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 *visitAnd(BinaryOperator &I);
184 Instruction *visitOr (BinaryOperator &I);
185 Instruction *visitXor(BinaryOperator &I);
186 Instruction *visitShl(BinaryOperator &I);
187 Instruction *visitAShr(BinaryOperator &I);
188 Instruction *visitLShr(BinaryOperator &I);
189 Instruction *commonShiftTransforms(BinaryOperator &I);
190 Instruction *FoldFCmp_IntToFP_Cst(FCmpInst &I, Instruction *LHSI,
192 Instruction *visitFCmpInst(FCmpInst &I);
193 Instruction *visitICmpInst(ICmpInst &I);
194 Instruction *visitICmpInstWithCastAndCast(ICmpInst &ICI);
195 Instruction *visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
198 Instruction *FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
199 ConstantInt *DivRHS);
201 Instruction *FoldGEPICmp(User *GEPLHS, Value *RHS,
202 ICmpInst::Predicate Cond, Instruction &I);
203 Instruction *FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
205 Instruction *commonCastTransforms(CastInst &CI);
206 Instruction *commonIntCastTransforms(CastInst &CI);
207 Instruction *commonPointerCastTransforms(CastInst &CI);
208 Instruction *visitTrunc(TruncInst &CI);
209 Instruction *visitZExt(ZExtInst &CI);
210 Instruction *visitSExt(SExtInst &CI);
211 Instruction *visitFPTrunc(FPTruncInst &CI);
212 Instruction *visitFPExt(CastInst &CI);
213 Instruction *visitFPToUI(FPToUIInst &FI);
214 Instruction *visitFPToSI(FPToSIInst &FI);
215 Instruction *visitUIToFP(CastInst &CI);
216 Instruction *visitSIToFP(CastInst &CI);
217 Instruction *visitPtrToInt(CastInst &CI);
218 Instruction *visitIntToPtr(IntToPtrInst &CI);
219 Instruction *visitBitCast(BitCastInst &CI);
220 Instruction *FoldSelectOpOp(SelectInst &SI, Instruction *TI,
222 Instruction *visitSelectInst(SelectInst &SI);
223 Instruction *visitSelectInstWithICmp(SelectInst &SI, ICmpInst *ICI);
224 Instruction *visitCallInst(CallInst &CI);
225 Instruction *visitInvokeInst(InvokeInst &II);
226 Instruction *visitPHINode(PHINode &PN);
227 Instruction *visitGetElementPtrInst(GetElementPtrInst &GEP);
228 Instruction *visitAllocationInst(AllocationInst &AI);
229 Instruction *visitFreeInst(FreeInst &FI);
230 Instruction *visitLoadInst(LoadInst &LI);
231 Instruction *visitStoreInst(StoreInst &SI);
232 Instruction *visitBranchInst(BranchInst &BI);
233 Instruction *visitSwitchInst(SwitchInst &SI);
234 Instruction *visitInsertElementInst(InsertElementInst &IE);
235 Instruction *visitExtractElementInst(ExtractElementInst &EI);
236 Instruction *visitShuffleVectorInst(ShuffleVectorInst &SVI);
237 Instruction *visitExtractValueInst(ExtractValueInst &EV);
239 // visitInstruction - Specify what to return for unhandled instructions...
240 Instruction *visitInstruction(Instruction &I) { return 0; }
243 Instruction *visitCallSite(CallSite CS);
244 bool transformConstExprCastCall(CallSite CS);
245 Instruction *transformCallThroughTrampoline(CallSite CS);
246 Instruction *transformZExtICmp(ICmpInst *ICI, Instruction &CI,
247 bool DoXform = true);
248 bool WillNotOverflowSignedAdd(Value *LHS, Value *RHS);
251 // InsertNewInstBefore - insert an instruction New before instruction Old
252 // in the program. Add the new instruction to the worklist.
254 Instruction *InsertNewInstBefore(Instruction *New, Instruction &Old) {
255 assert(New && New->getParent() == 0 &&
256 "New instruction already inserted into a basic block!");
257 BasicBlock *BB = Old.getParent();
258 BB->getInstList().insert(&Old, New); // Insert inst
263 /// InsertCastBefore - Insert a cast of V to TY before the instruction POS.
264 /// This also adds the cast to the worklist. Finally, this returns the
266 Value *InsertCastBefore(Instruction::CastOps opc, Value *V, const Type *Ty,
268 if (V->getType() == Ty) return V;
270 if (Constant *CV = dyn_cast<Constant>(V))
271 return ConstantExpr::getCast(opc, CV, Ty);
273 Instruction *C = CastInst::Create(opc, V, Ty, V->getName(), &Pos);
278 Value *InsertBitCastBefore(Value *V, const Type *Ty, Instruction &Pos) {
279 return InsertCastBefore(Instruction::BitCast, V, Ty, Pos);
283 // ReplaceInstUsesWith - This method is to be used when an instruction is
284 // found to be dead, replacable with another preexisting expression. Here
285 // we add all uses of I to the worklist, replace all uses of I with the new
286 // value, then return I, so that the inst combiner will know that I was
289 Instruction *ReplaceInstUsesWith(Instruction &I, Value *V) {
290 AddUsersToWorkList(I); // Add all modified instrs to worklist
292 I.replaceAllUsesWith(V);
295 // If we are replacing the instruction with itself, this must be in a
296 // segment of unreachable code, so just clobber the instruction.
297 I.replaceAllUsesWith(UndefValue::get(I.getType()));
302 // UpdateValueUsesWith - This method is to be used when an value is
303 // found to be replacable with another preexisting expression or was
304 // updated. Here we add all uses of I to the worklist, replace all uses of
305 // I with the new value (unless the instruction was just updated), then
306 // return true, so that the inst combiner will know that I was modified.
308 bool UpdateValueUsesWith(Value *Old, Value *New) {
309 AddUsersToWorkList(*Old); // Add all modified instrs to worklist
311 Old->replaceAllUsesWith(New);
312 if (Instruction *I = dyn_cast<Instruction>(Old))
314 if (Instruction *I = dyn_cast<Instruction>(New))
319 // EraseInstFromFunction - When dealing with an instruction that has side
320 // effects or produces a void value, we can't rely on DCE to delete the
321 // instruction. Instead, visit methods should return the value returned by
323 Instruction *EraseInstFromFunction(Instruction &I) {
324 assert(I.use_empty() && "Cannot erase instruction that is used!");
325 AddUsesToWorkList(I);
326 RemoveFromWorkList(&I);
328 return 0; // Don't do anything with FI
331 void ComputeMaskedBits(Value *V, const APInt &Mask, APInt &KnownZero,
332 APInt &KnownOne, unsigned Depth = 0) const {
333 return llvm::ComputeMaskedBits(V, Mask, KnownZero, KnownOne, TD, Depth);
336 bool MaskedValueIsZero(Value *V, const APInt &Mask,
337 unsigned Depth = 0) const {
338 return llvm::MaskedValueIsZero(V, Mask, TD, Depth);
340 unsigned ComputeNumSignBits(Value *Op, unsigned Depth = 0) const {
341 return llvm::ComputeNumSignBits(Op, TD, Depth);
345 /// InsertOperandCastBefore - This inserts a cast of V to DestTy before the
346 /// InsertBefore instruction. This is specialized a bit to avoid inserting
347 /// casts that are known to not do anything...
349 Value *InsertOperandCastBefore(Instruction::CastOps opcode,
350 Value *V, const Type *DestTy,
351 Instruction *InsertBefore);
353 /// SimplifyCommutative - This performs a few simplifications for
354 /// commutative operators.
355 bool SimplifyCommutative(BinaryOperator &I);
357 /// SimplifyCompare - This reorders the operands of a CmpInst to get them in
358 /// most-complex to least-complex order.
359 bool SimplifyCompare(CmpInst &I);
361 /// SimplifyDemandedBits - Attempts to replace V with a simpler value based
362 /// on the demanded bits.
363 bool SimplifyDemandedBits(Value *V, APInt DemandedMask,
364 APInt& KnownZero, APInt& KnownOne,
367 Value *SimplifyDemandedVectorElts(Value *V, uint64_t DemandedElts,
368 uint64_t &UndefElts, unsigned Depth = 0);
370 // FoldOpIntoPhi - Given a binary operator or cast instruction which has a
371 // PHI node as operand #0, see if we can fold the instruction into the PHI
372 // (which is only possible if all operands to the PHI are constants).
373 Instruction *FoldOpIntoPhi(Instruction &I);
375 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
376 // operator and they all are only used by the PHI, PHI together their
377 // inputs, and do the operation once, to the result of the PHI.
378 Instruction *FoldPHIArgOpIntoPHI(PHINode &PN);
379 Instruction *FoldPHIArgBinOpIntoPHI(PHINode &PN);
382 Instruction *OptAndOp(Instruction *Op, ConstantInt *OpRHS,
383 ConstantInt *AndRHS, BinaryOperator &TheAnd);
385 Value *FoldLogicalPlusAnd(Value *LHS, Value *RHS, ConstantInt *Mask,
386 bool isSub, Instruction &I);
387 Instruction *InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
388 bool isSigned, bool Inside, Instruction &IB);
389 Instruction *PromoteCastOfAllocation(BitCastInst &CI, AllocationInst &AI);
390 Instruction *MatchBSwap(BinaryOperator &I);
391 bool SimplifyStoreAtEndOfBlock(StoreInst &SI);
392 Instruction *SimplifyMemTransfer(MemIntrinsic *MI);
393 Instruction *SimplifyMemSet(MemSetInst *MI);
396 Value *EvaluateInDifferentType(Value *V, const Type *Ty, bool isSigned);
398 bool CanEvaluateInDifferentType(Value *V, const IntegerType *Ty,
400 int &NumCastsRemoved);
401 unsigned GetOrEnforceKnownAlignment(Value *V,
402 unsigned PrefAlign = 0);
407 char InstCombiner::ID = 0;
408 static RegisterPass<InstCombiner>
409 X("instcombine", "Combine redundant instructions");
411 // getComplexity: Assign a complexity or rank value to LLVM Values...
412 // 0 -> undef, 1 -> Const, 2 -> Other, 3 -> Arg, 3 -> Unary, 4 -> OtherInst
413 static unsigned getComplexity(Value *V) {
414 if (isa<Instruction>(V)) {
415 if (BinaryOperator::isNeg(V) || BinaryOperator::isNot(V))
419 if (isa<Argument>(V)) return 3;
420 return isa<Constant>(V) ? (isa<UndefValue>(V) ? 0 : 1) : 2;
423 // isOnlyUse - Return true if this instruction will be deleted if we stop using
425 static bool isOnlyUse(Value *V) {
426 return V->hasOneUse() || isa<Constant>(V);
429 // getPromotedType - Return the specified type promoted as it would be to pass
430 // though a va_arg area...
431 static const Type *getPromotedType(const Type *Ty) {
432 if (const IntegerType* ITy = dyn_cast<IntegerType>(Ty)) {
433 if (ITy->getBitWidth() < 32)
434 return Type::Int32Ty;
439 /// getBitCastOperand - If the specified operand is a CastInst or a constant
440 /// expression bitcast, return the operand value, otherwise return null.
441 static Value *getBitCastOperand(Value *V) {
442 if (BitCastInst *I = dyn_cast<BitCastInst>(V))
443 return I->getOperand(0);
444 else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
445 if (CE->getOpcode() == Instruction::BitCast)
446 return CE->getOperand(0);
450 /// This function is a wrapper around CastInst::isEliminableCastPair. It
451 /// simply extracts arguments and returns what that function returns.
452 static Instruction::CastOps
453 isEliminableCastPair(
454 const CastInst *CI, ///< The first cast instruction
455 unsigned opcode, ///< The opcode of the second cast instruction
456 const Type *DstTy, ///< The target type for the second cast instruction
457 TargetData *TD ///< The target data for pointer size
460 const Type *SrcTy = CI->getOperand(0)->getType(); // A from above
461 const Type *MidTy = CI->getType(); // B from above
463 // Get the opcodes of the two Cast instructions
464 Instruction::CastOps firstOp = Instruction::CastOps(CI->getOpcode());
465 Instruction::CastOps secondOp = Instruction::CastOps(opcode);
467 return Instruction::CastOps(
468 CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy,
469 DstTy, TD->getIntPtrType()));
472 /// ValueRequiresCast - Return true if the cast from "V to Ty" actually results
473 /// in any code being generated. It does not require codegen if V is simple
474 /// enough or if the cast can be folded into other casts.
475 static bool ValueRequiresCast(Instruction::CastOps opcode, const Value *V,
476 const Type *Ty, TargetData *TD) {
477 if (V->getType() == Ty || isa<Constant>(V)) return false;
479 // If this is another cast that can be eliminated, it isn't codegen either.
480 if (const CastInst *CI = dyn_cast<CastInst>(V))
481 if (isEliminableCastPair(CI, opcode, Ty, TD))
486 /// InsertOperandCastBefore - This inserts a cast of V to DestTy before the
487 /// InsertBefore instruction. This is specialized a bit to avoid inserting
488 /// casts that are known to not do anything...
490 Value *InstCombiner::InsertOperandCastBefore(Instruction::CastOps opcode,
491 Value *V, const Type *DestTy,
492 Instruction *InsertBefore) {
493 if (V->getType() == DestTy) return V;
494 if (Constant *C = dyn_cast<Constant>(V))
495 return ConstantExpr::getCast(opcode, C, DestTy);
497 return InsertCastBefore(opcode, V, DestTy, *InsertBefore);
500 // SimplifyCommutative - This performs a few simplifications for commutative
503 // 1. Order operands such that they are listed from right (least complex) to
504 // left (most complex). This puts constants before unary operators before
507 // 2. Transform: (op (op V, C1), C2) ==> (op V, (op C1, C2))
508 // 3. Transform: (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
510 bool InstCombiner::SimplifyCommutative(BinaryOperator &I) {
511 bool Changed = false;
512 if (getComplexity(I.getOperand(0)) < getComplexity(I.getOperand(1)))
513 Changed = !I.swapOperands();
515 if (!I.isAssociative()) return Changed;
516 Instruction::BinaryOps Opcode = I.getOpcode();
517 if (BinaryOperator *Op = dyn_cast<BinaryOperator>(I.getOperand(0)))
518 if (Op->getOpcode() == Opcode && isa<Constant>(Op->getOperand(1))) {
519 if (isa<Constant>(I.getOperand(1))) {
520 Constant *Folded = ConstantExpr::get(I.getOpcode(),
521 cast<Constant>(I.getOperand(1)),
522 cast<Constant>(Op->getOperand(1)));
523 I.setOperand(0, Op->getOperand(0));
524 I.setOperand(1, Folded);
526 } else if (BinaryOperator *Op1=dyn_cast<BinaryOperator>(I.getOperand(1)))
527 if (Op1->getOpcode() == Opcode && isa<Constant>(Op1->getOperand(1)) &&
528 isOnlyUse(Op) && isOnlyUse(Op1)) {
529 Constant *C1 = cast<Constant>(Op->getOperand(1));
530 Constant *C2 = cast<Constant>(Op1->getOperand(1));
532 // Fold (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
533 Constant *Folded = ConstantExpr::get(I.getOpcode(), C1, C2);
534 Instruction *New = BinaryOperator::Create(Opcode, Op->getOperand(0),
538 I.setOperand(0, New);
539 I.setOperand(1, Folded);
546 /// SimplifyCompare - For a CmpInst this function just orders the operands
547 /// so that theyare listed from right (least complex) to left (most complex).
548 /// This puts constants before unary operators before binary operators.
549 bool InstCombiner::SimplifyCompare(CmpInst &I) {
550 if (getComplexity(I.getOperand(0)) >= getComplexity(I.getOperand(1)))
553 // Compare instructions are not associative so there's nothing else we can do.
557 // dyn_castNegVal - Given a 'sub' instruction, return the RHS of the instruction
558 // if the LHS is a constant zero (which is the 'negate' form).
560 static inline Value *dyn_castNegVal(Value *V) {
561 if (BinaryOperator::isNeg(V))
562 return BinaryOperator::getNegArgument(V);
564 // Constants can be considered to be negated values if they can be folded.
565 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
566 return ConstantExpr::getNeg(C);
568 if (ConstantVector *C = dyn_cast<ConstantVector>(V))
569 if (C->getType()->getElementType()->isInteger())
570 return ConstantExpr::getNeg(C);
575 static inline Value *dyn_castNotVal(Value *V) {
576 if (BinaryOperator::isNot(V))
577 return BinaryOperator::getNotArgument(V);
579 // Constants can be considered to be not'ed values...
580 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
581 return ConstantInt::get(~C->getValue());
585 // dyn_castFoldableMul - If this value is a multiply that can be folded into
586 // other computations (because it has a constant operand), return the
587 // non-constant operand of the multiply, and set CST to point to the multiplier.
588 // Otherwise, return null.
590 static inline Value *dyn_castFoldableMul(Value *V, ConstantInt *&CST) {
591 if (V->hasOneUse() && V->getType()->isInteger())
592 if (Instruction *I = dyn_cast<Instruction>(V)) {
593 if (I->getOpcode() == Instruction::Mul)
594 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1))))
595 return I->getOperand(0);
596 if (I->getOpcode() == Instruction::Shl)
597 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1)))) {
598 // The multiplier is really 1 << CST.
599 uint32_t BitWidth = cast<IntegerType>(V->getType())->getBitWidth();
600 uint32_t CSTVal = CST->getLimitedValue(BitWidth);
601 CST = ConstantInt::get(APInt(BitWidth, 1).shl(CSTVal));
602 return I->getOperand(0);
608 /// dyn_castGetElementPtr - If this is a getelementptr instruction or constant
609 /// expression, return it.
610 static User *dyn_castGetElementPtr(Value *V) {
611 if (isa<GetElementPtrInst>(V)) return cast<User>(V);
612 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
613 if (CE->getOpcode() == Instruction::GetElementPtr)
614 return cast<User>(V);
618 /// getOpcode - If this is an Instruction or a ConstantExpr, return the
619 /// opcode value. Otherwise return UserOp1.
620 static unsigned getOpcode(const Value *V) {
621 if (const Instruction *I = dyn_cast<Instruction>(V))
622 return I->getOpcode();
623 if (const ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
624 return CE->getOpcode();
625 // Use UserOp1 to mean there's no opcode.
626 return Instruction::UserOp1;
629 /// AddOne - Add one to a ConstantInt
630 static ConstantInt *AddOne(ConstantInt *C) {
631 APInt Val(C->getValue());
632 return ConstantInt::get(++Val);
634 /// SubOne - Subtract one from a ConstantInt
635 static ConstantInt *SubOne(ConstantInt *C) {
636 APInt Val(C->getValue());
637 return ConstantInt::get(--Val);
639 /// Add - Add two ConstantInts together
640 static ConstantInt *Add(ConstantInt *C1, ConstantInt *C2) {
641 return ConstantInt::get(C1->getValue() + C2->getValue());
643 /// And - Bitwise AND two ConstantInts together
644 static ConstantInt *And(ConstantInt *C1, ConstantInt *C2) {
645 return ConstantInt::get(C1->getValue() & C2->getValue());
647 /// Subtract - Subtract one ConstantInt from another
648 static ConstantInt *Subtract(ConstantInt *C1, ConstantInt *C2) {
649 return ConstantInt::get(C1->getValue() - C2->getValue());
651 /// Multiply - Multiply two ConstantInts together
652 static ConstantInt *Multiply(ConstantInt *C1, ConstantInt *C2) {
653 return ConstantInt::get(C1->getValue() * C2->getValue());
655 /// MultiplyOverflows - True if the multiply can not be expressed in an int
657 static bool MultiplyOverflows(ConstantInt *C1, ConstantInt *C2, bool sign) {
658 uint32_t W = C1->getBitWidth();
659 APInt LHSExt = C1->getValue(), RHSExt = C2->getValue();
668 APInt MulExt = LHSExt * RHSExt;
671 APInt Min = APInt::getSignedMinValue(W).sext(W * 2);
672 APInt Max = APInt::getSignedMaxValue(W).sext(W * 2);
673 return MulExt.slt(Min) || MulExt.sgt(Max);
675 return MulExt.ugt(APInt::getLowBitsSet(W * 2, W));
679 /// ShrinkDemandedConstant - Check to see if the specified operand of the
680 /// specified instruction is a constant integer. If so, check to see if there
681 /// are any bits set in the constant that are not demanded. If so, shrink the
682 /// constant and return true.
683 static bool ShrinkDemandedConstant(Instruction *I, unsigned OpNo,
685 assert(I && "No instruction?");
686 assert(OpNo < I->getNumOperands() && "Operand index too large");
688 // If the operand is not a constant integer, nothing to do.
689 ConstantInt *OpC = dyn_cast<ConstantInt>(I->getOperand(OpNo));
690 if (!OpC) return false;
692 // If there are no bits set that aren't demanded, nothing to do.
693 Demanded.zextOrTrunc(OpC->getValue().getBitWidth());
694 if ((~Demanded & OpC->getValue()) == 0)
697 // This instruction is producing bits that are not demanded. Shrink the RHS.
698 Demanded &= OpC->getValue();
699 I->setOperand(OpNo, ConstantInt::get(Demanded));
703 // ComputeSignedMinMaxValuesFromKnownBits - Given a signed integer type and a
704 // set of known zero and one bits, compute the maximum and minimum values that
705 // could have the specified known zero and known one bits, returning them in
707 static void ComputeSignedMinMaxValuesFromKnownBits(const Type *Ty,
708 const APInt& KnownZero,
709 const APInt& KnownOne,
710 APInt& Min, APInt& Max) {
711 uint32_t BitWidth = cast<IntegerType>(Ty)->getBitWidth();
712 assert(KnownZero.getBitWidth() == BitWidth &&
713 KnownOne.getBitWidth() == BitWidth &&
714 Min.getBitWidth() == BitWidth && Max.getBitWidth() == BitWidth &&
715 "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth.");
716 APInt UnknownBits = ~(KnownZero|KnownOne);
718 // The minimum value is when all unknown bits are zeros, EXCEPT for the sign
719 // bit if it is unknown.
721 Max = KnownOne|UnknownBits;
723 if (UnknownBits[BitWidth-1]) { // Sign bit is unknown
725 Max.clear(BitWidth-1);
729 // ComputeUnsignedMinMaxValuesFromKnownBits - Given an unsigned integer type and
730 // a set of known zero and one bits, compute the maximum and minimum values that
731 // could have the specified known zero and known one bits, returning them in
733 static void ComputeUnsignedMinMaxValuesFromKnownBits(const Type *Ty,
734 const APInt &KnownZero,
735 const APInt &KnownOne,
736 APInt &Min, APInt &Max) {
737 uint32_t BitWidth = cast<IntegerType>(Ty)->getBitWidth(); BitWidth = BitWidth;
738 assert(KnownZero.getBitWidth() == BitWidth &&
739 KnownOne.getBitWidth() == BitWidth &&
740 Min.getBitWidth() == BitWidth && Max.getBitWidth() &&
741 "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth.");
742 APInt UnknownBits = ~(KnownZero|KnownOne);
744 // The minimum value is when the unknown bits are all zeros.
746 // The maximum value is when the unknown bits are all ones.
747 Max = KnownOne|UnknownBits;
750 /// SimplifyDemandedBits - This function attempts to replace V with a simpler
751 /// value based on the demanded bits. When this function is called, it is known
752 /// that only the bits set in DemandedMask of the result of V are ever used
753 /// downstream. Consequently, depending on the mask and V, it may be possible
754 /// to replace V with a constant or one of its operands. In such cases, this
755 /// function does the replacement and returns true. In all other cases, it
756 /// returns false after analyzing the expression and setting KnownOne and known
757 /// to be one in the expression. KnownZero contains all the bits that are known
758 /// to be zero in the expression. These are provided to potentially allow the
759 /// caller (which might recursively be SimplifyDemandedBits itself) to simplify
760 /// the expression. KnownOne and KnownZero always follow the invariant that
761 /// KnownOne & KnownZero == 0. That is, a bit can't be both 1 and 0. Note that
762 /// the bits in KnownOne and KnownZero may only be accurate for those bits set
763 /// in DemandedMask. Note also that the bitwidth of V, DemandedMask, KnownZero
764 /// and KnownOne must all be the same.
765 bool InstCombiner::SimplifyDemandedBits(Value *V, APInt DemandedMask,
766 APInt& KnownZero, APInt& KnownOne,
768 assert(V != 0 && "Null pointer of Value???");
769 assert(Depth <= 6 && "Limit Search Depth");
770 uint32_t BitWidth = DemandedMask.getBitWidth();
771 const IntegerType *VTy = cast<IntegerType>(V->getType());
772 assert(VTy->getBitWidth() == BitWidth &&
773 KnownZero.getBitWidth() == BitWidth &&
774 KnownOne.getBitWidth() == BitWidth &&
775 "Value *V, DemandedMask, KnownZero and KnownOne \
776 must have same BitWidth");
777 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
778 // We know all of the bits for a constant!
779 KnownOne = CI->getValue() & DemandedMask;
780 KnownZero = ~KnownOne & DemandedMask;
786 if (!V->hasOneUse()) { // Other users may use these bits.
787 if (Depth != 0) { // Not at the root.
788 // Just compute the KnownZero/KnownOne bits to simplify things downstream.
789 ComputeMaskedBits(V, DemandedMask, KnownZero, KnownOne, Depth);
792 // If this is the root being simplified, allow it to have multiple uses,
793 // just set the DemandedMask to all bits.
794 DemandedMask = APInt::getAllOnesValue(BitWidth);
795 } else if (DemandedMask == 0) { // Not demanding any bits from V.
796 if (V != UndefValue::get(VTy))
797 return UpdateValueUsesWith(V, UndefValue::get(VTy));
799 } else if (Depth == 6) { // Limit search depth.
803 Instruction *I = dyn_cast<Instruction>(V);
804 if (!I) return false; // Only analyze instructions.
806 APInt LHSKnownZero(BitWidth, 0), LHSKnownOne(BitWidth, 0);
807 APInt &RHSKnownZero = KnownZero, &RHSKnownOne = KnownOne;
808 switch (I->getOpcode()) {
810 ComputeMaskedBits(V, DemandedMask, RHSKnownZero, RHSKnownOne, Depth);
812 case Instruction::And:
813 // If either the LHS or the RHS are Zero, the result is zero.
814 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
815 RHSKnownZero, RHSKnownOne, Depth+1))
817 assert((RHSKnownZero & RHSKnownOne) == 0 &&
818 "Bits known to be one AND zero?");
820 // If something is known zero on the RHS, the bits aren't demanded on the
822 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask & ~RHSKnownZero,
823 LHSKnownZero, LHSKnownOne, Depth+1))
825 assert((LHSKnownZero & LHSKnownOne) == 0 &&
826 "Bits known to be one AND zero?");
828 // If all of the demanded bits are known 1 on one side, return the other.
829 // These bits cannot contribute to the result of the 'and'.
830 if ((DemandedMask & ~LHSKnownZero & RHSKnownOne) ==
831 (DemandedMask & ~LHSKnownZero))
832 return UpdateValueUsesWith(I, I->getOperand(0));
833 if ((DemandedMask & ~RHSKnownZero & LHSKnownOne) ==
834 (DemandedMask & ~RHSKnownZero))
835 return UpdateValueUsesWith(I, I->getOperand(1));
837 // If all of the demanded bits in the inputs are known zeros, return zero.
838 if ((DemandedMask & (RHSKnownZero|LHSKnownZero)) == DemandedMask)
839 return UpdateValueUsesWith(I, Constant::getNullValue(VTy));
841 // If the RHS is a constant, see if we can simplify it.
842 if (ShrinkDemandedConstant(I, 1, DemandedMask & ~LHSKnownZero))
843 return UpdateValueUsesWith(I, I);
845 // Output known-1 bits are only known if set in both the LHS & RHS.
846 RHSKnownOne &= LHSKnownOne;
847 // Output known-0 are known to be clear if zero in either the LHS | RHS.
848 RHSKnownZero |= LHSKnownZero;
850 case Instruction::Or:
851 // If either the LHS or the RHS are One, the result is One.
852 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
853 RHSKnownZero, RHSKnownOne, Depth+1))
855 assert((RHSKnownZero & RHSKnownOne) == 0 &&
856 "Bits known to be one AND zero?");
857 // If something is known one on the RHS, the bits aren't demanded on the
859 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask & ~RHSKnownOne,
860 LHSKnownZero, LHSKnownOne, Depth+1))
862 assert((LHSKnownZero & LHSKnownOne) == 0 &&
863 "Bits known to be one AND zero?");
865 // If all of the demanded bits are known zero on one side, return the other.
866 // These bits cannot contribute to the result of the 'or'.
867 if ((DemandedMask & ~LHSKnownOne & RHSKnownZero) ==
868 (DemandedMask & ~LHSKnownOne))
869 return UpdateValueUsesWith(I, I->getOperand(0));
870 if ((DemandedMask & ~RHSKnownOne & LHSKnownZero) ==
871 (DemandedMask & ~RHSKnownOne))
872 return UpdateValueUsesWith(I, I->getOperand(1));
874 // If all of the potentially set bits on one side are known to be set on
875 // the other side, just use the 'other' side.
876 if ((DemandedMask & (~RHSKnownZero) & LHSKnownOne) ==
877 (DemandedMask & (~RHSKnownZero)))
878 return UpdateValueUsesWith(I, I->getOperand(0));
879 if ((DemandedMask & (~LHSKnownZero) & RHSKnownOne) ==
880 (DemandedMask & (~LHSKnownZero)))
881 return UpdateValueUsesWith(I, I->getOperand(1));
883 // If the RHS is a constant, see if we can simplify it.
884 if (ShrinkDemandedConstant(I, 1, DemandedMask))
885 return UpdateValueUsesWith(I, I);
887 // Output known-0 bits are only known if clear in both the LHS & RHS.
888 RHSKnownZero &= LHSKnownZero;
889 // Output known-1 are known to be set if set in either the LHS | RHS.
890 RHSKnownOne |= LHSKnownOne;
892 case Instruction::Xor: {
893 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
894 RHSKnownZero, RHSKnownOne, Depth+1))
896 assert((RHSKnownZero & RHSKnownOne) == 0 &&
897 "Bits known to be one AND zero?");
898 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
899 LHSKnownZero, LHSKnownOne, Depth+1))
901 assert((LHSKnownZero & LHSKnownOne) == 0 &&
902 "Bits known to be one AND zero?");
904 // If all of the demanded bits are known zero on one side, return the other.
905 // These bits cannot contribute to the result of the 'xor'.
906 if ((DemandedMask & RHSKnownZero) == DemandedMask)
907 return UpdateValueUsesWith(I, I->getOperand(0));
908 if ((DemandedMask & LHSKnownZero) == DemandedMask)
909 return UpdateValueUsesWith(I, I->getOperand(1));
911 // Output known-0 bits are known if clear or set in both the LHS & RHS.
912 APInt KnownZeroOut = (RHSKnownZero & LHSKnownZero) |
913 (RHSKnownOne & LHSKnownOne);
914 // Output known-1 are known to be set if set in only one of the LHS, RHS.
915 APInt KnownOneOut = (RHSKnownZero & LHSKnownOne) |
916 (RHSKnownOne & LHSKnownZero);
918 // If all of the demanded bits are known to be zero on one side or the
919 // other, turn this into an *inclusive* or.
920 // e.g. (A & C1)^(B & C2) -> (A & C1)|(B & C2) iff C1&C2 == 0
921 if ((DemandedMask & ~RHSKnownZero & ~LHSKnownZero) == 0) {
923 BinaryOperator::CreateOr(I->getOperand(0), I->getOperand(1),
925 InsertNewInstBefore(Or, *I);
926 return UpdateValueUsesWith(I, Or);
929 // If all of the demanded bits on one side are known, and all of the set
930 // bits on that side are also known to be set on the other side, turn this
931 // into an AND, as we know the bits will be cleared.
932 // e.g. (X | C1) ^ C2 --> (X | C1) & ~C2 iff (C1&C2) == C2
933 if ((DemandedMask & (RHSKnownZero|RHSKnownOne)) == DemandedMask) {
935 if ((RHSKnownOne & LHSKnownOne) == RHSKnownOne) {
936 Constant *AndC = ConstantInt::get(~RHSKnownOne & DemandedMask);
938 BinaryOperator::CreateAnd(I->getOperand(0), AndC, "tmp");
939 InsertNewInstBefore(And, *I);
940 return UpdateValueUsesWith(I, And);
944 // If the RHS is a constant, see if we can simplify it.
945 // FIXME: for XOR, we prefer to force bits to 1 if they will make a -1.
946 if (ShrinkDemandedConstant(I, 1, DemandedMask))
947 return UpdateValueUsesWith(I, I);
949 RHSKnownZero = KnownZeroOut;
950 RHSKnownOne = KnownOneOut;
953 case Instruction::Select:
954 if (SimplifyDemandedBits(I->getOperand(2), DemandedMask,
955 RHSKnownZero, RHSKnownOne, Depth+1))
957 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
958 LHSKnownZero, LHSKnownOne, Depth+1))
960 assert((RHSKnownZero & RHSKnownOne) == 0 &&
961 "Bits known to be one AND zero?");
962 assert((LHSKnownZero & LHSKnownOne) == 0 &&
963 "Bits known to be one AND zero?");
965 // If the operands are constants, see if we can simplify them.
966 if (ShrinkDemandedConstant(I, 1, DemandedMask))
967 return UpdateValueUsesWith(I, I);
968 if (ShrinkDemandedConstant(I, 2, DemandedMask))
969 return UpdateValueUsesWith(I, I);
971 // Only known if known in both the LHS and RHS.
972 RHSKnownOne &= LHSKnownOne;
973 RHSKnownZero &= LHSKnownZero;
975 case Instruction::Trunc: {
977 cast<IntegerType>(I->getOperand(0)->getType())->getBitWidth();
978 DemandedMask.zext(truncBf);
979 RHSKnownZero.zext(truncBf);
980 RHSKnownOne.zext(truncBf);
981 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
982 RHSKnownZero, RHSKnownOne, Depth+1))
984 DemandedMask.trunc(BitWidth);
985 RHSKnownZero.trunc(BitWidth);
986 RHSKnownOne.trunc(BitWidth);
987 assert((RHSKnownZero & RHSKnownOne) == 0 &&
988 "Bits known to be one AND zero?");
991 case Instruction::BitCast:
992 if (!I->getOperand(0)->getType()->isInteger())
995 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
996 RHSKnownZero, RHSKnownOne, Depth+1))
998 assert((RHSKnownZero & RHSKnownOne) == 0 &&
999 "Bits known to be one AND zero?");
1001 case Instruction::ZExt: {
1002 // Compute the bits in the result that are not present in the input.
1003 const IntegerType *SrcTy = cast<IntegerType>(I->getOperand(0)->getType());
1004 uint32_t SrcBitWidth = SrcTy->getBitWidth();
1006 DemandedMask.trunc(SrcBitWidth);
1007 RHSKnownZero.trunc(SrcBitWidth);
1008 RHSKnownOne.trunc(SrcBitWidth);
1009 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1010 RHSKnownZero, RHSKnownOne, Depth+1))
1012 DemandedMask.zext(BitWidth);
1013 RHSKnownZero.zext(BitWidth);
1014 RHSKnownOne.zext(BitWidth);
1015 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1016 "Bits known to be one AND zero?");
1017 // The top bits are known to be zero.
1018 RHSKnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
1021 case Instruction::SExt: {
1022 // Compute the bits in the result that are not present in the input.
1023 const IntegerType *SrcTy = cast<IntegerType>(I->getOperand(0)->getType());
1024 uint32_t SrcBitWidth = SrcTy->getBitWidth();
1026 APInt InputDemandedBits = DemandedMask &
1027 APInt::getLowBitsSet(BitWidth, SrcBitWidth);
1029 APInt NewBits(APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth));
1030 // If any of the sign extended bits are demanded, we know that the sign
1032 if ((NewBits & DemandedMask) != 0)
1033 InputDemandedBits.set(SrcBitWidth-1);
1035 InputDemandedBits.trunc(SrcBitWidth);
1036 RHSKnownZero.trunc(SrcBitWidth);
1037 RHSKnownOne.trunc(SrcBitWidth);
1038 if (SimplifyDemandedBits(I->getOperand(0), InputDemandedBits,
1039 RHSKnownZero, RHSKnownOne, Depth+1))
1041 InputDemandedBits.zext(BitWidth);
1042 RHSKnownZero.zext(BitWidth);
1043 RHSKnownOne.zext(BitWidth);
1044 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1045 "Bits known to be one AND zero?");
1047 // If the sign bit of the input is known set or clear, then we know the
1048 // top bits of the result.
1050 // If the input sign bit is known zero, or if the NewBits are not demanded
1051 // convert this into a zero extension.
1052 if (RHSKnownZero[SrcBitWidth-1] || (NewBits & ~DemandedMask) == NewBits)
1054 // Convert to ZExt cast
1055 CastInst *NewCast = new ZExtInst(I->getOperand(0), VTy, I->getName(), I);
1056 return UpdateValueUsesWith(I, NewCast);
1057 } else if (RHSKnownOne[SrcBitWidth-1]) { // Input sign bit known set
1058 RHSKnownOne |= NewBits;
1062 case Instruction::Add: {
1063 // Figure out what the input bits are. If the top bits of the and result
1064 // are not demanded, then the add doesn't demand them from its input
1066 uint32_t NLZ = DemandedMask.countLeadingZeros();
1068 // If there is a constant on the RHS, there are a variety of xformations
1070 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
1071 // If null, this should be simplified elsewhere. Some of the xforms here
1072 // won't work if the RHS is zero.
1076 // If the top bit of the output is demanded, demand everything from the
1077 // input. Otherwise, we demand all the input bits except NLZ top bits.
1078 APInt InDemandedBits(APInt::getLowBitsSet(BitWidth, BitWidth - NLZ));
1080 // Find information about known zero/one bits in the input.
1081 if (SimplifyDemandedBits(I->getOperand(0), InDemandedBits,
1082 LHSKnownZero, LHSKnownOne, Depth+1))
1085 // If the RHS of the add has bits set that can't affect the input, reduce
1087 if (ShrinkDemandedConstant(I, 1, InDemandedBits))
1088 return UpdateValueUsesWith(I, I);
1090 // Avoid excess work.
1091 if (LHSKnownZero == 0 && LHSKnownOne == 0)
1094 // Turn it into OR if input bits are zero.
1095 if ((LHSKnownZero & RHS->getValue()) == RHS->getValue()) {
1097 BinaryOperator::CreateOr(I->getOperand(0), I->getOperand(1),
1099 InsertNewInstBefore(Or, *I);
1100 return UpdateValueUsesWith(I, Or);
1103 // We can say something about the output known-zero and known-one bits,
1104 // depending on potential carries from the input constant and the
1105 // unknowns. For example if the LHS is known to have at most the 0x0F0F0
1106 // bits set and the RHS constant is 0x01001, then we know we have a known
1107 // one mask of 0x00001 and a known zero mask of 0xE0F0E.
1109 // To compute this, we first compute the potential carry bits. These are
1110 // the bits which may be modified. I'm not aware of a better way to do
1112 const APInt& RHSVal = RHS->getValue();
1113 APInt CarryBits((~LHSKnownZero + RHSVal) ^ (~LHSKnownZero ^ RHSVal));
1115 // Now that we know which bits have carries, compute the known-1/0 sets.
1117 // Bits are known one if they are known zero in one operand and one in the
1118 // other, and there is no input carry.
1119 RHSKnownOne = ((LHSKnownZero & RHSVal) |
1120 (LHSKnownOne & ~RHSVal)) & ~CarryBits;
1122 // Bits are known zero if they are known zero in both operands and there
1123 // is no input carry.
1124 RHSKnownZero = LHSKnownZero & ~RHSVal & ~CarryBits;
1126 // If the high-bits of this ADD are not demanded, then it does not demand
1127 // the high bits of its LHS or RHS.
1128 if (DemandedMask[BitWidth-1] == 0) {
1129 // Right fill the mask of bits for this ADD to demand the most
1130 // significant bit and all those below it.
1131 APInt DemandedFromOps(APInt::getLowBitsSet(BitWidth, BitWidth-NLZ));
1132 if (SimplifyDemandedBits(I->getOperand(0), DemandedFromOps,
1133 LHSKnownZero, LHSKnownOne, Depth+1))
1135 if (SimplifyDemandedBits(I->getOperand(1), DemandedFromOps,
1136 LHSKnownZero, LHSKnownOne, Depth+1))
1142 case Instruction::Sub:
1143 // If the high-bits of this SUB are not demanded, then it does not demand
1144 // the high bits of its LHS or RHS.
1145 if (DemandedMask[BitWidth-1] == 0) {
1146 // Right fill the mask of bits for this SUB to demand the most
1147 // significant bit and all those below it.
1148 uint32_t NLZ = DemandedMask.countLeadingZeros();
1149 APInt DemandedFromOps(APInt::getLowBitsSet(BitWidth, BitWidth-NLZ));
1150 if (SimplifyDemandedBits(I->getOperand(0), DemandedFromOps,
1151 LHSKnownZero, LHSKnownOne, Depth+1))
1153 if (SimplifyDemandedBits(I->getOperand(1), DemandedFromOps,
1154 LHSKnownZero, LHSKnownOne, Depth+1))
1157 // Otherwise just hand the sub off to ComputeMaskedBits to fill in
1158 // the known zeros and ones.
1159 ComputeMaskedBits(V, DemandedMask, RHSKnownZero, RHSKnownOne, Depth);
1161 case Instruction::Shl:
1162 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1163 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
1164 APInt DemandedMaskIn(DemandedMask.lshr(ShiftAmt));
1165 if (SimplifyDemandedBits(I->getOperand(0), DemandedMaskIn,
1166 RHSKnownZero, RHSKnownOne, Depth+1))
1168 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1169 "Bits known to be one AND zero?");
1170 RHSKnownZero <<= ShiftAmt;
1171 RHSKnownOne <<= ShiftAmt;
1172 // low bits known zero.
1174 RHSKnownZero |= APInt::getLowBitsSet(BitWidth, ShiftAmt);
1177 case Instruction::LShr:
1178 // For a logical shift right
1179 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1180 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
1182 // Unsigned shift right.
1183 APInt DemandedMaskIn(DemandedMask.shl(ShiftAmt));
1184 if (SimplifyDemandedBits(I->getOperand(0), DemandedMaskIn,
1185 RHSKnownZero, RHSKnownOne, Depth+1))
1187 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1188 "Bits known to be one AND zero?");
1189 RHSKnownZero = APIntOps::lshr(RHSKnownZero, ShiftAmt);
1190 RHSKnownOne = APIntOps::lshr(RHSKnownOne, ShiftAmt);
1192 // Compute the new bits that are at the top now.
1193 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
1194 RHSKnownZero |= HighBits; // high bits known zero.
1198 case Instruction::AShr:
1199 // If this is an arithmetic shift right and only the low-bit is set, we can
1200 // always convert this into a logical shr, even if the shift amount is
1201 // variable. The low bit of the shift cannot be an input sign bit unless
1202 // the shift amount is >= the size of the datatype, which is undefined.
1203 if (DemandedMask == 1) {
1204 // Perform the logical shift right.
1205 Value *NewVal = BinaryOperator::CreateLShr(
1206 I->getOperand(0), I->getOperand(1), I->getName());
1207 InsertNewInstBefore(cast<Instruction>(NewVal), *I);
1208 return UpdateValueUsesWith(I, NewVal);
1211 // If the sign bit is the only bit demanded by this ashr, then there is no
1212 // need to do it, the shift doesn't change the high bit.
1213 if (DemandedMask.isSignBit())
1214 return UpdateValueUsesWith(I, I->getOperand(0));
1216 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1217 uint32_t ShiftAmt = SA->getLimitedValue(BitWidth);
1219 // Signed shift right.
1220 APInt DemandedMaskIn(DemandedMask.shl(ShiftAmt));
1221 // If any of the "high bits" are demanded, we should set the sign bit as
1223 if (DemandedMask.countLeadingZeros() <= ShiftAmt)
1224 DemandedMaskIn.set(BitWidth-1);
1225 if (SimplifyDemandedBits(I->getOperand(0),
1227 RHSKnownZero, RHSKnownOne, Depth+1))
1229 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1230 "Bits known to be one AND zero?");
1231 // Compute the new bits that are at the top now.
1232 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
1233 RHSKnownZero = APIntOps::lshr(RHSKnownZero, ShiftAmt);
1234 RHSKnownOne = APIntOps::lshr(RHSKnownOne, ShiftAmt);
1236 // Handle the sign bits.
1237 APInt SignBit(APInt::getSignBit(BitWidth));
1238 // Adjust to where it is now in the mask.
1239 SignBit = APIntOps::lshr(SignBit, ShiftAmt);
1241 // If the input sign bit is known to be zero, or if none of the top bits
1242 // are demanded, turn this into an unsigned shift right.
1243 if (BitWidth <= ShiftAmt || RHSKnownZero[BitWidth-ShiftAmt-1] ||
1244 (HighBits & ~DemandedMask) == HighBits) {
1245 // Perform the logical shift right.
1246 Value *NewVal = BinaryOperator::CreateLShr(
1247 I->getOperand(0), SA, I->getName());
1248 InsertNewInstBefore(cast<Instruction>(NewVal), *I);
1249 return UpdateValueUsesWith(I, NewVal);
1250 } else if ((RHSKnownOne & SignBit) != 0) { // New bits are known one.
1251 RHSKnownOne |= HighBits;
1255 case Instruction::SRem:
1256 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
1257 APInt RA = Rem->getValue();
1258 if (RA.isPowerOf2() || (-RA).isPowerOf2()) {
1259 if (DemandedMask.ule(RA)) // srem won't affect demanded bits
1260 return UpdateValueUsesWith(I, I->getOperand(0));
1262 APInt LowBits = RA.isStrictlyPositive() ? (RA - 1) : ~RA;
1263 APInt Mask2 = LowBits | APInt::getSignBit(BitWidth);
1264 if (SimplifyDemandedBits(I->getOperand(0), Mask2,
1265 LHSKnownZero, LHSKnownOne, Depth+1))
1268 if (LHSKnownZero[BitWidth-1] || ((LHSKnownZero & LowBits) == LowBits))
1269 LHSKnownZero |= ~LowBits;
1271 KnownZero |= LHSKnownZero & DemandedMask;
1273 assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?");
1277 case Instruction::URem: {
1278 APInt KnownZero2(BitWidth, 0), KnownOne2(BitWidth, 0);
1279 APInt AllOnes = APInt::getAllOnesValue(BitWidth);
1280 if (SimplifyDemandedBits(I->getOperand(0), AllOnes,
1281 KnownZero2, KnownOne2, Depth+1))
1284 uint32_t Leaders = KnownZero2.countLeadingOnes();
1285 if (SimplifyDemandedBits(I->getOperand(1), AllOnes,
1286 KnownZero2, KnownOne2, Depth+1))
1289 Leaders = std::max(Leaders,
1290 KnownZero2.countLeadingOnes());
1291 KnownZero = APInt::getHighBitsSet(BitWidth, Leaders) & DemandedMask;
1294 case Instruction::Call:
1295 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
1296 switch (II->getIntrinsicID()) {
1298 case Intrinsic::bswap: {
1299 // If the only bits demanded come from one byte of the bswap result,
1300 // just shift the input byte into position to eliminate the bswap.
1301 unsigned NLZ = DemandedMask.countLeadingZeros();
1302 unsigned NTZ = DemandedMask.countTrailingZeros();
1304 // Round NTZ down to the next byte. If we have 11 trailing zeros, then
1305 // we need all the bits down to bit 8. Likewise, round NLZ. If we
1306 // have 14 leading zeros, round to 8.
1309 // If we need exactly one byte, we can do this transformation.
1310 if (BitWidth-NLZ-NTZ == 8) {
1311 unsigned ResultBit = NTZ;
1312 unsigned InputBit = BitWidth-NTZ-8;
1314 // Replace this with either a left or right shift to get the byte into
1316 Instruction *NewVal;
1317 if (InputBit > ResultBit)
1318 NewVal = BinaryOperator::CreateLShr(I->getOperand(1),
1319 ConstantInt::get(I->getType(), InputBit-ResultBit));
1321 NewVal = BinaryOperator::CreateShl(I->getOperand(1),
1322 ConstantInt::get(I->getType(), ResultBit-InputBit));
1323 NewVal->takeName(I);
1324 InsertNewInstBefore(NewVal, *I);
1325 return UpdateValueUsesWith(I, NewVal);
1328 // TODO: Could compute known zero/one bits based on the input.
1333 ComputeMaskedBits(V, DemandedMask, RHSKnownZero, RHSKnownOne, Depth);
1337 // If the client is only demanding bits that we know, return the known
1339 if ((DemandedMask & (RHSKnownZero|RHSKnownOne)) == DemandedMask)
1340 return UpdateValueUsesWith(I, ConstantInt::get(RHSKnownOne));
1345 /// SimplifyDemandedVectorElts - The specified value producecs a vector with
1346 /// 64 or fewer elements. DemandedElts contains the set of elements that are
1347 /// actually used by the caller. This method analyzes which elements of the
1348 /// operand are undef and returns that information in UndefElts.
1350 /// If the information about demanded elements can be used to simplify the
1351 /// operation, the operation is simplified, then the resultant value is
1352 /// returned. This returns null if no change was made.
1353 Value *InstCombiner::SimplifyDemandedVectorElts(Value *V, uint64_t DemandedElts,
1354 uint64_t &UndefElts,
1356 unsigned VWidth = cast<VectorType>(V->getType())->getNumElements();
1357 assert(VWidth <= 64 && "Vector too wide to analyze!");
1358 uint64_t EltMask = ~0ULL >> (64-VWidth);
1359 assert((DemandedElts & ~EltMask) == 0 && "Invalid DemandedElts!");
1361 if (isa<UndefValue>(V)) {
1362 // If the entire vector is undefined, just return this info.
1363 UndefElts = EltMask;
1365 } else if (DemandedElts == 0) { // If nothing is demanded, provide undef.
1366 UndefElts = EltMask;
1367 return UndefValue::get(V->getType());
1371 if (ConstantVector *CP = dyn_cast<ConstantVector>(V)) {
1372 const Type *EltTy = cast<VectorType>(V->getType())->getElementType();
1373 Constant *Undef = UndefValue::get(EltTy);
1375 std::vector<Constant*> Elts;
1376 for (unsigned i = 0; i != VWidth; ++i)
1377 if (!(DemandedElts & (1ULL << i))) { // If not demanded, set to undef.
1378 Elts.push_back(Undef);
1379 UndefElts |= (1ULL << i);
1380 } else if (isa<UndefValue>(CP->getOperand(i))) { // Already undef.
1381 Elts.push_back(Undef);
1382 UndefElts |= (1ULL << i);
1383 } else { // Otherwise, defined.
1384 Elts.push_back(CP->getOperand(i));
1387 // If we changed the constant, return it.
1388 Constant *NewCP = ConstantVector::get(Elts);
1389 return NewCP != CP ? NewCP : 0;
1390 } else if (isa<ConstantAggregateZero>(V)) {
1391 // Simplify the CAZ to a ConstantVector where the non-demanded elements are
1393 const Type *EltTy = cast<VectorType>(V->getType())->getElementType();
1394 Constant *Zero = Constant::getNullValue(EltTy);
1395 Constant *Undef = UndefValue::get(EltTy);
1396 std::vector<Constant*> Elts;
1397 for (unsigned i = 0; i != VWidth; ++i)
1398 Elts.push_back((DemandedElts & (1ULL << i)) ? Zero : Undef);
1399 UndefElts = DemandedElts ^ EltMask;
1400 return ConstantVector::get(Elts);
1403 // Limit search depth.
1407 // If multiple users are using the root value, procede with
1408 // simplification conservatively assuming that all elements
1410 if (!V->hasOneUse()) {
1411 // Quit if we find multiple users of a non-root value though.
1412 // They'll be handled when it's their turn to be visited by
1413 // the main instcombine process.
1415 // TODO: Just compute the UndefElts information recursively.
1418 // Conservatively assume that all elements are needed.
1419 DemandedElts = EltMask;
1422 Instruction *I = dyn_cast<Instruction>(V);
1423 if (!I) return false; // Only analyze instructions.
1425 bool MadeChange = false;
1426 uint64_t UndefElts2;
1428 switch (I->getOpcode()) {
1431 case Instruction::InsertElement: {
1432 // If this is a variable index, we don't know which element it overwrites.
1433 // demand exactly the same input as we produce.
1434 ConstantInt *Idx = dyn_cast<ConstantInt>(I->getOperand(2));
1436 // Note that we can't propagate undef elt info, because we don't know
1437 // which elt is getting updated.
1438 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
1439 UndefElts2, Depth+1);
1440 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1444 // If this is inserting an element that isn't demanded, remove this
1446 unsigned IdxNo = Idx->getZExtValue();
1447 if (IdxNo >= VWidth || (DemandedElts & (1ULL << IdxNo)) == 0)
1448 return AddSoonDeadInstToWorklist(*I, 0);
1450 // Otherwise, the element inserted overwrites whatever was there, so the
1451 // input demanded set is simpler than the output set.
1452 TmpV = SimplifyDemandedVectorElts(I->getOperand(0),
1453 DemandedElts & ~(1ULL << IdxNo),
1454 UndefElts, Depth+1);
1455 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1457 // The inserted element is defined.
1458 UndefElts &= ~(1ULL << IdxNo);
1461 case Instruction::ShuffleVector: {
1462 ShuffleVectorInst *Shuffle = cast<ShuffleVectorInst>(I);
1463 uint64_t LeftDemanded = 0, RightDemanded = 0;
1464 for (unsigned i = 0; i < VWidth; i++) {
1465 if (DemandedElts & (1ULL << i)) {
1466 unsigned MaskVal = Shuffle->getMaskValue(i);
1467 if (MaskVal != -1u) {
1468 assert(MaskVal < VWidth * 2 &&
1469 "shufflevector mask index out of range!");
1470 if (MaskVal < VWidth)
1471 LeftDemanded |= 1ULL << MaskVal;
1473 RightDemanded |= 1ULL << (MaskVal - VWidth);
1478 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), LeftDemanded,
1479 UndefElts2, Depth+1);
1480 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1482 uint64_t UndefElts3;
1483 TmpV = SimplifyDemandedVectorElts(I->getOperand(1), RightDemanded,
1484 UndefElts3, Depth+1);
1485 if (TmpV) { I->setOperand(1, TmpV); MadeChange = true; }
1487 bool NewUndefElts = false;
1488 for (unsigned i = 0; i < VWidth; i++) {
1489 unsigned MaskVal = Shuffle->getMaskValue(i);
1490 if (MaskVal == -1u) {
1491 uint64_t NewBit = 1ULL << i;
1492 UndefElts |= NewBit;
1493 } else if (MaskVal < VWidth) {
1494 uint64_t NewBit = ((UndefElts2 >> MaskVal) & 1) << i;
1495 NewUndefElts |= NewBit;
1496 UndefElts |= NewBit;
1498 uint64_t NewBit = ((UndefElts3 >> (MaskVal - VWidth)) & 1) << i;
1499 NewUndefElts |= NewBit;
1500 UndefElts |= NewBit;
1505 // Add additional discovered undefs.
1506 std::vector<Constant*> Elts;
1507 for (unsigned i = 0; i < VWidth; ++i) {
1508 if (UndefElts & (1ULL << i))
1509 Elts.push_back(UndefValue::get(Type::Int32Ty));
1511 Elts.push_back(ConstantInt::get(Type::Int32Ty,
1512 Shuffle->getMaskValue(i)));
1514 I->setOperand(2, ConstantVector::get(Elts));
1519 case Instruction::BitCast: {
1520 // Vector->vector casts only.
1521 const VectorType *VTy = dyn_cast<VectorType>(I->getOperand(0)->getType());
1523 unsigned InVWidth = VTy->getNumElements();
1524 uint64_t InputDemandedElts = 0;
1527 if (VWidth == InVWidth) {
1528 // If we are converting from <4 x i32> -> <4 x f32>, we demand the same
1529 // elements as are demanded of us.
1531 InputDemandedElts = DemandedElts;
1532 } else if (VWidth > InVWidth) {
1536 // If there are more elements in the result than there are in the source,
1537 // then an input element is live if any of the corresponding output
1538 // elements are live.
1539 Ratio = VWidth/InVWidth;
1540 for (unsigned OutIdx = 0; OutIdx != VWidth; ++OutIdx) {
1541 if (DemandedElts & (1ULL << OutIdx))
1542 InputDemandedElts |= 1ULL << (OutIdx/Ratio);
1548 // If there are more elements in the source than there are in the result,
1549 // then an input element is live if the corresponding output element is
1551 Ratio = InVWidth/VWidth;
1552 for (unsigned InIdx = 0; InIdx != InVWidth; ++InIdx)
1553 if (DemandedElts & (1ULL << InIdx/Ratio))
1554 InputDemandedElts |= 1ULL << InIdx;
1557 // div/rem demand all inputs, because they don't want divide by zero.
1558 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), InputDemandedElts,
1559 UndefElts2, Depth+1);
1561 I->setOperand(0, TmpV);
1565 UndefElts = UndefElts2;
1566 if (VWidth > InVWidth) {
1567 assert(0 && "Unimp");
1568 // If there are more elements in the result than there are in the source,
1569 // then an output element is undef if the corresponding input element is
1571 for (unsigned OutIdx = 0; OutIdx != VWidth; ++OutIdx)
1572 if (UndefElts2 & (1ULL << (OutIdx/Ratio)))
1573 UndefElts |= 1ULL << OutIdx;
1574 } else if (VWidth < InVWidth) {
1575 assert(0 && "Unimp");
1576 // If there are more elements in the source than there are in the result,
1577 // then a result element is undef if all of the corresponding input
1578 // elements are undef.
1579 UndefElts = ~0ULL >> (64-VWidth); // Start out all undef.
1580 for (unsigned InIdx = 0; InIdx != InVWidth; ++InIdx)
1581 if ((UndefElts2 & (1ULL << InIdx)) == 0) // Not undef?
1582 UndefElts &= ~(1ULL << (InIdx/Ratio)); // Clear undef bit.
1586 case Instruction::And:
1587 case Instruction::Or:
1588 case Instruction::Xor:
1589 case Instruction::Add:
1590 case Instruction::Sub:
1591 case Instruction::Mul:
1592 // div/rem demand all inputs, because they don't want divide by zero.
1593 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
1594 UndefElts, Depth+1);
1595 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1596 TmpV = SimplifyDemandedVectorElts(I->getOperand(1), DemandedElts,
1597 UndefElts2, Depth+1);
1598 if (TmpV) { I->setOperand(1, TmpV); MadeChange = true; }
1600 // Output elements are undefined if both are undefined. Consider things
1601 // like undef&0. The result is known zero, not undef.
1602 UndefElts &= UndefElts2;
1605 case Instruction::Call: {
1606 IntrinsicInst *II = dyn_cast<IntrinsicInst>(I);
1608 switch (II->getIntrinsicID()) {
1611 // Binary vector operations that work column-wise. A dest element is a
1612 // function of the corresponding input elements from the two inputs.
1613 case Intrinsic::x86_sse_sub_ss:
1614 case Intrinsic::x86_sse_mul_ss:
1615 case Intrinsic::x86_sse_min_ss:
1616 case Intrinsic::x86_sse_max_ss:
1617 case Intrinsic::x86_sse2_sub_sd:
1618 case Intrinsic::x86_sse2_mul_sd:
1619 case Intrinsic::x86_sse2_min_sd:
1620 case Intrinsic::x86_sse2_max_sd:
1621 TmpV = SimplifyDemandedVectorElts(II->getOperand(1), DemandedElts,
1622 UndefElts, Depth+1);
1623 if (TmpV) { II->setOperand(1, TmpV); MadeChange = true; }
1624 TmpV = SimplifyDemandedVectorElts(II->getOperand(2), DemandedElts,
1625 UndefElts2, Depth+1);
1626 if (TmpV) { II->setOperand(2, TmpV); MadeChange = true; }
1628 // If only the low elt is demanded and this is a scalarizable intrinsic,
1629 // scalarize it now.
1630 if (DemandedElts == 1) {
1631 switch (II->getIntrinsicID()) {
1633 case Intrinsic::x86_sse_sub_ss:
1634 case Intrinsic::x86_sse_mul_ss:
1635 case Intrinsic::x86_sse2_sub_sd:
1636 case Intrinsic::x86_sse2_mul_sd:
1637 // TODO: Lower MIN/MAX/ABS/etc
1638 Value *LHS = II->getOperand(1);
1639 Value *RHS = II->getOperand(2);
1640 // Extract the element as scalars.
1641 LHS = InsertNewInstBefore(new ExtractElementInst(LHS, 0U,"tmp"), *II);
1642 RHS = InsertNewInstBefore(new ExtractElementInst(RHS, 0U,"tmp"), *II);
1644 switch (II->getIntrinsicID()) {
1645 default: assert(0 && "Case stmts out of sync!");
1646 case Intrinsic::x86_sse_sub_ss:
1647 case Intrinsic::x86_sse2_sub_sd:
1648 TmpV = InsertNewInstBefore(BinaryOperator::CreateSub(LHS, RHS,
1649 II->getName()), *II);
1651 case Intrinsic::x86_sse_mul_ss:
1652 case Intrinsic::x86_sse2_mul_sd:
1653 TmpV = InsertNewInstBefore(BinaryOperator::CreateMul(LHS, RHS,
1654 II->getName()), *II);
1659 InsertElementInst::Create(UndefValue::get(II->getType()), TmpV, 0U,
1661 InsertNewInstBefore(New, *II);
1662 AddSoonDeadInstToWorklist(*II, 0);
1667 // Output elements are undefined if both are undefined. Consider things
1668 // like undef&0. The result is known zero, not undef.
1669 UndefElts &= UndefElts2;
1675 return MadeChange ? I : 0;
1679 /// AssociativeOpt - Perform an optimization on an associative operator. This
1680 /// function is designed to check a chain of associative operators for a
1681 /// potential to apply a certain optimization. Since the optimization may be
1682 /// applicable if the expression was reassociated, this checks the chain, then
1683 /// reassociates the expression as necessary to expose the optimization
1684 /// opportunity. This makes use of a special Functor, which must define
1685 /// 'shouldApply' and 'apply' methods.
1687 template<typename Functor>
1688 static Instruction *AssociativeOpt(BinaryOperator &Root, const Functor &F) {
1689 unsigned Opcode = Root.getOpcode();
1690 Value *LHS = Root.getOperand(0);
1692 // Quick check, see if the immediate LHS matches...
1693 if (F.shouldApply(LHS))
1694 return F.apply(Root);
1696 // Otherwise, if the LHS is not of the same opcode as the root, return.
1697 Instruction *LHSI = dyn_cast<Instruction>(LHS);
1698 while (LHSI && LHSI->getOpcode() == Opcode && LHSI->hasOneUse()) {
1699 // Should we apply this transform to the RHS?
1700 bool ShouldApply = F.shouldApply(LHSI->getOperand(1));
1702 // If not to the RHS, check to see if we should apply to the LHS...
1703 if (!ShouldApply && F.shouldApply(LHSI->getOperand(0))) {
1704 cast<BinaryOperator>(LHSI)->swapOperands(); // Make the LHS the RHS
1708 // If the functor wants to apply the optimization to the RHS of LHSI,
1709 // reassociate the expression from ((? op A) op B) to (? op (A op B))
1711 // Now all of the instructions are in the current basic block, go ahead
1712 // and perform the reassociation.
1713 Instruction *TmpLHSI = cast<Instruction>(Root.getOperand(0));
1715 // First move the selected RHS to the LHS of the root...
1716 Root.setOperand(0, LHSI->getOperand(1));
1718 // Make what used to be the LHS of the root be the user of the root...
1719 Value *ExtraOperand = TmpLHSI->getOperand(1);
1720 if (&Root == TmpLHSI) {
1721 Root.replaceAllUsesWith(Constant::getNullValue(TmpLHSI->getType()));
1724 Root.replaceAllUsesWith(TmpLHSI); // Users now use TmpLHSI
1725 TmpLHSI->setOperand(1, &Root); // TmpLHSI now uses the root
1726 BasicBlock::iterator ARI = &Root; ++ARI;
1727 TmpLHSI->moveBefore(ARI); // Move TmpLHSI to after Root
1730 // Now propagate the ExtraOperand down the chain of instructions until we
1732 while (TmpLHSI != LHSI) {
1733 Instruction *NextLHSI = cast<Instruction>(TmpLHSI->getOperand(0));
1734 // Move the instruction to immediately before the chain we are
1735 // constructing to avoid breaking dominance properties.
1736 NextLHSI->moveBefore(ARI);
1739 Value *NextOp = NextLHSI->getOperand(1);
1740 NextLHSI->setOperand(1, ExtraOperand);
1742 ExtraOperand = NextOp;
1745 // Now that the instructions are reassociated, have the functor perform
1746 // the transformation...
1747 return F.apply(Root);
1750 LHSI = dyn_cast<Instruction>(LHSI->getOperand(0));
1757 // AddRHS - Implements: X + X --> X << 1
1760 AddRHS(Value *rhs) : RHS(rhs) {}
1761 bool shouldApply(Value *LHS) const { return LHS == RHS; }
1762 Instruction *apply(BinaryOperator &Add) const {
1763 return BinaryOperator::CreateShl(Add.getOperand(0),
1764 ConstantInt::get(Add.getType(), 1));
1768 // AddMaskingAnd - Implements (A & C1)+(B & C2) --> (A & C1)|(B & C2)
1770 struct AddMaskingAnd {
1772 AddMaskingAnd(Constant *c) : C2(c) {}
1773 bool shouldApply(Value *LHS) const {
1775 return match(LHS, m_And(m_Value(), m_ConstantInt(C1))) &&
1776 ConstantExpr::getAnd(C1, C2)->isNullValue();
1778 Instruction *apply(BinaryOperator &Add) const {
1779 return BinaryOperator::CreateOr(Add.getOperand(0), Add.getOperand(1));
1785 static Value *FoldOperationIntoSelectOperand(Instruction &I, Value *SO,
1787 if (CastInst *CI = dyn_cast<CastInst>(&I)) {
1788 if (Constant *SOC = dyn_cast<Constant>(SO))
1789 return ConstantExpr::getCast(CI->getOpcode(), SOC, I.getType());
1791 return IC->InsertNewInstBefore(CastInst::Create(
1792 CI->getOpcode(), SO, I.getType(), SO->getName() + ".cast"), I);
1795 // Figure out if the constant is the left or the right argument.
1796 bool ConstIsRHS = isa<Constant>(I.getOperand(1));
1797 Constant *ConstOperand = cast<Constant>(I.getOperand(ConstIsRHS));
1799 if (Constant *SOC = dyn_cast<Constant>(SO)) {
1801 return ConstantExpr::get(I.getOpcode(), SOC, ConstOperand);
1802 return ConstantExpr::get(I.getOpcode(), ConstOperand, SOC);
1805 Value *Op0 = SO, *Op1 = ConstOperand;
1807 std::swap(Op0, Op1);
1809 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
1810 New = BinaryOperator::Create(BO->getOpcode(), Op0, Op1,SO->getName()+".op");
1811 else if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1812 New = CmpInst::Create(CI->getOpcode(), CI->getPredicate(), Op0, Op1,
1813 SO->getName()+".cmp");
1815 assert(0 && "Unknown binary instruction type!");
1818 return IC->InsertNewInstBefore(New, I);
1821 // FoldOpIntoSelect - Given an instruction with a select as one operand and a
1822 // constant as the other operand, try to fold the binary operator into the
1823 // select arguments. This also works for Cast instructions, which obviously do
1824 // not have a second operand.
1825 static Instruction *FoldOpIntoSelect(Instruction &Op, SelectInst *SI,
1827 // Don't modify shared select instructions
1828 if (!SI->hasOneUse()) return 0;
1829 Value *TV = SI->getOperand(1);
1830 Value *FV = SI->getOperand(2);
1832 if (isa<Constant>(TV) || isa<Constant>(FV)) {
1833 // Bool selects with constant operands can be folded to logical ops.
1834 if (SI->getType() == Type::Int1Ty) return 0;
1836 Value *SelectTrueVal = FoldOperationIntoSelectOperand(Op, TV, IC);
1837 Value *SelectFalseVal = FoldOperationIntoSelectOperand(Op, FV, IC);
1839 return SelectInst::Create(SI->getCondition(), SelectTrueVal,
1846 /// FoldOpIntoPhi - Given a binary operator or cast instruction which has a PHI
1847 /// node as operand #0, see if we can fold the instruction into the PHI (which
1848 /// is only possible if all operands to the PHI are constants).
1849 Instruction *InstCombiner::FoldOpIntoPhi(Instruction &I) {
1850 PHINode *PN = cast<PHINode>(I.getOperand(0));
1851 unsigned NumPHIValues = PN->getNumIncomingValues();
1852 if (!PN->hasOneUse() || NumPHIValues == 0) return 0;
1854 // Check to see if all of the operands of the PHI are constants. If there is
1855 // one non-constant value, remember the BB it is. If there is more than one
1856 // or if *it* is a PHI, bail out.
1857 BasicBlock *NonConstBB = 0;
1858 for (unsigned i = 0; i != NumPHIValues; ++i)
1859 if (!isa<Constant>(PN->getIncomingValue(i))) {
1860 if (NonConstBB) return 0; // More than one non-const value.
1861 if (isa<PHINode>(PN->getIncomingValue(i))) return 0; // Itself a phi.
1862 NonConstBB = PN->getIncomingBlock(i);
1864 // If the incoming non-constant value is in I's block, we have an infinite
1866 if (NonConstBB == I.getParent())
1870 // If there is exactly one non-constant value, we can insert a copy of the
1871 // operation in that block. However, if this is a critical edge, we would be
1872 // inserting the computation one some other paths (e.g. inside a loop). Only
1873 // do this if the pred block is unconditionally branching into the phi block.
1875 BranchInst *BI = dyn_cast<BranchInst>(NonConstBB->getTerminator());
1876 if (!BI || !BI->isUnconditional()) return 0;
1879 // Okay, we can do the transformation: create the new PHI node.
1880 PHINode *NewPN = PHINode::Create(I.getType(), "");
1881 NewPN->reserveOperandSpace(PN->getNumOperands()/2);
1882 InsertNewInstBefore(NewPN, *PN);
1883 NewPN->takeName(PN);
1885 // Next, add all of the operands to the PHI.
1886 if (I.getNumOperands() == 2) {
1887 Constant *C = cast<Constant>(I.getOperand(1));
1888 for (unsigned i = 0; i != NumPHIValues; ++i) {
1890 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
1891 if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1892 InV = ConstantExpr::getCompare(CI->getPredicate(), InC, C);
1894 InV = ConstantExpr::get(I.getOpcode(), InC, C);
1896 assert(PN->getIncomingBlock(i) == NonConstBB);
1897 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
1898 InV = BinaryOperator::Create(BO->getOpcode(),
1899 PN->getIncomingValue(i), C, "phitmp",
1900 NonConstBB->getTerminator());
1901 else if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1902 InV = CmpInst::Create(CI->getOpcode(),
1904 PN->getIncomingValue(i), C, "phitmp",
1905 NonConstBB->getTerminator());
1907 assert(0 && "Unknown binop!");
1909 AddToWorkList(cast<Instruction>(InV));
1911 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
1914 CastInst *CI = cast<CastInst>(&I);
1915 const Type *RetTy = CI->getType();
1916 for (unsigned i = 0; i != NumPHIValues; ++i) {
1918 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
1919 InV = ConstantExpr::getCast(CI->getOpcode(), InC, RetTy);
1921 assert(PN->getIncomingBlock(i) == NonConstBB);
1922 InV = CastInst::Create(CI->getOpcode(), PN->getIncomingValue(i),
1923 I.getType(), "phitmp",
1924 NonConstBB->getTerminator());
1925 AddToWorkList(cast<Instruction>(InV));
1927 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
1930 return ReplaceInstUsesWith(I, NewPN);
1934 /// WillNotOverflowSignedAdd - Return true if we can prove that:
1935 /// (sext (add LHS, RHS)) === (add (sext LHS), (sext RHS))
1936 /// This basically requires proving that the add in the original type would not
1937 /// overflow to change the sign bit or have a carry out.
1938 bool InstCombiner::WillNotOverflowSignedAdd(Value *LHS, Value *RHS) {
1939 // There are different heuristics we can use for this. Here are some simple
1942 // Add has the property that adding any two 2's complement numbers can only
1943 // have one carry bit which can change a sign. As such, if LHS and RHS each
1944 // have at least two sign bits, we know that the addition of the two values will
1945 // sign extend fine.
1946 if (ComputeNumSignBits(LHS) > 1 && ComputeNumSignBits(RHS) > 1)
1950 // If one of the operands only has one non-zero bit, and if the other operand
1951 // has a known-zero bit in a more significant place than it (not including the
1952 // sign bit) the ripple may go up to and fill the zero, but won't change the
1953 // sign. For example, (X & ~4) + 1.
1961 Instruction *InstCombiner::visitAdd(BinaryOperator &I) {
1962 bool Changed = SimplifyCommutative(I);
1963 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
1965 if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
1966 // X + undef -> undef
1967 if (isa<UndefValue>(RHS))
1968 return ReplaceInstUsesWith(I, RHS);
1971 if (!I.getType()->isFPOrFPVector()) { // NOTE: -0 + +0 = +0.
1972 if (RHSC->isNullValue())
1973 return ReplaceInstUsesWith(I, LHS);
1974 } else if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHSC)) {
1975 if (CFP->isExactlyValue(ConstantFP::getNegativeZero
1976 (I.getType())->getValueAPF()))
1977 return ReplaceInstUsesWith(I, LHS);
1980 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHSC)) {
1981 // X + (signbit) --> X ^ signbit
1982 const APInt& Val = CI->getValue();
1983 uint32_t BitWidth = Val.getBitWidth();
1984 if (Val == APInt::getSignBit(BitWidth))
1985 return BinaryOperator::CreateXor(LHS, RHS);
1987 // See if SimplifyDemandedBits can simplify this. This handles stuff like
1988 // (X & 254)+1 -> (X&254)|1
1989 if (!isa<VectorType>(I.getType())) {
1990 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
1991 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
1992 KnownZero, KnownOne))
1997 if (isa<PHINode>(LHS))
1998 if (Instruction *NV = FoldOpIntoPhi(I))
2001 ConstantInt *XorRHS = 0;
2003 if (isa<ConstantInt>(RHSC) &&
2004 match(LHS, m_Xor(m_Value(XorLHS), m_ConstantInt(XorRHS)))) {
2005 uint32_t TySizeBits = I.getType()->getPrimitiveSizeInBits();
2006 const APInt& RHSVal = cast<ConstantInt>(RHSC)->getValue();
2008 uint32_t Size = TySizeBits / 2;
2009 APInt C0080Val(APInt(TySizeBits, 1ULL).shl(Size - 1));
2010 APInt CFF80Val(-C0080Val);
2012 if (TySizeBits > Size) {
2013 // If we have ADD(XOR(AND(X, 0xFF), 0x80), 0xF..F80), it's a sext.
2014 // If we have ADD(XOR(AND(X, 0xFF), 0xF..F80), 0x80), it's a sext.
2015 if ((RHSVal == CFF80Val && XorRHS->getValue() == C0080Val) ||
2016 (RHSVal == C0080Val && XorRHS->getValue() == CFF80Val)) {
2017 // This is a sign extend if the top bits are known zero.
2018 if (!MaskedValueIsZero(XorLHS,
2019 APInt::getHighBitsSet(TySizeBits, TySizeBits - Size)))
2020 Size = 0; // Not a sign ext, but can't be any others either.
2025 C0080Val = APIntOps::lshr(C0080Val, Size);
2026 CFF80Val = APIntOps::ashr(CFF80Val, Size);
2027 } while (Size >= 1);
2029 // FIXME: This shouldn't be necessary. When the backends can handle types
2030 // with funny bit widths then this switch statement should be removed. It
2031 // is just here to get the size of the "middle" type back up to something
2032 // that the back ends can handle.
2033 const Type *MiddleType = 0;
2036 case 32: MiddleType = Type::Int32Ty; break;
2037 case 16: MiddleType = Type::Int16Ty; break;
2038 case 8: MiddleType = Type::Int8Ty; break;
2041 Instruction *NewTrunc = new TruncInst(XorLHS, MiddleType, "sext");
2042 InsertNewInstBefore(NewTrunc, I);
2043 return new SExtInst(NewTrunc, I.getType(), I.getName());
2048 if (I.getType() == Type::Int1Ty)
2049 return BinaryOperator::CreateXor(LHS, RHS);
2052 if (I.getType()->isInteger()) {
2053 if (Instruction *Result = AssociativeOpt(I, AddRHS(RHS))) return Result;
2055 if (Instruction *RHSI = dyn_cast<Instruction>(RHS)) {
2056 if (RHSI->getOpcode() == Instruction::Sub)
2057 if (LHS == RHSI->getOperand(1)) // A + (B - A) --> B
2058 return ReplaceInstUsesWith(I, RHSI->getOperand(0));
2060 if (Instruction *LHSI = dyn_cast<Instruction>(LHS)) {
2061 if (LHSI->getOpcode() == Instruction::Sub)
2062 if (RHS == LHSI->getOperand(1)) // (B - A) + A --> B
2063 return ReplaceInstUsesWith(I, LHSI->getOperand(0));
2068 // -A + -B --> -(A + B)
2069 if (Value *LHSV = dyn_castNegVal(LHS)) {
2070 if (LHS->getType()->isIntOrIntVector()) {
2071 if (Value *RHSV = dyn_castNegVal(RHS)) {
2072 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSV, RHSV, "sum");
2073 InsertNewInstBefore(NewAdd, I);
2074 return BinaryOperator::CreateNeg(NewAdd);
2078 return BinaryOperator::CreateSub(RHS, LHSV);
2082 if (!isa<Constant>(RHS))
2083 if (Value *V = dyn_castNegVal(RHS))
2084 return BinaryOperator::CreateSub(LHS, V);
2088 if (Value *X = dyn_castFoldableMul(LHS, C2)) {
2089 if (X == RHS) // X*C + X --> X * (C+1)
2090 return BinaryOperator::CreateMul(RHS, AddOne(C2));
2092 // X*C1 + X*C2 --> X * (C1+C2)
2094 if (X == dyn_castFoldableMul(RHS, C1))
2095 return BinaryOperator::CreateMul(X, Add(C1, C2));
2098 // X + X*C --> X * (C+1)
2099 if (dyn_castFoldableMul(RHS, C2) == LHS)
2100 return BinaryOperator::CreateMul(LHS, AddOne(C2));
2102 // X + ~X --> -1 since ~X = -X-1
2103 if (dyn_castNotVal(LHS) == RHS || dyn_castNotVal(RHS) == LHS)
2104 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
2107 // (A & C1)+(B & C2) --> (A & C1)|(B & C2) iff C1&C2 == 0
2108 if (match(RHS, m_And(m_Value(), m_ConstantInt(C2))))
2109 if (Instruction *R = AssociativeOpt(I, AddMaskingAnd(C2)))
2112 // A+B --> A|B iff A and B have no bits set in common.
2113 if (const IntegerType *IT = dyn_cast<IntegerType>(I.getType())) {
2114 APInt Mask = APInt::getAllOnesValue(IT->getBitWidth());
2115 APInt LHSKnownOne(IT->getBitWidth(), 0);
2116 APInt LHSKnownZero(IT->getBitWidth(), 0);
2117 ComputeMaskedBits(LHS, Mask, LHSKnownZero, LHSKnownOne);
2118 if (LHSKnownZero != 0) {
2119 APInt RHSKnownOne(IT->getBitWidth(), 0);
2120 APInt RHSKnownZero(IT->getBitWidth(), 0);
2121 ComputeMaskedBits(RHS, Mask, RHSKnownZero, RHSKnownOne);
2123 // No bits in common -> bitwise or.
2124 if ((LHSKnownZero|RHSKnownZero).isAllOnesValue())
2125 return BinaryOperator::CreateOr(LHS, RHS);
2129 // W*X + Y*Z --> W * (X+Z) iff W == Y
2130 if (I.getType()->isIntOrIntVector()) {
2131 Value *W, *X, *Y, *Z;
2132 if (match(LHS, m_Mul(m_Value(W), m_Value(X))) &&
2133 match(RHS, m_Mul(m_Value(Y), m_Value(Z)))) {
2137 } else if (Y == X) {
2139 } else if (X == Z) {
2146 Value *NewAdd = InsertNewInstBefore(BinaryOperator::CreateAdd(X, Z,
2147 LHS->getName()), I);
2148 return BinaryOperator::CreateMul(W, NewAdd);
2153 if (ConstantInt *CRHS = dyn_cast<ConstantInt>(RHS)) {
2155 if (match(LHS, m_Not(m_Value(X)))) // ~X + C --> (C-1) - X
2156 return BinaryOperator::CreateSub(SubOne(CRHS), X);
2158 // (X & FF00) + xx00 -> (X+xx00) & FF00
2159 if (LHS->hasOneUse() && match(LHS, m_And(m_Value(X), m_ConstantInt(C2)))) {
2160 Constant *Anded = And(CRHS, C2);
2161 if (Anded == CRHS) {
2162 // See if all bits from the first bit set in the Add RHS up are included
2163 // in the mask. First, get the rightmost bit.
2164 const APInt& AddRHSV = CRHS->getValue();
2166 // Form a mask of all bits from the lowest bit added through the top.
2167 APInt AddRHSHighBits(~((AddRHSV & -AddRHSV)-1));
2169 // See if the and mask includes all of these bits.
2170 APInt AddRHSHighBitsAnd(AddRHSHighBits & C2->getValue());
2172 if (AddRHSHighBits == AddRHSHighBitsAnd) {
2173 // Okay, the xform is safe. Insert the new add pronto.
2174 Value *NewAdd = InsertNewInstBefore(BinaryOperator::CreateAdd(X, CRHS,
2175 LHS->getName()), I);
2176 return BinaryOperator::CreateAnd(NewAdd, C2);
2181 // Try to fold constant add into select arguments.
2182 if (SelectInst *SI = dyn_cast<SelectInst>(LHS))
2183 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2187 // add (cast *A to intptrtype) B ->
2188 // cast (GEP (cast *A to sbyte*) B) --> intptrtype
2190 CastInst *CI = dyn_cast<CastInst>(LHS);
2193 CI = dyn_cast<CastInst>(RHS);
2196 if (CI && CI->getType()->isSized() &&
2197 (CI->getType()->getPrimitiveSizeInBits() ==
2198 TD->getIntPtrType()->getPrimitiveSizeInBits())
2199 && isa<PointerType>(CI->getOperand(0)->getType())) {
2201 cast<PointerType>(CI->getOperand(0)->getType())->getAddressSpace();
2202 Value *I2 = InsertBitCastBefore(CI->getOperand(0),
2203 PointerType::get(Type::Int8Ty, AS), I);
2204 I2 = InsertNewInstBefore(GetElementPtrInst::Create(I2, Other, "ctg2"), I);
2205 return new PtrToIntInst(I2, CI->getType());
2209 // add (select X 0 (sub n A)) A --> select X A n
2211 SelectInst *SI = dyn_cast<SelectInst>(LHS);
2214 SI = dyn_cast<SelectInst>(RHS);
2217 if (SI && SI->hasOneUse()) {
2218 Value *TV = SI->getTrueValue();
2219 Value *FV = SI->getFalseValue();
2222 // Can we fold the add into the argument of the select?
2223 // We check both true and false select arguments for a matching subtract.
2224 if (match(FV, m_Zero()) && match(TV, m_Sub(m_Value(N), m_Value(A))) &&
2225 A == Other) // Fold the add into the true select value.
2226 return SelectInst::Create(SI->getCondition(), N, A);
2227 if (match(TV, m_Zero()) && match(FV, m_Sub(m_Value(N), m_Value(A))) &&
2228 A == Other) // Fold the add into the false select value.
2229 return SelectInst::Create(SI->getCondition(), A, N);
2233 // Check for X+0.0. Simplify it to X if we know X is not -0.0.
2234 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHS))
2235 if (CFP->getValueAPF().isPosZero() && CannotBeNegativeZero(LHS))
2236 return ReplaceInstUsesWith(I, LHS);
2238 // Check for (add (sext x), y), see if we can merge this into an
2239 // integer add followed by a sext.
2240 if (SExtInst *LHSConv = dyn_cast<SExtInst>(LHS)) {
2241 // (add (sext x), cst) --> (sext (add x, cst'))
2242 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS)) {
2244 ConstantExpr::getTrunc(RHSC, LHSConv->getOperand(0)->getType());
2245 if (LHSConv->hasOneUse() &&
2246 ConstantExpr::getSExt(CI, I.getType()) == RHSC &&
2247 WillNotOverflowSignedAdd(LHSConv->getOperand(0), CI)) {
2248 // Insert the new, smaller add.
2249 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSConv->getOperand(0),
2251 InsertNewInstBefore(NewAdd, I);
2252 return new SExtInst(NewAdd, I.getType());
2256 // (add (sext x), (sext y)) --> (sext (add int x, y))
2257 if (SExtInst *RHSConv = dyn_cast<SExtInst>(RHS)) {
2258 // Only do this if x/y have the same type, if at last one of them has a
2259 // single use (so we don't increase the number of sexts), and if the
2260 // integer add will not overflow.
2261 if (LHSConv->getOperand(0)->getType()==RHSConv->getOperand(0)->getType()&&
2262 (LHSConv->hasOneUse() || RHSConv->hasOneUse()) &&
2263 WillNotOverflowSignedAdd(LHSConv->getOperand(0),
2264 RHSConv->getOperand(0))) {
2265 // Insert the new integer add.
2266 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSConv->getOperand(0),
2267 RHSConv->getOperand(0),
2269 InsertNewInstBefore(NewAdd, I);
2270 return new SExtInst(NewAdd, I.getType());
2275 // Check for (add double (sitofp x), y), see if we can merge this into an
2276 // integer add followed by a promotion.
2277 if (SIToFPInst *LHSConv = dyn_cast<SIToFPInst>(LHS)) {
2278 // (add double (sitofp x), fpcst) --> (sitofp (add int x, intcst))
2279 // ... if the constant fits in the integer value. This is useful for things
2280 // like (double)(x & 1234) + 4.0 -> (double)((X & 1234)+4) which no longer
2281 // requires a constant pool load, and generally allows the add to be better
2283 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHS)) {
2285 ConstantExpr::getFPToSI(CFP, LHSConv->getOperand(0)->getType());
2286 if (LHSConv->hasOneUse() &&
2287 ConstantExpr::getSIToFP(CI, I.getType()) == CFP &&
2288 WillNotOverflowSignedAdd(LHSConv->getOperand(0), CI)) {
2289 // Insert the new integer add.
2290 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSConv->getOperand(0),
2292 InsertNewInstBefore(NewAdd, I);
2293 return new SIToFPInst(NewAdd, I.getType());
2297 // (add double (sitofp x), (sitofp y)) --> (sitofp (add int x, y))
2298 if (SIToFPInst *RHSConv = dyn_cast<SIToFPInst>(RHS)) {
2299 // Only do this if x/y have the same type, if at last one of them has a
2300 // single use (so we don't increase the number of int->fp conversions),
2301 // and if the integer add will not overflow.
2302 if (LHSConv->getOperand(0)->getType()==RHSConv->getOperand(0)->getType()&&
2303 (LHSConv->hasOneUse() || RHSConv->hasOneUse()) &&
2304 WillNotOverflowSignedAdd(LHSConv->getOperand(0),
2305 RHSConv->getOperand(0))) {
2306 // Insert the new integer add.
2307 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSConv->getOperand(0),
2308 RHSConv->getOperand(0),
2310 InsertNewInstBefore(NewAdd, I);
2311 return new SIToFPInst(NewAdd, I.getType());
2316 return Changed ? &I : 0;
2319 Instruction *InstCombiner::visitSub(BinaryOperator &I) {
2320 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2322 if (Op0 == Op1 && // sub X, X -> 0
2323 !I.getType()->isFPOrFPVector())
2324 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2326 // If this is a 'B = x-(-A)', change to B = x+A...
2327 if (Value *V = dyn_castNegVal(Op1))
2328 return BinaryOperator::CreateAdd(Op0, V);
2330 if (isa<UndefValue>(Op0))
2331 return ReplaceInstUsesWith(I, Op0); // undef - X -> undef
2332 if (isa<UndefValue>(Op1))
2333 return ReplaceInstUsesWith(I, Op1); // X - undef -> undef
2335 if (ConstantInt *C = dyn_cast<ConstantInt>(Op0)) {
2336 // Replace (-1 - A) with (~A)...
2337 if (C->isAllOnesValue())
2338 return BinaryOperator::CreateNot(Op1);
2340 // C - ~X == X + (1+C)
2342 if (match(Op1, m_Not(m_Value(X))))
2343 return BinaryOperator::CreateAdd(X, AddOne(C));
2345 // -(X >>u 31) -> (X >>s 31)
2346 // -(X >>s 31) -> (X >>u 31)
2348 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op1)) {
2349 if (SI->getOpcode() == Instruction::LShr) {
2350 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
2351 // Check to see if we are shifting out everything but the sign bit.
2352 if (CU->getLimitedValue(SI->getType()->getPrimitiveSizeInBits()) ==
2353 SI->getType()->getPrimitiveSizeInBits()-1) {
2354 // Ok, the transformation is safe. Insert AShr.
2355 return BinaryOperator::Create(Instruction::AShr,
2356 SI->getOperand(0), CU, SI->getName());
2360 else if (SI->getOpcode() == Instruction::AShr) {
2361 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
2362 // Check to see if we are shifting out everything but the sign bit.
2363 if (CU->getLimitedValue(SI->getType()->getPrimitiveSizeInBits()) ==
2364 SI->getType()->getPrimitiveSizeInBits()-1) {
2365 // Ok, the transformation is safe. Insert LShr.
2366 return BinaryOperator::CreateLShr(
2367 SI->getOperand(0), CU, SI->getName());
2374 // Try to fold constant sub into select arguments.
2375 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
2376 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2379 if (isa<PHINode>(Op0))
2380 if (Instruction *NV = FoldOpIntoPhi(I))
2384 if (I.getType() == Type::Int1Ty)
2385 return BinaryOperator::CreateXor(Op0, Op1);
2387 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
2388 if (Op1I->getOpcode() == Instruction::Add &&
2389 !Op0->getType()->isFPOrFPVector()) {
2390 if (Op1I->getOperand(0) == Op0) // X-(X+Y) == -Y
2391 return BinaryOperator::CreateNeg(Op1I->getOperand(1), I.getName());
2392 else if (Op1I->getOperand(1) == Op0) // X-(Y+X) == -Y
2393 return BinaryOperator::CreateNeg(Op1I->getOperand(0), I.getName());
2394 else if (ConstantInt *CI1 = dyn_cast<ConstantInt>(I.getOperand(0))) {
2395 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(Op1I->getOperand(1)))
2396 // C1-(X+C2) --> (C1-C2)-X
2397 return BinaryOperator::CreateSub(Subtract(CI1, CI2),
2398 Op1I->getOperand(0));
2402 if (Op1I->hasOneUse()) {
2403 // Replace (x - (y - z)) with (x + (z - y)) if the (y - z) subexpression
2404 // is not used by anyone else...
2406 if (Op1I->getOpcode() == Instruction::Sub &&
2407 !Op1I->getType()->isFPOrFPVector()) {
2408 // Swap the two operands of the subexpr...
2409 Value *IIOp0 = Op1I->getOperand(0), *IIOp1 = Op1I->getOperand(1);
2410 Op1I->setOperand(0, IIOp1);
2411 Op1I->setOperand(1, IIOp0);
2413 // Create the new top level add instruction...
2414 return BinaryOperator::CreateAdd(Op0, Op1);
2417 // Replace (A - (A & B)) with (A & ~B) if this is the only use of (A&B)...
2419 if (Op1I->getOpcode() == Instruction::And &&
2420 (Op1I->getOperand(0) == Op0 || Op1I->getOperand(1) == Op0)) {
2421 Value *OtherOp = Op1I->getOperand(Op1I->getOperand(0) == Op0);
2424 InsertNewInstBefore(BinaryOperator::CreateNot(OtherOp, "B.not"), I);
2425 return BinaryOperator::CreateAnd(Op0, NewNot);
2428 // 0 - (X sdiv C) -> (X sdiv -C)
2429 if (Op1I->getOpcode() == Instruction::SDiv)
2430 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
2432 if (Constant *DivRHS = dyn_cast<Constant>(Op1I->getOperand(1)))
2433 return BinaryOperator::CreateSDiv(Op1I->getOperand(0),
2434 ConstantExpr::getNeg(DivRHS));
2436 // X - X*C --> X * (1-C)
2437 ConstantInt *C2 = 0;
2438 if (dyn_castFoldableMul(Op1I, C2) == Op0) {
2439 Constant *CP1 = Subtract(ConstantInt::get(I.getType(), 1), C2);
2440 return BinaryOperator::CreateMul(Op0, CP1);
2443 // X - ((X / Y) * Y) --> X % Y
2444 if (Op1I->getOpcode() == Instruction::Mul)
2445 if (Instruction *I = dyn_cast<Instruction>(Op1I->getOperand(0)))
2446 if (Op0 == I->getOperand(0) &&
2447 Op1I->getOperand(1) == I->getOperand(1)) {
2448 if (I->getOpcode() == Instruction::SDiv)
2449 return BinaryOperator::CreateSRem(Op0, Op1I->getOperand(1));
2450 if (I->getOpcode() == Instruction::UDiv)
2451 return BinaryOperator::CreateURem(Op0, Op1I->getOperand(1));
2456 if (!Op0->getType()->isFPOrFPVector())
2457 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
2458 if (Op0I->getOpcode() == Instruction::Add) {
2459 if (Op0I->getOperand(0) == Op1) // (Y+X)-Y == X
2460 return ReplaceInstUsesWith(I, Op0I->getOperand(1));
2461 else if (Op0I->getOperand(1) == Op1) // (X+Y)-Y == X
2462 return ReplaceInstUsesWith(I, Op0I->getOperand(0));
2463 } else if (Op0I->getOpcode() == Instruction::Sub) {
2464 if (Op0I->getOperand(0) == Op1) // (X-Y)-X == -Y
2465 return BinaryOperator::CreateNeg(Op0I->getOperand(1), I.getName());
2470 if (Value *X = dyn_castFoldableMul(Op0, C1)) {
2471 if (X == Op1) // X*C - X --> X * (C-1)
2472 return BinaryOperator::CreateMul(Op1, SubOne(C1));
2474 ConstantInt *C2; // X*C1 - X*C2 -> X * (C1-C2)
2475 if (X == dyn_castFoldableMul(Op1, C2))
2476 return BinaryOperator::CreateMul(X, Subtract(C1, C2));
2481 /// isSignBitCheck - Given an exploded icmp instruction, return true if the
2482 /// comparison only checks the sign bit. If it only checks the sign bit, set
2483 /// TrueIfSigned if the result of the comparison is true when the input value is
2485 static bool isSignBitCheck(ICmpInst::Predicate pred, ConstantInt *RHS,
2486 bool &TrueIfSigned) {
2488 case ICmpInst::ICMP_SLT: // True if LHS s< 0
2489 TrueIfSigned = true;
2490 return RHS->isZero();
2491 case ICmpInst::ICMP_SLE: // True if LHS s<= RHS and RHS == -1
2492 TrueIfSigned = true;
2493 return RHS->isAllOnesValue();
2494 case ICmpInst::ICMP_SGT: // True if LHS s> -1
2495 TrueIfSigned = false;
2496 return RHS->isAllOnesValue();
2497 case ICmpInst::ICMP_UGT:
2498 // True if LHS u> RHS and RHS == high-bit-mask - 1
2499 TrueIfSigned = true;
2500 return RHS->getValue() ==
2501 APInt::getSignedMaxValue(RHS->getType()->getPrimitiveSizeInBits());
2502 case ICmpInst::ICMP_UGE:
2503 // True if LHS u>= RHS and RHS == high-bit-mask (2^7, 2^15, 2^31, etc)
2504 TrueIfSigned = true;
2505 return RHS->getValue().isSignBit();
2511 Instruction *InstCombiner::visitMul(BinaryOperator &I) {
2512 bool Changed = SimplifyCommutative(I);
2513 Value *Op0 = I.getOperand(0);
2515 if (isa<UndefValue>(I.getOperand(1))) // undef * X -> 0
2516 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2518 // Simplify mul instructions with a constant RHS...
2519 if (Constant *Op1 = dyn_cast<Constant>(I.getOperand(1))) {
2520 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2522 // ((X << C1)*C2) == (X * (C2 << C1))
2523 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op0))
2524 if (SI->getOpcode() == Instruction::Shl)
2525 if (Constant *ShOp = dyn_cast<Constant>(SI->getOperand(1)))
2526 return BinaryOperator::CreateMul(SI->getOperand(0),
2527 ConstantExpr::getShl(CI, ShOp));
2530 return ReplaceInstUsesWith(I, Op1); // X * 0 == 0
2531 if (CI->equalsInt(1)) // X * 1 == X
2532 return ReplaceInstUsesWith(I, Op0);
2533 if (CI->isAllOnesValue()) // X * -1 == 0 - X
2534 return BinaryOperator::CreateNeg(Op0, I.getName());
2536 const APInt& Val = cast<ConstantInt>(CI)->getValue();
2537 if (Val.isPowerOf2()) { // Replace X*(2^C) with X << C
2538 return BinaryOperator::CreateShl(Op0,
2539 ConstantInt::get(Op0->getType(), Val.logBase2()));
2541 } else if (ConstantFP *Op1F = dyn_cast<ConstantFP>(Op1)) {
2542 if (Op1F->isNullValue())
2543 return ReplaceInstUsesWith(I, Op1);
2545 // "In IEEE floating point, x*1 is not equivalent to x for nans. However,
2546 // ANSI says we can drop signals, so we can do this anyway." (from GCC)
2547 if (Op1F->isExactlyValue(1.0))
2548 return ReplaceInstUsesWith(I, Op0); // Eliminate 'mul double %X, 1.0'
2549 } else if (isa<VectorType>(Op1->getType())) {
2550 if (isa<ConstantAggregateZero>(Op1))
2551 return ReplaceInstUsesWith(I, Op1);
2553 // As above, vector X*splat(1.0) -> X in all defined cases.
2554 if (ConstantVector *Op1V = dyn_cast<ConstantVector>(Op1))
2555 if (ConstantFP *F = dyn_cast_or_null<ConstantFP>(Op1V->getSplatValue()))
2556 if (F->isExactlyValue(1.0))
2557 return ReplaceInstUsesWith(I, Op0);
2560 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
2561 if (Op0I->getOpcode() == Instruction::Add && Op0I->hasOneUse() &&
2562 isa<ConstantInt>(Op0I->getOperand(1)) && isa<ConstantInt>(Op1)) {
2563 // Canonicalize (X+C1)*C2 -> X*C2+C1*C2.
2564 Instruction *Add = BinaryOperator::CreateMul(Op0I->getOperand(0),
2566 InsertNewInstBefore(Add, I);
2567 Value *C1C2 = ConstantExpr::getMul(Op1,
2568 cast<Constant>(Op0I->getOperand(1)));
2569 return BinaryOperator::CreateAdd(Add, C1C2);
2573 // Try to fold constant mul into select arguments.
2574 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2575 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2578 if (isa<PHINode>(Op0))
2579 if (Instruction *NV = FoldOpIntoPhi(I))
2583 if (Value *Op0v = dyn_castNegVal(Op0)) // -X * -Y = X*Y
2584 if (Value *Op1v = dyn_castNegVal(I.getOperand(1)))
2585 return BinaryOperator::CreateMul(Op0v, Op1v);
2587 if (I.getType() == Type::Int1Ty)
2588 return BinaryOperator::CreateAnd(Op0, I.getOperand(1));
2590 // If one of the operands of the multiply is a cast from a boolean value, then
2591 // we know the bool is either zero or one, so this is a 'masking' multiply.
2592 // See if we can simplify things based on how the boolean was originally
2594 CastInst *BoolCast = 0;
2595 if (ZExtInst *CI = dyn_cast<ZExtInst>(Op0))
2596 if (CI->getOperand(0)->getType() == Type::Int1Ty)
2599 if (ZExtInst *CI = dyn_cast<ZExtInst>(I.getOperand(1)))
2600 if (CI->getOperand(0)->getType() == Type::Int1Ty)
2603 if (ICmpInst *SCI = dyn_cast<ICmpInst>(BoolCast->getOperand(0))) {
2604 Value *SCIOp0 = SCI->getOperand(0), *SCIOp1 = SCI->getOperand(1);
2605 const Type *SCOpTy = SCIOp0->getType();
2608 // If the icmp is true iff the sign bit of X is set, then convert this
2609 // multiply into a shift/and combination.
2610 if (isa<ConstantInt>(SCIOp1) &&
2611 isSignBitCheck(SCI->getPredicate(), cast<ConstantInt>(SCIOp1), TIS) &&
2613 // Shift the X value right to turn it into "all signbits".
2614 Constant *Amt = ConstantInt::get(SCIOp0->getType(),
2615 SCOpTy->getPrimitiveSizeInBits()-1);
2617 InsertNewInstBefore(
2618 BinaryOperator::Create(Instruction::AShr, SCIOp0, Amt,
2619 BoolCast->getOperand(0)->getName()+
2622 // If the multiply type is not the same as the source type, sign extend
2623 // or truncate to the multiply type.
2624 if (I.getType() != V->getType()) {
2625 uint32_t SrcBits = V->getType()->getPrimitiveSizeInBits();
2626 uint32_t DstBits = I.getType()->getPrimitiveSizeInBits();
2627 Instruction::CastOps opcode =
2628 (SrcBits == DstBits ? Instruction::BitCast :
2629 (SrcBits < DstBits ? Instruction::SExt : Instruction::Trunc));
2630 V = InsertCastBefore(opcode, V, I.getType(), I);
2633 Value *OtherOp = Op0 == BoolCast ? I.getOperand(1) : Op0;
2634 return BinaryOperator::CreateAnd(V, OtherOp);
2639 return Changed ? &I : 0;
2642 /// SimplifyDivRemOfSelect - Try to fold a divide or remainder of a select
2644 bool InstCombiner::SimplifyDivRemOfSelect(BinaryOperator &I) {
2645 SelectInst *SI = cast<SelectInst>(I.getOperand(1));
2647 // div/rem X, (Cond ? 0 : Y) -> div/rem X, Y
2648 int NonNullOperand = -1;
2649 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(1)))
2650 if (ST->isNullValue())
2652 // div/rem X, (Cond ? Y : 0) -> div/rem X, Y
2653 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(2)))
2654 if (ST->isNullValue())
2657 if (NonNullOperand == -1)
2660 Value *SelectCond = SI->getOperand(0);
2662 // Change the div/rem to use 'Y' instead of the select.
2663 I.setOperand(1, SI->getOperand(NonNullOperand));
2665 // Okay, we know we replace the operand of the div/rem with 'Y' with no
2666 // problem. However, the select, or the condition of the select may have
2667 // multiple uses. Based on our knowledge that the operand must be non-zero,
2668 // propagate the known value for the select into other uses of it, and
2669 // propagate a known value of the condition into its other users.
2671 // If the select and condition only have a single use, don't bother with this,
2673 if (SI->use_empty() && SelectCond->hasOneUse())
2676 // Scan the current block backward, looking for other uses of SI.
2677 BasicBlock::iterator BBI = &I, BBFront = I.getParent()->begin();
2679 while (BBI != BBFront) {
2681 // If we found a call to a function, we can't assume it will return, so
2682 // information from below it cannot be propagated above it.
2683 if (isa<CallInst>(BBI) && !isa<IntrinsicInst>(BBI))
2686 // Replace uses of the select or its condition with the known values.
2687 for (Instruction::op_iterator I = BBI->op_begin(), E = BBI->op_end();
2690 *I = SI->getOperand(NonNullOperand);
2692 } else if (*I == SelectCond) {
2693 *I = NonNullOperand == 1 ? ConstantInt::getTrue() :
2694 ConstantInt::getFalse();
2699 // If we past the instruction, quit looking for it.
2702 if (&*BBI == SelectCond)
2705 // If we ran out of things to eliminate, break out of the loop.
2706 if (SelectCond == 0 && SI == 0)
2714 /// This function implements the transforms on div instructions that work
2715 /// regardless of the kind of div instruction it is (udiv, sdiv, or fdiv). It is
2716 /// used by the visitors to those instructions.
2717 /// @brief Transforms common to all three div instructions
2718 Instruction *InstCombiner::commonDivTransforms(BinaryOperator &I) {
2719 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2721 // undef / X -> 0 for integer.
2722 // undef / X -> undef for FP (the undef could be a snan).
2723 if (isa<UndefValue>(Op0)) {
2724 if (Op0->getType()->isFPOrFPVector())
2725 return ReplaceInstUsesWith(I, Op0);
2726 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2729 // X / undef -> undef
2730 if (isa<UndefValue>(Op1))
2731 return ReplaceInstUsesWith(I, Op1);
2736 /// This function implements the transforms common to both integer division
2737 /// instructions (udiv and sdiv). It is called by the visitors to those integer
2738 /// division instructions.
2739 /// @brief Common integer divide transforms
2740 Instruction *InstCombiner::commonIDivTransforms(BinaryOperator &I) {
2741 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2743 // (sdiv X, X) --> 1 (udiv X, X) --> 1
2745 if (const VectorType *Ty = dyn_cast<VectorType>(I.getType())) {
2746 ConstantInt *CI = ConstantInt::get(Ty->getElementType(), 1);
2747 std::vector<Constant*> Elts(Ty->getNumElements(), CI);
2748 return ReplaceInstUsesWith(I, ConstantVector::get(Elts));
2751 ConstantInt *CI = ConstantInt::get(I.getType(), 1);
2752 return ReplaceInstUsesWith(I, CI);
2755 if (Instruction *Common = commonDivTransforms(I))
2758 // Handle cases involving: [su]div X, (select Cond, Y, Z)
2759 // This does not apply for fdiv.
2760 if (isa<SelectInst>(Op1) && SimplifyDivRemOfSelect(I))
2763 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2765 if (RHS->equalsInt(1))
2766 return ReplaceInstUsesWith(I, Op0);
2768 // (X / C1) / C2 -> X / (C1*C2)
2769 if (Instruction *LHS = dyn_cast<Instruction>(Op0))
2770 if (Instruction::BinaryOps(LHS->getOpcode()) == I.getOpcode())
2771 if (ConstantInt *LHSRHS = dyn_cast<ConstantInt>(LHS->getOperand(1))) {
2772 if (MultiplyOverflows(RHS, LHSRHS, I.getOpcode()==Instruction::SDiv))
2773 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2775 return BinaryOperator::Create(I.getOpcode(), LHS->getOperand(0),
2776 Multiply(RHS, LHSRHS));
2779 if (!RHS->isZero()) { // avoid X udiv 0
2780 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2781 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2783 if (isa<PHINode>(Op0))
2784 if (Instruction *NV = FoldOpIntoPhi(I))
2789 // 0 / X == 0, we don't need to preserve faults!
2790 if (ConstantInt *LHS = dyn_cast<ConstantInt>(Op0))
2791 if (LHS->equalsInt(0))
2792 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2794 // It can't be division by zero, hence it must be division by one.
2795 if (I.getType() == Type::Int1Ty)
2796 return ReplaceInstUsesWith(I, Op0);
2801 Instruction *InstCombiner::visitUDiv(BinaryOperator &I) {
2802 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2804 // Handle the integer div common cases
2805 if (Instruction *Common = commonIDivTransforms(I))
2808 // X udiv C^2 -> X >> C
2809 // Check to see if this is an unsigned division with an exact power of 2,
2810 // if so, convert to a right shift.
2811 if (ConstantInt *C = dyn_cast<ConstantInt>(Op1)) {
2812 if (C->getValue().isPowerOf2()) // 0 not included in isPowerOf2
2813 return BinaryOperator::CreateLShr(Op0,
2814 ConstantInt::get(Op0->getType(), C->getValue().logBase2()));
2817 // X udiv (C1 << N), where C1 is "1<<C2" --> X >> (N+C2)
2818 if (BinaryOperator *RHSI = dyn_cast<BinaryOperator>(I.getOperand(1))) {
2819 if (RHSI->getOpcode() == Instruction::Shl &&
2820 isa<ConstantInt>(RHSI->getOperand(0))) {
2821 const APInt& C1 = cast<ConstantInt>(RHSI->getOperand(0))->getValue();
2822 if (C1.isPowerOf2()) {
2823 Value *N = RHSI->getOperand(1);
2824 const Type *NTy = N->getType();
2825 if (uint32_t C2 = C1.logBase2()) {
2826 Constant *C2V = ConstantInt::get(NTy, C2);
2827 N = InsertNewInstBefore(BinaryOperator::CreateAdd(N, C2V, "tmp"), I);
2829 return BinaryOperator::CreateLShr(Op0, N);
2834 // udiv X, (Select Cond, C1, C2) --> Select Cond, (shr X, C1), (shr X, C2)
2835 // where C1&C2 are powers of two.
2836 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
2837 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
2838 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
2839 const APInt &TVA = STO->getValue(), &FVA = SFO->getValue();
2840 if (TVA.isPowerOf2() && FVA.isPowerOf2()) {
2841 // Compute the shift amounts
2842 uint32_t TSA = TVA.logBase2(), FSA = FVA.logBase2();
2843 // Construct the "on true" case of the select
2844 Constant *TC = ConstantInt::get(Op0->getType(), TSA);
2845 Instruction *TSI = BinaryOperator::CreateLShr(
2846 Op0, TC, SI->getName()+".t");
2847 TSI = InsertNewInstBefore(TSI, I);
2849 // Construct the "on false" case of the select
2850 Constant *FC = ConstantInt::get(Op0->getType(), FSA);
2851 Instruction *FSI = BinaryOperator::CreateLShr(
2852 Op0, FC, SI->getName()+".f");
2853 FSI = InsertNewInstBefore(FSI, I);
2855 // construct the select instruction and return it.
2856 return SelectInst::Create(SI->getOperand(0), TSI, FSI, SI->getName());
2862 Instruction *InstCombiner::visitSDiv(BinaryOperator &I) {
2863 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2865 // Handle the integer div common cases
2866 if (Instruction *Common = commonIDivTransforms(I))
2869 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2871 if (RHS->isAllOnesValue())
2872 return BinaryOperator::CreateNeg(Op0);
2875 if (Value *LHSNeg = dyn_castNegVal(Op0))
2876 return BinaryOperator::CreateSDiv(LHSNeg, ConstantExpr::getNeg(RHS));
2879 // If the sign bits of both operands are zero (i.e. we can prove they are
2880 // unsigned inputs), turn this into a udiv.
2881 if (I.getType()->isInteger()) {
2882 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()));
2883 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
2884 // X sdiv Y -> X udiv Y, iff X and Y don't have sign bit set
2885 return BinaryOperator::CreateUDiv(Op0, Op1, I.getName());
2892 Instruction *InstCombiner::visitFDiv(BinaryOperator &I) {
2893 return commonDivTransforms(I);
2896 /// This function implements the transforms on rem instructions that work
2897 /// regardless of the kind of rem instruction it is (urem, srem, or frem). It
2898 /// is used by the visitors to those instructions.
2899 /// @brief Transforms common to all three rem instructions
2900 Instruction *InstCombiner::commonRemTransforms(BinaryOperator &I) {
2901 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2903 // 0 % X == 0 for integer, we don't need to preserve faults!
2904 if (Constant *LHS = dyn_cast<Constant>(Op0))
2905 if (LHS->isNullValue())
2906 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2908 if (isa<UndefValue>(Op0)) { // undef % X -> 0
2909 if (I.getType()->isFPOrFPVector())
2910 return ReplaceInstUsesWith(I, Op0); // X % undef -> undef (could be SNaN)
2911 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2913 if (isa<UndefValue>(Op1))
2914 return ReplaceInstUsesWith(I, Op1); // X % undef -> undef
2916 // Handle cases involving: rem X, (select Cond, Y, Z)
2917 if (isa<SelectInst>(Op1) && SimplifyDivRemOfSelect(I))
2923 /// This function implements the transforms common to both integer remainder
2924 /// instructions (urem and srem). It is called by the visitors to those integer
2925 /// remainder instructions.
2926 /// @brief Common integer remainder transforms
2927 Instruction *InstCombiner::commonIRemTransforms(BinaryOperator &I) {
2928 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2930 if (Instruction *common = commonRemTransforms(I))
2933 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2934 // X % 0 == undef, we don't need to preserve faults!
2935 if (RHS->equalsInt(0))
2936 return ReplaceInstUsesWith(I, UndefValue::get(I.getType()));
2938 if (RHS->equalsInt(1)) // X % 1 == 0
2939 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2941 if (Instruction *Op0I = dyn_cast<Instruction>(Op0)) {
2942 if (SelectInst *SI = dyn_cast<SelectInst>(Op0I)) {
2943 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2945 } else if (isa<PHINode>(Op0I)) {
2946 if (Instruction *NV = FoldOpIntoPhi(I))
2950 // See if we can fold away this rem instruction.
2951 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
2952 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
2953 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
2954 KnownZero, KnownOne))
2962 Instruction *InstCombiner::visitURem(BinaryOperator &I) {
2963 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2965 if (Instruction *common = commonIRemTransforms(I))
2968 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2969 // X urem C^2 -> X and C
2970 // Check to see if this is an unsigned remainder with an exact power of 2,
2971 // if so, convert to a bitwise and.
2972 if (ConstantInt *C = dyn_cast<ConstantInt>(RHS))
2973 if (C->getValue().isPowerOf2())
2974 return BinaryOperator::CreateAnd(Op0, SubOne(C));
2977 if (Instruction *RHSI = dyn_cast<Instruction>(I.getOperand(1))) {
2978 // Turn A % (C << N), where C is 2^k, into A & ((C << N)-1)
2979 if (RHSI->getOpcode() == Instruction::Shl &&
2980 isa<ConstantInt>(RHSI->getOperand(0))) {
2981 if (cast<ConstantInt>(RHSI->getOperand(0))->getValue().isPowerOf2()) {
2982 Constant *N1 = ConstantInt::getAllOnesValue(I.getType());
2983 Value *Add = InsertNewInstBefore(BinaryOperator::CreateAdd(RHSI, N1,
2985 return BinaryOperator::CreateAnd(Op0, Add);
2990 // urem X, (select Cond, 2^C1, 2^C2) --> select Cond, (and X, C1), (and X, C2)
2991 // where C1&C2 are powers of two.
2992 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2993 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
2994 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
2995 // STO == 0 and SFO == 0 handled above.
2996 if ((STO->getValue().isPowerOf2()) &&
2997 (SFO->getValue().isPowerOf2())) {
2998 Value *TrueAnd = InsertNewInstBefore(
2999 BinaryOperator::CreateAnd(Op0, SubOne(STO), SI->getName()+".t"), I);
3000 Value *FalseAnd = InsertNewInstBefore(
3001 BinaryOperator::CreateAnd(Op0, SubOne(SFO), SI->getName()+".f"), I);
3002 return SelectInst::Create(SI->getOperand(0), TrueAnd, FalseAnd);
3010 Instruction *InstCombiner::visitSRem(BinaryOperator &I) {
3011 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3013 // Handle the integer rem common cases
3014 if (Instruction *common = commonIRemTransforms(I))
3017 if (Value *RHSNeg = dyn_castNegVal(Op1))
3018 if (!isa<Constant>(RHSNeg) ||
3019 (isa<ConstantInt>(RHSNeg) &&
3020 cast<ConstantInt>(RHSNeg)->getValue().isStrictlyPositive())) {
3022 AddUsesToWorkList(I);
3023 I.setOperand(1, RHSNeg);
3027 // If the sign bits of both operands are zero (i.e. we can prove they are
3028 // unsigned inputs), turn this into a urem.
3029 if (I.getType()->isInteger()) {
3030 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()));
3031 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
3032 // X srem Y -> X urem Y, iff X and Y don't have sign bit set
3033 return BinaryOperator::CreateURem(Op0, Op1, I.getName());
3040 Instruction *InstCombiner::visitFRem(BinaryOperator &I) {
3041 return commonRemTransforms(I);
3044 // isOneBitSet - Return true if there is exactly one bit set in the specified
3046 static bool isOneBitSet(const ConstantInt *CI) {
3047 return CI->getValue().isPowerOf2();
3050 // isHighOnes - Return true if the constant is of the form 1+0+.
3051 // This is the same as lowones(~X).
3052 static bool isHighOnes(const ConstantInt *CI) {
3053 return (~CI->getValue() + 1).isPowerOf2();
3056 /// getICmpCode - Encode a icmp predicate into a three bit mask. These bits
3057 /// are carefully arranged to allow folding of expressions such as:
3059 /// (A < B) | (A > B) --> (A != B)
3061 /// Note that this is only valid if the first and second predicates have the
3062 /// same sign. Is illegal to do: (A u< B) | (A s> B)
3064 /// Three bits are used to represent the condition, as follows:
3069 /// <=> Value Definition
3070 /// 000 0 Always false
3077 /// 111 7 Always true
3079 static unsigned getICmpCode(const ICmpInst *ICI) {
3080 switch (ICI->getPredicate()) {
3082 case ICmpInst::ICMP_UGT: return 1; // 001
3083 case ICmpInst::ICMP_SGT: return 1; // 001
3084 case ICmpInst::ICMP_EQ: return 2; // 010
3085 case ICmpInst::ICMP_UGE: return 3; // 011
3086 case ICmpInst::ICMP_SGE: return 3; // 011
3087 case ICmpInst::ICMP_ULT: return 4; // 100
3088 case ICmpInst::ICMP_SLT: return 4; // 100
3089 case ICmpInst::ICMP_NE: return 5; // 101
3090 case ICmpInst::ICMP_ULE: return 6; // 110
3091 case ICmpInst::ICMP_SLE: return 6; // 110
3094 assert(0 && "Invalid ICmp predicate!");
3099 /// getICmpValue - This is the complement of getICmpCode, which turns an
3100 /// opcode and two operands into either a constant true or false, or a brand
3101 /// new ICmp instruction. The sign is passed in to determine which kind
3102 /// of predicate to use in new icmp instructions.
3103 static Value *getICmpValue(bool sign, unsigned code, Value *LHS, Value *RHS) {
3105 default: assert(0 && "Illegal ICmp code!");
3106 case 0: return ConstantInt::getFalse();
3109 return new ICmpInst(ICmpInst::ICMP_SGT, LHS, RHS);
3111 return new ICmpInst(ICmpInst::ICMP_UGT, LHS, RHS);
3112 case 2: return new ICmpInst(ICmpInst::ICMP_EQ, LHS, RHS);
3115 return new ICmpInst(ICmpInst::ICMP_SGE, LHS, RHS);
3117 return new ICmpInst(ICmpInst::ICMP_UGE, LHS, RHS);
3120 return new ICmpInst(ICmpInst::ICMP_SLT, LHS, RHS);
3122 return new ICmpInst(ICmpInst::ICMP_ULT, LHS, RHS);
3123 case 5: return new ICmpInst(ICmpInst::ICMP_NE, LHS, RHS);
3126 return new ICmpInst(ICmpInst::ICMP_SLE, LHS, RHS);
3128 return new ICmpInst(ICmpInst::ICMP_ULE, LHS, RHS);
3129 case 7: return ConstantInt::getTrue();
3133 static bool PredicatesFoldable(ICmpInst::Predicate p1, ICmpInst::Predicate p2) {
3134 return (ICmpInst::isSignedPredicate(p1) == ICmpInst::isSignedPredicate(p2)) ||
3135 (ICmpInst::isSignedPredicate(p1) &&
3136 (p2 == ICmpInst::ICMP_EQ || p2 == ICmpInst::ICMP_NE)) ||
3137 (ICmpInst::isSignedPredicate(p2) &&
3138 (p1 == ICmpInst::ICMP_EQ || p1 == ICmpInst::ICMP_NE));
3142 // FoldICmpLogical - Implements (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
3143 struct FoldICmpLogical {
3146 ICmpInst::Predicate pred;
3147 FoldICmpLogical(InstCombiner &ic, ICmpInst *ICI)
3148 : IC(ic), LHS(ICI->getOperand(0)), RHS(ICI->getOperand(1)),
3149 pred(ICI->getPredicate()) {}
3150 bool shouldApply(Value *V) const {
3151 if (ICmpInst *ICI = dyn_cast<ICmpInst>(V))
3152 if (PredicatesFoldable(pred, ICI->getPredicate()))
3153 return ((ICI->getOperand(0) == LHS && ICI->getOperand(1) == RHS) ||
3154 (ICI->getOperand(0) == RHS && ICI->getOperand(1) == LHS));
3157 Instruction *apply(Instruction &Log) const {
3158 ICmpInst *ICI = cast<ICmpInst>(Log.getOperand(0));
3159 if (ICI->getOperand(0) != LHS) {
3160 assert(ICI->getOperand(1) == LHS);
3161 ICI->swapOperands(); // Swap the LHS and RHS of the ICmp
3164 ICmpInst *RHSICI = cast<ICmpInst>(Log.getOperand(1));
3165 unsigned LHSCode = getICmpCode(ICI);
3166 unsigned RHSCode = getICmpCode(RHSICI);
3168 switch (Log.getOpcode()) {
3169 case Instruction::And: Code = LHSCode & RHSCode; break;
3170 case Instruction::Or: Code = LHSCode | RHSCode; break;
3171 case Instruction::Xor: Code = LHSCode ^ RHSCode; break;
3172 default: assert(0 && "Illegal logical opcode!"); return 0;
3175 bool isSigned = ICmpInst::isSignedPredicate(RHSICI->getPredicate()) ||
3176 ICmpInst::isSignedPredicate(ICI->getPredicate());
3178 Value *RV = getICmpValue(isSigned, Code, LHS, RHS);
3179 if (Instruction *I = dyn_cast<Instruction>(RV))
3181 // Otherwise, it's a constant boolean value...
3182 return IC.ReplaceInstUsesWith(Log, RV);
3185 } // end anonymous namespace
3187 // OptAndOp - This handles expressions of the form ((val OP C1) & C2). Where
3188 // the Op parameter is 'OP', OpRHS is 'C1', and AndRHS is 'C2'. Op is
3189 // guaranteed to be a binary operator.
3190 Instruction *InstCombiner::OptAndOp(Instruction *Op,
3192 ConstantInt *AndRHS,
3193 BinaryOperator &TheAnd) {
3194 Value *X = Op->getOperand(0);
3195 Constant *Together = 0;
3197 Together = And(AndRHS, OpRHS);
3199 switch (Op->getOpcode()) {
3200 case Instruction::Xor:
3201 if (Op->hasOneUse()) {
3202 // (X ^ C1) & C2 --> (X & C2) ^ (C1&C2)
3203 Instruction *And = BinaryOperator::CreateAnd(X, AndRHS);
3204 InsertNewInstBefore(And, TheAnd);
3206 return BinaryOperator::CreateXor(And, Together);
3209 case Instruction::Or:
3210 if (Together == AndRHS) // (X | C) & C --> C
3211 return ReplaceInstUsesWith(TheAnd, AndRHS);
3213 if (Op->hasOneUse() && Together != OpRHS) {
3214 // (X | C1) & C2 --> (X | (C1&C2)) & C2
3215 Instruction *Or = BinaryOperator::CreateOr(X, Together);
3216 InsertNewInstBefore(Or, TheAnd);
3218 return BinaryOperator::CreateAnd(Or, AndRHS);
3221 case Instruction::Add:
3222 if (Op->hasOneUse()) {
3223 // Adding a one to a single bit bit-field should be turned into an XOR
3224 // of the bit. First thing to check is to see if this AND is with a
3225 // single bit constant.
3226 const APInt& AndRHSV = cast<ConstantInt>(AndRHS)->getValue();
3228 // If there is only one bit set...
3229 if (isOneBitSet(cast<ConstantInt>(AndRHS))) {
3230 // Ok, at this point, we know that we are masking the result of the
3231 // ADD down to exactly one bit. If the constant we are adding has
3232 // no bits set below this bit, then we can eliminate the ADD.
3233 const APInt& AddRHS = cast<ConstantInt>(OpRHS)->getValue();
3235 // Check to see if any bits below the one bit set in AndRHSV are set.
3236 if ((AddRHS & (AndRHSV-1)) == 0) {
3237 // If not, the only thing that can effect the output of the AND is
3238 // the bit specified by AndRHSV. If that bit is set, the effect of
3239 // the XOR is to toggle the bit. If it is clear, then the ADD has
3241 if ((AddRHS & AndRHSV) == 0) { // Bit is not set, noop
3242 TheAnd.setOperand(0, X);
3245 // Pull the XOR out of the AND.
3246 Instruction *NewAnd = BinaryOperator::CreateAnd(X, AndRHS);
3247 InsertNewInstBefore(NewAnd, TheAnd);
3248 NewAnd->takeName(Op);
3249 return BinaryOperator::CreateXor(NewAnd, AndRHS);
3256 case Instruction::Shl: {
3257 // We know that the AND will not produce any of the bits shifted in, so if
3258 // the anded constant includes them, clear them now!
3260 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3261 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3262 APInt ShlMask(APInt::getHighBitsSet(BitWidth, BitWidth-OpRHSVal));
3263 ConstantInt *CI = ConstantInt::get(AndRHS->getValue() & ShlMask);
3265 if (CI->getValue() == ShlMask) {
3266 // Masking out bits that the shift already masks
3267 return ReplaceInstUsesWith(TheAnd, Op); // No need for the and.
3268 } else if (CI != AndRHS) { // Reducing bits set in and.
3269 TheAnd.setOperand(1, CI);
3274 case Instruction::LShr:
3276 // We know that the AND will not produce any of the bits shifted in, so if
3277 // the anded constant includes them, clear them now! This only applies to
3278 // unsigned shifts, because a signed shr may bring in set bits!
3280 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3281 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3282 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
3283 ConstantInt *CI = ConstantInt::get(AndRHS->getValue() & ShrMask);
3285 if (CI->getValue() == ShrMask) {
3286 // Masking out bits that the shift already masks.
3287 return ReplaceInstUsesWith(TheAnd, Op);
3288 } else if (CI != AndRHS) {
3289 TheAnd.setOperand(1, CI); // Reduce bits set in and cst.
3294 case Instruction::AShr:
3296 // See if this is shifting in some sign extension, then masking it out
3298 if (Op->hasOneUse()) {
3299 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3300 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3301 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
3302 Constant *C = ConstantInt::get(AndRHS->getValue() & ShrMask);
3303 if (C == AndRHS) { // Masking out bits shifted in.
3304 // (Val ashr C1) & C2 -> (Val lshr C1) & C2
3305 // Make the argument unsigned.
3306 Value *ShVal = Op->getOperand(0);
3307 ShVal = InsertNewInstBefore(
3308 BinaryOperator::CreateLShr(ShVal, OpRHS,
3309 Op->getName()), TheAnd);
3310 return BinaryOperator::CreateAnd(ShVal, AndRHS, TheAnd.getName());
3319 /// InsertRangeTest - Emit a computation of: (V >= Lo && V < Hi) if Inside is
3320 /// true, otherwise (V < Lo || V >= Hi). In pratice, we emit the more efficient
3321 /// (V-Lo) <u Hi-Lo. This method expects that Lo <= Hi. isSigned indicates
3322 /// whether to treat the V, Lo and HI as signed or not. IB is the location to
3323 /// insert new instructions.
3324 Instruction *InstCombiner::InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
3325 bool isSigned, bool Inside,
3327 assert(cast<ConstantInt>(ConstantExpr::getICmp((isSigned ?
3328 ICmpInst::ICMP_SLE:ICmpInst::ICMP_ULE), Lo, Hi))->getZExtValue() &&
3329 "Lo is not <= Hi in range emission code!");
3332 if (Lo == Hi) // Trivially false.
3333 return new ICmpInst(ICmpInst::ICMP_NE, V, V);
3335 // V >= Min && V < Hi --> V < Hi
3336 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
3337 ICmpInst::Predicate pred = (isSigned ?
3338 ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT);
3339 return new ICmpInst(pred, V, Hi);
3342 // Emit V-Lo <u Hi-Lo
3343 Constant *NegLo = ConstantExpr::getNeg(Lo);
3344 Instruction *Add = BinaryOperator::CreateAdd(V, NegLo, V->getName()+".off");
3345 InsertNewInstBefore(Add, IB);
3346 Constant *UpperBound = ConstantExpr::getAdd(NegLo, Hi);
3347 return new ICmpInst(ICmpInst::ICMP_ULT, Add, UpperBound);
3350 if (Lo == Hi) // Trivially true.
3351 return new ICmpInst(ICmpInst::ICMP_EQ, V, V);
3353 // V < Min || V >= Hi -> V > Hi-1
3354 Hi = SubOne(cast<ConstantInt>(Hi));
3355 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
3356 ICmpInst::Predicate pred = (isSigned ?
3357 ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT);
3358 return new ICmpInst(pred, V, Hi);
3361 // Emit V-Lo >u Hi-1-Lo
3362 // Note that Hi has already had one subtracted from it, above.
3363 ConstantInt *NegLo = cast<ConstantInt>(ConstantExpr::getNeg(Lo));
3364 Instruction *Add = BinaryOperator::CreateAdd(V, NegLo, V->getName()+".off");
3365 InsertNewInstBefore(Add, IB);
3366 Constant *LowerBound = ConstantExpr::getAdd(NegLo, Hi);
3367 return new ICmpInst(ICmpInst::ICMP_UGT, Add, LowerBound);
3370 // isRunOfOnes - Returns true iff Val consists of one contiguous run of 1s with
3371 // any number of 0s on either side. The 1s are allowed to wrap from LSB to
3372 // MSB, so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs. 0x0F0F0000 is
3373 // not, since all 1s are not contiguous.
3374 static bool isRunOfOnes(ConstantInt *Val, uint32_t &MB, uint32_t &ME) {
3375 const APInt& V = Val->getValue();
3376 uint32_t BitWidth = Val->getType()->getBitWidth();
3377 if (!APIntOps::isShiftedMask(BitWidth, V)) return false;
3379 // look for the first zero bit after the run of ones
3380 MB = BitWidth - ((V - 1) ^ V).countLeadingZeros();
3381 // look for the first non-zero bit
3382 ME = V.getActiveBits();
3386 /// FoldLogicalPlusAnd - This is part of an expression (LHS +/- RHS) & Mask,
3387 /// where isSub determines whether the operator is a sub. If we can fold one of
3388 /// the following xforms:
3390 /// ((A & N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == Mask
3391 /// ((A | N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3392 /// ((A ^ N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3394 /// return (A +/- B).
3396 Value *InstCombiner::FoldLogicalPlusAnd(Value *LHS, Value *RHS,
3397 ConstantInt *Mask, bool isSub,
3399 Instruction *LHSI = dyn_cast<Instruction>(LHS);
3400 if (!LHSI || LHSI->getNumOperands() != 2 ||
3401 !isa<ConstantInt>(LHSI->getOperand(1))) return 0;
3403 ConstantInt *N = cast<ConstantInt>(LHSI->getOperand(1));
3405 switch (LHSI->getOpcode()) {
3407 case Instruction::And:
3408 if (And(N, Mask) == Mask) {
3409 // If the AndRHS is a power of two minus one (0+1+), this is simple.
3410 if ((Mask->getValue().countLeadingZeros() +
3411 Mask->getValue().countPopulation()) ==
3412 Mask->getValue().getBitWidth())
3415 // Otherwise, if Mask is 0+1+0+, and if B is known to have the low 0+
3416 // part, we don't need any explicit masks to take them out of A. If that
3417 // is all N is, ignore it.
3418 uint32_t MB = 0, ME = 0;
3419 if (isRunOfOnes(Mask, MB, ME)) { // begin/end bit of run, inclusive
3420 uint32_t BitWidth = cast<IntegerType>(RHS->getType())->getBitWidth();
3421 APInt Mask(APInt::getLowBitsSet(BitWidth, MB-1));
3422 if (MaskedValueIsZero(RHS, Mask))
3427 case Instruction::Or:
3428 case Instruction::Xor:
3429 // If the AndRHS is a power of two minus one (0+1+), and N&Mask == 0
3430 if ((Mask->getValue().countLeadingZeros() +
3431 Mask->getValue().countPopulation()) == Mask->getValue().getBitWidth()
3432 && And(N, Mask)->isZero())
3439 New = BinaryOperator::CreateSub(LHSI->getOperand(0), RHS, "fold");
3441 New = BinaryOperator::CreateAdd(LHSI->getOperand(0), RHS, "fold");
3442 return InsertNewInstBefore(New, I);
3445 Instruction *InstCombiner::visitAnd(BinaryOperator &I) {
3446 bool Changed = SimplifyCommutative(I);
3447 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3449 if (isa<UndefValue>(Op1)) // X & undef -> 0
3450 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3454 return ReplaceInstUsesWith(I, Op1);
3456 // See if we can simplify any instructions used by the instruction whose sole
3457 // purpose is to compute bits we don't care about.
3458 if (!isa<VectorType>(I.getType())) {
3459 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
3460 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
3461 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
3462 KnownZero, KnownOne))
3465 if (ConstantVector *CP = dyn_cast<ConstantVector>(Op1)) {
3466 if (CP->isAllOnesValue()) // X & <-1,-1> -> X
3467 return ReplaceInstUsesWith(I, I.getOperand(0));
3468 } else if (isa<ConstantAggregateZero>(Op1)) {
3469 return ReplaceInstUsesWith(I, Op1); // X & <0,0> -> <0,0>
3473 if (ConstantInt *AndRHS = dyn_cast<ConstantInt>(Op1)) {
3474 const APInt& AndRHSMask = AndRHS->getValue();
3475 APInt NotAndRHS(~AndRHSMask);
3477 // Optimize a variety of ((val OP C1) & C2) combinations...
3478 if (isa<BinaryOperator>(Op0)) {
3479 Instruction *Op0I = cast<Instruction>(Op0);
3480 Value *Op0LHS = Op0I->getOperand(0);
3481 Value *Op0RHS = Op0I->getOperand(1);
3482 switch (Op0I->getOpcode()) {
3483 case Instruction::Xor:
3484 case Instruction::Or:
3485 // If the mask is only needed on one incoming arm, push it up.
3486 if (Op0I->hasOneUse()) {
3487 if (MaskedValueIsZero(Op0LHS, NotAndRHS)) {
3488 // Not masking anything out for the LHS, move to RHS.
3489 Instruction *NewRHS = BinaryOperator::CreateAnd(Op0RHS, AndRHS,
3490 Op0RHS->getName()+".masked");
3491 InsertNewInstBefore(NewRHS, I);
3492 return BinaryOperator::Create(
3493 cast<BinaryOperator>(Op0I)->getOpcode(), Op0LHS, NewRHS);
3495 if (!isa<Constant>(Op0RHS) &&
3496 MaskedValueIsZero(Op0RHS, NotAndRHS)) {
3497 // Not masking anything out for the RHS, move to LHS.
3498 Instruction *NewLHS = BinaryOperator::CreateAnd(Op0LHS, AndRHS,
3499 Op0LHS->getName()+".masked");
3500 InsertNewInstBefore(NewLHS, I);
3501 return BinaryOperator::Create(
3502 cast<BinaryOperator>(Op0I)->getOpcode(), NewLHS, Op0RHS);
3507 case Instruction::Add:
3508 // ((A & N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == AndRHS.
3509 // ((A | N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
3510 // ((A ^ N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
3511 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, false, I))
3512 return BinaryOperator::CreateAnd(V, AndRHS);
3513 if (Value *V = FoldLogicalPlusAnd(Op0RHS, Op0LHS, AndRHS, false, I))
3514 return BinaryOperator::CreateAnd(V, AndRHS); // Add commutes
3517 case Instruction::Sub:
3518 // ((A & N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == AndRHS.
3519 // ((A | N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
3520 // ((A ^ N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
3521 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, true, I))
3522 return BinaryOperator::CreateAnd(V, AndRHS);
3524 // (A - N) & AndRHS -> -N & AndRHS iff A&AndRHS==0 and AndRHS
3525 // has 1's for all bits that the subtraction with A might affect.
3526 if (Op0I->hasOneUse()) {
3527 uint32_t BitWidth = AndRHSMask.getBitWidth();
3528 uint32_t Zeros = AndRHSMask.countLeadingZeros();
3529 APInt Mask = APInt::getLowBitsSet(BitWidth, BitWidth - Zeros);
3531 ConstantInt *A = dyn_cast<ConstantInt>(Op0LHS);
3532 if (!(A && A->isZero()) && // avoid infinite recursion.
3533 MaskedValueIsZero(Op0LHS, Mask)) {
3534 Instruction *NewNeg = BinaryOperator::CreateNeg(Op0RHS);
3535 InsertNewInstBefore(NewNeg, I);
3536 return BinaryOperator::CreateAnd(NewNeg, AndRHS);
3541 case Instruction::Shl:
3542 case Instruction::LShr:
3543 // (1 << x) & 1 --> zext(x == 0)
3544 // (1 >> x) & 1 --> zext(x == 0)
3545 if (AndRHSMask == 1 && Op0LHS == AndRHS) {
3546 Instruction *NewICmp = new ICmpInst(ICmpInst::ICMP_EQ, Op0RHS,
3547 Constant::getNullValue(I.getType()));
3548 InsertNewInstBefore(NewICmp, I);
3549 return new ZExtInst(NewICmp, I.getType());
3554 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
3555 if (Instruction *Res = OptAndOp(Op0I, Op0CI, AndRHS, I))
3557 } else if (CastInst *CI = dyn_cast<CastInst>(Op0)) {
3558 // If this is an integer truncation or change from signed-to-unsigned, and
3559 // if the source is an and/or with immediate, transform it. This
3560 // frequently occurs for bitfield accesses.
3561 if (Instruction *CastOp = dyn_cast<Instruction>(CI->getOperand(0))) {
3562 if ((isa<TruncInst>(CI) || isa<BitCastInst>(CI)) &&
3563 CastOp->getNumOperands() == 2)
3564 if (ConstantInt *AndCI = dyn_cast<ConstantInt>(CastOp->getOperand(1))) {
3565 if (CastOp->getOpcode() == Instruction::And) {
3566 // Change: and (cast (and X, C1) to T), C2
3567 // into : and (cast X to T), trunc_or_bitcast(C1)&C2
3568 // This will fold the two constants together, which may allow
3569 // other simplifications.
3570 Instruction *NewCast = CastInst::CreateTruncOrBitCast(
3571 CastOp->getOperand(0), I.getType(),
3572 CastOp->getName()+".shrunk");
3573 NewCast = InsertNewInstBefore(NewCast, I);
3574 // trunc_or_bitcast(C1)&C2
3575 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
3576 C3 = ConstantExpr::getAnd(C3, AndRHS);
3577 return BinaryOperator::CreateAnd(NewCast, C3);
3578 } else if (CastOp->getOpcode() == Instruction::Or) {
3579 // Change: and (cast (or X, C1) to T), C2
3580 // into : trunc(C1)&C2 iff trunc(C1)&C2 == C2
3581 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
3582 if (ConstantExpr::getAnd(C3, AndRHS) == AndRHS) // trunc(C1)&C2
3583 return ReplaceInstUsesWith(I, AndRHS);
3589 // Try to fold constant and into select arguments.
3590 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3591 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3593 if (isa<PHINode>(Op0))
3594 if (Instruction *NV = FoldOpIntoPhi(I))
3598 Value *Op0NotVal = dyn_castNotVal(Op0);
3599 Value *Op1NotVal = dyn_castNotVal(Op1);
3601 if (Op0NotVal == Op1 || Op1NotVal == Op0) // A & ~A == ~A & A == 0
3602 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3604 // (~A & ~B) == (~(A | B)) - De Morgan's Law
3605 if (Op0NotVal && Op1NotVal && isOnlyUse(Op0) && isOnlyUse(Op1)) {
3606 Instruction *Or = BinaryOperator::CreateOr(Op0NotVal, Op1NotVal,
3607 I.getName()+".demorgan");
3608 InsertNewInstBefore(Or, I);
3609 return BinaryOperator::CreateNot(Or);
3613 Value *A = 0, *B = 0, *C = 0, *D = 0;
3614 if (match(Op0, m_Or(m_Value(A), m_Value(B)))) {
3615 if (A == Op1 || B == Op1) // (A | ?) & A --> A
3616 return ReplaceInstUsesWith(I, Op1);
3618 // (A|B) & ~(A&B) -> A^B
3619 if (match(Op1, m_Not(m_And(m_Value(C), m_Value(D))))) {
3620 if ((A == C && B == D) || (A == D && B == C))
3621 return BinaryOperator::CreateXor(A, B);
3625 if (match(Op1, m_Or(m_Value(A), m_Value(B)))) {
3626 if (A == Op0 || B == Op0) // A & (A | ?) --> A
3627 return ReplaceInstUsesWith(I, Op0);
3629 // ~(A&B) & (A|B) -> A^B
3630 if (match(Op0, m_Not(m_And(m_Value(C), m_Value(D))))) {
3631 if ((A == C && B == D) || (A == D && B == C))
3632 return BinaryOperator::CreateXor(A, B);
3636 if (Op0->hasOneUse() &&
3637 match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
3638 if (A == Op1) { // (A^B)&A -> A&(A^B)
3639 I.swapOperands(); // Simplify below
3640 std::swap(Op0, Op1);
3641 } else if (B == Op1) { // (A^B)&B -> B&(B^A)
3642 cast<BinaryOperator>(Op0)->swapOperands();
3643 I.swapOperands(); // Simplify below
3644 std::swap(Op0, Op1);
3647 if (Op1->hasOneUse() &&
3648 match(Op1, m_Xor(m_Value(A), m_Value(B)))) {
3649 if (B == Op0) { // B&(A^B) -> B&(B^A)
3650 cast<BinaryOperator>(Op1)->swapOperands();
3653 if (A == Op0) { // A&(A^B) -> A & ~B
3654 Instruction *NotB = BinaryOperator::CreateNot(B, "tmp");
3655 InsertNewInstBefore(NotB, I);
3656 return BinaryOperator::CreateAnd(A, NotB);
3661 { // (icmp ult A, C) & (icmp ult B, C) --> (icmp ult (A|B), C)
3662 // where C is a power of 2
3664 ConstantInt *C1, *C2;
3665 ICmpInst::Predicate LHSCC = ICmpInst::BAD_ICMP_PREDICATE;
3666 ICmpInst::Predicate RHSCC = ICmpInst::BAD_ICMP_PREDICATE;
3667 if (match(&I, m_And(m_ICmp(LHSCC, m_Value(A), m_ConstantInt(C1)),
3668 m_ICmp(RHSCC, m_Value(B), m_ConstantInt(C2)))))
3669 if (C1 == C2 && LHSCC == RHSCC && LHSCC == ICmpInst::ICMP_ULT &&
3670 C1->getValue().isPowerOf2()) {
3671 Instruction *NewOr = BinaryOperator::CreateOr(A, B);
3672 InsertNewInstBefore(NewOr, I);
3673 return new ICmpInst(LHSCC, NewOr, C1);
3677 if (ICmpInst *RHS = dyn_cast<ICmpInst>(Op1)) {
3678 // (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
3679 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
3682 Value *LHSVal, *RHSVal;
3683 ConstantInt *LHSCst, *RHSCst;
3684 ICmpInst::Predicate LHSCC, RHSCC;
3685 if (match(Op0, m_ICmp(LHSCC, m_Value(LHSVal), m_ConstantInt(LHSCst))))
3686 if (match(RHS, m_ICmp(RHSCC, m_Value(RHSVal), m_ConstantInt(RHSCst))))
3687 if (LHSVal == RHSVal && // Found (X icmp C1) & (X icmp C2)
3688 // ICMP_[GL]E X, CST is folded to ICMP_[GL]T elsewhere.
3689 LHSCC != ICmpInst::ICMP_UGE && LHSCC != ICmpInst::ICMP_ULE &&
3690 RHSCC != ICmpInst::ICMP_UGE && RHSCC != ICmpInst::ICMP_ULE &&
3691 LHSCC != ICmpInst::ICMP_SGE && LHSCC != ICmpInst::ICMP_SLE &&
3692 RHSCC != ICmpInst::ICMP_SGE && RHSCC != ICmpInst::ICMP_SLE &&
3694 // Don't try to fold ICMP_SLT + ICMP_ULT.
3695 (ICmpInst::isEquality(LHSCC) || ICmpInst::isEquality(RHSCC) ||
3696 ICmpInst::isSignedPredicate(LHSCC) ==
3697 ICmpInst::isSignedPredicate(RHSCC))) {
3698 // Ensure that the larger constant is on the RHS.
3699 ICmpInst::Predicate GT;
3700 if (ICmpInst::isSignedPredicate(LHSCC) ||
3701 (ICmpInst::isEquality(LHSCC) &&
3702 ICmpInst::isSignedPredicate(RHSCC)))
3703 GT = ICmpInst::ICMP_SGT;
3705 GT = ICmpInst::ICMP_UGT;
3707 Constant *Cmp = ConstantExpr::getICmp(GT, LHSCst, RHSCst);
3708 ICmpInst *LHS = cast<ICmpInst>(Op0);
3709 if (cast<ConstantInt>(Cmp)->getZExtValue()) {
3710 std::swap(LHS, RHS);
3711 std::swap(LHSCst, RHSCst);
3712 std::swap(LHSCC, RHSCC);
3715 // At this point, we know we have have two icmp instructions
3716 // comparing a value against two constants and and'ing the result
3717 // together. Because of the above check, we know that we only have
3718 // icmp eq, icmp ne, icmp [su]lt, and icmp [SU]gt here. We also know
3719 // (from the FoldICmpLogical check above), that the two constants
3720 // are not equal and that the larger constant is on the RHS
3721 assert(LHSCst != RHSCst && "Compares not folded above?");
3724 default: assert(0 && "Unknown integer condition code!");
3725 case ICmpInst::ICMP_EQ:
3727 default: assert(0 && "Unknown integer condition code!");
3728 case ICmpInst::ICMP_EQ: // (X == 13 & X == 15) -> false
3729 case ICmpInst::ICMP_UGT: // (X == 13 & X > 15) -> false
3730 case ICmpInst::ICMP_SGT: // (X == 13 & X > 15) -> false
3731 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3732 case ICmpInst::ICMP_NE: // (X == 13 & X != 15) -> X == 13
3733 case ICmpInst::ICMP_ULT: // (X == 13 & X < 15) -> X == 13
3734 case ICmpInst::ICMP_SLT: // (X == 13 & X < 15) -> X == 13
3735 return ReplaceInstUsesWith(I, LHS);
3737 case ICmpInst::ICMP_NE:
3739 default: assert(0 && "Unknown integer condition code!");
3740 case ICmpInst::ICMP_ULT:
3741 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X u< 14) -> X < 13
3742 return new ICmpInst(ICmpInst::ICMP_ULT, LHSVal, LHSCst);
3743 break; // (X != 13 & X u< 15) -> no change
3744 case ICmpInst::ICMP_SLT:
3745 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X s< 14) -> X < 13
3746 return new ICmpInst(ICmpInst::ICMP_SLT, LHSVal, LHSCst);
3747 break; // (X != 13 & X s< 15) -> no change
3748 case ICmpInst::ICMP_EQ: // (X != 13 & X == 15) -> X == 15
3749 case ICmpInst::ICMP_UGT: // (X != 13 & X u> 15) -> X u> 15
3750 case ICmpInst::ICMP_SGT: // (X != 13 & X s> 15) -> X s> 15
3751 return ReplaceInstUsesWith(I, RHS);
3752 case ICmpInst::ICMP_NE:
3753 if (LHSCst == SubOne(RHSCst)){// (X != 13 & X != 14) -> X-13 >u 1
3754 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
3755 Instruction *Add = BinaryOperator::CreateAdd(LHSVal, AddCST,
3756 LHSVal->getName()+".off");
3757 InsertNewInstBefore(Add, I);
3758 return new ICmpInst(ICmpInst::ICMP_UGT, Add,
3759 ConstantInt::get(Add->getType(), 1));
3761 break; // (X != 13 & X != 15) -> no change
3764 case ICmpInst::ICMP_ULT:
3766 default: assert(0 && "Unknown integer condition code!");
3767 case ICmpInst::ICMP_EQ: // (X u< 13 & X == 15) -> false
3768 case ICmpInst::ICMP_UGT: // (X u< 13 & X u> 15) -> false
3769 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3770 case ICmpInst::ICMP_SGT: // (X u< 13 & X s> 15) -> no change
3772 case ICmpInst::ICMP_NE: // (X u< 13 & X != 15) -> X u< 13
3773 case ICmpInst::ICMP_ULT: // (X u< 13 & X u< 15) -> X u< 13
3774 return ReplaceInstUsesWith(I, LHS);
3775 case ICmpInst::ICMP_SLT: // (X u< 13 & X s< 15) -> no change
3779 case ICmpInst::ICMP_SLT:
3781 default: assert(0 && "Unknown integer condition code!");
3782 case ICmpInst::ICMP_EQ: // (X s< 13 & X == 15) -> false
3783 case ICmpInst::ICMP_SGT: // (X s< 13 & X s> 15) -> false
3784 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3785 case ICmpInst::ICMP_UGT: // (X s< 13 & X u> 15) -> no change
3787 case ICmpInst::ICMP_NE: // (X s< 13 & X != 15) -> X < 13
3788 case ICmpInst::ICMP_SLT: // (X s< 13 & X s< 15) -> X < 13
3789 return ReplaceInstUsesWith(I, LHS);
3790 case ICmpInst::ICMP_ULT: // (X s< 13 & X u< 15) -> no change
3794 case ICmpInst::ICMP_UGT:
3796 default: assert(0 && "Unknown integer condition code!");
3797 case ICmpInst::ICMP_EQ: // (X u> 13 & X == 15) -> X == 15
3798 case ICmpInst::ICMP_UGT: // (X u> 13 & X u> 15) -> X u> 15
3799 return ReplaceInstUsesWith(I, RHS);
3800 case ICmpInst::ICMP_SGT: // (X u> 13 & X s> 15) -> no change
3802 case ICmpInst::ICMP_NE:
3803 if (RHSCst == AddOne(LHSCst)) // (X u> 13 & X != 14) -> X u> 14
3804 return new ICmpInst(LHSCC, LHSVal, RHSCst);
3805 break; // (X u> 13 & X != 15) -> no change
3806 case ICmpInst::ICMP_ULT: // (X u> 13 & X u< 15) ->(X-14) <u 1
3807 return InsertRangeTest(LHSVal, AddOne(LHSCst), RHSCst, false,
3809 case ICmpInst::ICMP_SLT: // (X u> 13 & X s< 15) -> no change
3813 case ICmpInst::ICMP_SGT:
3815 default: assert(0 && "Unknown integer condition code!");
3816 case ICmpInst::ICMP_EQ: // (X s> 13 & X == 15) -> X == 15
3817 case ICmpInst::ICMP_SGT: // (X s> 13 & X s> 15) -> X s> 15
3818 return ReplaceInstUsesWith(I, RHS);
3819 case ICmpInst::ICMP_UGT: // (X s> 13 & X u> 15) -> no change
3821 case ICmpInst::ICMP_NE:
3822 if (RHSCst == AddOne(LHSCst)) // (X s> 13 & X != 14) -> X s> 14
3823 return new ICmpInst(LHSCC, LHSVal, RHSCst);
3824 break; // (X s> 13 & X != 15) -> no change
3825 case ICmpInst::ICMP_SLT: // (X s> 13 & X s< 15) ->(X-14) s< 1
3826 return InsertRangeTest(LHSVal, AddOne(LHSCst), RHSCst, true,
3828 case ICmpInst::ICMP_ULT: // (X s> 13 & X u< 15) -> no change
3836 // fold (and (cast A), (cast B)) -> (cast (and A, B))
3837 if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
3838 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
3839 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind ?
3840 const Type *SrcTy = Op0C->getOperand(0)->getType();
3841 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
3842 // Only do this if the casts both really cause code to be generated.
3843 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
3845 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
3847 Instruction *NewOp = BinaryOperator::CreateAnd(Op0C->getOperand(0),
3848 Op1C->getOperand(0),
3850 InsertNewInstBefore(NewOp, I);
3851 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
3855 // (X >> Z) & (Y >> Z) -> (X&Y) >> Z for all shifts.
3856 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
3857 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
3858 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
3859 SI0->getOperand(1) == SI1->getOperand(1) &&
3860 (SI0->hasOneUse() || SI1->hasOneUse())) {
3861 Instruction *NewOp =
3862 InsertNewInstBefore(BinaryOperator::CreateAnd(SI0->getOperand(0),
3864 SI0->getName()), I);
3865 return BinaryOperator::Create(SI1->getOpcode(), NewOp,
3866 SI1->getOperand(1));
3870 // (fcmp ord x, c) & (fcmp ord y, c) -> (fcmp ord x, y)
3871 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0))) {
3872 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1))) {
3873 if (LHS->getPredicate() == FCmpInst::FCMP_ORD &&
3874 RHS->getPredicate() == FCmpInst::FCMP_ORD)
3875 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
3876 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
3877 // If either of the constants are nans, then the whole thing returns
3879 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
3880 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3881 return new FCmpInst(FCmpInst::FCMP_ORD, LHS->getOperand(0),
3882 RHS->getOperand(0));
3887 return Changed ? &I : 0;
3890 /// CollectBSwapParts - Look to see if the specified value defines a single byte
3891 /// in the result. If it does, and if the specified byte hasn't been filled in
3892 /// yet, fill it in and return false.
3893 static bool CollectBSwapParts(Value *V, SmallVector<Value*, 8> &ByteValues) {
3894 Instruction *I = dyn_cast<Instruction>(V);
3895 if (I == 0) return true;
3897 // If this is an or instruction, it is an inner node of the bswap.
3898 if (I->getOpcode() == Instruction::Or)
3899 return CollectBSwapParts(I->getOperand(0), ByteValues) ||
3900 CollectBSwapParts(I->getOperand(1), ByteValues);
3902 uint32_t BitWidth = I->getType()->getPrimitiveSizeInBits();
3903 // If this is a shift by a constant int, and it is "24", then its operand
3904 // defines a byte. We only handle unsigned types here.
3905 if (I->isShift() && isa<ConstantInt>(I->getOperand(1))) {
3906 // Not shifting the entire input by N-1 bytes?
3907 if (cast<ConstantInt>(I->getOperand(1))->getLimitedValue(BitWidth) !=
3908 8*(ByteValues.size()-1))
3912 if (I->getOpcode() == Instruction::Shl) {
3913 // X << 24 defines the top byte with the lowest of the input bytes.
3914 DestNo = ByteValues.size()-1;
3916 // X >>u 24 defines the low byte with the highest of the input bytes.
3920 // If the destination byte value is already defined, the values are or'd
3921 // together, which isn't a bswap (unless it's an or of the same bits).
3922 if (ByteValues[DestNo] && ByteValues[DestNo] != I->getOperand(0))
3924 ByteValues[DestNo] = I->getOperand(0);
3928 // Otherwise, we can only handle and(shift X, imm), imm). Bail out of if we
3930 Value *Shift = 0, *ShiftLHS = 0;
3931 ConstantInt *AndAmt = 0, *ShiftAmt = 0;
3932 if (!match(I, m_And(m_Value(Shift), m_ConstantInt(AndAmt))) ||
3933 !match(Shift, m_Shift(m_Value(ShiftLHS), m_ConstantInt(ShiftAmt))))
3935 Instruction *SI = cast<Instruction>(Shift);
3937 // Make sure that the shift amount is by a multiple of 8 and isn't too big.
3938 if (ShiftAmt->getLimitedValue(BitWidth) & 7 ||
3939 ShiftAmt->getLimitedValue(BitWidth) > 8*ByteValues.size())
3942 // Turn 0xFF -> 0, 0xFF00 -> 1, 0xFF0000 -> 2, etc.
3944 if (AndAmt->getValue().getActiveBits() > 64)
3946 uint64_t AndAmtVal = AndAmt->getZExtValue();
3947 for (DestByte = 0; DestByte != ByteValues.size(); ++DestByte)
3948 if (AndAmtVal == uint64_t(0xFF) << 8*DestByte)
3950 // Unknown mask for bswap.
3951 if (DestByte == ByteValues.size()) return true;
3953 unsigned ShiftBytes = ShiftAmt->getZExtValue()/8;
3955 if (SI->getOpcode() == Instruction::Shl)
3956 SrcByte = DestByte - ShiftBytes;
3958 SrcByte = DestByte + ShiftBytes;
3960 // If the SrcByte isn't a bswapped value from the DestByte, reject it.
3961 if (SrcByte != ByteValues.size()-DestByte-1)
3964 // If the destination byte value is already defined, the values are or'd
3965 // together, which isn't a bswap (unless it's an or of the same bits).
3966 if (ByteValues[DestByte] && ByteValues[DestByte] != SI->getOperand(0))
3968 ByteValues[DestByte] = SI->getOperand(0);
3972 /// MatchBSwap - Given an OR instruction, check to see if this is a bswap idiom.
3973 /// If so, insert the new bswap intrinsic and return it.
3974 Instruction *InstCombiner::MatchBSwap(BinaryOperator &I) {
3975 const IntegerType *ITy = dyn_cast<IntegerType>(I.getType());
3976 if (!ITy || ITy->getBitWidth() % 16)
3977 return 0; // Can only bswap pairs of bytes. Can't do vectors.
3979 /// ByteValues - For each byte of the result, we keep track of which value
3980 /// defines each byte.
3981 SmallVector<Value*, 8> ByteValues;
3982 ByteValues.resize(ITy->getBitWidth()/8);
3984 // Try to find all the pieces corresponding to the bswap.
3985 if (CollectBSwapParts(I.getOperand(0), ByteValues) ||
3986 CollectBSwapParts(I.getOperand(1), ByteValues))
3989 // Check to see if all of the bytes come from the same value.
3990 Value *V = ByteValues[0];
3991 if (V == 0) return 0; // Didn't find a byte? Must be zero.
3993 // Check to make sure that all of the bytes come from the same value.
3994 for (unsigned i = 1, e = ByteValues.size(); i != e; ++i)
3995 if (ByteValues[i] != V)
3997 const Type *Tys[] = { ITy };
3998 Module *M = I.getParent()->getParent()->getParent();
3999 Function *F = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 1);
4000 return CallInst::Create(F, V);
4004 Instruction *InstCombiner::visitOr(BinaryOperator &I) {
4005 bool Changed = SimplifyCommutative(I);
4006 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4008 if (isa<UndefValue>(Op1)) // X | undef -> -1
4009 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4013 return ReplaceInstUsesWith(I, Op0);
4015 // See if we can simplify any instructions used by the instruction whose sole
4016 // purpose is to compute bits we don't care about.
4017 if (!isa<VectorType>(I.getType())) {
4018 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
4019 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
4020 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
4021 KnownZero, KnownOne))
4023 } else if (isa<ConstantAggregateZero>(Op1)) {
4024 return ReplaceInstUsesWith(I, Op0); // X | <0,0> -> X
4025 } else if (ConstantVector *CP = dyn_cast<ConstantVector>(Op1)) {
4026 if (CP->isAllOnesValue()) // X | <-1,-1> -> <-1,-1>
4027 return ReplaceInstUsesWith(I, I.getOperand(1));
4033 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
4034 ConstantInt *C1 = 0; Value *X = 0;
4035 // (X & C1) | C2 --> (X | C2) & (C1|C2)
4036 if (match(Op0, m_And(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
4037 Instruction *Or = BinaryOperator::CreateOr(X, RHS);
4038 InsertNewInstBefore(Or, I);
4040 return BinaryOperator::CreateAnd(Or,
4041 ConstantInt::get(RHS->getValue() | C1->getValue()));
4044 // (X ^ C1) | C2 --> (X | C2) ^ (C1&~C2)
4045 if (match(Op0, m_Xor(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
4046 Instruction *Or = BinaryOperator::CreateOr(X, RHS);
4047 InsertNewInstBefore(Or, I);
4049 return BinaryOperator::CreateXor(Or,
4050 ConstantInt::get(C1->getValue() & ~RHS->getValue()));
4053 // Try to fold constant and into select arguments.
4054 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
4055 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
4057 if (isa<PHINode>(Op0))
4058 if (Instruction *NV = FoldOpIntoPhi(I))
4062 Value *A = 0, *B = 0;
4063 ConstantInt *C1 = 0, *C2 = 0;
4065 if (match(Op0, m_And(m_Value(A), m_Value(B))))
4066 if (A == Op1 || B == Op1) // (A & ?) | A --> A
4067 return ReplaceInstUsesWith(I, Op1);
4068 if (match(Op1, m_And(m_Value(A), m_Value(B))))
4069 if (A == Op0 || B == Op0) // A | (A & ?) --> A
4070 return ReplaceInstUsesWith(I, Op0);
4072 // (A | B) | C and A | (B | C) -> bswap if possible.
4073 // (A >> B) | (C << D) and (A << B) | (B >> C) -> bswap if possible.
4074 if (match(Op0, m_Or(m_Value(), m_Value())) ||
4075 match(Op1, m_Or(m_Value(), m_Value())) ||
4076 (match(Op0, m_Shift(m_Value(), m_Value())) &&
4077 match(Op1, m_Shift(m_Value(), m_Value())))) {
4078 if (Instruction *BSwap = MatchBSwap(I))
4082 // (X^C)|Y -> (X|Y)^C iff Y&C == 0
4083 if (Op0->hasOneUse() && match(Op0, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
4084 MaskedValueIsZero(Op1, C1->getValue())) {
4085 Instruction *NOr = BinaryOperator::CreateOr(A, Op1);
4086 InsertNewInstBefore(NOr, I);
4088 return BinaryOperator::CreateXor(NOr, C1);
4091 // Y|(X^C) -> (X|Y)^C iff Y&C == 0
4092 if (Op1->hasOneUse() && match(Op1, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
4093 MaskedValueIsZero(Op0, C1->getValue())) {
4094 Instruction *NOr = BinaryOperator::CreateOr(A, Op0);
4095 InsertNewInstBefore(NOr, I);
4097 return BinaryOperator::CreateXor(NOr, C1);
4101 Value *C = 0, *D = 0;
4102 if (match(Op0, m_And(m_Value(A), m_Value(C))) &&
4103 match(Op1, m_And(m_Value(B), m_Value(D)))) {
4104 Value *V1 = 0, *V2 = 0, *V3 = 0;
4105 C1 = dyn_cast<ConstantInt>(C);
4106 C2 = dyn_cast<ConstantInt>(D);
4107 if (C1 && C2) { // (A & C1)|(B & C2)
4108 // If we have: ((V + N) & C1) | (V & C2)
4109 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
4110 // replace with V+N.
4111 if (C1->getValue() == ~C2->getValue()) {
4112 if ((C2->getValue() & (C2->getValue()+1)) == 0 && // C2 == 0+1+
4113 match(A, m_Add(m_Value(V1), m_Value(V2)))) {
4114 // Add commutes, try both ways.
4115 if (V1 == B && MaskedValueIsZero(V2, C2->getValue()))
4116 return ReplaceInstUsesWith(I, A);
4117 if (V2 == B && MaskedValueIsZero(V1, C2->getValue()))
4118 return ReplaceInstUsesWith(I, A);
4120 // Or commutes, try both ways.
4121 if ((C1->getValue() & (C1->getValue()+1)) == 0 &&
4122 match(B, m_Add(m_Value(V1), m_Value(V2)))) {
4123 // Add commutes, try both ways.
4124 if (V1 == A && MaskedValueIsZero(V2, C1->getValue()))
4125 return ReplaceInstUsesWith(I, B);
4126 if (V2 == A && MaskedValueIsZero(V1, C1->getValue()))
4127 return ReplaceInstUsesWith(I, B);
4130 V1 = 0; V2 = 0; V3 = 0;
4133 // Check to see if we have any common things being and'ed. If so, find the
4134 // terms for V1 & (V2|V3).
4135 if (isOnlyUse(Op0) || isOnlyUse(Op1)) {
4136 if (A == B) // (A & C)|(A & D) == A & (C|D)
4137 V1 = A, V2 = C, V3 = D;
4138 else if (A == D) // (A & C)|(B & A) == A & (B|C)
4139 V1 = A, V2 = B, V3 = C;
4140 else if (C == B) // (A & C)|(C & D) == C & (A|D)
4141 V1 = C, V2 = A, V3 = D;
4142 else if (C == D) // (A & C)|(B & C) == C & (A|B)
4143 V1 = C, V2 = A, V3 = B;
4147 InsertNewInstBefore(BinaryOperator::CreateOr(V2, V3, "tmp"), I);
4148 return BinaryOperator::CreateAnd(V1, Or);
4153 // (X >> Z) | (Y >> Z) -> (X|Y) >> Z for all shifts.
4154 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
4155 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
4156 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
4157 SI0->getOperand(1) == SI1->getOperand(1) &&
4158 (SI0->hasOneUse() || SI1->hasOneUse())) {
4159 Instruction *NewOp =
4160 InsertNewInstBefore(BinaryOperator::CreateOr(SI0->getOperand(0),
4162 SI0->getName()), I);
4163 return BinaryOperator::Create(SI1->getOpcode(), NewOp,
4164 SI1->getOperand(1));
4168 if (match(Op0, m_Not(m_Value(A)))) { // ~A | Op1
4169 if (A == Op1) // ~A | A == -1
4170 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4174 // Note, A is still live here!
4175 if (match(Op1, m_Not(m_Value(B)))) { // Op0 | ~B
4177 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4179 // (~A | ~B) == (~(A & B)) - De Morgan's Law
4180 if (A && isOnlyUse(Op0) && isOnlyUse(Op1)) {
4181 Value *And = InsertNewInstBefore(BinaryOperator::CreateAnd(A, B,
4182 I.getName()+".demorgan"), I);
4183 return BinaryOperator::CreateNot(And);
4187 // (icmp1 A, B) | (icmp2 A, B) --> (icmp3 A, B)
4188 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1))) {
4189 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
4192 Value *LHSVal, *RHSVal;
4193 ConstantInt *LHSCst, *RHSCst;
4194 ICmpInst::Predicate LHSCC, RHSCC;
4195 if (match(Op0, m_ICmp(LHSCC, m_Value(LHSVal), m_ConstantInt(LHSCst))))
4196 if (match(RHS, m_ICmp(RHSCC, m_Value(RHSVal), m_ConstantInt(RHSCst))))
4197 if (LHSVal == RHSVal && // Found (X icmp C1) | (X icmp C2)
4198 // icmp [us][gl]e x, cst is folded to icmp [us][gl]t elsewhere.
4199 LHSCC != ICmpInst::ICMP_UGE && LHSCC != ICmpInst::ICMP_ULE &&
4200 RHSCC != ICmpInst::ICMP_UGE && RHSCC != ICmpInst::ICMP_ULE &&
4201 LHSCC != ICmpInst::ICMP_SGE && LHSCC != ICmpInst::ICMP_SLE &&
4202 RHSCC != ICmpInst::ICMP_SGE && RHSCC != ICmpInst::ICMP_SLE &&
4203 // We can't fold (ugt x, C) | (sgt x, C2).
4204 PredicatesFoldable(LHSCC, RHSCC)) {
4205 // Ensure that the larger constant is on the RHS.
4206 ICmpInst *LHS = cast<ICmpInst>(Op0);
4208 if (ICmpInst::isSignedPredicate(LHSCC))
4209 NeedsSwap = LHSCst->getValue().sgt(RHSCst->getValue());
4211 NeedsSwap = LHSCst->getValue().ugt(RHSCst->getValue());
4214 std::swap(LHS, RHS);
4215 std::swap(LHSCst, RHSCst);
4216 std::swap(LHSCC, RHSCC);
4219 // At this point, we know we have have two icmp instructions
4220 // comparing a value against two constants and or'ing the result
4221 // together. Because of the above check, we know that we only have
4222 // ICMP_EQ, ICMP_NE, ICMP_LT, and ICMP_GT here. We also know (from the
4223 // FoldICmpLogical check above), that the two constants are not
4225 assert(LHSCst != RHSCst && "Compares not folded above?");
4228 default: assert(0 && "Unknown integer condition code!");
4229 case ICmpInst::ICMP_EQ:
4231 default: assert(0 && "Unknown integer condition code!");
4232 case ICmpInst::ICMP_EQ:
4233 if (LHSCst == SubOne(RHSCst)) {// (X == 13 | X == 14) -> X-13 <u 2
4234 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
4235 Instruction *Add = BinaryOperator::CreateAdd(LHSVal, AddCST,
4236 LHSVal->getName()+".off");
4237 InsertNewInstBefore(Add, I);
4238 AddCST = Subtract(AddOne(RHSCst), LHSCst);
4239 return new ICmpInst(ICmpInst::ICMP_ULT, Add, AddCST);
4241 break; // (X == 13 | X == 15) -> no change
4242 case ICmpInst::ICMP_UGT: // (X == 13 | X u> 14) -> no change
4243 case ICmpInst::ICMP_SGT: // (X == 13 | X s> 14) -> no change
4245 case ICmpInst::ICMP_NE: // (X == 13 | X != 15) -> X != 15
4246 case ICmpInst::ICMP_ULT: // (X == 13 | X u< 15) -> X u< 15
4247 case ICmpInst::ICMP_SLT: // (X == 13 | X s< 15) -> X s< 15
4248 return ReplaceInstUsesWith(I, RHS);
4251 case ICmpInst::ICMP_NE:
4253 default: assert(0 && "Unknown integer condition code!");
4254 case ICmpInst::ICMP_EQ: // (X != 13 | X == 15) -> X != 13
4255 case ICmpInst::ICMP_UGT: // (X != 13 | X u> 15) -> X != 13
4256 case ICmpInst::ICMP_SGT: // (X != 13 | X s> 15) -> X != 13
4257 return ReplaceInstUsesWith(I, LHS);
4258 case ICmpInst::ICMP_NE: // (X != 13 | X != 15) -> true
4259 case ICmpInst::ICMP_ULT: // (X != 13 | X u< 15) -> true
4260 case ICmpInst::ICMP_SLT: // (X != 13 | X s< 15) -> true
4261 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4264 case ICmpInst::ICMP_ULT:
4266 default: assert(0 && "Unknown integer condition code!");
4267 case ICmpInst::ICMP_EQ: // (X u< 13 | X == 14) -> no change
4269 case ICmpInst::ICMP_UGT: // (X u< 13 | X u> 15) ->(X-13) u> 2
4270 // If RHSCst is [us]MAXINT, it is always false. Not handling
4271 // this can cause overflow.
4272 if (RHSCst->isMaxValue(false))
4273 return ReplaceInstUsesWith(I, LHS);
4274 return InsertRangeTest(LHSVal, LHSCst, AddOne(RHSCst), false,
4276 case ICmpInst::ICMP_SGT: // (X u< 13 | X s> 15) -> no change
4278 case ICmpInst::ICMP_NE: // (X u< 13 | X != 15) -> X != 15
4279 case ICmpInst::ICMP_ULT: // (X u< 13 | X u< 15) -> X u< 15
4280 return ReplaceInstUsesWith(I, RHS);
4281 case ICmpInst::ICMP_SLT: // (X u< 13 | X s< 15) -> no change
4285 case ICmpInst::ICMP_SLT:
4287 default: assert(0 && "Unknown integer condition code!");
4288 case ICmpInst::ICMP_EQ: // (X s< 13 | X == 14) -> no change
4290 case ICmpInst::ICMP_SGT: // (X s< 13 | X s> 15) ->(X-13) s> 2
4291 // If RHSCst is [us]MAXINT, it is always false. Not handling
4292 // this can cause overflow.
4293 if (RHSCst->isMaxValue(true))
4294 return ReplaceInstUsesWith(I, LHS);
4295 return InsertRangeTest(LHSVal, LHSCst, AddOne(RHSCst), true,
4297 case ICmpInst::ICMP_UGT: // (X s< 13 | X u> 15) -> no change
4299 case ICmpInst::ICMP_NE: // (X s< 13 | X != 15) -> X != 15
4300 case ICmpInst::ICMP_SLT: // (X s< 13 | X s< 15) -> X s< 15
4301 return ReplaceInstUsesWith(I, RHS);
4302 case ICmpInst::ICMP_ULT: // (X s< 13 | X u< 15) -> no change
4306 case ICmpInst::ICMP_UGT:
4308 default: assert(0 && "Unknown integer condition code!");
4309 case ICmpInst::ICMP_EQ: // (X u> 13 | X == 15) -> X u> 13
4310 case ICmpInst::ICMP_UGT: // (X u> 13 | X u> 15) -> X u> 13
4311 return ReplaceInstUsesWith(I, LHS);
4312 case ICmpInst::ICMP_SGT: // (X u> 13 | X s> 15) -> no change
4314 case ICmpInst::ICMP_NE: // (X u> 13 | X != 15) -> true
4315 case ICmpInst::ICMP_ULT: // (X u> 13 | X u< 15) -> true
4316 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4317 case ICmpInst::ICMP_SLT: // (X u> 13 | X s< 15) -> no change
4321 case ICmpInst::ICMP_SGT:
4323 default: assert(0 && "Unknown integer condition code!");
4324 case ICmpInst::ICMP_EQ: // (X s> 13 | X == 15) -> X > 13
4325 case ICmpInst::ICMP_SGT: // (X s> 13 | X s> 15) -> X > 13
4326 return ReplaceInstUsesWith(I, LHS);
4327 case ICmpInst::ICMP_UGT: // (X s> 13 | X u> 15) -> no change
4329 case ICmpInst::ICMP_NE: // (X s> 13 | X != 15) -> true
4330 case ICmpInst::ICMP_SLT: // (X s> 13 | X s< 15) -> true
4331 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4332 case ICmpInst::ICMP_ULT: // (X s> 13 | X u< 15) -> no change
4340 // fold (or (cast A), (cast B)) -> (cast (or A, B))
4341 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
4342 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4343 if (Op0C->getOpcode() == Op1C->getOpcode()) {// same cast kind ?
4344 if (!isa<ICmpInst>(Op0C->getOperand(0)) ||
4345 !isa<ICmpInst>(Op1C->getOperand(0))) {
4346 const Type *SrcTy = Op0C->getOperand(0)->getType();
4347 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
4348 // Only do this if the casts both really cause code to be
4350 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4352 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4354 Instruction *NewOp = BinaryOperator::CreateOr(Op0C->getOperand(0),
4355 Op1C->getOperand(0),
4357 InsertNewInstBefore(NewOp, I);
4358 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
4365 // (fcmp uno x, c) | (fcmp uno y, c) -> (fcmp uno x, y)
4366 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0))) {
4367 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1))) {
4368 if (LHS->getPredicate() == FCmpInst::FCMP_UNO &&
4369 RHS->getPredicate() == FCmpInst::FCMP_UNO &&
4370 LHS->getOperand(0)->getType() == RHS->getOperand(0)->getType())
4371 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
4372 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
4373 // If either of the constants are nans, then the whole thing returns
4375 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
4376 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4378 // Otherwise, no need to compare the two constants, compare the
4380 return new FCmpInst(FCmpInst::FCMP_UNO, LHS->getOperand(0),
4381 RHS->getOperand(0));
4386 return Changed ? &I : 0;
4391 // XorSelf - Implements: X ^ X --> 0
4394 XorSelf(Value *rhs) : RHS(rhs) {}
4395 bool shouldApply(Value *LHS) const { return LHS == RHS; }
4396 Instruction *apply(BinaryOperator &Xor) const {
4403 Instruction *InstCombiner::visitXor(BinaryOperator &I) {
4404 bool Changed = SimplifyCommutative(I);
4405 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4407 if (isa<UndefValue>(Op1)) {
4408 if (isa<UndefValue>(Op0))
4409 // Handle undef ^ undef -> 0 special case. This is a common
4411 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
4412 return ReplaceInstUsesWith(I, Op1); // X ^ undef -> undef
4415 // xor X, X = 0, even if X is nested in a sequence of Xor's.
4416 if (Instruction *Result = AssociativeOpt(I, XorSelf(Op1))) {
4417 assert(Result == &I && "AssociativeOpt didn't work?"); Result=Result;
4418 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
4421 // See if we can simplify any instructions used by the instruction whose sole
4422 // purpose is to compute bits we don't care about.
4423 if (!isa<VectorType>(I.getType())) {
4424 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
4425 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
4426 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
4427 KnownZero, KnownOne))
4429 } else if (isa<ConstantAggregateZero>(Op1)) {
4430 return ReplaceInstUsesWith(I, Op0); // X ^ <0,0> -> X
4433 // Is this a ~ operation?
4434 if (Value *NotOp = dyn_castNotVal(&I)) {
4435 // ~(~X & Y) --> (X | ~Y) - De Morgan's Law
4436 // ~(~X | Y) === (X & ~Y) - De Morgan's Law
4437 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(NotOp)) {
4438 if (Op0I->getOpcode() == Instruction::And ||
4439 Op0I->getOpcode() == Instruction::Or) {
4440 if (dyn_castNotVal(Op0I->getOperand(1))) Op0I->swapOperands();
4441 if (Value *Op0NotVal = dyn_castNotVal(Op0I->getOperand(0))) {
4443 BinaryOperator::CreateNot(Op0I->getOperand(1),
4444 Op0I->getOperand(1)->getName()+".not");
4445 InsertNewInstBefore(NotY, I);
4446 if (Op0I->getOpcode() == Instruction::And)
4447 return BinaryOperator::CreateOr(Op0NotVal, NotY);
4449 return BinaryOperator::CreateAnd(Op0NotVal, NotY);
4456 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
4457 // xor (cmp A, B), true = not (cmp A, B) = !cmp A, B
4458 if (RHS == ConstantInt::getTrue() && Op0->hasOneUse()) {
4459 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Op0))
4460 return new ICmpInst(ICI->getInversePredicate(),
4461 ICI->getOperand(0), ICI->getOperand(1));
4463 if (FCmpInst *FCI = dyn_cast<FCmpInst>(Op0))
4464 return new FCmpInst(FCI->getInversePredicate(),
4465 FCI->getOperand(0), FCI->getOperand(1));
4468 // fold (xor(zext(cmp)), 1) and (xor(sext(cmp)), -1) to ext(!cmp).
4469 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
4470 if (CmpInst *CI = dyn_cast<CmpInst>(Op0C->getOperand(0))) {
4471 if (CI->hasOneUse() && Op0C->hasOneUse()) {
4472 Instruction::CastOps Opcode = Op0C->getOpcode();
4473 if (Opcode == Instruction::ZExt || Opcode == Instruction::SExt) {
4474 if (RHS == ConstantExpr::getCast(Opcode, ConstantInt::getTrue(),
4475 Op0C->getDestTy())) {
4476 Instruction *NewCI = InsertNewInstBefore(CmpInst::Create(
4477 CI->getOpcode(), CI->getInversePredicate(),
4478 CI->getOperand(0), CI->getOperand(1)), I);
4479 NewCI->takeName(CI);
4480 return CastInst::Create(Opcode, NewCI, Op0C->getType());
4487 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
4488 // ~(c-X) == X-c-1 == X+(-c-1)
4489 if (Op0I->getOpcode() == Instruction::Sub && RHS->isAllOnesValue())
4490 if (Constant *Op0I0C = dyn_cast<Constant>(Op0I->getOperand(0))) {
4491 Constant *NegOp0I0C = ConstantExpr::getNeg(Op0I0C);
4492 Constant *ConstantRHS = ConstantExpr::getSub(NegOp0I0C,
4493 ConstantInt::get(I.getType(), 1));
4494 return BinaryOperator::CreateAdd(Op0I->getOperand(1), ConstantRHS);
4497 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) {
4498 if (Op0I->getOpcode() == Instruction::Add) {
4499 // ~(X-c) --> (-c-1)-X
4500 if (RHS->isAllOnesValue()) {
4501 Constant *NegOp0CI = ConstantExpr::getNeg(Op0CI);
4502 return BinaryOperator::CreateSub(
4503 ConstantExpr::getSub(NegOp0CI,
4504 ConstantInt::get(I.getType(), 1)),
4505 Op0I->getOperand(0));
4506 } else if (RHS->getValue().isSignBit()) {
4507 // (X + C) ^ signbit -> (X + C + signbit)
4508 Constant *C = ConstantInt::get(RHS->getValue() + Op0CI->getValue());
4509 return BinaryOperator::CreateAdd(Op0I->getOperand(0), C);
4512 } else if (Op0I->getOpcode() == Instruction::Or) {
4513 // (X|C1)^C2 -> X^(C1|C2) iff X&~C1 == 0
4514 if (MaskedValueIsZero(Op0I->getOperand(0), Op0CI->getValue())) {
4515 Constant *NewRHS = ConstantExpr::getOr(Op0CI, RHS);
4516 // Anything in both C1 and C2 is known to be zero, remove it from
4518 Constant *CommonBits = And(Op0CI, RHS);
4519 NewRHS = ConstantExpr::getAnd(NewRHS,
4520 ConstantExpr::getNot(CommonBits));
4521 AddToWorkList(Op0I);
4522 I.setOperand(0, Op0I->getOperand(0));
4523 I.setOperand(1, NewRHS);
4530 // Try to fold constant and into select arguments.
4531 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
4532 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
4534 if (isa<PHINode>(Op0))
4535 if (Instruction *NV = FoldOpIntoPhi(I))
4539 if (Value *X = dyn_castNotVal(Op0)) // ~A ^ A == -1
4541 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4543 if (Value *X = dyn_castNotVal(Op1)) // A ^ ~A == -1
4545 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4548 BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1);
4551 if (match(Op1I, m_Or(m_Value(A), m_Value(B)))) {
4552 if (A == Op0) { // B^(B|A) == (A|B)^B
4553 Op1I->swapOperands();
4555 std::swap(Op0, Op1);
4556 } else if (B == Op0) { // B^(A|B) == (A|B)^B
4557 I.swapOperands(); // Simplified below.
4558 std::swap(Op0, Op1);
4560 } else if (match(Op1I, m_Xor(m_Value(A), m_Value(B)))) {
4561 if (Op0 == A) // A^(A^B) == B
4562 return ReplaceInstUsesWith(I, B);
4563 else if (Op0 == B) // A^(B^A) == B
4564 return ReplaceInstUsesWith(I, A);
4565 } else if (match(Op1I, m_And(m_Value(A), m_Value(B))) && Op1I->hasOneUse()){
4566 if (A == Op0) { // A^(A&B) -> A^(B&A)
4567 Op1I->swapOperands();
4570 if (B == Op0) { // A^(B&A) -> (B&A)^A
4571 I.swapOperands(); // Simplified below.
4572 std::swap(Op0, Op1);
4577 BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0);
4580 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) && Op0I->hasOneUse()) {
4581 if (A == Op1) // (B|A)^B == (A|B)^B
4583 if (B == Op1) { // (A|B)^B == A & ~B
4585 InsertNewInstBefore(BinaryOperator::CreateNot(Op1, "tmp"), I);
4586 return BinaryOperator::CreateAnd(A, NotB);
4588 } else if (match(Op0I, m_Xor(m_Value(A), m_Value(B)))) {
4589 if (Op1 == A) // (A^B)^A == B
4590 return ReplaceInstUsesWith(I, B);
4591 else if (Op1 == B) // (B^A)^A == B
4592 return ReplaceInstUsesWith(I, A);
4593 } else if (match(Op0I, m_And(m_Value(A), m_Value(B))) && Op0I->hasOneUse()){
4594 if (A == Op1) // (A&B)^A -> (B&A)^A
4596 if (B == Op1 && // (B&A)^A == ~B & A
4597 !isa<ConstantInt>(Op1)) { // Canonical form is (B&C)^C
4599 InsertNewInstBefore(BinaryOperator::CreateNot(A, "tmp"), I);
4600 return BinaryOperator::CreateAnd(N, Op1);
4605 // (X >> Z) ^ (Y >> Z) -> (X^Y) >> Z for all shifts.
4606 if (Op0I && Op1I && Op0I->isShift() &&
4607 Op0I->getOpcode() == Op1I->getOpcode() &&
4608 Op0I->getOperand(1) == Op1I->getOperand(1) &&
4609 (Op1I->hasOneUse() || Op1I->hasOneUse())) {
4610 Instruction *NewOp =
4611 InsertNewInstBefore(BinaryOperator::CreateXor(Op0I->getOperand(0),
4612 Op1I->getOperand(0),
4613 Op0I->getName()), I);
4614 return BinaryOperator::Create(Op1I->getOpcode(), NewOp,
4615 Op1I->getOperand(1));
4619 Value *A, *B, *C, *D;
4620 // (A & B)^(A | B) -> A ^ B
4621 if (match(Op0I, m_And(m_Value(A), m_Value(B))) &&
4622 match(Op1I, m_Or(m_Value(C), m_Value(D)))) {
4623 if ((A == C && B == D) || (A == D && B == C))
4624 return BinaryOperator::CreateXor(A, B);
4626 // (A | B)^(A & B) -> A ^ B
4627 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) &&
4628 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
4629 if ((A == C && B == D) || (A == D && B == C))
4630 return BinaryOperator::CreateXor(A, B);
4634 if ((Op0I->hasOneUse() || Op1I->hasOneUse()) &&
4635 match(Op0I, m_And(m_Value(A), m_Value(B))) &&
4636 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
4637 // (X & Y)^(X & Y) -> (Y^Z) & X
4638 Value *X = 0, *Y = 0, *Z = 0;
4640 X = A, Y = B, Z = D;
4642 X = A, Y = B, Z = C;
4644 X = B, Y = A, Z = D;
4646 X = B, Y = A, Z = C;
4649 Instruction *NewOp =
4650 InsertNewInstBefore(BinaryOperator::CreateXor(Y, Z, Op0->getName()), I);
4651 return BinaryOperator::CreateAnd(NewOp, X);
4656 // (icmp1 A, B) ^ (icmp2 A, B) --> (icmp3 A, B)
4657 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1)))
4658 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
4661 // fold (xor (cast A), (cast B)) -> (cast (xor A, B))
4662 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
4663 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4664 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind?
4665 const Type *SrcTy = Op0C->getOperand(0)->getType();
4666 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
4667 // Only do this if the casts both really cause code to be generated.
4668 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4670 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4672 Instruction *NewOp = BinaryOperator::CreateXor(Op0C->getOperand(0),
4673 Op1C->getOperand(0),
4675 InsertNewInstBefore(NewOp, I);
4676 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
4681 return Changed ? &I : 0;
4684 /// AddWithOverflow - Compute Result = In1+In2, returning true if the result
4685 /// overflowed for this type.
4686 static bool AddWithOverflow(ConstantInt *&Result, ConstantInt *In1,
4687 ConstantInt *In2, bool IsSigned = false) {
4688 Result = cast<ConstantInt>(Add(In1, In2));
4691 if (In2->getValue().isNegative())
4692 return Result->getValue().sgt(In1->getValue());
4694 return Result->getValue().slt(In1->getValue());
4696 return Result->getValue().ult(In1->getValue());
4699 /// SubWithOverflow - Compute Result = In1-In2, returning true if the result
4700 /// overflowed for this type.
4701 static bool SubWithOverflow(ConstantInt *&Result, ConstantInt *In1,
4702 ConstantInt *In2, bool IsSigned = false) {
4703 Result = cast<ConstantInt>(Subtract(In1, In2));
4706 if (In2->getValue().isNegative())
4707 return Result->getValue().slt(In1->getValue());
4709 return Result->getValue().sgt(In1->getValue());
4711 return Result->getValue().ugt(In1->getValue());
4714 /// EmitGEPOffset - Given a getelementptr instruction/constantexpr, emit the
4715 /// code necessary to compute the offset from the base pointer (without adding
4716 /// in the base pointer). Return the result as a signed integer of intptr size.
4717 static Value *EmitGEPOffset(User *GEP, Instruction &I, InstCombiner &IC) {
4718 TargetData &TD = IC.getTargetData();
4719 gep_type_iterator GTI = gep_type_begin(GEP);
4720 const Type *IntPtrTy = TD.getIntPtrType();
4721 Value *Result = Constant::getNullValue(IntPtrTy);
4723 // Build a mask for high order bits.
4724 unsigned IntPtrWidth = TD.getPointerSizeInBits();
4725 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth);
4727 for (User::op_iterator i = GEP->op_begin() + 1, e = GEP->op_end(); i != e;
4730 uint64_t Size = TD.getABITypeSize(GTI.getIndexedType()) & PtrSizeMask;
4731 if (ConstantInt *OpC = dyn_cast<ConstantInt>(Op)) {
4732 if (OpC->isZero()) continue;
4734 // Handle a struct index, which adds its field offset to the pointer.
4735 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
4736 Size = TD.getStructLayout(STy)->getElementOffset(OpC->getZExtValue());
4738 if (ConstantInt *RC = dyn_cast<ConstantInt>(Result))
4739 Result = ConstantInt::get(RC->getValue() + APInt(IntPtrWidth, Size));
4741 Result = IC.InsertNewInstBefore(
4742 BinaryOperator::CreateAdd(Result,
4743 ConstantInt::get(IntPtrTy, Size),
4744 GEP->getName()+".offs"), I);
4748 Constant *Scale = ConstantInt::get(IntPtrTy, Size);
4749 Constant *OC = ConstantExpr::getIntegerCast(OpC, IntPtrTy, true /*SExt*/);
4750 Scale = ConstantExpr::getMul(OC, Scale);
4751 if (Constant *RC = dyn_cast<Constant>(Result))
4752 Result = ConstantExpr::getAdd(RC, Scale);
4754 // Emit an add instruction.
4755 Result = IC.InsertNewInstBefore(
4756 BinaryOperator::CreateAdd(Result, Scale,
4757 GEP->getName()+".offs"), I);
4761 // Convert to correct type.
4762 if (Op->getType() != IntPtrTy) {
4763 if (Constant *OpC = dyn_cast<Constant>(Op))
4764 Op = ConstantExpr::getSExt(OpC, IntPtrTy);
4766 Op = IC.InsertNewInstBefore(new SExtInst(Op, IntPtrTy,
4767 Op->getName()+".c"), I);
4770 Constant *Scale = ConstantInt::get(IntPtrTy, Size);
4771 if (Constant *OpC = dyn_cast<Constant>(Op))
4772 Op = ConstantExpr::getMul(OpC, Scale);
4773 else // We'll let instcombine(mul) convert this to a shl if possible.
4774 Op = IC.InsertNewInstBefore(BinaryOperator::CreateMul(Op, Scale,
4775 GEP->getName()+".idx"), I);
4778 // Emit an add instruction.
4779 if (isa<Constant>(Op) && isa<Constant>(Result))
4780 Result = ConstantExpr::getAdd(cast<Constant>(Op),
4781 cast<Constant>(Result));
4783 Result = IC.InsertNewInstBefore(BinaryOperator::CreateAdd(Op, Result,
4784 GEP->getName()+".offs"), I);
4790 /// EvaluateGEPOffsetExpression - Return an value that can be used to compare of
4791 /// the *offset* implied by GEP to zero. For example, if we have &A[i], we want
4792 /// to return 'i' for "icmp ne i, 0". Note that, in general, indices can be
4793 /// complex, and scales are involved. The above expression would also be legal
4794 /// to codegen as "icmp ne (i*4), 0" (assuming A is a pointer to i32). This
4795 /// later form is less amenable to optimization though, and we are allowed to
4796 /// generate the first by knowing that pointer arithmetic doesn't overflow.
4798 /// If we can't emit an optimized form for this expression, this returns null.
4800 static Value *EvaluateGEPOffsetExpression(User *GEP, Instruction &I,
4802 TargetData &TD = IC.getTargetData();
4803 gep_type_iterator GTI = gep_type_begin(GEP);
4805 // Check to see if this gep only has a single variable index. If so, and if
4806 // any constant indices are a multiple of its scale, then we can compute this
4807 // in terms of the scale of the variable index. For example, if the GEP
4808 // implies an offset of "12 + i*4", then we can codegen this as "3 + i",
4809 // because the expression will cross zero at the same point.
4810 unsigned i, e = GEP->getNumOperands();
4812 for (i = 1; i != e; ++i, ++GTI) {
4813 if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i))) {
4814 // Compute the aggregate offset of constant indices.
4815 if (CI->isZero()) continue;
4817 // Handle a struct index, which adds its field offset to the pointer.
4818 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
4819 Offset += TD.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
4821 uint64_t Size = TD.getABITypeSize(GTI.getIndexedType());
4822 Offset += Size*CI->getSExtValue();
4825 // Found our variable index.
4830 // If there are no variable indices, we must have a constant offset, just
4831 // evaluate it the general way.
4832 if (i == e) return 0;
4834 Value *VariableIdx = GEP->getOperand(i);
4835 // Determine the scale factor of the variable element. For example, this is
4836 // 4 if the variable index is into an array of i32.
4837 uint64_t VariableScale = TD.getABITypeSize(GTI.getIndexedType());
4839 // Verify that there are no other variable indices. If so, emit the hard way.
4840 for (++i, ++GTI; i != e; ++i, ++GTI) {
4841 ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i));
4844 // Compute the aggregate offset of constant indices.
4845 if (CI->isZero()) continue;
4847 // Handle a struct index, which adds its field offset to the pointer.
4848 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
4849 Offset += TD.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
4851 uint64_t Size = TD.getABITypeSize(GTI.getIndexedType());
4852 Offset += Size*CI->getSExtValue();
4856 // Okay, we know we have a single variable index, which must be a
4857 // pointer/array/vector index. If there is no offset, life is simple, return
4859 unsigned IntPtrWidth = TD.getPointerSizeInBits();
4861 // Cast to intptrty in case a truncation occurs. If an extension is needed,
4862 // we don't need to bother extending: the extension won't affect where the
4863 // computation crosses zero.
4864 if (VariableIdx->getType()->getPrimitiveSizeInBits() > IntPtrWidth)
4865 VariableIdx = new TruncInst(VariableIdx, TD.getIntPtrType(),
4866 VariableIdx->getNameStart(), &I);
4870 // Otherwise, there is an index. The computation we will do will be modulo
4871 // the pointer size, so get it.
4872 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth);
4874 Offset &= PtrSizeMask;
4875 VariableScale &= PtrSizeMask;
4877 // To do this transformation, any constant index must be a multiple of the
4878 // variable scale factor. For example, we can evaluate "12 + 4*i" as "3 + i",
4879 // but we can't evaluate "10 + 3*i" in terms of i. Check that the offset is a
4880 // multiple of the variable scale.
4881 int64_t NewOffs = Offset / (int64_t)VariableScale;
4882 if (Offset != NewOffs*(int64_t)VariableScale)
4885 // Okay, we can do this evaluation. Start by converting the index to intptr.
4886 const Type *IntPtrTy = TD.getIntPtrType();
4887 if (VariableIdx->getType() != IntPtrTy)
4888 VariableIdx = CastInst::CreateIntegerCast(VariableIdx, IntPtrTy,
4890 VariableIdx->getNameStart(), &I);
4891 Constant *OffsetVal = ConstantInt::get(IntPtrTy, NewOffs);
4892 return BinaryOperator::CreateAdd(VariableIdx, OffsetVal, "offset", &I);
4896 /// FoldGEPICmp - Fold comparisons between a GEP instruction and something
4897 /// else. At this point we know that the GEP is on the LHS of the comparison.
4898 Instruction *InstCombiner::FoldGEPICmp(User *GEPLHS, Value *RHS,
4899 ICmpInst::Predicate Cond,
4901 assert(dyn_castGetElementPtr(GEPLHS) && "LHS is not a getelementptr!");
4903 // Look through bitcasts.
4904 if (BitCastInst *BCI = dyn_cast<BitCastInst>(RHS))
4905 RHS = BCI->getOperand(0);
4907 Value *PtrBase = GEPLHS->getOperand(0);
4908 if (PtrBase == RHS) {
4909 // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
4910 // This transformation (ignoring the base and scales) is valid because we
4911 // know pointers can't overflow. See if we can output an optimized form.
4912 Value *Offset = EvaluateGEPOffsetExpression(GEPLHS, I, *this);
4914 // If not, synthesize the offset the hard way.
4916 Offset = EmitGEPOffset(GEPLHS, I, *this);
4917 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Offset,
4918 Constant::getNullValue(Offset->getType()));
4919 } else if (User *GEPRHS = dyn_castGetElementPtr(RHS)) {
4920 // If the base pointers are different, but the indices are the same, just
4921 // compare the base pointer.
4922 if (PtrBase != GEPRHS->getOperand(0)) {
4923 bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands();
4924 IndicesTheSame &= GEPLHS->getOperand(0)->getType() ==
4925 GEPRHS->getOperand(0)->getType();
4927 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
4928 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
4929 IndicesTheSame = false;
4933 // If all indices are the same, just compare the base pointers.
4935 return new ICmpInst(ICmpInst::getSignedPredicate(Cond),
4936 GEPLHS->getOperand(0), GEPRHS->getOperand(0));
4938 // Otherwise, the base pointers are different and the indices are
4939 // different, bail out.
4943 // If one of the GEPs has all zero indices, recurse.
4944 bool AllZeros = true;
4945 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
4946 if (!isa<Constant>(GEPLHS->getOperand(i)) ||
4947 !cast<Constant>(GEPLHS->getOperand(i))->isNullValue()) {
4952 return FoldGEPICmp(GEPRHS, GEPLHS->getOperand(0),
4953 ICmpInst::getSwappedPredicate(Cond), I);
4955 // If the other GEP has all zero indices, recurse.
4957 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
4958 if (!isa<Constant>(GEPRHS->getOperand(i)) ||
4959 !cast<Constant>(GEPRHS->getOperand(i))->isNullValue()) {
4964 return FoldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I);
4966 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
4967 // If the GEPs only differ by one index, compare it.
4968 unsigned NumDifferences = 0; // Keep track of # differences.
4969 unsigned DiffOperand = 0; // The operand that differs.
4970 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
4971 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
4972 if (GEPLHS->getOperand(i)->getType()->getPrimitiveSizeInBits() !=
4973 GEPRHS->getOperand(i)->getType()->getPrimitiveSizeInBits()) {
4974 // Irreconcilable differences.
4978 if (NumDifferences++) break;
4983 if (NumDifferences == 0) // SAME GEP?
4984 return ReplaceInstUsesWith(I, // No comparison is needed here.
4985 ConstantInt::get(Type::Int1Ty,
4986 ICmpInst::isTrueWhenEqual(Cond)));
4988 else if (NumDifferences == 1) {
4989 Value *LHSV = GEPLHS->getOperand(DiffOperand);
4990 Value *RHSV = GEPRHS->getOperand(DiffOperand);
4991 // Make sure we do a signed comparison here.
4992 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), LHSV, RHSV);
4996 // Only lower this if the icmp is the only user of the GEP or if we expect
4997 // the result to fold to a constant!
4998 if ((isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
4999 (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
5000 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
5001 Value *L = EmitGEPOffset(GEPLHS, I, *this);
5002 Value *R = EmitGEPOffset(GEPRHS, I, *this);
5003 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), L, R);
5009 /// FoldFCmp_IntToFP_Cst - Fold fcmp ([us]itofp x, cst) if possible.
5011 Instruction *InstCombiner::FoldFCmp_IntToFP_Cst(FCmpInst &I,
5014 if (!isa<ConstantFP>(RHSC)) return 0;
5015 const APFloat &RHS = cast<ConstantFP>(RHSC)->getValueAPF();
5017 // Get the width of the mantissa. We don't want to hack on conversions that
5018 // might lose information from the integer, e.g. "i64 -> float"
5019 int MantissaWidth = LHSI->getType()->getFPMantissaWidth();
5020 if (MantissaWidth == -1) return 0; // Unknown.
5022 // Check to see that the input is converted from an integer type that is small
5023 // enough that preserves all bits. TODO: check here for "known" sign bits.
5024 // This would allow us to handle (fptosi (x >>s 62) to float) if x is i64 f.e.
5025 unsigned InputSize = LHSI->getOperand(0)->getType()->getPrimitiveSizeInBits();
5027 // If this is a uitofp instruction, we need an extra bit to hold the sign.
5028 if (isa<UIToFPInst>(LHSI))
5031 // If the conversion would lose info, don't hack on this.
5032 if ((int)InputSize > MantissaWidth)
5035 // Otherwise, we can potentially simplify the comparison. We know that it
5036 // will always come through as an integer value and we know the constant is
5037 // not a NAN (it would have been previously simplified).
5038 assert(!RHS.isNaN() && "NaN comparison not already folded!");
5040 ICmpInst::Predicate Pred;
5041 switch (I.getPredicate()) {
5042 default: assert(0 && "Unexpected predicate!");
5043 case FCmpInst::FCMP_UEQ:
5044 case FCmpInst::FCMP_OEQ: Pred = ICmpInst::ICMP_EQ; break;
5045 case FCmpInst::FCMP_UGT:
5046 case FCmpInst::FCMP_OGT: Pred = ICmpInst::ICMP_SGT; break;
5047 case FCmpInst::FCMP_UGE:
5048 case FCmpInst::FCMP_OGE: Pred = ICmpInst::ICMP_SGE; break;
5049 case FCmpInst::FCMP_ULT:
5050 case FCmpInst::FCMP_OLT: Pred = ICmpInst::ICMP_SLT; break;
5051 case FCmpInst::FCMP_ULE:
5052 case FCmpInst::FCMP_OLE: Pred = ICmpInst::ICMP_SLE; break;
5053 case FCmpInst::FCMP_UNE:
5054 case FCmpInst::FCMP_ONE: Pred = ICmpInst::ICMP_NE; break;
5055 case FCmpInst::FCMP_ORD:
5056 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
5057 case FCmpInst::FCMP_UNO:
5058 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
5061 const IntegerType *IntTy = cast<IntegerType>(LHSI->getOperand(0)->getType());
5063 // Now we know that the APFloat is a normal number, zero or inf.
5065 // See if the FP constant is too large for the integer. For example,
5066 // comparing an i8 to 300.0.
5067 unsigned IntWidth = IntTy->getPrimitiveSizeInBits();
5069 // If the RHS value is > SignedMax, fold the comparison. This handles +INF
5070 // and large values.
5071 APFloat SMax(RHS.getSemantics(), APFloat::fcZero, false);
5072 SMax.convertFromAPInt(APInt::getSignedMaxValue(IntWidth), true,
5073 APFloat::rmNearestTiesToEven);
5074 if (SMax.compare(RHS) == APFloat::cmpLessThan) { // smax < 13123.0
5075 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SLT ||
5076 Pred == ICmpInst::ICMP_SLE)
5077 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
5078 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
5081 // See if the RHS value is < SignedMin.
5082 APFloat SMin(RHS.getSemantics(), APFloat::fcZero, false);
5083 SMin.convertFromAPInt(APInt::getSignedMinValue(IntWidth), true,
5084 APFloat::rmNearestTiesToEven);
5085 if (SMin.compare(RHS) == APFloat::cmpGreaterThan) { // smin > 12312.0
5086 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SGT ||
5087 Pred == ICmpInst::ICMP_SGE)
5088 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
5089 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
5092 // Okay, now we know that the FP constant fits in the range [SMIN, SMAX] but
5093 // it may still be fractional. See if it is fractional by casting the FP
5094 // value to the integer value and back, checking for equality. Don't do this
5095 // for zero, because -0.0 is not fractional.
5096 Constant *RHSInt = ConstantExpr::getFPToSI(RHSC, IntTy);
5097 if (!RHS.isZero() &&
5098 ConstantExpr::getSIToFP(RHSInt, RHSC->getType()) != RHSC) {
5099 // If we had a comparison against a fractional value, we have to adjust
5100 // the compare predicate and sometimes the value. RHSC is rounded towards
5101 // zero at this point.
5103 default: assert(0 && "Unexpected integer comparison!");
5104 case ICmpInst::ICMP_NE: // (float)int != 4.4 --> true
5105 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
5106 case ICmpInst::ICMP_EQ: // (float)int == 4.4 --> false
5107 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
5108 case ICmpInst::ICMP_SLE:
5109 // (float)int <= 4.4 --> int <= 4
5110 // (float)int <= -4.4 --> int < -4
5111 if (RHS.isNegative())
5112 Pred = ICmpInst::ICMP_SLT;
5114 case ICmpInst::ICMP_SLT:
5115 // (float)int < -4.4 --> int < -4
5116 // (float)int < 4.4 --> int <= 4
5117 if (!RHS.isNegative())
5118 Pred = ICmpInst::ICMP_SLE;
5120 case ICmpInst::ICMP_SGT:
5121 // (float)int > 4.4 --> int > 4
5122 // (float)int > -4.4 --> int >= -4
5123 if (RHS.isNegative())
5124 Pred = ICmpInst::ICMP_SGE;
5126 case ICmpInst::ICMP_SGE:
5127 // (float)int >= -4.4 --> int >= -4
5128 // (float)int >= 4.4 --> int > 4
5129 if (!RHS.isNegative())
5130 Pred = ICmpInst::ICMP_SGT;
5135 // Lower this FP comparison into an appropriate integer version of the
5137 return new ICmpInst(Pred, LHSI->getOperand(0), RHSInt);
5140 Instruction *InstCombiner::visitFCmpInst(FCmpInst &I) {
5141 bool Changed = SimplifyCompare(I);
5142 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5144 // Fold trivial predicates.
5145 if (I.getPredicate() == FCmpInst::FCMP_FALSE)
5146 return ReplaceInstUsesWith(I, Constant::getNullValue(Type::Int1Ty));
5147 if (I.getPredicate() == FCmpInst::FCMP_TRUE)
5148 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
5150 // Simplify 'fcmp pred X, X'
5152 switch (I.getPredicate()) {
5153 default: assert(0 && "Unknown predicate!");
5154 case FCmpInst::FCMP_UEQ: // True if unordered or equal
5155 case FCmpInst::FCMP_UGE: // True if unordered, greater than, or equal
5156 case FCmpInst::FCMP_ULE: // True if unordered, less than, or equal
5157 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
5158 case FCmpInst::FCMP_OGT: // True if ordered and greater than
5159 case FCmpInst::FCMP_OLT: // True if ordered and less than
5160 case FCmpInst::FCMP_ONE: // True if ordered and operands are unequal
5161 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
5163 case FCmpInst::FCMP_UNO: // True if unordered: isnan(X) | isnan(Y)
5164 case FCmpInst::FCMP_ULT: // True if unordered or less than
5165 case FCmpInst::FCMP_UGT: // True if unordered or greater than
5166 case FCmpInst::FCMP_UNE: // True if unordered or not equal
5167 // Canonicalize these to be 'fcmp uno %X, 0.0'.
5168 I.setPredicate(FCmpInst::FCMP_UNO);
5169 I.setOperand(1, Constant::getNullValue(Op0->getType()));
5172 case FCmpInst::FCMP_ORD: // True if ordered (no nans)
5173 case FCmpInst::FCMP_OEQ: // True if ordered and equal
5174 case FCmpInst::FCMP_OGE: // True if ordered and greater than or equal
5175 case FCmpInst::FCMP_OLE: // True if ordered and less than or equal
5176 // Canonicalize these to be 'fcmp ord %X, 0.0'.
5177 I.setPredicate(FCmpInst::FCMP_ORD);
5178 I.setOperand(1, Constant::getNullValue(Op0->getType()));
5183 if (isa<UndefValue>(Op1)) // fcmp pred X, undef -> undef
5184 return ReplaceInstUsesWith(I, UndefValue::get(Type::Int1Ty));
5186 // Handle fcmp with constant RHS
5187 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
5188 // If the constant is a nan, see if we can fold the comparison based on it.
5189 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHSC)) {
5190 if (CFP->getValueAPF().isNaN()) {
5191 if (FCmpInst::isOrdered(I.getPredicate())) // True if ordered and...
5192 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
5193 assert(FCmpInst::isUnordered(I.getPredicate()) &&
5194 "Comparison must be either ordered or unordered!");
5195 // True if unordered.
5196 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
5200 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
5201 switch (LHSI->getOpcode()) {
5202 case Instruction::PHI:
5203 // Only fold fcmp into the PHI if the phi and fcmp are in the same
5204 // block. If in the same block, we're encouraging jump threading. If
5205 // not, we are just pessimizing the code by making an i1 phi.
5206 if (LHSI->getParent() == I.getParent())
5207 if (Instruction *NV = FoldOpIntoPhi(I))
5210 case Instruction::SIToFP:
5211 case Instruction::UIToFP:
5212 if (Instruction *NV = FoldFCmp_IntToFP_Cst(I, LHSI, RHSC))
5215 case Instruction::Select:
5216 // If either operand of the select is a constant, we can fold the
5217 // comparison into the select arms, which will cause one to be
5218 // constant folded and the select turned into a bitwise or.
5219 Value *Op1 = 0, *Op2 = 0;
5220 if (LHSI->hasOneUse()) {
5221 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
5222 // Fold the known value into the constant operand.
5223 Op1 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
5224 // Insert a new FCmp of the other select operand.
5225 Op2 = InsertNewInstBefore(new FCmpInst(I.getPredicate(),
5226 LHSI->getOperand(2), RHSC,
5228 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
5229 // Fold the known value into the constant operand.
5230 Op2 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
5231 // Insert a new FCmp of the other select operand.
5232 Op1 = InsertNewInstBefore(new FCmpInst(I.getPredicate(),
5233 LHSI->getOperand(1), RHSC,
5239 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
5244 return Changed ? &I : 0;
5247 Instruction *InstCombiner::visitICmpInst(ICmpInst &I) {
5248 bool Changed = SimplifyCompare(I);
5249 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5250 const Type *Ty = Op0->getType();
5254 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
5255 I.isTrueWhenEqual()));
5257 if (isa<UndefValue>(Op1)) // X icmp undef -> undef
5258 return ReplaceInstUsesWith(I, UndefValue::get(Type::Int1Ty));
5260 // icmp <global/alloca*/null>, <global/alloca*/null> - Global/Stack value
5261 // addresses never equal each other! We already know that Op0 != Op1.
5262 if ((isa<GlobalValue>(Op0) || isa<AllocaInst>(Op0) ||
5263 isa<ConstantPointerNull>(Op0)) &&
5264 (isa<GlobalValue>(Op1) || isa<AllocaInst>(Op1) ||
5265 isa<ConstantPointerNull>(Op1)))
5266 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
5267 !I.isTrueWhenEqual()));
5269 // icmp's with boolean values can always be turned into bitwise operations
5270 if (Ty == Type::Int1Ty) {
5271 switch (I.getPredicate()) {
5272 default: assert(0 && "Invalid icmp instruction!");
5273 case ICmpInst::ICMP_EQ: { // icmp eq i1 A, B -> ~(A^B)
5274 Instruction *Xor = BinaryOperator::CreateXor(Op0, Op1, I.getName()+"tmp");
5275 InsertNewInstBefore(Xor, I);
5276 return BinaryOperator::CreateNot(Xor);
5278 case ICmpInst::ICMP_NE: // icmp eq i1 A, B -> A^B
5279 return BinaryOperator::CreateXor(Op0, Op1);
5281 case ICmpInst::ICMP_UGT:
5282 std::swap(Op0, Op1); // Change icmp ugt -> icmp ult
5284 case ICmpInst::ICMP_ULT:{ // icmp ult i1 A, B -> ~A & B
5285 Instruction *Not = BinaryOperator::CreateNot(Op0, I.getName()+"tmp");
5286 InsertNewInstBefore(Not, I);
5287 return BinaryOperator::CreateAnd(Not, Op1);
5289 case ICmpInst::ICMP_SGT:
5290 std::swap(Op0, Op1); // Change icmp sgt -> icmp slt
5292 case ICmpInst::ICMP_SLT: { // icmp slt i1 A, B -> A & ~B
5293 Instruction *Not = BinaryOperator::CreateNot(Op1, I.getName()+"tmp");
5294 InsertNewInstBefore(Not, I);
5295 return BinaryOperator::CreateAnd(Not, Op0);
5297 case ICmpInst::ICMP_UGE:
5298 std::swap(Op0, Op1); // Change icmp uge -> icmp ule
5300 case ICmpInst::ICMP_ULE: { // icmp ule i1 A, B -> ~A | B
5301 Instruction *Not = BinaryOperator::CreateNot(Op0, I.getName()+"tmp");
5302 InsertNewInstBefore(Not, I);
5303 return BinaryOperator::CreateOr(Not, Op1);
5305 case ICmpInst::ICMP_SGE:
5306 std::swap(Op0, Op1); // Change icmp sge -> icmp sle
5308 case ICmpInst::ICMP_SLE: { // icmp sle i1 A, B -> A | ~B
5309 Instruction *Not = BinaryOperator::CreateNot(Op1, I.getName()+"tmp");
5310 InsertNewInstBefore(Not, I);
5311 return BinaryOperator::CreateOr(Not, Op0);
5316 // See if we are doing a comparison with a constant.
5317 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
5320 // (icmp ne/eq (sub A B) 0) -> (icmp ne/eq A, B)
5321 if (I.isEquality() && CI->isNullValue() &&
5322 match(Op0, m_Sub(m_Value(A), m_Value(B)))) {
5323 // (icmp cond A B) if cond is equality
5324 return new ICmpInst(I.getPredicate(), A, B);
5327 // If we have an icmp le or icmp ge instruction, turn it into the
5328 // appropriate icmp lt or icmp gt instruction. This allows us to rely on
5329 // them being folded in the code below.
5330 switch (I.getPredicate()) {
5332 case ICmpInst::ICMP_ULE:
5333 if (CI->isMaxValue(false)) // A <=u MAX -> TRUE
5334 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5335 return new ICmpInst(ICmpInst::ICMP_ULT, Op0, AddOne(CI));
5336 case ICmpInst::ICMP_SLE:
5337 if (CI->isMaxValue(true)) // A <=s MAX -> TRUE
5338 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5339 return new ICmpInst(ICmpInst::ICMP_SLT, Op0, AddOne(CI));
5340 case ICmpInst::ICMP_UGE:
5341 if (CI->isMinValue(false)) // A >=u MIN -> TRUE
5342 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5343 return new ICmpInst( ICmpInst::ICMP_UGT, Op0, SubOne(CI));
5344 case ICmpInst::ICMP_SGE:
5345 if (CI->isMinValue(true)) // A >=s MIN -> TRUE
5346 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5347 return new ICmpInst(ICmpInst::ICMP_SGT, Op0, SubOne(CI));
5350 // See if we can fold the comparison based on range information we can get
5351 // by checking whether bits are known to be zero or one in the input.
5352 uint32_t BitWidth = cast<IntegerType>(Ty)->getBitWidth();
5353 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
5355 // If this comparison is a normal comparison, it demands all
5356 // bits, if it is a sign bit comparison, it only demands the sign bit.
5358 bool isSignBit = isSignBitCheck(I.getPredicate(), CI, UnusedBit);
5360 if (SimplifyDemandedBits(Op0,
5361 isSignBit ? APInt::getSignBit(BitWidth)
5362 : APInt::getAllOnesValue(BitWidth),
5363 KnownZero, KnownOne, 0))
5366 // Given the known and unknown bits, compute a range that the LHS could be
5367 // in. Compute the Min, Max and RHS values based on the known bits. For the
5368 // EQ and NE we use unsigned values.
5369 APInt Min(BitWidth, 0), Max(BitWidth, 0);
5370 if (ICmpInst::isSignedPredicate(I.getPredicate()))
5371 ComputeSignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne, Min, Max);
5373 ComputeUnsignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne,Min,Max);
5375 // If Min and Max are known to be the same, then SimplifyDemandedBits
5376 // figured out that the LHS is a constant. Just constant fold this now so
5377 // that code below can assume that Min != Max.
5379 return ReplaceInstUsesWith(I, ConstantExpr::getICmp(I.getPredicate(),
5380 ConstantInt::get(Min),
5383 // Based on the range information we know about the LHS, see if we can
5384 // simplify this comparison. For example, (x&4) < 8 is always true.
5385 const APInt &RHSVal = CI->getValue();
5386 switch (I.getPredicate()) { // LE/GE have been folded already.
5387 default: assert(0 && "Unknown icmp opcode!");
5388 case ICmpInst::ICMP_EQ:
5389 if (Max.ult(RHSVal) || Min.ugt(RHSVal))
5390 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5392 case ICmpInst::ICMP_NE:
5393 if (Max.ult(RHSVal) || Min.ugt(RHSVal))
5394 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5396 case ICmpInst::ICMP_ULT:
5397 if (Max.ult(RHSVal)) // A <u C -> true iff max(A) < C
5398 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5399 if (Min.uge(RHSVal)) // A <u C -> false iff min(A) >= C
5400 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5401 if (RHSVal == Max) // A <u MAX -> A != MAX
5402 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5403 if (RHSVal == Min+1) // A <u MIN+1 -> A == MIN
5404 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, SubOne(CI));
5406 // (x <u 2147483648) -> (x >s -1) -> true if sign bit clear
5407 if (CI->isMinValue(true))
5408 return new ICmpInst(ICmpInst::ICMP_SGT, Op0,
5409 ConstantInt::getAllOnesValue(Op0->getType()));
5411 case ICmpInst::ICMP_UGT:
5412 if (Min.ugt(RHSVal)) // A >u C -> true iff min(A) > C
5413 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5414 if (Max.ule(RHSVal)) // A >u C -> false iff max(A) <= C
5415 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5417 if (RHSVal == Min) // A >u MIN -> A != MIN
5418 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5419 if (RHSVal == Max-1) // A >u MAX-1 -> A == MAX
5420 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, AddOne(CI));
5422 // (x >u 2147483647) -> (x <s 0) -> true if sign bit set
5423 if (CI->isMaxValue(true))
5424 return new ICmpInst(ICmpInst::ICMP_SLT, Op0,
5425 ConstantInt::getNullValue(Op0->getType()));
5427 case ICmpInst::ICMP_SLT:
5428 if (Max.slt(RHSVal)) // A <s C -> true iff max(A) < C
5429 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5430 if (Min.sge(RHSVal)) // A <s C -> false iff min(A) >= C
5431 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5432 if (RHSVal == Max) // A <s MAX -> A != MAX
5433 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5434 if (RHSVal == Min+1) // A <s MIN+1 -> A == MIN
5435 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, SubOne(CI));
5437 case ICmpInst::ICMP_SGT:
5438 if (Min.sgt(RHSVal)) // A >s C -> true iff min(A) > C
5439 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5440 if (Max.sle(RHSVal)) // A >s C -> false iff max(A) <= C
5441 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5443 if (RHSVal == Min) // A >s MIN -> A != MIN
5444 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5445 if (RHSVal == Max-1) // A >s MAX-1 -> A == MAX
5446 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, AddOne(CI));
5451 // Test if the ICmpInst instruction is used exclusively by a select as
5452 // part of a minimum or maximum operation. If so, refrain from doing
5453 // any other folding. This helps out other analyses which understand
5454 // non-obfuscated minimum and maximum idioms, such as ScalarEvolution
5455 // and CodeGen. And in this case, at least one of the comparison
5456 // operands has at least one user besides the compare (the select),
5457 // which would often largely negate the benefit of folding anyway.
5459 if (SelectInst *SI = dyn_cast<SelectInst>(*I.use_begin()))
5460 if ((SI->getOperand(1) == Op0 && SI->getOperand(2) == Op1) ||
5461 (SI->getOperand(2) == Op0 && SI->getOperand(1) == Op1))
5464 // See if we are doing a comparison between a constant and an instruction that
5465 // can be folded into the comparison.
5466 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
5467 // Since the RHS is a ConstantInt (CI), if the left hand side is an
5468 // instruction, see if that instruction also has constants so that the
5469 // instruction can be folded into the icmp
5470 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
5471 if (Instruction *Res = visitICmpInstWithInstAndIntCst(I, LHSI, CI))
5475 // Handle icmp with constant (but not simple integer constant) RHS
5476 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
5477 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
5478 switch (LHSI->getOpcode()) {
5479 case Instruction::GetElementPtr:
5480 if (RHSC->isNullValue()) {
5481 // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
5482 bool isAllZeros = true;
5483 for (unsigned i = 1, e = LHSI->getNumOperands(); i != e; ++i)
5484 if (!isa<Constant>(LHSI->getOperand(i)) ||
5485 !cast<Constant>(LHSI->getOperand(i))->isNullValue()) {
5490 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
5491 Constant::getNullValue(LHSI->getOperand(0)->getType()));
5495 case Instruction::PHI:
5496 // Only fold icmp into the PHI if the phi and fcmp are in the same
5497 // block. If in the same block, we're encouraging jump threading. If
5498 // not, we are just pessimizing the code by making an i1 phi.
5499 if (LHSI->getParent() == I.getParent())
5500 if (Instruction *NV = FoldOpIntoPhi(I))
5503 case Instruction::Select: {
5504 // If either operand of the select is a constant, we can fold the
5505 // comparison into the select arms, which will cause one to be
5506 // constant folded and the select turned into a bitwise or.
5507 Value *Op1 = 0, *Op2 = 0;
5508 if (LHSI->hasOneUse()) {
5509 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
5510 // Fold the known value into the constant operand.
5511 Op1 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
5512 // Insert a new ICmp of the other select operand.
5513 Op2 = InsertNewInstBefore(new ICmpInst(I.getPredicate(),
5514 LHSI->getOperand(2), RHSC,
5516 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
5517 // Fold the known value into the constant operand.
5518 Op2 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
5519 // Insert a new ICmp of the other select operand.
5520 Op1 = InsertNewInstBefore(new ICmpInst(I.getPredicate(),
5521 LHSI->getOperand(1), RHSC,
5527 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
5530 case Instruction::Malloc:
5531 // If we have (malloc != null), and if the malloc has a single use, we
5532 // can assume it is successful and remove the malloc.
5533 if (LHSI->hasOneUse() && isa<ConstantPointerNull>(RHSC)) {
5534 AddToWorkList(LHSI);
5535 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
5536 !I.isTrueWhenEqual()));
5542 // If we can optimize a 'icmp GEP, P' or 'icmp P, GEP', do so now.
5543 if (User *GEP = dyn_castGetElementPtr(Op0))
5544 if (Instruction *NI = FoldGEPICmp(GEP, Op1, I.getPredicate(), I))
5546 if (User *GEP = dyn_castGetElementPtr(Op1))
5547 if (Instruction *NI = FoldGEPICmp(GEP, Op0,
5548 ICmpInst::getSwappedPredicate(I.getPredicate()), I))
5551 // Test to see if the operands of the icmp are casted versions of other
5552 // values. If the ptr->ptr cast can be stripped off both arguments, we do so
5554 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op0)) {
5555 if (isa<PointerType>(Op0->getType()) &&
5556 (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
5557 // We keep moving the cast from the left operand over to the right
5558 // operand, where it can often be eliminated completely.
5559 Op0 = CI->getOperand(0);
5561 // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast
5562 // so eliminate it as well.
5563 if (BitCastInst *CI2 = dyn_cast<BitCastInst>(Op1))
5564 Op1 = CI2->getOperand(0);
5566 // If Op1 is a constant, we can fold the cast into the constant.
5567 if (Op0->getType() != Op1->getType()) {
5568 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
5569 Op1 = ConstantExpr::getBitCast(Op1C, Op0->getType());
5571 // Otherwise, cast the RHS right before the icmp
5572 Op1 = InsertBitCastBefore(Op1, Op0->getType(), I);
5575 return new ICmpInst(I.getPredicate(), Op0, Op1);
5579 if (isa<CastInst>(Op0)) {
5580 // Handle the special case of: icmp (cast bool to X), <cst>
5581 // This comes up when you have code like
5584 // For generality, we handle any zero-extension of any operand comparison
5585 // with a constant or another cast from the same type.
5586 if (isa<ConstantInt>(Op1) || isa<CastInst>(Op1))
5587 if (Instruction *R = visitICmpInstWithCastAndCast(I))
5591 // See if it's the same type of instruction on the left and right.
5592 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
5593 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
5594 if (Op0I->getOpcode() == Op1I->getOpcode() && Op0I->hasOneUse() &&
5595 Op1I->hasOneUse() && Op0I->getOperand(1) == Op1I->getOperand(1) &&
5597 switch (Op0I->getOpcode()) {
5599 case Instruction::Add:
5600 case Instruction::Sub:
5601 case Instruction::Xor:
5602 // a+x icmp eq/ne b+x --> a icmp b
5603 return new ICmpInst(I.getPredicate(), Op0I->getOperand(0),
5604 Op1I->getOperand(0));
5606 case Instruction::Mul:
5607 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) {
5608 // a * Cst icmp eq/ne b * Cst --> a & Mask icmp b & Mask
5609 // Mask = -1 >> count-trailing-zeros(Cst).
5610 if (!CI->isZero() && !CI->isOne()) {
5611 const APInt &AP = CI->getValue();
5612 ConstantInt *Mask = ConstantInt::get(
5613 APInt::getLowBitsSet(AP.getBitWidth(),
5615 AP.countTrailingZeros()));
5616 Instruction *And1 = BinaryOperator::CreateAnd(Op0I->getOperand(0),
5618 Instruction *And2 = BinaryOperator::CreateAnd(Op1I->getOperand(0),
5620 InsertNewInstBefore(And1, I);
5621 InsertNewInstBefore(And2, I);
5622 return new ICmpInst(I.getPredicate(), And1, And2);
5631 // ~x < ~y --> y < x
5633 if (match(Op0, m_Not(m_Value(A))) &&
5634 match(Op1, m_Not(m_Value(B))))
5635 return new ICmpInst(I.getPredicate(), B, A);
5638 if (I.isEquality()) {
5639 Value *A, *B, *C, *D;
5641 // -x == -y --> x == y
5642 if (match(Op0, m_Neg(m_Value(A))) &&
5643 match(Op1, m_Neg(m_Value(B))))
5644 return new ICmpInst(I.getPredicate(), A, B);
5646 if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
5647 if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0
5648 Value *OtherVal = A == Op1 ? B : A;
5649 return new ICmpInst(I.getPredicate(), OtherVal,
5650 Constant::getNullValue(A->getType()));
5653 if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) {
5654 // A^c1 == C^c2 --> A == C^(c1^c2)
5655 if (ConstantInt *C1 = dyn_cast<ConstantInt>(B))
5656 if (ConstantInt *C2 = dyn_cast<ConstantInt>(D))
5657 if (Op1->hasOneUse()) {
5658 Constant *NC = ConstantInt::get(C1->getValue() ^ C2->getValue());
5659 Instruction *Xor = BinaryOperator::CreateXor(C, NC, "tmp");
5660 return new ICmpInst(I.getPredicate(), A,
5661 InsertNewInstBefore(Xor, I));
5664 // A^B == A^D -> B == D
5665 if (A == C) return new ICmpInst(I.getPredicate(), B, D);
5666 if (A == D) return new ICmpInst(I.getPredicate(), B, C);
5667 if (B == C) return new ICmpInst(I.getPredicate(), A, D);
5668 if (B == D) return new ICmpInst(I.getPredicate(), A, C);
5672 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
5673 (A == Op0 || B == Op0)) {
5674 // A == (A^B) -> B == 0
5675 Value *OtherVal = A == Op0 ? B : A;
5676 return new ICmpInst(I.getPredicate(), OtherVal,
5677 Constant::getNullValue(A->getType()));
5679 if (match(Op0, m_Sub(m_Value(A), m_Value(B))) && A == Op1) {
5680 // (A-B) == A -> B == 0
5681 return new ICmpInst(I.getPredicate(), B,
5682 Constant::getNullValue(B->getType()));
5684 if (match(Op1, m_Sub(m_Value(A), m_Value(B))) && A == Op0) {
5685 // A == (A-B) -> B == 0
5686 return new ICmpInst(I.getPredicate(), B,
5687 Constant::getNullValue(B->getType()));
5690 // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
5691 if (Op0->hasOneUse() && Op1->hasOneUse() &&
5692 match(Op0, m_And(m_Value(A), m_Value(B))) &&
5693 match(Op1, m_And(m_Value(C), m_Value(D)))) {
5694 Value *X = 0, *Y = 0, *Z = 0;
5697 X = B; Y = D; Z = A;
5698 } else if (A == D) {
5699 X = B; Y = C; Z = A;
5700 } else if (B == C) {
5701 X = A; Y = D; Z = B;
5702 } else if (B == D) {
5703 X = A; Y = C; Z = B;
5706 if (X) { // Build (X^Y) & Z
5707 Op1 = InsertNewInstBefore(BinaryOperator::CreateXor(X, Y, "tmp"), I);
5708 Op1 = InsertNewInstBefore(BinaryOperator::CreateAnd(Op1, Z, "tmp"), I);
5709 I.setOperand(0, Op1);
5710 I.setOperand(1, Constant::getNullValue(Op1->getType()));
5715 return Changed ? &I : 0;
5719 /// FoldICmpDivCst - Fold "icmp pred, ([su]div X, DivRHS), CmpRHS" where DivRHS
5720 /// and CmpRHS are both known to be integer constants.
5721 Instruction *InstCombiner::FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
5722 ConstantInt *DivRHS) {
5723 ConstantInt *CmpRHS = cast<ConstantInt>(ICI.getOperand(1));
5724 const APInt &CmpRHSV = CmpRHS->getValue();
5726 // FIXME: If the operand types don't match the type of the divide
5727 // then don't attempt this transform. The code below doesn't have the
5728 // logic to deal with a signed divide and an unsigned compare (and
5729 // vice versa). This is because (x /s C1) <s C2 produces different
5730 // results than (x /s C1) <u C2 or (x /u C1) <s C2 or even
5731 // (x /u C1) <u C2. Simply casting the operands and result won't
5732 // work. :( The if statement below tests that condition and bails
5734 bool DivIsSigned = DivI->getOpcode() == Instruction::SDiv;
5735 if (!ICI.isEquality() && DivIsSigned != ICI.isSignedPredicate())
5737 if (DivRHS->isZero())
5738 return 0; // The ProdOV computation fails on divide by zero.
5740 // Compute Prod = CI * DivRHS. We are essentially solving an equation
5741 // of form X/C1=C2. We solve for X by multiplying C1 (DivRHS) and
5742 // C2 (CI). By solving for X we can turn this into a range check
5743 // instead of computing a divide.
5744 ConstantInt *Prod = Multiply(CmpRHS, DivRHS);
5746 // Determine if the product overflows by seeing if the product is
5747 // not equal to the divide. Make sure we do the same kind of divide
5748 // as in the LHS instruction that we're folding.
5749 bool ProdOV = (DivIsSigned ? ConstantExpr::getSDiv(Prod, DivRHS) :
5750 ConstantExpr::getUDiv(Prod, DivRHS)) != CmpRHS;
5752 // Get the ICmp opcode
5753 ICmpInst::Predicate Pred = ICI.getPredicate();
5755 // Figure out the interval that is being checked. For example, a comparison
5756 // like "X /u 5 == 0" is really checking that X is in the interval [0, 5).
5757 // Compute this interval based on the constants involved and the signedness of
5758 // the compare/divide. This computes a half-open interval, keeping track of
5759 // whether either value in the interval overflows. After analysis each
5760 // overflow variable is set to 0 if it's corresponding bound variable is valid
5761 // -1 if overflowed off the bottom end, or +1 if overflowed off the top end.
5762 int LoOverflow = 0, HiOverflow = 0;
5763 ConstantInt *LoBound = 0, *HiBound = 0;
5766 if (!DivIsSigned) { // udiv
5767 // e.g. X/5 op 3 --> [15, 20)
5769 HiOverflow = LoOverflow = ProdOV;
5771 HiOverflow = AddWithOverflow(HiBound, LoBound, DivRHS, false);
5772 } else if (DivRHS->getValue().isStrictlyPositive()) { // Divisor is > 0.
5773 if (CmpRHSV == 0) { // (X / pos) op 0
5774 // Can't overflow. e.g. X/2 op 0 --> [-1, 2)
5775 LoBound = cast<ConstantInt>(ConstantExpr::getNeg(SubOne(DivRHS)));
5777 } else if (CmpRHSV.isStrictlyPositive()) { // (X / pos) op pos
5778 LoBound = Prod; // e.g. X/5 op 3 --> [15, 20)
5779 HiOverflow = LoOverflow = ProdOV;
5781 HiOverflow = AddWithOverflow(HiBound, Prod, DivRHS, true);
5782 } else { // (X / pos) op neg
5783 // e.g. X/5 op -3 --> [-15-4, -15+1) --> [-19, -14)
5784 Constant *DivRHSH = ConstantExpr::getNeg(SubOne(DivRHS));
5785 LoOverflow = AddWithOverflow(LoBound, Prod,
5786 cast<ConstantInt>(DivRHSH), true) ? -1 : 0;
5787 HiBound = AddOne(Prod);
5788 HiOverflow = ProdOV ? -1 : 0;
5790 } else if (DivRHS->getValue().isNegative()) { // Divisor is < 0.
5791 if (CmpRHSV == 0) { // (X / neg) op 0
5792 // e.g. X/-5 op 0 --> [-4, 5)
5793 LoBound = AddOne(DivRHS);
5794 HiBound = cast<ConstantInt>(ConstantExpr::getNeg(DivRHS));
5795 if (HiBound == DivRHS) { // -INTMIN = INTMIN
5796 HiOverflow = 1; // [INTMIN+1, overflow)
5797 HiBound = 0; // e.g. X/INTMIN = 0 --> X > INTMIN
5799 } else if (CmpRHSV.isStrictlyPositive()) { // (X / neg) op pos
5800 // e.g. X/-5 op 3 --> [-19, -14)
5801 HiOverflow = LoOverflow = ProdOV ? -1 : 0;
5803 LoOverflow = AddWithOverflow(LoBound, Prod, AddOne(DivRHS), true) ?-1:0;
5804 HiBound = AddOne(Prod);
5805 } else { // (X / neg) op neg
5806 // e.g. X/-5 op -3 --> [15, 20)
5808 LoOverflow = HiOverflow = ProdOV ? 1 : 0;
5810 HiOverflow = SubWithOverflow(HiBound, Prod, DivRHS, true);
5813 // Dividing by a negative swaps the condition. LT <-> GT
5814 Pred = ICmpInst::getSwappedPredicate(Pred);
5817 Value *X = DivI->getOperand(0);
5819 default: assert(0 && "Unhandled icmp opcode!");
5820 case ICmpInst::ICMP_EQ:
5821 if (LoOverflow && HiOverflow)
5822 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5823 else if (HiOverflow)
5824 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
5825 ICmpInst::ICMP_UGE, X, LoBound);
5826 else if (LoOverflow)
5827 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
5828 ICmpInst::ICMP_ULT, X, HiBound);
5830 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned, true, ICI);
5831 case ICmpInst::ICMP_NE:
5832 if (LoOverflow && HiOverflow)
5833 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5834 else if (HiOverflow)
5835 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
5836 ICmpInst::ICMP_ULT, X, LoBound);
5837 else if (LoOverflow)
5838 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
5839 ICmpInst::ICMP_UGE, X, HiBound);
5841 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned, false, ICI);
5842 case ICmpInst::ICMP_ULT:
5843 case ICmpInst::ICMP_SLT:
5844 if (LoOverflow == +1) // Low bound is greater than input range.
5845 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5846 if (LoOverflow == -1) // Low bound is less than input range.
5847 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5848 return new ICmpInst(Pred, X, LoBound);
5849 case ICmpInst::ICMP_UGT:
5850 case ICmpInst::ICMP_SGT:
5851 if (HiOverflow == +1) // High bound greater than input range.
5852 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5853 else if (HiOverflow == -1) // High bound less than input range.
5854 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5855 if (Pred == ICmpInst::ICMP_UGT)
5856 return new ICmpInst(ICmpInst::ICMP_UGE, X, HiBound);
5858 return new ICmpInst(ICmpInst::ICMP_SGE, X, HiBound);
5863 /// visitICmpInstWithInstAndIntCst - Handle "icmp (instr, intcst)".
5865 Instruction *InstCombiner::visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
5868 const APInt &RHSV = RHS->getValue();
5870 switch (LHSI->getOpcode()) {
5871 case Instruction::Xor: // (icmp pred (xor X, XorCST), CI)
5872 if (ConstantInt *XorCST = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
5873 // If this is a comparison that tests the signbit (X < 0) or (x > -1),
5875 if ((ICI.getPredicate() == ICmpInst::ICMP_SLT && RHSV == 0) ||
5876 (ICI.getPredicate() == ICmpInst::ICMP_SGT && RHSV.isAllOnesValue())) {
5877 Value *CompareVal = LHSI->getOperand(0);
5879 // If the sign bit of the XorCST is not set, there is no change to
5880 // the operation, just stop using the Xor.
5881 if (!XorCST->getValue().isNegative()) {
5882 ICI.setOperand(0, CompareVal);
5883 AddToWorkList(LHSI);
5887 // Was the old condition true if the operand is positive?
5888 bool isTrueIfPositive = ICI.getPredicate() == ICmpInst::ICMP_SGT;
5890 // If so, the new one isn't.
5891 isTrueIfPositive ^= true;
5893 if (isTrueIfPositive)
5894 return new ICmpInst(ICmpInst::ICMP_SGT, CompareVal, SubOne(RHS));
5896 return new ICmpInst(ICmpInst::ICMP_SLT, CompareVal, AddOne(RHS));
5900 case Instruction::And: // (icmp pred (and X, AndCST), RHS)
5901 if (LHSI->hasOneUse() && isa<ConstantInt>(LHSI->getOperand(1)) &&
5902 LHSI->getOperand(0)->hasOneUse()) {
5903 ConstantInt *AndCST = cast<ConstantInt>(LHSI->getOperand(1));
5905 // If the LHS is an AND of a truncating cast, we can widen the
5906 // and/compare to be the input width without changing the value
5907 // produced, eliminating a cast.
5908 if (TruncInst *Cast = dyn_cast<TruncInst>(LHSI->getOperand(0))) {
5909 // We can do this transformation if either the AND constant does not
5910 // have its sign bit set or if it is an equality comparison.
5911 // Extending a relational comparison when we're checking the sign
5912 // bit would not work.
5913 if (Cast->hasOneUse() &&
5914 (ICI.isEquality() ||
5915 (AndCST->getValue().isNonNegative() && RHSV.isNonNegative()))) {
5917 cast<IntegerType>(Cast->getOperand(0)->getType())->getBitWidth();
5918 APInt NewCST = AndCST->getValue();
5919 NewCST.zext(BitWidth);
5921 NewCI.zext(BitWidth);
5922 Instruction *NewAnd =
5923 BinaryOperator::CreateAnd(Cast->getOperand(0),
5924 ConstantInt::get(NewCST),LHSI->getName());
5925 InsertNewInstBefore(NewAnd, ICI);
5926 return new ICmpInst(ICI.getPredicate(), NewAnd,
5927 ConstantInt::get(NewCI));
5931 // If this is: (X >> C1) & C2 != C3 (where any shift and any compare
5932 // could exist), turn it into (X & (C2 << C1)) != (C3 << C1). This
5933 // happens a LOT in code produced by the C front-end, for bitfield
5935 BinaryOperator *Shift = dyn_cast<BinaryOperator>(LHSI->getOperand(0));
5936 if (Shift && !Shift->isShift())
5940 ShAmt = Shift ? dyn_cast<ConstantInt>(Shift->getOperand(1)) : 0;
5941 const Type *Ty = Shift ? Shift->getType() : 0; // Type of the shift.
5942 const Type *AndTy = AndCST->getType(); // Type of the and.
5944 // We can fold this as long as we can't shift unknown bits
5945 // into the mask. This can only happen with signed shift
5946 // rights, as they sign-extend.
5948 bool CanFold = Shift->isLogicalShift();
5950 // To test for the bad case of the signed shr, see if any
5951 // of the bits shifted in could be tested after the mask.
5952 uint32_t TyBits = Ty->getPrimitiveSizeInBits();
5953 int ShAmtVal = TyBits - ShAmt->getLimitedValue(TyBits);
5955 uint32_t BitWidth = AndTy->getPrimitiveSizeInBits();
5956 if ((APInt::getHighBitsSet(BitWidth, BitWidth-ShAmtVal) &
5957 AndCST->getValue()) == 0)
5963 if (Shift->getOpcode() == Instruction::Shl)
5964 NewCst = ConstantExpr::getLShr(RHS, ShAmt);
5966 NewCst = ConstantExpr::getShl(RHS, ShAmt);
5968 // Check to see if we are shifting out any of the bits being
5970 if (ConstantExpr::get(Shift->getOpcode(), NewCst, ShAmt) != RHS) {
5971 // If we shifted bits out, the fold is not going to work out.
5972 // As a special case, check to see if this means that the
5973 // result is always true or false now.
5974 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
5975 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5976 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
5977 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5979 ICI.setOperand(1, NewCst);
5980 Constant *NewAndCST;
5981 if (Shift->getOpcode() == Instruction::Shl)
5982 NewAndCST = ConstantExpr::getLShr(AndCST, ShAmt);
5984 NewAndCST = ConstantExpr::getShl(AndCST, ShAmt);
5985 LHSI->setOperand(1, NewAndCST);
5986 LHSI->setOperand(0, Shift->getOperand(0));
5987 AddToWorkList(Shift); // Shift is dead.
5988 AddUsesToWorkList(ICI);
5994 // Turn ((X >> Y) & C) == 0 into (X & (C << Y)) == 0. The later is
5995 // preferable because it allows the C<<Y expression to be hoisted out
5996 // of a loop if Y is invariant and X is not.
5997 if (Shift && Shift->hasOneUse() && RHSV == 0 &&
5998 ICI.isEquality() && !Shift->isArithmeticShift() &&
5999 isa<Instruction>(Shift->getOperand(0))) {
6002 if (Shift->getOpcode() == Instruction::LShr) {
6003 NS = BinaryOperator::CreateShl(AndCST,
6004 Shift->getOperand(1), "tmp");
6006 // Insert a logical shift.
6007 NS = BinaryOperator::CreateLShr(AndCST,
6008 Shift->getOperand(1), "tmp");
6010 InsertNewInstBefore(cast<Instruction>(NS), ICI);
6012 // Compute X & (C << Y).
6013 Instruction *NewAnd =
6014 BinaryOperator::CreateAnd(Shift->getOperand(0), NS, LHSI->getName());
6015 InsertNewInstBefore(NewAnd, ICI);
6017 ICI.setOperand(0, NewAnd);
6023 case Instruction::Shl: { // (icmp pred (shl X, ShAmt), CI)
6024 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
6027 uint32_t TypeBits = RHSV.getBitWidth();
6029 // Check that the shift amount is in range. If not, don't perform
6030 // undefined shifts. When the shift is visited it will be
6032 if (ShAmt->uge(TypeBits))
6035 if (ICI.isEquality()) {
6036 // If we are comparing against bits always shifted out, the
6037 // comparison cannot succeed.
6039 ConstantExpr::getShl(ConstantExpr::getLShr(RHS, ShAmt), ShAmt);
6040 if (Comp != RHS) {// Comparing against a bit that we know is zero.
6041 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
6042 Constant *Cst = ConstantInt::get(Type::Int1Ty, IsICMP_NE);
6043 return ReplaceInstUsesWith(ICI, Cst);
6046 if (LHSI->hasOneUse()) {
6047 // Otherwise strength reduce the shift into an and.
6048 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
6050 ConstantInt::get(APInt::getLowBitsSet(TypeBits, TypeBits-ShAmtVal));
6053 BinaryOperator::CreateAnd(LHSI->getOperand(0),
6054 Mask, LHSI->getName()+".mask");
6055 Value *And = InsertNewInstBefore(AndI, ICI);
6056 return new ICmpInst(ICI.getPredicate(), And,
6057 ConstantInt::get(RHSV.lshr(ShAmtVal)));
6061 // Otherwise, if this is a comparison of the sign bit, simplify to and/test.
6062 bool TrueIfSigned = false;
6063 if (LHSI->hasOneUse() &&
6064 isSignBitCheck(ICI.getPredicate(), RHS, TrueIfSigned)) {
6065 // (X << 31) <s 0 --> (X&1) != 0
6066 Constant *Mask = ConstantInt::get(APInt(TypeBits, 1) <<
6067 (TypeBits-ShAmt->getZExtValue()-1));
6069 BinaryOperator::CreateAnd(LHSI->getOperand(0),
6070 Mask, LHSI->getName()+".mask");
6071 Value *And = InsertNewInstBefore(AndI, ICI);
6073 return new ICmpInst(TrueIfSigned ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ,
6074 And, Constant::getNullValue(And->getType()));
6079 case Instruction::LShr: // (icmp pred (shr X, ShAmt), CI)
6080 case Instruction::AShr: {
6081 // Only handle equality comparisons of shift-by-constant.
6082 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
6083 if (!ShAmt || !ICI.isEquality()) break;
6085 // Check that the shift amount is in range. If not, don't perform
6086 // undefined shifts. When the shift is visited it will be
6088 uint32_t TypeBits = RHSV.getBitWidth();
6089 if (ShAmt->uge(TypeBits))
6092 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
6094 // If we are comparing against bits always shifted out, the
6095 // comparison cannot succeed.
6096 APInt Comp = RHSV << ShAmtVal;
6097 if (LHSI->getOpcode() == Instruction::LShr)
6098 Comp = Comp.lshr(ShAmtVal);
6100 Comp = Comp.ashr(ShAmtVal);
6102 if (Comp != RHSV) { // Comparing against a bit that we know is zero.
6103 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
6104 Constant *Cst = ConstantInt::get(Type::Int1Ty, IsICMP_NE);
6105 return ReplaceInstUsesWith(ICI, Cst);
6108 // Otherwise, check to see if the bits shifted out are known to be zero.
6109 // If so, we can compare against the unshifted value:
6110 // (X & 4) >> 1 == 2 --> (X & 4) == 4.
6111 if (LHSI->hasOneUse() &&
6112 MaskedValueIsZero(LHSI->getOperand(0),
6113 APInt::getLowBitsSet(Comp.getBitWidth(), ShAmtVal))) {
6114 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
6115 ConstantExpr::getShl(RHS, ShAmt));
6118 if (LHSI->hasOneUse()) {
6119 // Otherwise strength reduce the shift into an and.
6120 APInt Val(APInt::getHighBitsSet(TypeBits, TypeBits - ShAmtVal));
6121 Constant *Mask = ConstantInt::get(Val);
6124 BinaryOperator::CreateAnd(LHSI->getOperand(0),
6125 Mask, LHSI->getName()+".mask");
6126 Value *And = InsertNewInstBefore(AndI, ICI);
6127 return new ICmpInst(ICI.getPredicate(), And,
6128 ConstantExpr::getShl(RHS, ShAmt));
6133 case Instruction::SDiv:
6134 case Instruction::UDiv:
6135 // Fold: icmp pred ([us]div X, C1), C2 -> range test
6136 // Fold this div into the comparison, producing a range check.
6137 // Determine, based on the divide type, what the range is being
6138 // checked. If there is an overflow on the low or high side, remember
6139 // it, otherwise compute the range [low, hi) bounding the new value.
6140 // See: InsertRangeTest above for the kinds of replacements possible.
6141 if (ConstantInt *DivRHS = dyn_cast<ConstantInt>(LHSI->getOperand(1)))
6142 if (Instruction *R = FoldICmpDivCst(ICI, cast<BinaryOperator>(LHSI),
6147 case Instruction::Add:
6148 // Fold: icmp pred (add, X, C1), C2
6150 if (!ICI.isEquality()) {
6151 ConstantInt *LHSC = dyn_cast<ConstantInt>(LHSI->getOperand(1));
6153 const APInt &LHSV = LHSC->getValue();
6155 ConstantRange CR = ICI.makeConstantRange(ICI.getPredicate(), RHSV)
6158 if (ICI.isSignedPredicate()) {
6159 if (CR.getLower().isSignBit()) {
6160 return new ICmpInst(ICmpInst::ICMP_SLT, LHSI->getOperand(0),
6161 ConstantInt::get(CR.getUpper()));
6162 } else if (CR.getUpper().isSignBit()) {
6163 return new ICmpInst(ICmpInst::ICMP_SGE, LHSI->getOperand(0),
6164 ConstantInt::get(CR.getLower()));
6167 if (CR.getLower().isMinValue()) {
6168 return new ICmpInst(ICmpInst::ICMP_ULT, LHSI->getOperand(0),
6169 ConstantInt::get(CR.getUpper()));
6170 } else if (CR.getUpper().isMinValue()) {
6171 return new ICmpInst(ICmpInst::ICMP_UGE, LHSI->getOperand(0),
6172 ConstantInt::get(CR.getLower()));
6179 // Simplify icmp_eq and icmp_ne instructions with integer constant RHS.
6180 if (ICI.isEquality()) {
6181 bool isICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
6183 // If the first operand is (add|sub|and|or|xor|rem) with a constant, and
6184 // the second operand is a constant, simplify a bit.
6185 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(LHSI)) {
6186 switch (BO->getOpcode()) {
6187 case Instruction::SRem:
6188 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
6189 if (RHSV == 0 && isa<ConstantInt>(BO->getOperand(1)) &&BO->hasOneUse()){
6190 const APInt &V = cast<ConstantInt>(BO->getOperand(1))->getValue();
6191 if (V.sgt(APInt(V.getBitWidth(), 1)) && V.isPowerOf2()) {
6192 Instruction *NewRem =
6193 BinaryOperator::CreateURem(BO->getOperand(0), BO->getOperand(1),
6195 InsertNewInstBefore(NewRem, ICI);
6196 return new ICmpInst(ICI.getPredicate(), NewRem,
6197 Constant::getNullValue(BO->getType()));
6201 case Instruction::Add:
6202 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
6203 if (ConstantInt *BOp1C = dyn_cast<ConstantInt>(BO->getOperand(1))) {
6204 if (BO->hasOneUse())
6205 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
6206 Subtract(RHS, BOp1C));
6207 } else if (RHSV == 0) {
6208 // Replace ((add A, B) != 0) with (A != -B) if A or B is
6209 // efficiently invertible, or if the add has just this one use.
6210 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
6212 if (Value *NegVal = dyn_castNegVal(BOp1))
6213 return new ICmpInst(ICI.getPredicate(), BOp0, NegVal);
6214 else if (Value *NegVal = dyn_castNegVal(BOp0))
6215 return new ICmpInst(ICI.getPredicate(), NegVal, BOp1);
6216 else if (BO->hasOneUse()) {
6217 Instruction *Neg = BinaryOperator::CreateNeg(BOp1);
6218 InsertNewInstBefore(Neg, ICI);
6220 return new ICmpInst(ICI.getPredicate(), BOp0, Neg);
6224 case Instruction::Xor:
6225 // For the xor case, we can xor two constants together, eliminating
6226 // the explicit xor.
6227 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1)))
6228 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
6229 ConstantExpr::getXor(RHS, BOC));
6232 case Instruction::Sub:
6233 // Replace (([sub|xor] A, B) != 0) with (A != B)
6235 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
6239 case Instruction::Or:
6240 // If bits are being or'd in that are not present in the constant we
6241 // are comparing against, then the comparison could never succeed!
6242 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1))) {
6243 Constant *NotCI = ConstantExpr::getNot(RHS);
6244 if (!ConstantExpr::getAnd(BOC, NotCI)->isNullValue())
6245 return ReplaceInstUsesWith(ICI, ConstantInt::get(Type::Int1Ty,
6250 case Instruction::And:
6251 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
6252 // If bits are being compared against that are and'd out, then the
6253 // comparison can never succeed!
6254 if ((RHSV & ~BOC->getValue()) != 0)
6255 return ReplaceInstUsesWith(ICI, ConstantInt::get(Type::Int1Ty,
6258 // If we have ((X & C) == C), turn it into ((X & C) != 0).
6259 if (RHS == BOC && RHSV.isPowerOf2())
6260 return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ :
6261 ICmpInst::ICMP_NE, LHSI,
6262 Constant::getNullValue(RHS->getType()));
6264 // Replace (and X, (1 << size(X)-1) != 0) with x s< 0
6265 if (BOC->getValue().isSignBit()) {
6266 Value *X = BO->getOperand(0);
6267 Constant *Zero = Constant::getNullValue(X->getType());
6268 ICmpInst::Predicate pred = isICMP_NE ?
6269 ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
6270 return new ICmpInst(pred, X, Zero);
6273 // ((X & ~7) == 0) --> X < 8
6274 if (RHSV == 0 && isHighOnes(BOC)) {
6275 Value *X = BO->getOperand(0);
6276 Constant *NegX = ConstantExpr::getNeg(BOC);
6277 ICmpInst::Predicate pred = isICMP_NE ?
6278 ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
6279 return new ICmpInst(pred, X, NegX);
6284 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(LHSI)) {
6285 // Handle icmp {eq|ne} <intrinsic>, intcst.
6286 if (II->getIntrinsicID() == Intrinsic::bswap) {
6288 ICI.setOperand(0, II->getOperand(1));
6289 ICI.setOperand(1, ConstantInt::get(RHSV.byteSwap()));
6293 } else { // Not a ICMP_EQ/ICMP_NE
6294 // If the LHS is a cast from an integral value of the same size,
6295 // then since we know the RHS is a constant, try to simlify.
6296 if (CastInst *Cast = dyn_cast<CastInst>(LHSI)) {
6297 Value *CastOp = Cast->getOperand(0);
6298 const Type *SrcTy = CastOp->getType();
6299 uint32_t SrcTySize = SrcTy->getPrimitiveSizeInBits();
6300 if (SrcTy->isInteger() &&
6301 SrcTySize == Cast->getType()->getPrimitiveSizeInBits()) {
6302 // If this is an unsigned comparison, try to make the comparison use
6303 // smaller constant values.
6304 if (ICI.getPredicate() == ICmpInst::ICMP_ULT && RHSV.isSignBit()) {
6305 // X u< 128 => X s> -1
6306 return new ICmpInst(ICmpInst::ICMP_SGT, CastOp,
6307 ConstantInt::get(APInt::getAllOnesValue(SrcTySize)));
6308 } else if (ICI.getPredicate() == ICmpInst::ICMP_UGT &&
6309 RHSV == APInt::getSignedMaxValue(SrcTySize)) {
6310 // X u> 127 => X s< 0
6311 return new ICmpInst(ICmpInst::ICMP_SLT, CastOp,
6312 Constant::getNullValue(SrcTy));
6320 /// visitICmpInstWithCastAndCast - Handle icmp (cast x to y), (cast/cst).
6321 /// We only handle extending casts so far.
6323 Instruction *InstCombiner::visitICmpInstWithCastAndCast(ICmpInst &ICI) {
6324 const CastInst *LHSCI = cast<CastInst>(ICI.getOperand(0));
6325 Value *LHSCIOp = LHSCI->getOperand(0);
6326 const Type *SrcTy = LHSCIOp->getType();
6327 const Type *DestTy = LHSCI->getType();
6330 // Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the
6331 // integer type is the same size as the pointer type.
6332 if (LHSCI->getOpcode() == Instruction::PtrToInt &&
6333 getTargetData().getPointerSizeInBits() ==
6334 cast<IntegerType>(DestTy)->getBitWidth()) {
6336 if (Constant *RHSC = dyn_cast<Constant>(ICI.getOperand(1))) {
6337 RHSOp = ConstantExpr::getIntToPtr(RHSC, SrcTy);
6338 } else if (PtrToIntInst *RHSC = dyn_cast<PtrToIntInst>(ICI.getOperand(1))) {
6339 RHSOp = RHSC->getOperand(0);
6340 // If the pointer types don't match, insert a bitcast.
6341 if (LHSCIOp->getType() != RHSOp->getType())
6342 RHSOp = InsertBitCastBefore(RHSOp, LHSCIOp->getType(), ICI);
6346 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSOp);
6349 // The code below only handles extension cast instructions, so far.
6351 if (LHSCI->getOpcode() != Instruction::ZExt &&
6352 LHSCI->getOpcode() != Instruction::SExt)
6355 bool isSignedExt = LHSCI->getOpcode() == Instruction::SExt;
6356 bool isSignedCmp = ICI.isSignedPredicate();
6358 if (CastInst *CI = dyn_cast<CastInst>(ICI.getOperand(1))) {
6359 // Not an extension from the same type?
6360 RHSCIOp = CI->getOperand(0);
6361 if (RHSCIOp->getType() != LHSCIOp->getType())
6364 // If the signedness of the two casts doesn't agree (i.e. one is a sext
6365 // and the other is a zext), then we can't handle this.
6366 if (CI->getOpcode() != LHSCI->getOpcode())
6369 // Deal with equality cases early.
6370 if (ICI.isEquality())
6371 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
6373 // A signed comparison of sign extended values simplifies into a
6374 // signed comparison.
6375 if (isSignedCmp && isSignedExt)
6376 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
6378 // The other three cases all fold into an unsigned comparison.
6379 return new ICmpInst(ICI.getUnsignedPredicate(), LHSCIOp, RHSCIOp);
6382 // If we aren't dealing with a constant on the RHS, exit early
6383 ConstantInt *CI = dyn_cast<ConstantInt>(ICI.getOperand(1));
6387 // Compute the constant that would happen if we truncated to SrcTy then
6388 // reextended to DestTy.
6389 Constant *Res1 = ConstantExpr::getTrunc(CI, SrcTy);
6390 Constant *Res2 = ConstantExpr::getCast(LHSCI->getOpcode(), Res1, DestTy);
6392 // If the re-extended constant didn't change...
6394 // Make sure that sign of the Cmp and the sign of the Cast are the same.
6395 // For example, we might have:
6396 // %A = sext short %X to uint
6397 // %B = icmp ugt uint %A, 1330
6398 // It is incorrect to transform this into
6399 // %B = icmp ugt short %X, 1330
6400 // because %A may have negative value.
6402 // However, we allow this when the compare is EQ/NE, because they are
6404 if (isSignedExt == isSignedCmp || ICI.isEquality())
6405 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
6409 // The re-extended constant changed so the constant cannot be represented
6410 // in the shorter type. Consequently, we cannot emit a simple comparison.
6412 // First, handle some easy cases. We know the result cannot be equal at this
6413 // point so handle the ICI.isEquality() cases
6414 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
6415 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
6416 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
6417 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
6419 // Evaluate the comparison for LT (we invert for GT below). LE and GE cases
6420 // should have been folded away previously and not enter in here.
6423 // We're performing a signed comparison.
6424 if (cast<ConstantInt>(CI)->getValue().isNegative())
6425 Result = ConstantInt::getFalse(); // X < (small) --> false
6427 Result = ConstantInt::getTrue(); // X < (large) --> true
6429 // We're performing an unsigned comparison.
6431 // We're performing an unsigned comp with a sign extended value.
6432 // This is true if the input is >= 0. [aka >s -1]
6433 Constant *NegOne = ConstantInt::getAllOnesValue(SrcTy);
6434 Result = InsertNewInstBefore(new ICmpInst(ICmpInst::ICMP_SGT, LHSCIOp,
6435 NegOne, ICI.getName()), ICI);
6437 // Unsigned extend & unsigned compare -> always true.
6438 Result = ConstantInt::getTrue();
6442 // Finally, return the value computed.
6443 if (ICI.getPredicate() == ICmpInst::ICMP_ULT ||
6444 ICI.getPredicate() == ICmpInst::ICMP_SLT)
6445 return ReplaceInstUsesWith(ICI, Result);
6447 assert((ICI.getPredicate()==ICmpInst::ICMP_UGT ||
6448 ICI.getPredicate()==ICmpInst::ICMP_SGT) &&
6449 "ICmp should be folded!");
6450 if (Constant *CI = dyn_cast<Constant>(Result))
6451 return ReplaceInstUsesWith(ICI, ConstantExpr::getNot(CI));
6452 return BinaryOperator::CreateNot(Result);
6455 Instruction *InstCombiner::visitShl(BinaryOperator &I) {
6456 return commonShiftTransforms(I);
6459 Instruction *InstCombiner::visitLShr(BinaryOperator &I) {
6460 return commonShiftTransforms(I);
6463 Instruction *InstCombiner::visitAShr(BinaryOperator &I) {
6464 if (Instruction *R = commonShiftTransforms(I))
6467 Value *Op0 = I.getOperand(0);
6469 // ashr int -1, X = -1 (for any arithmetic shift rights of ~0)
6470 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
6471 if (CSI->isAllOnesValue())
6472 return ReplaceInstUsesWith(I, CSI);
6474 // See if we can turn a signed shr into an unsigned shr.
6475 if (!isa<VectorType>(I.getType()) &&
6476 MaskedValueIsZero(Op0,
6477 APInt::getSignBit(I.getType()->getPrimitiveSizeInBits())))
6478 return BinaryOperator::CreateLShr(Op0, I.getOperand(1));
6483 Instruction *InstCombiner::commonShiftTransforms(BinaryOperator &I) {
6484 assert(I.getOperand(1)->getType() == I.getOperand(0)->getType());
6485 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
6487 // shl X, 0 == X and shr X, 0 == X
6488 // shl 0, X == 0 and shr 0, X == 0
6489 if (Op1 == Constant::getNullValue(Op1->getType()) ||
6490 Op0 == Constant::getNullValue(Op0->getType()))
6491 return ReplaceInstUsesWith(I, Op0);
6493 if (isa<UndefValue>(Op0)) {
6494 if (I.getOpcode() == Instruction::AShr) // undef >>s X -> undef
6495 return ReplaceInstUsesWith(I, Op0);
6496 else // undef << X -> 0, undef >>u X -> 0
6497 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
6499 if (isa<UndefValue>(Op1)) {
6500 if (I.getOpcode() == Instruction::AShr) // X >>s undef -> X
6501 return ReplaceInstUsesWith(I, Op0);
6502 else // X << undef, X >>u undef -> 0
6503 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
6506 // Try to fold constant and into select arguments.
6507 if (isa<Constant>(Op0))
6508 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
6509 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
6512 if (ConstantInt *CUI = dyn_cast<ConstantInt>(Op1))
6513 if (Instruction *Res = FoldShiftByConstant(Op0, CUI, I))
6518 Instruction *InstCombiner::FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
6519 BinaryOperator &I) {
6520 bool isLeftShift = I.getOpcode() == Instruction::Shl;
6522 // See if we can simplify any instructions used by the instruction whose sole
6523 // purpose is to compute bits we don't care about.
6524 uint32_t TypeBits = Op0->getType()->getPrimitiveSizeInBits();
6525 APInt KnownZero(TypeBits, 0), KnownOne(TypeBits, 0);
6526 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(TypeBits),
6527 KnownZero, KnownOne))
6530 // shl uint X, 32 = 0 and shr ubyte Y, 9 = 0, ... just don't eliminate shr
6531 // of a signed value.
6533 if (Op1->uge(TypeBits)) {
6534 if (I.getOpcode() != Instruction::AShr)
6535 return ReplaceInstUsesWith(I, Constant::getNullValue(Op0->getType()));
6537 I.setOperand(1, ConstantInt::get(I.getType(), TypeBits-1));
6542 // ((X*C1) << C2) == (X * (C1 << C2))
6543 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0))
6544 if (BO->getOpcode() == Instruction::Mul && isLeftShift)
6545 if (Constant *BOOp = dyn_cast<Constant>(BO->getOperand(1)))
6546 return BinaryOperator::CreateMul(BO->getOperand(0),
6547 ConstantExpr::getShl(BOOp, Op1));
6549 // Try to fold constant and into select arguments.
6550 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
6551 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
6553 if (isa<PHINode>(Op0))
6554 if (Instruction *NV = FoldOpIntoPhi(I))
6557 // Fold shift2(trunc(shift1(x,c1)), c2) -> trunc(shift2(shift1(x,c1),c2))
6558 if (TruncInst *TI = dyn_cast<TruncInst>(Op0)) {
6559 Instruction *TrOp = dyn_cast<Instruction>(TI->getOperand(0));
6560 // If 'shift2' is an ashr, we would have to get the sign bit into a funny
6561 // place. Don't try to do this transformation in this case. Also, we
6562 // require that the input operand is a shift-by-constant so that we have
6563 // confidence that the shifts will get folded together. We could do this
6564 // xform in more cases, but it is unlikely to be profitable.
6565 if (TrOp && I.isLogicalShift() && TrOp->isShift() &&
6566 isa<ConstantInt>(TrOp->getOperand(1))) {
6567 // Okay, we'll do this xform. Make the shift of shift.
6568 Constant *ShAmt = ConstantExpr::getZExt(Op1, TrOp->getType());
6569 Instruction *NSh = BinaryOperator::Create(I.getOpcode(), TrOp, ShAmt,
6571 InsertNewInstBefore(NSh, I); // (shift2 (shift1 & 0x00FF), c2)
6573 // For logical shifts, the truncation has the effect of making the high
6574 // part of the register be zeros. Emulate this by inserting an AND to
6575 // clear the top bits as needed. This 'and' will usually be zapped by
6576 // other xforms later if dead.
6577 unsigned SrcSize = TrOp->getType()->getPrimitiveSizeInBits();
6578 unsigned DstSize = TI->getType()->getPrimitiveSizeInBits();
6579 APInt MaskV(APInt::getLowBitsSet(SrcSize, DstSize));
6581 // The mask we constructed says what the trunc would do if occurring
6582 // between the shifts. We want to know the effect *after* the second
6583 // shift. We know that it is a logical shift by a constant, so adjust the
6584 // mask as appropriate.
6585 if (I.getOpcode() == Instruction::Shl)
6586 MaskV <<= Op1->getZExtValue();
6588 assert(I.getOpcode() == Instruction::LShr && "Unknown logical shift");
6589 MaskV = MaskV.lshr(Op1->getZExtValue());
6592 Instruction *And = BinaryOperator::CreateAnd(NSh, ConstantInt::get(MaskV),
6594 InsertNewInstBefore(And, I); // shift1 & 0x00FF
6596 // Return the value truncated to the interesting size.
6597 return new TruncInst(And, I.getType());
6601 if (Op0->hasOneUse()) {
6602 if (BinaryOperator *Op0BO = dyn_cast<BinaryOperator>(Op0)) {
6603 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
6606 switch (Op0BO->getOpcode()) {
6608 case Instruction::Add:
6609 case Instruction::And:
6610 case Instruction::Or:
6611 case Instruction::Xor: {
6612 // These operators commute.
6613 // Turn (Y + (X >> C)) << C -> (X + (Y << C)) & (~0 << C)
6614 if (isLeftShift && Op0BO->getOperand(1)->hasOneUse() &&
6615 match(Op0BO->getOperand(1),
6616 m_Shr(m_Value(V1), m_ConstantInt(CC))) && CC == Op1) {
6617 Instruction *YS = BinaryOperator::CreateShl(
6618 Op0BO->getOperand(0), Op1,
6620 InsertNewInstBefore(YS, I); // (Y << C)
6622 BinaryOperator::Create(Op0BO->getOpcode(), YS, V1,
6623 Op0BO->getOperand(1)->getName());
6624 InsertNewInstBefore(X, I); // (X + (Y << C))
6625 uint32_t Op1Val = Op1->getLimitedValue(TypeBits);
6626 return BinaryOperator::CreateAnd(X, ConstantInt::get(
6627 APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
6630 // Turn (Y + ((X >> C) & CC)) << C -> ((X & (CC << C)) + (Y << C))
6631 Value *Op0BOOp1 = Op0BO->getOperand(1);
6632 if (isLeftShift && Op0BOOp1->hasOneUse() &&
6634 m_And(m_Shr(m_Value(V1), m_Value(V2)),m_ConstantInt(CC))) &&
6635 cast<BinaryOperator>(Op0BOOp1)->getOperand(0)->hasOneUse() &&
6637 Instruction *YS = BinaryOperator::CreateShl(
6638 Op0BO->getOperand(0), Op1,
6640 InsertNewInstBefore(YS, I); // (Y << C)
6642 BinaryOperator::CreateAnd(V1, ConstantExpr::getShl(CC, Op1),
6643 V1->getName()+".mask");
6644 InsertNewInstBefore(XM, I); // X & (CC << C)
6646 return BinaryOperator::Create(Op0BO->getOpcode(), YS, XM);
6651 case Instruction::Sub: {
6652 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
6653 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
6654 match(Op0BO->getOperand(0),
6655 m_Shr(m_Value(V1), m_ConstantInt(CC))) && CC == Op1) {
6656 Instruction *YS = BinaryOperator::CreateShl(
6657 Op0BO->getOperand(1), Op1,
6659 InsertNewInstBefore(YS, I); // (Y << C)
6661 BinaryOperator::Create(Op0BO->getOpcode(), V1, YS,
6662 Op0BO->getOperand(0)->getName());
6663 InsertNewInstBefore(X, I); // (X + (Y << C))
6664 uint32_t Op1Val = Op1->getLimitedValue(TypeBits);
6665 return BinaryOperator::CreateAnd(X, ConstantInt::get(
6666 APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
6669 // Turn (((X >> C)&CC) + Y) << C -> (X + (Y << C)) & (CC << C)
6670 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
6671 match(Op0BO->getOperand(0),
6672 m_And(m_Shr(m_Value(V1), m_Value(V2)),
6673 m_ConstantInt(CC))) && V2 == Op1 &&
6674 cast<BinaryOperator>(Op0BO->getOperand(0))
6675 ->getOperand(0)->hasOneUse()) {
6676 Instruction *YS = BinaryOperator::CreateShl(
6677 Op0BO->getOperand(1), Op1,
6679 InsertNewInstBefore(YS, I); // (Y << C)
6681 BinaryOperator::CreateAnd(V1, ConstantExpr::getShl(CC, Op1),
6682 V1->getName()+".mask");
6683 InsertNewInstBefore(XM, I); // X & (CC << C)
6685 return BinaryOperator::Create(Op0BO->getOpcode(), XM, YS);
6693 // If the operand is an bitwise operator with a constant RHS, and the
6694 // shift is the only use, we can pull it out of the shift.
6695 if (ConstantInt *Op0C = dyn_cast<ConstantInt>(Op0BO->getOperand(1))) {
6696 bool isValid = true; // Valid only for And, Or, Xor
6697 bool highBitSet = false; // Transform if high bit of constant set?
6699 switch (Op0BO->getOpcode()) {
6700 default: isValid = false; break; // Do not perform transform!
6701 case Instruction::Add:
6702 isValid = isLeftShift;
6704 case Instruction::Or:
6705 case Instruction::Xor:
6708 case Instruction::And:
6713 // If this is a signed shift right, and the high bit is modified
6714 // by the logical operation, do not perform the transformation.
6715 // The highBitSet boolean indicates the value of the high bit of
6716 // the constant which would cause it to be modified for this
6719 if (isValid && I.getOpcode() == Instruction::AShr)
6720 isValid = Op0C->getValue()[TypeBits-1] == highBitSet;
6723 Constant *NewRHS = ConstantExpr::get(I.getOpcode(), Op0C, Op1);
6725 Instruction *NewShift =
6726 BinaryOperator::Create(I.getOpcode(), Op0BO->getOperand(0), Op1);
6727 InsertNewInstBefore(NewShift, I);
6728 NewShift->takeName(Op0BO);
6730 return BinaryOperator::Create(Op0BO->getOpcode(), NewShift,
6737 // Find out if this is a shift of a shift by a constant.
6738 BinaryOperator *ShiftOp = dyn_cast<BinaryOperator>(Op0);
6739 if (ShiftOp && !ShiftOp->isShift())
6742 if (ShiftOp && isa<ConstantInt>(ShiftOp->getOperand(1))) {
6743 ConstantInt *ShiftAmt1C = cast<ConstantInt>(ShiftOp->getOperand(1));
6744 uint32_t ShiftAmt1 = ShiftAmt1C->getLimitedValue(TypeBits);
6745 uint32_t ShiftAmt2 = Op1->getLimitedValue(TypeBits);
6746 assert(ShiftAmt2 != 0 && "Should have been simplified earlier");
6747 if (ShiftAmt1 == 0) return 0; // Will be simplified in the future.
6748 Value *X = ShiftOp->getOperand(0);
6750 uint32_t AmtSum = ShiftAmt1+ShiftAmt2; // Fold into one big shift.
6751 if (AmtSum > TypeBits)
6754 const IntegerType *Ty = cast<IntegerType>(I.getType());
6756 // Check for (X << c1) << c2 and (X >> c1) >> c2
6757 if (I.getOpcode() == ShiftOp->getOpcode()) {
6758 return BinaryOperator::Create(I.getOpcode(), X,
6759 ConstantInt::get(Ty, AmtSum));
6760 } else if (ShiftOp->getOpcode() == Instruction::LShr &&
6761 I.getOpcode() == Instruction::AShr) {
6762 // ((X >>u C1) >>s C2) -> (X >>u (C1+C2)) since C1 != 0.
6763 return BinaryOperator::CreateLShr(X, ConstantInt::get(Ty, AmtSum));
6764 } else if (ShiftOp->getOpcode() == Instruction::AShr &&
6765 I.getOpcode() == Instruction::LShr) {
6766 // ((X >>s C1) >>u C2) -> ((X >>s (C1+C2)) & mask) since C1 != 0.
6767 Instruction *Shift =
6768 BinaryOperator::CreateAShr(X, ConstantInt::get(Ty, AmtSum));
6769 InsertNewInstBefore(Shift, I);
6771 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
6772 return BinaryOperator::CreateAnd(Shift, ConstantInt::get(Mask));
6775 // Okay, if we get here, one shift must be left, and the other shift must be
6776 // right. See if the amounts are equal.
6777 if (ShiftAmt1 == ShiftAmt2) {
6778 // If we have ((X >>? C) << C), turn this into X & (-1 << C).
6779 if (I.getOpcode() == Instruction::Shl) {
6780 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt1));
6781 return BinaryOperator::CreateAnd(X, ConstantInt::get(Mask));
6783 // If we have ((X << C) >>u C), turn this into X & (-1 >>u C).
6784 if (I.getOpcode() == Instruction::LShr) {
6785 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt1));
6786 return BinaryOperator::CreateAnd(X, ConstantInt::get(Mask));
6788 // We can simplify ((X << C) >>s C) into a trunc + sext.
6789 // NOTE: we could do this for any C, but that would make 'unusual' integer
6790 // types. For now, just stick to ones well-supported by the code
6792 const Type *SExtType = 0;
6793 switch (Ty->getBitWidth() - ShiftAmt1) {
6800 SExtType = IntegerType::get(Ty->getBitWidth() - ShiftAmt1);
6805 Instruction *NewTrunc = new TruncInst(X, SExtType, "sext");
6806 InsertNewInstBefore(NewTrunc, I);
6807 return new SExtInst(NewTrunc, Ty);
6809 // Otherwise, we can't handle it yet.
6810 } else if (ShiftAmt1 < ShiftAmt2) {
6811 uint32_t ShiftDiff = ShiftAmt2-ShiftAmt1;
6813 // (X >>? C1) << C2 --> X << (C2-C1) & (-1 << C2)
6814 if (I.getOpcode() == Instruction::Shl) {
6815 assert(ShiftOp->getOpcode() == Instruction::LShr ||
6816 ShiftOp->getOpcode() == Instruction::AShr);
6817 Instruction *Shift =
6818 BinaryOperator::CreateShl(X, ConstantInt::get(Ty, ShiftDiff));
6819 InsertNewInstBefore(Shift, I);
6821 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
6822 return BinaryOperator::CreateAnd(Shift, ConstantInt::get(Mask));
6825 // (X << C1) >>u C2 --> X >>u (C2-C1) & (-1 >> C2)
6826 if (I.getOpcode() == Instruction::LShr) {
6827 assert(ShiftOp->getOpcode() == Instruction::Shl);
6828 Instruction *Shift =
6829 BinaryOperator::CreateLShr(X, ConstantInt::get(Ty, ShiftDiff));
6830 InsertNewInstBefore(Shift, I);
6832 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
6833 return BinaryOperator::CreateAnd(Shift, ConstantInt::get(Mask));
6836 // We can't handle (X << C1) >>s C2, it shifts arbitrary bits in.
6838 assert(ShiftAmt2 < ShiftAmt1);
6839 uint32_t ShiftDiff = ShiftAmt1-ShiftAmt2;
6841 // (X >>? C1) << C2 --> X >>? (C1-C2) & (-1 << C2)
6842 if (I.getOpcode() == Instruction::Shl) {
6843 assert(ShiftOp->getOpcode() == Instruction::LShr ||
6844 ShiftOp->getOpcode() == Instruction::AShr);
6845 Instruction *Shift =
6846 BinaryOperator::Create(ShiftOp->getOpcode(), X,
6847 ConstantInt::get(Ty, ShiftDiff));
6848 InsertNewInstBefore(Shift, I);
6850 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
6851 return BinaryOperator::CreateAnd(Shift, ConstantInt::get(Mask));
6854 // (X << C1) >>u C2 --> X << (C1-C2) & (-1 >> C2)
6855 if (I.getOpcode() == Instruction::LShr) {
6856 assert(ShiftOp->getOpcode() == Instruction::Shl);
6857 Instruction *Shift =
6858 BinaryOperator::CreateShl(X, ConstantInt::get(Ty, ShiftDiff));
6859 InsertNewInstBefore(Shift, I);
6861 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
6862 return BinaryOperator::CreateAnd(Shift, ConstantInt::get(Mask));
6865 // We can't handle (X << C1) >>a C2, it shifts arbitrary bits in.
6872 /// DecomposeSimpleLinearExpr - Analyze 'Val', seeing if it is a simple linear
6873 /// expression. If so, decompose it, returning some value X, such that Val is
6876 static Value *DecomposeSimpleLinearExpr(Value *Val, unsigned &Scale,
6878 assert(Val->getType() == Type::Int32Ty && "Unexpected allocation size type!");
6879 if (ConstantInt *CI = dyn_cast<ConstantInt>(Val)) {
6880 Offset = CI->getZExtValue();
6882 return ConstantInt::get(Type::Int32Ty, 0);
6883 } else if (BinaryOperator *I = dyn_cast<BinaryOperator>(Val)) {
6884 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
6885 if (I->getOpcode() == Instruction::Shl) {
6886 // This is a value scaled by '1 << the shift amt'.
6887 Scale = 1U << RHS->getZExtValue();
6889 return I->getOperand(0);
6890 } else if (I->getOpcode() == Instruction::Mul) {
6891 // This value is scaled by 'RHS'.
6892 Scale = RHS->getZExtValue();
6894 return I->getOperand(0);
6895 } else if (I->getOpcode() == Instruction::Add) {
6896 // We have X+C. Check to see if we really have (X*C2)+C1,
6897 // where C1 is divisible by C2.
6900 DecomposeSimpleLinearExpr(I->getOperand(0), SubScale, Offset);
6901 Offset += RHS->getZExtValue();
6908 // Otherwise, we can't look past this.
6915 /// PromoteCastOfAllocation - If we find a cast of an allocation instruction,
6916 /// try to eliminate the cast by moving the type information into the alloc.
6917 Instruction *InstCombiner::PromoteCastOfAllocation(BitCastInst &CI,
6918 AllocationInst &AI) {
6919 const PointerType *PTy = cast<PointerType>(CI.getType());
6921 // Remove any uses of AI that are dead.
6922 assert(!CI.use_empty() && "Dead instructions should be removed earlier!");
6924 for (Value::use_iterator UI = AI.use_begin(), E = AI.use_end(); UI != E; ) {
6925 Instruction *User = cast<Instruction>(*UI++);
6926 if (isInstructionTriviallyDead(User)) {
6927 while (UI != E && *UI == User)
6928 ++UI; // If this instruction uses AI more than once, don't break UI.
6931 DOUT << "IC: DCE: " << *User;
6932 EraseInstFromFunction(*User);
6936 // Get the type really allocated and the type casted to.
6937 const Type *AllocElTy = AI.getAllocatedType();
6938 const Type *CastElTy = PTy->getElementType();
6939 if (!AllocElTy->isSized() || !CastElTy->isSized()) return 0;
6941 unsigned AllocElTyAlign = TD->getABITypeAlignment(AllocElTy);
6942 unsigned CastElTyAlign = TD->getABITypeAlignment(CastElTy);
6943 if (CastElTyAlign < AllocElTyAlign) return 0;
6945 // If the allocation has multiple uses, only promote it if we are strictly
6946 // increasing the alignment of the resultant allocation. If we keep it the
6947 // same, we open the door to infinite loops of various kinds.
6948 if (!AI.hasOneUse() && CastElTyAlign == AllocElTyAlign) return 0;
6950 uint64_t AllocElTySize = TD->getABITypeSize(AllocElTy);
6951 uint64_t CastElTySize = TD->getABITypeSize(CastElTy);
6952 if (CastElTySize == 0 || AllocElTySize == 0) return 0;
6954 // See if we can satisfy the modulus by pulling a scale out of the array
6956 unsigned ArraySizeScale;
6958 Value *NumElements = // See if the array size is a decomposable linear expr.
6959 DecomposeSimpleLinearExpr(AI.getOperand(0), ArraySizeScale, ArrayOffset);
6961 // If we can now satisfy the modulus, by using a non-1 scale, we really can
6963 if ((AllocElTySize*ArraySizeScale) % CastElTySize != 0 ||
6964 (AllocElTySize*ArrayOffset ) % CastElTySize != 0) return 0;
6966 unsigned Scale = (AllocElTySize*ArraySizeScale)/CastElTySize;
6971 // If the allocation size is constant, form a constant mul expression
6972 Amt = ConstantInt::get(Type::Int32Ty, Scale);
6973 if (isa<ConstantInt>(NumElements))
6974 Amt = Multiply(cast<ConstantInt>(NumElements), cast<ConstantInt>(Amt));
6975 // otherwise multiply the amount and the number of elements
6976 else if (Scale != 1) {
6977 Instruction *Tmp = BinaryOperator::CreateMul(Amt, NumElements, "tmp");
6978 Amt = InsertNewInstBefore(Tmp, AI);
6982 if (int Offset = (AllocElTySize*ArrayOffset)/CastElTySize) {
6983 Value *Off = ConstantInt::get(Type::Int32Ty, Offset, true);
6984 Instruction *Tmp = BinaryOperator::CreateAdd(Amt, Off, "tmp");
6985 Amt = InsertNewInstBefore(Tmp, AI);
6988 AllocationInst *New;
6989 if (isa<MallocInst>(AI))
6990 New = new MallocInst(CastElTy, Amt, AI.getAlignment());
6992 New = new AllocaInst(CastElTy, Amt, AI.getAlignment());
6993 InsertNewInstBefore(New, AI);
6996 // If the allocation has multiple uses, insert a cast and change all things
6997 // that used it to use the new cast. This will also hack on CI, but it will
6999 if (!AI.hasOneUse()) {
7000 AddUsesToWorkList(AI);
7001 // New is the allocation instruction, pointer typed. AI is the original
7002 // allocation instruction, also pointer typed. Thus, cast to use is BitCast.
7003 CastInst *NewCast = new BitCastInst(New, AI.getType(), "tmpcast");
7004 InsertNewInstBefore(NewCast, AI);
7005 AI.replaceAllUsesWith(NewCast);
7007 return ReplaceInstUsesWith(CI, New);
7010 /// CanEvaluateInDifferentType - Return true if we can take the specified value
7011 /// and return it as type Ty without inserting any new casts and without
7012 /// changing the computed value. This is used by code that tries to decide
7013 /// whether promoting or shrinking integer operations to wider or smaller types
7014 /// will allow us to eliminate a truncate or extend.
7016 /// This is a truncation operation if Ty is smaller than V->getType(), or an
7017 /// extension operation if Ty is larger.
7019 /// If CastOpc is a truncation, then Ty will be a type smaller than V. We
7020 /// should return true if trunc(V) can be computed by computing V in the smaller
7021 /// type. If V is an instruction, then trunc(inst(x,y)) can be computed as
7022 /// inst(trunc(x),trunc(y)), which only makes sense if x and y can be
7023 /// efficiently truncated.
7025 /// If CastOpc is a sext or zext, we are asking if the low bits of the value can
7026 /// bit computed in a larger type, which is then and'd or sext_in_reg'd to get
7027 /// the final result.
7028 bool InstCombiner::CanEvaluateInDifferentType(Value *V, const IntegerType *Ty,
7030 int &NumCastsRemoved) {
7031 // We can always evaluate constants in another type.
7032 if (isa<ConstantInt>(V))
7035 Instruction *I = dyn_cast<Instruction>(V);
7036 if (!I) return false;
7038 const IntegerType *OrigTy = cast<IntegerType>(V->getType());
7040 // If this is an extension or truncate, we can often eliminate it.
7041 if (isa<TruncInst>(I) || isa<ZExtInst>(I) || isa<SExtInst>(I)) {
7042 // If this is a cast from the destination type, we can trivially eliminate
7043 // it, and this will remove a cast overall.
7044 if (I->getOperand(0)->getType() == Ty) {
7045 // If the first operand is itself a cast, and is eliminable, do not count
7046 // this as an eliminable cast. We would prefer to eliminate those two
7048 if (!isa<CastInst>(I->getOperand(0)) && I->hasOneUse())
7054 // We can't extend or shrink something that has multiple uses: doing so would
7055 // require duplicating the instruction in general, which isn't profitable.
7056 if (!I->hasOneUse()) return false;
7058 switch (I->getOpcode()) {
7059 case Instruction::Add:
7060 case Instruction::Sub:
7061 case Instruction::Mul:
7062 case Instruction::And:
7063 case Instruction::Or:
7064 case Instruction::Xor:
7065 // These operators can all arbitrarily be extended or truncated.
7066 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
7068 CanEvaluateInDifferentType(I->getOperand(1), Ty, CastOpc,
7071 case Instruction::Shl:
7072 // If we are truncating the result of this SHL, and if it's a shift of a
7073 // constant amount, we can always perform a SHL in a smaller type.
7074 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
7075 uint32_t BitWidth = Ty->getBitWidth();
7076 if (BitWidth < OrigTy->getBitWidth() &&
7077 CI->getLimitedValue(BitWidth) < BitWidth)
7078 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
7082 case Instruction::LShr:
7083 // If this is a truncate of a logical shr, we can truncate it to a smaller
7084 // lshr iff we know that the bits we would otherwise be shifting in are
7086 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
7087 uint32_t OrigBitWidth = OrigTy->getBitWidth();
7088 uint32_t BitWidth = Ty->getBitWidth();
7089 if (BitWidth < OrigBitWidth &&
7090 MaskedValueIsZero(I->getOperand(0),
7091 APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth)) &&
7092 CI->getLimitedValue(BitWidth) < BitWidth) {
7093 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
7098 case Instruction::ZExt:
7099 case Instruction::SExt:
7100 case Instruction::Trunc:
7101 // If this is the same kind of case as our original (e.g. zext+zext), we
7102 // can safely replace it. Note that replacing it does not reduce the number
7103 // of casts in the input.
7104 if (I->getOpcode() == CastOpc)
7107 case Instruction::Select: {
7108 SelectInst *SI = cast<SelectInst>(I);
7109 return CanEvaluateInDifferentType(SI->getTrueValue(), Ty, CastOpc,
7111 CanEvaluateInDifferentType(SI->getFalseValue(), Ty, CastOpc,
7114 case Instruction::PHI: {
7115 // We can change a phi if we can change all operands.
7116 PHINode *PN = cast<PHINode>(I);
7117 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
7118 if (!CanEvaluateInDifferentType(PN->getIncomingValue(i), Ty, CastOpc,
7124 // TODO: Can handle more cases here.
7131 /// EvaluateInDifferentType - Given an expression that
7132 /// CanEvaluateInDifferentType returns true for, actually insert the code to
7133 /// evaluate the expression.
7134 Value *InstCombiner::EvaluateInDifferentType(Value *V, const Type *Ty,
7136 if (Constant *C = dyn_cast<Constant>(V))
7137 return ConstantExpr::getIntegerCast(C, Ty, isSigned /*Sext or ZExt*/);
7139 // Otherwise, it must be an instruction.
7140 Instruction *I = cast<Instruction>(V);
7141 Instruction *Res = 0;
7142 switch (I->getOpcode()) {
7143 case Instruction::Add:
7144 case Instruction::Sub:
7145 case Instruction::Mul:
7146 case Instruction::And:
7147 case Instruction::Or:
7148 case Instruction::Xor:
7149 case Instruction::AShr:
7150 case Instruction::LShr:
7151 case Instruction::Shl: {
7152 Value *LHS = EvaluateInDifferentType(I->getOperand(0), Ty, isSigned);
7153 Value *RHS = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
7154 Res = BinaryOperator::Create((Instruction::BinaryOps)I->getOpcode(),
7158 case Instruction::Trunc:
7159 case Instruction::ZExt:
7160 case Instruction::SExt:
7161 // If the source type of the cast is the type we're trying for then we can
7162 // just return the source. There's no need to insert it because it is not
7164 if (I->getOperand(0)->getType() == Ty)
7165 return I->getOperand(0);
7167 // Otherwise, must be the same type of cast, so just reinsert a new one.
7168 Res = CastInst::Create(cast<CastInst>(I)->getOpcode(), I->getOperand(0),
7171 case Instruction::Select: {
7172 Value *True = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
7173 Value *False = EvaluateInDifferentType(I->getOperand(2), Ty, isSigned);
7174 Res = SelectInst::Create(I->getOperand(0), True, False);
7177 case Instruction::PHI: {
7178 PHINode *OPN = cast<PHINode>(I);
7179 PHINode *NPN = PHINode::Create(Ty);
7180 for (unsigned i = 0, e = OPN->getNumIncomingValues(); i != e; ++i) {
7181 Value *V =EvaluateInDifferentType(OPN->getIncomingValue(i), Ty, isSigned);
7182 NPN->addIncoming(V, OPN->getIncomingBlock(i));
7188 // TODO: Can handle more cases here.
7189 assert(0 && "Unreachable!");
7194 return InsertNewInstBefore(Res, *I);
7197 /// @brief Implement the transforms common to all CastInst visitors.
7198 Instruction *InstCombiner::commonCastTransforms(CastInst &CI) {
7199 Value *Src = CI.getOperand(0);
7201 // Many cases of "cast of a cast" are eliminable. If it's eliminable we just
7202 // eliminate it now.
7203 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
7204 if (Instruction::CastOps opc =
7205 isEliminableCastPair(CSrc, CI.getOpcode(), CI.getType(), TD)) {
7206 // The first cast (CSrc) is eliminable so we need to fix up or replace
7207 // the second cast (CI). CSrc will then have a good chance of being dead.
7208 return CastInst::Create(opc, CSrc->getOperand(0), CI.getType());
7212 // If we are casting a select then fold the cast into the select
7213 if (SelectInst *SI = dyn_cast<SelectInst>(Src))
7214 if (Instruction *NV = FoldOpIntoSelect(CI, SI, this))
7217 // If we are casting a PHI then fold the cast into the PHI
7218 if (isa<PHINode>(Src))
7219 if (Instruction *NV = FoldOpIntoPhi(CI))
7225 /// @brief Implement the transforms for cast of pointer (bitcast/ptrtoint)
7226 Instruction *InstCombiner::commonPointerCastTransforms(CastInst &CI) {
7227 Value *Src = CI.getOperand(0);
7229 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Src)) {
7230 // If casting the result of a getelementptr instruction with no offset, turn
7231 // this into a cast of the original pointer!
7232 if (GEP->hasAllZeroIndices()) {
7233 // Changing the cast operand is usually not a good idea but it is safe
7234 // here because the pointer operand is being replaced with another
7235 // pointer operand so the opcode doesn't need to change.
7237 CI.setOperand(0, GEP->getOperand(0));
7241 // If the GEP has a single use, and the base pointer is a bitcast, and the
7242 // GEP computes a constant offset, see if we can convert these three
7243 // instructions into fewer. This typically happens with unions and other
7244 // non-type-safe code.
7245 if (GEP->hasOneUse() && isa<BitCastInst>(GEP->getOperand(0))) {
7246 if (GEP->hasAllConstantIndices()) {
7247 // We are guaranteed to get a constant from EmitGEPOffset.
7248 ConstantInt *OffsetV = cast<ConstantInt>(EmitGEPOffset(GEP, CI, *this));
7249 int64_t Offset = OffsetV->getSExtValue();
7251 // Get the base pointer input of the bitcast, and the type it points to.
7252 Value *OrigBase = cast<BitCastInst>(GEP->getOperand(0))->getOperand(0);
7253 const Type *GEPIdxTy =
7254 cast<PointerType>(OrigBase->getType())->getElementType();
7255 if (GEPIdxTy->isSized()) {
7256 SmallVector<Value*, 8> NewIndices;
7258 // Start with the index over the outer type. Note that the type size
7259 // might be zero (even if the offset isn't zero) if the indexed type
7260 // is something like [0 x {int, int}]
7261 const Type *IntPtrTy = TD->getIntPtrType();
7262 int64_t FirstIdx = 0;
7263 if (int64_t TySize = TD->getABITypeSize(GEPIdxTy)) {
7264 FirstIdx = Offset/TySize;
7267 // Handle silly modulus not returning values values [0..TySize).
7271 assert(Offset >= 0);
7273 assert((uint64_t)Offset < (uint64_t)TySize &&"Out of range offset");
7276 NewIndices.push_back(ConstantInt::get(IntPtrTy, FirstIdx));
7278 // Index into the types. If we fail, set OrigBase to null.
7280 if (const StructType *STy = dyn_cast<StructType>(GEPIdxTy)) {
7281 const StructLayout *SL = TD->getStructLayout(STy);
7282 if (Offset < (int64_t)SL->getSizeInBytes()) {
7283 unsigned Elt = SL->getElementContainingOffset(Offset);
7284 NewIndices.push_back(ConstantInt::get(Type::Int32Ty, Elt));
7286 Offset -= SL->getElementOffset(Elt);
7287 GEPIdxTy = STy->getElementType(Elt);
7289 // Otherwise, we can't index into this, bail out.
7293 } else if (isa<ArrayType>(GEPIdxTy) || isa<VectorType>(GEPIdxTy)) {
7294 const SequentialType *STy = cast<SequentialType>(GEPIdxTy);
7295 if (uint64_t EltSize = TD->getABITypeSize(STy->getElementType())){
7296 NewIndices.push_back(ConstantInt::get(IntPtrTy,Offset/EltSize));
7299 NewIndices.push_back(ConstantInt::get(IntPtrTy, 0));
7301 GEPIdxTy = STy->getElementType();
7303 // Otherwise, we can't index into this, bail out.
7309 // If we were able to index down into an element, create the GEP
7310 // and bitcast the result. This eliminates one bitcast, potentially
7312 Instruction *NGEP = GetElementPtrInst::Create(OrigBase,
7314 NewIndices.end(), "");
7315 InsertNewInstBefore(NGEP, CI);
7316 NGEP->takeName(GEP);
7318 if (isa<BitCastInst>(CI))
7319 return new BitCastInst(NGEP, CI.getType());
7320 assert(isa<PtrToIntInst>(CI));
7321 return new PtrToIntInst(NGEP, CI.getType());
7328 return commonCastTransforms(CI);
7333 /// Only the TRUNC, ZEXT, SEXT, and BITCAST can both operand and result as
7334 /// integer types. This function implements the common transforms for all those
7336 /// @brief Implement the transforms common to CastInst with integer operands
7337 Instruction *InstCombiner::commonIntCastTransforms(CastInst &CI) {
7338 if (Instruction *Result = commonCastTransforms(CI))
7341 Value *Src = CI.getOperand(0);
7342 const Type *SrcTy = Src->getType();
7343 const Type *DestTy = CI.getType();
7344 uint32_t SrcBitSize = SrcTy->getPrimitiveSizeInBits();
7345 uint32_t DestBitSize = DestTy->getPrimitiveSizeInBits();
7347 // See if we can simplify any instructions used by the LHS whose sole
7348 // purpose is to compute bits we don't care about.
7349 APInt KnownZero(DestBitSize, 0), KnownOne(DestBitSize, 0);
7350 if (SimplifyDemandedBits(&CI, APInt::getAllOnesValue(DestBitSize),
7351 KnownZero, KnownOne))
7354 // If the source isn't an instruction or has more than one use then we
7355 // can't do anything more.
7356 Instruction *SrcI = dyn_cast<Instruction>(Src);
7357 if (!SrcI || !Src->hasOneUse())
7360 // Attempt to propagate the cast into the instruction for int->int casts.
7361 int NumCastsRemoved = 0;
7362 if (!isa<BitCastInst>(CI) &&
7363 CanEvaluateInDifferentType(SrcI, cast<IntegerType>(DestTy),
7364 CI.getOpcode(), NumCastsRemoved)) {
7365 // If this cast is a truncate, evaluting in a different type always
7366 // eliminates the cast, so it is always a win. If this is a zero-extension,
7367 // we need to do an AND to maintain the clear top-part of the computation,
7368 // so we require that the input have eliminated at least one cast. If this
7369 // is a sign extension, we insert two new casts (to do the extension) so we
7370 // require that two casts have been eliminated.
7372 switch (CI.getOpcode()) {
7374 // All the others use floating point so we shouldn't actually
7375 // get here because of the check above.
7376 assert(0 && "Unknown cast type");
7377 case Instruction::Trunc:
7380 case Instruction::ZExt:
7381 DoXForm = NumCastsRemoved >= 1;
7383 case Instruction::SExt:
7384 DoXForm = NumCastsRemoved >= 2;
7389 Value *Res = EvaluateInDifferentType(SrcI, DestTy,
7390 CI.getOpcode() == Instruction::SExt);
7391 assert(Res->getType() == DestTy);
7392 switch (CI.getOpcode()) {
7393 default: assert(0 && "Unknown cast type!");
7394 case Instruction::Trunc:
7395 case Instruction::BitCast:
7396 // Just replace this cast with the result.
7397 return ReplaceInstUsesWith(CI, Res);
7398 case Instruction::ZExt: {
7399 // We need to emit an AND to clear the high bits.
7400 assert(SrcBitSize < DestBitSize && "Not a zext?");
7401 Constant *C = ConstantInt::get(APInt::getLowBitsSet(DestBitSize,
7403 return BinaryOperator::CreateAnd(Res, C);
7405 case Instruction::SExt:
7406 // We need to emit a cast to truncate, then a cast to sext.
7407 return CastInst::Create(Instruction::SExt,
7408 InsertCastBefore(Instruction::Trunc, Res, Src->getType(),
7414 Value *Op0 = SrcI->getNumOperands() > 0 ? SrcI->getOperand(0) : 0;
7415 Value *Op1 = SrcI->getNumOperands() > 1 ? SrcI->getOperand(1) : 0;
7417 switch (SrcI->getOpcode()) {
7418 case Instruction::Add:
7419 case Instruction::Mul:
7420 case Instruction::And:
7421 case Instruction::Or:
7422 case Instruction::Xor:
7423 // If we are discarding information, rewrite.
7424 if (DestBitSize <= SrcBitSize && DestBitSize != 1) {
7425 // Don't insert two casts if they cannot be eliminated. We allow
7426 // two casts to be inserted if the sizes are the same. This could
7427 // only be converting signedness, which is a noop.
7428 if (DestBitSize == SrcBitSize ||
7429 !ValueRequiresCast(CI.getOpcode(), Op1, DestTy,TD) ||
7430 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
7431 Instruction::CastOps opcode = CI.getOpcode();
7432 Value *Op0c = InsertOperandCastBefore(opcode, Op0, DestTy, SrcI);
7433 Value *Op1c = InsertOperandCastBefore(opcode, Op1, DestTy, SrcI);
7434 return BinaryOperator::Create(
7435 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
7439 // cast (xor bool X, true) to int --> xor (cast bool X to int), 1
7440 if (isa<ZExtInst>(CI) && SrcBitSize == 1 &&
7441 SrcI->getOpcode() == Instruction::Xor &&
7442 Op1 == ConstantInt::getTrue() &&
7443 (!Op0->hasOneUse() || !isa<CmpInst>(Op0))) {
7444 Value *New = InsertOperandCastBefore(Instruction::ZExt, Op0, DestTy, &CI);
7445 return BinaryOperator::CreateXor(New, ConstantInt::get(CI.getType(), 1));
7448 case Instruction::SDiv:
7449 case Instruction::UDiv:
7450 case Instruction::SRem:
7451 case Instruction::URem:
7452 // If we are just changing the sign, rewrite.
7453 if (DestBitSize == SrcBitSize) {
7454 // Don't insert two casts if they cannot be eliminated. We allow
7455 // two casts to be inserted if the sizes are the same. This could
7456 // only be converting signedness, which is a noop.
7457 if (!ValueRequiresCast(CI.getOpcode(), Op1, DestTy, TD) ||
7458 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
7459 Value *Op0c = InsertOperandCastBefore(Instruction::BitCast,
7461 Value *Op1c = InsertOperandCastBefore(Instruction::BitCast,
7463 return BinaryOperator::Create(
7464 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
7469 case Instruction::Shl:
7470 // Allow changing the sign of the source operand. Do not allow
7471 // changing the size of the shift, UNLESS the shift amount is a
7472 // constant. We must not change variable sized shifts to a smaller
7473 // size, because it is undefined to shift more bits out than exist
7475 if (DestBitSize == SrcBitSize ||
7476 (DestBitSize < SrcBitSize && isa<Constant>(Op1))) {
7477 Instruction::CastOps opcode = (DestBitSize == SrcBitSize ?
7478 Instruction::BitCast : Instruction::Trunc);
7479 Value *Op0c = InsertOperandCastBefore(opcode, Op0, DestTy, SrcI);
7480 Value *Op1c = InsertOperandCastBefore(opcode, Op1, DestTy, SrcI);
7481 return BinaryOperator::CreateShl(Op0c, Op1c);
7484 case Instruction::AShr:
7485 // If this is a signed shr, and if all bits shifted in are about to be
7486 // truncated off, turn it into an unsigned shr to allow greater
7488 if (DestBitSize < SrcBitSize &&
7489 isa<ConstantInt>(Op1)) {
7490 uint32_t ShiftAmt = cast<ConstantInt>(Op1)->getLimitedValue(SrcBitSize);
7491 if (SrcBitSize > ShiftAmt && SrcBitSize-ShiftAmt >= DestBitSize) {
7492 // Insert the new logical shift right.
7493 return BinaryOperator::CreateLShr(Op0, Op1);
7501 Instruction *InstCombiner::visitTrunc(TruncInst &CI) {
7502 if (Instruction *Result = commonIntCastTransforms(CI))
7505 Value *Src = CI.getOperand(0);
7506 const Type *Ty = CI.getType();
7507 uint32_t DestBitWidth = Ty->getPrimitiveSizeInBits();
7508 uint32_t SrcBitWidth = cast<IntegerType>(Src->getType())->getBitWidth();
7510 if (Instruction *SrcI = dyn_cast<Instruction>(Src)) {
7511 switch (SrcI->getOpcode()) {
7513 case Instruction::LShr:
7514 // We can shrink lshr to something smaller if we know the bits shifted in
7515 // are already zeros.
7516 if (ConstantInt *ShAmtV = dyn_cast<ConstantInt>(SrcI->getOperand(1))) {
7517 uint32_t ShAmt = ShAmtV->getLimitedValue(SrcBitWidth);
7519 // Get a mask for the bits shifting in.
7520 APInt Mask(APInt::getLowBitsSet(SrcBitWidth, ShAmt).shl(DestBitWidth));
7521 Value* SrcIOp0 = SrcI->getOperand(0);
7522 if (SrcI->hasOneUse() && MaskedValueIsZero(SrcIOp0, Mask)) {
7523 if (ShAmt >= DestBitWidth) // All zeros.
7524 return ReplaceInstUsesWith(CI, Constant::getNullValue(Ty));
7526 // Okay, we can shrink this. Truncate the input, then return a new
7528 Value *V1 = InsertCastBefore(Instruction::Trunc, SrcIOp0, Ty, CI);
7529 Value *V2 = InsertCastBefore(Instruction::Trunc, SrcI->getOperand(1),
7531 return BinaryOperator::CreateLShr(V1, V2);
7533 } else { // This is a variable shr.
7535 // Turn 'trunc (lshr X, Y) to bool' into '(X & (1 << Y)) != 0'. This is
7536 // more LLVM instructions, but allows '1 << Y' to be hoisted if
7537 // loop-invariant and CSE'd.
7538 if (CI.getType() == Type::Int1Ty && SrcI->hasOneUse()) {
7539 Value *One = ConstantInt::get(SrcI->getType(), 1);
7541 Value *V = InsertNewInstBefore(
7542 BinaryOperator::CreateShl(One, SrcI->getOperand(1),
7544 V = InsertNewInstBefore(BinaryOperator::CreateAnd(V,
7545 SrcI->getOperand(0),
7547 Value *Zero = Constant::getNullValue(V->getType());
7548 return new ICmpInst(ICmpInst::ICMP_NE, V, Zero);
7558 /// transformZExtICmp - Transform (zext icmp) to bitwise / integer operations
7559 /// in order to eliminate the icmp.
7560 Instruction *InstCombiner::transformZExtICmp(ICmpInst *ICI, Instruction &CI,
7562 // If we are just checking for a icmp eq of a single bit and zext'ing it
7563 // to an integer, then shift the bit to the appropriate place and then
7564 // cast to integer to avoid the comparison.
7565 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(ICI->getOperand(1))) {
7566 const APInt &Op1CV = Op1C->getValue();
7568 // zext (x <s 0) to i32 --> x>>u31 true if signbit set.
7569 // zext (x >s -1) to i32 --> (x>>u31)^1 true if signbit clear.
7570 if ((ICI->getPredicate() == ICmpInst::ICMP_SLT && Op1CV == 0) ||
7571 (ICI->getPredicate() == ICmpInst::ICMP_SGT &&Op1CV.isAllOnesValue())) {
7572 if (!DoXform) return ICI;
7574 Value *In = ICI->getOperand(0);
7575 Value *Sh = ConstantInt::get(In->getType(),
7576 In->getType()->getPrimitiveSizeInBits()-1);
7577 In = InsertNewInstBefore(BinaryOperator::CreateLShr(In, Sh,
7578 In->getName()+".lobit"),
7580 if (In->getType() != CI.getType())
7581 In = CastInst::CreateIntegerCast(In, CI.getType(),
7582 false/*ZExt*/, "tmp", &CI);
7584 if (ICI->getPredicate() == ICmpInst::ICMP_SGT) {
7585 Constant *One = ConstantInt::get(In->getType(), 1);
7586 In = InsertNewInstBefore(BinaryOperator::CreateXor(In, One,
7587 In->getName()+".not"),
7591 return ReplaceInstUsesWith(CI, In);
7596 // zext (X == 0) to i32 --> X^1 iff X has only the low bit set.
7597 // zext (X == 0) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
7598 // zext (X == 1) to i32 --> X iff X has only the low bit set.
7599 // zext (X == 2) to i32 --> X>>1 iff X has only the 2nd bit set.
7600 // zext (X != 0) to i32 --> X iff X has only the low bit set.
7601 // zext (X != 0) to i32 --> X>>1 iff X has only the 2nd bit set.
7602 // zext (X != 1) to i32 --> X^1 iff X has only the low bit set.
7603 // zext (X != 2) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
7604 if ((Op1CV == 0 || Op1CV.isPowerOf2()) &&
7605 // This only works for EQ and NE
7606 ICI->isEquality()) {
7607 // If Op1C some other power of two, convert:
7608 uint32_t BitWidth = Op1C->getType()->getBitWidth();
7609 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
7610 APInt TypeMask(APInt::getAllOnesValue(BitWidth));
7611 ComputeMaskedBits(ICI->getOperand(0), TypeMask, KnownZero, KnownOne);
7613 APInt KnownZeroMask(~KnownZero);
7614 if (KnownZeroMask.isPowerOf2()) { // Exactly 1 possible 1?
7615 if (!DoXform) return ICI;
7617 bool isNE = ICI->getPredicate() == ICmpInst::ICMP_NE;
7618 if (Op1CV != 0 && (Op1CV != KnownZeroMask)) {
7619 // (X&4) == 2 --> false
7620 // (X&4) != 2 --> true
7621 Constant *Res = ConstantInt::get(Type::Int1Ty, isNE);
7622 Res = ConstantExpr::getZExt(Res, CI.getType());
7623 return ReplaceInstUsesWith(CI, Res);
7626 uint32_t ShiftAmt = KnownZeroMask.logBase2();
7627 Value *In = ICI->getOperand(0);
7629 // Perform a logical shr by shiftamt.
7630 // Insert the shift to put the result in the low bit.
7631 In = InsertNewInstBefore(BinaryOperator::CreateLShr(In,
7632 ConstantInt::get(In->getType(), ShiftAmt),
7633 In->getName()+".lobit"), CI);
7636 if ((Op1CV != 0) == isNE) { // Toggle the low bit.
7637 Constant *One = ConstantInt::get(In->getType(), 1);
7638 In = BinaryOperator::CreateXor(In, One, "tmp");
7639 InsertNewInstBefore(cast<Instruction>(In), CI);
7642 if (CI.getType() == In->getType())
7643 return ReplaceInstUsesWith(CI, In);
7645 return CastInst::CreateIntegerCast(In, CI.getType(), false/*ZExt*/);
7653 Instruction *InstCombiner::visitZExt(ZExtInst &CI) {
7654 // If one of the common conversion will work ..
7655 if (Instruction *Result = commonIntCastTransforms(CI))
7658 Value *Src = CI.getOperand(0);
7660 // If this is a cast of a cast
7661 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
7662 // If this is a TRUNC followed by a ZEXT then we are dealing with integral
7663 // types and if the sizes are just right we can convert this into a logical
7664 // 'and' which will be much cheaper than the pair of casts.
7665 if (isa<TruncInst>(CSrc)) {
7666 // Get the sizes of the types involved
7667 Value *A = CSrc->getOperand(0);
7668 uint32_t SrcSize = A->getType()->getPrimitiveSizeInBits();
7669 uint32_t MidSize = CSrc->getType()->getPrimitiveSizeInBits();
7670 uint32_t DstSize = CI.getType()->getPrimitiveSizeInBits();
7671 // If we're actually extending zero bits and the trunc is a no-op
7672 if (MidSize < DstSize && SrcSize == DstSize) {
7673 // Replace both of the casts with an And of the type mask.
7674 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
7675 Constant *AndConst = ConstantInt::get(AndValue);
7677 BinaryOperator::CreateAnd(CSrc->getOperand(0), AndConst);
7678 // Unfortunately, if the type changed, we need to cast it back.
7679 if (And->getType() != CI.getType()) {
7680 And->setName(CSrc->getName()+".mask");
7681 InsertNewInstBefore(And, CI);
7682 And = CastInst::CreateIntegerCast(And, CI.getType(), false/*ZExt*/);
7689 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src))
7690 return transformZExtICmp(ICI, CI);
7692 BinaryOperator *SrcI = dyn_cast<BinaryOperator>(Src);
7693 if (SrcI && SrcI->getOpcode() == Instruction::Or) {
7694 // zext (or icmp, icmp) --> or (zext icmp), (zext icmp) if at least one
7695 // of the (zext icmp) will be transformed.
7696 ICmpInst *LHS = dyn_cast<ICmpInst>(SrcI->getOperand(0));
7697 ICmpInst *RHS = dyn_cast<ICmpInst>(SrcI->getOperand(1));
7698 if (LHS && RHS && LHS->hasOneUse() && RHS->hasOneUse() &&
7699 (transformZExtICmp(LHS, CI, false) ||
7700 transformZExtICmp(RHS, CI, false))) {
7701 Value *LCast = InsertCastBefore(Instruction::ZExt, LHS, CI.getType(), CI);
7702 Value *RCast = InsertCastBefore(Instruction::ZExt, RHS, CI.getType(), CI);
7703 return BinaryOperator::Create(Instruction::Or, LCast, RCast);
7710 Instruction *InstCombiner::visitSExt(SExtInst &CI) {
7711 if (Instruction *I = commonIntCastTransforms(CI))
7714 Value *Src = CI.getOperand(0);
7716 // sext (x <s 0) -> ashr x, 31 -> all ones if signed
7717 // sext (x >s -1) -> ashr x, 31 -> all ones if not signed
7718 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src)) {
7719 // If we are just checking for a icmp eq of a single bit and zext'ing it
7720 // to an integer, then shift the bit to the appropriate place and then
7721 // cast to integer to avoid the comparison.
7722 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(ICI->getOperand(1))) {
7723 const APInt &Op1CV = Op1C->getValue();
7725 // sext (x <s 0) to i32 --> x>>s31 true if signbit set.
7726 // sext (x >s -1) to i32 --> (x>>s31)^-1 true if signbit clear.
7727 if ((ICI->getPredicate() == ICmpInst::ICMP_SLT && Op1CV == 0) ||
7728 (ICI->getPredicate() == ICmpInst::ICMP_SGT &&Op1CV.isAllOnesValue())){
7729 Value *In = ICI->getOperand(0);
7730 Value *Sh = ConstantInt::get(In->getType(),
7731 In->getType()->getPrimitiveSizeInBits()-1);
7732 In = InsertNewInstBefore(BinaryOperator::CreateAShr(In, Sh,
7733 In->getName()+".lobit"),
7735 if (In->getType() != CI.getType())
7736 In = CastInst::CreateIntegerCast(In, CI.getType(),
7737 true/*SExt*/, "tmp", &CI);
7739 if (ICI->getPredicate() == ICmpInst::ICMP_SGT)
7740 In = InsertNewInstBefore(BinaryOperator::CreateNot(In,
7741 In->getName()+".not"), CI);
7743 return ReplaceInstUsesWith(CI, In);
7748 // See if the value being truncated is already sign extended. If so, just
7749 // eliminate the trunc/sext pair.
7750 if (getOpcode(Src) == Instruction::Trunc) {
7751 Value *Op = cast<User>(Src)->getOperand(0);
7752 unsigned OpBits = cast<IntegerType>(Op->getType())->getBitWidth();
7753 unsigned MidBits = cast<IntegerType>(Src->getType())->getBitWidth();
7754 unsigned DestBits = cast<IntegerType>(CI.getType())->getBitWidth();
7755 unsigned NumSignBits = ComputeNumSignBits(Op);
7757 if (OpBits == DestBits) {
7758 // Op is i32, Mid is i8, and Dest is i32. If Op has more than 24 sign
7759 // bits, it is already ready.
7760 if (NumSignBits > DestBits-MidBits)
7761 return ReplaceInstUsesWith(CI, Op);
7762 } else if (OpBits < DestBits) {
7763 // Op is i32, Mid is i8, and Dest is i64. If Op has more than 24 sign
7764 // bits, just sext from i32.
7765 if (NumSignBits > OpBits-MidBits)
7766 return new SExtInst(Op, CI.getType(), "tmp");
7768 // Op is i64, Mid is i8, and Dest is i32. If Op has more than 56 sign
7769 // bits, just truncate to i32.
7770 if (NumSignBits > OpBits-MidBits)
7771 return new TruncInst(Op, CI.getType(), "tmp");
7775 // If the input is a shl/ashr pair of a same constant, then this is a sign
7776 // extension from a smaller value. If we could trust arbitrary bitwidth
7777 // integers, we could turn this into a truncate to the smaller bit and then
7778 // use a sext for the whole extension. Since we don't, look deeper and check
7779 // for a truncate. If the source and dest are the same type, eliminate the
7780 // trunc and extend and just do shifts. For example, turn:
7781 // %a = trunc i32 %i to i8
7782 // %b = shl i8 %a, 6
7783 // %c = ashr i8 %b, 6
7784 // %d = sext i8 %c to i32
7786 // %a = shl i32 %i, 30
7787 // %d = ashr i32 %a, 30
7789 ConstantInt *BA = 0, *CA = 0;
7790 if (match(Src, m_AShr(m_Shl(m_Value(A), m_ConstantInt(BA)),
7791 m_ConstantInt(CA))) &&
7792 BA == CA && isa<TruncInst>(A)) {
7793 Value *I = cast<TruncInst>(A)->getOperand(0);
7794 if (I->getType() == CI.getType()) {
7795 unsigned MidSize = Src->getType()->getPrimitiveSizeInBits();
7796 unsigned SrcDstSize = CI.getType()->getPrimitiveSizeInBits();
7797 unsigned ShAmt = CA->getZExtValue()+SrcDstSize-MidSize;
7798 Constant *ShAmtV = ConstantInt::get(CI.getType(), ShAmt);
7799 I = InsertNewInstBefore(BinaryOperator::CreateShl(I, ShAmtV,
7801 return BinaryOperator::CreateAShr(I, ShAmtV);
7808 /// FitsInFPType - Return a Constant* for the specified FP constant if it fits
7809 /// in the specified FP type without changing its value.
7810 static Constant *FitsInFPType(ConstantFP *CFP, const fltSemantics &Sem) {
7811 APFloat F = CFP->getValueAPF();
7812 if (F.convert(Sem, APFloat::rmNearestTiesToEven) == APFloat::opOK)
7813 return ConstantFP::get(F);
7817 /// LookThroughFPExtensions - If this is an fp extension instruction, look
7818 /// through it until we get the source value.
7819 static Value *LookThroughFPExtensions(Value *V) {
7820 if (Instruction *I = dyn_cast<Instruction>(V))
7821 if (I->getOpcode() == Instruction::FPExt)
7822 return LookThroughFPExtensions(I->getOperand(0));
7824 // If this value is a constant, return the constant in the smallest FP type
7825 // that can accurately represent it. This allows us to turn
7826 // (float)((double)X+2.0) into x+2.0f.
7827 if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
7828 if (CFP->getType() == Type::PPC_FP128Ty)
7829 return V; // No constant folding of this.
7830 // See if the value can be truncated to float and then reextended.
7831 if (Value *V = FitsInFPType(CFP, APFloat::IEEEsingle))
7833 if (CFP->getType() == Type::DoubleTy)
7834 return V; // Won't shrink.
7835 if (Value *V = FitsInFPType(CFP, APFloat::IEEEdouble))
7837 // Don't try to shrink to various long double types.
7843 Instruction *InstCombiner::visitFPTrunc(FPTruncInst &CI) {
7844 if (Instruction *I = commonCastTransforms(CI))
7847 // If we have fptrunc(add (fpextend x), (fpextend y)), where x and y are
7848 // smaller than the destination type, we can eliminate the truncate by doing
7849 // the add as the smaller type. This applies to add/sub/mul/div as well as
7850 // many builtins (sqrt, etc).
7851 BinaryOperator *OpI = dyn_cast<BinaryOperator>(CI.getOperand(0));
7852 if (OpI && OpI->hasOneUse()) {
7853 switch (OpI->getOpcode()) {
7855 case Instruction::Add:
7856 case Instruction::Sub:
7857 case Instruction::Mul:
7858 case Instruction::FDiv:
7859 case Instruction::FRem:
7860 const Type *SrcTy = OpI->getType();
7861 Value *LHSTrunc = LookThroughFPExtensions(OpI->getOperand(0));
7862 Value *RHSTrunc = LookThroughFPExtensions(OpI->getOperand(1));
7863 if (LHSTrunc->getType() != SrcTy &&
7864 RHSTrunc->getType() != SrcTy) {
7865 unsigned DstSize = CI.getType()->getPrimitiveSizeInBits();
7866 // If the source types were both smaller than the destination type of
7867 // the cast, do this xform.
7868 if (LHSTrunc->getType()->getPrimitiveSizeInBits() <= DstSize &&
7869 RHSTrunc->getType()->getPrimitiveSizeInBits() <= DstSize) {
7870 LHSTrunc = InsertCastBefore(Instruction::FPExt, LHSTrunc,
7872 RHSTrunc = InsertCastBefore(Instruction::FPExt, RHSTrunc,
7874 return BinaryOperator::Create(OpI->getOpcode(), LHSTrunc, RHSTrunc);
7883 Instruction *InstCombiner::visitFPExt(CastInst &CI) {
7884 return commonCastTransforms(CI);
7887 Instruction *InstCombiner::visitFPToUI(FPToUIInst &FI) {
7888 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0));
7890 return commonCastTransforms(FI);
7892 // fptoui(uitofp(X)) --> X
7893 // fptoui(sitofp(X)) --> X
7894 // This is safe if the intermediate type has enough bits in its mantissa to
7895 // accurately represent all values of X. For example, do not do this with
7896 // i64->float->i64. This is also safe for sitofp case, because any negative
7897 // 'X' value would cause an undefined result for the fptoui.
7898 if ((isa<UIToFPInst>(OpI) || isa<SIToFPInst>(OpI)) &&
7899 OpI->getOperand(0)->getType() == FI.getType() &&
7900 (int)FI.getType()->getPrimitiveSizeInBits() < /*extra bit for sign */
7901 OpI->getType()->getFPMantissaWidth())
7902 return ReplaceInstUsesWith(FI, OpI->getOperand(0));
7904 return commonCastTransforms(FI);
7907 Instruction *InstCombiner::visitFPToSI(FPToSIInst &FI) {
7908 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0));
7910 return commonCastTransforms(FI);
7912 // fptosi(sitofp(X)) --> X
7913 // fptosi(uitofp(X)) --> X
7914 // This is safe if the intermediate type has enough bits in its mantissa to
7915 // accurately represent all values of X. For example, do not do this with
7916 // i64->float->i64. This is also safe for sitofp case, because any negative
7917 // 'X' value would cause an undefined result for the fptoui.
7918 if ((isa<UIToFPInst>(OpI) || isa<SIToFPInst>(OpI)) &&
7919 OpI->getOperand(0)->getType() == FI.getType() &&
7920 (int)FI.getType()->getPrimitiveSizeInBits() <=
7921 OpI->getType()->getFPMantissaWidth())
7922 return ReplaceInstUsesWith(FI, OpI->getOperand(0));
7924 return commonCastTransforms(FI);
7927 Instruction *InstCombiner::visitUIToFP(CastInst &CI) {
7928 return commonCastTransforms(CI);
7931 Instruction *InstCombiner::visitSIToFP(CastInst &CI) {
7932 return commonCastTransforms(CI);
7935 Instruction *InstCombiner::visitPtrToInt(CastInst &CI) {
7936 return commonPointerCastTransforms(CI);
7939 Instruction *InstCombiner::visitIntToPtr(IntToPtrInst &CI) {
7940 if (Instruction *I = commonCastTransforms(CI))
7943 const Type *DestPointee = cast<PointerType>(CI.getType())->getElementType();
7944 if (!DestPointee->isSized()) return 0;
7946 // If this is inttoptr(add (ptrtoint x), cst), try to turn this into a GEP.
7949 if (match(CI.getOperand(0), m_Add(m_Cast<PtrToIntInst>(m_Value(X)),
7950 m_ConstantInt(Cst)))) {
7951 // If the source and destination operands have the same type, see if this
7952 // is a single-index GEP.
7953 if (X->getType() == CI.getType()) {
7954 // Get the size of the pointee type.
7955 uint64_t Size = TD->getABITypeSize(DestPointee);
7957 // Convert the constant to intptr type.
7958 APInt Offset = Cst->getValue();
7959 Offset.sextOrTrunc(TD->getPointerSizeInBits());
7961 // If Offset is evenly divisible by Size, we can do this xform.
7962 if (Size && !APIntOps::srem(Offset, APInt(Offset.getBitWidth(), Size))){
7963 Offset = APIntOps::sdiv(Offset, APInt(Offset.getBitWidth(), Size));
7964 return GetElementPtrInst::Create(X, ConstantInt::get(Offset));
7967 // TODO: Could handle other cases, e.g. where add is indexing into field of
7969 } else if (CI.getOperand(0)->hasOneUse() &&
7970 match(CI.getOperand(0), m_Add(m_Value(X), m_ConstantInt(Cst)))) {
7971 // Otherwise, if this is inttoptr(add x, cst), try to turn this into an
7972 // "inttoptr+GEP" instead of "add+intptr".
7974 // Get the size of the pointee type.
7975 uint64_t Size = TD->getABITypeSize(DestPointee);
7977 // Convert the constant to intptr type.
7978 APInt Offset = Cst->getValue();
7979 Offset.sextOrTrunc(TD->getPointerSizeInBits());
7981 // If Offset is evenly divisible by Size, we can do this xform.
7982 if (Size && !APIntOps::srem(Offset, APInt(Offset.getBitWidth(), Size))){
7983 Offset = APIntOps::sdiv(Offset, APInt(Offset.getBitWidth(), Size));
7985 Instruction *P = InsertNewInstBefore(new IntToPtrInst(X, CI.getType(),
7987 return GetElementPtrInst::Create(P, ConstantInt::get(Offset), "tmp");
7993 Instruction *InstCombiner::visitBitCast(BitCastInst &CI) {
7994 // If the operands are integer typed then apply the integer transforms,
7995 // otherwise just apply the common ones.
7996 Value *Src = CI.getOperand(0);
7997 const Type *SrcTy = Src->getType();
7998 const Type *DestTy = CI.getType();
8000 if (SrcTy->isInteger() && DestTy->isInteger()) {
8001 if (Instruction *Result = commonIntCastTransforms(CI))
8003 } else if (isa<PointerType>(SrcTy)) {
8004 if (Instruction *I = commonPointerCastTransforms(CI))
8007 if (Instruction *Result = commonCastTransforms(CI))
8012 // Get rid of casts from one type to the same type. These are useless and can
8013 // be replaced by the operand.
8014 if (DestTy == Src->getType())
8015 return ReplaceInstUsesWith(CI, Src);
8017 if (const PointerType *DstPTy = dyn_cast<PointerType>(DestTy)) {
8018 const PointerType *SrcPTy = cast<PointerType>(SrcTy);
8019 const Type *DstElTy = DstPTy->getElementType();
8020 const Type *SrcElTy = SrcPTy->getElementType();
8022 // If the address spaces don't match, don't eliminate the bitcast, which is
8023 // required for changing types.
8024 if (SrcPTy->getAddressSpace() != DstPTy->getAddressSpace())
8027 // If we are casting a malloc or alloca to a pointer to a type of the same
8028 // size, rewrite the allocation instruction to allocate the "right" type.
8029 if (AllocationInst *AI = dyn_cast<AllocationInst>(Src))
8030 if (Instruction *V = PromoteCastOfAllocation(CI, *AI))
8033 // If the source and destination are pointers, and this cast is equivalent
8034 // to a getelementptr X, 0, 0, 0... turn it into the appropriate gep.
8035 // This can enhance SROA and other transforms that want type-safe pointers.
8036 Constant *ZeroUInt = Constant::getNullValue(Type::Int32Ty);
8037 unsigned NumZeros = 0;
8038 while (SrcElTy != DstElTy &&
8039 isa<CompositeType>(SrcElTy) && !isa<PointerType>(SrcElTy) &&
8040 SrcElTy->getNumContainedTypes() /* not "{}" */) {
8041 SrcElTy = cast<CompositeType>(SrcElTy)->getTypeAtIndex(ZeroUInt);
8045 // If we found a path from the src to dest, create the getelementptr now.
8046 if (SrcElTy == DstElTy) {
8047 SmallVector<Value*, 8> Idxs(NumZeros+1, ZeroUInt);
8048 return GetElementPtrInst::Create(Src, Idxs.begin(), Idxs.end(), "",
8049 ((Instruction*) NULL));
8053 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(Src)) {
8054 if (SVI->hasOneUse()) {
8055 // Okay, we have (bitconvert (shuffle ..)). Check to see if this is
8056 // a bitconvert to a vector with the same # elts.
8057 if (isa<VectorType>(DestTy) &&
8058 cast<VectorType>(DestTy)->getNumElements() ==
8059 SVI->getType()->getNumElements()) {
8061 // If either of the operands is a cast from CI.getType(), then
8062 // evaluating the shuffle in the casted destination's type will allow
8063 // us to eliminate at least one cast.
8064 if (((Tmp = dyn_cast<CastInst>(SVI->getOperand(0))) &&
8065 Tmp->getOperand(0)->getType() == DestTy) ||
8066 ((Tmp = dyn_cast<CastInst>(SVI->getOperand(1))) &&
8067 Tmp->getOperand(0)->getType() == DestTy)) {
8068 Value *LHS = InsertOperandCastBefore(Instruction::BitCast,
8069 SVI->getOperand(0), DestTy, &CI);
8070 Value *RHS = InsertOperandCastBefore(Instruction::BitCast,
8071 SVI->getOperand(1), DestTy, &CI);
8072 // Return a new shuffle vector. Use the same element ID's, as we
8073 // know the vector types match #elts.
8074 return new ShuffleVectorInst(LHS, RHS, SVI->getOperand(2));
8082 /// GetSelectFoldableOperands - We want to turn code that looks like this:
8084 /// %D = select %cond, %C, %A
8086 /// %C = select %cond, %B, 0
8089 /// Assuming that the specified instruction is an operand to the select, return
8090 /// a bitmask indicating which operands of this instruction are foldable if they
8091 /// equal the other incoming value of the select.
8093 static unsigned GetSelectFoldableOperands(Instruction *I) {
8094 switch (I->getOpcode()) {
8095 case Instruction::Add:
8096 case Instruction::Mul:
8097 case Instruction::And:
8098 case Instruction::Or:
8099 case Instruction::Xor:
8100 return 3; // Can fold through either operand.
8101 case Instruction::Sub: // Can only fold on the amount subtracted.
8102 case Instruction::Shl: // Can only fold on the shift amount.
8103 case Instruction::LShr:
8104 case Instruction::AShr:
8107 return 0; // Cannot fold
8111 /// GetSelectFoldableConstant - For the same transformation as the previous
8112 /// function, return the identity constant that goes into the select.
8113 static Constant *GetSelectFoldableConstant(Instruction *I) {
8114 switch (I->getOpcode()) {
8115 default: assert(0 && "This cannot happen!"); abort();
8116 case Instruction::Add:
8117 case Instruction::Sub:
8118 case Instruction::Or:
8119 case Instruction::Xor:
8120 case Instruction::Shl:
8121 case Instruction::LShr:
8122 case Instruction::AShr:
8123 return Constant::getNullValue(I->getType());
8124 case Instruction::And:
8125 return Constant::getAllOnesValue(I->getType());
8126 case Instruction::Mul:
8127 return ConstantInt::get(I->getType(), 1);
8131 /// FoldSelectOpOp - Here we have (select c, TI, FI), and we know that TI and FI
8132 /// have the same opcode and only one use each. Try to simplify this.
8133 Instruction *InstCombiner::FoldSelectOpOp(SelectInst &SI, Instruction *TI,
8135 if (TI->getNumOperands() == 1) {
8136 // If this is a non-volatile load or a cast from the same type,
8139 if (TI->getOperand(0)->getType() != FI->getOperand(0)->getType())
8142 return 0; // unknown unary op.
8145 // Fold this by inserting a select from the input values.
8146 SelectInst *NewSI = SelectInst::Create(SI.getCondition(), TI->getOperand(0),
8147 FI->getOperand(0), SI.getName()+".v");
8148 InsertNewInstBefore(NewSI, SI);
8149 return CastInst::Create(Instruction::CastOps(TI->getOpcode()), NewSI,
8153 // Only handle binary operators here.
8154 if (!isa<BinaryOperator>(TI))
8157 // Figure out if the operations have any operands in common.
8158 Value *MatchOp, *OtherOpT, *OtherOpF;
8160 if (TI->getOperand(0) == FI->getOperand(0)) {
8161 MatchOp = TI->getOperand(0);
8162 OtherOpT = TI->getOperand(1);
8163 OtherOpF = FI->getOperand(1);
8164 MatchIsOpZero = true;
8165 } else if (TI->getOperand(1) == FI->getOperand(1)) {
8166 MatchOp = TI->getOperand(1);
8167 OtherOpT = TI->getOperand(0);
8168 OtherOpF = FI->getOperand(0);
8169 MatchIsOpZero = false;
8170 } else if (!TI->isCommutative()) {
8172 } else if (TI->getOperand(0) == FI->getOperand(1)) {
8173 MatchOp = TI->getOperand(0);
8174 OtherOpT = TI->getOperand(1);
8175 OtherOpF = FI->getOperand(0);
8176 MatchIsOpZero = true;
8177 } else if (TI->getOperand(1) == FI->getOperand(0)) {
8178 MatchOp = TI->getOperand(1);
8179 OtherOpT = TI->getOperand(0);
8180 OtherOpF = FI->getOperand(1);
8181 MatchIsOpZero = true;
8186 // If we reach here, they do have operations in common.
8187 SelectInst *NewSI = SelectInst::Create(SI.getCondition(), OtherOpT,
8188 OtherOpF, SI.getName()+".v");
8189 InsertNewInstBefore(NewSI, SI);
8191 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TI)) {
8193 return BinaryOperator::Create(BO->getOpcode(), MatchOp, NewSI);
8195 return BinaryOperator::Create(BO->getOpcode(), NewSI, MatchOp);
8197 assert(0 && "Shouldn't get here");
8201 /// visitSelectInstWithICmp - Visit a SelectInst that has an
8202 /// ICmpInst as its first operand.
8204 Instruction *InstCombiner::visitSelectInstWithICmp(SelectInst &SI,
8206 bool Changed = false;
8207 ICmpInst::Predicate Pred = ICI->getPredicate();
8208 Value *CmpLHS = ICI->getOperand(0);
8209 Value *CmpRHS = ICI->getOperand(1);
8210 Value *TrueVal = SI.getTrueValue();
8211 Value *FalseVal = SI.getFalseValue();
8213 // Check cases where the comparison is with a constant that
8214 // can be adjusted to fit the min/max idiom. We may edit ICI in
8215 // place here, so make sure the select is the only user.
8216 if (ICI->hasOneUse())
8217 if (ConstantInt *CI = dyn_cast<ConstantInt>(CmpRHS))
8220 case ICmpInst::ICMP_ULT:
8221 case ICmpInst::ICMP_SLT: {
8222 // X < MIN ? T : F --> F
8223 if (CI->isMinValue(Pred == ICmpInst::ICMP_SLT))
8224 return ReplaceInstUsesWith(SI, FalseVal);
8225 // X < C ? X : C-1 --> X > C-1 ? C-1 : X
8226 Constant *AdjustedRHS = SubOne(CI);
8227 if ((CmpLHS == TrueVal && AdjustedRHS == FalseVal) ||
8228 (CmpLHS == FalseVal && AdjustedRHS == TrueVal)) {
8229 Pred = ICmpInst::getSwappedPredicate(Pred);
8230 CmpRHS = AdjustedRHS;
8231 std::swap(FalseVal, TrueVal);
8232 ICI->setPredicate(Pred);
8233 ICI->setOperand(1, CmpRHS);
8234 SI.setOperand(1, TrueVal);
8235 SI.setOperand(2, FalseVal);
8240 case ICmpInst::ICMP_UGT:
8241 case ICmpInst::ICMP_SGT: {
8242 // X > MAX ? T : F --> F
8243 if (CI->isMaxValue(Pred == ICmpInst::ICMP_SGT))
8244 return ReplaceInstUsesWith(SI, FalseVal);
8245 // X > C ? X : C+1 --> X < C+1 ? C+1 : X
8246 Constant *AdjustedRHS = AddOne(CI);
8247 if ((CmpLHS == TrueVal && AdjustedRHS == FalseVal) ||
8248 (CmpLHS == FalseVal && AdjustedRHS == TrueVal)) {
8249 Pred = ICmpInst::getSwappedPredicate(Pred);
8250 CmpRHS = AdjustedRHS;
8251 std::swap(FalseVal, TrueVal);
8252 ICI->setPredicate(Pred);
8253 ICI->setOperand(1, CmpRHS);
8254 SI.setOperand(1, TrueVal);
8255 SI.setOperand(2, FalseVal);
8262 if (CmpLHS == TrueVal && CmpRHS == FalseVal) {
8263 // Transform (X == Y) ? X : Y -> Y
8264 if (Pred == ICmpInst::ICMP_EQ)
8265 return ReplaceInstUsesWith(SI, FalseVal);
8266 // Transform (X != Y) ? X : Y -> X
8267 if (Pred == ICmpInst::ICMP_NE)
8268 return ReplaceInstUsesWith(SI, TrueVal);
8269 /// NOTE: if we wanted to, this is where to detect integer MIN/MAX
8271 } else if (CmpLHS == FalseVal && CmpRHS == TrueVal) {
8272 // Transform (X == Y) ? Y : X -> X
8273 if (Pred == ICmpInst::ICMP_EQ)
8274 return ReplaceInstUsesWith(SI, FalseVal);
8275 // Transform (X != Y) ? Y : X -> Y
8276 if (Pred == ICmpInst::ICMP_NE)
8277 return ReplaceInstUsesWith(SI, TrueVal);
8278 /// NOTE: if we wanted to, this is where to detect integer MIN/MAX
8281 /// NOTE: if we wanted to, this is where to detect integer ABS
8283 return Changed ? &SI : 0;
8286 Instruction *InstCombiner::visitSelectInst(SelectInst &SI) {
8287 Value *CondVal = SI.getCondition();
8288 Value *TrueVal = SI.getTrueValue();
8289 Value *FalseVal = SI.getFalseValue();
8291 // select true, X, Y -> X
8292 // select false, X, Y -> Y
8293 if (ConstantInt *C = dyn_cast<ConstantInt>(CondVal))
8294 return ReplaceInstUsesWith(SI, C->getZExtValue() ? TrueVal : FalseVal);
8296 // select C, X, X -> X
8297 if (TrueVal == FalseVal)
8298 return ReplaceInstUsesWith(SI, TrueVal);
8300 if (isa<UndefValue>(TrueVal)) // select C, undef, X -> X
8301 return ReplaceInstUsesWith(SI, FalseVal);
8302 if (isa<UndefValue>(FalseVal)) // select C, X, undef -> X
8303 return ReplaceInstUsesWith(SI, TrueVal);
8304 if (isa<UndefValue>(CondVal)) { // select undef, X, Y -> X or Y
8305 if (isa<Constant>(TrueVal))
8306 return ReplaceInstUsesWith(SI, TrueVal);
8308 return ReplaceInstUsesWith(SI, FalseVal);
8311 if (SI.getType() == Type::Int1Ty) {
8312 if (ConstantInt *C = dyn_cast<ConstantInt>(TrueVal)) {
8313 if (C->getZExtValue()) {
8314 // Change: A = select B, true, C --> A = or B, C
8315 return BinaryOperator::CreateOr(CondVal, FalseVal);
8317 // Change: A = select B, false, C --> A = and !B, C
8319 InsertNewInstBefore(BinaryOperator::CreateNot(CondVal,
8320 "not."+CondVal->getName()), SI);
8321 return BinaryOperator::CreateAnd(NotCond, FalseVal);
8323 } else if (ConstantInt *C = dyn_cast<ConstantInt>(FalseVal)) {
8324 if (C->getZExtValue() == false) {
8325 // Change: A = select B, C, false --> A = and B, C
8326 return BinaryOperator::CreateAnd(CondVal, TrueVal);
8328 // Change: A = select B, C, true --> A = or !B, C
8330 InsertNewInstBefore(BinaryOperator::CreateNot(CondVal,
8331 "not."+CondVal->getName()), SI);
8332 return BinaryOperator::CreateOr(NotCond, TrueVal);
8336 // select a, b, a -> a&b
8337 // select a, a, b -> a|b
8338 if (CondVal == TrueVal)
8339 return BinaryOperator::CreateOr(CondVal, FalseVal);
8340 else if (CondVal == FalseVal)
8341 return BinaryOperator::CreateAnd(CondVal, TrueVal);
8344 // Selecting between two integer constants?
8345 if (ConstantInt *TrueValC = dyn_cast<ConstantInt>(TrueVal))
8346 if (ConstantInt *FalseValC = dyn_cast<ConstantInt>(FalseVal)) {
8347 // select C, 1, 0 -> zext C to int
8348 if (FalseValC->isZero() && TrueValC->getValue() == 1) {
8349 return CastInst::Create(Instruction::ZExt, CondVal, SI.getType());
8350 } else if (TrueValC->isZero() && FalseValC->getValue() == 1) {
8351 // select C, 0, 1 -> zext !C to int
8353 InsertNewInstBefore(BinaryOperator::CreateNot(CondVal,
8354 "not."+CondVal->getName()), SI);
8355 return CastInst::Create(Instruction::ZExt, NotCond, SI.getType());
8358 // FIXME: Turn select 0/-1 and -1/0 into sext from condition!
8360 if (ICmpInst *IC = dyn_cast<ICmpInst>(SI.getCondition())) {
8362 // (x <s 0) ? -1 : 0 -> ashr x, 31
8363 if (TrueValC->isAllOnesValue() && FalseValC->isZero())
8364 if (ConstantInt *CmpCst = dyn_cast<ConstantInt>(IC->getOperand(1))) {
8365 if (IC->getPredicate() == ICmpInst::ICMP_SLT && CmpCst->isZero()) {
8366 // The comparison constant and the result are not neccessarily the
8367 // same width. Make an all-ones value by inserting a AShr.
8368 Value *X = IC->getOperand(0);
8369 uint32_t Bits = X->getType()->getPrimitiveSizeInBits();
8370 Constant *ShAmt = ConstantInt::get(X->getType(), Bits-1);
8371 Instruction *SRA = BinaryOperator::Create(Instruction::AShr, X,
8373 InsertNewInstBefore(SRA, SI);
8375 // Finally, convert to the type of the select RHS. We figure out
8376 // if this requires a SExt, Trunc or BitCast based on the sizes.
8377 Instruction::CastOps opc = Instruction::BitCast;
8378 uint32_t SRASize = SRA->getType()->getPrimitiveSizeInBits();
8379 uint32_t SISize = SI.getType()->getPrimitiveSizeInBits();
8380 if (SRASize < SISize)
8381 opc = Instruction::SExt;
8382 else if (SRASize > SISize)
8383 opc = Instruction::Trunc;
8384 return CastInst::Create(opc, SRA, SI.getType());
8389 // If one of the constants is zero (we know they can't both be) and we
8390 // have an icmp instruction with zero, and we have an 'and' with the
8391 // non-constant value, eliminate this whole mess. This corresponds to
8392 // cases like this: ((X & 27) ? 27 : 0)
8393 if (TrueValC->isZero() || FalseValC->isZero())
8394 if (IC->isEquality() && isa<ConstantInt>(IC->getOperand(1)) &&
8395 cast<Constant>(IC->getOperand(1))->isNullValue())
8396 if (Instruction *ICA = dyn_cast<Instruction>(IC->getOperand(0)))
8397 if (ICA->getOpcode() == Instruction::And &&
8398 isa<ConstantInt>(ICA->getOperand(1)) &&
8399 (ICA->getOperand(1) == TrueValC ||
8400 ICA->getOperand(1) == FalseValC) &&
8401 isOneBitSet(cast<ConstantInt>(ICA->getOperand(1)))) {
8402 // Okay, now we know that everything is set up, we just don't
8403 // know whether we have a icmp_ne or icmp_eq and whether the
8404 // true or false val is the zero.
8405 bool ShouldNotVal = !TrueValC->isZero();
8406 ShouldNotVal ^= IC->getPredicate() == ICmpInst::ICMP_NE;
8409 V = InsertNewInstBefore(BinaryOperator::Create(
8410 Instruction::Xor, V, ICA->getOperand(1)), SI);
8411 return ReplaceInstUsesWith(SI, V);
8416 // See if we are selecting two values based on a comparison of the two values.
8417 if (FCmpInst *FCI = dyn_cast<FCmpInst>(CondVal)) {
8418 if (FCI->getOperand(0) == TrueVal && FCI->getOperand(1) == FalseVal) {
8419 // Transform (X == Y) ? X : Y -> Y
8420 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ) {
8421 // This is not safe in general for floating point:
8422 // consider X== -0, Y== +0.
8423 // It becomes safe if either operand is a nonzero constant.
8424 ConstantFP *CFPt, *CFPf;
8425 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) &&
8426 !CFPt->getValueAPF().isZero()) ||
8427 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) &&
8428 !CFPf->getValueAPF().isZero()))
8429 return ReplaceInstUsesWith(SI, FalseVal);
8431 // Transform (X != Y) ? X : Y -> X
8432 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
8433 return ReplaceInstUsesWith(SI, TrueVal);
8434 // NOTE: if we wanted to, this is where to detect MIN/MAX
8436 } else if (FCI->getOperand(0) == FalseVal && FCI->getOperand(1) == TrueVal){
8437 // Transform (X == Y) ? Y : X -> X
8438 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ) {
8439 // This is not safe in general for floating point:
8440 // consider X== -0, Y== +0.
8441 // It becomes safe if either operand is a nonzero constant.
8442 ConstantFP *CFPt, *CFPf;
8443 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) &&
8444 !CFPt->getValueAPF().isZero()) ||
8445 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) &&
8446 !CFPf->getValueAPF().isZero()))
8447 return ReplaceInstUsesWith(SI, FalseVal);
8449 // Transform (X != Y) ? Y : X -> Y
8450 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
8451 return ReplaceInstUsesWith(SI, TrueVal);
8452 // NOTE: if we wanted to, this is where to detect MIN/MAX
8454 // NOTE: if we wanted to, this is where to detect ABS
8457 // See if we are selecting two values based on a comparison of the two values.
8458 if (ICmpInst *ICI = dyn_cast<ICmpInst>(CondVal))
8459 if (Instruction *Result = visitSelectInstWithICmp(SI, ICI))
8462 if (Instruction *TI = dyn_cast<Instruction>(TrueVal))
8463 if (Instruction *FI = dyn_cast<Instruction>(FalseVal))
8464 if (TI->hasOneUse() && FI->hasOneUse()) {
8465 Instruction *AddOp = 0, *SubOp = 0;
8467 // Turn (select C, (op X, Y), (op X, Z)) -> (op X, (select C, Y, Z))
8468 if (TI->getOpcode() == FI->getOpcode())
8469 if (Instruction *IV = FoldSelectOpOp(SI, TI, FI))
8472 // Turn select C, (X+Y), (X-Y) --> (X+(select C, Y, (-Y))). This is
8473 // even legal for FP.
8474 if (TI->getOpcode() == Instruction::Sub &&
8475 FI->getOpcode() == Instruction::Add) {
8476 AddOp = FI; SubOp = TI;
8477 } else if (FI->getOpcode() == Instruction::Sub &&
8478 TI->getOpcode() == Instruction::Add) {
8479 AddOp = TI; SubOp = FI;
8483 Value *OtherAddOp = 0;
8484 if (SubOp->getOperand(0) == AddOp->getOperand(0)) {
8485 OtherAddOp = AddOp->getOperand(1);
8486 } else if (SubOp->getOperand(0) == AddOp->getOperand(1)) {
8487 OtherAddOp = AddOp->getOperand(0);
8491 // So at this point we know we have (Y -> OtherAddOp):
8492 // select C, (add X, Y), (sub X, Z)
8493 Value *NegVal; // Compute -Z
8494 if (Constant *C = dyn_cast<Constant>(SubOp->getOperand(1))) {
8495 NegVal = ConstantExpr::getNeg(C);
8497 NegVal = InsertNewInstBefore(
8498 BinaryOperator::CreateNeg(SubOp->getOperand(1), "tmp"), SI);
8501 Value *NewTrueOp = OtherAddOp;
8502 Value *NewFalseOp = NegVal;
8504 std::swap(NewTrueOp, NewFalseOp);
8505 Instruction *NewSel =
8506 SelectInst::Create(CondVal, NewTrueOp,
8507 NewFalseOp, SI.getName() + ".p");
8509 NewSel = InsertNewInstBefore(NewSel, SI);
8510 return BinaryOperator::CreateAdd(SubOp->getOperand(0), NewSel);
8515 // See if we can fold the select into one of our operands.
8516 if (SI.getType()->isInteger()) {
8517 // See the comment above GetSelectFoldableOperands for a description of the
8518 // transformation we are doing here.
8519 if (Instruction *TVI = dyn_cast<Instruction>(TrueVal))
8520 if (TVI->hasOneUse() && TVI->getNumOperands() == 2 &&
8521 !isa<Constant>(FalseVal))
8522 if (unsigned SFO = GetSelectFoldableOperands(TVI)) {
8523 unsigned OpToFold = 0;
8524 if ((SFO & 1) && FalseVal == TVI->getOperand(0)) {
8526 } else if ((SFO & 2) && FalseVal == TVI->getOperand(1)) {
8531 Constant *C = GetSelectFoldableConstant(TVI);
8532 Instruction *NewSel =
8533 SelectInst::Create(SI.getCondition(),
8534 TVI->getOperand(2-OpToFold), C);
8535 InsertNewInstBefore(NewSel, SI);
8536 NewSel->takeName(TVI);
8537 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TVI))
8538 return BinaryOperator::Create(BO->getOpcode(), FalseVal, NewSel);
8540 assert(0 && "Unknown instruction!!");
8545 if (Instruction *FVI = dyn_cast<Instruction>(FalseVal))
8546 if (FVI->hasOneUse() && FVI->getNumOperands() == 2 &&
8547 !isa<Constant>(TrueVal))
8548 if (unsigned SFO = GetSelectFoldableOperands(FVI)) {
8549 unsigned OpToFold = 0;
8550 if ((SFO & 1) && TrueVal == FVI->getOperand(0)) {
8552 } else if ((SFO & 2) && TrueVal == FVI->getOperand(1)) {
8557 Constant *C = GetSelectFoldableConstant(FVI);
8558 Instruction *NewSel =
8559 SelectInst::Create(SI.getCondition(), C,
8560 FVI->getOperand(2-OpToFold));
8561 InsertNewInstBefore(NewSel, SI);
8562 NewSel->takeName(FVI);
8563 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(FVI))
8564 return BinaryOperator::Create(BO->getOpcode(), TrueVal, NewSel);
8566 assert(0 && "Unknown instruction!!");
8571 if (BinaryOperator::isNot(CondVal)) {
8572 SI.setOperand(0, BinaryOperator::getNotArgument(CondVal));
8573 SI.setOperand(1, FalseVal);
8574 SI.setOperand(2, TrueVal);
8581 /// EnforceKnownAlignment - If the specified pointer points to an object that
8582 /// we control, modify the object's alignment to PrefAlign. This isn't
8583 /// often possible though. If alignment is important, a more reliable approach
8584 /// is to simply align all global variables and allocation instructions to
8585 /// their preferred alignment from the beginning.
8587 static unsigned EnforceKnownAlignment(Value *V,
8588 unsigned Align, unsigned PrefAlign) {
8590 User *U = dyn_cast<User>(V);
8591 if (!U) return Align;
8593 switch (getOpcode(U)) {
8595 case Instruction::BitCast:
8596 return EnforceKnownAlignment(U->getOperand(0), Align, PrefAlign);
8597 case Instruction::GetElementPtr: {
8598 // If all indexes are zero, it is just the alignment of the base pointer.
8599 bool AllZeroOperands = true;
8600 for (User::op_iterator i = U->op_begin() + 1, e = U->op_end(); i != e; ++i)
8601 if (!isa<Constant>(*i) ||
8602 !cast<Constant>(*i)->isNullValue()) {
8603 AllZeroOperands = false;
8607 if (AllZeroOperands) {
8608 // Treat this like a bitcast.
8609 return EnforceKnownAlignment(U->getOperand(0), Align, PrefAlign);
8615 if (GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
8616 // If there is a large requested alignment and we can, bump up the alignment
8618 if (!GV->isDeclaration()) {
8619 GV->setAlignment(PrefAlign);
8622 } else if (AllocationInst *AI = dyn_cast<AllocationInst>(V)) {
8623 // If there is a requested alignment and if this is an alloca, round up. We
8624 // don't do this for malloc, because some systems can't respect the request.
8625 if (isa<AllocaInst>(AI)) {
8626 AI->setAlignment(PrefAlign);
8634 /// GetOrEnforceKnownAlignment - If the specified pointer has an alignment that
8635 /// we can determine, return it, otherwise return 0. If PrefAlign is specified,
8636 /// and it is more than the alignment of the ultimate object, see if we can
8637 /// increase the alignment of the ultimate object, making this check succeed.
8638 unsigned InstCombiner::GetOrEnforceKnownAlignment(Value *V,
8639 unsigned PrefAlign) {
8640 unsigned BitWidth = TD ? TD->getTypeSizeInBits(V->getType()) :
8641 sizeof(PrefAlign) * CHAR_BIT;
8642 APInt Mask = APInt::getAllOnesValue(BitWidth);
8643 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
8644 ComputeMaskedBits(V, Mask, KnownZero, KnownOne);
8645 unsigned TrailZ = KnownZero.countTrailingOnes();
8646 unsigned Align = 1u << std::min(BitWidth - 1, TrailZ);
8648 if (PrefAlign > Align)
8649 Align = EnforceKnownAlignment(V, Align, PrefAlign);
8651 // We don't need to make any adjustment.
8655 Instruction *InstCombiner::SimplifyMemTransfer(MemIntrinsic *MI) {
8656 unsigned DstAlign = GetOrEnforceKnownAlignment(MI->getOperand(1));
8657 unsigned SrcAlign = GetOrEnforceKnownAlignment(MI->getOperand(2));
8658 unsigned MinAlign = std::min(DstAlign, SrcAlign);
8659 unsigned CopyAlign = MI->getAlignment()->getZExtValue();
8661 if (CopyAlign < MinAlign) {
8662 MI->setAlignment(ConstantInt::get(Type::Int32Ty, MinAlign));
8666 // If MemCpyInst length is 1/2/4/8 bytes then replace memcpy with
8668 ConstantInt *MemOpLength = dyn_cast<ConstantInt>(MI->getOperand(3));
8669 if (MemOpLength == 0) return 0;
8671 // Source and destination pointer types are always "i8*" for intrinsic. See
8672 // if the size is something we can handle with a single primitive load/store.
8673 // A single load+store correctly handles overlapping memory in the memmove
8675 unsigned Size = MemOpLength->getZExtValue();
8676 if (Size == 0) return MI; // Delete this mem transfer.
8678 if (Size > 8 || (Size&(Size-1)))
8679 return 0; // If not 1/2/4/8 bytes, exit.
8681 // Use an integer load+store unless we can find something better.
8682 Type *NewPtrTy = PointerType::getUnqual(IntegerType::get(Size<<3));
8684 // Memcpy forces the use of i8* for the source and destination. That means
8685 // that if you're using memcpy to move one double around, you'll get a cast
8686 // from double* to i8*. We'd much rather use a double load+store rather than
8687 // an i64 load+store, here because this improves the odds that the source or
8688 // dest address will be promotable. See if we can find a better type than the
8689 // integer datatype.
8690 if (Value *Op = getBitCastOperand(MI->getOperand(1))) {
8691 const Type *SrcETy = cast<PointerType>(Op->getType())->getElementType();
8692 if (SrcETy->isSized() && TD->getTypeStoreSize(SrcETy) == Size) {
8693 // The SrcETy might be something like {{{double}}} or [1 x double]. Rip
8694 // down through these levels if so.
8695 while (!SrcETy->isSingleValueType()) {
8696 if (const StructType *STy = dyn_cast<StructType>(SrcETy)) {
8697 if (STy->getNumElements() == 1)
8698 SrcETy = STy->getElementType(0);
8701 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(SrcETy)) {
8702 if (ATy->getNumElements() == 1)
8703 SrcETy = ATy->getElementType();
8710 if (SrcETy->isSingleValueType())
8711 NewPtrTy = PointerType::getUnqual(SrcETy);
8716 // If the memcpy/memmove provides better alignment info than we can
8718 SrcAlign = std::max(SrcAlign, CopyAlign);
8719 DstAlign = std::max(DstAlign, CopyAlign);
8721 Value *Src = InsertBitCastBefore(MI->getOperand(2), NewPtrTy, *MI);
8722 Value *Dest = InsertBitCastBefore(MI->getOperand(1), NewPtrTy, *MI);
8723 Instruction *L = new LoadInst(Src, "tmp", false, SrcAlign);
8724 InsertNewInstBefore(L, *MI);
8725 InsertNewInstBefore(new StoreInst(L, Dest, false, DstAlign), *MI);
8727 // Set the size of the copy to 0, it will be deleted on the next iteration.
8728 MI->setOperand(3, Constant::getNullValue(MemOpLength->getType()));
8732 Instruction *InstCombiner::SimplifyMemSet(MemSetInst *MI) {
8733 unsigned Alignment = GetOrEnforceKnownAlignment(MI->getDest());
8734 if (MI->getAlignment()->getZExtValue() < Alignment) {
8735 MI->setAlignment(ConstantInt::get(Type::Int32Ty, Alignment));
8739 // Extract the length and alignment and fill if they are constant.
8740 ConstantInt *LenC = dyn_cast<ConstantInt>(MI->getLength());
8741 ConstantInt *FillC = dyn_cast<ConstantInt>(MI->getValue());
8742 if (!LenC || !FillC || FillC->getType() != Type::Int8Ty)
8744 uint64_t Len = LenC->getZExtValue();
8745 Alignment = MI->getAlignment()->getZExtValue();
8747 // If the length is zero, this is a no-op
8748 if (Len == 0) return MI; // memset(d,c,0,a) -> noop
8750 // memset(s,c,n) -> store s, c (for n=1,2,4,8)
8751 if (Len <= 8 && isPowerOf2_32((uint32_t)Len)) {
8752 const Type *ITy = IntegerType::get(Len*8); // n=1 -> i8.
8754 Value *Dest = MI->getDest();
8755 Dest = InsertBitCastBefore(Dest, PointerType::getUnqual(ITy), *MI);
8757 // Alignment 0 is identity for alignment 1 for memset, but not store.
8758 if (Alignment == 0) Alignment = 1;
8760 // Extract the fill value and store.
8761 uint64_t Fill = FillC->getZExtValue()*0x0101010101010101ULL;
8762 InsertNewInstBefore(new StoreInst(ConstantInt::get(ITy, Fill), Dest, false,
8765 // Set the size of the copy to 0, it will be deleted on the next iteration.
8766 MI->setLength(Constant::getNullValue(LenC->getType()));
8774 /// visitCallInst - CallInst simplification. This mostly only handles folding
8775 /// of intrinsic instructions. For normal calls, it allows visitCallSite to do
8776 /// the heavy lifting.
8778 Instruction *InstCombiner::visitCallInst(CallInst &CI) {
8779 IntrinsicInst *II = dyn_cast<IntrinsicInst>(&CI);
8780 if (!II) return visitCallSite(&CI);
8782 // Intrinsics cannot occur in an invoke, so handle them here instead of in
8784 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(II)) {
8785 bool Changed = false;
8787 // memmove/cpy/set of zero bytes is a noop.
8788 if (Constant *NumBytes = dyn_cast<Constant>(MI->getLength())) {
8789 if (NumBytes->isNullValue()) return EraseInstFromFunction(CI);
8791 if (ConstantInt *CI = dyn_cast<ConstantInt>(NumBytes))
8792 if (CI->getZExtValue() == 1) {
8793 // Replace the instruction with just byte operations. We would
8794 // transform other cases to loads/stores, but we don't know if
8795 // alignment is sufficient.
8799 // If we have a memmove and the source operation is a constant global,
8800 // then the source and dest pointers can't alias, so we can change this
8801 // into a call to memcpy.
8802 if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(MI)) {
8803 if (GlobalVariable *GVSrc = dyn_cast<GlobalVariable>(MMI->getSource()))
8804 if (GVSrc->isConstant()) {
8805 Module *M = CI.getParent()->getParent()->getParent();
8806 Intrinsic::ID MemCpyID;
8807 if (CI.getOperand(3)->getType() == Type::Int32Ty)
8808 MemCpyID = Intrinsic::memcpy_i32;
8810 MemCpyID = Intrinsic::memcpy_i64;
8811 CI.setOperand(0, Intrinsic::getDeclaration(M, MemCpyID));
8815 // memmove(x,x,size) -> noop.
8816 if (MMI->getSource() == MMI->getDest())
8817 return EraseInstFromFunction(CI);
8820 // If we can determine a pointer alignment that is bigger than currently
8821 // set, update the alignment.
8822 if (isa<MemCpyInst>(MI) || isa<MemMoveInst>(MI)) {
8823 if (Instruction *I = SimplifyMemTransfer(MI))
8825 } else if (MemSetInst *MSI = dyn_cast<MemSetInst>(MI)) {
8826 if (Instruction *I = SimplifyMemSet(MSI))
8830 if (Changed) return II;
8833 switch (II->getIntrinsicID()) {
8835 case Intrinsic::bswap:
8836 // bswap(bswap(x)) -> x
8837 if (IntrinsicInst *Operand = dyn_cast<IntrinsicInst>(II->getOperand(1)))
8838 if (Operand->getIntrinsicID() == Intrinsic::bswap)
8839 return ReplaceInstUsesWith(CI, Operand->getOperand(1));
8841 case Intrinsic::ppc_altivec_lvx:
8842 case Intrinsic::ppc_altivec_lvxl:
8843 case Intrinsic::x86_sse_loadu_ps:
8844 case Intrinsic::x86_sse2_loadu_pd:
8845 case Intrinsic::x86_sse2_loadu_dq:
8846 // Turn PPC lvx -> load if the pointer is known aligned.
8847 // Turn X86 loadups -> load if the pointer is known aligned.
8848 if (GetOrEnforceKnownAlignment(II->getOperand(1), 16) >= 16) {
8849 Value *Ptr = InsertBitCastBefore(II->getOperand(1),
8850 PointerType::getUnqual(II->getType()),
8852 return new LoadInst(Ptr);
8855 case Intrinsic::ppc_altivec_stvx:
8856 case Intrinsic::ppc_altivec_stvxl:
8857 // Turn stvx -> store if the pointer is known aligned.
8858 if (GetOrEnforceKnownAlignment(II->getOperand(2), 16) >= 16) {
8859 const Type *OpPtrTy =
8860 PointerType::getUnqual(II->getOperand(1)->getType());
8861 Value *Ptr = InsertBitCastBefore(II->getOperand(2), OpPtrTy, CI);
8862 return new StoreInst(II->getOperand(1), Ptr);
8865 case Intrinsic::x86_sse_storeu_ps:
8866 case Intrinsic::x86_sse2_storeu_pd:
8867 case Intrinsic::x86_sse2_storeu_dq:
8868 // Turn X86 storeu -> store if the pointer is known aligned.
8869 if (GetOrEnforceKnownAlignment(II->getOperand(1), 16) >= 16) {
8870 const Type *OpPtrTy =
8871 PointerType::getUnqual(II->getOperand(2)->getType());
8872 Value *Ptr = InsertBitCastBefore(II->getOperand(1), OpPtrTy, CI);
8873 return new StoreInst(II->getOperand(2), Ptr);
8877 case Intrinsic::x86_sse_cvttss2si: {
8878 // These intrinsics only demands the 0th element of its input vector. If
8879 // we can simplify the input based on that, do so now.
8881 if (Value *V = SimplifyDemandedVectorElts(II->getOperand(1), 1,
8883 II->setOperand(1, V);
8889 case Intrinsic::ppc_altivec_vperm:
8890 // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant.
8891 if (ConstantVector *Mask = dyn_cast<ConstantVector>(II->getOperand(3))) {
8892 assert(Mask->getNumOperands() == 16 && "Bad type for intrinsic!");
8894 // Check that all of the elements are integer constants or undefs.
8895 bool AllEltsOk = true;
8896 for (unsigned i = 0; i != 16; ++i) {
8897 if (!isa<ConstantInt>(Mask->getOperand(i)) &&
8898 !isa<UndefValue>(Mask->getOperand(i))) {
8905 // Cast the input vectors to byte vectors.
8906 Value *Op0 =InsertBitCastBefore(II->getOperand(1),Mask->getType(),CI);
8907 Value *Op1 =InsertBitCastBefore(II->getOperand(2),Mask->getType(),CI);
8908 Value *Result = UndefValue::get(Op0->getType());
8910 // Only extract each element once.
8911 Value *ExtractedElts[32];
8912 memset(ExtractedElts, 0, sizeof(ExtractedElts));
8914 for (unsigned i = 0; i != 16; ++i) {
8915 if (isa<UndefValue>(Mask->getOperand(i)))
8917 unsigned Idx=cast<ConstantInt>(Mask->getOperand(i))->getZExtValue();
8918 Idx &= 31; // Match the hardware behavior.
8920 if (ExtractedElts[Idx] == 0) {
8922 new ExtractElementInst(Idx < 16 ? Op0 : Op1, Idx&15, "tmp");
8923 InsertNewInstBefore(Elt, CI);
8924 ExtractedElts[Idx] = Elt;
8927 // Insert this value into the result vector.
8928 Result = InsertElementInst::Create(Result, ExtractedElts[Idx],
8930 InsertNewInstBefore(cast<Instruction>(Result), CI);
8932 return CastInst::Create(Instruction::BitCast, Result, CI.getType());
8937 case Intrinsic::stackrestore: {
8938 // If the save is right next to the restore, remove the restore. This can
8939 // happen when variable allocas are DCE'd.
8940 if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getOperand(1))) {
8941 if (SS->getIntrinsicID() == Intrinsic::stacksave) {
8942 BasicBlock::iterator BI = SS;
8944 return EraseInstFromFunction(CI);
8948 // Scan down this block to see if there is another stack restore in the
8949 // same block without an intervening call/alloca.
8950 BasicBlock::iterator BI = II;
8951 TerminatorInst *TI = II->getParent()->getTerminator();
8952 bool CannotRemove = false;
8953 for (++BI; &*BI != TI; ++BI) {
8954 if (isa<AllocaInst>(BI)) {
8955 CannotRemove = true;
8958 if (CallInst *BCI = dyn_cast<CallInst>(BI)) {
8959 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(BCI)) {
8960 // If there is a stackrestore below this one, remove this one.
8961 if (II->getIntrinsicID() == Intrinsic::stackrestore)
8962 return EraseInstFromFunction(CI);
8963 // Otherwise, ignore the intrinsic.
8965 // If we found a non-intrinsic call, we can't remove the stack
8967 CannotRemove = true;
8973 // If the stack restore is in a return/unwind block and if there are no
8974 // allocas or calls between the restore and the return, nuke the restore.
8975 if (!CannotRemove && (isa<ReturnInst>(TI) || isa<UnwindInst>(TI)))
8976 return EraseInstFromFunction(CI);
8981 return visitCallSite(II);
8984 // InvokeInst simplification
8986 Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) {
8987 return visitCallSite(&II);
8990 /// isSafeToEliminateVarargsCast - If this cast does not affect the value
8991 /// passed through the varargs area, we can eliminate the use of the cast.
8992 static bool isSafeToEliminateVarargsCast(const CallSite CS,
8993 const CastInst * const CI,
8994 const TargetData * const TD,
8996 if (!CI->isLosslessCast())
8999 // The size of ByVal arguments is derived from the type, so we
9000 // can't change to a type with a different size. If the size were
9001 // passed explicitly we could avoid this check.
9002 if (!CS.paramHasAttr(ix, ParamAttr::ByVal))
9006 cast<PointerType>(CI->getOperand(0)->getType())->getElementType();
9007 const Type* DstTy = cast<PointerType>(CI->getType())->getElementType();
9008 if (!SrcTy->isSized() || !DstTy->isSized())
9010 if (TD->getABITypeSize(SrcTy) != TD->getABITypeSize(DstTy))
9015 // visitCallSite - Improvements for call and invoke instructions.
9017 Instruction *InstCombiner::visitCallSite(CallSite CS) {
9018 bool Changed = false;
9020 // If the callee is a constexpr cast of a function, attempt to move the cast
9021 // to the arguments of the call/invoke.
9022 if (transformConstExprCastCall(CS)) return 0;
9024 Value *Callee = CS.getCalledValue();
9026 if (Function *CalleeF = dyn_cast<Function>(Callee))
9027 if (CalleeF->getCallingConv() != CS.getCallingConv()) {
9028 Instruction *OldCall = CS.getInstruction();
9029 // If the call and callee calling conventions don't match, this call must
9030 // be unreachable, as the call is undefined.
9031 new StoreInst(ConstantInt::getTrue(),
9032 UndefValue::get(PointerType::getUnqual(Type::Int1Ty)),
9034 if (!OldCall->use_empty())
9035 OldCall->replaceAllUsesWith(UndefValue::get(OldCall->getType()));
9036 if (isa<CallInst>(OldCall)) // Not worth removing an invoke here.
9037 return EraseInstFromFunction(*OldCall);
9041 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) {
9042 // This instruction is not reachable, just remove it. We insert a store to
9043 // undef so that we know that this code is not reachable, despite the fact
9044 // that we can't modify the CFG here.
9045 new StoreInst(ConstantInt::getTrue(),
9046 UndefValue::get(PointerType::getUnqual(Type::Int1Ty)),
9047 CS.getInstruction());
9049 if (!CS.getInstruction()->use_empty())
9050 CS.getInstruction()->
9051 replaceAllUsesWith(UndefValue::get(CS.getInstruction()->getType()));
9053 if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) {
9054 // Don't break the CFG, insert a dummy cond branch.
9055 BranchInst::Create(II->getNormalDest(), II->getUnwindDest(),
9056 ConstantInt::getTrue(), II);
9058 return EraseInstFromFunction(*CS.getInstruction());
9061 if (BitCastInst *BC = dyn_cast<BitCastInst>(Callee))
9062 if (IntrinsicInst *In = dyn_cast<IntrinsicInst>(BC->getOperand(0)))
9063 if (In->getIntrinsicID() == Intrinsic::init_trampoline)
9064 return transformCallThroughTrampoline(CS);
9066 const PointerType *PTy = cast<PointerType>(Callee->getType());
9067 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
9068 if (FTy->isVarArg()) {
9069 int ix = FTy->getNumParams() + (isa<InvokeInst>(Callee) ? 3 : 1);
9070 // See if we can optimize any arguments passed through the varargs area of
9072 for (CallSite::arg_iterator I = CS.arg_begin()+FTy->getNumParams(),
9073 E = CS.arg_end(); I != E; ++I, ++ix) {
9074 CastInst *CI = dyn_cast<CastInst>(*I);
9075 if (CI && isSafeToEliminateVarargsCast(CS, CI, TD, ix)) {
9076 *I = CI->getOperand(0);
9082 if (isa<InlineAsm>(Callee) && !CS.doesNotThrow()) {
9083 // Inline asm calls cannot throw - mark them 'nounwind'.
9084 CS.setDoesNotThrow();
9088 return Changed ? CS.getInstruction() : 0;
9091 // transformConstExprCastCall - If the callee is a constexpr cast of a function,
9092 // attempt to move the cast to the arguments of the call/invoke.
9094 bool InstCombiner::transformConstExprCastCall(CallSite CS) {
9095 if (!isa<ConstantExpr>(CS.getCalledValue())) return false;
9096 ConstantExpr *CE = cast<ConstantExpr>(CS.getCalledValue());
9097 if (CE->getOpcode() != Instruction::BitCast ||
9098 !isa<Function>(CE->getOperand(0)))
9100 Function *Callee = cast<Function>(CE->getOperand(0));
9101 Instruction *Caller = CS.getInstruction();
9102 const PAListPtr &CallerPAL = CS.getParamAttrs();
9104 // Okay, this is a cast from a function to a different type. Unless doing so
9105 // would cause a type conversion of one of our arguments, change this call to
9106 // be a direct call with arguments casted to the appropriate types.
9108 const FunctionType *FT = Callee->getFunctionType();
9109 const Type *OldRetTy = Caller->getType();
9110 const Type *NewRetTy = FT->getReturnType();
9112 if (isa<StructType>(NewRetTy))
9113 return false; // TODO: Handle multiple return values.
9115 // Check to see if we are changing the return type...
9116 if (OldRetTy != NewRetTy) {
9117 if (Callee->isDeclaration() &&
9118 // Conversion is ok if changing from one pointer type to another or from
9119 // a pointer to an integer of the same size.
9120 !((isa<PointerType>(OldRetTy) || OldRetTy == TD->getIntPtrType()) &&
9121 (isa<PointerType>(NewRetTy) || NewRetTy == TD->getIntPtrType())))
9122 return false; // Cannot transform this return value.
9124 if (!Caller->use_empty() &&
9125 // void -> non-void is handled specially
9126 NewRetTy != Type::VoidTy && !CastInst::isCastable(NewRetTy, OldRetTy))
9127 return false; // Cannot transform this return value.
9129 if (!CallerPAL.isEmpty() && !Caller->use_empty()) {
9130 ParameterAttributes RAttrs = CallerPAL.getParamAttrs(0);
9131 if (RAttrs & ParamAttr::typeIncompatible(NewRetTy))
9132 return false; // Attribute not compatible with transformed value.
9135 // If the callsite is an invoke instruction, and the return value is used by
9136 // a PHI node in a successor, we cannot change the return type of the call
9137 // because there is no place to put the cast instruction (without breaking
9138 // the critical edge). Bail out in this case.
9139 if (!Caller->use_empty())
9140 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller))
9141 for (Value::use_iterator UI = II->use_begin(), E = II->use_end();
9143 if (PHINode *PN = dyn_cast<PHINode>(*UI))
9144 if (PN->getParent() == II->getNormalDest() ||
9145 PN->getParent() == II->getUnwindDest())
9149 unsigned NumActualArgs = unsigned(CS.arg_end()-CS.arg_begin());
9150 unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs);
9152 CallSite::arg_iterator AI = CS.arg_begin();
9153 for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) {
9154 const Type *ParamTy = FT->getParamType(i);
9155 const Type *ActTy = (*AI)->getType();
9157 if (!CastInst::isCastable(ActTy, ParamTy))
9158 return false; // Cannot transform this parameter value.
9160 if (CallerPAL.getParamAttrs(i + 1) & ParamAttr::typeIncompatible(ParamTy))
9161 return false; // Attribute not compatible with transformed value.
9163 // Converting from one pointer type to another or between a pointer and an
9164 // integer of the same size is safe even if we do not have a body.
9165 bool isConvertible = ActTy == ParamTy ||
9166 ((isa<PointerType>(ParamTy) || ParamTy == TD->getIntPtrType()) &&
9167 (isa<PointerType>(ActTy) || ActTy == TD->getIntPtrType()));
9168 if (Callee->isDeclaration() && !isConvertible) return false;
9171 if (FT->getNumParams() < NumActualArgs && !FT->isVarArg() &&
9172 Callee->isDeclaration())
9173 return false; // Do not delete arguments unless we have a function body.
9175 if (FT->getNumParams() < NumActualArgs && FT->isVarArg() &&
9176 !CallerPAL.isEmpty())
9177 // In this case we have more arguments than the new function type, but we
9178 // won't be dropping them. Check that these extra arguments have attributes
9179 // that are compatible with being a vararg call argument.
9180 for (unsigned i = CallerPAL.getNumSlots(); i; --i) {
9181 if (CallerPAL.getSlot(i - 1).Index <= FT->getNumParams())
9183 ParameterAttributes PAttrs = CallerPAL.getSlot(i - 1).Attrs;
9184 if (PAttrs & ParamAttr::VarArgsIncompatible)
9188 // Okay, we decided that this is a safe thing to do: go ahead and start
9189 // inserting cast instructions as necessary...
9190 std::vector<Value*> Args;
9191 Args.reserve(NumActualArgs);
9192 SmallVector<ParamAttrsWithIndex, 8> attrVec;
9193 attrVec.reserve(NumCommonArgs);
9195 // Get any return attributes.
9196 ParameterAttributes RAttrs = CallerPAL.getParamAttrs(0);
9198 // If the return value is not being used, the type may not be compatible
9199 // with the existing attributes. Wipe out any problematic attributes.
9200 RAttrs &= ~ParamAttr::typeIncompatible(NewRetTy);
9202 // Add the new return attributes.
9204 attrVec.push_back(ParamAttrsWithIndex::get(0, RAttrs));
9206 AI = CS.arg_begin();
9207 for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) {
9208 const Type *ParamTy = FT->getParamType(i);
9209 if ((*AI)->getType() == ParamTy) {
9210 Args.push_back(*AI);
9212 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI,
9213 false, ParamTy, false);
9214 CastInst *NewCast = CastInst::Create(opcode, *AI, ParamTy, "tmp");
9215 Args.push_back(InsertNewInstBefore(NewCast, *Caller));
9218 // Add any parameter attributes.
9219 if (ParameterAttributes PAttrs = CallerPAL.getParamAttrs(i + 1))
9220 attrVec.push_back(ParamAttrsWithIndex::get(i + 1, PAttrs));
9223 // If the function takes more arguments than the call was taking, add them
9225 for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i)
9226 Args.push_back(Constant::getNullValue(FT->getParamType(i)));
9228 // If we are removing arguments to the function, emit an obnoxious warning...
9229 if (FT->getNumParams() < NumActualArgs) {
9230 if (!FT->isVarArg()) {
9231 cerr << "WARNING: While resolving call to function '"
9232 << Callee->getName() << "' arguments were dropped!\n";
9234 // Add all of the arguments in their promoted form to the arg list...
9235 for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) {
9236 const Type *PTy = getPromotedType((*AI)->getType());
9237 if (PTy != (*AI)->getType()) {
9238 // Must promote to pass through va_arg area!
9239 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI, false,
9241 Instruction *Cast = CastInst::Create(opcode, *AI, PTy, "tmp");
9242 InsertNewInstBefore(Cast, *Caller);
9243 Args.push_back(Cast);
9245 Args.push_back(*AI);
9248 // Add any parameter attributes.
9249 if (ParameterAttributes PAttrs = CallerPAL.getParamAttrs(i + 1))
9250 attrVec.push_back(ParamAttrsWithIndex::get(i + 1, PAttrs));
9255 if (NewRetTy == Type::VoidTy)
9256 Caller->setName(""); // Void type should not have a name.
9258 const PAListPtr &NewCallerPAL = PAListPtr::get(attrVec.begin(),attrVec.end());
9261 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
9262 NC = InvokeInst::Create(Callee, II->getNormalDest(), II->getUnwindDest(),
9263 Args.begin(), Args.end(),
9264 Caller->getName(), Caller);
9265 cast<InvokeInst>(NC)->setCallingConv(II->getCallingConv());
9266 cast<InvokeInst>(NC)->setParamAttrs(NewCallerPAL);
9268 NC = CallInst::Create(Callee, Args.begin(), Args.end(),
9269 Caller->getName(), Caller);
9270 CallInst *CI = cast<CallInst>(Caller);
9271 if (CI->isTailCall())
9272 cast<CallInst>(NC)->setTailCall();
9273 cast<CallInst>(NC)->setCallingConv(CI->getCallingConv());
9274 cast<CallInst>(NC)->setParamAttrs(NewCallerPAL);
9277 // Insert a cast of the return type as necessary.
9279 if (OldRetTy != NV->getType() && !Caller->use_empty()) {
9280 if (NV->getType() != Type::VoidTy) {
9281 Instruction::CastOps opcode = CastInst::getCastOpcode(NC, false,
9283 NV = NC = CastInst::Create(opcode, NC, OldRetTy, "tmp");
9285 // If this is an invoke instruction, we should insert it after the first
9286 // non-phi, instruction in the normal successor block.
9287 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
9288 BasicBlock::iterator I = II->getNormalDest()->getFirstNonPHI();
9289 InsertNewInstBefore(NC, *I);
9291 // Otherwise, it's a call, just insert cast right after the call instr
9292 InsertNewInstBefore(NC, *Caller);
9294 AddUsersToWorkList(*Caller);
9296 NV = UndefValue::get(Caller->getType());
9300 if (Caller->getType() != Type::VoidTy && !Caller->use_empty())
9301 Caller->replaceAllUsesWith(NV);
9302 Caller->eraseFromParent();
9303 RemoveFromWorkList(Caller);
9307 // transformCallThroughTrampoline - Turn a call to a function created by the
9308 // init_trampoline intrinsic into a direct call to the underlying function.
9310 Instruction *InstCombiner::transformCallThroughTrampoline(CallSite CS) {
9311 Value *Callee = CS.getCalledValue();
9312 const PointerType *PTy = cast<PointerType>(Callee->getType());
9313 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
9314 const PAListPtr &Attrs = CS.getParamAttrs();
9316 // If the call already has the 'nest' attribute somewhere then give up -
9317 // otherwise 'nest' would occur twice after splicing in the chain.
9318 if (Attrs.hasAttrSomewhere(ParamAttr::Nest))
9321 IntrinsicInst *Tramp =
9322 cast<IntrinsicInst>(cast<BitCastInst>(Callee)->getOperand(0));
9324 Function *NestF = cast<Function>(Tramp->getOperand(2)->stripPointerCasts());
9325 const PointerType *NestFPTy = cast<PointerType>(NestF->getType());
9326 const FunctionType *NestFTy = cast<FunctionType>(NestFPTy->getElementType());
9328 const PAListPtr &NestAttrs = NestF->getParamAttrs();
9329 if (!NestAttrs.isEmpty()) {
9330 unsigned NestIdx = 1;
9331 const Type *NestTy = 0;
9332 ParameterAttributes NestAttr = ParamAttr::None;
9334 // Look for a parameter marked with the 'nest' attribute.
9335 for (FunctionType::param_iterator I = NestFTy->param_begin(),
9336 E = NestFTy->param_end(); I != E; ++NestIdx, ++I)
9337 if (NestAttrs.paramHasAttr(NestIdx, ParamAttr::Nest)) {
9338 // Record the parameter type and any other attributes.
9340 NestAttr = NestAttrs.getParamAttrs(NestIdx);
9345 Instruction *Caller = CS.getInstruction();
9346 std::vector<Value*> NewArgs;
9347 NewArgs.reserve(unsigned(CS.arg_end()-CS.arg_begin())+1);
9349 SmallVector<ParamAttrsWithIndex, 8> NewAttrs;
9350 NewAttrs.reserve(Attrs.getNumSlots() + 1);
9352 // Insert the nest argument into the call argument list, which may
9353 // mean appending it. Likewise for attributes.
9355 // Add any function result attributes.
9356 if (ParameterAttributes Attr = Attrs.getParamAttrs(0))
9357 NewAttrs.push_back(ParamAttrsWithIndex::get(0, Attr));
9361 CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
9363 if (Idx == NestIdx) {
9364 // Add the chain argument and attributes.
9365 Value *NestVal = Tramp->getOperand(3);
9366 if (NestVal->getType() != NestTy)
9367 NestVal = new BitCastInst(NestVal, NestTy, "nest", Caller);
9368 NewArgs.push_back(NestVal);
9369 NewAttrs.push_back(ParamAttrsWithIndex::get(NestIdx, NestAttr));
9375 // Add the original argument and attributes.
9376 NewArgs.push_back(*I);
9377 if (ParameterAttributes Attr = Attrs.getParamAttrs(Idx))
9379 (ParamAttrsWithIndex::get(Idx + (Idx >= NestIdx), Attr));
9385 // The trampoline may have been bitcast to a bogus type (FTy).
9386 // Handle this by synthesizing a new function type, equal to FTy
9387 // with the chain parameter inserted.
9389 std::vector<const Type*> NewTypes;
9390 NewTypes.reserve(FTy->getNumParams()+1);
9392 // Insert the chain's type into the list of parameter types, which may
9393 // mean appending it.
9396 FunctionType::param_iterator I = FTy->param_begin(),
9397 E = FTy->param_end();
9401 // Add the chain's type.
9402 NewTypes.push_back(NestTy);
9407 // Add the original type.
9408 NewTypes.push_back(*I);
9414 // Replace the trampoline call with a direct call. Let the generic
9415 // code sort out any function type mismatches.
9416 FunctionType *NewFTy =
9417 FunctionType::get(FTy->getReturnType(), NewTypes, FTy->isVarArg());
9418 Constant *NewCallee = NestF->getType() == PointerType::getUnqual(NewFTy) ?
9419 NestF : ConstantExpr::getBitCast(NestF, PointerType::getUnqual(NewFTy));
9420 const PAListPtr &NewPAL = PAListPtr::get(NewAttrs.begin(),NewAttrs.end());
9422 Instruction *NewCaller;
9423 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
9424 NewCaller = InvokeInst::Create(NewCallee,
9425 II->getNormalDest(), II->getUnwindDest(),
9426 NewArgs.begin(), NewArgs.end(),
9427 Caller->getName(), Caller);
9428 cast<InvokeInst>(NewCaller)->setCallingConv(II->getCallingConv());
9429 cast<InvokeInst>(NewCaller)->setParamAttrs(NewPAL);
9431 NewCaller = CallInst::Create(NewCallee, NewArgs.begin(), NewArgs.end(),
9432 Caller->getName(), Caller);
9433 if (cast<CallInst>(Caller)->isTailCall())
9434 cast<CallInst>(NewCaller)->setTailCall();
9435 cast<CallInst>(NewCaller)->
9436 setCallingConv(cast<CallInst>(Caller)->getCallingConv());
9437 cast<CallInst>(NewCaller)->setParamAttrs(NewPAL);
9439 if (Caller->getType() != Type::VoidTy && !Caller->use_empty())
9440 Caller->replaceAllUsesWith(NewCaller);
9441 Caller->eraseFromParent();
9442 RemoveFromWorkList(Caller);
9447 // Replace the trampoline call with a direct call. Since there is no 'nest'
9448 // parameter, there is no need to adjust the argument list. Let the generic
9449 // code sort out any function type mismatches.
9450 Constant *NewCallee =
9451 NestF->getType() == PTy ? NestF : ConstantExpr::getBitCast(NestF, PTy);
9452 CS.setCalledFunction(NewCallee);
9453 return CS.getInstruction();
9456 /// FoldPHIArgBinOpIntoPHI - If we have something like phi [add (a,b), add(c,d)]
9457 /// and if a/b/c/d and the add's all have a single use, turn this into two phi's
9458 /// and a single binop.
9459 Instruction *InstCombiner::FoldPHIArgBinOpIntoPHI(PHINode &PN) {
9460 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
9461 assert(isa<BinaryOperator>(FirstInst) || isa<GetElementPtrInst>(FirstInst) ||
9462 isa<CmpInst>(FirstInst));
9463 unsigned Opc = FirstInst->getOpcode();
9464 Value *LHSVal = FirstInst->getOperand(0);
9465 Value *RHSVal = FirstInst->getOperand(1);
9467 const Type *LHSType = LHSVal->getType();
9468 const Type *RHSType = RHSVal->getType();
9470 // Scan to see if all operands are the same opcode, all have one use, and all
9471 // kill their operands (i.e. the operands have one use).
9472 for (unsigned i = 0; i != PN.getNumIncomingValues(); ++i) {
9473 Instruction *I = dyn_cast<Instruction>(PN.getIncomingValue(i));
9474 if (!I || I->getOpcode() != Opc || !I->hasOneUse() ||
9475 // Verify type of the LHS matches so we don't fold cmp's of different
9476 // types or GEP's with different index types.
9477 I->getOperand(0)->getType() != LHSType ||
9478 I->getOperand(1)->getType() != RHSType)
9481 // If they are CmpInst instructions, check their predicates
9482 if (Opc == Instruction::ICmp || Opc == Instruction::FCmp)
9483 if (cast<CmpInst>(I)->getPredicate() !=
9484 cast<CmpInst>(FirstInst)->getPredicate())
9487 // Keep track of which operand needs a phi node.
9488 if (I->getOperand(0) != LHSVal) LHSVal = 0;
9489 if (I->getOperand(1) != RHSVal) RHSVal = 0;
9492 // Otherwise, this is safe to transform, determine if it is profitable.
9494 // If this is a GEP, and if the index (not the pointer) needs a PHI, bail out.
9495 // Indexes are often folded into load/store instructions, so we don't want to
9496 // hide them behind a phi.
9497 if (isa<GetElementPtrInst>(FirstInst) && RHSVal == 0)
9500 Value *InLHS = FirstInst->getOperand(0);
9501 Value *InRHS = FirstInst->getOperand(1);
9502 PHINode *NewLHS = 0, *NewRHS = 0;
9504 NewLHS = PHINode::Create(LHSType,
9505 FirstInst->getOperand(0)->getName() + ".pn");
9506 NewLHS->reserveOperandSpace(PN.getNumOperands()/2);
9507 NewLHS->addIncoming(InLHS, PN.getIncomingBlock(0));
9508 InsertNewInstBefore(NewLHS, PN);
9513 NewRHS = PHINode::Create(RHSType,
9514 FirstInst->getOperand(1)->getName() + ".pn");
9515 NewRHS->reserveOperandSpace(PN.getNumOperands()/2);
9516 NewRHS->addIncoming(InRHS, PN.getIncomingBlock(0));
9517 InsertNewInstBefore(NewRHS, PN);
9521 // Add all operands to the new PHIs.
9522 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
9524 Value *NewInLHS =cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
9525 NewLHS->addIncoming(NewInLHS, PN.getIncomingBlock(i));
9528 Value *NewInRHS =cast<Instruction>(PN.getIncomingValue(i))->getOperand(1);
9529 NewRHS->addIncoming(NewInRHS, PN.getIncomingBlock(i));
9533 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
9534 return BinaryOperator::Create(BinOp->getOpcode(), LHSVal, RHSVal);
9535 else if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst))
9536 return CmpInst::Create(CIOp->getOpcode(), CIOp->getPredicate(), LHSVal,
9539 assert(isa<GetElementPtrInst>(FirstInst));
9540 return GetElementPtrInst::Create(LHSVal, RHSVal);
9544 /// isSafeToSinkLoad - Return true if we know that it is safe sink the load out
9545 /// of the block that defines it. This means that it must be obvious the value
9546 /// of the load is not changed from the point of the load to the end of the
9549 /// Finally, it is safe, but not profitable, to sink a load targetting a
9550 /// non-address-taken alloca. Doing so will cause us to not promote the alloca
9552 static bool isSafeToSinkLoad(LoadInst *L) {
9553 BasicBlock::iterator BBI = L, E = L->getParent()->end();
9555 for (++BBI; BBI != E; ++BBI)
9556 if (BBI->mayWriteToMemory())
9559 // Check for non-address taken alloca. If not address-taken already, it isn't
9560 // profitable to do this xform.
9561 if (AllocaInst *AI = dyn_cast<AllocaInst>(L->getOperand(0))) {
9562 bool isAddressTaken = false;
9563 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
9565 if (isa<LoadInst>(UI)) continue;
9566 if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
9567 // If storing TO the alloca, then the address isn't taken.
9568 if (SI->getOperand(1) == AI) continue;
9570 isAddressTaken = true;
9574 if (!isAddressTaken)
9582 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
9583 // operator and they all are only used by the PHI, PHI together their
9584 // inputs, and do the operation once, to the result of the PHI.
9585 Instruction *InstCombiner::FoldPHIArgOpIntoPHI(PHINode &PN) {
9586 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
9588 // Scan the instruction, looking for input operations that can be folded away.
9589 // If all input operands to the phi are the same instruction (e.g. a cast from
9590 // the same type or "+42") we can pull the operation through the PHI, reducing
9591 // code size and simplifying code.
9592 Constant *ConstantOp = 0;
9593 const Type *CastSrcTy = 0;
9594 bool isVolatile = false;
9595 if (isa<CastInst>(FirstInst)) {
9596 CastSrcTy = FirstInst->getOperand(0)->getType();
9597 } else if (isa<BinaryOperator>(FirstInst) || isa<CmpInst>(FirstInst)) {
9598 // Can fold binop, compare or shift here if the RHS is a constant,
9599 // otherwise call FoldPHIArgBinOpIntoPHI.
9600 ConstantOp = dyn_cast<Constant>(FirstInst->getOperand(1));
9601 if (ConstantOp == 0)
9602 return FoldPHIArgBinOpIntoPHI(PN);
9603 } else if (LoadInst *LI = dyn_cast<LoadInst>(FirstInst)) {
9604 isVolatile = LI->isVolatile();
9605 // We can't sink the load if the loaded value could be modified between the
9606 // load and the PHI.
9607 if (LI->getParent() != PN.getIncomingBlock(0) ||
9608 !isSafeToSinkLoad(LI))
9611 // If the PHI is of volatile loads and the load block has multiple
9612 // successors, sinking it would remove a load of the volatile value from
9613 // the path through the other successor.
9615 LI->getParent()->getTerminator()->getNumSuccessors() != 1)
9618 } else if (isa<GetElementPtrInst>(FirstInst)) {
9619 if (FirstInst->getNumOperands() == 2)
9620 return FoldPHIArgBinOpIntoPHI(PN);
9621 // Can't handle general GEPs yet.
9624 return 0; // Cannot fold this operation.
9627 // Check to see if all arguments are the same operation.
9628 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
9629 if (!isa<Instruction>(PN.getIncomingValue(i))) return 0;
9630 Instruction *I = cast<Instruction>(PN.getIncomingValue(i));
9631 if (!I->hasOneUse() || !I->isSameOperationAs(FirstInst))
9634 if (I->getOperand(0)->getType() != CastSrcTy)
9635 return 0; // Cast operation must match.
9636 } else if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
9637 // We can't sink the load if the loaded value could be modified between
9638 // the load and the PHI.
9639 if (LI->isVolatile() != isVolatile ||
9640 LI->getParent() != PN.getIncomingBlock(i) ||
9641 !isSafeToSinkLoad(LI))
9644 // If the PHI is of volatile loads and the load block has multiple
9645 // successors, sinking it would remove a load of the volatile value from
9646 // the path through the other successor.
9648 LI->getParent()->getTerminator()->getNumSuccessors() != 1)
9652 } else if (I->getOperand(1) != ConstantOp) {
9657 // Okay, they are all the same operation. Create a new PHI node of the
9658 // correct type, and PHI together all of the LHS's of the instructions.
9659 PHINode *NewPN = PHINode::Create(FirstInst->getOperand(0)->getType(),
9660 PN.getName()+".in");
9661 NewPN->reserveOperandSpace(PN.getNumOperands()/2);
9663 Value *InVal = FirstInst->getOperand(0);
9664 NewPN->addIncoming(InVal, PN.getIncomingBlock(0));
9666 // Add all operands to the new PHI.
9667 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
9668 Value *NewInVal = cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
9669 if (NewInVal != InVal)
9671 NewPN->addIncoming(NewInVal, PN.getIncomingBlock(i));
9676 // The new PHI unions all of the same values together. This is really
9677 // common, so we handle it intelligently here for compile-time speed.
9681 InsertNewInstBefore(NewPN, PN);
9685 // Insert and return the new operation.
9686 if (CastInst* FirstCI = dyn_cast<CastInst>(FirstInst))
9687 return CastInst::Create(FirstCI->getOpcode(), PhiVal, PN.getType());
9688 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
9689 return BinaryOperator::Create(BinOp->getOpcode(), PhiVal, ConstantOp);
9690 if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst))
9691 return CmpInst::Create(CIOp->getOpcode(), CIOp->getPredicate(),
9692 PhiVal, ConstantOp);
9693 assert(isa<LoadInst>(FirstInst) && "Unknown operation");
9695 // If this was a volatile load that we are merging, make sure to loop through
9696 // and mark all the input loads as non-volatile. If we don't do this, we will
9697 // insert a new volatile load and the old ones will not be deletable.
9699 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i)
9700 cast<LoadInst>(PN.getIncomingValue(i))->setVolatile(false);
9702 return new LoadInst(PhiVal, "", isVolatile);
9705 /// DeadPHICycle - Return true if this PHI node is only used by a PHI node cycle
9707 static bool DeadPHICycle(PHINode *PN,
9708 SmallPtrSet<PHINode*, 16> &PotentiallyDeadPHIs) {
9709 if (PN->use_empty()) return true;
9710 if (!PN->hasOneUse()) return false;
9712 // Remember this node, and if we find the cycle, return.
9713 if (!PotentiallyDeadPHIs.insert(PN))
9716 // Don't scan crazily complex things.
9717 if (PotentiallyDeadPHIs.size() == 16)
9720 if (PHINode *PU = dyn_cast<PHINode>(PN->use_back()))
9721 return DeadPHICycle(PU, PotentiallyDeadPHIs);
9726 /// PHIsEqualValue - Return true if this phi node is always equal to
9727 /// NonPhiInVal. This happens with mutually cyclic phi nodes like:
9728 /// z = some value; x = phi (y, z); y = phi (x, z)
9729 static bool PHIsEqualValue(PHINode *PN, Value *NonPhiInVal,
9730 SmallPtrSet<PHINode*, 16> &ValueEqualPHIs) {
9731 // See if we already saw this PHI node.
9732 if (!ValueEqualPHIs.insert(PN))
9735 // Don't scan crazily complex things.
9736 if (ValueEqualPHIs.size() == 16)
9739 // Scan the operands to see if they are either phi nodes or are equal to
9741 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
9742 Value *Op = PN->getIncomingValue(i);
9743 if (PHINode *OpPN = dyn_cast<PHINode>(Op)) {
9744 if (!PHIsEqualValue(OpPN, NonPhiInVal, ValueEqualPHIs))
9746 } else if (Op != NonPhiInVal)
9754 // PHINode simplification
9756 Instruction *InstCombiner::visitPHINode(PHINode &PN) {
9757 // If LCSSA is around, don't mess with Phi nodes
9758 if (MustPreserveLCSSA) return 0;
9760 if (Value *V = PN.hasConstantValue())
9761 return ReplaceInstUsesWith(PN, V);
9763 // If all PHI operands are the same operation, pull them through the PHI,
9764 // reducing code size.
9765 if (isa<Instruction>(PN.getIncomingValue(0)) &&
9766 PN.getIncomingValue(0)->hasOneUse())
9767 if (Instruction *Result = FoldPHIArgOpIntoPHI(PN))
9770 // If this is a trivial cycle in the PHI node graph, remove it. Basically, if
9771 // this PHI only has a single use (a PHI), and if that PHI only has one use (a
9772 // PHI)... break the cycle.
9773 if (PN.hasOneUse()) {
9774 Instruction *PHIUser = cast<Instruction>(PN.use_back());
9775 if (PHINode *PU = dyn_cast<PHINode>(PHIUser)) {
9776 SmallPtrSet<PHINode*, 16> PotentiallyDeadPHIs;
9777 PotentiallyDeadPHIs.insert(&PN);
9778 if (DeadPHICycle(PU, PotentiallyDeadPHIs))
9779 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
9782 // If this phi has a single use, and if that use just computes a value for
9783 // the next iteration of a loop, delete the phi. This occurs with unused
9784 // induction variables, e.g. "for (int j = 0; ; ++j);". Detecting this
9785 // common case here is good because the only other things that catch this
9786 // are induction variable analysis (sometimes) and ADCE, which is only run
9788 if (PHIUser->hasOneUse() &&
9789 (isa<BinaryOperator>(PHIUser) || isa<GetElementPtrInst>(PHIUser)) &&
9790 PHIUser->use_back() == &PN) {
9791 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
9795 // We sometimes end up with phi cycles that non-obviously end up being the
9796 // same value, for example:
9797 // z = some value; x = phi (y, z); y = phi (x, z)
9798 // where the phi nodes don't necessarily need to be in the same block. Do a
9799 // quick check to see if the PHI node only contains a single non-phi value, if
9800 // so, scan to see if the phi cycle is actually equal to that value.
9802 unsigned InValNo = 0, NumOperandVals = PN.getNumIncomingValues();
9803 // Scan for the first non-phi operand.
9804 while (InValNo != NumOperandVals &&
9805 isa<PHINode>(PN.getIncomingValue(InValNo)))
9808 if (InValNo != NumOperandVals) {
9809 Value *NonPhiInVal = PN.getOperand(InValNo);
9811 // Scan the rest of the operands to see if there are any conflicts, if so
9812 // there is no need to recursively scan other phis.
9813 for (++InValNo; InValNo != NumOperandVals; ++InValNo) {
9814 Value *OpVal = PN.getIncomingValue(InValNo);
9815 if (OpVal != NonPhiInVal && !isa<PHINode>(OpVal))
9819 // If we scanned over all operands, then we have one unique value plus
9820 // phi values. Scan PHI nodes to see if they all merge in each other or
9822 if (InValNo == NumOperandVals) {
9823 SmallPtrSet<PHINode*, 16> ValueEqualPHIs;
9824 if (PHIsEqualValue(&PN, NonPhiInVal, ValueEqualPHIs))
9825 return ReplaceInstUsesWith(PN, NonPhiInVal);
9832 static Value *InsertCastToIntPtrTy(Value *V, const Type *DTy,
9833 Instruction *InsertPoint,
9835 unsigned PtrSize = DTy->getPrimitiveSizeInBits();
9836 unsigned VTySize = V->getType()->getPrimitiveSizeInBits();
9837 // We must cast correctly to the pointer type. Ensure that we
9838 // sign extend the integer value if it is smaller as this is
9839 // used for address computation.
9840 Instruction::CastOps opcode =
9841 (VTySize < PtrSize ? Instruction::SExt :
9842 (VTySize == PtrSize ? Instruction::BitCast : Instruction::Trunc));
9843 return IC->InsertCastBefore(opcode, V, DTy, *InsertPoint);
9847 Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) {
9848 Value *PtrOp = GEP.getOperand(0);
9849 // Is it 'getelementptr %P, i32 0' or 'getelementptr %P'
9850 // If so, eliminate the noop.
9851 if (GEP.getNumOperands() == 1)
9852 return ReplaceInstUsesWith(GEP, PtrOp);
9854 if (isa<UndefValue>(GEP.getOperand(0)))
9855 return ReplaceInstUsesWith(GEP, UndefValue::get(GEP.getType()));
9857 bool HasZeroPointerIndex = false;
9858 if (Constant *C = dyn_cast<Constant>(GEP.getOperand(1)))
9859 HasZeroPointerIndex = C->isNullValue();
9861 if (GEP.getNumOperands() == 2 && HasZeroPointerIndex)
9862 return ReplaceInstUsesWith(GEP, PtrOp);
9864 // Eliminate unneeded casts for indices.
9865 bool MadeChange = false;
9867 gep_type_iterator GTI = gep_type_begin(GEP);
9868 for (User::op_iterator i = GEP.op_begin() + 1, e = GEP.op_end();
9869 i != e; ++i, ++GTI) {
9870 if (isa<SequentialType>(*GTI)) {
9871 if (CastInst *CI = dyn_cast<CastInst>(*i)) {
9872 if (CI->getOpcode() == Instruction::ZExt ||
9873 CI->getOpcode() == Instruction::SExt) {
9874 const Type *SrcTy = CI->getOperand(0)->getType();
9875 // We can eliminate a cast from i32 to i64 iff the target
9876 // is a 32-bit pointer target.
9877 if (SrcTy->getPrimitiveSizeInBits() >= TD->getPointerSizeInBits()) {
9879 *i = CI->getOperand(0);
9883 // If we are using a wider index than needed for this platform, shrink it
9884 // to what we need. If narrower, sign-extend it to what we need.
9885 // If the incoming value needs a cast instruction,
9886 // insert it. This explicit cast can make subsequent optimizations more
9889 if (TD->getTypeSizeInBits(Op->getType()) > TD->getPointerSizeInBits()) {
9890 if (Constant *C = dyn_cast<Constant>(Op)) {
9891 *i = ConstantExpr::getTrunc(C, TD->getIntPtrType());
9894 Op = InsertCastBefore(Instruction::Trunc, Op, TD->getIntPtrType(),
9899 } else if (TD->getTypeSizeInBits(Op->getType()) < TD->getPointerSizeInBits()) {
9900 if (Constant *C = dyn_cast<Constant>(Op)) {
9901 *i = ConstantExpr::getSExt(C, TD->getIntPtrType());
9904 Op = InsertCastBefore(Instruction::SExt, Op, TD->getIntPtrType(),
9912 if (MadeChange) return &GEP;
9914 // If this GEP instruction doesn't move the pointer, and if the input operand
9915 // is a bitcast of another pointer, just replace the GEP with a bitcast of the
9916 // real input to the dest type.
9917 if (GEP.hasAllZeroIndices()) {
9918 if (BitCastInst *BCI = dyn_cast<BitCastInst>(GEP.getOperand(0))) {
9919 // If the bitcast is of an allocation, and the allocation will be
9920 // converted to match the type of the cast, don't touch this.
9921 if (isa<AllocationInst>(BCI->getOperand(0))) {
9922 // See if the bitcast simplifies, if so, don't nuke this GEP yet.
9923 if (Instruction *I = visitBitCast(*BCI)) {
9926 BCI->getParent()->getInstList().insert(BCI, I);
9927 ReplaceInstUsesWith(*BCI, I);
9932 return new BitCastInst(BCI->getOperand(0), GEP.getType());
9936 // Combine Indices - If the source pointer to this getelementptr instruction
9937 // is a getelementptr instruction, combine the indices of the two
9938 // getelementptr instructions into a single instruction.
9940 SmallVector<Value*, 8> SrcGEPOperands;
9941 if (User *Src = dyn_castGetElementPtr(PtrOp))
9942 SrcGEPOperands.append(Src->op_begin(), Src->op_end());
9944 if (!SrcGEPOperands.empty()) {
9945 // Note that if our source is a gep chain itself that we wait for that
9946 // chain to be resolved before we perform this transformation. This
9947 // avoids us creating a TON of code in some cases.
9949 if (isa<GetElementPtrInst>(SrcGEPOperands[0]) &&
9950 cast<Instruction>(SrcGEPOperands[0])->getNumOperands() == 2)
9951 return 0; // Wait until our source is folded to completion.
9953 SmallVector<Value*, 8> Indices;
9955 // Find out whether the last index in the source GEP is a sequential idx.
9956 bool EndsWithSequential = false;
9957 for (gep_type_iterator I = gep_type_begin(*cast<User>(PtrOp)),
9958 E = gep_type_end(*cast<User>(PtrOp)); I != E; ++I)
9959 EndsWithSequential = !isa<StructType>(*I);
9961 // Can we combine the two pointer arithmetics offsets?
9962 if (EndsWithSequential) {
9963 // Replace: gep (gep %P, long B), long A, ...
9964 // With: T = long A+B; gep %P, T, ...
9966 Value *Sum, *SO1 = SrcGEPOperands.back(), *GO1 = GEP.getOperand(1);
9967 if (SO1 == Constant::getNullValue(SO1->getType())) {
9969 } else if (GO1 == Constant::getNullValue(GO1->getType())) {
9972 // If they aren't the same type, convert both to an integer of the
9973 // target's pointer size.
9974 if (SO1->getType() != GO1->getType()) {
9975 if (Constant *SO1C = dyn_cast<Constant>(SO1)) {
9976 SO1 = ConstantExpr::getIntegerCast(SO1C, GO1->getType(), true);
9977 } else if (Constant *GO1C = dyn_cast<Constant>(GO1)) {
9978 GO1 = ConstantExpr::getIntegerCast(GO1C, SO1->getType(), true);
9980 unsigned PS = TD->getPointerSizeInBits();
9981 if (TD->getTypeSizeInBits(SO1->getType()) == PS) {
9982 // Convert GO1 to SO1's type.
9983 GO1 = InsertCastToIntPtrTy(GO1, SO1->getType(), &GEP, this);
9985 } else if (TD->getTypeSizeInBits(GO1->getType()) == PS) {
9986 // Convert SO1 to GO1's type.
9987 SO1 = InsertCastToIntPtrTy(SO1, GO1->getType(), &GEP, this);
9989 const Type *PT = TD->getIntPtrType();
9990 SO1 = InsertCastToIntPtrTy(SO1, PT, &GEP, this);
9991 GO1 = InsertCastToIntPtrTy(GO1, PT, &GEP, this);
9995 if (isa<Constant>(SO1) && isa<Constant>(GO1))
9996 Sum = ConstantExpr::getAdd(cast<Constant>(SO1), cast<Constant>(GO1));
9998 Sum = BinaryOperator::CreateAdd(SO1, GO1, PtrOp->getName()+".sum");
9999 InsertNewInstBefore(cast<Instruction>(Sum), GEP);
10003 // Recycle the GEP we already have if possible.
10004 if (SrcGEPOperands.size() == 2) {
10005 GEP.setOperand(0, SrcGEPOperands[0]);
10006 GEP.setOperand(1, Sum);
10009 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
10010 SrcGEPOperands.end()-1);
10011 Indices.push_back(Sum);
10012 Indices.insert(Indices.end(), GEP.op_begin()+2, GEP.op_end());
10014 } else if (isa<Constant>(*GEP.idx_begin()) &&
10015 cast<Constant>(*GEP.idx_begin())->isNullValue() &&
10016 SrcGEPOperands.size() != 1) {
10017 // Otherwise we can do the fold if the first index of the GEP is a zero
10018 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
10019 SrcGEPOperands.end());
10020 Indices.insert(Indices.end(), GEP.idx_begin()+1, GEP.idx_end());
10023 if (!Indices.empty())
10024 return GetElementPtrInst::Create(SrcGEPOperands[0], Indices.begin(),
10025 Indices.end(), GEP.getName());
10027 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(PtrOp)) {
10028 // GEP of global variable. If all of the indices for this GEP are
10029 // constants, we can promote this to a constexpr instead of an instruction.
10031 // Scan for nonconstants...
10032 SmallVector<Constant*, 8> Indices;
10033 User::op_iterator I = GEP.idx_begin(), E = GEP.idx_end();
10034 for (; I != E && isa<Constant>(*I); ++I)
10035 Indices.push_back(cast<Constant>(*I));
10037 if (I == E) { // If they are all constants...
10038 Constant *CE = ConstantExpr::getGetElementPtr(GV,
10039 &Indices[0],Indices.size());
10041 // Replace all uses of the GEP with the new constexpr...
10042 return ReplaceInstUsesWith(GEP, CE);
10044 } else if (Value *X = getBitCastOperand(PtrOp)) { // Is the operand a cast?
10045 if (!isa<PointerType>(X->getType())) {
10046 // Not interesting. Source pointer must be a cast from pointer.
10047 } else if (HasZeroPointerIndex) {
10048 // transform: GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ...
10049 // into : GEP [10 x i8]* X, i32 0, ...
10051 // This occurs when the program declares an array extern like "int X[];"
10053 const PointerType *CPTy = cast<PointerType>(PtrOp->getType());
10054 const PointerType *XTy = cast<PointerType>(X->getType());
10055 if (const ArrayType *XATy =
10056 dyn_cast<ArrayType>(XTy->getElementType()))
10057 if (const ArrayType *CATy =
10058 dyn_cast<ArrayType>(CPTy->getElementType()))
10059 if (CATy->getElementType() == XATy->getElementType()) {
10060 // At this point, we know that the cast source type is a pointer
10061 // to an array of the same type as the destination pointer
10062 // array. Because the array type is never stepped over (there
10063 // is a leading zero) we can fold the cast into this GEP.
10064 GEP.setOperand(0, X);
10067 } else if (GEP.getNumOperands() == 2) {
10068 // Transform things like:
10069 // %t = getelementptr i32* bitcast ([2 x i32]* %str to i32*), i32 %V
10070 // into: %t1 = getelementptr [2 x i32]* %str, i32 0, i32 %V; bitcast
10071 const Type *SrcElTy = cast<PointerType>(X->getType())->getElementType();
10072 const Type *ResElTy=cast<PointerType>(PtrOp->getType())->getElementType();
10073 if (isa<ArrayType>(SrcElTy) &&
10074 TD->getABITypeSize(cast<ArrayType>(SrcElTy)->getElementType()) ==
10075 TD->getABITypeSize(ResElTy)) {
10077 Idx[0] = Constant::getNullValue(Type::Int32Ty);
10078 Idx[1] = GEP.getOperand(1);
10079 Value *V = InsertNewInstBefore(
10080 GetElementPtrInst::Create(X, Idx, Idx + 2, GEP.getName()), GEP);
10081 // V and GEP are both pointer types --> BitCast
10082 return new BitCastInst(V, GEP.getType());
10085 // Transform things like:
10086 // getelementptr i8* bitcast ([100 x double]* X to i8*), i32 %tmp
10087 // (where tmp = 8*tmp2) into:
10088 // getelementptr [100 x double]* %arr, i32 0, i32 %tmp2; bitcast
10090 if (isa<ArrayType>(SrcElTy) && ResElTy == Type::Int8Ty) {
10091 uint64_t ArrayEltSize =
10092 TD->getABITypeSize(cast<ArrayType>(SrcElTy)->getElementType());
10094 // Check to see if "tmp" is a scale by a multiple of ArrayEltSize. We
10095 // allow either a mul, shift, or constant here.
10097 ConstantInt *Scale = 0;
10098 if (ArrayEltSize == 1) {
10099 NewIdx = GEP.getOperand(1);
10100 Scale = ConstantInt::get(NewIdx->getType(), 1);
10101 } else if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP.getOperand(1))) {
10102 NewIdx = ConstantInt::get(CI->getType(), 1);
10104 } else if (Instruction *Inst =dyn_cast<Instruction>(GEP.getOperand(1))){
10105 if (Inst->getOpcode() == Instruction::Shl &&
10106 isa<ConstantInt>(Inst->getOperand(1))) {
10107 ConstantInt *ShAmt = cast<ConstantInt>(Inst->getOperand(1));
10108 uint32_t ShAmtVal = ShAmt->getLimitedValue(64);
10109 Scale = ConstantInt::get(Inst->getType(), 1ULL << ShAmtVal);
10110 NewIdx = Inst->getOperand(0);
10111 } else if (Inst->getOpcode() == Instruction::Mul &&
10112 isa<ConstantInt>(Inst->getOperand(1))) {
10113 Scale = cast<ConstantInt>(Inst->getOperand(1));
10114 NewIdx = Inst->getOperand(0);
10118 // If the index will be to exactly the right offset with the scale taken
10119 // out, perform the transformation. Note, we don't know whether Scale is
10120 // signed or not. We'll use unsigned version of division/modulo
10121 // operation after making sure Scale doesn't have the sign bit set.
10122 if (Scale && Scale->getSExtValue() >= 0LL &&
10123 Scale->getZExtValue() % ArrayEltSize == 0) {
10124 Scale = ConstantInt::get(Scale->getType(),
10125 Scale->getZExtValue() / ArrayEltSize);
10126 if (Scale->getZExtValue() != 1) {
10127 Constant *C = ConstantExpr::getIntegerCast(Scale, NewIdx->getType(),
10129 Instruction *Sc = BinaryOperator::CreateMul(NewIdx, C, "idxscale");
10130 NewIdx = InsertNewInstBefore(Sc, GEP);
10133 // Insert the new GEP instruction.
10135 Idx[0] = Constant::getNullValue(Type::Int32Ty);
10137 Instruction *NewGEP =
10138 GetElementPtrInst::Create(X, Idx, Idx + 2, GEP.getName());
10139 NewGEP = InsertNewInstBefore(NewGEP, GEP);
10140 // The NewGEP must be pointer typed, so must the old one -> BitCast
10141 return new BitCastInst(NewGEP, GEP.getType());
10150 Instruction *InstCombiner::visitAllocationInst(AllocationInst &AI) {
10151 // Convert: malloc Ty, C - where C is a constant != 1 into: malloc [C x Ty], 1
10152 if (AI.isArrayAllocation()) { // Check C != 1
10153 if (const ConstantInt *C = dyn_cast<ConstantInt>(AI.getArraySize())) {
10154 const Type *NewTy =
10155 ArrayType::get(AI.getAllocatedType(), C->getZExtValue());
10156 AllocationInst *New = 0;
10158 // Create and insert the replacement instruction...
10159 if (isa<MallocInst>(AI))
10160 New = new MallocInst(NewTy, 0, AI.getAlignment(), AI.getName());
10162 assert(isa<AllocaInst>(AI) && "Unknown type of allocation inst!");
10163 New = new AllocaInst(NewTy, 0, AI.getAlignment(), AI.getName());
10166 InsertNewInstBefore(New, AI);
10168 // Scan to the end of the allocation instructions, to skip over a block of
10169 // allocas if possible...
10171 BasicBlock::iterator It = New;
10172 while (isa<AllocationInst>(*It)) ++It;
10174 // Now that I is pointing to the first non-allocation-inst in the block,
10175 // insert our getelementptr instruction...
10177 Value *NullIdx = Constant::getNullValue(Type::Int32Ty);
10181 Value *V = GetElementPtrInst::Create(New, Idx, Idx + 2,
10182 New->getName()+".sub", It);
10184 // Now make everything use the getelementptr instead of the original
10186 return ReplaceInstUsesWith(AI, V);
10187 } else if (isa<UndefValue>(AI.getArraySize())) {
10188 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
10192 // If alloca'ing a zero byte object, replace the alloca with a null pointer.
10193 // Note that we only do this for alloca's, because malloc should allocate and
10194 // return a unique pointer, even for a zero byte allocation.
10195 if (isa<AllocaInst>(AI) && AI.getAllocatedType()->isSized() &&
10196 TD->getABITypeSize(AI.getAllocatedType()) == 0)
10197 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
10202 Instruction *InstCombiner::visitFreeInst(FreeInst &FI) {
10203 Value *Op = FI.getOperand(0);
10205 // free undef -> unreachable.
10206 if (isa<UndefValue>(Op)) {
10207 // Insert a new store to null because we cannot modify the CFG here.
10208 new StoreInst(ConstantInt::getTrue(),
10209 UndefValue::get(PointerType::getUnqual(Type::Int1Ty)), &FI);
10210 return EraseInstFromFunction(FI);
10213 // If we have 'free null' delete the instruction. This can happen in stl code
10214 // when lots of inlining happens.
10215 if (isa<ConstantPointerNull>(Op))
10216 return EraseInstFromFunction(FI);
10218 // Change free <ty>* (cast <ty2>* X to <ty>*) into free <ty2>* X
10219 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op)) {
10220 FI.setOperand(0, CI->getOperand(0));
10224 // Change free (gep X, 0,0,0,0) into free(X)
10225 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op)) {
10226 if (GEPI->hasAllZeroIndices()) {
10227 AddToWorkList(GEPI);
10228 FI.setOperand(0, GEPI->getOperand(0));
10233 // Change free(malloc) into nothing, if the malloc has a single use.
10234 if (MallocInst *MI = dyn_cast<MallocInst>(Op))
10235 if (MI->hasOneUse()) {
10236 EraseInstFromFunction(FI);
10237 return EraseInstFromFunction(*MI);
10244 /// InstCombineLoadCast - Fold 'load (cast P)' -> cast (load P)' when possible.
10245 static Instruction *InstCombineLoadCast(InstCombiner &IC, LoadInst &LI,
10246 const TargetData *TD) {
10247 User *CI = cast<User>(LI.getOperand(0));
10248 Value *CastOp = CI->getOperand(0);
10250 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(CI)) {
10251 // Instead of loading constant c string, use corresponding integer value
10252 // directly if string length is small enough.
10254 if (GetConstantStringInfo(CE->getOperand(0), Str) && !Str.empty()) {
10255 unsigned len = Str.length();
10256 const Type *Ty = cast<PointerType>(CE->getType())->getElementType();
10257 unsigned numBits = Ty->getPrimitiveSizeInBits();
10258 // Replace LI with immediate integer store.
10259 if ((numBits >> 3) == len + 1) {
10260 APInt StrVal(numBits, 0);
10261 APInt SingleChar(numBits, 0);
10262 if (TD->isLittleEndian()) {
10263 for (signed i = len-1; i >= 0; i--) {
10264 SingleChar = (uint64_t) Str[i];
10265 StrVal = (StrVal << 8) | SingleChar;
10268 for (unsigned i = 0; i < len; i++) {
10269 SingleChar = (uint64_t) Str[i];
10270 StrVal = (StrVal << 8) | SingleChar;
10272 // Append NULL at the end.
10274 StrVal = (StrVal << 8) | SingleChar;
10276 Value *NL = ConstantInt::get(StrVal);
10277 return IC.ReplaceInstUsesWith(LI, NL);
10282 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
10283 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
10284 const Type *SrcPTy = SrcTy->getElementType();
10286 if (DestPTy->isInteger() || isa<PointerType>(DestPTy) ||
10287 isa<VectorType>(DestPTy)) {
10288 // If the source is an array, the code below will not succeed. Check to
10289 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
10291 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
10292 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
10293 if (ASrcTy->getNumElements() != 0) {
10295 Idxs[0] = Idxs[1] = Constant::getNullValue(Type::Int32Ty);
10296 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs, 2);
10297 SrcTy = cast<PointerType>(CastOp->getType());
10298 SrcPTy = SrcTy->getElementType();
10301 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy) ||
10302 isa<VectorType>(SrcPTy)) &&
10303 // Do not allow turning this into a load of an integer, which is then
10304 // casted to a pointer, this pessimizes pointer analysis a lot.
10305 (isa<PointerType>(SrcPTy) == isa<PointerType>(LI.getType())) &&
10306 IC.getTargetData().getTypeSizeInBits(SrcPTy) ==
10307 IC.getTargetData().getTypeSizeInBits(DestPTy)) {
10309 // Okay, we are casting from one integer or pointer type to another of
10310 // the same size. Instead of casting the pointer before the load, cast
10311 // the result of the loaded value.
10312 Value *NewLoad = IC.InsertNewInstBefore(new LoadInst(CastOp,
10314 LI.isVolatile()),LI);
10315 // Now cast the result of the load.
10316 return new BitCastInst(NewLoad, LI.getType());
10323 /// isSafeToLoadUnconditionally - Return true if we know that executing a load
10324 /// from this value cannot trap. If it is not obviously safe to load from the
10325 /// specified pointer, we do a quick local scan of the basic block containing
10326 /// ScanFrom, to determine if the address is already accessed.
10327 static bool isSafeToLoadUnconditionally(Value *V, Instruction *ScanFrom) {
10328 // If it is an alloca it is always safe to load from.
10329 if (isa<AllocaInst>(V)) return true;
10331 // If it is a global variable it is mostly safe to load from.
10332 if (const GlobalValue *GV = dyn_cast<GlobalVariable>(V))
10333 // Don't try to evaluate aliases. External weak GV can be null.
10334 return !isa<GlobalAlias>(GV) && !GV->hasExternalWeakLinkage();
10336 // Otherwise, be a little bit agressive by scanning the local block where we
10337 // want to check to see if the pointer is already being loaded or stored
10338 // from/to. If so, the previous load or store would have already trapped,
10339 // so there is no harm doing an extra load (also, CSE will later eliminate
10340 // the load entirely).
10341 BasicBlock::iterator BBI = ScanFrom, E = ScanFrom->getParent()->begin();
10346 // If we see a free or a call (which might do a free) the pointer could be
10348 if (isa<FreeInst>(BBI) || isa<CallInst>(BBI))
10351 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
10352 if (LI->getOperand(0) == V) return true;
10353 } else if (StoreInst *SI = dyn_cast<StoreInst>(BBI)) {
10354 if (SI->getOperand(1) == V) return true;
10361 /// GetUnderlyingObject - Trace through a series of getelementptrs and bitcasts
10362 /// until we find the underlying object a pointer is referring to or something
10363 /// we don't understand. Note that the returned pointer may be offset from the
10364 /// input, because we ignore GEP indices.
10365 static Value *GetUnderlyingObject(Value *Ptr) {
10367 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr)) {
10368 if (CE->getOpcode() == Instruction::BitCast ||
10369 CE->getOpcode() == Instruction::GetElementPtr)
10370 Ptr = CE->getOperand(0);
10373 } else if (BitCastInst *BCI = dyn_cast<BitCastInst>(Ptr)) {
10374 Ptr = BCI->getOperand(0);
10375 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr)) {
10376 Ptr = GEP->getOperand(0);
10383 Instruction *InstCombiner::visitLoadInst(LoadInst &LI) {
10384 Value *Op = LI.getOperand(0);
10386 // Attempt to improve the alignment.
10387 unsigned KnownAlign = GetOrEnforceKnownAlignment(Op);
10389 (LI.getAlignment() == 0 ? TD->getABITypeAlignment(LI.getType()) :
10390 LI.getAlignment()))
10391 LI.setAlignment(KnownAlign);
10393 // load (cast X) --> cast (load X) iff safe
10394 if (isa<CastInst>(Op))
10395 if (Instruction *Res = InstCombineLoadCast(*this, LI, TD))
10398 // None of the following transforms are legal for volatile loads.
10399 if (LI.isVolatile()) return 0;
10401 if (&LI.getParent()->front() != &LI) {
10402 BasicBlock::iterator BBI = &LI; --BBI;
10403 // If the instruction immediately before this is a store to the same
10404 // address, do a simple form of store->load forwarding.
10405 if (StoreInst *SI = dyn_cast<StoreInst>(BBI))
10406 if (SI->getOperand(1) == LI.getOperand(0))
10407 return ReplaceInstUsesWith(LI, SI->getOperand(0));
10408 if (LoadInst *LIB = dyn_cast<LoadInst>(BBI))
10409 if (LIB->getOperand(0) == LI.getOperand(0))
10410 return ReplaceInstUsesWith(LI, LIB);
10413 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op)) {
10414 const Value *GEPI0 = GEPI->getOperand(0);
10415 // TODO: Consider a target hook for valid address spaces for this xform.
10416 if (isa<ConstantPointerNull>(GEPI0) &&
10417 cast<PointerType>(GEPI0->getType())->getAddressSpace() == 0) {
10418 // Insert a new store to null instruction before the load to indicate
10419 // that this code is not reachable. We do this instead of inserting
10420 // an unreachable instruction directly because we cannot modify the
10422 new StoreInst(UndefValue::get(LI.getType()),
10423 Constant::getNullValue(Op->getType()), &LI);
10424 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
10428 if (Constant *C = dyn_cast<Constant>(Op)) {
10429 // load null/undef -> undef
10430 // TODO: Consider a target hook for valid address spaces for this xform.
10431 if (isa<UndefValue>(C) || (C->isNullValue() &&
10432 cast<PointerType>(Op->getType())->getAddressSpace() == 0)) {
10433 // Insert a new store to null instruction before the load to indicate that
10434 // this code is not reachable. We do this instead of inserting an
10435 // unreachable instruction directly because we cannot modify the CFG.
10436 new StoreInst(UndefValue::get(LI.getType()),
10437 Constant::getNullValue(Op->getType()), &LI);
10438 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
10441 // Instcombine load (constant global) into the value loaded.
10442 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Op))
10443 if (GV->isConstant() && !GV->isDeclaration())
10444 return ReplaceInstUsesWith(LI, GV->getInitializer());
10446 // Instcombine load (constantexpr_GEP global, 0, ...) into the value loaded.
10447 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Op)) {
10448 if (CE->getOpcode() == Instruction::GetElementPtr) {
10449 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(CE->getOperand(0)))
10450 if (GV->isConstant() && !GV->isDeclaration())
10452 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE))
10453 return ReplaceInstUsesWith(LI, V);
10454 if (CE->getOperand(0)->isNullValue()) {
10455 // Insert a new store to null instruction before the load to indicate
10456 // that this code is not reachable. We do this instead of inserting
10457 // an unreachable instruction directly because we cannot modify the
10459 new StoreInst(UndefValue::get(LI.getType()),
10460 Constant::getNullValue(Op->getType()), &LI);
10461 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
10464 } else if (CE->isCast()) {
10465 if (Instruction *Res = InstCombineLoadCast(*this, LI, TD))
10471 // If this load comes from anywhere in a constant global, and if the global
10472 // is all undef or zero, we know what it loads.
10473 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GetUnderlyingObject(Op))) {
10474 if (GV->isConstant() && GV->hasInitializer()) {
10475 if (GV->getInitializer()->isNullValue())
10476 return ReplaceInstUsesWith(LI, Constant::getNullValue(LI.getType()));
10477 else if (isa<UndefValue>(GV->getInitializer()))
10478 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
10482 if (Op->hasOneUse()) {
10483 // Change select and PHI nodes to select values instead of addresses: this
10484 // helps alias analysis out a lot, allows many others simplifications, and
10485 // exposes redundancy in the code.
10487 // Note that we cannot do the transformation unless we know that the
10488 // introduced loads cannot trap! Something like this is valid as long as
10489 // the condition is always false: load (select bool %C, int* null, int* %G),
10490 // but it would not be valid if we transformed it to load from null
10491 // unconditionally.
10493 if (SelectInst *SI = dyn_cast<SelectInst>(Op)) {
10494 // load (select (Cond, &V1, &V2)) --> select(Cond, load &V1, load &V2).
10495 if (isSafeToLoadUnconditionally(SI->getOperand(1), SI) &&
10496 isSafeToLoadUnconditionally(SI->getOperand(2), SI)) {
10497 Value *V1 = InsertNewInstBefore(new LoadInst(SI->getOperand(1),
10498 SI->getOperand(1)->getName()+".val"), LI);
10499 Value *V2 = InsertNewInstBefore(new LoadInst(SI->getOperand(2),
10500 SI->getOperand(2)->getName()+".val"), LI);
10501 return SelectInst::Create(SI->getCondition(), V1, V2);
10504 // load (select (cond, null, P)) -> load P
10505 if (Constant *C = dyn_cast<Constant>(SI->getOperand(1)))
10506 if (C->isNullValue()) {
10507 LI.setOperand(0, SI->getOperand(2));
10511 // load (select (cond, P, null)) -> load P
10512 if (Constant *C = dyn_cast<Constant>(SI->getOperand(2)))
10513 if (C->isNullValue()) {
10514 LI.setOperand(0, SI->getOperand(1));
10522 /// InstCombineStoreToCast - Fold store V, (cast P) -> store (cast V), P
10524 static Instruction *InstCombineStoreToCast(InstCombiner &IC, StoreInst &SI) {
10525 User *CI = cast<User>(SI.getOperand(1));
10526 Value *CastOp = CI->getOperand(0);
10528 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
10529 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
10530 const Type *SrcPTy = SrcTy->getElementType();
10532 if (DestPTy->isInteger() || isa<PointerType>(DestPTy)) {
10533 // If the source is an array, the code below will not succeed. Check to
10534 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
10536 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
10537 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
10538 if (ASrcTy->getNumElements() != 0) {
10540 Idxs[0] = Idxs[1] = Constant::getNullValue(Type::Int32Ty);
10541 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs, 2);
10542 SrcTy = cast<PointerType>(CastOp->getType());
10543 SrcPTy = SrcTy->getElementType();
10546 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy)) &&
10547 IC.getTargetData().getTypeSizeInBits(SrcPTy) ==
10548 IC.getTargetData().getTypeSizeInBits(DestPTy)) {
10550 // Okay, we are casting from one integer or pointer type to another of
10551 // the same size. Instead of casting the pointer before
10552 // the store, cast the value to be stored.
10554 Value *SIOp0 = SI.getOperand(0);
10555 Instruction::CastOps opcode = Instruction::BitCast;
10556 const Type* CastSrcTy = SIOp0->getType();
10557 const Type* CastDstTy = SrcPTy;
10558 if (isa<PointerType>(CastDstTy)) {
10559 if (CastSrcTy->isInteger())
10560 opcode = Instruction::IntToPtr;
10561 } else if (isa<IntegerType>(CastDstTy)) {
10562 if (isa<PointerType>(SIOp0->getType()))
10563 opcode = Instruction::PtrToInt;
10565 if (Constant *C = dyn_cast<Constant>(SIOp0))
10566 NewCast = ConstantExpr::getCast(opcode, C, CastDstTy);
10568 NewCast = IC.InsertNewInstBefore(
10569 CastInst::Create(opcode, SIOp0, CastDstTy, SIOp0->getName()+".c"),
10571 return new StoreInst(NewCast, CastOp);
10578 Instruction *InstCombiner::visitStoreInst(StoreInst &SI) {
10579 Value *Val = SI.getOperand(0);
10580 Value *Ptr = SI.getOperand(1);
10582 if (isa<UndefValue>(Ptr)) { // store X, undef -> noop (even if volatile)
10583 EraseInstFromFunction(SI);
10588 // If the RHS is an alloca with a single use, zapify the store, making the
10590 if (Ptr->hasOneUse() && !SI.isVolatile()) {
10591 if (isa<AllocaInst>(Ptr)) {
10592 EraseInstFromFunction(SI);
10597 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr))
10598 if (isa<AllocaInst>(GEP->getOperand(0)) &&
10599 GEP->getOperand(0)->hasOneUse()) {
10600 EraseInstFromFunction(SI);
10606 // Attempt to improve the alignment.
10607 unsigned KnownAlign = GetOrEnforceKnownAlignment(Ptr);
10609 (SI.getAlignment() == 0 ? TD->getABITypeAlignment(Val->getType()) :
10610 SI.getAlignment()))
10611 SI.setAlignment(KnownAlign);
10613 // Do really simple DSE, to catch cases where there are several consequtive
10614 // stores to the same location, separated by a few arithmetic operations. This
10615 // situation often occurs with bitfield accesses.
10616 BasicBlock::iterator BBI = &SI;
10617 for (unsigned ScanInsts = 6; BBI != SI.getParent()->begin() && ScanInsts;
10621 if (StoreInst *PrevSI = dyn_cast<StoreInst>(BBI)) {
10622 // Prev store isn't volatile, and stores to the same location?
10623 if (!PrevSI->isVolatile() && PrevSI->getOperand(1) == SI.getOperand(1)) {
10626 EraseInstFromFunction(*PrevSI);
10632 // If this is a load, we have to stop. However, if the loaded value is from
10633 // the pointer we're loading and is producing the pointer we're storing,
10634 // then *this* store is dead (X = load P; store X -> P).
10635 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
10636 if (LI == Val && LI->getOperand(0) == Ptr && !SI.isVolatile()) {
10637 EraseInstFromFunction(SI);
10641 // Otherwise, this is a load from some other location. Stores before it
10642 // may not be dead.
10646 // Don't skip over loads or things that can modify memory.
10647 if (BBI->mayWriteToMemory() || BBI->mayReadFromMemory())
10652 if (SI.isVolatile()) return 0; // Don't hack volatile stores.
10654 // store X, null -> turns into 'unreachable' in SimplifyCFG
10655 if (isa<ConstantPointerNull>(Ptr)) {
10656 if (!isa<UndefValue>(Val)) {
10657 SI.setOperand(0, UndefValue::get(Val->getType()));
10658 if (Instruction *U = dyn_cast<Instruction>(Val))
10659 AddToWorkList(U); // Dropped a use.
10662 return 0; // Do not modify these!
10665 // store undef, Ptr -> noop
10666 if (isa<UndefValue>(Val)) {
10667 EraseInstFromFunction(SI);
10672 // If the pointer destination is a cast, see if we can fold the cast into the
10674 if (isa<CastInst>(Ptr))
10675 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
10677 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
10679 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
10683 // If this store is the last instruction in the basic block, and if the block
10684 // ends with an unconditional branch, try to move it to the successor block.
10686 if (BranchInst *BI = dyn_cast<BranchInst>(BBI))
10687 if (BI->isUnconditional())
10688 if (SimplifyStoreAtEndOfBlock(SI))
10689 return 0; // xform done!
10694 /// SimplifyStoreAtEndOfBlock - Turn things like:
10695 /// if () { *P = v1; } else { *P = v2 }
10696 /// into a phi node with a store in the successor.
10698 /// Simplify things like:
10699 /// *P = v1; if () { *P = v2; }
10700 /// into a phi node with a store in the successor.
10702 bool InstCombiner::SimplifyStoreAtEndOfBlock(StoreInst &SI) {
10703 BasicBlock *StoreBB = SI.getParent();
10705 // Check to see if the successor block has exactly two incoming edges. If
10706 // so, see if the other predecessor contains a store to the same location.
10707 // if so, insert a PHI node (if needed) and move the stores down.
10708 BasicBlock *DestBB = StoreBB->getTerminator()->getSuccessor(0);
10710 // Determine whether Dest has exactly two predecessors and, if so, compute
10711 // the other predecessor.
10712 pred_iterator PI = pred_begin(DestBB);
10713 BasicBlock *OtherBB = 0;
10714 if (*PI != StoreBB)
10717 if (PI == pred_end(DestBB))
10720 if (*PI != StoreBB) {
10725 if (++PI != pred_end(DestBB))
10728 // Bail out if all the relevant blocks aren't distinct (this can happen,
10729 // for example, if SI is in an infinite loop)
10730 if (StoreBB == DestBB || OtherBB == DestBB)
10733 // Verify that the other block ends in a branch and is not otherwise empty.
10734 BasicBlock::iterator BBI = OtherBB->getTerminator();
10735 BranchInst *OtherBr = dyn_cast<BranchInst>(BBI);
10736 if (!OtherBr || BBI == OtherBB->begin())
10739 // If the other block ends in an unconditional branch, check for the 'if then
10740 // else' case. there is an instruction before the branch.
10741 StoreInst *OtherStore = 0;
10742 if (OtherBr->isUnconditional()) {
10743 // If this isn't a store, or isn't a store to the same location, bail out.
10745 OtherStore = dyn_cast<StoreInst>(BBI);
10746 if (!OtherStore || OtherStore->getOperand(1) != SI.getOperand(1))
10749 // Otherwise, the other block ended with a conditional branch. If one of the
10750 // destinations is StoreBB, then we have the if/then case.
10751 if (OtherBr->getSuccessor(0) != StoreBB &&
10752 OtherBr->getSuccessor(1) != StoreBB)
10755 // Okay, we know that OtherBr now goes to Dest and StoreBB, so this is an
10756 // if/then triangle. See if there is a store to the same ptr as SI that
10757 // lives in OtherBB.
10759 // Check to see if we find the matching store.
10760 if ((OtherStore = dyn_cast<StoreInst>(BBI))) {
10761 if (OtherStore->getOperand(1) != SI.getOperand(1))
10765 // If we find something that may be using or overwriting the stored
10766 // value, or if we run out of instructions, we can't do the xform.
10767 if (BBI->mayReadFromMemory() || BBI->mayWriteToMemory() ||
10768 BBI == OtherBB->begin())
10772 // In order to eliminate the store in OtherBr, we have to
10773 // make sure nothing reads or overwrites the stored value in
10775 for (BasicBlock::iterator I = StoreBB->begin(); &*I != &SI; ++I) {
10776 // FIXME: This should really be AA driven.
10777 if (I->mayReadFromMemory() || I->mayWriteToMemory())
10782 // Insert a PHI node now if we need it.
10783 Value *MergedVal = OtherStore->getOperand(0);
10784 if (MergedVal != SI.getOperand(0)) {
10785 PHINode *PN = PHINode::Create(MergedVal->getType(), "storemerge");
10786 PN->reserveOperandSpace(2);
10787 PN->addIncoming(SI.getOperand(0), SI.getParent());
10788 PN->addIncoming(OtherStore->getOperand(0), OtherBB);
10789 MergedVal = InsertNewInstBefore(PN, DestBB->front());
10792 // Advance to a place where it is safe to insert the new store and
10794 BBI = DestBB->getFirstNonPHI();
10795 InsertNewInstBefore(new StoreInst(MergedVal, SI.getOperand(1),
10796 OtherStore->isVolatile()), *BBI);
10798 // Nuke the old stores.
10799 EraseInstFromFunction(SI);
10800 EraseInstFromFunction(*OtherStore);
10806 Instruction *InstCombiner::visitBranchInst(BranchInst &BI) {
10807 // Change br (not X), label True, label False to: br X, label False, True
10809 BasicBlock *TrueDest;
10810 BasicBlock *FalseDest;
10811 if (match(&BI, m_Br(m_Not(m_Value(X)), TrueDest, FalseDest)) &&
10812 !isa<Constant>(X)) {
10813 // Swap Destinations and condition...
10814 BI.setCondition(X);
10815 BI.setSuccessor(0, FalseDest);
10816 BI.setSuccessor(1, TrueDest);
10820 // Cannonicalize fcmp_one -> fcmp_oeq
10821 FCmpInst::Predicate FPred; Value *Y;
10822 if (match(&BI, m_Br(m_FCmp(FPred, m_Value(X), m_Value(Y)),
10823 TrueDest, FalseDest)))
10824 if ((FPred == FCmpInst::FCMP_ONE || FPred == FCmpInst::FCMP_OLE ||
10825 FPred == FCmpInst::FCMP_OGE) && BI.getCondition()->hasOneUse()) {
10826 FCmpInst *I = cast<FCmpInst>(BI.getCondition());
10827 FCmpInst::Predicate NewPred = FCmpInst::getInversePredicate(FPred);
10828 Instruction *NewSCC = new FCmpInst(NewPred, X, Y, "", I);
10829 NewSCC->takeName(I);
10830 // Swap Destinations and condition...
10831 BI.setCondition(NewSCC);
10832 BI.setSuccessor(0, FalseDest);
10833 BI.setSuccessor(1, TrueDest);
10834 RemoveFromWorkList(I);
10835 I->eraseFromParent();
10836 AddToWorkList(NewSCC);
10840 // Cannonicalize icmp_ne -> icmp_eq
10841 ICmpInst::Predicate IPred;
10842 if (match(&BI, m_Br(m_ICmp(IPred, m_Value(X), m_Value(Y)),
10843 TrueDest, FalseDest)))
10844 if ((IPred == ICmpInst::ICMP_NE || IPred == ICmpInst::ICMP_ULE ||
10845 IPred == ICmpInst::ICMP_SLE || IPred == ICmpInst::ICMP_UGE ||
10846 IPred == ICmpInst::ICMP_SGE) && BI.getCondition()->hasOneUse()) {
10847 ICmpInst *I = cast<ICmpInst>(BI.getCondition());
10848 ICmpInst::Predicate NewPred = ICmpInst::getInversePredicate(IPred);
10849 Instruction *NewSCC = new ICmpInst(NewPred, X, Y, "", I);
10850 NewSCC->takeName(I);
10851 // Swap Destinations and condition...
10852 BI.setCondition(NewSCC);
10853 BI.setSuccessor(0, FalseDest);
10854 BI.setSuccessor(1, TrueDest);
10855 RemoveFromWorkList(I);
10856 I->eraseFromParent();;
10857 AddToWorkList(NewSCC);
10864 Instruction *InstCombiner::visitSwitchInst(SwitchInst &SI) {
10865 Value *Cond = SI.getCondition();
10866 if (Instruction *I = dyn_cast<Instruction>(Cond)) {
10867 if (I->getOpcode() == Instruction::Add)
10868 if (ConstantInt *AddRHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
10869 // change 'switch (X+4) case 1:' into 'switch (X) case -3'
10870 for (unsigned i = 2, e = SI.getNumOperands(); i != e; i += 2)
10871 SI.setOperand(i,ConstantExpr::getSub(cast<Constant>(SI.getOperand(i)),
10873 SI.setOperand(0, I->getOperand(0));
10881 Instruction *InstCombiner::visitExtractValueInst(ExtractValueInst &EV) {
10882 Value *Agg = EV.getAggregateOperand();
10884 if (!EV.hasIndices())
10885 return ReplaceInstUsesWith(EV, Agg);
10887 if (Constant *C = dyn_cast<Constant>(Agg)) {
10888 if (isa<UndefValue>(C))
10889 return ReplaceInstUsesWith(EV, UndefValue::get(EV.getType()));
10891 if (isa<ConstantAggregateZero>(C))
10892 return ReplaceInstUsesWith(EV, Constant::getNullValue(EV.getType()));
10894 if (isa<ConstantArray>(C) || isa<ConstantStruct>(C)) {
10895 // Extract the element indexed by the first index out of the constant
10896 Value *V = C->getOperand(*EV.idx_begin());
10897 if (EV.getNumIndices() > 1)
10898 // Extract the remaining indices out of the constant indexed by the
10900 return ExtractValueInst::Create(V, EV.idx_begin() + 1, EV.idx_end());
10902 return ReplaceInstUsesWith(EV, V);
10904 return 0; // Can't handle other constants
10906 if (InsertValueInst *IV = dyn_cast<InsertValueInst>(Agg)) {
10907 // We're extracting from an insertvalue instruction, compare the indices
10908 const unsigned *exti, *exte, *insi, *inse;
10909 for (exti = EV.idx_begin(), insi = IV->idx_begin(),
10910 exte = EV.idx_end(), inse = IV->idx_end();
10911 exti != exte && insi != inse;
10913 if (*insi != *exti)
10914 // The insert and extract both reference distinctly different elements.
10915 // This means the extract is not influenced by the insert, and we can
10916 // replace the aggregate operand of the extract with the aggregate
10917 // operand of the insert. i.e., replace
10918 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1
10919 // %E = extractvalue { i32, { i32 } } %I, 0
10921 // %E = extractvalue { i32, { i32 } } %A, 0
10922 return ExtractValueInst::Create(IV->getAggregateOperand(),
10923 EV.idx_begin(), EV.idx_end());
10925 if (exti == exte && insi == inse)
10926 // Both iterators are at the end: Index lists are identical. Replace
10927 // %B = insertvalue { i32, { i32 } } %A, i32 42, 1, 0
10928 // %C = extractvalue { i32, { i32 } } %B, 1, 0
10930 return ReplaceInstUsesWith(EV, IV->getInsertedValueOperand());
10931 if (exti == exte) {
10932 // The extract list is a prefix of the insert list. i.e. replace
10933 // %I = insertvalue { i32, { i32 } } %A, i32 42, 1, 0
10934 // %E = extractvalue { i32, { i32 } } %I, 1
10936 // %X = extractvalue { i32, { i32 } } %A, 1
10937 // %E = insertvalue { i32 } %X, i32 42, 0
10938 // by switching the order of the insert and extract (though the
10939 // insertvalue should be left in, since it may have other uses).
10940 Value *NewEV = InsertNewInstBefore(
10941 ExtractValueInst::Create(IV->getAggregateOperand(),
10942 EV.idx_begin(), EV.idx_end()),
10944 return InsertValueInst::Create(NewEV, IV->getInsertedValueOperand(),
10948 // The insert list is a prefix of the extract list
10949 // We can simply remove the common indices from the extract and make it
10950 // operate on the inserted value instead of the insertvalue result.
10952 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1
10953 // %E = extractvalue { i32, { i32 } } %I, 1, 0
10955 // %E extractvalue { i32 } { i32 42 }, 0
10956 return ExtractValueInst::Create(IV->getInsertedValueOperand(),
10959 // Can't simplify extracts from other values. Note that nested extracts are
10960 // already simplified implicitely by the above (extract ( extract (insert) )
10961 // will be translated into extract ( insert ( extract ) ) first and then just
10962 // the value inserted, if appropriate).
10966 /// CheapToScalarize - Return true if the value is cheaper to scalarize than it
10967 /// is to leave as a vector operation.
10968 static bool CheapToScalarize(Value *V, bool isConstant) {
10969 if (isa<ConstantAggregateZero>(V))
10971 if (ConstantVector *C = dyn_cast<ConstantVector>(V)) {
10972 if (isConstant) return true;
10973 // If all elts are the same, we can extract.
10974 Constant *Op0 = C->getOperand(0);
10975 for (unsigned i = 1; i < C->getNumOperands(); ++i)
10976 if (C->getOperand(i) != Op0)
10980 Instruction *I = dyn_cast<Instruction>(V);
10981 if (!I) return false;
10983 // Insert element gets simplified to the inserted element or is deleted if
10984 // this is constant idx extract element and its a constant idx insertelt.
10985 if (I->getOpcode() == Instruction::InsertElement && isConstant &&
10986 isa<ConstantInt>(I->getOperand(2)))
10988 if (I->getOpcode() == Instruction::Load && I->hasOneUse())
10990 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I))
10991 if (BO->hasOneUse() &&
10992 (CheapToScalarize(BO->getOperand(0), isConstant) ||
10993 CheapToScalarize(BO->getOperand(1), isConstant)))
10995 if (CmpInst *CI = dyn_cast<CmpInst>(I))
10996 if (CI->hasOneUse() &&
10997 (CheapToScalarize(CI->getOperand(0), isConstant) ||
10998 CheapToScalarize(CI->getOperand(1), isConstant)))
11004 /// Read and decode a shufflevector mask.
11006 /// It turns undef elements into values that are larger than the number of
11007 /// elements in the input.
11008 static std::vector<unsigned> getShuffleMask(const ShuffleVectorInst *SVI) {
11009 unsigned NElts = SVI->getType()->getNumElements();
11010 if (isa<ConstantAggregateZero>(SVI->getOperand(2)))
11011 return std::vector<unsigned>(NElts, 0);
11012 if (isa<UndefValue>(SVI->getOperand(2)))
11013 return std::vector<unsigned>(NElts, 2*NElts);
11015 std::vector<unsigned> Result;
11016 const ConstantVector *CP = cast<ConstantVector>(SVI->getOperand(2));
11017 for (User::const_op_iterator i = CP->op_begin(), e = CP->op_end(); i!=e; ++i)
11018 if (isa<UndefValue>(*i))
11019 Result.push_back(NElts*2); // undef -> 8
11021 Result.push_back(cast<ConstantInt>(*i)->getZExtValue());
11025 /// FindScalarElement - Given a vector and an element number, see if the scalar
11026 /// value is already around as a register, for example if it were inserted then
11027 /// extracted from the vector.
11028 static Value *FindScalarElement(Value *V, unsigned EltNo) {
11029 assert(isa<VectorType>(V->getType()) && "Not looking at a vector?");
11030 const VectorType *PTy = cast<VectorType>(V->getType());
11031 unsigned Width = PTy->getNumElements();
11032 if (EltNo >= Width) // Out of range access.
11033 return UndefValue::get(PTy->getElementType());
11035 if (isa<UndefValue>(V))
11036 return UndefValue::get(PTy->getElementType());
11037 else if (isa<ConstantAggregateZero>(V))
11038 return Constant::getNullValue(PTy->getElementType());
11039 else if (ConstantVector *CP = dyn_cast<ConstantVector>(V))
11040 return CP->getOperand(EltNo);
11041 else if (InsertElementInst *III = dyn_cast<InsertElementInst>(V)) {
11042 // If this is an insert to a variable element, we don't know what it is.
11043 if (!isa<ConstantInt>(III->getOperand(2)))
11045 unsigned IIElt = cast<ConstantInt>(III->getOperand(2))->getZExtValue();
11047 // If this is an insert to the element we are looking for, return the
11049 if (EltNo == IIElt)
11050 return III->getOperand(1);
11052 // Otherwise, the insertelement doesn't modify the value, recurse on its
11054 return FindScalarElement(III->getOperand(0), EltNo);
11055 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(V)) {
11056 unsigned InEl = getShuffleMask(SVI)[EltNo];
11058 return FindScalarElement(SVI->getOperand(0), InEl);
11059 else if (InEl < Width*2)
11060 return FindScalarElement(SVI->getOperand(1), InEl - Width);
11062 return UndefValue::get(PTy->getElementType());
11065 // Otherwise, we don't know.
11069 Instruction *InstCombiner::visitExtractElementInst(ExtractElementInst &EI) {
11070 // If vector val is undef, replace extract with scalar undef.
11071 if (isa<UndefValue>(EI.getOperand(0)))
11072 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
11074 // If vector val is constant 0, replace extract with scalar 0.
11075 if (isa<ConstantAggregateZero>(EI.getOperand(0)))
11076 return ReplaceInstUsesWith(EI, Constant::getNullValue(EI.getType()));
11078 if (ConstantVector *C = dyn_cast<ConstantVector>(EI.getOperand(0))) {
11079 // If vector val is constant with all elements the same, replace EI with
11080 // that element. When the elements are not identical, we cannot replace yet
11081 // (we do that below, but only when the index is constant).
11082 Constant *op0 = C->getOperand(0);
11083 for (unsigned i = 1; i < C->getNumOperands(); ++i)
11084 if (C->getOperand(i) != op0) {
11089 return ReplaceInstUsesWith(EI, op0);
11092 // If extracting a specified index from the vector, see if we can recursively
11093 // find a previously computed scalar that was inserted into the vector.
11094 if (ConstantInt *IdxC = dyn_cast<ConstantInt>(EI.getOperand(1))) {
11095 unsigned IndexVal = IdxC->getZExtValue();
11096 unsigned VectorWidth =
11097 cast<VectorType>(EI.getOperand(0)->getType())->getNumElements();
11099 // If this is extracting an invalid index, turn this into undef, to avoid
11100 // crashing the code below.
11101 if (IndexVal >= VectorWidth)
11102 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
11104 // This instruction only demands the single element from the input vector.
11105 // If the input vector has a single use, simplify it based on this use
11107 if (EI.getOperand(0)->hasOneUse() && VectorWidth != 1) {
11108 uint64_t UndefElts;
11109 if (Value *V = SimplifyDemandedVectorElts(EI.getOperand(0),
11112 EI.setOperand(0, V);
11117 if (Value *Elt = FindScalarElement(EI.getOperand(0), IndexVal))
11118 return ReplaceInstUsesWith(EI, Elt);
11120 // If the this extractelement is directly using a bitcast from a vector of
11121 // the same number of elements, see if we can find the source element from
11122 // it. In this case, we will end up needing to bitcast the scalars.
11123 if (BitCastInst *BCI = dyn_cast<BitCastInst>(EI.getOperand(0))) {
11124 if (const VectorType *VT =
11125 dyn_cast<VectorType>(BCI->getOperand(0)->getType()))
11126 if (VT->getNumElements() == VectorWidth)
11127 if (Value *Elt = FindScalarElement(BCI->getOperand(0), IndexVal))
11128 return new BitCastInst(Elt, EI.getType());
11132 if (Instruction *I = dyn_cast<Instruction>(EI.getOperand(0))) {
11133 if (I->hasOneUse()) {
11134 // Push extractelement into predecessor operation if legal and
11135 // profitable to do so
11136 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
11137 bool isConstantElt = isa<ConstantInt>(EI.getOperand(1));
11138 if (CheapToScalarize(BO, isConstantElt)) {
11139 ExtractElementInst *newEI0 =
11140 new ExtractElementInst(BO->getOperand(0), EI.getOperand(1),
11141 EI.getName()+".lhs");
11142 ExtractElementInst *newEI1 =
11143 new ExtractElementInst(BO->getOperand(1), EI.getOperand(1),
11144 EI.getName()+".rhs");
11145 InsertNewInstBefore(newEI0, EI);
11146 InsertNewInstBefore(newEI1, EI);
11147 return BinaryOperator::Create(BO->getOpcode(), newEI0, newEI1);
11149 } else if (isa<LoadInst>(I)) {
11151 cast<PointerType>(I->getOperand(0)->getType())->getAddressSpace();
11152 Value *Ptr = InsertBitCastBefore(I->getOperand(0),
11153 PointerType::get(EI.getType(), AS),EI);
11154 GetElementPtrInst *GEP =
11155 GetElementPtrInst::Create(Ptr, EI.getOperand(1), I->getName()+".gep");
11156 InsertNewInstBefore(GEP, EI);
11157 return new LoadInst(GEP);
11160 if (InsertElementInst *IE = dyn_cast<InsertElementInst>(I)) {
11161 // Extracting the inserted element?
11162 if (IE->getOperand(2) == EI.getOperand(1))
11163 return ReplaceInstUsesWith(EI, IE->getOperand(1));
11164 // If the inserted and extracted elements are constants, they must not
11165 // be the same value, extract from the pre-inserted value instead.
11166 if (isa<Constant>(IE->getOperand(2)) &&
11167 isa<Constant>(EI.getOperand(1))) {
11168 AddUsesToWorkList(EI);
11169 EI.setOperand(0, IE->getOperand(0));
11172 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I)) {
11173 // If this is extracting an element from a shufflevector, figure out where
11174 // it came from and extract from the appropriate input element instead.
11175 if (ConstantInt *Elt = dyn_cast<ConstantInt>(EI.getOperand(1))) {
11176 unsigned SrcIdx = getShuffleMask(SVI)[Elt->getZExtValue()];
11178 if (SrcIdx < SVI->getType()->getNumElements())
11179 Src = SVI->getOperand(0);
11180 else if (SrcIdx < SVI->getType()->getNumElements()*2) {
11181 SrcIdx -= SVI->getType()->getNumElements();
11182 Src = SVI->getOperand(1);
11184 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
11186 return new ExtractElementInst(Src, SrcIdx);
11193 /// CollectSingleShuffleElements - If V is a shuffle of values that ONLY returns
11194 /// elements from either LHS or RHS, return the shuffle mask and true.
11195 /// Otherwise, return false.
11196 static bool CollectSingleShuffleElements(Value *V, Value *LHS, Value *RHS,
11197 std::vector<Constant*> &Mask) {
11198 assert(V->getType() == LHS->getType() && V->getType() == RHS->getType() &&
11199 "Invalid CollectSingleShuffleElements");
11200 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
11202 if (isa<UndefValue>(V)) {
11203 Mask.assign(NumElts, UndefValue::get(Type::Int32Ty));
11205 } else if (V == LHS) {
11206 for (unsigned i = 0; i != NumElts; ++i)
11207 Mask.push_back(ConstantInt::get(Type::Int32Ty, i));
11209 } else if (V == RHS) {
11210 for (unsigned i = 0; i != NumElts; ++i)
11211 Mask.push_back(ConstantInt::get(Type::Int32Ty, i+NumElts));
11213 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
11214 // If this is an insert of an extract from some other vector, include it.
11215 Value *VecOp = IEI->getOperand(0);
11216 Value *ScalarOp = IEI->getOperand(1);
11217 Value *IdxOp = IEI->getOperand(2);
11219 if (!isa<ConstantInt>(IdxOp))
11221 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
11223 if (isa<UndefValue>(ScalarOp)) { // inserting undef into vector.
11224 // Okay, we can handle this if the vector we are insertinting into is
11225 // transitively ok.
11226 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
11227 // If so, update the mask to reflect the inserted undef.
11228 Mask[InsertedIdx] = UndefValue::get(Type::Int32Ty);
11231 } else if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)){
11232 if (isa<ConstantInt>(EI->getOperand(1)) &&
11233 EI->getOperand(0)->getType() == V->getType()) {
11234 unsigned ExtractedIdx =
11235 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
11237 // This must be extracting from either LHS or RHS.
11238 if (EI->getOperand(0) == LHS || EI->getOperand(0) == RHS) {
11239 // Okay, we can handle this if the vector we are insertinting into is
11240 // transitively ok.
11241 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
11242 // If so, update the mask to reflect the inserted value.
11243 if (EI->getOperand(0) == LHS) {
11244 Mask[InsertedIdx % NumElts] =
11245 ConstantInt::get(Type::Int32Ty, ExtractedIdx);
11247 assert(EI->getOperand(0) == RHS);
11248 Mask[InsertedIdx % NumElts] =
11249 ConstantInt::get(Type::Int32Ty, ExtractedIdx+NumElts);
11258 // TODO: Handle shufflevector here!
11263 /// CollectShuffleElements - We are building a shuffle of V, using RHS as the
11264 /// RHS of the shuffle instruction, if it is not null. Return a shuffle mask
11265 /// that computes V and the LHS value of the shuffle.
11266 static Value *CollectShuffleElements(Value *V, std::vector<Constant*> &Mask,
11268 assert(isa<VectorType>(V->getType()) &&
11269 (RHS == 0 || V->getType() == RHS->getType()) &&
11270 "Invalid shuffle!");
11271 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
11273 if (isa<UndefValue>(V)) {
11274 Mask.assign(NumElts, UndefValue::get(Type::Int32Ty));
11276 } else if (isa<ConstantAggregateZero>(V)) {
11277 Mask.assign(NumElts, ConstantInt::get(Type::Int32Ty, 0));
11279 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
11280 // If this is an insert of an extract from some other vector, include it.
11281 Value *VecOp = IEI->getOperand(0);
11282 Value *ScalarOp = IEI->getOperand(1);
11283 Value *IdxOp = IEI->getOperand(2);
11285 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
11286 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
11287 EI->getOperand(0)->getType() == V->getType()) {
11288 unsigned ExtractedIdx =
11289 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
11290 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
11292 // Either the extracted from or inserted into vector must be RHSVec,
11293 // otherwise we'd end up with a shuffle of three inputs.
11294 if (EI->getOperand(0) == RHS || RHS == 0) {
11295 RHS = EI->getOperand(0);
11296 Value *V = CollectShuffleElements(VecOp, Mask, RHS);
11297 Mask[InsertedIdx % NumElts] =
11298 ConstantInt::get(Type::Int32Ty, NumElts+ExtractedIdx);
11302 if (VecOp == RHS) {
11303 Value *V = CollectShuffleElements(EI->getOperand(0), Mask, RHS);
11304 // Everything but the extracted element is replaced with the RHS.
11305 for (unsigned i = 0; i != NumElts; ++i) {
11306 if (i != InsertedIdx)
11307 Mask[i] = ConstantInt::get(Type::Int32Ty, NumElts+i);
11312 // If this insertelement is a chain that comes from exactly these two
11313 // vectors, return the vector and the effective shuffle.
11314 if (CollectSingleShuffleElements(IEI, EI->getOperand(0), RHS, Mask))
11315 return EI->getOperand(0);
11320 // TODO: Handle shufflevector here!
11322 // Otherwise, can't do anything fancy. Return an identity vector.
11323 for (unsigned i = 0; i != NumElts; ++i)
11324 Mask.push_back(ConstantInt::get(Type::Int32Ty, i));
11328 Instruction *InstCombiner::visitInsertElementInst(InsertElementInst &IE) {
11329 Value *VecOp = IE.getOperand(0);
11330 Value *ScalarOp = IE.getOperand(1);
11331 Value *IdxOp = IE.getOperand(2);
11333 // Inserting an undef or into an undefined place, remove this.
11334 if (isa<UndefValue>(ScalarOp) || isa<UndefValue>(IdxOp))
11335 ReplaceInstUsesWith(IE, VecOp);
11337 // If the inserted element was extracted from some other vector, and if the
11338 // indexes are constant, try to turn this into a shufflevector operation.
11339 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
11340 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
11341 EI->getOperand(0)->getType() == IE.getType()) {
11342 unsigned NumVectorElts = IE.getType()->getNumElements();
11343 unsigned ExtractedIdx =
11344 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
11345 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
11347 if (ExtractedIdx >= NumVectorElts) // Out of range extract.
11348 return ReplaceInstUsesWith(IE, VecOp);
11350 if (InsertedIdx >= NumVectorElts) // Out of range insert.
11351 return ReplaceInstUsesWith(IE, UndefValue::get(IE.getType()));
11353 // If we are extracting a value from a vector, then inserting it right
11354 // back into the same place, just use the input vector.
11355 if (EI->getOperand(0) == VecOp && ExtractedIdx == InsertedIdx)
11356 return ReplaceInstUsesWith(IE, VecOp);
11358 // We could theoretically do this for ANY input. However, doing so could
11359 // turn chains of insertelement instructions into a chain of shufflevector
11360 // instructions, and right now we do not merge shufflevectors. As such,
11361 // only do this in a situation where it is clear that there is benefit.
11362 if (isa<UndefValue>(VecOp) || isa<ConstantAggregateZero>(VecOp)) {
11363 // Turn this into shuffle(EIOp0, VecOp, Mask). The result has all of
11364 // the values of VecOp, except then one read from EIOp0.
11365 // Build a new shuffle mask.
11366 std::vector<Constant*> Mask;
11367 if (isa<UndefValue>(VecOp))
11368 Mask.assign(NumVectorElts, UndefValue::get(Type::Int32Ty));
11370 assert(isa<ConstantAggregateZero>(VecOp) && "Unknown thing");
11371 Mask.assign(NumVectorElts, ConstantInt::get(Type::Int32Ty,
11374 Mask[InsertedIdx] = ConstantInt::get(Type::Int32Ty, ExtractedIdx);
11375 return new ShuffleVectorInst(EI->getOperand(0), VecOp,
11376 ConstantVector::get(Mask));
11379 // If this insertelement isn't used by some other insertelement, turn it
11380 // (and any insertelements it points to), into one big shuffle.
11381 if (!IE.hasOneUse() || !isa<InsertElementInst>(IE.use_back())) {
11382 std::vector<Constant*> Mask;
11384 Value *LHS = CollectShuffleElements(&IE, Mask, RHS);
11385 if (RHS == 0) RHS = UndefValue::get(LHS->getType());
11386 // We now have a shuffle of LHS, RHS, Mask.
11387 return new ShuffleVectorInst(LHS, RHS, ConstantVector::get(Mask));
11396 Instruction *InstCombiner::visitShuffleVectorInst(ShuffleVectorInst &SVI) {
11397 Value *LHS = SVI.getOperand(0);
11398 Value *RHS = SVI.getOperand(1);
11399 std::vector<unsigned> Mask = getShuffleMask(&SVI);
11401 bool MadeChange = false;
11403 // Undefined shuffle mask -> undefined value.
11404 if (isa<UndefValue>(SVI.getOperand(2)))
11405 return ReplaceInstUsesWith(SVI, UndefValue::get(SVI.getType()));
11407 uint64_t UndefElts;
11408 unsigned VWidth = cast<VectorType>(SVI.getType())->getNumElements();
11409 uint64_t AllOnesEltMask = ~0ULL >> (64-VWidth);
11410 if (VWidth <= 64 &&
11411 SimplifyDemandedVectorElts(&SVI, AllOnesEltMask, UndefElts)) {
11412 LHS = SVI.getOperand(0);
11413 RHS = SVI.getOperand(1);
11417 // Canonicalize shuffle(x ,x,mask) -> shuffle(x, undef,mask')
11418 // Canonicalize shuffle(undef,x,mask) -> shuffle(x, undef,mask').
11419 if (LHS == RHS || isa<UndefValue>(LHS)) {
11420 if (isa<UndefValue>(LHS) && LHS == RHS) {
11421 // shuffle(undef,undef,mask) -> undef.
11422 return ReplaceInstUsesWith(SVI, LHS);
11425 // Remap any references to RHS to use LHS.
11426 std::vector<Constant*> Elts;
11427 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
11428 if (Mask[i] >= 2*e)
11429 Elts.push_back(UndefValue::get(Type::Int32Ty));
11431 if ((Mask[i] >= e && isa<UndefValue>(RHS)) ||
11432 (Mask[i] < e && isa<UndefValue>(LHS))) {
11433 Mask[i] = 2*e; // Turn into undef.
11434 Elts.push_back(UndefValue::get(Type::Int32Ty));
11436 Mask[i] = Mask[i] % e; // Force to LHS.
11437 Elts.push_back(ConstantInt::get(Type::Int32Ty, Mask[i]));
11441 SVI.setOperand(0, SVI.getOperand(1));
11442 SVI.setOperand(1, UndefValue::get(RHS->getType()));
11443 SVI.setOperand(2, ConstantVector::get(Elts));
11444 LHS = SVI.getOperand(0);
11445 RHS = SVI.getOperand(1);
11449 // Analyze the shuffle, are the LHS or RHS and identity shuffles?
11450 bool isLHSID = true, isRHSID = true;
11452 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
11453 if (Mask[i] >= e*2) continue; // Ignore undef values.
11454 // Is this an identity shuffle of the LHS value?
11455 isLHSID &= (Mask[i] == i);
11457 // Is this an identity shuffle of the RHS value?
11458 isRHSID &= (Mask[i]-e == i);
11461 // Eliminate identity shuffles.
11462 if (isLHSID) return ReplaceInstUsesWith(SVI, LHS);
11463 if (isRHSID) return ReplaceInstUsesWith(SVI, RHS);
11465 // If the LHS is a shufflevector itself, see if we can combine it with this
11466 // one without producing an unusual shuffle. Here we are really conservative:
11467 // we are absolutely afraid of producing a shuffle mask not in the input
11468 // program, because the code gen may not be smart enough to turn a merged
11469 // shuffle into two specific shuffles: it may produce worse code. As such,
11470 // we only merge two shuffles if the result is one of the two input shuffle
11471 // masks. In this case, merging the shuffles just removes one instruction,
11472 // which we know is safe. This is good for things like turning:
11473 // (splat(splat)) -> splat.
11474 if (ShuffleVectorInst *LHSSVI = dyn_cast<ShuffleVectorInst>(LHS)) {
11475 if (isa<UndefValue>(RHS)) {
11476 std::vector<unsigned> LHSMask = getShuffleMask(LHSSVI);
11478 std::vector<unsigned> NewMask;
11479 for (unsigned i = 0, e = Mask.size(); i != e; ++i)
11480 if (Mask[i] >= 2*e)
11481 NewMask.push_back(2*e);
11483 NewMask.push_back(LHSMask[Mask[i]]);
11485 // If the result mask is equal to the src shuffle or this shuffle mask, do
11486 // the replacement.
11487 if (NewMask == LHSMask || NewMask == Mask) {
11488 std::vector<Constant*> Elts;
11489 for (unsigned i = 0, e = NewMask.size(); i != e; ++i) {
11490 if (NewMask[i] >= e*2) {
11491 Elts.push_back(UndefValue::get(Type::Int32Ty));
11493 Elts.push_back(ConstantInt::get(Type::Int32Ty, NewMask[i]));
11496 return new ShuffleVectorInst(LHSSVI->getOperand(0),
11497 LHSSVI->getOperand(1),
11498 ConstantVector::get(Elts));
11503 return MadeChange ? &SVI : 0;
11509 /// TryToSinkInstruction - Try to move the specified instruction from its
11510 /// current block into the beginning of DestBlock, which can only happen if it's
11511 /// safe to move the instruction past all of the instructions between it and the
11512 /// end of its block.
11513 static bool TryToSinkInstruction(Instruction *I, BasicBlock *DestBlock) {
11514 assert(I->hasOneUse() && "Invariants didn't hold!");
11516 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
11517 if (isa<PHINode>(I) || I->mayWriteToMemory() || isa<TerminatorInst>(I))
11520 // Do not sink alloca instructions out of the entry block.
11521 if (isa<AllocaInst>(I) && I->getParent() ==
11522 &DestBlock->getParent()->getEntryBlock())
11525 // We can only sink load instructions if there is nothing between the load and
11526 // the end of block that could change the value.
11527 if (I->mayReadFromMemory()) {
11528 for (BasicBlock::iterator Scan = I, E = I->getParent()->end();
11530 if (Scan->mayWriteToMemory())
11534 BasicBlock::iterator InsertPos = DestBlock->getFirstNonPHI();
11536 I->moveBefore(InsertPos);
11542 /// AddReachableCodeToWorklist - Walk the function in depth-first order, adding
11543 /// all reachable code to the worklist.
11545 /// This has a couple of tricks to make the code faster and more powerful. In
11546 /// particular, we constant fold and DCE instructions as we go, to avoid adding
11547 /// them to the worklist (this significantly speeds up instcombine on code where
11548 /// many instructions are dead or constant). Additionally, if we find a branch
11549 /// whose condition is a known constant, we only visit the reachable successors.
11551 static void AddReachableCodeToWorklist(BasicBlock *BB,
11552 SmallPtrSet<BasicBlock*, 64> &Visited,
11554 const TargetData *TD) {
11555 SmallVector<BasicBlock*, 256> Worklist;
11556 Worklist.push_back(BB);
11558 while (!Worklist.empty()) {
11559 BB = Worklist.back();
11560 Worklist.pop_back();
11562 // We have now visited this block! If we've already been here, ignore it.
11563 if (!Visited.insert(BB)) continue;
11565 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) {
11566 Instruction *Inst = BBI++;
11568 // DCE instruction if trivially dead.
11569 if (isInstructionTriviallyDead(Inst)) {
11571 DOUT << "IC: DCE: " << *Inst;
11572 Inst->eraseFromParent();
11576 // ConstantProp instruction if trivially constant.
11577 if (Constant *C = ConstantFoldInstruction(Inst, TD)) {
11578 DOUT << "IC: ConstFold to: " << *C << " from: " << *Inst;
11579 Inst->replaceAllUsesWith(C);
11581 Inst->eraseFromParent();
11585 IC.AddToWorkList(Inst);
11588 // Recursively visit successors. If this is a branch or switch on a
11589 // constant, only visit the reachable successor.
11590 TerminatorInst *TI = BB->getTerminator();
11591 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
11592 if (BI->isConditional() && isa<ConstantInt>(BI->getCondition())) {
11593 bool CondVal = cast<ConstantInt>(BI->getCondition())->getZExtValue();
11594 BasicBlock *ReachableBB = BI->getSuccessor(!CondVal);
11595 Worklist.push_back(ReachableBB);
11598 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
11599 if (ConstantInt *Cond = dyn_cast<ConstantInt>(SI->getCondition())) {
11600 // See if this is an explicit destination.
11601 for (unsigned i = 1, e = SI->getNumSuccessors(); i != e; ++i)
11602 if (SI->getCaseValue(i) == Cond) {
11603 BasicBlock *ReachableBB = SI->getSuccessor(i);
11604 Worklist.push_back(ReachableBB);
11608 // Otherwise it is the default destination.
11609 Worklist.push_back(SI->getSuccessor(0));
11614 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
11615 Worklist.push_back(TI->getSuccessor(i));
11619 bool InstCombiner::DoOneIteration(Function &F, unsigned Iteration) {
11620 bool Changed = false;
11621 TD = &getAnalysis<TargetData>();
11623 DEBUG(DOUT << "\n\nINSTCOMBINE ITERATION #" << Iteration << " on "
11624 << F.getNameStr() << "\n");
11627 // Do a depth-first traversal of the function, populate the worklist with
11628 // the reachable instructions. Ignore blocks that are not reachable. Keep
11629 // track of which blocks we visit.
11630 SmallPtrSet<BasicBlock*, 64> Visited;
11631 AddReachableCodeToWorklist(F.begin(), Visited, *this, TD);
11633 // Do a quick scan over the function. If we find any blocks that are
11634 // unreachable, remove any instructions inside of them. This prevents
11635 // the instcombine code from having to deal with some bad special cases.
11636 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
11637 if (!Visited.count(BB)) {
11638 Instruction *Term = BB->getTerminator();
11639 while (Term != BB->begin()) { // Remove instrs bottom-up
11640 BasicBlock::iterator I = Term; --I;
11642 DOUT << "IC: DCE: " << *I;
11645 if (!I->use_empty())
11646 I->replaceAllUsesWith(UndefValue::get(I->getType()));
11647 I->eraseFromParent();
11652 while (!Worklist.empty()) {
11653 Instruction *I = RemoveOneFromWorkList();
11654 if (I == 0) continue; // skip null values.
11656 // Check to see if we can DCE the instruction.
11657 if (isInstructionTriviallyDead(I)) {
11658 // Add operands to the worklist.
11659 if (I->getNumOperands() < 4)
11660 AddUsesToWorkList(*I);
11663 DOUT << "IC: DCE: " << *I;
11665 I->eraseFromParent();
11666 RemoveFromWorkList(I);
11670 // Instruction isn't dead, see if we can constant propagate it.
11671 if (Constant *C = ConstantFoldInstruction(I, TD)) {
11672 DOUT << "IC: ConstFold to: " << *C << " from: " << *I;
11674 // Add operands to the worklist.
11675 AddUsesToWorkList(*I);
11676 ReplaceInstUsesWith(*I, C);
11679 I->eraseFromParent();
11680 RemoveFromWorkList(I);
11684 if (TD && I->getType()->getTypeID() == Type::VoidTyID) {
11685 // See if we can constant fold its operands.
11686 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i) {
11687 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(i)) {
11688 if (Constant *NewC = ConstantFoldConstantExpression(CE, TD))
11694 // See if we can trivially sink this instruction to a successor basic block.
11695 if (I->hasOneUse()) {
11696 BasicBlock *BB = I->getParent();
11697 BasicBlock *UserParent = cast<Instruction>(I->use_back())->getParent();
11698 if (UserParent != BB) {
11699 bool UserIsSuccessor = false;
11700 // See if the user is one of our successors.
11701 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI)
11702 if (*SI == UserParent) {
11703 UserIsSuccessor = true;
11707 // If the user is one of our immediate successors, and if that successor
11708 // only has us as a predecessors (we'd have to split the critical edge
11709 // otherwise), we can keep going.
11710 if (UserIsSuccessor && !isa<PHINode>(I->use_back()) &&
11711 next(pred_begin(UserParent)) == pred_end(UserParent))
11712 // Okay, the CFG is simple enough, try to sink this instruction.
11713 Changed |= TryToSinkInstruction(I, UserParent);
11717 // Now that we have an instruction, try combining it to simplify it...
11721 DEBUG(std::ostringstream SS; I->print(SS); OrigI = SS.str(););
11722 if (Instruction *Result = visit(*I)) {
11724 // Should we replace the old instruction with a new one?
11726 DOUT << "IC: Old = " << *I
11727 << " New = " << *Result;
11729 // Everything uses the new instruction now.
11730 I->replaceAllUsesWith(Result);
11732 // Push the new instruction and any users onto the worklist.
11733 AddToWorkList(Result);
11734 AddUsersToWorkList(*Result);
11736 // Move the name to the new instruction first.
11737 Result->takeName(I);
11739 // Insert the new instruction into the basic block...
11740 BasicBlock *InstParent = I->getParent();
11741 BasicBlock::iterator InsertPos = I;
11743 if (!isa<PHINode>(Result)) // If combining a PHI, don't insert
11744 while (isa<PHINode>(InsertPos)) // middle of a block of PHIs.
11747 InstParent->getInstList().insert(InsertPos, Result);
11749 // Make sure that we reprocess all operands now that we reduced their
11751 AddUsesToWorkList(*I);
11753 // Instructions can end up on the worklist more than once. Make sure
11754 // we do not process an instruction that has been deleted.
11755 RemoveFromWorkList(I);
11757 // Erase the old instruction.
11758 InstParent->getInstList().erase(I);
11761 DOUT << "IC: Mod = " << OrigI
11762 << " New = " << *I;
11765 // If the instruction was modified, it's possible that it is now dead.
11766 // if so, remove it.
11767 if (isInstructionTriviallyDead(I)) {
11768 // Make sure we process all operands now that we are reducing their
11770 AddUsesToWorkList(*I);
11772 // Instructions may end up in the worklist more than once. Erase all
11773 // occurrences of this instruction.
11774 RemoveFromWorkList(I);
11775 I->eraseFromParent();
11778 AddUsersToWorkList(*I);
11785 assert(WorklistMap.empty() && "Worklist empty, but map not?");
11787 // Do an explicit clear, this shrinks the map if needed.
11788 WorklistMap.clear();
11793 bool InstCombiner::runOnFunction(Function &F) {
11794 MustPreserveLCSSA = mustPreserveAnalysisID(LCSSAID);
11796 bool EverMadeChange = false;
11798 // Iterate while there is work to do.
11799 unsigned Iteration = 0;
11800 while (DoOneIteration(F, Iteration++))
11801 EverMadeChange = true;
11802 return EverMadeChange;
11805 FunctionPass *llvm::createInstructionCombiningPass() {
11806 return new InstCombiner();