159 // Instruction isn't dead, see if we can constant propagate it.
12160 if (Constant *C = ConstantFoldInstruction(I, TD)) {
12161 DOUT << "IC: ConstFold to: " << *C << " from: " << *I;
12163 // Add operands to the worklist.
12164 AddUsesToWorkList(*I);
12165 ReplaceInstUsesWith(*I, C);
12168 I->eraseFromParent();
12169 RemoveFromWorkList(I);
12173 if (TD && I->getType()->getTypeID() == Type::VoidTyID) {
12174 // See if we can constant fold its operands.
12175 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i) {
12176 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(i)) {
12177 if (Constant *NewC = ConstantFoldConstantExpression(CE, TD))
12183 // See if we can trivially sink this instruction to a successor basic block.
12184 if (I->hasOneUse()) {
12185 BasicBlock *BB = I->getParent();
12186 BasicBlock *UserParent = cast<Instruction>(I->use_back())->getParent();
12187 if (UserParent != BB) {
12188 bool UserIsSuccessor = false;
12189 // See if the user is one of our successors.
12190 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI)
12191 if (*SI == UserParent) {
12192 UserIsSuccessor = true;
12196 // If the user is one of our immediate successors, and if that successor
12197 // only has us as a predecessors (we'd have to split the critical edge
12198 // otherwise), we can keep going.
12199 if (UserIsSuccessor && !isa<PHINode>(I->use_back()) &&
12200 next(pred_begin(UserParent)) == pred_end(UserParent))
12201 // Okay, the CFG is simple enough, try to sink this instruction.
12202 Changed |= TryToSinkInstruction(I, UserParent);
12206 // Now that we have an instruction, try combining it to simplify it...
12210 DEBUG(std::ostringstream SS; I->print(SS); OrigI = SS.str(););
12211 if (Instruction *Result = visit(*I)) {
12213 // Should we replace the old instruction with a new one?
12215 DOUT << "IC: Old = " << *I
12216 << " New = " << *Result;
12218 // Everything uses the new instruction now.
12219 I->replaceAllUsesWith(Result);
12221 // Push the new instruction and any users onto the worklist.
12222 AddToWorkList(Result);
12223 AddUsersToWorkList(*Result);
12225 // Move the name to the new instruction first.
12226 Result->takeName(I);
12228 // Insert the new instruction into the basic block...
12229 BasicBlock *InstParent = I->getParent();
12230 BasicBlock::iterator InsertPos = I;
12232 if (!isa<PHINode>(Result)) // If combining a PHI, don't insert
12233 while (isa<PHINode>(InsertPos)) // middle of a block of PHIs.
12236 InstParent->getInstList().insert(InsertPos, Result);
12238 // Make sure that we reprocess all operands now that we reduced their
12240 AddUsesToWorkList(*I);
12242 // Instructions can end up on the worklist more than once. Make sure
12243 // we do not process an instruction that has been deleted.
12244 RemoveFromWorkList(I);
12246 // Erase the old instruction.
12247 InstParent->getInstList().erase(I);
12250 DOUT << "IC: Mod = " << OrigI
12251 << " New = " << *I;
12254 // If the instruction was modified, it's possible that it is now dead.
12255 // if so, remove it.
12256 if (isInstructionTriviallyDead(I)) {
12257 // Make sure we process all operands now that we are reducing their
12259 AddUsesToWorkList(*I);
12261 // Instructions may end up in the worklist more than once. Erase all
12262 // occurrences of this instruction.
12263 RemoveFromWorkList(I);
12264 I->eraseFromParent();
12267 AddUsersToWorkList(*I);
12274 assert(WorklistMap.empty() && "Worklist empty, but map not?");
12276 // Do an explicit clear, this shrinks the map if needed.
12277 WorklistMap.clear();
12282 bool InstCombiner::runOnFunction(Function &F) {
12283 MustPreserveLCSSA = mustPreserveAnalysisID(LCSSAID);
12285 bool EverMadeChange = false;
12287 // Iterate while there is work to do.
12288 unsigned Iteration = 0;
12289 while (DoOneIteration(F, Iteration++))
12290 EverMadeChange = true;
12291 return EverMadeChange;
12294 FunctionPass *llvm::createInstructionCombiningPass() {
12295 return new InstCombiner();