1 //===- InstructionCombining.cpp - Combine multiple instructions -----------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // InstructionCombining - Combine instructions to form fewer, simple
11 // instructions. This pass does not modify the CFG. This pass is where
12 // algebraic simplification happens.
14 // This pass combines things like:
20 // This is a simple worklist driven algorithm.
22 // This pass guarantees that the following canonicalizations are performed on
24 // 1. If a binary operator has a constant operand, it is moved to the RHS
25 // 2. Bitwise operators with constant operands are always grouped so that
26 // shifts are performed first, then or's, then and's, then xor's.
27 // 3. Compare instructions are converted from <,>,<=,>= to ==,!= if possible
28 // 4. All cmp instructions on boolean values are replaced with logical ops
29 // 5. add X, X is represented as (X*2) => (X << 1)
30 // 6. Multiplies with a power-of-two constant argument are transformed into
34 //===----------------------------------------------------------------------===//
36 #define DEBUG_TYPE "instcombine"
37 #include "llvm/Transforms/Scalar.h"
38 #include "llvm/IntrinsicInst.h"
39 #include "llvm/Pass.h"
40 #include "llvm/DerivedTypes.h"
41 #include "llvm/GlobalVariable.h"
42 #include "llvm/Analysis/ConstantFolding.h"
43 #include "llvm/Analysis/ValueTracking.h"
44 #include "llvm/Target/TargetData.h"
45 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
46 #include "llvm/Transforms/Utils/Local.h"
47 #include "llvm/Support/CallSite.h"
48 #include "llvm/Support/ConstantRange.h"
49 #include "llvm/Support/Debug.h"
50 #include "llvm/Support/GetElementPtrTypeIterator.h"
51 #include "llvm/Support/InstVisitor.h"
52 #include "llvm/Support/MathExtras.h"
53 #include "llvm/Support/PatternMatch.h"
54 #include "llvm/Support/Compiler.h"
55 #include "llvm/ADT/DenseMap.h"
56 #include "llvm/ADT/SmallVector.h"
57 #include "llvm/ADT/SmallPtrSet.h"
58 #include "llvm/ADT/Statistic.h"
59 #include "llvm/ADT/STLExtras.h"
64 using namespace llvm::PatternMatch;
66 STATISTIC(NumCombined , "Number of insts combined");
67 STATISTIC(NumConstProp, "Number of constant folds");
68 STATISTIC(NumDeadInst , "Number of dead inst eliminated");
69 STATISTIC(NumDeadStore, "Number of dead stores eliminated");
70 STATISTIC(NumSunkInst , "Number of instructions sunk");
73 class VISIBILITY_HIDDEN InstCombiner
74 : public FunctionPass,
75 public InstVisitor<InstCombiner, Instruction*> {
76 // Worklist of all of the instructions that need to be simplified.
77 SmallVector<Instruction*, 256> Worklist;
78 DenseMap<Instruction*, unsigned> WorklistMap;
80 bool MustPreserveLCSSA;
82 static char ID; // Pass identification, replacement for typeid
83 InstCombiner() : FunctionPass(&ID) {}
85 /// AddToWorkList - Add the specified instruction to the worklist if it
86 /// isn't already in it.
87 void AddToWorkList(Instruction *I) {
88 if (WorklistMap.insert(std::make_pair(I, Worklist.size())).second)
89 Worklist.push_back(I);
92 // RemoveFromWorkList - remove I from the worklist if it exists.
93 void RemoveFromWorkList(Instruction *I) {
94 DenseMap<Instruction*, unsigned>::iterator It = WorklistMap.find(I);
95 if (It == WorklistMap.end()) return; // Not in worklist.
97 // Don't bother moving everything down, just null out the slot.
98 Worklist[It->second] = 0;
100 WorklistMap.erase(It);
103 Instruction *RemoveOneFromWorkList() {
104 Instruction *I = Worklist.back();
106 WorklistMap.erase(I);
111 /// AddUsersToWorkList - When an instruction is simplified, add all users of
112 /// the instruction to the work lists because they might get more simplified
115 void AddUsersToWorkList(Value &I) {
116 for (Value::use_iterator UI = I.use_begin(), UE = I.use_end();
118 AddToWorkList(cast<Instruction>(*UI));
121 /// AddUsesToWorkList - When an instruction is simplified, add operands to
122 /// the work lists because they might get more simplified now.
124 void AddUsesToWorkList(Instruction &I) {
125 for (User::op_iterator i = I.op_begin(), e = I.op_end(); i != e; ++i)
126 if (Instruction *Op = dyn_cast<Instruction>(*i))
130 /// AddSoonDeadInstToWorklist - The specified instruction is about to become
131 /// dead. Add all of its operands to the worklist, turning them into
132 /// undef's to reduce the number of uses of those instructions.
134 /// Return the specified operand before it is turned into an undef.
136 Value *AddSoonDeadInstToWorklist(Instruction &I, unsigned op) {
137 Value *R = I.getOperand(op);
139 for (User::op_iterator i = I.op_begin(), e = I.op_end(); i != e; ++i)
140 if (Instruction *Op = dyn_cast<Instruction>(*i)) {
142 // Set the operand to undef to drop the use.
143 *i = UndefValue::get(Op->getType());
150 virtual bool runOnFunction(Function &F);
152 bool DoOneIteration(Function &F, unsigned ItNum);
154 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
155 AU.addRequired<TargetData>();
156 AU.addPreservedID(LCSSAID);
157 AU.setPreservesCFG();
160 TargetData &getTargetData() const { return *TD; }
162 // Visitation implementation - Implement instruction combining for different
163 // instruction types. The semantics are as follows:
165 // null - No change was made
166 // I - Change was made, I is still valid, I may be dead though
167 // otherwise - Change was made, replace I with returned instruction
169 Instruction *visitAdd(BinaryOperator &I);
170 Instruction *visitSub(BinaryOperator &I);
171 Instruction *visitMul(BinaryOperator &I);
172 Instruction *visitURem(BinaryOperator &I);
173 Instruction *visitSRem(BinaryOperator &I);
174 Instruction *visitFRem(BinaryOperator &I);
175 bool SimplifyDivRemOfSelect(BinaryOperator &I);
176 Instruction *commonRemTransforms(BinaryOperator &I);
177 Instruction *commonIRemTransforms(BinaryOperator &I);
178 Instruction *commonDivTransforms(BinaryOperator &I);
179 Instruction *commonIDivTransforms(BinaryOperator &I);
180 Instruction *visitUDiv(BinaryOperator &I);
181 Instruction *visitSDiv(BinaryOperator &I);
182 Instruction *visitFDiv(BinaryOperator &I);
183 Instruction *FoldAndOfICmps(Instruction &I, ICmpInst *LHS, ICmpInst *RHS);
184 Instruction *visitAnd(BinaryOperator &I);
185 Instruction *FoldOrOfICmps(Instruction &I, ICmpInst *LHS, ICmpInst *RHS);
186 Instruction *FoldOrWithConstants(BinaryOperator &I, Value *Op,
187 Value *A, Value *B, Value *C);
188 Instruction *visitOr (BinaryOperator &I);
189 Instruction *visitXor(BinaryOperator &I);
190 Instruction *visitShl(BinaryOperator &I);
191 Instruction *visitAShr(BinaryOperator &I);
192 Instruction *visitLShr(BinaryOperator &I);
193 Instruction *commonShiftTransforms(BinaryOperator &I);
194 Instruction *FoldFCmp_IntToFP_Cst(FCmpInst &I, Instruction *LHSI,
196 Instruction *visitFCmpInst(FCmpInst &I);
197 Instruction *visitICmpInst(ICmpInst &I);
198 Instruction *visitICmpInstWithCastAndCast(ICmpInst &ICI);
199 Instruction *visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
202 Instruction *FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
203 ConstantInt *DivRHS);
205 Instruction *FoldGEPICmp(User *GEPLHS, Value *RHS,
206 ICmpInst::Predicate Cond, Instruction &I);
207 Instruction *FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
209 Instruction *commonCastTransforms(CastInst &CI);
210 Instruction *commonIntCastTransforms(CastInst &CI);
211 Instruction *commonPointerCastTransforms(CastInst &CI);
212 Instruction *visitTrunc(TruncInst &CI);
213 Instruction *visitZExt(ZExtInst &CI);
214 Instruction *visitSExt(SExtInst &CI);
215 Instruction *visitFPTrunc(FPTruncInst &CI);
216 Instruction *visitFPExt(CastInst &CI);
217 Instruction *visitFPToUI(FPToUIInst &FI);
218 Instruction *visitFPToSI(FPToSIInst &FI);
219 Instruction *visitUIToFP(CastInst &CI);
220 Instruction *visitSIToFP(CastInst &CI);
221 Instruction *visitPtrToInt(CastInst &CI);
222 Instruction *visitIntToPtr(IntToPtrInst &CI);
223 Instruction *visitBitCast(BitCastInst &CI);
224 Instruction *FoldSelectOpOp(SelectInst &SI, Instruction *TI,
226 Instruction *visitSelectInst(SelectInst &SI);
227 Instruction *visitSelectInstWithICmp(SelectInst &SI, ICmpInst *ICI);
228 Instruction *visitCallInst(CallInst &CI);
229 Instruction *visitInvokeInst(InvokeInst &II);
230 Instruction *visitPHINode(PHINode &PN);
231 Instruction *visitGetElementPtrInst(GetElementPtrInst &GEP);
232 Instruction *visitAllocationInst(AllocationInst &AI);
233 Instruction *visitFreeInst(FreeInst &FI);
234 Instruction *visitLoadInst(LoadInst &LI);
235 Instruction *visitStoreInst(StoreInst &SI);
236 Instruction *visitBranchInst(BranchInst &BI);
237 Instruction *visitSwitchInst(SwitchInst &SI);
238 Instruction *visitInsertElementInst(InsertElementInst &IE);
239 Instruction *visitExtractElementInst(ExtractElementInst &EI);
240 Instruction *visitShuffleVectorInst(ShuffleVectorInst &SVI);
241 Instruction *visitExtractValueInst(ExtractValueInst &EV);
243 // visitInstruction - Specify what to return for unhandled instructions...
244 Instruction *visitInstruction(Instruction &I) { return 0; }
247 Instruction *visitCallSite(CallSite CS);
248 bool transformConstExprCastCall(CallSite CS);
249 Instruction *transformCallThroughTrampoline(CallSite CS);
250 Instruction *transformZExtICmp(ICmpInst *ICI, Instruction &CI,
251 bool DoXform = true);
252 bool WillNotOverflowSignedAdd(Value *LHS, Value *RHS);
255 // InsertNewInstBefore - insert an instruction New before instruction Old
256 // in the program. Add the new instruction to the worklist.
258 Instruction *InsertNewInstBefore(Instruction *New, Instruction &Old) {
259 assert(New && New->getParent() == 0 &&
260 "New instruction already inserted into a basic block!");
261 BasicBlock *BB = Old.getParent();
262 BB->getInstList().insert(&Old, New); // Insert inst
267 /// InsertCastBefore - Insert a cast of V to TY before the instruction POS.
268 /// This also adds the cast to the worklist. Finally, this returns the
270 Value *InsertCastBefore(Instruction::CastOps opc, Value *V, const Type *Ty,
272 if (V->getType() == Ty) return V;
274 if (Constant *CV = dyn_cast<Constant>(V))
275 return ConstantExpr::getCast(opc, CV, Ty);
277 Instruction *C = CastInst::Create(opc, V, Ty, V->getName(), &Pos);
282 Value *InsertBitCastBefore(Value *V, const Type *Ty, Instruction &Pos) {
283 return InsertCastBefore(Instruction::BitCast, V, Ty, Pos);
287 // ReplaceInstUsesWith - This method is to be used when an instruction is
288 // found to be dead, replacable with another preexisting expression. Here
289 // we add all uses of I to the worklist, replace all uses of I with the new
290 // value, then return I, so that the inst combiner will know that I was
293 Instruction *ReplaceInstUsesWith(Instruction &I, Value *V) {
294 AddUsersToWorkList(I); // Add all modified instrs to worklist
296 I.replaceAllUsesWith(V);
299 // If we are replacing the instruction with itself, this must be in a
300 // segment of unreachable code, so just clobber the instruction.
301 I.replaceAllUsesWith(UndefValue::get(I.getType()));
306 // EraseInstFromFunction - When dealing with an instruction that has side
307 // effects or produces a void value, we can't rely on DCE to delete the
308 // instruction. Instead, visit methods should return the value returned by
310 Instruction *EraseInstFromFunction(Instruction &I) {
311 assert(I.use_empty() && "Cannot erase instruction that is used!");
312 AddUsesToWorkList(I);
313 RemoveFromWorkList(&I);
315 return 0; // Don't do anything with FI
318 void ComputeMaskedBits(Value *V, const APInt &Mask, APInt &KnownZero,
319 APInt &KnownOne, unsigned Depth = 0) const {
320 return llvm::ComputeMaskedBits(V, Mask, KnownZero, KnownOne, TD, Depth);
323 bool MaskedValueIsZero(Value *V, const APInt &Mask,
324 unsigned Depth = 0) const {
325 return llvm::MaskedValueIsZero(V, Mask, TD, Depth);
327 unsigned ComputeNumSignBits(Value *Op, unsigned Depth = 0) const {
328 return llvm::ComputeNumSignBits(Op, TD, Depth);
333 /// SimplifyCommutative - This performs a few simplifications for
334 /// commutative operators.
335 bool SimplifyCommutative(BinaryOperator &I);
337 /// SimplifyCompare - This reorders the operands of a CmpInst to get them in
338 /// most-complex to least-complex order.
339 bool SimplifyCompare(CmpInst &I);
341 /// SimplifyDemandedUseBits - Attempts to replace V with a simpler value
342 /// based on the demanded bits.
343 Value *SimplifyDemandedUseBits(Value *V, APInt DemandedMask,
344 APInt& KnownZero, APInt& KnownOne,
346 bool SimplifyDemandedBits(Use &U, APInt DemandedMask,
347 APInt& KnownZero, APInt& KnownOne,
350 /// SimplifyDemandedInstructionBits - Inst is an integer instruction that
351 /// SimplifyDemandedBits knows about. See if the instruction has any
352 /// properties that allow us to simplify its operands.
353 bool SimplifyDemandedInstructionBits(Instruction &Inst);
355 Value *SimplifyDemandedVectorElts(Value *V, uint64_t DemandedElts,
356 uint64_t &UndefElts, unsigned Depth = 0);
358 // FoldOpIntoPhi - Given a binary operator or cast instruction which has a
359 // PHI node as operand #0, see if we can fold the instruction into the PHI
360 // (which is only possible if all operands to the PHI are constants).
361 Instruction *FoldOpIntoPhi(Instruction &I);
363 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
364 // operator and they all are only used by the PHI, PHI together their
365 // inputs, and do the operation once, to the result of the PHI.
366 Instruction *FoldPHIArgOpIntoPHI(PHINode &PN);
367 Instruction *FoldPHIArgBinOpIntoPHI(PHINode &PN);
368 Instruction *FoldPHIArgGEPIntoPHI(PHINode &PN);
371 Instruction *OptAndOp(Instruction *Op, ConstantInt *OpRHS,
372 ConstantInt *AndRHS, BinaryOperator &TheAnd);
374 Value *FoldLogicalPlusAnd(Value *LHS, Value *RHS, ConstantInt *Mask,
375 bool isSub, Instruction &I);
376 Instruction *InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
377 bool isSigned, bool Inside, Instruction &IB);
378 Instruction *PromoteCastOfAllocation(BitCastInst &CI, AllocationInst &AI);
379 Instruction *MatchBSwap(BinaryOperator &I);
380 bool SimplifyStoreAtEndOfBlock(StoreInst &SI);
381 Instruction *SimplifyMemTransfer(MemIntrinsic *MI);
382 Instruction *SimplifyMemSet(MemSetInst *MI);
385 Value *EvaluateInDifferentType(Value *V, const Type *Ty, bool isSigned);
387 bool CanEvaluateInDifferentType(Value *V, const IntegerType *Ty,
388 unsigned CastOpc, int &NumCastsRemoved);
389 unsigned GetOrEnforceKnownAlignment(Value *V,
390 unsigned PrefAlign = 0);
395 char InstCombiner::ID = 0;
396 static RegisterPass<InstCombiner>
397 X("instcombine", "Combine redundant instructions");
399 // getComplexity: Assign a complexity or rank value to LLVM Values...
400 // 0 -> undef, 1 -> Const, 2 -> Other, 3 -> Arg, 3 -> Unary, 4 -> OtherInst
401 static unsigned getComplexity(Value *V) {
402 if (isa<Instruction>(V)) {
403 if (BinaryOperator::isNeg(V) || BinaryOperator::isNot(V))
407 if (isa<Argument>(V)) return 3;
408 return isa<Constant>(V) ? (isa<UndefValue>(V) ? 0 : 1) : 2;
411 // isOnlyUse - Return true if this instruction will be deleted if we stop using
413 static bool isOnlyUse(Value *V) {
414 return V->hasOneUse() || isa<Constant>(V);
417 // getPromotedType - Return the specified type promoted as it would be to pass
418 // though a va_arg area...
419 static const Type *getPromotedType(const Type *Ty) {
420 if (const IntegerType* ITy = dyn_cast<IntegerType>(Ty)) {
421 if (ITy->getBitWidth() < 32)
422 return Type::Int32Ty;
427 /// getBitCastOperand - If the specified operand is a CastInst, a constant
428 /// expression bitcast, or a GetElementPtrInst with all zero indices, return the
429 /// operand value, otherwise return null.
430 static Value *getBitCastOperand(Value *V) {
431 if (BitCastInst *I = dyn_cast<BitCastInst>(V))
433 return I->getOperand(0);
434 else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(V)) {
435 // GetElementPtrInst?
436 if (GEP->hasAllZeroIndices())
437 return GEP->getOperand(0);
438 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) {
439 if (CE->getOpcode() == Instruction::BitCast)
440 // BitCast ConstantExp?
441 return CE->getOperand(0);
442 else if (CE->getOpcode() == Instruction::GetElementPtr) {
443 // GetElementPtr ConstantExp?
444 for (User::op_iterator I = CE->op_begin() + 1, E = CE->op_end();
446 ConstantInt *CI = dyn_cast<ConstantInt>(I);
447 if (!CI || !CI->isZero())
448 // Any non-zero indices? Not cast-like.
451 // All-zero indices? This is just like casting.
452 return CE->getOperand(0);
458 /// This function is a wrapper around CastInst::isEliminableCastPair. It
459 /// simply extracts arguments and returns what that function returns.
460 static Instruction::CastOps
461 isEliminableCastPair(
462 const CastInst *CI, ///< The first cast instruction
463 unsigned opcode, ///< The opcode of the second cast instruction
464 const Type *DstTy, ///< The target type for the second cast instruction
465 TargetData *TD ///< The target data for pointer size
468 const Type *SrcTy = CI->getOperand(0)->getType(); // A from above
469 const Type *MidTy = CI->getType(); // B from above
471 // Get the opcodes of the two Cast instructions
472 Instruction::CastOps firstOp = Instruction::CastOps(CI->getOpcode());
473 Instruction::CastOps secondOp = Instruction::CastOps(opcode);
475 return Instruction::CastOps(
476 CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy,
477 DstTy, TD->getIntPtrType()));
480 /// ValueRequiresCast - Return true if the cast from "V to Ty" actually results
481 /// in any code being generated. It does not require codegen if V is simple
482 /// enough or if the cast can be folded into other casts.
483 static bool ValueRequiresCast(Instruction::CastOps opcode, const Value *V,
484 const Type *Ty, TargetData *TD) {
485 if (V->getType() == Ty || isa<Constant>(V)) return false;
487 // If this is another cast that can be eliminated, it isn't codegen either.
488 if (const CastInst *CI = dyn_cast<CastInst>(V))
489 if (isEliminableCastPair(CI, opcode, Ty, TD))
494 // SimplifyCommutative - This performs a few simplifications for commutative
497 // 1. Order operands such that they are listed from right (least complex) to
498 // left (most complex). This puts constants before unary operators before
501 // 2. Transform: (op (op V, C1), C2) ==> (op V, (op C1, C2))
502 // 3. Transform: (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
504 bool InstCombiner::SimplifyCommutative(BinaryOperator &I) {
505 bool Changed = false;
506 if (getComplexity(I.getOperand(0)) < getComplexity(I.getOperand(1)))
507 Changed = !I.swapOperands();
509 if (!I.isAssociative()) return Changed;
510 Instruction::BinaryOps Opcode = I.getOpcode();
511 if (BinaryOperator *Op = dyn_cast<BinaryOperator>(I.getOperand(0)))
512 if (Op->getOpcode() == Opcode && isa<Constant>(Op->getOperand(1))) {
513 if (isa<Constant>(I.getOperand(1))) {
514 Constant *Folded = ConstantExpr::get(I.getOpcode(),
515 cast<Constant>(I.getOperand(1)),
516 cast<Constant>(Op->getOperand(1)));
517 I.setOperand(0, Op->getOperand(0));
518 I.setOperand(1, Folded);
520 } else if (BinaryOperator *Op1=dyn_cast<BinaryOperator>(I.getOperand(1)))
521 if (Op1->getOpcode() == Opcode && isa<Constant>(Op1->getOperand(1)) &&
522 isOnlyUse(Op) && isOnlyUse(Op1)) {
523 Constant *C1 = cast<Constant>(Op->getOperand(1));
524 Constant *C2 = cast<Constant>(Op1->getOperand(1));
526 // Fold (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
527 Constant *Folded = ConstantExpr::get(I.getOpcode(), C1, C2);
528 Instruction *New = BinaryOperator::Create(Opcode, Op->getOperand(0),
532 I.setOperand(0, New);
533 I.setOperand(1, Folded);
540 /// SimplifyCompare - For a CmpInst this function just orders the operands
541 /// so that theyare listed from right (least complex) to left (most complex).
542 /// This puts constants before unary operators before binary operators.
543 bool InstCombiner::SimplifyCompare(CmpInst &I) {
544 if (getComplexity(I.getOperand(0)) >= getComplexity(I.getOperand(1)))
547 // Compare instructions are not associative so there's nothing else we can do.
551 // dyn_castNegVal - Given a 'sub' instruction, return the RHS of the instruction
552 // if the LHS is a constant zero (which is the 'negate' form).
554 static inline Value *dyn_castNegVal(Value *V) {
555 if (BinaryOperator::isNeg(V))
556 return BinaryOperator::getNegArgument(V);
558 // Constants can be considered to be negated values if they can be folded.
559 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
560 return ConstantExpr::getNeg(C);
562 if (ConstantVector *C = dyn_cast<ConstantVector>(V))
563 if (C->getType()->getElementType()->isInteger())
564 return ConstantExpr::getNeg(C);
569 static inline Value *dyn_castNotVal(Value *V) {
570 if (BinaryOperator::isNot(V))
571 return BinaryOperator::getNotArgument(V);
573 // Constants can be considered to be not'ed values...
574 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
575 return ConstantInt::get(~C->getValue());
579 // dyn_castFoldableMul - If this value is a multiply that can be folded into
580 // other computations (because it has a constant operand), return the
581 // non-constant operand of the multiply, and set CST to point to the multiplier.
582 // Otherwise, return null.
584 static inline Value *dyn_castFoldableMul(Value *V, ConstantInt *&CST) {
585 if (V->hasOneUse() && V->getType()->isInteger())
586 if (Instruction *I = dyn_cast<Instruction>(V)) {
587 if (I->getOpcode() == Instruction::Mul)
588 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1))))
589 return I->getOperand(0);
590 if (I->getOpcode() == Instruction::Shl)
591 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1)))) {
592 // The multiplier is really 1 << CST.
593 uint32_t BitWidth = cast<IntegerType>(V->getType())->getBitWidth();
594 uint32_t CSTVal = CST->getLimitedValue(BitWidth);
595 CST = ConstantInt::get(APInt(BitWidth, 1).shl(CSTVal));
596 return I->getOperand(0);
602 /// dyn_castGetElementPtr - If this is a getelementptr instruction or constant
603 /// expression, return it.
604 static User *dyn_castGetElementPtr(Value *V) {
605 if (isa<GetElementPtrInst>(V)) return cast<User>(V);
606 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
607 if (CE->getOpcode() == Instruction::GetElementPtr)
608 return cast<User>(V);
612 /// getOpcode - If this is an Instruction or a ConstantExpr, return the
613 /// opcode value. Otherwise return UserOp1.
614 static unsigned getOpcode(const Value *V) {
615 if (const Instruction *I = dyn_cast<Instruction>(V))
616 return I->getOpcode();
617 if (const ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
618 return CE->getOpcode();
619 // Use UserOp1 to mean there's no opcode.
620 return Instruction::UserOp1;
623 /// AddOne - Add one to a ConstantInt
624 static ConstantInt *AddOne(ConstantInt *C) {
625 APInt Val(C->getValue());
626 return ConstantInt::get(++Val);
628 /// SubOne - Subtract one from a ConstantInt
629 static ConstantInt *SubOne(ConstantInt *C) {
630 APInt Val(C->getValue());
631 return ConstantInt::get(--Val);
633 /// Add - Add two ConstantInts together
634 static ConstantInt *Add(ConstantInt *C1, ConstantInt *C2) {
635 return ConstantInt::get(C1->getValue() + C2->getValue());
637 /// And - Bitwise AND two ConstantInts together
638 static ConstantInt *And(ConstantInt *C1, ConstantInt *C2) {
639 return ConstantInt::get(C1->getValue() & C2->getValue());
641 /// Subtract - Subtract one ConstantInt from another
642 static ConstantInt *Subtract(ConstantInt *C1, ConstantInt *C2) {
643 return ConstantInt::get(C1->getValue() - C2->getValue());
645 /// Multiply - Multiply two ConstantInts together
646 static ConstantInt *Multiply(ConstantInt *C1, ConstantInt *C2) {
647 return ConstantInt::get(C1->getValue() * C2->getValue());
649 /// MultiplyOverflows - True if the multiply can not be expressed in an int
651 static bool MultiplyOverflows(ConstantInt *C1, ConstantInt *C2, bool sign) {
652 uint32_t W = C1->getBitWidth();
653 APInt LHSExt = C1->getValue(), RHSExt = C2->getValue();
662 APInt MulExt = LHSExt * RHSExt;
665 APInt Min = APInt::getSignedMinValue(W).sext(W * 2);
666 APInt Max = APInt::getSignedMaxValue(W).sext(W * 2);
667 return MulExt.slt(Min) || MulExt.sgt(Max);
669 return MulExt.ugt(APInt::getLowBitsSet(W * 2, W));
673 /// ShrinkDemandedConstant - Check to see if the specified operand of the
674 /// specified instruction is a constant integer. If so, check to see if there
675 /// are any bits set in the constant that are not demanded. If so, shrink the
676 /// constant and return true.
677 static bool ShrinkDemandedConstant(Instruction *I, unsigned OpNo,
679 assert(I && "No instruction?");
680 assert(OpNo < I->getNumOperands() && "Operand index too large");
682 // If the operand is not a constant integer, nothing to do.
683 ConstantInt *OpC = dyn_cast<ConstantInt>(I->getOperand(OpNo));
684 if (!OpC) return false;
686 // If there are no bits set that aren't demanded, nothing to do.
687 Demanded.zextOrTrunc(OpC->getValue().getBitWidth());
688 if ((~Demanded & OpC->getValue()) == 0)
691 // This instruction is producing bits that are not demanded. Shrink the RHS.
692 Demanded &= OpC->getValue();
693 I->setOperand(OpNo, ConstantInt::get(Demanded));
697 // ComputeSignedMinMaxValuesFromKnownBits - Given a signed integer type and a
698 // set of known zero and one bits, compute the maximum and minimum values that
699 // could have the specified known zero and known one bits, returning them in
701 static void ComputeSignedMinMaxValuesFromKnownBits(const Type *Ty,
702 const APInt& KnownZero,
703 const APInt& KnownOne,
704 APInt& Min, APInt& Max) {
705 uint32_t BitWidth = cast<IntegerType>(Ty)->getBitWidth();
706 assert(KnownZero.getBitWidth() == BitWidth &&
707 KnownOne.getBitWidth() == BitWidth &&
708 Min.getBitWidth() == BitWidth && Max.getBitWidth() == BitWidth &&
709 "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth.");
710 APInt UnknownBits = ~(KnownZero|KnownOne);
712 // The minimum value is when all unknown bits are zeros, EXCEPT for the sign
713 // bit if it is unknown.
715 Max = KnownOne|UnknownBits;
717 if (UnknownBits[BitWidth-1]) { // Sign bit is unknown
719 Max.clear(BitWidth-1);
723 // ComputeUnsignedMinMaxValuesFromKnownBits - Given an unsigned integer type and
724 // a set of known zero and one bits, compute the maximum and minimum values that
725 // could have the specified known zero and known one bits, returning them in
727 static void ComputeUnsignedMinMaxValuesFromKnownBits(const Type *Ty,
728 const APInt &KnownZero,
729 const APInt &KnownOne,
730 APInt &Min, APInt &Max) {
731 uint32_t BitWidth = cast<IntegerType>(Ty)->getBitWidth(); BitWidth = BitWidth;
732 assert(KnownZero.getBitWidth() == BitWidth &&
733 KnownOne.getBitWidth() == BitWidth &&
734 Min.getBitWidth() == BitWidth && Max.getBitWidth() &&
735 "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth.");
736 APInt UnknownBits = ~(KnownZero|KnownOne);
738 // The minimum value is when the unknown bits are all zeros.
740 // The maximum value is when the unknown bits are all ones.
741 Max = KnownOne|UnknownBits;
744 /// SimplifyDemandedInstructionBits - Inst is an integer instruction that
745 /// SimplifyDemandedBits knows about. See if the instruction has any
746 /// properties that allow us to simplify its operands.
747 bool InstCombiner::SimplifyDemandedInstructionBits(Instruction &Inst) {
748 unsigned BitWidth = cast<IntegerType>(Inst.getType())->getBitWidth();
749 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
750 APInt DemandedMask(APInt::getAllOnesValue(BitWidth));
752 Value *V = SimplifyDemandedUseBits(&Inst, DemandedMask,
753 KnownZero, KnownOne, 0);
754 if (V == 0) return false;
755 if (V == &Inst) return true;
756 ReplaceInstUsesWith(Inst, V);
760 /// SimplifyDemandedBits - This form of SimplifyDemandedBits simplifies the
761 /// specified instruction operand if possible, updating it in place. It returns
762 /// true if it made any change and false otherwise.
763 bool InstCombiner::SimplifyDemandedBits(Use &U, APInt DemandedMask,
764 APInt &KnownZero, APInt &KnownOne,
766 Value *NewVal = SimplifyDemandedUseBits(U.get(), DemandedMask,
767 KnownZero, KnownOne, Depth);
768 if (NewVal == 0) return false;
774 /// SimplifyDemandedUseBits - This function attempts to replace V with a simpler
775 /// value based on the demanded bits. When this function is called, it is known
776 /// that only the bits set in DemandedMask of the result of V are ever used
777 /// downstream. Consequently, depending on the mask and V, it may be possible
778 /// to replace V with a constant or one of its operands. In such cases, this
779 /// function does the replacement and returns true. In all other cases, it
780 /// returns false after analyzing the expression and setting KnownOne and known
781 /// to be one in the expression. KnownZero contains all the bits that are known
782 /// to be zero in the expression. These are provided to potentially allow the
783 /// caller (which might recursively be SimplifyDemandedBits itself) to simplify
784 /// the expression. KnownOne and KnownZero always follow the invariant that
785 /// KnownOne & KnownZero == 0. That is, a bit can't be both 1 and 0. Note that
786 /// the bits in KnownOne and KnownZero may only be accurate for those bits set
787 /// in DemandedMask. Note also that the bitwidth of V, DemandedMask, KnownZero
788 /// and KnownOne must all be the same.
790 /// This returns null if it did not change anything and it permits no
791 /// simplification. This returns V itself if it did some simplification of V's
792 /// operands based on the information about what bits are demanded. This returns
793 /// some other non-null value if it found out that V is equal to another value
794 /// in the context where the specified bits are demanded, but not for all users.
795 Value *InstCombiner::SimplifyDemandedUseBits(Value *V, APInt DemandedMask,
796 APInt &KnownZero, APInt &KnownOne,
798 assert(V != 0 && "Null pointer of Value???");
799 assert(Depth <= 6 && "Limit Search Depth");
800 uint32_t BitWidth = DemandedMask.getBitWidth();
801 const IntegerType *VTy = cast<IntegerType>(V->getType());
802 assert(VTy->getBitWidth() == BitWidth &&
803 KnownZero.getBitWidth() == BitWidth &&
804 KnownOne.getBitWidth() == BitWidth &&
805 "Value *V, DemandedMask, KnownZero and KnownOne \
806 must have same BitWidth");
807 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
808 // We know all of the bits for a constant!
809 KnownOne = CI->getValue() & DemandedMask;
810 KnownZero = ~KnownOne & DemandedMask;
816 if (DemandedMask == 0) { // Not demanding any bits from V.
817 if (isa<UndefValue>(V))
819 return UndefValue::get(VTy);
822 if (Depth == 6) // Limit search depth.
825 Instruction *I = dyn_cast<Instruction>(V);
826 if (!I) return 0; // Only analyze instructions.
828 APInt LHSKnownZero(BitWidth, 0), LHSKnownOne(BitWidth, 0);
829 APInt &RHSKnownZero = KnownZero, &RHSKnownOne = KnownOne;
831 // If there are multiple uses of this value and we aren't at the root, then
832 // we can't do any simplifications of the operands, because DemandedMask
833 // only reflects the bits demanded by *one* of the users.
834 if (Depth != 0 && !I->hasOneUse()) {
835 // Despite the fact that we can't simplify this instruction in all User's
836 // context, we can at least compute the knownzero/knownone bits, and we can
837 // do simplifications that apply to *just* the one user if we know that
838 // this instruction has a simpler value in that context.
839 if (I->getOpcode() == Instruction::And) {
840 // If either the LHS or the RHS are Zero, the result is zero.
841 ComputeMaskedBits(I->getOperand(1), DemandedMask,
842 RHSKnownZero, RHSKnownOne, Depth+1);
843 ComputeMaskedBits(I->getOperand(0), DemandedMask & ~RHSKnownZero,
844 LHSKnownZero, LHSKnownOne, Depth+1);
846 // If all of the demanded bits are known 1 on one side, return the other.
847 // These bits cannot contribute to the result of the 'and' in this
849 if ((DemandedMask & ~LHSKnownZero & RHSKnownOne) ==
850 (DemandedMask & ~LHSKnownZero))
851 return I->getOperand(0);
852 if ((DemandedMask & ~RHSKnownZero & LHSKnownOne) ==
853 (DemandedMask & ~RHSKnownZero))
854 return I->getOperand(1);
856 // If all of the demanded bits in the inputs are known zeros, return zero.
857 if ((DemandedMask & (RHSKnownZero|LHSKnownZero)) == DemandedMask)
858 return Constant::getNullValue(VTy);
860 } else if (I->getOpcode() == Instruction::Or) {
861 // We can simplify (X|Y) -> X or Y in the user's context if we know that
862 // only bits from X or Y are demanded.
864 // If either the LHS or the RHS are One, the result is One.
865 ComputeMaskedBits(I->getOperand(1), DemandedMask,
866 RHSKnownZero, RHSKnownOne, Depth+1);
867 ComputeMaskedBits(I->getOperand(0), DemandedMask & ~RHSKnownOne,
868 LHSKnownZero, LHSKnownOne, Depth+1);
870 // If all of the demanded bits are known zero on one side, return the
871 // other. These bits cannot contribute to the result of the 'or' in this
873 if ((DemandedMask & ~LHSKnownOne & RHSKnownZero) ==
874 (DemandedMask & ~LHSKnownOne))
875 return I->getOperand(0);
876 if ((DemandedMask & ~RHSKnownOne & LHSKnownZero) ==
877 (DemandedMask & ~RHSKnownOne))
878 return I->getOperand(1);
880 // If all of the potentially set bits on one side are known to be set on
881 // the other side, just use the 'other' side.
882 if ((DemandedMask & (~RHSKnownZero) & LHSKnownOne) ==
883 (DemandedMask & (~RHSKnownZero)))
884 return I->getOperand(0);
885 if ((DemandedMask & (~LHSKnownZero) & RHSKnownOne) ==
886 (DemandedMask & (~LHSKnownZero)))
887 return I->getOperand(1);
890 // Compute the KnownZero/KnownOne bits to simplify things downstream.
891 ComputeMaskedBits(I, DemandedMask, KnownZero, KnownOne, Depth);
895 // If this is the root being simplified, allow it to have multiple uses,
896 // just set the DemandedMask to all bits so that we can try to simplify the
897 // operands. This allows visitTruncInst (for example) to simplify the
898 // operand of a trunc without duplicating all the logic below.
899 if (Depth == 0 && !V->hasOneUse())
900 DemandedMask = APInt::getAllOnesValue(BitWidth);
902 switch (I->getOpcode()) {
904 ComputeMaskedBits(I, DemandedMask, RHSKnownZero, RHSKnownOne, Depth);
906 case Instruction::And:
907 // If either the LHS or the RHS are Zero, the result is zero.
908 if (SimplifyDemandedBits(I->getOperandUse(1), DemandedMask,
909 RHSKnownZero, RHSKnownOne, Depth+1) ||
910 SimplifyDemandedBits(I->getOperandUse(0), DemandedMask & ~RHSKnownZero,
911 LHSKnownZero, LHSKnownOne, Depth+1))
913 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
914 assert(!(LHSKnownZero & LHSKnownOne) && "Bits known to be one AND zero?");
916 // If all of the demanded bits are known 1 on one side, return the other.
917 // These bits cannot contribute to the result of the 'and'.
918 if ((DemandedMask & ~LHSKnownZero & RHSKnownOne) ==
919 (DemandedMask & ~LHSKnownZero))
920 return I->getOperand(0);
921 if ((DemandedMask & ~RHSKnownZero & LHSKnownOne) ==
922 (DemandedMask & ~RHSKnownZero))
923 return I->getOperand(1);
925 // If all of the demanded bits in the inputs are known zeros, return zero.
926 if ((DemandedMask & (RHSKnownZero|LHSKnownZero)) == DemandedMask)
927 return Constant::getNullValue(VTy);
929 // If the RHS is a constant, see if we can simplify it.
930 if (ShrinkDemandedConstant(I, 1, DemandedMask & ~LHSKnownZero))
933 // Output known-1 bits are only known if set in both the LHS & RHS.
934 RHSKnownOne &= LHSKnownOne;
935 // Output known-0 are known to be clear if zero in either the LHS | RHS.
936 RHSKnownZero |= LHSKnownZero;
938 case Instruction::Or:
939 // If either the LHS or the RHS are One, the result is One.
940 if (SimplifyDemandedBits(I->getOperandUse(1), DemandedMask,
941 RHSKnownZero, RHSKnownOne, Depth+1) ||
942 SimplifyDemandedBits(I->getOperandUse(0), DemandedMask & ~RHSKnownOne,
943 LHSKnownZero, LHSKnownOne, Depth+1))
945 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
946 assert(!(LHSKnownZero & LHSKnownOne) && "Bits known to be one AND zero?");
948 // If all of the demanded bits are known zero on one side, return the other.
949 // These bits cannot contribute to the result of the 'or'.
950 if ((DemandedMask & ~LHSKnownOne & RHSKnownZero) ==
951 (DemandedMask & ~LHSKnownOne))
952 return I->getOperand(0);
953 if ((DemandedMask & ~RHSKnownOne & LHSKnownZero) ==
954 (DemandedMask & ~RHSKnownOne))
955 return I->getOperand(1);
957 // If all of the potentially set bits on one side are known to be set on
958 // the other side, just use the 'other' side.
959 if ((DemandedMask & (~RHSKnownZero) & LHSKnownOne) ==
960 (DemandedMask & (~RHSKnownZero)))
961 return I->getOperand(0);
962 if ((DemandedMask & (~LHSKnownZero) & RHSKnownOne) ==
963 (DemandedMask & (~LHSKnownZero)))
964 return I->getOperand(1);
966 // If the RHS is a constant, see if we can simplify it.
967 if (ShrinkDemandedConstant(I, 1, DemandedMask))
970 // Output known-0 bits are only known if clear in both the LHS & RHS.
971 RHSKnownZero &= LHSKnownZero;
972 // Output known-1 are known to be set if set in either the LHS | RHS.
973 RHSKnownOne |= LHSKnownOne;
975 case Instruction::Xor: {
976 if (SimplifyDemandedBits(I->getOperandUse(1), DemandedMask,
977 RHSKnownZero, RHSKnownOne, Depth+1) ||
978 SimplifyDemandedBits(I->getOperandUse(0), DemandedMask,
979 LHSKnownZero, LHSKnownOne, Depth+1))
981 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
982 assert(!(LHSKnownZero & LHSKnownOne) && "Bits known to be one AND zero?");
984 // If all of the demanded bits are known zero on one side, return the other.
985 // These bits cannot contribute to the result of the 'xor'.
986 if ((DemandedMask & RHSKnownZero) == DemandedMask)
987 return I->getOperand(0);
988 if ((DemandedMask & LHSKnownZero) == DemandedMask)
989 return I->getOperand(1);
991 // Output known-0 bits are known if clear or set in both the LHS & RHS.
992 APInt KnownZeroOut = (RHSKnownZero & LHSKnownZero) |
993 (RHSKnownOne & LHSKnownOne);
994 // Output known-1 are known to be set if set in only one of the LHS, RHS.
995 APInt KnownOneOut = (RHSKnownZero & LHSKnownOne) |
996 (RHSKnownOne & LHSKnownZero);
998 // If all of the demanded bits are known to be zero on one side or the
999 // other, turn this into an *inclusive* or.
1000 // e.g. (A & C1)^(B & C2) -> (A & C1)|(B & C2) iff C1&C2 == 0
1001 if ((DemandedMask & ~RHSKnownZero & ~LHSKnownZero) == 0) {
1003 BinaryOperator::CreateOr(I->getOperand(0), I->getOperand(1),
1005 return InsertNewInstBefore(Or, *I);
1008 // If all of the demanded bits on one side are known, and all of the set
1009 // bits on that side are also known to be set on the other side, turn this
1010 // into an AND, as we know the bits will be cleared.
1011 // e.g. (X | C1) ^ C2 --> (X | C1) & ~C2 iff (C1&C2) == C2
1012 if ((DemandedMask & (RHSKnownZero|RHSKnownOne)) == DemandedMask) {
1014 if ((RHSKnownOne & LHSKnownOne) == RHSKnownOne) {
1015 Constant *AndC = ConstantInt::get(~RHSKnownOne & DemandedMask);
1017 BinaryOperator::CreateAnd(I->getOperand(0), AndC, "tmp");
1018 return InsertNewInstBefore(And, *I);
1022 // If the RHS is a constant, see if we can simplify it.
1023 // FIXME: for XOR, we prefer to force bits to 1 if they will make a -1.
1024 if (ShrinkDemandedConstant(I, 1, DemandedMask))
1027 RHSKnownZero = KnownZeroOut;
1028 RHSKnownOne = KnownOneOut;
1031 case Instruction::Select:
1032 if (SimplifyDemandedBits(I->getOperandUse(2), DemandedMask,
1033 RHSKnownZero, RHSKnownOne, Depth+1) ||
1034 SimplifyDemandedBits(I->getOperandUse(1), DemandedMask,
1035 LHSKnownZero, LHSKnownOne, Depth+1))
1037 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1038 assert(!(LHSKnownZero & LHSKnownOne) && "Bits known to be one AND zero?");
1040 // If the operands are constants, see if we can simplify them.
1041 if (ShrinkDemandedConstant(I, 1, DemandedMask) ||
1042 ShrinkDemandedConstant(I, 2, DemandedMask))
1045 // Only known if known in both the LHS and RHS.
1046 RHSKnownOne &= LHSKnownOne;
1047 RHSKnownZero &= LHSKnownZero;
1049 case Instruction::Trunc: {
1050 unsigned truncBf = I->getOperand(0)->getType()->getPrimitiveSizeInBits();
1051 DemandedMask.zext(truncBf);
1052 RHSKnownZero.zext(truncBf);
1053 RHSKnownOne.zext(truncBf);
1054 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedMask,
1055 RHSKnownZero, RHSKnownOne, Depth+1))
1057 DemandedMask.trunc(BitWidth);
1058 RHSKnownZero.trunc(BitWidth);
1059 RHSKnownOne.trunc(BitWidth);
1060 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1063 case Instruction::BitCast:
1064 if (!I->getOperand(0)->getType()->isInteger())
1065 return false; // vector->int or fp->int?
1066 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedMask,
1067 RHSKnownZero, RHSKnownOne, Depth+1))
1069 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1071 case Instruction::ZExt: {
1072 // Compute the bits in the result that are not present in the input.
1073 unsigned SrcBitWidth =I->getOperand(0)->getType()->getPrimitiveSizeInBits();
1075 DemandedMask.trunc(SrcBitWidth);
1076 RHSKnownZero.trunc(SrcBitWidth);
1077 RHSKnownOne.trunc(SrcBitWidth);
1078 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedMask,
1079 RHSKnownZero, RHSKnownOne, Depth+1))
1081 DemandedMask.zext(BitWidth);
1082 RHSKnownZero.zext(BitWidth);
1083 RHSKnownOne.zext(BitWidth);
1084 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1085 // The top bits are known to be zero.
1086 RHSKnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
1089 case Instruction::SExt: {
1090 // Compute the bits in the result that are not present in the input.
1091 unsigned SrcBitWidth =I->getOperand(0)->getType()->getPrimitiveSizeInBits();
1093 APInt InputDemandedBits = DemandedMask &
1094 APInt::getLowBitsSet(BitWidth, SrcBitWidth);
1096 APInt NewBits(APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth));
1097 // If any of the sign extended bits are demanded, we know that the sign
1099 if ((NewBits & DemandedMask) != 0)
1100 InputDemandedBits.set(SrcBitWidth-1);
1102 InputDemandedBits.trunc(SrcBitWidth);
1103 RHSKnownZero.trunc(SrcBitWidth);
1104 RHSKnownOne.trunc(SrcBitWidth);
1105 if (SimplifyDemandedBits(I->getOperandUse(0), InputDemandedBits,
1106 RHSKnownZero, RHSKnownOne, Depth+1))
1108 InputDemandedBits.zext(BitWidth);
1109 RHSKnownZero.zext(BitWidth);
1110 RHSKnownOne.zext(BitWidth);
1111 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1113 // If the sign bit of the input is known set or clear, then we know the
1114 // top bits of the result.
1116 // If the input sign bit is known zero, or if the NewBits are not demanded
1117 // convert this into a zero extension.
1118 if (RHSKnownZero[SrcBitWidth-1] || (NewBits & ~DemandedMask) == NewBits) {
1119 // Convert to ZExt cast
1120 CastInst *NewCast = new ZExtInst(I->getOperand(0), VTy, I->getName());
1121 return InsertNewInstBefore(NewCast, *I);
1122 } else if (RHSKnownOne[SrcBitWidth-1]) { // Input sign bit known set
1123 RHSKnownOne |= NewBits;
1127 case Instruction::Add: {
1128 // Figure out what the input bits are. If the top bits of the and result
1129 // are not demanded, then the add doesn't demand them from its input
1131 unsigned NLZ = DemandedMask.countLeadingZeros();
1133 // If there is a constant on the RHS, there are a variety of xformations
1135 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
1136 // If null, this should be simplified elsewhere. Some of the xforms here
1137 // won't work if the RHS is zero.
1141 // If the top bit of the output is demanded, demand everything from the
1142 // input. Otherwise, we demand all the input bits except NLZ top bits.
1143 APInt InDemandedBits(APInt::getLowBitsSet(BitWidth, BitWidth - NLZ));
1145 // Find information about known zero/one bits in the input.
1146 if (SimplifyDemandedBits(I->getOperandUse(0), InDemandedBits,
1147 LHSKnownZero, LHSKnownOne, Depth+1))
1150 // If the RHS of the add has bits set that can't affect the input, reduce
1152 if (ShrinkDemandedConstant(I, 1, InDemandedBits))
1155 // Avoid excess work.
1156 if (LHSKnownZero == 0 && LHSKnownOne == 0)
1159 // Turn it into OR if input bits are zero.
1160 if ((LHSKnownZero & RHS->getValue()) == RHS->getValue()) {
1162 BinaryOperator::CreateOr(I->getOperand(0), I->getOperand(1),
1164 return InsertNewInstBefore(Or, *I);
1167 // We can say something about the output known-zero and known-one bits,
1168 // depending on potential carries from the input constant and the
1169 // unknowns. For example if the LHS is known to have at most the 0x0F0F0
1170 // bits set and the RHS constant is 0x01001, then we know we have a known
1171 // one mask of 0x00001 and a known zero mask of 0xE0F0E.
1173 // To compute this, we first compute the potential carry bits. These are
1174 // the bits which may be modified. I'm not aware of a better way to do
1176 const APInt &RHSVal = RHS->getValue();
1177 APInt CarryBits((~LHSKnownZero + RHSVal) ^ (~LHSKnownZero ^ RHSVal));
1179 // Now that we know which bits have carries, compute the known-1/0 sets.
1181 // Bits are known one if they are known zero in one operand and one in the
1182 // other, and there is no input carry.
1183 RHSKnownOne = ((LHSKnownZero & RHSVal) |
1184 (LHSKnownOne & ~RHSVal)) & ~CarryBits;
1186 // Bits are known zero if they are known zero in both operands and there
1187 // is no input carry.
1188 RHSKnownZero = LHSKnownZero & ~RHSVal & ~CarryBits;
1190 // If the high-bits of this ADD are not demanded, then it does not demand
1191 // the high bits of its LHS or RHS.
1192 if (DemandedMask[BitWidth-1] == 0) {
1193 // Right fill the mask of bits for this ADD to demand the most
1194 // significant bit and all those below it.
1195 APInt DemandedFromOps(APInt::getLowBitsSet(BitWidth, BitWidth-NLZ));
1196 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedFromOps,
1197 LHSKnownZero, LHSKnownOne, Depth+1) ||
1198 SimplifyDemandedBits(I->getOperandUse(1), DemandedFromOps,
1199 LHSKnownZero, LHSKnownOne, Depth+1))
1205 case Instruction::Sub:
1206 // If the high-bits of this SUB are not demanded, then it does not demand
1207 // the high bits of its LHS or RHS.
1208 if (DemandedMask[BitWidth-1] == 0) {
1209 // Right fill the mask of bits for this SUB to demand the most
1210 // significant bit and all those below it.
1211 uint32_t NLZ = DemandedMask.countLeadingZeros();
1212 APInt DemandedFromOps(APInt::getLowBitsSet(BitWidth, BitWidth-NLZ));
1213 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedFromOps,
1214 LHSKnownZero, LHSKnownOne, Depth+1) ||
1215 SimplifyDemandedBits(I->getOperandUse(1), DemandedFromOps,
1216 LHSKnownZero, LHSKnownOne, Depth+1))
1219 // Otherwise just hand the sub off to ComputeMaskedBits to fill in
1220 // the known zeros and ones.
1221 ComputeMaskedBits(V, DemandedMask, RHSKnownZero, RHSKnownOne, Depth);
1223 case Instruction::Shl:
1224 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1225 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
1226 APInt DemandedMaskIn(DemandedMask.lshr(ShiftAmt));
1227 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedMaskIn,
1228 RHSKnownZero, RHSKnownOne, Depth+1))
1230 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1231 RHSKnownZero <<= ShiftAmt;
1232 RHSKnownOne <<= ShiftAmt;
1233 // low bits known zero.
1235 RHSKnownZero |= APInt::getLowBitsSet(BitWidth, ShiftAmt);
1238 case Instruction::LShr:
1239 // For a logical shift right
1240 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1241 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
1243 // Unsigned shift right.
1244 APInt DemandedMaskIn(DemandedMask.shl(ShiftAmt));
1245 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedMaskIn,
1246 RHSKnownZero, RHSKnownOne, Depth+1))
1248 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1249 RHSKnownZero = APIntOps::lshr(RHSKnownZero, ShiftAmt);
1250 RHSKnownOne = APIntOps::lshr(RHSKnownOne, ShiftAmt);
1252 // Compute the new bits that are at the top now.
1253 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
1254 RHSKnownZero |= HighBits; // high bits known zero.
1258 case Instruction::AShr:
1259 // If this is an arithmetic shift right and only the low-bit is set, we can
1260 // always convert this into a logical shr, even if the shift amount is
1261 // variable. The low bit of the shift cannot be an input sign bit unless
1262 // the shift amount is >= the size of the datatype, which is undefined.
1263 if (DemandedMask == 1) {
1264 // Perform the logical shift right.
1265 Instruction *NewVal = BinaryOperator::CreateLShr(
1266 I->getOperand(0), I->getOperand(1), I->getName());
1267 return InsertNewInstBefore(NewVal, *I);
1270 // If the sign bit is the only bit demanded by this ashr, then there is no
1271 // need to do it, the shift doesn't change the high bit.
1272 if (DemandedMask.isSignBit())
1273 return I->getOperand(0);
1275 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1276 uint32_t ShiftAmt = SA->getLimitedValue(BitWidth);
1278 // Signed shift right.
1279 APInt DemandedMaskIn(DemandedMask.shl(ShiftAmt));
1280 // If any of the "high bits" are demanded, we should set the sign bit as
1282 if (DemandedMask.countLeadingZeros() <= ShiftAmt)
1283 DemandedMaskIn.set(BitWidth-1);
1284 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedMaskIn,
1285 RHSKnownZero, RHSKnownOne, Depth+1))
1287 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1288 // Compute the new bits that are at the top now.
1289 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
1290 RHSKnownZero = APIntOps::lshr(RHSKnownZero, ShiftAmt);
1291 RHSKnownOne = APIntOps::lshr(RHSKnownOne, ShiftAmt);
1293 // Handle the sign bits.
1294 APInt SignBit(APInt::getSignBit(BitWidth));
1295 // Adjust to where it is now in the mask.
1296 SignBit = APIntOps::lshr(SignBit, ShiftAmt);
1298 // If the input sign bit is known to be zero, or if none of the top bits
1299 // are demanded, turn this into an unsigned shift right.
1300 if (BitWidth <= ShiftAmt || RHSKnownZero[BitWidth-ShiftAmt-1] ||
1301 (HighBits & ~DemandedMask) == HighBits) {
1302 // Perform the logical shift right.
1303 Instruction *NewVal = BinaryOperator::CreateLShr(
1304 I->getOperand(0), SA, I->getName());
1305 return InsertNewInstBefore(NewVal, *I);
1306 } else if ((RHSKnownOne & SignBit) != 0) { // New bits are known one.
1307 RHSKnownOne |= HighBits;
1311 case Instruction::SRem:
1312 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
1313 APInt RA = Rem->getValue().abs();
1314 if (RA.isPowerOf2()) {
1315 if (DemandedMask.ule(RA)) // srem won't affect demanded bits
1316 return I->getOperand(0);
1318 APInt LowBits = RA - 1;
1319 APInt Mask2 = LowBits | APInt::getSignBit(BitWidth);
1320 if (SimplifyDemandedBits(I->getOperandUse(0), Mask2,
1321 LHSKnownZero, LHSKnownOne, Depth+1))
1324 if (LHSKnownZero[BitWidth-1] || ((LHSKnownZero & LowBits) == LowBits))
1325 LHSKnownZero |= ~LowBits;
1327 KnownZero |= LHSKnownZero & DemandedMask;
1329 assert(!(KnownZero & KnownOne) && "Bits known to be one AND zero?");
1333 case Instruction::URem: {
1334 APInt KnownZero2(BitWidth, 0), KnownOne2(BitWidth, 0);
1335 APInt AllOnes = APInt::getAllOnesValue(BitWidth);
1336 if (SimplifyDemandedBits(I->getOperandUse(0), AllOnes,
1337 KnownZero2, KnownOne2, Depth+1) ||
1338 SimplifyDemandedBits(I->getOperandUse(1), AllOnes,
1339 KnownZero2, KnownOne2, Depth+1))
1342 unsigned Leaders = KnownZero2.countLeadingOnes();
1343 Leaders = std::max(Leaders,
1344 KnownZero2.countLeadingOnes());
1345 KnownZero = APInt::getHighBitsSet(BitWidth, Leaders) & DemandedMask;
1348 case Instruction::Call:
1349 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
1350 switch (II->getIntrinsicID()) {
1352 case Intrinsic::bswap: {
1353 // If the only bits demanded come from one byte of the bswap result,
1354 // just shift the input byte into position to eliminate the bswap.
1355 unsigned NLZ = DemandedMask.countLeadingZeros();
1356 unsigned NTZ = DemandedMask.countTrailingZeros();
1358 // Round NTZ down to the next byte. If we have 11 trailing zeros, then
1359 // we need all the bits down to bit 8. Likewise, round NLZ. If we
1360 // have 14 leading zeros, round to 8.
1363 // If we need exactly one byte, we can do this transformation.
1364 if (BitWidth-NLZ-NTZ == 8) {
1365 unsigned ResultBit = NTZ;
1366 unsigned InputBit = BitWidth-NTZ-8;
1368 // Replace this with either a left or right shift to get the byte into
1370 Instruction *NewVal;
1371 if (InputBit > ResultBit)
1372 NewVal = BinaryOperator::CreateLShr(I->getOperand(1),
1373 ConstantInt::get(I->getType(), InputBit-ResultBit));
1375 NewVal = BinaryOperator::CreateShl(I->getOperand(1),
1376 ConstantInt::get(I->getType(), ResultBit-InputBit));
1377 NewVal->takeName(I);
1378 return InsertNewInstBefore(NewVal, *I);
1381 // TODO: Could compute known zero/one bits based on the input.
1386 ComputeMaskedBits(V, DemandedMask, RHSKnownZero, RHSKnownOne, Depth);
1390 // If the client is only demanding bits that we know, return the known
1392 if ((DemandedMask & (RHSKnownZero|RHSKnownOne)) == DemandedMask)
1393 return ConstantInt::get(RHSKnownOne);
1398 /// SimplifyDemandedVectorElts - The specified value produces a vector with
1399 /// 64 or fewer elements. DemandedElts contains the set of elements that are
1400 /// actually used by the caller. This method analyzes which elements of the
1401 /// operand are undef and returns that information in UndefElts.
1403 /// If the information about demanded elements can be used to simplify the
1404 /// operation, the operation is simplified, then the resultant value is
1405 /// returned. This returns null if no change was made.
1406 Value *InstCombiner::SimplifyDemandedVectorElts(Value *V, uint64_t DemandedElts,
1407 uint64_t &UndefElts,
1409 unsigned VWidth = cast<VectorType>(V->getType())->getNumElements();
1410 assert(VWidth <= 64 && "Vector too wide to analyze!");
1411 uint64_t EltMask = ~0ULL >> (64-VWidth);
1412 assert((DemandedElts & ~EltMask) == 0 && "Invalid DemandedElts!");
1414 if (isa<UndefValue>(V)) {
1415 // If the entire vector is undefined, just return this info.
1416 UndefElts = EltMask;
1418 } else if (DemandedElts == 0) { // If nothing is demanded, provide undef.
1419 UndefElts = EltMask;
1420 return UndefValue::get(V->getType());
1424 if (ConstantVector *CP = dyn_cast<ConstantVector>(V)) {
1425 const Type *EltTy = cast<VectorType>(V->getType())->getElementType();
1426 Constant *Undef = UndefValue::get(EltTy);
1428 std::vector<Constant*> Elts;
1429 for (unsigned i = 0; i != VWidth; ++i)
1430 if (!(DemandedElts & (1ULL << i))) { // If not demanded, set to undef.
1431 Elts.push_back(Undef);
1432 UndefElts |= (1ULL << i);
1433 } else if (isa<UndefValue>(CP->getOperand(i))) { // Already undef.
1434 Elts.push_back(Undef);
1435 UndefElts |= (1ULL << i);
1436 } else { // Otherwise, defined.
1437 Elts.push_back(CP->getOperand(i));
1440 // If we changed the constant, return it.
1441 Constant *NewCP = ConstantVector::get(Elts);
1442 return NewCP != CP ? NewCP : 0;
1443 } else if (isa<ConstantAggregateZero>(V)) {
1444 // Simplify the CAZ to a ConstantVector where the non-demanded elements are
1447 // Check if this is identity. If so, return 0 since we are not simplifying
1449 if (DemandedElts == ((1ULL << VWidth) -1))
1452 const Type *EltTy = cast<VectorType>(V->getType())->getElementType();
1453 Constant *Zero = Constant::getNullValue(EltTy);
1454 Constant *Undef = UndefValue::get(EltTy);
1455 std::vector<Constant*> Elts;
1456 for (unsigned i = 0; i != VWidth; ++i)
1457 Elts.push_back((DemandedElts & (1ULL << i)) ? Zero : Undef);
1458 UndefElts = DemandedElts ^ EltMask;
1459 return ConstantVector::get(Elts);
1462 // Limit search depth.
1466 // If multiple users are using the root value, procede with
1467 // simplification conservatively assuming that all elements
1469 if (!V->hasOneUse()) {
1470 // Quit if we find multiple users of a non-root value though.
1471 // They'll be handled when it's their turn to be visited by
1472 // the main instcombine process.
1474 // TODO: Just compute the UndefElts information recursively.
1477 // Conservatively assume that all elements are needed.
1478 DemandedElts = EltMask;
1481 Instruction *I = dyn_cast<Instruction>(V);
1482 if (!I) return false; // Only analyze instructions.
1484 bool MadeChange = false;
1485 uint64_t UndefElts2;
1487 switch (I->getOpcode()) {
1490 case Instruction::InsertElement: {
1491 // If this is a variable index, we don't know which element it overwrites.
1492 // demand exactly the same input as we produce.
1493 ConstantInt *Idx = dyn_cast<ConstantInt>(I->getOperand(2));
1495 // Note that we can't propagate undef elt info, because we don't know
1496 // which elt is getting updated.
1497 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
1498 UndefElts2, Depth+1);
1499 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1503 // If this is inserting an element that isn't demanded, remove this
1505 unsigned IdxNo = Idx->getZExtValue();
1506 if (IdxNo >= VWidth || (DemandedElts & (1ULL << IdxNo)) == 0)
1507 return AddSoonDeadInstToWorklist(*I, 0);
1509 // Otherwise, the element inserted overwrites whatever was there, so the
1510 // input demanded set is simpler than the output set.
1511 TmpV = SimplifyDemandedVectorElts(I->getOperand(0),
1512 DemandedElts & ~(1ULL << IdxNo),
1513 UndefElts, Depth+1);
1514 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1516 // The inserted element is defined.
1517 UndefElts &= ~(1ULL << IdxNo);
1520 case Instruction::ShuffleVector: {
1521 ShuffleVectorInst *Shuffle = cast<ShuffleVectorInst>(I);
1522 uint64_t LHSVWidth =
1523 cast<VectorType>(Shuffle->getOperand(0)->getType())->getNumElements();
1524 uint64_t LeftDemanded = 0, RightDemanded = 0;
1525 for (unsigned i = 0; i < VWidth; i++) {
1526 if (DemandedElts & (1ULL << i)) {
1527 unsigned MaskVal = Shuffle->getMaskValue(i);
1528 if (MaskVal != -1u) {
1529 assert(MaskVal < LHSVWidth * 2 &&
1530 "shufflevector mask index out of range!");
1531 if (MaskVal < LHSVWidth)
1532 LeftDemanded |= 1ULL << MaskVal;
1534 RightDemanded |= 1ULL << (MaskVal - LHSVWidth);
1539 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), LeftDemanded,
1540 UndefElts2, Depth+1);
1541 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1543 uint64_t UndefElts3;
1544 TmpV = SimplifyDemandedVectorElts(I->getOperand(1), RightDemanded,
1545 UndefElts3, Depth+1);
1546 if (TmpV) { I->setOperand(1, TmpV); MadeChange = true; }
1548 bool NewUndefElts = false;
1549 for (unsigned i = 0; i < VWidth; i++) {
1550 unsigned MaskVal = Shuffle->getMaskValue(i);
1551 if (MaskVal == -1u) {
1552 uint64_t NewBit = 1ULL << i;
1553 UndefElts |= NewBit;
1554 } else if (MaskVal < LHSVWidth) {
1555 uint64_t NewBit = ((UndefElts2 >> MaskVal) & 1) << i;
1556 NewUndefElts |= NewBit;
1557 UndefElts |= NewBit;
1559 uint64_t NewBit = ((UndefElts3 >> (MaskVal - LHSVWidth)) & 1) << i;
1560 NewUndefElts |= NewBit;
1561 UndefElts |= NewBit;
1566 // Add additional discovered undefs.
1567 std::vector<Constant*> Elts;
1568 for (unsigned i = 0; i < VWidth; ++i) {
1569 if (UndefElts & (1ULL << i))
1570 Elts.push_back(UndefValue::get(Type::Int32Ty));
1572 Elts.push_back(ConstantInt::get(Type::Int32Ty,
1573 Shuffle->getMaskValue(i)));
1575 I->setOperand(2, ConstantVector::get(Elts));
1580 case Instruction::BitCast: {
1581 // Vector->vector casts only.
1582 const VectorType *VTy = dyn_cast<VectorType>(I->getOperand(0)->getType());
1584 unsigned InVWidth = VTy->getNumElements();
1585 uint64_t InputDemandedElts = 0;
1588 if (VWidth == InVWidth) {
1589 // If we are converting from <4 x i32> -> <4 x f32>, we demand the same
1590 // elements as are demanded of us.
1592 InputDemandedElts = DemandedElts;
1593 } else if (VWidth > InVWidth) {
1597 // If there are more elements in the result than there are in the source,
1598 // then an input element is live if any of the corresponding output
1599 // elements are live.
1600 Ratio = VWidth/InVWidth;
1601 for (unsigned OutIdx = 0; OutIdx != VWidth; ++OutIdx) {
1602 if (DemandedElts & (1ULL << OutIdx))
1603 InputDemandedElts |= 1ULL << (OutIdx/Ratio);
1609 // If there are more elements in the source than there are in the result,
1610 // then an input element is live if the corresponding output element is
1612 Ratio = InVWidth/VWidth;
1613 for (unsigned InIdx = 0; InIdx != InVWidth; ++InIdx)
1614 if (DemandedElts & (1ULL << InIdx/Ratio))
1615 InputDemandedElts |= 1ULL << InIdx;
1618 // div/rem demand all inputs, because they don't want divide by zero.
1619 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), InputDemandedElts,
1620 UndefElts2, Depth+1);
1622 I->setOperand(0, TmpV);
1626 UndefElts = UndefElts2;
1627 if (VWidth > InVWidth) {
1628 assert(0 && "Unimp");
1629 // If there are more elements in the result than there are in the source,
1630 // then an output element is undef if the corresponding input element is
1632 for (unsigned OutIdx = 0; OutIdx != VWidth; ++OutIdx)
1633 if (UndefElts2 & (1ULL << (OutIdx/Ratio)))
1634 UndefElts |= 1ULL << OutIdx;
1635 } else if (VWidth < InVWidth) {
1636 assert(0 && "Unimp");
1637 // If there are more elements in the source than there are in the result,
1638 // then a result element is undef if all of the corresponding input
1639 // elements are undef.
1640 UndefElts = ~0ULL >> (64-VWidth); // Start out all undef.
1641 for (unsigned InIdx = 0; InIdx != InVWidth; ++InIdx)
1642 if ((UndefElts2 & (1ULL << InIdx)) == 0) // Not undef?
1643 UndefElts &= ~(1ULL << (InIdx/Ratio)); // Clear undef bit.
1647 case Instruction::And:
1648 case Instruction::Or:
1649 case Instruction::Xor:
1650 case Instruction::Add:
1651 case Instruction::Sub:
1652 case Instruction::Mul:
1653 // div/rem demand all inputs, because they don't want divide by zero.
1654 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
1655 UndefElts, Depth+1);
1656 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1657 TmpV = SimplifyDemandedVectorElts(I->getOperand(1), DemandedElts,
1658 UndefElts2, Depth+1);
1659 if (TmpV) { I->setOperand(1, TmpV); MadeChange = true; }
1661 // Output elements are undefined if both are undefined. Consider things
1662 // like undef&0. The result is known zero, not undef.
1663 UndefElts &= UndefElts2;
1666 case Instruction::Call: {
1667 IntrinsicInst *II = dyn_cast<IntrinsicInst>(I);
1669 switch (II->getIntrinsicID()) {
1672 // Binary vector operations that work column-wise. A dest element is a
1673 // function of the corresponding input elements from the two inputs.
1674 case Intrinsic::x86_sse_sub_ss:
1675 case Intrinsic::x86_sse_mul_ss:
1676 case Intrinsic::x86_sse_min_ss:
1677 case Intrinsic::x86_sse_max_ss:
1678 case Intrinsic::x86_sse2_sub_sd:
1679 case Intrinsic::x86_sse2_mul_sd:
1680 case Intrinsic::x86_sse2_min_sd:
1681 case Intrinsic::x86_sse2_max_sd:
1682 TmpV = SimplifyDemandedVectorElts(II->getOperand(1), DemandedElts,
1683 UndefElts, Depth+1);
1684 if (TmpV) { II->setOperand(1, TmpV); MadeChange = true; }
1685 TmpV = SimplifyDemandedVectorElts(II->getOperand(2), DemandedElts,
1686 UndefElts2, Depth+1);
1687 if (TmpV) { II->setOperand(2, TmpV); MadeChange = true; }
1689 // If only the low elt is demanded and this is a scalarizable intrinsic,
1690 // scalarize it now.
1691 if (DemandedElts == 1) {
1692 switch (II->getIntrinsicID()) {
1694 case Intrinsic::x86_sse_sub_ss:
1695 case Intrinsic::x86_sse_mul_ss:
1696 case Intrinsic::x86_sse2_sub_sd:
1697 case Intrinsic::x86_sse2_mul_sd:
1698 // TODO: Lower MIN/MAX/ABS/etc
1699 Value *LHS = II->getOperand(1);
1700 Value *RHS = II->getOperand(2);
1701 // Extract the element as scalars.
1702 LHS = InsertNewInstBefore(new ExtractElementInst(LHS, 0U,"tmp"), *II);
1703 RHS = InsertNewInstBefore(new ExtractElementInst(RHS, 0U,"tmp"), *II);
1705 switch (II->getIntrinsicID()) {
1706 default: assert(0 && "Case stmts out of sync!");
1707 case Intrinsic::x86_sse_sub_ss:
1708 case Intrinsic::x86_sse2_sub_sd:
1709 TmpV = InsertNewInstBefore(BinaryOperator::CreateSub(LHS, RHS,
1710 II->getName()), *II);
1712 case Intrinsic::x86_sse_mul_ss:
1713 case Intrinsic::x86_sse2_mul_sd:
1714 TmpV = InsertNewInstBefore(BinaryOperator::CreateMul(LHS, RHS,
1715 II->getName()), *II);
1720 InsertElementInst::Create(UndefValue::get(II->getType()), TmpV, 0U,
1722 InsertNewInstBefore(New, *II);
1723 AddSoonDeadInstToWorklist(*II, 0);
1728 // Output elements are undefined if both are undefined. Consider things
1729 // like undef&0. The result is known zero, not undef.
1730 UndefElts &= UndefElts2;
1736 return MadeChange ? I : 0;
1740 /// AssociativeOpt - Perform an optimization on an associative operator. This
1741 /// function is designed to check a chain of associative operators for a
1742 /// potential to apply a certain optimization. Since the optimization may be
1743 /// applicable if the expression was reassociated, this checks the chain, then
1744 /// reassociates the expression as necessary to expose the optimization
1745 /// opportunity. This makes use of a special Functor, which must define
1746 /// 'shouldApply' and 'apply' methods.
1748 template<typename Functor>
1749 static Instruction *AssociativeOpt(BinaryOperator &Root, const Functor &F) {
1750 unsigned Opcode = Root.getOpcode();
1751 Value *LHS = Root.getOperand(0);
1753 // Quick check, see if the immediate LHS matches...
1754 if (F.shouldApply(LHS))
1755 return F.apply(Root);
1757 // Otherwise, if the LHS is not of the same opcode as the root, return.
1758 Instruction *LHSI = dyn_cast<Instruction>(LHS);
1759 while (LHSI && LHSI->getOpcode() == Opcode && LHSI->hasOneUse()) {
1760 // Should we apply this transform to the RHS?
1761 bool ShouldApply = F.shouldApply(LHSI->getOperand(1));
1763 // If not to the RHS, check to see if we should apply to the LHS...
1764 if (!ShouldApply && F.shouldApply(LHSI->getOperand(0))) {
1765 cast<BinaryOperator>(LHSI)->swapOperands(); // Make the LHS the RHS
1769 // If the functor wants to apply the optimization to the RHS of LHSI,
1770 // reassociate the expression from ((? op A) op B) to (? op (A op B))
1772 // Now all of the instructions are in the current basic block, go ahead
1773 // and perform the reassociation.
1774 Instruction *TmpLHSI = cast<Instruction>(Root.getOperand(0));
1776 // First move the selected RHS to the LHS of the root...
1777 Root.setOperand(0, LHSI->getOperand(1));
1779 // Make what used to be the LHS of the root be the user of the root...
1780 Value *ExtraOperand = TmpLHSI->getOperand(1);
1781 if (&Root == TmpLHSI) {
1782 Root.replaceAllUsesWith(Constant::getNullValue(TmpLHSI->getType()));
1785 Root.replaceAllUsesWith(TmpLHSI); // Users now use TmpLHSI
1786 TmpLHSI->setOperand(1, &Root); // TmpLHSI now uses the root
1787 BasicBlock::iterator ARI = &Root; ++ARI;
1788 TmpLHSI->moveBefore(ARI); // Move TmpLHSI to after Root
1791 // Now propagate the ExtraOperand down the chain of instructions until we
1793 while (TmpLHSI != LHSI) {
1794 Instruction *NextLHSI = cast<Instruction>(TmpLHSI->getOperand(0));
1795 // Move the instruction to immediately before the chain we are
1796 // constructing to avoid breaking dominance properties.
1797 NextLHSI->moveBefore(ARI);
1800 Value *NextOp = NextLHSI->getOperand(1);
1801 NextLHSI->setOperand(1, ExtraOperand);
1803 ExtraOperand = NextOp;
1806 // Now that the instructions are reassociated, have the functor perform
1807 // the transformation...
1808 return F.apply(Root);
1811 LHSI = dyn_cast<Instruction>(LHSI->getOperand(0));
1818 // AddRHS - Implements: X + X --> X << 1
1821 AddRHS(Value *rhs) : RHS(rhs) {}
1822 bool shouldApply(Value *LHS) const { return LHS == RHS; }
1823 Instruction *apply(BinaryOperator &Add) const {
1824 return BinaryOperator::CreateShl(Add.getOperand(0),
1825 ConstantInt::get(Add.getType(), 1));
1829 // AddMaskingAnd - Implements (A & C1)+(B & C2) --> (A & C1)|(B & C2)
1831 struct AddMaskingAnd {
1833 AddMaskingAnd(Constant *c) : C2(c) {}
1834 bool shouldApply(Value *LHS) const {
1836 return match(LHS, m_And(m_Value(), m_ConstantInt(C1))) &&
1837 ConstantExpr::getAnd(C1, C2)->isNullValue();
1839 Instruction *apply(BinaryOperator &Add) const {
1840 return BinaryOperator::CreateOr(Add.getOperand(0), Add.getOperand(1));
1846 static Value *FoldOperationIntoSelectOperand(Instruction &I, Value *SO,
1848 if (CastInst *CI = dyn_cast<CastInst>(&I)) {
1849 return IC->InsertCastBefore(CI->getOpcode(), SO, I.getType(), I);
1852 // Figure out if the constant is the left or the right argument.
1853 bool ConstIsRHS = isa<Constant>(I.getOperand(1));
1854 Constant *ConstOperand = cast<Constant>(I.getOperand(ConstIsRHS));
1856 if (Constant *SOC = dyn_cast<Constant>(SO)) {
1858 return ConstantExpr::get(I.getOpcode(), SOC, ConstOperand);
1859 return ConstantExpr::get(I.getOpcode(), ConstOperand, SOC);
1862 Value *Op0 = SO, *Op1 = ConstOperand;
1864 std::swap(Op0, Op1);
1866 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
1867 New = BinaryOperator::Create(BO->getOpcode(), Op0, Op1,SO->getName()+".op");
1868 else if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1869 New = CmpInst::Create(CI->getOpcode(), CI->getPredicate(), Op0, Op1,
1870 SO->getName()+".cmp");
1872 assert(0 && "Unknown binary instruction type!");
1875 return IC->InsertNewInstBefore(New, I);
1878 // FoldOpIntoSelect - Given an instruction with a select as one operand and a
1879 // constant as the other operand, try to fold the binary operator into the
1880 // select arguments. This also works for Cast instructions, which obviously do
1881 // not have a second operand.
1882 static Instruction *FoldOpIntoSelect(Instruction &Op, SelectInst *SI,
1884 // Don't modify shared select instructions
1885 if (!SI->hasOneUse()) return 0;
1886 Value *TV = SI->getOperand(1);
1887 Value *FV = SI->getOperand(2);
1889 if (isa<Constant>(TV) || isa<Constant>(FV)) {
1890 // Bool selects with constant operands can be folded to logical ops.
1891 if (SI->getType() == Type::Int1Ty) return 0;
1893 Value *SelectTrueVal = FoldOperationIntoSelectOperand(Op, TV, IC);
1894 Value *SelectFalseVal = FoldOperationIntoSelectOperand(Op, FV, IC);
1896 return SelectInst::Create(SI->getCondition(), SelectTrueVal,
1903 /// FoldOpIntoPhi - Given a binary operator or cast instruction which has a PHI
1904 /// node as operand #0, see if we can fold the instruction into the PHI (which
1905 /// is only possible if all operands to the PHI are constants).
1906 Instruction *InstCombiner::FoldOpIntoPhi(Instruction &I) {
1907 PHINode *PN = cast<PHINode>(I.getOperand(0));
1908 unsigned NumPHIValues = PN->getNumIncomingValues();
1909 if (!PN->hasOneUse() || NumPHIValues == 0) return 0;
1911 // Check to see if all of the operands of the PHI are constants. If there is
1912 // one non-constant value, remember the BB it is. If there is more than one
1913 // or if *it* is a PHI, bail out.
1914 BasicBlock *NonConstBB = 0;
1915 for (unsigned i = 0; i != NumPHIValues; ++i)
1916 if (!isa<Constant>(PN->getIncomingValue(i))) {
1917 if (NonConstBB) return 0; // More than one non-const value.
1918 if (isa<PHINode>(PN->getIncomingValue(i))) return 0; // Itself a phi.
1919 NonConstBB = PN->getIncomingBlock(i);
1921 // If the incoming non-constant value is in I's block, we have an infinite
1923 if (NonConstBB == I.getParent())
1927 // If there is exactly one non-constant value, we can insert a copy of the
1928 // operation in that block. However, if this is a critical edge, we would be
1929 // inserting the computation one some other paths (e.g. inside a loop). Only
1930 // do this if the pred block is unconditionally branching into the phi block.
1932 BranchInst *BI = dyn_cast<BranchInst>(NonConstBB->getTerminator());
1933 if (!BI || !BI->isUnconditional()) return 0;
1936 // Okay, we can do the transformation: create the new PHI node.
1937 PHINode *NewPN = PHINode::Create(I.getType(), "");
1938 NewPN->reserveOperandSpace(PN->getNumOperands()/2);
1939 InsertNewInstBefore(NewPN, *PN);
1940 NewPN->takeName(PN);
1942 // Next, add all of the operands to the PHI.
1943 if (I.getNumOperands() == 2) {
1944 Constant *C = cast<Constant>(I.getOperand(1));
1945 for (unsigned i = 0; i != NumPHIValues; ++i) {
1947 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
1948 if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1949 InV = ConstantExpr::getCompare(CI->getPredicate(), InC, C);
1951 InV = ConstantExpr::get(I.getOpcode(), InC, C);
1953 assert(PN->getIncomingBlock(i) == NonConstBB);
1954 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
1955 InV = BinaryOperator::Create(BO->getOpcode(),
1956 PN->getIncomingValue(i), C, "phitmp",
1957 NonConstBB->getTerminator());
1958 else if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1959 InV = CmpInst::Create(CI->getOpcode(),
1961 PN->getIncomingValue(i), C, "phitmp",
1962 NonConstBB->getTerminator());
1964 assert(0 && "Unknown binop!");
1966 AddToWorkList(cast<Instruction>(InV));
1968 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
1971 CastInst *CI = cast<CastInst>(&I);
1972 const Type *RetTy = CI->getType();
1973 for (unsigned i = 0; i != NumPHIValues; ++i) {
1975 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
1976 InV = ConstantExpr::getCast(CI->getOpcode(), InC, RetTy);
1978 assert(PN->getIncomingBlock(i) == NonConstBB);
1979 InV = CastInst::Create(CI->getOpcode(), PN->getIncomingValue(i),
1980 I.getType(), "phitmp",
1981 NonConstBB->getTerminator());
1982 AddToWorkList(cast<Instruction>(InV));
1984 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
1987 return ReplaceInstUsesWith(I, NewPN);
1991 /// WillNotOverflowSignedAdd - Return true if we can prove that:
1992 /// (sext (add LHS, RHS)) === (add (sext LHS), (sext RHS))
1993 /// This basically requires proving that the add in the original type would not
1994 /// overflow to change the sign bit or have a carry out.
1995 bool InstCombiner::WillNotOverflowSignedAdd(Value *LHS, Value *RHS) {
1996 // There are different heuristics we can use for this. Here are some simple
1999 // Add has the property that adding any two 2's complement numbers can only
2000 // have one carry bit which can change a sign. As such, if LHS and RHS each
2001 // have at least two sign bits, we know that the addition of the two values will
2002 // sign extend fine.
2003 if (ComputeNumSignBits(LHS) > 1 && ComputeNumSignBits(RHS) > 1)
2007 // If one of the operands only has one non-zero bit, and if the other operand
2008 // has a known-zero bit in a more significant place than it (not including the
2009 // sign bit) the ripple may go up to and fill the zero, but won't change the
2010 // sign. For example, (X & ~4) + 1.
2018 Instruction *InstCombiner::visitAdd(BinaryOperator &I) {
2019 bool Changed = SimplifyCommutative(I);
2020 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
2022 if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
2023 // X + undef -> undef
2024 if (isa<UndefValue>(RHS))
2025 return ReplaceInstUsesWith(I, RHS);
2028 if (!I.getType()->isFPOrFPVector()) { // NOTE: -0 + +0 = +0.
2029 if (RHSC->isNullValue())
2030 return ReplaceInstUsesWith(I, LHS);
2031 } else if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHSC)) {
2032 if (CFP->isExactlyValue(ConstantFP::getNegativeZero
2033 (I.getType())->getValueAPF()))
2034 return ReplaceInstUsesWith(I, LHS);
2037 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHSC)) {
2038 // X + (signbit) --> X ^ signbit
2039 const APInt& Val = CI->getValue();
2040 uint32_t BitWidth = Val.getBitWidth();
2041 if (Val == APInt::getSignBit(BitWidth))
2042 return BinaryOperator::CreateXor(LHS, RHS);
2044 // See if SimplifyDemandedBits can simplify this. This handles stuff like
2045 // (X & 254)+1 -> (X&254)|1
2046 if (!isa<VectorType>(I.getType()) && SimplifyDemandedInstructionBits(I))
2049 // zext(i1) - 1 -> select i1, 0, -1
2050 if (ZExtInst *ZI = dyn_cast<ZExtInst>(LHS))
2051 if (CI->isAllOnesValue() &&
2052 ZI->getOperand(0)->getType() == Type::Int1Ty)
2053 return SelectInst::Create(ZI->getOperand(0),
2054 Constant::getNullValue(I.getType()),
2055 ConstantInt::getAllOnesValue(I.getType()));
2058 if (isa<PHINode>(LHS))
2059 if (Instruction *NV = FoldOpIntoPhi(I))
2062 ConstantInt *XorRHS = 0;
2064 if (isa<ConstantInt>(RHSC) &&
2065 match(LHS, m_Xor(m_Value(XorLHS), m_ConstantInt(XorRHS)))) {
2066 uint32_t TySizeBits = I.getType()->getPrimitiveSizeInBits();
2067 const APInt& RHSVal = cast<ConstantInt>(RHSC)->getValue();
2069 uint32_t Size = TySizeBits / 2;
2070 APInt C0080Val(APInt(TySizeBits, 1ULL).shl(Size - 1));
2071 APInt CFF80Val(-C0080Val);
2073 if (TySizeBits > Size) {
2074 // If we have ADD(XOR(AND(X, 0xFF), 0x80), 0xF..F80), it's a sext.
2075 // If we have ADD(XOR(AND(X, 0xFF), 0xF..F80), 0x80), it's a sext.
2076 if ((RHSVal == CFF80Val && XorRHS->getValue() == C0080Val) ||
2077 (RHSVal == C0080Val && XorRHS->getValue() == CFF80Val)) {
2078 // This is a sign extend if the top bits are known zero.
2079 if (!MaskedValueIsZero(XorLHS,
2080 APInt::getHighBitsSet(TySizeBits, TySizeBits - Size)))
2081 Size = 0; // Not a sign ext, but can't be any others either.
2086 C0080Val = APIntOps::lshr(C0080Val, Size);
2087 CFF80Val = APIntOps::ashr(CFF80Val, Size);
2088 } while (Size >= 1);
2090 // FIXME: This shouldn't be necessary. When the backends can handle types
2091 // with funny bit widths then this switch statement should be removed. It
2092 // is just here to get the size of the "middle" type back up to something
2093 // that the back ends can handle.
2094 const Type *MiddleType = 0;
2097 case 32: MiddleType = Type::Int32Ty; break;
2098 case 16: MiddleType = Type::Int16Ty; break;
2099 case 8: MiddleType = Type::Int8Ty; break;
2102 Instruction *NewTrunc = new TruncInst(XorLHS, MiddleType, "sext");
2103 InsertNewInstBefore(NewTrunc, I);
2104 return new SExtInst(NewTrunc, I.getType(), I.getName());
2109 if (I.getType() == Type::Int1Ty)
2110 return BinaryOperator::CreateXor(LHS, RHS);
2113 if (I.getType()->isInteger()) {
2114 if (Instruction *Result = AssociativeOpt(I, AddRHS(RHS))) return Result;
2116 if (Instruction *RHSI = dyn_cast<Instruction>(RHS)) {
2117 if (RHSI->getOpcode() == Instruction::Sub)
2118 if (LHS == RHSI->getOperand(1)) // A + (B - A) --> B
2119 return ReplaceInstUsesWith(I, RHSI->getOperand(0));
2121 if (Instruction *LHSI = dyn_cast<Instruction>(LHS)) {
2122 if (LHSI->getOpcode() == Instruction::Sub)
2123 if (RHS == LHSI->getOperand(1)) // (B - A) + A --> B
2124 return ReplaceInstUsesWith(I, LHSI->getOperand(0));
2129 // -A + -B --> -(A + B)
2130 if (Value *LHSV = dyn_castNegVal(LHS)) {
2131 if (LHS->getType()->isIntOrIntVector()) {
2132 if (Value *RHSV = dyn_castNegVal(RHS)) {
2133 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSV, RHSV, "sum");
2134 InsertNewInstBefore(NewAdd, I);
2135 return BinaryOperator::CreateNeg(NewAdd);
2139 return BinaryOperator::CreateSub(RHS, LHSV);
2143 if (!isa<Constant>(RHS))
2144 if (Value *V = dyn_castNegVal(RHS))
2145 return BinaryOperator::CreateSub(LHS, V);
2149 if (Value *X = dyn_castFoldableMul(LHS, C2)) {
2150 if (X == RHS) // X*C + X --> X * (C+1)
2151 return BinaryOperator::CreateMul(RHS, AddOne(C2));
2153 // X*C1 + X*C2 --> X * (C1+C2)
2155 if (X == dyn_castFoldableMul(RHS, C1))
2156 return BinaryOperator::CreateMul(X, Add(C1, C2));
2159 // X + X*C --> X * (C+1)
2160 if (dyn_castFoldableMul(RHS, C2) == LHS)
2161 return BinaryOperator::CreateMul(LHS, AddOne(C2));
2163 // X + ~X --> -1 since ~X = -X-1
2164 if (dyn_castNotVal(LHS) == RHS || dyn_castNotVal(RHS) == LHS)
2165 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
2168 // (A & C1)+(B & C2) --> (A & C1)|(B & C2) iff C1&C2 == 0
2169 if (match(RHS, m_And(m_Value(), m_ConstantInt(C2))))
2170 if (Instruction *R = AssociativeOpt(I, AddMaskingAnd(C2)))
2173 // A+B --> A|B iff A and B have no bits set in common.
2174 if (const IntegerType *IT = dyn_cast<IntegerType>(I.getType())) {
2175 APInt Mask = APInt::getAllOnesValue(IT->getBitWidth());
2176 APInt LHSKnownOne(IT->getBitWidth(), 0);
2177 APInt LHSKnownZero(IT->getBitWidth(), 0);
2178 ComputeMaskedBits(LHS, Mask, LHSKnownZero, LHSKnownOne);
2179 if (LHSKnownZero != 0) {
2180 APInt RHSKnownOne(IT->getBitWidth(), 0);
2181 APInt RHSKnownZero(IT->getBitWidth(), 0);
2182 ComputeMaskedBits(RHS, Mask, RHSKnownZero, RHSKnownOne);
2184 // No bits in common -> bitwise or.
2185 if ((LHSKnownZero|RHSKnownZero).isAllOnesValue())
2186 return BinaryOperator::CreateOr(LHS, RHS);
2190 // W*X + Y*Z --> W * (X+Z) iff W == Y
2191 if (I.getType()->isIntOrIntVector()) {
2192 Value *W, *X, *Y, *Z;
2193 if (match(LHS, m_Mul(m_Value(W), m_Value(X))) &&
2194 match(RHS, m_Mul(m_Value(Y), m_Value(Z)))) {
2198 } else if (Y == X) {
2200 } else if (X == Z) {
2207 Value *NewAdd = InsertNewInstBefore(BinaryOperator::CreateAdd(X, Z,
2208 LHS->getName()), I);
2209 return BinaryOperator::CreateMul(W, NewAdd);
2214 if (ConstantInt *CRHS = dyn_cast<ConstantInt>(RHS)) {
2216 if (match(LHS, m_Not(m_Value(X)))) // ~X + C --> (C-1) - X
2217 return BinaryOperator::CreateSub(SubOne(CRHS), X);
2219 // (X & FF00) + xx00 -> (X+xx00) & FF00
2220 if (LHS->hasOneUse() && match(LHS, m_And(m_Value(X), m_ConstantInt(C2)))) {
2221 Constant *Anded = And(CRHS, C2);
2222 if (Anded == CRHS) {
2223 // See if all bits from the first bit set in the Add RHS up are included
2224 // in the mask. First, get the rightmost bit.
2225 const APInt& AddRHSV = CRHS->getValue();
2227 // Form a mask of all bits from the lowest bit added through the top.
2228 APInt AddRHSHighBits(~((AddRHSV & -AddRHSV)-1));
2230 // See if the and mask includes all of these bits.
2231 APInt AddRHSHighBitsAnd(AddRHSHighBits & C2->getValue());
2233 if (AddRHSHighBits == AddRHSHighBitsAnd) {
2234 // Okay, the xform is safe. Insert the new add pronto.
2235 Value *NewAdd = InsertNewInstBefore(BinaryOperator::CreateAdd(X, CRHS,
2236 LHS->getName()), I);
2237 return BinaryOperator::CreateAnd(NewAdd, C2);
2242 // Try to fold constant add into select arguments.
2243 if (SelectInst *SI = dyn_cast<SelectInst>(LHS))
2244 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2248 // add (cast *A to intptrtype) B ->
2249 // cast (GEP (cast *A to sbyte*) B) --> intptrtype
2251 CastInst *CI = dyn_cast<CastInst>(LHS);
2254 CI = dyn_cast<CastInst>(RHS);
2257 if (CI && CI->getType()->isSized() &&
2258 (CI->getType()->getPrimitiveSizeInBits() ==
2259 TD->getIntPtrType()->getPrimitiveSizeInBits())
2260 && isa<PointerType>(CI->getOperand(0)->getType())) {
2262 cast<PointerType>(CI->getOperand(0)->getType())->getAddressSpace();
2263 Value *I2 = InsertBitCastBefore(CI->getOperand(0),
2264 PointerType::get(Type::Int8Ty, AS), I);
2265 I2 = InsertNewInstBefore(GetElementPtrInst::Create(I2, Other, "ctg2"), I);
2266 return new PtrToIntInst(I2, CI->getType());
2270 // add (select X 0 (sub n A)) A --> select X A n
2272 SelectInst *SI = dyn_cast<SelectInst>(LHS);
2275 SI = dyn_cast<SelectInst>(RHS);
2278 if (SI && SI->hasOneUse()) {
2279 Value *TV = SI->getTrueValue();
2280 Value *FV = SI->getFalseValue();
2283 // Can we fold the add into the argument of the select?
2284 // We check both true and false select arguments for a matching subtract.
2285 if (match(FV, m_Zero()) && match(TV, m_Sub(m_Value(N), m_Specific(A))))
2286 // Fold the add into the true select value.
2287 return SelectInst::Create(SI->getCondition(), N, A);
2288 if (match(TV, m_Zero()) && match(FV, m_Sub(m_Value(N), m_Specific(A))))
2289 // Fold the add into the false select value.
2290 return SelectInst::Create(SI->getCondition(), A, N);
2294 // Check for X+0.0. Simplify it to X if we know X is not -0.0.
2295 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHS))
2296 if (CFP->getValueAPF().isPosZero() && CannotBeNegativeZero(LHS))
2297 return ReplaceInstUsesWith(I, LHS);
2299 // Check for (add (sext x), y), see if we can merge this into an
2300 // integer add followed by a sext.
2301 if (SExtInst *LHSConv = dyn_cast<SExtInst>(LHS)) {
2302 // (add (sext x), cst) --> (sext (add x, cst'))
2303 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS)) {
2305 ConstantExpr::getTrunc(RHSC, LHSConv->getOperand(0)->getType());
2306 if (LHSConv->hasOneUse() &&
2307 ConstantExpr::getSExt(CI, I.getType()) == RHSC &&
2308 WillNotOverflowSignedAdd(LHSConv->getOperand(0), CI)) {
2309 // Insert the new, smaller add.
2310 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSConv->getOperand(0),
2312 InsertNewInstBefore(NewAdd, I);
2313 return new SExtInst(NewAdd, I.getType());
2317 // (add (sext x), (sext y)) --> (sext (add int x, y))
2318 if (SExtInst *RHSConv = dyn_cast<SExtInst>(RHS)) {
2319 // Only do this if x/y have the same type, if at last one of them has a
2320 // single use (so we don't increase the number of sexts), and if the
2321 // integer add will not overflow.
2322 if (LHSConv->getOperand(0)->getType()==RHSConv->getOperand(0)->getType()&&
2323 (LHSConv->hasOneUse() || RHSConv->hasOneUse()) &&
2324 WillNotOverflowSignedAdd(LHSConv->getOperand(0),
2325 RHSConv->getOperand(0))) {
2326 // Insert the new integer add.
2327 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSConv->getOperand(0),
2328 RHSConv->getOperand(0),
2330 InsertNewInstBefore(NewAdd, I);
2331 return new SExtInst(NewAdd, I.getType());
2336 // Check for (add double (sitofp x), y), see if we can merge this into an
2337 // integer add followed by a promotion.
2338 if (SIToFPInst *LHSConv = dyn_cast<SIToFPInst>(LHS)) {
2339 // (add double (sitofp x), fpcst) --> (sitofp (add int x, intcst))
2340 // ... if the constant fits in the integer value. This is useful for things
2341 // like (double)(x & 1234) + 4.0 -> (double)((X & 1234)+4) which no longer
2342 // requires a constant pool load, and generally allows the add to be better
2344 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHS)) {
2346 ConstantExpr::getFPToSI(CFP, LHSConv->getOperand(0)->getType());
2347 if (LHSConv->hasOneUse() &&
2348 ConstantExpr::getSIToFP(CI, I.getType()) == CFP &&
2349 WillNotOverflowSignedAdd(LHSConv->getOperand(0), CI)) {
2350 // Insert the new integer add.
2351 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSConv->getOperand(0),
2353 InsertNewInstBefore(NewAdd, I);
2354 return new SIToFPInst(NewAdd, I.getType());
2358 // (add double (sitofp x), (sitofp y)) --> (sitofp (add int x, y))
2359 if (SIToFPInst *RHSConv = dyn_cast<SIToFPInst>(RHS)) {
2360 // Only do this if x/y have the same type, if at last one of them has a
2361 // single use (so we don't increase the number of int->fp conversions),
2362 // and if the integer add will not overflow.
2363 if (LHSConv->getOperand(0)->getType()==RHSConv->getOperand(0)->getType()&&
2364 (LHSConv->hasOneUse() || RHSConv->hasOneUse()) &&
2365 WillNotOverflowSignedAdd(LHSConv->getOperand(0),
2366 RHSConv->getOperand(0))) {
2367 // Insert the new integer add.
2368 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSConv->getOperand(0),
2369 RHSConv->getOperand(0),
2371 InsertNewInstBefore(NewAdd, I);
2372 return new SIToFPInst(NewAdd, I.getType());
2377 return Changed ? &I : 0;
2380 Instruction *InstCombiner::visitSub(BinaryOperator &I) {
2381 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2383 if (Op0 == Op1 && // sub X, X -> 0
2384 !I.getType()->isFPOrFPVector())
2385 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2387 // If this is a 'B = x-(-A)', change to B = x+A...
2388 if (Value *V = dyn_castNegVal(Op1))
2389 return BinaryOperator::CreateAdd(Op0, V);
2391 if (isa<UndefValue>(Op0))
2392 return ReplaceInstUsesWith(I, Op0); // undef - X -> undef
2393 if (isa<UndefValue>(Op1))
2394 return ReplaceInstUsesWith(I, Op1); // X - undef -> undef
2396 if (ConstantInt *C = dyn_cast<ConstantInt>(Op0)) {
2397 // Replace (-1 - A) with (~A)...
2398 if (C->isAllOnesValue())
2399 return BinaryOperator::CreateNot(Op1);
2401 // C - ~X == X + (1+C)
2403 if (match(Op1, m_Not(m_Value(X))))
2404 return BinaryOperator::CreateAdd(X, AddOne(C));
2406 // -(X >>u 31) -> (X >>s 31)
2407 // -(X >>s 31) -> (X >>u 31)
2409 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op1)) {
2410 if (SI->getOpcode() == Instruction::LShr) {
2411 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
2412 // Check to see if we are shifting out everything but the sign bit.
2413 if (CU->getLimitedValue(SI->getType()->getPrimitiveSizeInBits()) ==
2414 SI->getType()->getPrimitiveSizeInBits()-1) {
2415 // Ok, the transformation is safe. Insert AShr.
2416 return BinaryOperator::Create(Instruction::AShr,
2417 SI->getOperand(0), CU, SI->getName());
2421 else if (SI->getOpcode() == Instruction::AShr) {
2422 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
2423 // Check to see if we are shifting out everything but the sign bit.
2424 if (CU->getLimitedValue(SI->getType()->getPrimitiveSizeInBits()) ==
2425 SI->getType()->getPrimitiveSizeInBits()-1) {
2426 // Ok, the transformation is safe. Insert LShr.
2427 return BinaryOperator::CreateLShr(
2428 SI->getOperand(0), CU, SI->getName());
2435 // Try to fold constant sub into select arguments.
2436 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
2437 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2441 if (I.getType() == Type::Int1Ty)
2442 return BinaryOperator::CreateXor(Op0, Op1);
2444 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
2445 if (Op1I->getOpcode() == Instruction::Add &&
2446 !Op0->getType()->isFPOrFPVector()) {
2447 if (Op1I->getOperand(0) == Op0) // X-(X+Y) == -Y
2448 return BinaryOperator::CreateNeg(Op1I->getOperand(1), I.getName());
2449 else if (Op1I->getOperand(1) == Op0) // X-(Y+X) == -Y
2450 return BinaryOperator::CreateNeg(Op1I->getOperand(0), I.getName());
2451 else if (ConstantInt *CI1 = dyn_cast<ConstantInt>(I.getOperand(0))) {
2452 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(Op1I->getOperand(1)))
2453 // C1-(X+C2) --> (C1-C2)-X
2454 return BinaryOperator::CreateSub(Subtract(CI1, CI2),
2455 Op1I->getOperand(0));
2459 if (Op1I->hasOneUse()) {
2460 // Replace (x - (y - z)) with (x + (z - y)) if the (y - z) subexpression
2461 // is not used by anyone else...
2463 if (Op1I->getOpcode() == Instruction::Sub &&
2464 !Op1I->getType()->isFPOrFPVector()) {
2465 // Swap the two operands of the subexpr...
2466 Value *IIOp0 = Op1I->getOperand(0), *IIOp1 = Op1I->getOperand(1);
2467 Op1I->setOperand(0, IIOp1);
2468 Op1I->setOperand(1, IIOp0);
2470 // Create the new top level add instruction...
2471 return BinaryOperator::CreateAdd(Op0, Op1);
2474 // Replace (A - (A & B)) with (A & ~B) if this is the only use of (A&B)...
2476 if (Op1I->getOpcode() == Instruction::And &&
2477 (Op1I->getOperand(0) == Op0 || Op1I->getOperand(1) == Op0)) {
2478 Value *OtherOp = Op1I->getOperand(Op1I->getOperand(0) == Op0);
2481 InsertNewInstBefore(BinaryOperator::CreateNot(OtherOp, "B.not"), I);
2482 return BinaryOperator::CreateAnd(Op0, NewNot);
2485 // 0 - (X sdiv C) -> (X sdiv -C)
2486 if (Op1I->getOpcode() == Instruction::SDiv)
2487 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
2489 if (Constant *DivRHS = dyn_cast<Constant>(Op1I->getOperand(1)))
2490 return BinaryOperator::CreateSDiv(Op1I->getOperand(0),
2491 ConstantExpr::getNeg(DivRHS));
2493 // X - X*C --> X * (1-C)
2494 ConstantInt *C2 = 0;
2495 if (dyn_castFoldableMul(Op1I, C2) == Op0) {
2496 Constant *CP1 = Subtract(ConstantInt::get(I.getType(), 1), C2);
2497 return BinaryOperator::CreateMul(Op0, CP1);
2502 if (!Op0->getType()->isFPOrFPVector())
2503 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
2504 if (Op0I->getOpcode() == Instruction::Add) {
2505 if (Op0I->getOperand(0) == Op1) // (Y+X)-Y == X
2506 return ReplaceInstUsesWith(I, Op0I->getOperand(1));
2507 else if (Op0I->getOperand(1) == Op1) // (X+Y)-Y == X
2508 return ReplaceInstUsesWith(I, Op0I->getOperand(0));
2509 } else if (Op0I->getOpcode() == Instruction::Sub) {
2510 if (Op0I->getOperand(0) == Op1) // (X-Y)-X == -Y
2511 return BinaryOperator::CreateNeg(Op0I->getOperand(1), I.getName());
2516 if (Value *X = dyn_castFoldableMul(Op0, C1)) {
2517 if (X == Op1) // X*C - X --> X * (C-1)
2518 return BinaryOperator::CreateMul(Op1, SubOne(C1));
2520 ConstantInt *C2; // X*C1 - X*C2 -> X * (C1-C2)
2521 if (X == dyn_castFoldableMul(Op1, C2))
2522 return BinaryOperator::CreateMul(X, Subtract(C1, C2));
2527 /// isSignBitCheck - Given an exploded icmp instruction, return true if the
2528 /// comparison only checks the sign bit. If it only checks the sign bit, set
2529 /// TrueIfSigned if the result of the comparison is true when the input value is
2531 static bool isSignBitCheck(ICmpInst::Predicate pred, ConstantInt *RHS,
2532 bool &TrueIfSigned) {
2534 case ICmpInst::ICMP_SLT: // True if LHS s< 0
2535 TrueIfSigned = true;
2536 return RHS->isZero();
2537 case ICmpInst::ICMP_SLE: // True if LHS s<= RHS and RHS == -1
2538 TrueIfSigned = true;
2539 return RHS->isAllOnesValue();
2540 case ICmpInst::ICMP_SGT: // True if LHS s> -1
2541 TrueIfSigned = false;
2542 return RHS->isAllOnesValue();
2543 case ICmpInst::ICMP_UGT:
2544 // True if LHS u> RHS and RHS == high-bit-mask - 1
2545 TrueIfSigned = true;
2546 return RHS->getValue() ==
2547 APInt::getSignedMaxValue(RHS->getType()->getPrimitiveSizeInBits());
2548 case ICmpInst::ICMP_UGE:
2549 // True if LHS u>= RHS and RHS == high-bit-mask (2^7, 2^15, 2^31, etc)
2550 TrueIfSigned = true;
2551 return RHS->getValue().isSignBit();
2557 Instruction *InstCombiner::visitMul(BinaryOperator &I) {
2558 bool Changed = SimplifyCommutative(I);
2559 Value *Op0 = I.getOperand(0);
2561 if (isa<UndefValue>(I.getOperand(1))) // undef * X -> 0
2562 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2564 // Simplify mul instructions with a constant RHS...
2565 if (Constant *Op1 = dyn_cast<Constant>(I.getOperand(1))) {
2566 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2568 // ((X << C1)*C2) == (X * (C2 << C1))
2569 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op0))
2570 if (SI->getOpcode() == Instruction::Shl)
2571 if (Constant *ShOp = dyn_cast<Constant>(SI->getOperand(1)))
2572 return BinaryOperator::CreateMul(SI->getOperand(0),
2573 ConstantExpr::getShl(CI, ShOp));
2576 return ReplaceInstUsesWith(I, Op1); // X * 0 == 0
2577 if (CI->equalsInt(1)) // X * 1 == X
2578 return ReplaceInstUsesWith(I, Op0);
2579 if (CI->isAllOnesValue()) // X * -1 == 0 - X
2580 return BinaryOperator::CreateNeg(Op0, I.getName());
2582 const APInt& Val = cast<ConstantInt>(CI)->getValue();
2583 if (Val.isPowerOf2()) { // Replace X*(2^C) with X << C
2584 return BinaryOperator::CreateShl(Op0,
2585 ConstantInt::get(Op0->getType(), Val.logBase2()));
2587 } else if (ConstantFP *Op1F = dyn_cast<ConstantFP>(Op1)) {
2588 if (Op1F->isNullValue())
2589 return ReplaceInstUsesWith(I, Op1);
2591 // "In IEEE floating point, x*1 is not equivalent to x for nans. However,
2592 // ANSI says we can drop signals, so we can do this anyway." (from GCC)
2593 if (Op1F->isExactlyValue(1.0))
2594 return ReplaceInstUsesWith(I, Op0); // Eliminate 'mul double %X, 1.0'
2595 } else if (isa<VectorType>(Op1->getType())) {
2596 if (isa<ConstantAggregateZero>(Op1))
2597 return ReplaceInstUsesWith(I, Op1);
2599 if (ConstantVector *Op1V = dyn_cast<ConstantVector>(Op1)) {
2600 if (Op1V->isAllOnesValue()) // X * -1 == 0 - X
2601 return BinaryOperator::CreateNeg(Op0, I.getName());
2603 // As above, vector X*splat(1.0) -> X in all defined cases.
2604 if (Constant *Splat = Op1V->getSplatValue()) {
2605 if (ConstantFP *F = dyn_cast<ConstantFP>(Splat))
2606 if (F->isExactlyValue(1.0))
2607 return ReplaceInstUsesWith(I, Op0);
2608 if (ConstantInt *CI = dyn_cast<ConstantInt>(Splat))
2609 if (CI->equalsInt(1))
2610 return ReplaceInstUsesWith(I, Op0);
2615 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
2616 if (Op0I->getOpcode() == Instruction::Add && Op0I->hasOneUse() &&
2617 isa<ConstantInt>(Op0I->getOperand(1)) && isa<ConstantInt>(Op1)) {
2618 // Canonicalize (X+C1)*C2 -> X*C2+C1*C2.
2619 Instruction *Add = BinaryOperator::CreateMul(Op0I->getOperand(0),
2621 InsertNewInstBefore(Add, I);
2622 Value *C1C2 = ConstantExpr::getMul(Op1,
2623 cast<Constant>(Op0I->getOperand(1)));
2624 return BinaryOperator::CreateAdd(Add, C1C2);
2628 // Try to fold constant mul into select arguments.
2629 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2630 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2633 if (isa<PHINode>(Op0))
2634 if (Instruction *NV = FoldOpIntoPhi(I))
2638 if (Value *Op0v = dyn_castNegVal(Op0)) // -X * -Y = X*Y
2639 if (Value *Op1v = dyn_castNegVal(I.getOperand(1)))
2640 return BinaryOperator::CreateMul(Op0v, Op1v);
2642 // (X / Y) * Y = X - (X % Y)
2643 // (X / Y) * -Y = (X % Y) - X
2645 Value *Op1 = I.getOperand(1);
2646 BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0);
2648 (BO->getOpcode() != Instruction::UDiv &&
2649 BO->getOpcode() != Instruction::SDiv)) {
2651 BO = dyn_cast<BinaryOperator>(I.getOperand(1));
2653 Value *Neg = dyn_castNegVal(Op1);
2654 if (BO && BO->hasOneUse() &&
2655 (BO->getOperand(1) == Op1 || BO->getOperand(1) == Neg) &&
2656 (BO->getOpcode() == Instruction::UDiv ||
2657 BO->getOpcode() == Instruction::SDiv)) {
2658 Value *Op0BO = BO->getOperand(0), *Op1BO = BO->getOperand(1);
2661 if (BO->getOpcode() == Instruction::UDiv)
2662 Rem = BinaryOperator::CreateURem(Op0BO, Op1BO);
2664 Rem = BinaryOperator::CreateSRem(Op0BO, Op1BO);
2666 InsertNewInstBefore(Rem, I);
2670 return BinaryOperator::CreateSub(Op0BO, Rem);
2672 return BinaryOperator::CreateSub(Rem, Op0BO);
2676 if (I.getType() == Type::Int1Ty)
2677 return BinaryOperator::CreateAnd(Op0, I.getOperand(1));
2679 // If one of the operands of the multiply is a cast from a boolean value, then
2680 // we know the bool is either zero or one, so this is a 'masking' multiply.
2681 // See if we can simplify things based on how the boolean was originally
2683 CastInst *BoolCast = 0;
2684 if (ZExtInst *CI = dyn_cast<ZExtInst>(Op0))
2685 if (CI->getOperand(0)->getType() == Type::Int1Ty)
2688 if (ZExtInst *CI = dyn_cast<ZExtInst>(I.getOperand(1)))
2689 if (CI->getOperand(0)->getType() == Type::Int1Ty)
2692 if (ICmpInst *SCI = dyn_cast<ICmpInst>(BoolCast->getOperand(0))) {
2693 Value *SCIOp0 = SCI->getOperand(0), *SCIOp1 = SCI->getOperand(1);
2694 const Type *SCOpTy = SCIOp0->getType();
2697 // If the icmp is true iff the sign bit of X is set, then convert this
2698 // multiply into a shift/and combination.
2699 if (isa<ConstantInt>(SCIOp1) &&
2700 isSignBitCheck(SCI->getPredicate(), cast<ConstantInt>(SCIOp1), TIS) &&
2702 // Shift the X value right to turn it into "all signbits".
2703 Constant *Amt = ConstantInt::get(SCIOp0->getType(),
2704 SCOpTy->getPrimitiveSizeInBits()-1);
2706 InsertNewInstBefore(
2707 BinaryOperator::Create(Instruction::AShr, SCIOp0, Amt,
2708 BoolCast->getOperand(0)->getName()+
2711 // If the multiply type is not the same as the source type, sign extend
2712 // or truncate to the multiply type.
2713 if (I.getType() != V->getType()) {
2714 uint32_t SrcBits = V->getType()->getPrimitiveSizeInBits();
2715 uint32_t DstBits = I.getType()->getPrimitiveSizeInBits();
2716 Instruction::CastOps opcode =
2717 (SrcBits == DstBits ? Instruction::BitCast :
2718 (SrcBits < DstBits ? Instruction::SExt : Instruction::Trunc));
2719 V = InsertCastBefore(opcode, V, I.getType(), I);
2722 Value *OtherOp = Op0 == BoolCast ? I.getOperand(1) : Op0;
2723 return BinaryOperator::CreateAnd(V, OtherOp);
2728 return Changed ? &I : 0;
2731 /// SimplifyDivRemOfSelect - Try to fold a divide or remainder of a select
2733 bool InstCombiner::SimplifyDivRemOfSelect(BinaryOperator &I) {
2734 SelectInst *SI = cast<SelectInst>(I.getOperand(1));
2736 // div/rem X, (Cond ? 0 : Y) -> div/rem X, Y
2737 int NonNullOperand = -1;
2738 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(1)))
2739 if (ST->isNullValue())
2741 // div/rem X, (Cond ? Y : 0) -> div/rem X, Y
2742 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(2)))
2743 if (ST->isNullValue())
2746 if (NonNullOperand == -1)
2749 Value *SelectCond = SI->getOperand(0);
2751 // Change the div/rem to use 'Y' instead of the select.
2752 I.setOperand(1, SI->getOperand(NonNullOperand));
2754 // Okay, we know we replace the operand of the div/rem with 'Y' with no
2755 // problem. However, the select, or the condition of the select may have
2756 // multiple uses. Based on our knowledge that the operand must be non-zero,
2757 // propagate the known value for the select into other uses of it, and
2758 // propagate a known value of the condition into its other users.
2760 // If the select and condition only have a single use, don't bother with this,
2762 if (SI->use_empty() && SelectCond->hasOneUse())
2765 // Scan the current block backward, looking for other uses of SI.
2766 BasicBlock::iterator BBI = &I, BBFront = I.getParent()->begin();
2768 while (BBI != BBFront) {
2770 // If we found a call to a function, we can't assume it will return, so
2771 // information from below it cannot be propagated above it.
2772 if (isa<CallInst>(BBI) && !isa<IntrinsicInst>(BBI))
2775 // Replace uses of the select or its condition with the known values.
2776 for (Instruction::op_iterator I = BBI->op_begin(), E = BBI->op_end();
2779 *I = SI->getOperand(NonNullOperand);
2781 } else if (*I == SelectCond) {
2782 *I = NonNullOperand == 1 ? ConstantInt::getTrue() :
2783 ConstantInt::getFalse();
2788 // If we past the instruction, quit looking for it.
2791 if (&*BBI == SelectCond)
2794 // If we ran out of things to eliminate, break out of the loop.
2795 if (SelectCond == 0 && SI == 0)
2803 /// This function implements the transforms on div instructions that work
2804 /// regardless of the kind of div instruction it is (udiv, sdiv, or fdiv). It is
2805 /// used by the visitors to those instructions.
2806 /// @brief Transforms common to all three div instructions
2807 Instruction *InstCombiner::commonDivTransforms(BinaryOperator &I) {
2808 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2810 // undef / X -> 0 for integer.
2811 // undef / X -> undef for FP (the undef could be a snan).
2812 if (isa<UndefValue>(Op0)) {
2813 if (Op0->getType()->isFPOrFPVector())
2814 return ReplaceInstUsesWith(I, Op0);
2815 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2818 // X / undef -> undef
2819 if (isa<UndefValue>(Op1))
2820 return ReplaceInstUsesWith(I, Op1);
2825 /// This function implements the transforms common to both integer division
2826 /// instructions (udiv and sdiv). It is called by the visitors to those integer
2827 /// division instructions.
2828 /// @brief Common integer divide transforms
2829 Instruction *InstCombiner::commonIDivTransforms(BinaryOperator &I) {
2830 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2832 // (sdiv X, X) --> 1 (udiv X, X) --> 1
2834 if (const VectorType *Ty = dyn_cast<VectorType>(I.getType())) {
2835 ConstantInt *CI = ConstantInt::get(Ty->getElementType(), 1);
2836 std::vector<Constant*> Elts(Ty->getNumElements(), CI);
2837 return ReplaceInstUsesWith(I, ConstantVector::get(Elts));
2840 ConstantInt *CI = ConstantInt::get(I.getType(), 1);
2841 return ReplaceInstUsesWith(I, CI);
2844 if (Instruction *Common = commonDivTransforms(I))
2847 // Handle cases involving: [su]div X, (select Cond, Y, Z)
2848 // This does not apply for fdiv.
2849 if (isa<SelectInst>(Op1) && SimplifyDivRemOfSelect(I))
2852 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2854 if (RHS->equalsInt(1))
2855 return ReplaceInstUsesWith(I, Op0);
2857 // (X / C1) / C2 -> X / (C1*C2)
2858 if (Instruction *LHS = dyn_cast<Instruction>(Op0))
2859 if (Instruction::BinaryOps(LHS->getOpcode()) == I.getOpcode())
2860 if (ConstantInt *LHSRHS = dyn_cast<ConstantInt>(LHS->getOperand(1))) {
2861 if (MultiplyOverflows(RHS, LHSRHS, I.getOpcode()==Instruction::SDiv))
2862 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2864 return BinaryOperator::Create(I.getOpcode(), LHS->getOperand(0),
2865 Multiply(RHS, LHSRHS));
2868 if (!RHS->isZero()) { // avoid X udiv 0
2869 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2870 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2872 if (isa<PHINode>(Op0))
2873 if (Instruction *NV = FoldOpIntoPhi(I))
2878 // 0 / X == 0, we don't need to preserve faults!
2879 if (ConstantInt *LHS = dyn_cast<ConstantInt>(Op0))
2880 if (LHS->equalsInt(0))
2881 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2883 // It can't be division by zero, hence it must be division by one.
2884 if (I.getType() == Type::Int1Ty)
2885 return ReplaceInstUsesWith(I, Op0);
2887 if (ConstantVector *Op1V = dyn_cast<ConstantVector>(Op1)) {
2888 if (ConstantInt *X = cast_or_null<ConstantInt>(Op1V->getSplatValue()))
2891 return ReplaceInstUsesWith(I, Op0);
2897 Instruction *InstCombiner::visitUDiv(BinaryOperator &I) {
2898 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2900 // Handle the integer div common cases
2901 if (Instruction *Common = commonIDivTransforms(I))
2904 if (ConstantInt *C = dyn_cast<ConstantInt>(Op1)) {
2905 // X udiv C^2 -> X >> C
2906 // Check to see if this is an unsigned division with an exact power of 2,
2907 // if so, convert to a right shift.
2908 if (C->getValue().isPowerOf2()) // 0 not included in isPowerOf2
2909 return BinaryOperator::CreateLShr(Op0,
2910 ConstantInt::get(Op0->getType(), C->getValue().logBase2()));
2912 // X udiv C, where C >= signbit
2913 if (C->getValue().isNegative()) {
2914 Value *IC = InsertNewInstBefore(new ICmpInst(ICmpInst::ICMP_ULT, Op0, C),
2916 return SelectInst::Create(IC, Constant::getNullValue(I.getType()),
2917 ConstantInt::get(I.getType(), 1));
2921 // X udiv (C1 << N), where C1 is "1<<C2" --> X >> (N+C2)
2922 if (BinaryOperator *RHSI = dyn_cast<BinaryOperator>(I.getOperand(1))) {
2923 if (RHSI->getOpcode() == Instruction::Shl &&
2924 isa<ConstantInt>(RHSI->getOperand(0))) {
2925 const APInt& C1 = cast<ConstantInt>(RHSI->getOperand(0))->getValue();
2926 if (C1.isPowerOf2()) {
2927 Value *N = RHSI->getOperand(1);
2928 const Type *NTy = N->getType();
2929 if (uint32_t C2 = C1.logBase2()) {
2930 Constant *C2V = ConstantInt::get(NTy, C2);
2931 N = InsertNewInstBefore(BinaryOperator::CreateAdd(N, C2V, "tmp"), I);
2933 return BinaryOperator::CreateLShr(Op0, N);
2938 // udiv X, (Select Cond, C1, C2) --> Select Cond, (shr X, C1), (shr X, C2)
2939 // where C1&C2 are powers of two.
2940 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
2941 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
2942 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
2943 const APInt &TVA = STO->getValue(), &FVA = SFO->getValue();
2944 if (TVA.isPowerOf2() && FVA.isPowerOf2()) {
2945 // Compute the shift amounts
2946 uint32_t TSA = TVA.logBase2(), FSA = FVA.logBase2();
2947 // Construct the "on true" case of the select
2948 Constant *TC = ConstantInt::get(Op0->getType(), TSA);
2949 Instruction *TSI = BinaryOperator::CreateLShr(
2950 Op0, TC, SI->getName()+".t");
2951 TSI = InsertNewInstBefore(TSI, I);
2953 // Construct the "on false" case of the select
2954 Constant *FC = ConstantInt::get(Op0->getType(), FSA);
2955 Instruction *FSI = BinaryOperator::CreateLShr(
2956 Op0, FC, SI->getName()+".f");
2957 FSI = InsertNewInstBefore(FSI, I);
2959 // construct the select instruction and return it.
2960 return SelectInst::Create(SI->getOperand(0), TSI, FSI, SI->getName());
2966 Instruction *InstCombiner::visitSDiv(BinaryOperator &I) {
2967 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2969 // Handle the integer div common cases
2970 if (Instruction *Common = commonIDivTransforms(I))
2973 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2975 if (RHS->isAllOnesValue())
2976 return BinaryOperator::CreateNeg(Op0);
2979 // If the sign bits of both operands are zero (i.e. we can prove they are
2980 // unsigned inputs), turn this into a udiv.
2981 if (I.getType()->isInteger()) {
2982 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()));
2983 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
2984 // X sdiv Y -> X udiv Y, iff X and Y don't have sign bit set
2985 return BinaryOperator::CreateUDiv(Op0, Op1, I.getName());
2992 Instruction *InstCombiner::visitFDiv(BinaryOperator &I) {
2993 return commonDivTransforms(I);
2996 /// This function implements the transforms on rem instructions that work
2997 /// regardless of the kind of rem instruction it is (urem, srem, or frem). It
2998 /// is used by the visitors to those instructions.
2999 /// @brief Transforms common to all three rem instructions
3000 Instruction *InstCombiner::commonRemTransforms(BinaryOperator &I) {
3001 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3003 if (isa<UndefValue>(Op0)) { // undef % X -> 0
3004 if (I.getType()->isFPOrFPVector())
3005 return ReplaceInstUsesWith(I, Op0); // X % undef -> undef (could be SNaN)
3006 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3008 if (isa<UndefValue>(Op1))
3009 return ReplaceInstUsesWith(I, Op1); // X % undef -> undef
3011 // Handle cases involving: rem X, (select Cond, Y, Z)
3012 if (isa<SelectInst>(Op1) && SimplifyDivRemOfSelect(I))
3018 /// This function implements the transforms common to both integer remainder
3019 /// instructions (urem and srem). It is called by the visitors to those integer
3020 /// remainder instructions.
3021 /// @brief Common integer remainder transforms
3022 Instruction *InstCombiner::commonIRemTransforms(BinaryOperator &I) {
3023 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3025 if (Instruction *common = commonRemTransforms(I))
3028 // 0 % X == 0 for integer, we don't need to preserve faults!
3029 if (Constant *LHS = dyn_cast<Constant>(Op0))
3030 if (LHS->isNullValue())
3031 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3033 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3034 // X % 0 == undef, we don't need to preserve faults!
3035 if (RHS->equalsInt(0))
3036 return ReplaceInstUsesWith(I, UndefValue::get(I.getType()));
3038 if (RHS->equalsInt(1)) // X % 1 == 0
3039 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3041 if (Instruction *Op0I = dyn_cast<Instruction>(Op0)) {
3042 if (SelectInst *SI = dyn_cast<SelectInst>(Op0I)) {
3043 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3045 } else if (isa<PHINode>(Op0I)) {
3046 if (Instruction *NV = FoldOpIntoPhi(I))
3050 // See if we can fold away this rem instruction.
3051 if (SimplifyDemandedInstructionBits(I))
3059 Instruction *InstCombiner::visitURem(BinaryOperator &I) {
3060 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3062 if (Instruction *common = commonIRemTransforms(I))
3065 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3066 // X urem C^2 -> X and C
3067 // Check to see if this is an unsigned remainder with an exact power of 2,
3068 // if so, convert to a bitwise and.
3069 if (ConstantInt *C = dyn_cast<ConstantInt>(RHS))
3070 if (C->getValue().isPowerOf2())
3071 return BinaryOperator::CreateAnd(Op0, SubOne(C));
3074 if (Instruction *RHSI = dyn_cast<Instruction>(I.getOperand(1))) {
3075 // Turn A % (C << N), where C is 2^k, into A & ((C << N)-1)
3076 if (RHSI->getOpcode() == Instruction::Shl &&
3077 isa<ConstantInt>(RHSI->getOperand(0))) {
3078 if (cast<ConstantInt>(RHSI->getOperand(0))->getValue().isPowerOf2()) {
3079 Constant *N1 = ConstantInt::getAllOnesValue(I.getType());
3080 Value *Add = InsertNewInstBefore(BinaryOperator::CreateAdd(RHSI, N1,
3082 return BinaryOperator::CreateAnd(Op0, Add);
3087 // urem X, (select Cond, 2^C1, 2^C2) --> select Cond, (and X, C1), (and X, C2)
3088 // where C1&C2 are powers of two.
3089 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
3090 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
3091 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
3092 // STO == 0 and SFO == 0 handled above.
3093 if ((STO->getValue().isPowerOf2()) &&
3094 (SFO->getValue().isPowerOf2())) {
3095 Value *TrueAnd = InsertNewInstBefore(
3096 BinaryOperator::CreateAnd(Op0, SubOne(STO), SI->getName()+".t"), I);
3097 Value *FalseAnd = InsertNewInstBefore(
3098 BinaryOperator::CreateAnd(Op0, SubOne(SFO), SI->getName()+".f"), I);
3099 return SelectInst::Create(SI->getOperand(0), TrueAnd, FalseAnd);
3107 Instruction *InstCombiner::visitSRem(BinaryOperator &I) {
3108 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3110 // Handle the integer rem common cases
3111 if (Instruction *common = commonIRemTransforms(I))
3114 if (Value *RHSNeg = dyn_castNegVal(Op1))
3115 if (!isa<Constant>(RHSNeg) ||
3116 (isa<ConstantInt>(RHSNeg) &&
3117 cast<ConstantInt>(RHSNeg)->getValue().isStrictlyPositive())) {
3119 AddUsesToWorkList(I);
3120 I.setOperand(1, RHSNeg);
3124 // If the sign bits of both operands are zero (i.e. we can prove they are
3125 // unsigned inputs), turn this into a urem.
3126 if (I.getType()->isInteger()) {
3127 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()));
3128 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
3129 // X srem Y -> X urem Y, iff X and Y don't have sign bit set
3130 return BinaryOperator::CreateURem(Op0, Op1, I.getName());
3134 // If it's a constant vector, flip any negative values positive.
3135 if (ConstantVector *RHSV = dyn_cast<ConstantVector>(Op1)) {
3136 unsigned VWidth = RHSV->getNumOperands();
3138 bool hasNegative = false;
3139 for (unsigned i = 0; !hasNegative && i != VWidth; ++i)
3140 if (ConstantInt *RHS = dyn_cast<ConstantInt>(RHSV->getOperand(i)))
3141 if (RHS->getValue().isNegative())
3145 std::vector<Constant *> Elts(VWidth);
3146 for (unsigned i = 0; i != VWidth; ++i) {
3147 if (ConstantInt *RHS = dyn_cast<ConstantInt>(RHSV->getOperand(i))) {
3148 if (RHS->getValue().isNegative())
3149 Elts[i] = cast<ConstantInt>(ConstantExpr::getNeg(RHS));
3155 Constant *NewRHSV = ConstantVector::get(Elts);
3156 if (NewRHSV != RHSV) {
3157 AddUsesToWorkList(I);
3158 I.setOperand(1, NewRHSV);
3167 Instruction *InstCombiner::visitFRem(BinaryOperator &I) {
3168 return commonRemTransforms(I);
3171 // isOneBitSet - Return true if there is exactly one bit set in the specified
3173 static bool isOneBitSet(const ConstantInt *CI) {
3174 return CI->getValue().isPowerOf2();
3177 // isHighOnes - Return true if the constant is of the form 1+0+.
3178 // This is the same as lowones(~X).
3179 static bool isHighOnes(const ConstantInt *CI) {
3180 return (~CI->getValue() + 1).isPowerOf2();
3183 /// getICmpCode - Encode a icmp predicate into a three bit mask. These bits
3184 /// are carefully arranged to allow folding of expressions such as:
3186 /// (A < B) | (A > B) --> (A != B)
3188 /// Note that this is only valid if the first and second predicates have the
3189 /// same sign. Is illegal to do: (A u< B) | (A s> B)
3191 /// Three bits are used to represent the condition, as follows:
3196 /// <=> Value Definition
3197 /// 000 0 Always false
3204 /// 111 7 Always true
3206 static unsigned getICmpCode(const ICmpInst *ICI) {
3207 switch (ICI->getPredicate()) {
3209 case ICmpInst::ICMP_UGT: return 1; // 001
3210 case ICmpInst::ICMP_SGT: return 1; // 001
3211 case ICmpInst::ICMP_EQ: return 2; // 010
3212 case ICmpInst::ICMP_UGE: return 3; // 011
3213 case ICmpInst::ICMP_SGE: return 3; // 011
3214 case ICmpInst::ICMP_ULT: return 4; // 100
3215 case ICmpInst::ICMP_SLT: return 4; // 100
3216 case ICmpInst::ICMP_NE: return 5; // 101
3217 case ICmpInst::ICMP_ULE: return 6; // 110
3218 case ICmpInst::ICMP_SLE: return 6; // 110
3221 assert(0 && "Invalid ICmp predicate!");
3226 /// getFCmpCode - Similar to getICmpCode but for FCmpInst. This encodes a fcmp
3227 /// predicate into a three bit mask. It also returns whether it is an ordered
3228 /// predicate by reference.
3229 static unsigned getFCmpCode(FCmpInst::Predicate CC, bool &isOrdered) {
3232 case FCmpInst::FCMP_ORD: isOrdered = true; return 0; // 000
3233 case FCmpInst::FCMP_UNO: return 0; // 000
3234 case FCmpInst::FCMP_OGT: isOrdered = true; return 1; // 001
3235 case FCmpInst::FCMP_UGT: return 1; // 001
3236 case FCmpInst::FCMP_OEQ: isOrdered = true; return 2; // 010
3237 case FCmpInst::FCMP_UEQ: return 2; // 010
3238 case FCmpInst::FCMP_OGE: isOrdered = true; return 3; // 011
3239 case FCmpInst::FCMP_UGE: return 3; // 011
3240 case FCmpInst::FCMP_OLT: isOrdered = true; return 4; // 100
3241 case FCmpInst::FCMP_ULT: return 4; // 100
3242 case FCmpInst::FCMP_ONE: isOrdered = true; return 5; // 101
3243 case FCmpInst::FCMP_UNE: return 5; // 101
3244 case FCmpInst::FCMP_OLE: isOrdered = true; return 6; // 110
3245 case FCmpInst::FCMP_ULE: return 6; // 110
3248 // Not expecting FCMP_FALSE and FCMP_TRUE;
3249 assert(0 && "Unexpected FCmp predicate!");
3254 /// getICmpValue - This is the complement of getICmpCode, which turns an
3255 /// opcode and two operands into either a constant true or false, or a brand
3256 /// new ICmp instruction. The sign is passed in to determine which kind
3257 /// of predicate to use in the new icmp instruction.
3258 static Value *getICmpValue(bool sign, unsigned code, Value *LHS, Value *RHS) {
3260 default: assert(0 && "Illegal ICmp code!");
3261 case 0: return ConstantInt::getFalse();
3264 return new ICmpInst(ICmpInst::ICMP_SGT, LHS, RHS);
3266 return new ICmpInst(ICmpInst::ICMP_UGT, LHS, RHS);
3267 case 2: return new ICmpInst(ICmpInst::ICMP_EQ, LHS, RHS);
3270 return new ICmpInst(ICmpInst::ICMP_SGE, LHS, RHS);
3272 return new ICmpInst(ICmpInst::ICMP_UGE, LHS, RHS);
3275 return new ICmpInst(ICmpInst::ICMP_SLT, LHS, RHS);
3277 return new ICmpInst(ICmpInst::ICMP_ULT, LHS, RHS);
3278 case 5: return new ICmpInst(ICmpInst::ICMP_NE, LHS, RHS);
3281 return new ICmpInst(ICmpInst::ICMP_SLE, LHS, RHS);
3283 return new ICmpInst(ICmpInst::ICMP_ULE, LHS, RHS);
3284 case 7: return ConstantInt::getTrue();
3288 /// getFCmpValue - This is the complement of getFCmpCode, which turns an
3289 /// opcode and two operands into either a FCmp instruction. isordered is passed
3290 /// in to determine which kind of predicate to use in the new fcmp instruction.
3291 static Value *getFCmpValue(bool isordered, unsigned code,
3292 Value *LHS, Value *RHS) {
3294 default: assert(0 && "Illegal FCmp code!");
3297 return new FCmpInst(FCmpInst::FCMP_ORD, LHS, RHS);
3299 return new FCmpInst(FCmpInst::FCMP_UNO, LHS, RHS);
3302 return new FCmpInst(FCmpInst::FCMP_OGT, LHS, RHS);
3304 return new FCmpInst(FCmpInst::FCMP_UGT, LHS, RHS);
3307 return new FCmpInst(FCmpInst::FCMP_OEQ, LHS, RHS);
3309 return new FCmpInst(FCmpInst::FCMP_UEQ, LHS, RHS);
3312 return new FCmpInst(FCmpInst::FCMP_OGE, LHS, RHS);
3314 return new FCmpInst(FCmpInst::FCMP_UGE, LHS, RHS);
3317 return new FCmpInst(FCmpInst::FCMP_OLT, LHS, RHS);
3319 return new FCmpInst(FCmpInst::FCMP_ULT, LHS, RHS);
3322 return new FCmpInst(FCmpInst::FCMP_ONE, LHS, RHS);
3324 return new FCmpInst(FCmpInst::FCMP_UNE, LHS, RHS);
3327 return new FCmpInst(FCmpInst::FCMP_OLE, LHS, RHS);
3329 return new FCmpInst(FCmpInst::FCMP_ULE, LHS, RHS);
3330 case 7: return ConstantInt::getTrue();
3334 /// PredicatesFoldable - Return true if both predicates match sign or if at
3335 /// least one of them is an equality comparison (which is signless).
3336 static bool PredicatesFoldable(ICmpInst::Predicate p1, ICmpInst::Predicate p2) {
3337 return (ICmpInst::isSignedPredicate(p1) == ICmpInst::isSignedPredicate(p2)) ||
3338 (ICmpInst::isSignedPredicate(p1) && ICmpInst::isEquality(p2)) ||
3339 (ICmpInst::isSignedPredicate(p2) && ICmpInst::isEquality(p1));
3343 // FoldICmpLogical - Implements (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
3344 struct FoldICmpLogical {
3347 ICmpInst::Predicate pred;
3348 FoldICmpLogical(InstCombiner &ic, ICmpInst *ICI)
3349 : IC(ic), LHS(ICI->getOperand(0)), RHS(ICI->getOperand(1)),
3350 pred(ICI->getPredicate()) {}
3351 bool shouldApply(Value *V) const {
3352 if (ICmpInst *ICI = dyn_cast<ICmpInst>(V))
3353 if (PredicatesFoldable(pred, ICI->getPredicate()))
3354 return ((ICI->getOperand(0) == LHS && ICI->getOperand(1) == RHS) ||
3355 (ICI->getOperand(0) == RHS && ICI->getOperand(1) == LHS));
3358 Instruction *apply(Instruction &Log) const {
3359 ICmpInst *ICI = cast<ICmpInst>(Log.getOperand(0));
3360 if (ICI->getOperand(0) != LHS) {
3361 assert(ICI->getOperand(1) == LHS);
3362 ICI->swapOperands(); // Swap the LHS and RHS of the ICmp
3365 ICmpInst *RHSICI = cast<ICmpInst>(Log.getOperand(1));
3366 unsigned LHSCode = getICmpCode(ICI);
3367 unsigned RHSCode = getICmpCode(RHSICI);
3369 switch (Log.getOpcode()) {
3370 case Instruction::And: Code = LHSCode & RHSCode; break;
3371 case Instruction::Or: Code = LHSCode | RHSCode; break;
3372 case Instruction::Xor: Code = LHSCode ^ RHSCode; break;
3373 default: assert(0 && "Illegal logical opcode!"); return 0;
3376 bool isSigned = ICmpInst::isSignedPredicate(RHSICI->getPredicate()) ||
3377 ICmpInst::isSignedPredicate(ICI->getPredicate());
3379 Value *RV = getICmpValue(isSigned, Code, LHS, RHS);
3380 if (Instruction *I = dyn_cast<Instruction>(RV))
3382 // Otherwise, it's a constant boolean value...
3383 return IC.ReplaceInstUsesWith(Log, RV);
3386 } // end anonymous namespace
3388 // OptAndOp - This handles expressions of the form ((val OP C1) & C2). Where
3389 // the Op parameter is 'OP', OpRHS is 'C1', and AndRHS is 'C2'. Op is
3390 // guaranteed to be a binary operator.
3391 Instruction *InstCombiner::OptAndOp(Instruction *Op,
3393 ConstantInt *AndRHS,
3394 BinaryOperator &TheAnd) {
3395 Value *X = Op->getOperand(0);
3396 Constant *Together = 0;
3398 Together = And(AndRHS, OpRHS);
3400 switch (Op->getOpcode()) {
3401 case Instruction::Xor:
3402 if (Op->hasOneUse()) {
3403 // (X ^ C1) & C2 --> (X & C2) ^ (C1&C2)
3404 Instruction *And = BinaryOperator::CreateAnd(X, AndRHS);
3405 InsertNewInstBefore(And, TheAnd);
3407 return BinaryOperator::CreateXor(And, Together);
3410 case Instruction::Or:
3411 if (Together == AndRHS) // (X | C) & C --> C
3412 return ReplaceInstUsesWith(TheAnd, AndRHS);
3414 if (Op->hasOneUse() && Together != OpRHS) {
3415 // (X | C1) & C2 --> (X | (C1&C2)) & C2
3416 Instruction *Or = BinaryOperator::CreateOr(X, Together);
3417 InsertNewInstBefore(Or, TheAnd);
3419 return BinaryOperator::CreateAnd(Or, AndRHS);
3422 case Instruction::Add:
3423 if (Op->hasOneUse()) {
3424 // Adding a one to a single bit bit-field should be turned into an XOR
3425 // of the bit. First thing to check is to see if this AND is with a
3426 // single bit constant.
3427 const APInt& AndRHSV = cast<ConstantInt>(AndRHS)->getValue();
3429 // If there is only one bit set...
3430 if (isOneBitSet(cast<ConstantInt>(AndRHS))) {
3431 // Ok, at this point, we know that we are masking the result of the
3432 // ADD down to exactly one bit. If the constant we are adding has
3433 // no bits set below this bit, then we can eliminate the ADD.
3434 const APInt& AddRHS = cast<ConstantInt>(OpRHS)->getValue();
3436 // Check to see if any bits below the one bit set in AndRHSV are set.
3437 if ((AddRHS & (AndRHSV-1)) == 0) {
3438 // If not, the only thing that can effect the output of the AND is
3439 // the bit specified by AndRHSV. If that bit is set, the effect of
3440 // the XOR is to toggle the bit. If it is clear, then the ADD has
3442 if ((AddRHS & AndRHSV) == 0) { // Bit is not set, noop
3443 TheAnd.setOperand(0, X);
3446 // Pull the XOR out of the AND.
3447 Instruction *NewAnd = BinaryOperator::CreateAnd(X, AndRHS);
3448 InsertNewInstBefore(NewAnd, TheAnd);
3449 NewAnd->takeName(Op);
3450 return BinaryOperator::CreateXor(NewAnd, AndRHS);
3457 case Instruction::Shl: {
3458 // We know that the AND will not produce any of the bits shifted in, so if
3459 // the anded constant includes them, clear them now!
3461 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3462 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3463 APInt ShlMask(APInt::getHighBitsSet(BitWidth, BitWidth-OpRHSVal));
3464 ConstantInt *CI = ConstantInt::get(AndRHS->getValue() & ShlMask);
3466 if (CI->getValue() == ShlMask) {
3467 // Masking out bits that the shift already masks
3468 return ReplaceInstUsesWith(TheAnd, Op); // No need for the and.
3469 } else if (CI != AndRHS) { // Reducing bits set in and.
3470 TheAnd.setOperand(1, CI);
3475 case Instruction::LShr:
3477 // We know that the AND will not produce any of the bits shifted in, so if
3478 // the anded constant includes them, clear them now! This only applies to
3479 // unsigned shifts, because a signed shr may bring in set bits!
3481 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3482 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3483 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
3484 ConstantInt *CI = ConstantInt::get(AndRHS->getValue() & ShrMask);
3486 if (CI->getValue() == ShrMask) {
3487 // Masking out bits that the shift already masks.
3488 return ReplaceInstUsesWith(TheAnd, Op);
3489 } else if (CI != AndRHS) {
3490 TheAnd.setOperand(1, CI); // Reduce bits set in and cst.
3495 case Instruction::AShr:
3497 // See if this is shifting in some sign extension, then masking it out
3499 if (Op->hasOneUse()) {
3500 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3501 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3502 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
3503 Constant *C = ConstantInt::get(AndRHS->getValue() & ShrMask);
3504 if (C == AndRHS) { // Masking out bits shifted in.
3505 // (Val ashr C1) & C2 -> (Val lshr C1) & C2
3506 // Make the argument unsigned.
3507 Value *ShVal = Op->getOperand(0);
3508 ShVal = InsertNewInstBefore(
3509 BinaryOperator::CreateLShr(ShVal, OpRHS,
3510 Op->getName()), TheAnd);
3511 return BinaryOperator::CreateAnd(ShVal, AndRHS, TheAnd.getName());
3520 /// InsertRangeTest - Emit a computation of: (V >= Lo && V < Hi) if Inside is
3521 /// true, otherwise (V < Lo || V >= Hi). In pratice, we emit the more efficient
3522 /// (V-Lo) <u Hi-Lo. This method expects that Lo <= Hi. isSigned indicates
3523 /// whether to treat the V, Lo and HI as signed or not. IB is the location to
3524 /// insert new instructions.
3525 Instruction *InstCombiner::InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
3526 bool isSigned, bool Inside,
3528 assert(cast<ConstantInt>(ConstantExpr::getICmp((isSigned ?
3529 ICmpInst::ICMP_SLE:ICmpInst::ICMP_ULE), Lo, Hi))->getZExtValue() &&
3530 "Lo is not <= Hi in range emission code!");
3533 if (Lo == Hi) // Trivially false.
3534 return new ICmpInst(ICmpInst::ICMP_NE, V, V);
3536 // V >= Min && V < Hi --> V < Hi
3537 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
3538 ICmpInst::Predicate pred = (isSigned ?
3539 ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT);
3540 return new ICmpInst(pred, V, Hi);
3543 // Emit V-Lo <u Hi-Lo
3544 Constant *NegLo = ConstantExpr::getNeg(Lo);
3545 Instruction *Add = BinaryOperator::CreateAdd(V, NegLo, V->getName()+".off");
3546 InsertNewInstBefore(Add, IB);
3547 Constant *UpperBound = ConstantExpr::getAdd(NegLo, Hi);
3548 return new ICmpInst(ICmpInst::ICMP_ULT, Add, UpperBound);
3551 if (Lo == Hi) // Trivially true.
3552 return new ICmpInst(ICmpInst::ICMP_EQ, V, V);
3554 // V < Min || V >= Hi -> V > Hi-1
3555 Hi = SubOne(cast<ConstantInt>(Hi));
3556 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
3557 ICmpInst::Predicate pred = (isSigned ?
3558 ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT);
3559 return new ICmpInst(pred, V, Hi);
3562 // Emit V-Lo >u Hi-1-Lo
3563 // Note that Hi has already had one subtracted from it, above.
3564 ConstantInt *NegLo = cast<ConstantInt>(ConstantExpr::getNeg(Lo));
3565 Instruction *Add = BinaryOperator::CreateAdd(V, NegLo, V->getName()+".off");
3566 InsertNewInstBefore(Add, IB);
3567 Constant *LowerBound = ConstantExpr::getAdd(NegLo, Hi);
3568 return new ICmpInst(ICmpInst::ICMP_UGT, Add, LowerBound);
3571 // isRunOfOnes - Returns true iff Val consists of one contiguous run of 1s with
3572 // any number of 0s on either side. The 1s are allowed to wrap from LSB to
3573 // MSB, so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs. 0x0F0F0000 is
3574 // not, since all 1s are not contiguous.
3575 static bool isRunOfOnes(ConstantInt *Val, uint32_t &MB, uint32_t &ME) {
3576 const APInt& V = Val->getValue();
3577 uint32_t BitWidth = Val->getType()->getBitWidth();
3578 if (!APIntOps::isShiftedMask(BitWidth, V)) return false;
3580 // look for the first zero bit after the run of ones
3581 MB = BitWidth - ((V - 1) ^ V).countLeadingZeros();
3582 // look for the first non-zero bit
3583 ME = V.getActiveBits();
3587 /// FoldLogicalPlusAnd - This is part of an expression (LHS +/- RHS) & Mask,
3588 /// where isSub determines whether the operator is a sub. If we can fold one of
3589 /// the following xforms:
3591 /// ((A & N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == Mask
3592 /// ((A | N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3593 /// ((A ^ N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3595 /// return (A +/- B).
3597 Value *InstCombiner::FoldLogicalPlusAnd(Value *LHS, Value *RHS,
3598 ConstantInt *Mask, bool isSub,
3600 Instruction *LHSI = dyn_cast<Instruction>(LHS);
3601 if (!LHSI || LHSI->getNumOperands() != 2 ||
3602 !isa<ConstantInt>(LHSI->getOperand(1))) return 0;
3604 ConstantInt *N = cast<ConstantInt>(LHSI->getOperand(1));
3606 switch (LHSI->getOpcode()) {
3608 case Instruction::And:
3609 if (And(N, Mask) == Mask) {
3610 // If the AndRHS is a power of two minus one (0+1+), this is simple.
3611 if ((Mask->getValue().countLeadingZeros() +
3612 Mask->getValue().countPopulation()) ==
3613 Mask->getValue().getBitWidth())
3616 // Otherwise, if Mask is 0+1+0+, and if B is known to have the low 0+
3617 // part, we don't need any explicit masks to take them out of A. If that
3618 // is all N is, ignore it.
3619 uint32_t MB = 0, ME = 0;
3620 if (isRunOfOnes(Mask, MB, ME)) { // begin/end bit of run, inclusive
3621 uint32_t BitWidth = cast<IntegerType>(RHS->getType())->getBitWidth();
3622 APInt Mask(APInt::getLowBitsSet(BitWidth, MB-1));
3623 if (MaskedValueIsZero(RHS, Mask))
3628 case Instruction::Or:
3629 case Instruction::Xor:
3630 // If the AndRHS is a power of two minus one (0+1+), and N&Mask == 0
3631 if ((Mask->getValue().countLeadingZeros() +
3632 Mask->getValue().countPopulation()) == Mask->getValue().getBitWidth()
3633 && And(N, Mask)->isZero())
3640 New = BinaryOperator::CreateSub(LHSI->getOperand(0), RHS, "fold");
3642 New = BinaryOperator::CreateAdd(LHSI->getOperand(0), RHS, "fold");
3643 return InsertNewInstBefore(New, I);
3646 /// FoldAndOfICmps - Fold (icmp)&(icmp) if possible.
3647 Instruction *InstCombiner::FoldAndOfICmps(Instruction &I,
3648 ICmpInst *LHS, ICmpInst *RHS) {
3650 ConstantInt *LHSCst, *RHSCst;
3651 ICmpInst::Predicate LHSCC, RHSCC;
3653 // This only handles icmp of constants: (icmp1 A, C1) & (icmp2 B, C2).
3654 if (!match(LHS, m_ICmp(LHSCC, m_Value(Val), m_ConstantInt(LHSCst))) ||
3655 !match(RHS, m_ICmp(RHSCC, m_Value(Val2), m_ConstantInt(RHSCst))))
3658 // (icmp ult A, C) & (icmp ult B, C) --> (icmp ult (A|B), C)
3659 // where C is a power of 2
3660 if (LHSCst == RHSCst && LHSCC == RHSCC && LHSCC == ICmpInst::ICMP_ULT &&
3661 LHSCst->getValue().isPowerOf2()) {
3662 Instruction *NewOr = BinaryOperator::CreateOr(Val, Val2);
3663 InsertNewInstBefore(NewOr, I);
3664 return new ICmpInst(LHSCC, NewOr, LHSCst);
3667 // From here on, we only handle:
3668 // (icmp1 A, C1) & (icmp2 A, C2) --> something simpler.
3669 if (Val != Val2) return 0;
3671 // ICMP_[US][GL]E X, CST is folded to ICMP_[US][GL]T elsewhere.
3672 if (LHSCC == ICmpInst::ICMP_UGE || LHSCC == ICmpInst::ICMP_ULE ||
3673 RHSCC == ICmpInst::ICMP_UGE || RHSCC == ICmpInst::ICMP_ULE ||
3674 LHSCC == ICmpInst::ICMP_SGE || LHSCC == ICmpInst::ICMP_SLE ||
3675 RHSCC == ICmpInst::ICMP_SGE || RHSCC == ICmpInst::ICMP_SLE)
3678 // We can't fold (ugt x, C) & (sgt x, C2).
3679 if (!PredicatesFoldable(LHSCC, RHSCC))
3682 // Ensure that the larger constant is on the RHS.
3684 if (ICmpInst::isSignedPredicate(LHSCC) ||
3685 (ICmpInst::isEquality(LHSCC) &&
3686 ICmpInst::isSignedPredicate(RHSCC)))
3687 ShouldSwap = LHSCst->getValue().sgt(RHSCst->getValue());
3689 ShouldSwap = LHSCst->getValue().ugt(RHSCst->getValue());
3692 std::swap(LHS, RHS);
3693 std::swap(LHSCst, RHSCst);
3694 std::swap(LHSCC, RHSCC);
3697 // At this point, we know we have have two icmp instructions
3698 // comparing a value against two constants and and'ing the result
3699 // together. Because of the above check, we know that we only have
3700 // icmp eq, icmp ne, icmp [su]lt, and icmp [SU]gt here. We also know
3701 // (from the FoldICmpLogical check above), that the two constants
3702 // are not equal and that the larger constant is on the RHS
3703 assert(LHSCst != RHSCst && "Compares not folded above?");
3706 default: assert(0 && "Unknown integer condition code!");
3707 case ICmpInst::ICMP_EQ:
3709 default: assert(0 && "Unknown integer condition code!");
3710 case ICmpInst::ICMP_EQ: // (X == 13 & X == 15) -> false
3711 case ICmpInst::ICMP_UGT: // (X == 13 & X > 15) -> false
3712 case ICmpInst::ICMP_SGT: // (X == 13 & X > 15) -> false
3713 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3714 case ICmpInst::ICMP_NE: // (X == 13 & X != 15) -> X == 13
3715 case ICmpInst::ICMP_ULT: // (X == 13 & X < 15) -> X == 13
3716 case ICmpInst::ICMP_SLT: // (X == 13 & X < 15) -> X == 13
3717 return ReplaceInstUsesWith(I, LHS);
3719 case ICmpInst::ICMP_NE:
3721 default: assert(0 && "Unknown integer condition code!");
3722 case ICmpInst::ICMP_ULT:
3723 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X u< 14) -> X < 13
3724 return new ICmpInst(ICmpInst::ICMP_ULT, Val, LHSCst);
3725 break; // (X != 13 & X u< 15) -> no change
3726 case ICmpInst::ICMP_SLT:
3727 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X s< 14) -> X < 13
3728 return new ICmpInst(ICmpInst::ICMP_SLT, Val, LHSCst);
3729 break; // (X != 13 & X s< 15) -> no change
3730 case ICmpInst::ICMP_EQ: // (X != 13 & X == 15) -> X == 15
3731 case ICmpInst::ICMP_UGT: // (X != 13 & X u> 15) -> X u> 15
3732 case ICmpInst::ICMP_SGT: // (X != 13 & X s> 15) -> X s> 15
3733 return ReplaceInstUsesWith(I, RHS);
3734 case ICmpInst::ICMP_NE:
3735 if (LHSCst == SubOne(RHSCst)){// (X != 13 & X != 14) -> X-13 >u 1
3736 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
3737 Instruction *Add = BinaryOperator::CreateAdd(Val, AddCST,
3738 Val->getName()+".off");
3739 InsertNewInstBefore(Add, I);
3740 return new ICmpInst(ICmpInst::ICMP_UGT, Add,
3741 ConstantInt::get(Add->getType(), 1));
3743 break; // (X != 13 & X != 15) -> no change
3746 case ICmpInst::ICMP_ULT:
3748 default: assert(0 && "Unknown integer condition code!");
3749 case ICmpInst::ICMP_EQ: // (X u< 13 & X == 15) -> false
3750 case ICmpInst::ICMP_UGT: // (X u< 13 & X u> 15) -> false
3751 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3752 case ICmpInst::ICMP_SGT: // (X u< 13 & X s> 15) -> no change
3754 case ICmpInst::ICMP_NE: // (X u< 13 & X != 15) -> X u< 13
3755 case ICmpInst::ICMP_ULT: // (X u< 13 & X u< 15) -> X u< 13
3756 return ReplaceInstUsesWith(I, LHS);
3757 case ICmpInst::ICMP_SLT: // (X u< 13 & X s< 15) -> no change
3761 case ICmpInst::ICMP_SLT:
3763 default: assert(0 && "Unknown integer condition code!");
3764 case ICmpInst::ICMP_EQ: // (X s< 13 & X == 15) -> false
3765 case ICmpInst::ICMP_SGT: // (X s< 13 & X s> 15) -> false
3766 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3767 case ICmpInst::ICMP_UGT: // (X s< 13 & X u> 15) -> no change
3769 case ICmpInst::ICMP_NE: // (X s< 13 & X != 15) -> X < 13
3770 case ICmpInst::ICMP_SLT: // (X s< 13 & X s< 15) -> X < 13
3771 return ReplaceInstUsesWith(I, LHS);
3772 case ICmpInst::ICMP_ULT: // (X s< 13 & X u< 15) -> no change
3776 case ICmpInst::ICMP_UGT:
3778 default: assert(0 && "Unknown integer condition code!");
3779 case ICmpInst::ICMP_EQ: // (X u> 13 & X == 15) -> X == 15
3780 case ICmpInst::ICMP_UGT: // (X u> 13 & X u> 15) -> X u> 15
3781 return ReplaceInstUsesWith(I, RHS);
3782 case ICmpInst::ICMP_SGT: // (X u> 13 & X s> 15) -> no change
3784 case ICmpInst::ICMP_NE:
3785 if (RHSCst == AddOne(LHSCst)) // (X u> 13 & X != 14) -> X u> 14
3786 return new ICmpInst(LHSCC, Val, RHSCst);
3787 break; // (X u> 13 & X != 15) -> no change
3788 case ICmpInst::ICMP_ULT: // (X u> 13 & X u< 15) -> (X-14) <u 1
3789 return InsertRangeTest(Val, AddOne(LHSCst), RHSCst, false, true, I);
3790 case ICmpInst::ICMP_SLT: // (X u> 13 & X s< 15) -> no change
3794 case ICmpInst::ICMP_SGT:
3796 default: assert(0 && "Unknown integer condition code!");
3797 case ICmpInst::ICMP_EQ: // (X s> 13 & X == 15) -> X == 15
3798 case ICmpInst::ICMP_SGT: // (X s> 13 & X s> 15) -> X s> 15
3799 return ReplaceInstUsesWith(I, RHS);
3800 case ICmpInst::ICMP_UGT: // (X s> 13 & X u> 15) -> no change
3802 case ICmpInst::ICMP_NE:
3803 if (RHSCst == AddOne(LHSCst)) // (X s> 13 & X != 14) -> X s> 14
3804 return new ICmpInst(LHSCC, Val, RHSCst);
3805 break; // (X s> 13 & X != 15) -> no change
3806 case ICmpInst::ICMP_SLT: // (X s> 13 & X s< 15) -> (X-14) s< 1
3807 return InsertRangeTest(Val, AddOne(LHSCst), RHSCst, true, true, I);
3808 case ICmpInst::ICMP_ULT: // (X s> 13 & X u< 15) -> no change
3818 Instruction *InstCombiner::visitAnd(BinaryOperator &I) {
3819 bool Changed = SimplifyCommutative(I);
3820 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3822 if (isa<UndefValue>(Op1)) // X & undef -> 0
3823 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3827 return ReplaceInstUsesWith(I, Op1);
3829 // See if we can simplify any instructions used by the instruction whose sole
3830 // purpose is to compute bits we don't care about.
3831 if (!isa<VectorType>(I.getType())) {
3832 if (SimplifyDemandedInstructionBits(I))
3835 if (ConstantVector *CP = dyn_cast<ConstantVector>(Op1)) {
3836 if (CP->isAllOnesValue()) // X & <-1,-1> -> X
3837 return ReplaceInstUsesWith(I, I.getOperand(0));
3838 } else if (isa<ConstantAggregateZero>(Op1)) {
3839 return ReplaceInstUsesWith(I, Op1); // X & <0,0> -> <0,0>
3843 if (ConstantInt *AndRHS = dyn_cast<ConstantInt>(Op1)) {
3844 const APInt& AndRHSMask = AndRHS->getValue();
3845 APInt NotAndRHS(~AndRHSMask);
3847 // Optimize a variety of ((val OP C1) & C2) combinations...
3848 if (isa<BinaryOperator>(Op0)) {
3849 Instruction *Op0I = cast<Instruction>(Op0);
3850 Value *Op0LHS = Op0I->getOperand(0);
3851 Value *Op0RHS = Op0I->getOperand(1);
3852 switch (Op0I->getOpcode()) {
3853 case Instruction::Xor:
3854 case Instruction::Or:
3855 // If the mask is only needed on one incoming arm, push it up.
3856 if (Op0I->hasOneUse()) {
3857 if (MaskedValueIsZero(Op0LHS, NotAndRHS)) {
3858 // Not masking anything out for the LHS, move to RHS.
3859 Instruction *NewRHS = BinaryOperator::CreateAnd(Op0RHS, AndRHS,
3860 Op0RHS->getName()+".masked");
3861 InsertNewInstBefore(NewRHS, I);
3862 return BinaryOperator::Create(
3863 cast<BinaryOperator>(Op0I)->getOpcode(), Op0LHS, NewRHS);
3865 if (!isa<Constant>(Op0RHS) &&
3866 MaskedValueIsZero(Op0RHS, NotAndRHS)) {
3867 // Not masking anything out for the RHS, move to LHS.
3868 Instruction *NewLHS = BinaryOperator::CreateAnd(Op0LHS, AndRHS,
3869 Op0LHS->getName()+".masked");
3870 InsertNewInstBefore(NewLHS, I);
3871 return BinaryOperator::Create(
3872 cast<BinaryOperator>(Op0I)->getOpcode(), NewLHS, Op0RHS);
3877 case Instruction::Add:
3878 // ((A & N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == AndRHS.
3879 // ((A | N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
3880 // ((A ^ N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
3881 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, false, I))
3882 return BinaryOperator::CreateAnd(V, AndRHS);
3883 if (Value *V = FoldLogicalPlusAnd(Op0RHS, Op0LHS, AndRHS, false, I))
3884 return BinaryOperator::CreateAnd(V, AndRHS); // Add commutes
3887 case Instruction::Sub:
3888 // ((A & N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == AndRHS.
3889 // ((A | N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
3890 // ((A ^ N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
3891 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, true, I))
3892 return BinaryOperator::CreateAnd(V, AndRHS);
3894 // (A - N) & AndRHS -> -N & AndRHS iff A&AndRHS==0 and AndRHS
3895 // has 1's for all bits that the subtraction with A might affect.
3896 if (Op0I->hasOneUse()) {
3897 uint32_t BitWidth = AndRHSMask.getBitWidth();
3898 uint32_t Zeros = AndRHSMask.countLeadingZeros();
3899 APInt Mask = APInt::getLowBitsSet(BitWidth, BitWidth - Zeros);
3901 ConstantInt *A = dyn_cast<ConstantInt>(Op0LHS);
3902 if (!(A && A->isZero()) && // avoid infinite recursion.
3903 MaskedValueIsZero(Op0LHS, Mask)) {
3904 Instruction *NewNeg = BinaryOperator::CreateNeg(Op0RHS);
3905 InsertNewInstBefore(NewNeg, I);
3906 return BinaryOperator::CreateAnd(NewNeg, AndRHS);
3911 case Instruction::Shl:
3912 case Instruction::LShr:
3913 // (1 << x) & 1 --> zext(x == 0)
3914 // (1 >> x) & 1 --> zext(x == 0)
3915 if (AndRHSMask == 1 && Op0LHS == AndRHS) {
3916 Instruction *NewICmp = new ICmpInst(ICmpInst::ICMP_EQ, Op0RHS,
3917 Constant::getNullValue(I.getType()));
3918 InsertNewInstBefore(NewICmp, I);
3919 return new ZExtInst(NewICmp, I.getType());
3924 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
3925 if (Instruction *Res = OptAndOp(Op0I, Op0CI, AndRHS, I))
3927 } else if (CastInst *CI = dyn_cast<CastInst>(Op0)) {
3928 // If this is an integer truncation or change from signed-to-unsigned, and
3929 // if the source is an and/or with immediate, transform it. This
3930 // frequently occurs for bitfield accesses.
3931 if (Instruction *CastOp = dyn_cast<Instruction>(CI->getOperand(0))) {
3932 if ((isa<TruncInst>(CI) || isa<BitCastInst>(CI)) &&
3933 CastOp->getNumOperands() == 2)
3934 if (ConstantInt *AndCI = dyn_cast<ConstantInt>(CastOp->getOperand(1))) {
3935 if (CastOp->getOpcode() == Instruction::And) {
3936 // Change: and (cast (and X, C1) to T), C2
3937 // into : and (cast X to T), trunc_or_bitcast(C1)&C2
3938 // This will fold the two constants together, which may allow
3939 // other simplifications.
3940 Instruction *NewCast = CastInst::CreateTruncOrBitCast(
3941 CastOp->getOperand(0), I.getType(),
3942 CastOp->getName()+".shrunk");
3943 NewCast = InsertNewInstBefore(NewCast, I);
3944 // trunc_or_bitcast(C1)&C2
3945 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
3946 C3 = ConstantExpr::getAnd(C3, AndRHS);
3947 return BinaryOperator::CreateAnd(NewCast, C3);
3948 } else if (CastOp->getOpcode() == Instruction::Or) {
3949 // Change: and (cast (or X, C1) to T), C2
3950 // into : trunc(C1)&C2 iff trunc(C1)&C2 == C2
3951 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
3952 if (ConstantExpr::getAnd(C3, AndRHS) == AndRHS) // trunc(C1)&C2
3953 return ReplaceInstUsesWith(I, AndRHS);
3959 // Try to fold constant and into select arguments.
3960 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3961 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3963 if (isa<PHINode>(Op0))
3964 if (Instruction *NV = FoldOpIntoPhi(I))
3968 Value *Op0NotVal = dyn_castNotVal(Op0);
3969 Value *Op1NotVal = dyn_castNotVal(Op1);
3971 if (Op0NotVal == Op1 || Op1NotVal == Op0) // A & ~A == ~A & A == 0
3972 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3974 // (~A & ~B) == (~(A | B)) - De Morgan's Law
3975 if (Op0NotVal && Op1NotVal && isOnlyUse(Op0) && isOnlyUse(Op1)) {
3976 Instruction *Or = BinaryOperator::CreateOr(Op0NotVal, Op1NotVal,
3977 I.getName()+".demorgan");
3978 InsertNewInstBefore(Or, I);
3979 return BinaryOperator::CreateNot(Or);
3983 Value *A = 0, *B = 0, *C = 0, *D = 0;
3984 if (match(Op0, m_Or(m_Value(A), m_Value(B)))) {
3985 if (A == Op1 || B == Op1) // (A | ?) & A --> A
3986 return ReplaceInstUsesWith(I, Op1);
3988 // (A|B) & ~(A&B) -> A^B
3989 if (match(Op1, m_Not(m_And(m_Value(C), m_Value(D))))) {
3990 if ((A == C && B == D) || (A == D && B == C))
3991 return BinaryOperator::CreateXor(A, B);
3995 if (match(Op1, m_Or(m_Value(A), m_Value(B)))) {
3996 if (A == Op0 || B == Op0) // A & (A | ?) --> A
3997 return ReplaceInstUsesWith(I, Op0);
3999 // ~(A&B) & (A|B) -> A^B
4000 if (match(Op0, m_Not(m_And(m_Value(C), m_Value(D))))) {
4001 if ((A == C && B == D) || (A == D && B == C))
4002 return BinaryOperator::CreateXor(A, B);
4006 if (Op0->hasOneUse() &&
4007 match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
4008 if (A == Op1) { // (A^B)&A -> A&(A^B)
4009 I.swapOperands(); // Simplify below
4010 std::swap(Op0, Op1);
4011 } else if (B == Op1) { // (A^B)&B -> B&(B^A)
4012 cast<BinaryOperator>(Op0)->swapOperands();
4013 I.swapOperands(); // Simplify below
4014 std::swap(Op0, Op1);
4018 if (Op1->hasOneUse() &&
4019 match(Op1, m_Xor(m_Value(A), m_Value(B)))) {
4020 if (B == Op0) { // B&(A^B) -> B&(B^A)
4021 cast<BinaryOperator>(Op1)->swapOperands();
4024 if (A == Op0) { // A&(A^B) -> A & ~B
4025 Instruction *NotB = BinaryOperator::CreateNot(B, "tmp");
4026 InsertNewInstBefore(NotB, I);
4027 return BinaryOperator::CreateAnd(A, NotB);
4031 // (A&((~A)|B)) -> A&B
4032 if (match(Op0, m_Or(m_Not(m_Specific(Op1)), m_Value(A))) ||
4033 match(Op0, m_Or(m_Value(A), m_Not(m_Specific(Op1)))))
4034 return BinaryOperator::CreateAnd(A, Op1);
4035 if (match(Op1, m_Or(m_Not(m_Specific(Op0)), m_Value(A))) ||
4036 match(Op1, m_Or(m_Value(A), m_Not(m_Specific(Op0)))))
4037 return BinaryOperator::CreateAnd(A, Op0);
4040 if (ICmpInst *RHS = dyn_cast<ICmpInst>(Op1)) {
4041 // (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
4042 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
4045 if (ICmpInst *LHS = dyn_cast<ICmpInst>(Op0))
4046 if (Instruction *Res = FoldAndOfICmps(I, LHS, RHS))
4050 // fold (and (cast A), (cast B)) -> (cast (and A, B))
4051 if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
4052 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4053 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind ?
4054 const Type *SrcTy = Op0C->getOperand(0)->getType();
4055 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
4056 // Only do this if the casts both really cause code to be generated.
4057 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4059 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4061 Instruction *NewOp = BinaryOperator::CreateAnd(Op0C->getOperand(0),
4062 Op1C->getOperand(0),
4064 InsertNewInstBefore(NewOp, I);
4065 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
4069 // (X >> Z) & (Y >> Z) -> (X&Y) >> Z for all shifts.
4070 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
4071 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
4072 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
4073 SI0->getOperand(1) == SI1->getOperand(1) &&
4074 (SI0->hasOneUse() || SI1->hasOneUse())) {
4075 Instruction *NewOp =
4076 InsertNewInstBefore(BinaryOperator::CreateAnd(SI0->getOperand(0),
4078 SI0->getName()), I);
4079 return BinaryOperator::Create(SI1->getOpcode(), NewOp,
4080 SI1->getOperand(1));
4084 // If and'ing two fcmp, try combine them into one.
4085 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0))) {
4086 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1))) {
4087 if (LHS->getPredicate() == FCmpInst::FCMP_ORD &&
4088 RHS->getPredicate() == FCmpInst::FCMP_ORD) {
4089 // (fcmp ord x, c) & (fcmp ord y, c) -> (fcmp ord x, y)
4090 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
4091 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
4092 // If either of the constants are nans, then the whole thing returns
4094 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
4095 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4096 return new FCmpInst(FCmpInst::FCMP_ORD, LHS->getOperand(0),
4097 RHS->getOperand(0));
4100 Value *Op0LHS, *Op0RHS, *Op1LHS, *Op1RHS;
4101 FCmpInst::Predicate Op0CC, Op1CC;
4102 if (match(Op0, m_FCmp(Op0CC, m_Value(Op0LHS), m_Value(Op0RHS))) &&
4103 match(Op1, m_FCmp(Op1CC, m_Value(Op1LHS), m_Value(Op1RHS)))) {
4104 if (Op0LHS == Op1RHS && Op0RHS == Op1LHS) {
4105 // Swap RHS operands to match LHS.
4106 Op1CC = FCmpInst::getSwappedPredicate(Op1CC);
4107 std::swap(Op1LHS, Op1RHS);
4109 if (Op0LHS == Op1LHS && Op0RHS == Op1RHS) {
4110 // Simplify (fcmp cc0 x, y) & (fcmp cc1 x, y).
4112 return new FCmpInst((FCmpInst::Predicate)Op0CC, Op0LHS, Op0RHS);
4113 else if (Op0CC == FCmpInst::FCMP_FALSE ||
4114 Op1CC == FCmpInst::FCMP_FALSE)
4115 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4116 else if (Op0CC == FCmpInst::FCMP_TRUE)
4117 return ReplaceInstUsesWith(I, Op1);
4118 else if (Op1CC == FCmpInst::FCMP_TRUE)
4119 return ReplaceInstUsesWith(I, Op0);
4122 unsigned Op0Pred = getFCmpCode(Op0CC, Op0Ordered);
4123 unsigned Op1Pred = getFCmpCode(Op1CC, Op1Ordered);
4125 std::swap(Op0, Op1);
4126 std::swap(Op0Pred, Op1Pred);
4127 std::swap(Op0Ordered, Op1Ordered);
4130 // uno && ueq -> uno && (uno || eq) -> ueq
4131 // ord && olt -> ord && (ord && lt) -> olt
4132 if (Op0Ordered == Op1Ordered)
4133 return ReplaceInstUsesWith(I, Op1);
4134 // uno && oeq -> uno && (ord && eq) -> false
4135 // uno && ord -> false
4137 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4138 // ord && ueq -> ord && (uno || eq) -> oeq
4139 return cast<Instruction>(getFCmpValue(true, Op1Pred,
4148 return Changed ? &I : 0;
4151 /// CollectBSwapParts - Analyze the specified subexpression and see if it is
4152 /// capable of providing pieces of a bswap. The subexpression provides pieces
4153 /// of a bswap if it is proven that each of the non-zero bytes in the output of
4154 /// the expression came from the corresponding "byte swapped" byte in some other
4155 /// value. For example, if the current subexpression is "(shl i32 %X, 24)" then
4156 /// we know that the expression deposits the low byte of %X into the high byte
4157 /// of the bswap result and that all other bytes are zero. This expression is
4158 /// accepted, the high byte of ByteValues is set to X to indicate a correct
4161 /// This function returns true if the match was unsuccessful and false if so.
4162 /// On entry to the function the "OverallLeftShift" is a signed integer value
4163 /// indicating the number of bytes that the subexpression is later shifted. For
4164 /// example, if the expression is later right shifted by 16 bits, the
4165 /// OverallLeftShift value would be -2 on entry. This is used to specify which
4166 /// byte of ByteValues is actually being set.
4168 /// Similarly, ByteMask is a bitmask where a bit is clear if its corresponding
4169 /// byte is masked to zero by a user. For example, in (X & 255), X will be
4170 /// processed with a bytemask of 1. Because bytemask is 32-bits, this limits
4171 /// this function to working on up to 32-byte (256 bit) values. ByteMask is
4172 /// always in the local (OverallLeftShift) coordinate space.
4174 static bool CollectBSwapParts(Value *V, int OverallLeftShift, uint32_t ByteMask,
4175 SmallVector<Value*, 8> &ByteValues) {
4176 if (Instruction *I = dyn_cast<Instruction>(V)) {
4177 // If this is an or instruction, it may be an inner node of the bswap.
4178 if (I->getOpcode() == Instruction::Or) {
4179 return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
4181 CollectBSwapParts(I->getOperand(1), OverallLeftShift, ByteMask,
4185 // If this is a logical shift by a constant multiple of 8, recurse with
4186 // OverallLeftShift and ByteMask adjusted.
4187 if (I->isLogicalShift() && isa<ConstantInt>(I->getOperand(1))) {
4189 cast<ConstantInt>(I->getOperand(1))->getLimitedValue(~0U);
4190 // Ensure the shift amount is defined and of a byte value.
4191 if ((ShAmt & 7) || (ShAmt > 8*ByteValues.size()))
4194 unsigned ByteShift = ShAmt >> 3;
4195 if (I->getOpcode() == Instruction::Shl) {
4196 // X << 2 -> collect(X, +2)
4197 OverallLeftShift += ByteShift;
4198 ByteMask >>= ByteShift;
4200 // X >>u 2 -> collect(X, -2)
4201 OverallLeftShift -= ByteShift;
4202 ByteMask <<= ByteShift;
4203 ByteMask &= (~0U >> (32-ByteValues.size()));
4206 if (OverallLeftShift >= (int)ByteValues.size()) return true;
4207 if (OverallLeftShift <= -(int)ByteValues.size()) return true;
4209 return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
4213 // If this is a logical 'and' with a mask that clears bytes, clear the
4214 // corresponding bytes in ByteMask.
4215 if (I->getOpcode() == Instruction::And &&
4216 isa<ConstantInt>(I->getOperand(1))) {
4217 // Scan every byte of the and mask, seeing if the byte is either 0 or 255.
4218 unsigned NumBytes = ByteValues.size();
4219 APInt Byte(I->getType()->getPrimitiveSizeInBits(), 255);
4220 const APInt &AndMask = cast<ConstantInt>(I->getOperand(1))->getValue();
4222 for (unsigned i = 0; i != NumBytes; ++i, Byte <<= 8) {
4223 // If this byte is masked out by a later operation, we don't care what
4225 if ((ByteMask & (1 << i)) == 0)
4228 // If the AndMask is all zeros for this byte, clear the bit.
4229 APInt MaskB = AndMask & Byte;
4231 ByteMask &= ~(1U << i);
4235 // If the AndMask is not all ones for this byte, it's not a bytezap.
4239 // Otherwise, this byte is kept.
4242 return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
4247 // Okay, we got to something that isn't a shift, 'or' or 'and'. This must be
4248 // the input value to the bswap. Some observations: 1) if more than one byte
4249 // is demanded from this input, then it could not be successfully assembled
4250 // into a byteswap. At least one of the two bytes would not be aligned with
4251 // their ultimate destination.
4252 if (!isPowerOf2_32(ByteMask)) return true;
4253 unsigned InputByteNo = CountTrailingZeros_32(ByteMask);
4255 // 2) The input and ultimate destinations must line up: if byte 3 of an i32
4256 // is demanded, it needs to go into byte 0 of the result. This means that the
4257 // byte needs to be shifted until it lands in the right byte bucket. The
4258 // shift amount depends on the position: if the byte is coming from the high
4259 // part of the value (e.g. byte 3) then it must be shifted right. If from the
4260 // low part, it must be shifted left.
4261 unsigned DestByteNo = InputByteNo + OverallLeftShift;
4262 if (InputByteNo < ByteValues.size()/2) {
4263 if (ByteValues.size()-1-DestByteNo != InputByteNo)
4266 if (ByteValues.size()-1-DestByteNo != InputByteNo)
4270 // If the destination byte value is already defined, the values are or'd
4271 // together, which isn't a bswap (unless it's an or of the same bits).
4272 if (ByteValues[DestByteNo] && ByteValues[DestByteNo] != V)
4274 ByteValues[DestByteNo] = V;
4278 /// MatchBSwap - Given an OR instruction, check to see if this is a bswap idiom.
4279 /// If so, insert the new bswap intrinsic and return it.
4280 Instruction *InstCombiner::MatchBSwap(BinaryOperator &I) {
4281 const IntegerType *ITy = dyn_cast<IntegerType>(I.getType());
4282 if (!ITy || ITy->getBitWidth() % 16 ||
4283 // ByteMask only allows up to 32-byte values.
4284 ITy->getBitWidth() > 32*8)
4285 return 0; // Can only bswap pairs of bytes. Can't do vectors.
4287 /// ByteValues - For each byte of the result, we keep track of which value
4288 /// defines each byte.
4289 SmallVector<Value*, 8> ByteValues;
4290 ByteValues.resize(ITy->getBitWidth()/8);
4292 // Try to find all the pieces corresponding to the bswap.
4293 uint32_t ByteMask = ~0U >> (32-ByteValues.size());
4294 if (CollectBSwapParts(&I, 0, ByteMask, ByteValues))
4297 // Check to see if all of the bytes come from the same value.
4298 Value *V = ByteValues[0];
4299 if (V == 0) return 0; // Didn't find a byte? Must be zero.
4301 // Check to make sure that all of the bytes come from the same value.
4302 for (unsigned i = 1, e = ByteValues.size(); i != e; ++i)
4303 if (ByteValues[i] != V)
4305 const Type *Tys[] = { ITy };
4306 Module *M = I.getParent()->getParent()->getParent();
4307 Function *F = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 1);
4308 return CallInst::Create(F, V);
4311 /// MatchSelectFromAndOr - We have an expression of the form (A&C)|(B&D). Check
4312 /// If A is (cond?-1:0) and either B or D is ~(cond?-1,0) or (cond?0,-1), then
4313 /// we can simplify this expression to "cond ? C : D or B".
4314 static Instruction *MatchSelectFromAndOr(Value *A, Value *B,
4315 Value *C, Value *D) {
4316 // If A is not a select of -1/0, this cannot match.
4318 if (!match(A, m_SelectCst<-1, 0>(m_Value(Cond))))
4321 // ((cond?-1:0)&C) | (B&(cond?0:-1)) -> cond ? C : B.
4322 if (match(D, m_SelectCst<0, -1>(m_Specific(Cond))))
4323 return SelectInst::Create(Cond, C, B);
4324 if (match(D, m_Not(m_SelectCst<-1, 0>(m_Specific(Cond)))))
4325 return SelectInst::Create(Cond, C, B);
4326 // ((cond?-1:0)&C) | ((cond?0:-1)&D) -> cond ? C : D.
4327 if (match(B, m_SelectCst<0, -1>(m_Specific(Cond))))
4328 return SelectInst::Create(Cond, C, D);
4329 if (match(B, m_Not(m_SelectCst<-1, 0>(m_Specific(Cond)))))
4330 return SelectInst::Create(Cond, C, D);
4334 /// FoldOrOfICmps - Fold (icmp)|(icmp) if possible.
4335 Instruction *InstCombiner::FoldOrOfICmps(Instruction &I,
4336 ICmpInst *LHS, ICmpInst *RHS) {
4338 ConstantInt *LHSCst, *RHSCst;
4339 ICmpInst::Predicate LHSCC, RHSCC;
4341 // This only handles icmp of constants: (icmp1 A, C1) | (icmp2 B, C2).
4342 if (!match(LHS, m_ICmp(LHSCC, m_Value(Val), m_ConstantInt(LHSCst))) ||
4343 !match(RHS, m_ICmp(RHSCC, m_Value(Val2), m_ConstantInt(RHSCst))))
4346 // From here on, we only handle:
4347 // (icmp1 A, C1) | (icmp2 A, C2) --> something simpler.
4348 if (Val != Val2) return 0;
4350 // ICMP_[US][GL]E X, CST is folded to ICMP_[US][GL]T elsewhere.
4351 if (LHSCC == ICmpInst::ICMP_UGE || LHSCC == ICmpInst::ICMP_ULE ||
4352 RHSCC == ICmpInst::ICMP_UGE || RHSCC == ICmpInst::ICMP_ULE ||
4353 LHSCC == ICmpInst::ICMP_SGE || LHSCC == ICmpInst::ICMP_SLE ||
4354 RHSCC == ICmpInst::ICMP_SGE || RHSCC == ICmpInst::ICMP_SLE)
4357 // We can't fold (ugt x, C) | (sgt x, C2).
4358 if (!PredicatesFoldable(LHSCC, RHSCC))
4361 // Ensure that the larger constant is on the RHS.
4363 if (ICmpInst::isSignedPredicate(LHSCC) ||
4364 (ICmpInst::isEquality(LHSCC) &&
4365 ICmpInst::isSignedPredicate(RHSCC)))
4366 ShouldSwap = LHSCst->getValue().sgt(RHSCst->getValue());
4368 ShouldSwap = LHSCst->getValue().ugt(RHSCst->getValue());
4371 std::swap(LHS, RHS);
4372 std::swap(LHSCst, RHSCst);
4373 std::swap(LHSCC, RHSCC);
4376 // At this point, we know we have have two icmp instructions
4377 // comparing a value against two constants and or'ing the result
4378 // together. Because of the above check, we know that we only have
4379 // ICMP_EQ, ICMP_NE, ICMP_LT, and ICMP_GT here. We also know (from the
4380 // FoldICmpLogical check above), that the two constants are not
4382 assert(LHSCst != RHSCst && "Compares not folded above?");
4385 default: assert(0 && "Unknown integer condition code!");
4386 case ICmpInst::ICMP_EQ:
4388 default: assert(0 && "Unknown integer condition code!");
4389 case ICmpInst::ICMP_EQ:
4390 if (LHSCst == SubOne(RHSCst)) { // (X == 13 | X == 14) -> X-13 <u 2
4391 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
4392 Instruction *Add = BinaryOperator::CreateAdd(Val, AddCST,
4393 Val->getName()+".off");
4394 InsertNewInstBefore(Add, I);
4395 AddCST = Subtract(AddOne(RHSCst), LHSCst);
4396 return new ICmpInst(ICmpInst::ICMP_ULT, Add, AddCST);
4398 break; // (X == 13 | X == 15) -> no change
4399 case ICmpInst::ICMP_UGT: // (X == 13 | X u> 14) -> no change
4400 case ICmpInst::ICMP_SGT: // (X == 13 | X s> 14) -> no change
4402 case ICmpInst::ICMP_NE: // (X == 13 | X != 15) -> X != 15
4403 case ICmpInst::ICMP_ULT: // (X == 13 | X u< 15) -> X u< 15
4404 case ICmpInst::ICMP_SLT: // (X == 13 | X s< 15) -> X s< 15
4405 return ReplaceInstUsesWith(I, RHS);
4408 case ICmpInst::ICMP_NE:
4410 default: assert(0 && "Unknown integer condition code!");
4411 case ICmpInst::ICMP_EQ: // (X != 13 | X == 15) -> X != 13
4412 case ICmpInst::ICMP_UGT: // (X != 13 | X u> 15) -> X != 13
4413 case ICmpInst::ICMP_SGT: // (X != 13 | X s> 15) -> X != 13
4414 return ReplaceInstUsesWith(I, LHS);
4415 case ICmpInst::ICMP_NE: // (X != 13 | X != 15) -> true
4416 case ICmpInst::ICMP_ULT: // (X != 13 | X u< 15) -> true
4417 case ICmpInst::ICMP_SLT: // (X != 13 | X s< 15) -> true
4418 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4421 case ICmpInst::ICMP_ULT:
4423 default: assert(0 && "Unknown integer condition code!");
4424 case ICmpInst::ICMP_EQ: // (X u< 13 | X == 14) -> no change
4426 case ICmpInst::ICMP_UGT: // (X u< 13 | X u> 15) -> (X-13) u> 2
4427 // If RHSCst is [us]MAXINT, it is always false. Not handling
4428 // this can cause overflow.
4429 if (RHSCst->isMaxValue(false))
4430 return ReplaceInstUsesWith(I, LHS);
4431 return InsertRangeTest(Val, LHSCst, AddOne(RHSCst), false, false, I);
4432 case ICmpInst::ICMP_SGT: // (X u< 13 | X s> 15) -> no change
4434 case ICmpInst::ICMP_NE: // (X u< 13 | X != 15) -> X != 15
4435 case ICmpInst::ICMP_ULT: // (X u< 13 | X u< 15) -> X u< 15
4436 return ReplaceInstUsesWith(I, RHS);
4437 case ICmpInst::ICMP_SLT: // (X u< 13 | X s< 15) -> no change
4441 case ICmpInst::ICMP_SLT:
4443 default: assert(0 && "Unknown integer condition code!");
4444 case ICmpInst::ICMP_EQ: // (X s< 13 | X == 14) -> no change
4446 case ICmpInst::ICMP_SGT: // (X s< 13 | X s> 15) -> (X-13) s> 2
4447 // If RHSCst is [us]MAXINT, it is always false. Not handling
4448 // this can cause overflow.
4449 if (RHSCst->isMaxValue(true))
4450 return ReplaceInstUsesWith(I, LHS);
4451 return InsertRangeTest(Val, LHSCst, AddOne(RHSCst), true, false, I);
4452 case ICmpInst::ICMP_UGT: // (X s< 13 | X u> 15) -> no change
4454 case ICmpInst::ICMP_NE: // (X s< 13 | X != 15) -> X != 15
4455 case ICmpInst::ICMP_SLT: // (X s< 13 | X s< 15) -> X s< 15
4456 return ReplaceInstUsesWith(I, RHS);
4457 case ICmpInst::ICMP_ULT: // (X s< 13 | X u< 15) -> no change
4461 case ICmpInst::ICMP_UGT:
4463 default: assert(0 && "Unknown integer condition code!");
4464 case ICmpInst::ICMP_EQ: // (X u> 13 | X == 15) -> X u> 13
4465 case ICmpInst::ICMP_UGT: // (X u> 13 | X u> 15) -> X u> 13
4466 return ReplaceInstUsesWith(I, LHS);
4467 case ICmpInst::ICMP_SGT: // (X u> 13 | X s> 15) -> no change
4469 case ICmpInst::ICMP_NE: // (X u> 13 | X != 15) -> true
4470 case ICmpInst::ICMP_ULT: // (X u> 13 | X u< 15) -> true
4471 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4472 case ICmpInst::ICMP_SLT: // (X u> 13 | X s< 15) -> no change
4476 case ICmpInst::ICMP_SGT:
4478 default: assert(0 && "Unknown integer condition code!");
4479 case ICmpInst::ICMP_EQ: // (X s> 13 | X == 15) -> X > 13
4480 case ICmpInst::ICMP_SGT: // (X s> 13 | X s> 15) -> X > 13
4481 return ReplaceInstUsesWith(I, LHS);
4482 case ICmpInst::ICMP_UGT: // (X s> 13 | X u> 15) -> no change
4484 case ICmpInst::ICMP_NE: // (X s> 13 | X != 15) -> true
4485 case ICmpInst::ICMP_SLT: // (X s> 13 | X s< 15) -> true
4486 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4487 case ICmpInst::ICMP_ULT: // (X s> 13 | X u< 15) -> no change
4495 /// FoldOrWithConstants - This helper function folds:
4497 /// ((A | B) & C1) | (B & C2)
4503 /// when the XOR of the two constants is "all ones" (-1).
4504 Instruction *InstCombiner::FoldOrWithConstants(BinaryOperator &I, Value *Op,
4505 Value *A, Value *B, Value *C) {
4506 ConstantInt *CI1 = dyn_cast<ConstantInt>(C);
4510 ConstantInt *CI2 = 0;
4511 if (!match(Op, m_And(m_Value(V1), m_ConstantInt(CI2)))) return 0;
4513 APInt Xor = CI1->getValue() ^ CI2->getValue();
4514 if (!Xor.isAllOnesValue()) return 0;
4516 if (V1 == A || V1 == B) {
4517 Instruction *NewOp =
4518 InsertNewInstBefore(BinaryOperator::CreateAnd((V1 == A) ? B : A, CI1), I);
4519 return BinaryOperator::CreateOr(NewOp, V1);
4525 Instruction *InstCombiner::visitOr(BinaryOperator &I) {
4526 bool Changed = SimplifyCommutative(I);
4527 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4529 if (isa<UndefValue>(Op1)) // X | undef -> -1
4530 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4534 return ReplaceInstUsesWith(I, Op0);
4536 // See if we can simplify any instructions used by the instruction whose sole
4537 // purpose is to compute bits we don't care about.
4538 if (!isa<VectorType>(I.getType())) {
4539 if (SimplifyDemandedInstructionBits(I))
4541 } else if (isa<ConstantAggregateZero>(Op1)) {
4542 return ReplaceInstUsesWith(I, Op0); // X | <0,0> -> X
4543 } else if (ConstantVector *CP = dyn_cast<ConstantVector>(Op1)) {
4544 if (CP->isAllOnesValue()) // X | <-1,-1> -> <-1,-1>
4545 return ReplaceInstUsesWith(I, I.getOperand(1));
4551 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
4552 ConstantInt *C1 = 0; Value *X = 0;
4553 // (X & C1) | C2 --> (X | C2) & (C1|C2)
4554 if (match(Op0, m_And(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
4555 Instruction *Or = BinaryOperator::CreateOr(X, RHS);
4556 InsertNewInstBefore(Or, I);
4558 return BinaryOperator::CreateAnd(Or,
4559 ConstantInt::get(RHS->getValue() | C1->getValue()));
4562 // (X ^ C1) | C2 --> (X | C2) ^ (C1&~C2)
4563 if (match(Op0, m_Xor(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
4564 Instruction *Or = BinaryOperator::CreateOr(X, RHS);
4565 InsertNewInstBefore(Or, I);
4567 return BinaryOperator::CreateXor(Or,
4568 ConstantInt::get(C1->getValue() & ~RHS->getValue()));
4571 // Try to fold constant and into select arguments.
4572 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
4573 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
4575 if (isa<PHINode>(Op0))
4576 if (Instruction *NV = FoldOpIntoPhi(I))
4580 Value *A = 0, *B = 0;
4581 ConstantInt *C1 = 0, *C2 = 0;
4583 if (match(Op0, m_And(m_Value(A), m_Value(B))))
4584 if (A == Op1 || B == Op1) // (A & ?) | A --> A
4585 return ReplaceInstUsesWith(I, Op1);
4586 if (match(Op1, m_And(m_Value(A), m_Value(B))))
4587 if (A == Op0 || B == Op0) // A | (A & ?) --> A
4588 return ReplaceInstUsesWith(I, Op0);
4590 // (A | B) | C and A | (B | C) -> bswap if possible.
4591 // (A >> B) | (C << D) and (A << B) | (B >> C) -> bswap if possible.
4592 if (match(Op0, m_Or(m_Value(), m_Value())) ||
4593 match(Op1, m_Or(m_Value(), m_Value())) ||
4594 (match(Op0, m_Shift(m_Value(), m_Value())) &&
4595 match(Op1, m_Shift(m_Value(), m_Value())))) {
4596 if (Instruction *BSwap = MatchBSwap(I))
4600 // (X^C)|Y -> (X|Y)^C iff Y&C == 0
4601 if (Op0->hasOneUse() && match(Op0, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
4602 MaskedValueIsZero(Op1, C1->getValue())) {
4603 Instruction *NOr = BinaryOperator::CreateOr(A, Op1);
4604 InsertNewInstBefore(NOr, I);
4606 return BinaryOperator::CreateXor(NOr, C1);
4609 // Y|(X^C) -> (X|Y)^C iff Y&C == 0
4610 if (Op1->hasOneUse() && match(Op1, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
4611 MaskedValueIsZero(Op0, C1->getValue())) {
4612 Instruction *NOr = BinaryOperator::CreateOr(A, Op0);
4613 InsertNewInstBefore(NOr, I);
4615 return BinaryOperator::CreateXor(NOr, C1);
4619 Value *C = 0, *D = 0;
4620 if (match(Op0, m_And(m_Value(A), m_Value(C))) &&
4621 match(Op1, m_And(m_Value(B), m_Value(D)))) {
4622 Value *V1 = 0, *V2 = 0, *V3 = 0;
4623 C1 = dyn_cast<ConstantInt>(C);
4624 C2 = dyn_cast<ConstantInt>(D);
4625 if (C1 && C2) { // (A & C1)|(B & C2)
4626 // If we have: ((V + N) & C1) | (V & C2)
4627 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
4628 // replace with V+N.
4629 if (C1->getValue() == ~C2->getValue()) {
4630 if ((C2->getValue() & (C2->getValue()+1)) == 0 && // C2 == 0+1+
4631 match(A, m_Add(m_Value(V1), m_Value(V2)))) {
4632 // Add commutes, try both ways.
4633 if (V1 == B && MaskedValueIsZero(V2, C2->getValue()))
4634 return ReplaceInstUsesWith(I, A);
4635 if (V2 == B && MaskedValueIsZero(V1, C2->getValue()))
4636 return ReplaceInstUsesWith(I, A);
4638 // Or commutes, try both ways.
4639 if ((C1->getValue() & (C1->getValue()+1)) == 0 &&
4640 match(B, m_Add(m_Value(V1), m_Value(V2)))) {
4641 // Add commutes, try both ways.
4642 if (V1 == A && MaskedValueIsZero(V2, C1->getValue()))
4643 return ReplaceInstUsesWith(I, B);
4644 if (V2 == A && MaskedValueIsZero(V1, C1->getValue()))
4645 return ReplaceInstUsesWith(I, B);
4648 V1 = 0; V2 = 0; V3 = 0;
4651 // Check to see if we have any common things being and'ed. If so, find the
4652 // terms for V1 & (V2|V3).
4653 if (isOnlyUse(Op0) || isOnlyUse(Op1)) {
4654 if (A == B) // (A & C)|(A & D) == A & (C|D)
4655 V1 = A, V2 = C, V3 = D;
4656 else if (A == D) // (A & C)|(B & A) == A & (B|C)
4657 V1 = A, V2 = B, V3 = C;
4658 else if (C == B) // (A & C)|(C & D) == C & (A|D)
4659 V1 = C, V2 = A, V3 = D;
4660 else if (C == D) // (A & C)|(B & C) == C & (A|B)
4661 V1 = C, V2 = A, V3 = B;
4665 InsertNewInstBefore(BinaryOperator::CreateOr(V2, V3, "tmp"), I);
4666 return BinaryOperator::CreateAnd(V1, Or);
4670 // (A & (C0?-1:0)) | (B & ~(C0?-1:0)) -> C0 ? A : B, and commuted variants
4671 if (Instruction *Match = MatchSelectFromAndOr(A, B, C, D))
4673 if (Instruction *Match = MatchSelectFromAndOr(B, A, D, C))
4675 if (Instruction *Match = MatchSelectFromAndOr(C, B, A, D))
4677 if (Instruction *Match = MatchSelectFromAndOr(D, A, B, C))
4680 // ((A&~B)|(~A&B)) -> A^B
4681 if ((match(C, m_Not(m_Specific(D))) &&
4682 match(B, m_Not(m_Specific(A)))))
4683 return BinaryOperator::CreateXor(A, D);
4684 // ((~B&A)|(~A&B)) -> A^B
4685 if ((match(A, m_Not(m_Specific(D))) &&
4686 match(B, m_Not(m_Specific(C)))))
4687 return BinaryOperator::CreateXor(C, D);
4688 // ((A&~B)|(B&~A)) -> A^B
4689 if ((match(C, m_Not(m_Specific(B))) &&
4690 match(D, m_Not(m_Specific(A)))))
4691 return BinaryOperator::CreateXor(A, B);
4692 // ((~B&A)|(B&~A)) -> A^B
4693 if ((match(A, m_Not(m_Specific(B))) &&
4694 match(D, m_Not(m_Specific(C)))))
4695 return BinaryOperator::CreateXor(C, B);
4698 // (X >> Z) | (Y >> Z) -> (X|Y) >> Z for all shifts.
4699 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
4700 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
4701 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
4702 SI0->getOperand(1) == SI1->getOperand(1) &&
4703 (SI0->hasOneUse() || SI1->hasOneUse())) {
4704 Instruction *NewOp =
4705 InsertNewInstBefore(BinaryOperator::CreateOr(SI0->getOperand(0),
4707 SI0->getName()), I);
4708 return BinaryOperator::Create(SI1->getOpcode(), NewOp,
4709 SI1->getOperand(1));
4713 // ((A|B)&1)|(B&-2) -> (A&1) | B
4714 if (match(Op0, m_And(m_Or(m_Value(A), m_Value(B)), m_Value(C))) ||
4715 match(Op0, m_And(m_Value(C), m_Or(m_Value(A), m_Value(B))))) {
4716 Instruction *Ret = FoldOrWithConstants(I, Op1, A, B, C);
4717 if (Ret) return Ret;
4719 // (B&-2)|((A|B)&1) -> (A&1) | B
4720 if (match(Op1, m_And(m_Or(m_Value(A), m_Value(B)), m_Value(C))) ||
4721 match(Op1, m_And(m_Value(C), m_Or(m_Value(A), m_Value(B))))) {
4722 Instruction *Ret = FoldOrWithConstants(I, Op0, A, B, C);
4723 if (Ret) return Ret;
4726 if (match(Op0, m_Not(m_Value(A)))) { // ~A | Op1
4727 if (A == Op1) // ~A | A == -1
4728 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4732 // Note, A is still live here!
4733 if (match(Op1, m_Not(m_Value(B)))) { // Op0 | ~B
4735 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4737 // (~A | ~B) == (~(A & B)) - De Morgan's Law
4738 if (A && isOnlyUse(Op0) && isOnlyUse(Op1)) {
4739 Value *And = InsertNewInstBefore(BinaryOperator::CreateAnd(A, B,
4740 I.getName()+".demorgan"), I);
4741 return BinaryOperator::CreateNot(And);
4745 // (icmp1 A, B) | (icmp2 A, B) --> (icmp3 A, B)
4746 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1))) {
4747 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
4750 if (ICmpInst *LHS = dyn_cast<ICmpInst>(I.getOperand(0)))
4751 if (Instruction *Res = FoldOrOfICmps(I, LHS, RHS))
4755 // fold (or (cast A), (cast B)) -> (cast (or A, B))
4756 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
4757 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4758 if (Op0C->getOpcode() == Op1C->getOpcode()) {// same cast kind ?
4759 if (!isa<ICmpInst>(Op0C->getOperand(0)) ||
4760 !isa<ICmpInst>(Op1C->getOperand(0))) {
4761 const Type *SrcTy = Op0C->getOperand(0)->getType();
4762 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
4763 // Only do this if the casts both really cause code to be
4765 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4767 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4769 Instruction *NewOp = BinaryOperator::CreateOr(Op0C->getOperand(0),
4770 Op1C->getOperand(0),
4772 InsertNewInstBefore(NewOp, I);
4773 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
4780 // (fcmp uno x, c) | (fcmp uno y, c) -> (fcmp uno x, y)
4781 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0))) {
4782 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1))) {
4783 if (LHS->getPredicate() == FCmpInst::FCMP_UNO &&
4784 RHS->getPredicate() == FCmpInst::FCMP_UNO &&
4785 LHS->getOperand(0)->getType() == RHS->getOperand(0)->getType()) {
4786 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
4787 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
4788 // If either of the constants are nans, then the whole thing returns
4790 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
4791 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4793 // Otherwise, no need to compare the two constants, compare the
4795 return new FCmpInst(FCmpInst::FCMP_UNO, LHS->getOperand(0),
4796 RHS->getOperand(0));
4799 Value *Op0LHS, *Op0RHS, *Op1LHS, *Op1RHS;
4800 FCmpInst::Predicate Op0CC, Op1CC;
4801 if (match(Op0, m_FCmp(Op0CC, m_Value(Op0LHS), m_Value(Op0RHS))) &&
4802 match(Op1, m_FCmp(Op1CC, m_Value(Op1LHS), m_Value(Op1RHS)))) {
4803 if (Op0LHS == Op1RHS && Op0RHS == Op1LHS) {
4804 // Swap RHS operands to match LHS.
4805 Op1CC = FCmpInst::getSwappedPredicate(Op1CC);
4806 std::swap(Op1LHS, Op1RHS);
4808 if (Op0LHS == Op1LHS && Op0RHS == Op1RHS) {
4809 // Simplify (fcmp cc0 x, y) | (fcmp cc1 x, y).
4811 return new FCmpInst((FCmpInst::Predicate)Op0CC, Op0LHS, Op0RHS);
4812 else if (Op0CC == FCmpInst::FCMP_TRUE ||
4813 Op1CC == FCmpInst::FCMP_TRUE)
4814 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4815 else if (Op0CC == FCmpInst::FCMP_FALSE)
4816 return ReplaceInstUsesWith(I, Op1);
4817 else if (Op1CC == FCmpInst::FCMP_FALSE)
4818 return ReplaceInstUsesWith(I, Op0);
4821 unsigned Op0Pred = getFCmpCode(Op0CC, Op0Ordered);
4822 unsigned Op1Pred = getFCmpCode(Op1CC, Op1Ordered);
4823 if (Op0Ordered == Op1Ordered) {
4824 // If both are ordered or unordered, return a new fcmp with
4825 // or'ed predicates.
4826 Value *RV = getFCmpValue(Op0Ordered, Op0Pred|Op1Pred,
4828 if (Instruction *I = dyn_cast<Instruction>(RV))
4830 // Otherwise, it's a constant boolean value...
4831 return ReplaceInstUsesWith(I, RV);
4839 return Changed ? &I : 0;
4844 // XorSelf - Implements: X ^ X --> 0
4847 XorSelf(Value *rhs) : RHS(rhs) {}
4848 bool shouldApply(Value *LHS) const { return LHS == RHS; }
4849 Instruction *apply(BinaryOperator &Xor) const {
4856 Instruction *InstCombiner::visitXor(BinaryOperator &I) {
4857 bool Changed = SimplifyCommutative(I);
4858 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4860 if (isa<UndefValue>(Op1)) {
4861 if (isa<UndefValue>(Op0))
4862 // Handle undef ^ undef -> 0 special case. This is a common
4864 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
4865 return ReplaceInstUsesWith(I, Op1); // X ^ undef -> undef
4868 // xor X, X = 0, even if X is nested in a sequence of Xor's.
4869 if (Instruction *Result = AssociativeOpt(I, XorSelf(Op1))) {
4870 assert(Result == &I && "AssociativeOpt didn't work?"); Result=Result;
4871 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
4874 // See if we can simplify any instructions used by the instruction whose sole
4875 // purpose is to compute bits we don't care about.
4876 if (!isa<VectorType>(I.getType())) {
4877 if (SimplifyDemandedInstructionBits(I))
4879 } else if (isa<ConstantAggregateZero>(Op1)) {
4880 return ReplaceInstUsesWith(I, Op0); // X ^ <0,0> -> X
4883 // Is this a ~ operation?
4884 if (Value *NotOp = dyn_castNotVal(&I)) {
4885 // ~(~X & Y) --> (X | ~Y) - De Morgan's Law
4886 // ~(~X | Y) === (X & ~Y) - De Morgan's Law
4887 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(NotOp)) {
4888 if (Op0I->getOpcode() == Instruction::And ||
4889 Op0I->getOpcode() == Instruction::Or) {
4890 if (dyn_castNotVal(Op0I->getOperand(1))) Op0I->swapOperands();
4891 if (Value *Op0NotVal = dyn_castNotVal(Op0I->getOperand(0))) {
4893 BinaryOperator::CreateNot(Op0I->getOperand(1),
4894 Op0I->getOperand(1)->getName()+".not");
4895 InsertNewInstBefore(NotY, I);
4896 if (Op0I->getOpcode() == Instruction::And)
4897 return BinaryOperator::CreateOr(Op0NotVal, NotY);
4899 return BinaryOperator::CreateAnd(Op0NotVal, NotY);
4906 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
4907 if (RHS == ConstantInt::getTrue() && Op0->hasOneUse()) {
4908 // xor (cmp A, B), true = not (cmp A, B) = !cmp A, B
4909 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Op0))
4910 return new ICmpInst(ICI->getInversePredicate(),
4911 ICI->getOperand(0), ICI->getOperand(1));
4913 if (FCmpInst *FCI = dyn_cast<FCmpInst>(Op0))
4914 return new FCmpInst(FCI->getInversePredicate(),
4915 FCI->getOperand(0), FCI->getOperand(1));
4918 // fold (xor(zext(cmp)), 1) and (xor(sext(cmp)), -1) to ext(!cmp).
4919 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
4920 if (CmpInst *CI = dyn_cast<CmpInst>(Op0C->getOperand(0))) {
4921 if (CI->hasOneUse() && Op0C->hasOneUse()) {
4922 Instruction::CastOps Opcode = Op0C->getOpcode();
4923 if (Opcode == Instruction::ZExt || Opcode == Instruction::SExt) {
4924 if (RHS == ConstantExpr::getCast(Opcode, ConstantInt::getTrue(),
4925 Op0C->getDestTy())) {
4926 Instruction *NewCI = InsertNewInstBefore(CmpInst::Create(
4927 CI->getOpcode(), CI->getInversePredicate(),
4928 CI->getOperand(0), CI->getOperand(1)), I);
4929 NewCI->takeName(CI);
4930 return CastInst::Create(Opcode, NewCI, Op0C->getType());
4937 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
4938 // ~(c-X) == X-c-1 == X+(-c-1)
4939 if (Op0I->getOpcode() == Instruction::Sub && RHS->isAllOnesValue())
4940 if (Constant *Op0I0C = dyn_cast<Constant>(Op0I->getOperand(0))) {
4941 Constant *NegOp0I0C = ConstantExpr::getNeg(Op0I0C);
4942 Constant *ConstantRHS = ConstantExpr::getSub(NegOp0I0C,
4943 ConstantInt::get(I.getType(), 1));
4944 return BinaryOperator::CreateAdd(Op0I->getOperand(1), ConstantRHS);
4947 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) {
4948 if (Op0I->getOpcode() == Instruction::Add) {
4949 // ~(X-c) --> (-c-1)-X
4950 if (RHS->isAllOnesValue()) {
4951 Constant *NegOp0CI = ConstantExpr::getNeg(Op0CI);
4952 return BinaryOperator::CreateSub(
4953 ConstantExpr::getSub(NegOp0CI,
4954 ConstantInt::get(I.getType(), 1)),
4955 Op0I->getOperand(0));
4956 } else if (RHS->getValue().isSignBit()) {
4957 // (X + C) ^ signbit -> (X + C + signbit)
4958 Constant *C = ConstantInt::get(RHS->getValue() + Op0CI->getValue());
4959 return BinaryOperator::CreateAdd(Op0I->getOperand(0), C);
4962 } else if (Op0I->getOpcode() == Instruction::Or) {
4963 // (X|C1)^C2 -> X^(C1|C2) iff X&~C1 == 0
4964 if (MaskedValueIsZero(Op0I->getOperand(0), Op0CI->getValue())) {
4965 Constant *NewRHS = ConstantExpr::getOr(Op0CI, RHS);
4966 // Anything in both C1 and C2 is known to be zero, remove it from
4968 Constant *CommonBits = And(Op0CI, RHS);
4969 NewRHS = ConstantExpr::getAnd(NewRHS,
4970 ConstantExpr::getNot(CommonBits));
4971 AddToWorkList(Op0I);
4972 I.setOperand(0, Op0I->getOperand(0));
4973 I.setOperand(1, NewRHS);
4980 // Try to fold constant and into select arguments.
4981 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
4982 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
4984 if (isa<PHINode>(Op0))
4985 if (Instruction *NV = FoldOpIntoPhi(I))
4989 if (Value *X = dyn_castNotVal(Op0)) // ~A ^ A == -1
4991 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4993 if (Value *X = dyn_castNotVal(Op1)) // A ^ ~A == -1
4995 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4998 BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1);
5001 if (match(Op1I, m_Or(m_Value(A), m_Value(B)))) {
5002 if (A == Op0) { // B^(B|A) == (A|B)^B
5003 Op1I->swapOperands();
5005 std::swap(Op0, Op1);
5006 } else if (B == Op0) { // B^(A|B) == (A|B)^B
5007 I.swapOperands(); // Simplified below.
5008 std::swap(Op0, Op1);
5010 } else if (match(Op1I, m_Xor(m_Specific(Op0), m_Value(B)))) {
5011 return ReplaceInstUsesWith(I, B); // A^(A^B) == B
5012 } else if (match(Op1I, m_Xor(m_Value(A), m_Specific(Op0)))) {
5013 return ReplaceInstUsesWith(I, A); // A^(B^A) == B
5014 } else if (match(Op1I, m_And(m_Value(A), m_Value(B))) && Op1I->hasOneUse()){
5015 if (A == Op0) { // A^(A&B) -> A^(B&A)
5016 Op1I->swapOperands();
5019 if (B == Op0) { // A^(B&A) -> (B&A)^A
5020 I.swapOperands(); // Simplified below.
5021 std::swap(Op0, Op1);
5026 BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0);
5029 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) && Op0I->hasOneUse()) {
5030 if (A == Op1) // (B|A)^B == (A|B)^B
5032 if (B == Op1) { // (A|B)^B == A & ~B
5034 InsertNewInstBefore(BinaryOperator::CreateNot(Op1, "tmp"), I);
5035 return BinaryOperator::CreateAnd(A, NotB);
5037 } else if (match(Op0I, m_Xor(m_Specific(Op1), m_Value(B)))) {
5038 return ReplaceInstUsesWith(I, B); // (A^B)^A == B
5039 } else if (match(Op0I, m_Xor(m_Value(A), m_Specific(Op1)))) {
5040 return ReplaceInstUsesWith(I, A); // (B^A)^A == B
5041 } else if (match(Op0I, m_And(m_Value(A), m_Value(B))) && Op0I->hasOneUse()){
5042 if (A == Op1) // (A&B)^A -> (B&A)^A
5044 if (B == Op1 && // (B&A)^A == ~B & A
5045 !isa<ConstantInt>(Op1)) { // Canonical form is (B&C)^C
5047 InsertNewInstBefore(BinaryOperator::CreateNot(A, "tmp"), I);
5048 return BinaryOperator::CreateAnd(N, Op1);
5053 // (X >> Z) ^ (Y >> Z) -> (X^Y) >> Z for all shifts.
5054 if (Op0I && Op1I && Op0I->isShift() &&
5055 Op0I->getOpcode() == Op1I->getOpcode() &&
5056 Op0I->getOperand(1) == Op1I->getOperand(1) &&
5057 (Op1I->hasOneUse() || Op1I->hasOneUse())) {
5058 Instruction *NewOp =
5059 InsertNewInstBefore(BinaryOperator::CreateXor(Op0I->getOperand(0),
5060 Op1I->getOperand(0),
5061 Op0I->getName()), I);
5062 return BinaryOperator::Create(Op1I->getOpcode(), NewOp,
5063 Op1I->getOperand(1));
5067 Value *A, *B, *C, *D;
5068 // (A & B)^(A | B) -> A ^ B
5069 if (match(Op0I, m_And(m_Value(A), m_Value(B))) &&
5070 match(Op1I, m_Or(m_Value(C), m_Value(D)))) {
5071 if ((A == C && B == D) || (A == D && B == C))
5072 return BinaryOperator::CreateXor(A, B);
5074 // (A | B)^(A & B) -> A ^ B
5075 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) &&
5076 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
5077 if ((A == C && B == D) || (A == D && B == C))
5078 return BinaryOperator::CreateXor(A, B);
5082 if ((Op0I->hasOneUse() || Op1I->hasOneUse()) &&
5083 match(Op0I, m_And(m_Value(A), m_Value(B))) &&
5084 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
5085 // (X & Y)^(X & Y) -> (Y^Z) & X
5086 Value *X = 0, *Y = 0, *Z = 0;
5088 X = A, Y = B, Z = D;
5090 X = A, Y = B, Z = C;
5092 X = B, Y = A, Z = D;
5094 X = B, Y = A, Z = C;
5097 Instruction *NewOp =
5098 InsertNewInstBefore(BinaryOperator::CreateXor(Y, Z, Op0->getName()), I);
5099 return BinaryOperator::CreateAnd(NewOp, X);
5104 // (icmp1 A, B) ^ (icmp2 A, B) --> (icmp3 A, B)
5105 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1)))
5106 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
5109 // fold (xor (cast A), (cast B)) -> (cast (xor A, B))
5110 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
5111 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
5112 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind?
5113 const Type *SrcTy = Op0C->getOperand(0)->getType();
5114 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
5115 // Only do this if the casts both really cause code to be generated.
5116 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
5118 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
5120 Instruction *NewOp = BinaryOperator::CreateXor(Op0C->getOperand(0),
5121 Op1C->getOperand(0),
5123 InsertNewInstBefore(NewOp, I);
5124 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
5129 return Changed ? &I : 0;
5132 /// AddWithOverflow - Compute Result = In1+In2, returning true if the result
5133 /// overflowed for this type.
5134 static bool AddWithOverflow(ConstantInt *&Result, ConstantInt *In1,
5135 ConstantInt *In2, bool IsSigned = false) {
5136 Result = cast<ConstantInt>(Add(In1, In2));
5139 if (In2->getValue().isNegative())
5140 return Result->getValue().sgt(In1->getValue());
5142 return Result->getValue().slt(In1->getValue());
5144 return Result->getValue().ult(In1->getValue());
5147 /// SubWithOverflow - Compute Result = In1-In2, returning true if the result
5148 /// overflowed for this type.
5149 static bool SubWithOverflow(ConstantInt *&Result, ConstantInt *In1,
5150 ConstantInt *In2, bool IsSigned = false) {
5151 Result = cast<ConstantInt>(Subtract(In1, In2));
5154 if (In2->getValue().isNegative())
5155 return Result->getValue().slt(In1->getValue());
5157 return Result->getValue().sgt(In1->getValue());
5159 return Result->getValue().ugt(In1->getValue());
5162 /// EmitGEPOffset - Given a getelementptr instruction/constantexpr, emit the
5163 /// code necessary to compute the offset from the base pointer (without adding
5164 /// in the base pointer). Return the result as a signed integer of intptr size.
5165 static Value *EmitGEPOffset(User *GEP, Instruction &I, InstCombiner &IC) {
5166 TargetData &TD = IC.getTargetData();
5167 gep_type_iterator GTI = gep_type_begin(GEP);
5168 const Type *IntPtrTy = TD.getIntPtrType();
5169 Value *Result = Constant::getNullValue(IntPtrTy);
5171 // Build a mask for high order bits.
5172 unsigned IntPtrWidth = TD.getPointerSizeInBits();
5173 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth);
5175 for (User::op_iterator i = GEP->op_begin() + 1, e = GEP->op_end(); i != e;
5178 uint64_t Size = TD.getTypePaddedSize(GTI.getIndexedType()) & PtrSizeMask;
5179 if (ConstantInt *OpC = dyn_cast<ConstantInt>(Op)) {
5180 if (OpC->isZero()) continue;
5182 // Handle a struct index, which adds its field offset to the pointer.
5183 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
5184 Size = TD.getStructLayout(STy)->getElementOffset(OpC->getZExtValue());
5186 if (ConstantInt *RC = dyn_cast<ConstantInt>(Result))
5187 Result = ConstantInt::get(RC->getValue() + APInt(IntPtrWidth, Size));
5189 Result = IC.InsertNewInstBefore(
5190 BinaryOperator::CreateAdd(Result,
5191 ConstantInt::get(IntPtrTy, Size),
5192 GEP->getName()+".offs"), I);
5196 Constant *Scale = ConstantInt::get(IntPtrTy, Size);
5197 Constant *OC = ConstantExpr::getIntegerCast(OpC, IntPtrTy, true /*SExt*/);
5198 Scale = ConstantExpr::getMul(OC, Scale);
5199 if (Constant *RC = dyn_cast<Constant>(Result))
5200 Result = ConstantExpr::getAdd(RC, Scale);
5202 // Emit an add instruction.
5203 Result = IC.InsertNewInstBefore(
5204 BinaryOperator::CreateAdd(Result, Scale,
5205 GEP->getName()+".offs"), I);
5209 // Convert to correct type.
5210 if (Op->getType() != IntPtrTy) {
5211 if (Constant *OpC = dyn_cast<Constant>(Op))
5212 Op = ConstantExpr::getSExt(OpC, IntPtrTy);
5214 Op = IC.InsertNewInstBefore(new SExtInst(Op, IntPtrTy,
5215 Op->getName()+".c"), I);
5218 Constant *Scale = ConstantInt::get(IntPtrTy, Size);
5219 if (Constant *OpC = dyn_cast<Constant>(Op))
5220 Op = ConstantExpr::getMul(OpC, Scale);
5221 else // We'll let instcombine(mul) convert this to a shl if possible.
5222 Op = IC.InsertNewInstBefore(BinaryOperator::CreateMul(Op, Scale,
5223 GEP->getName()+".idx"), I);
5226 // Emit an add instruction.
5227 if (isa<Constant>(Op) && isa<Constant>(Result))
5228 Result = ConstantExpr::getAdd(cast<Constant>(Op),
5229 cast<Constant>(Result));
5231 Result = IC.InsertNewInstBefore(BinaryOperator::CreateAdd(Op, Result,
5232 GEP->getName()+".offs"), I);
5238 /// EvaluateGEPOffsetExpression - Return an value that can be used to compare of
5239 /// the *offset* implied by GEP to zero. For example, if we have &A[i], we want
5240 /// to return 'i' for "icmp ne i, 0". Note that, in general, indices can be
5241 /// complex, and scales are involved. The above expression would also be legal
5242 /// to codegen as "icmp ne (i*4), 0" (assuming A is a pointer to i32). This
5243 /// later form is less amenable to optimization though, and we are allowed to
5244 /// generate the first by knowing that pointer arithmetic doesn't overflow.
5246 /// If we can't emit an optimized form for this expression, this returns null.
5248 static Value *EvaluateGEPOffsetExpression(User *GEP, Instruction &I,
5250 TargetData &TD = IC.getTargetData();
5251 gep_type_iterator GTI = gep_type_begin(GEP);
5253 // Check to see if this gep only has a single variable index. If so, and if
5254 // any constant indices are a multiple of its scale, then we can compute this
5255 // in terms of the scale of the variable index. For example, if the GEP
5256 // implies an offset of "12 + i*4", then we can codegen this as "3 + i",
5257 // because the expression will cross zero at the same point.
5258 unsigned i, e = GEP->getNumOperands();
5260 for (i = 1; i != e; ++i, ++GTI) {
5261 if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i))) {
5262 // Compute the aggregate offset of constant indices.
5263 if (CI->isZero()) continue;
5265 // Handle a struct index, which adds its field offset to the pointer.
5266 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
5267 Offset += TD.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
5269 uint64_t Size = TD.getTypePaddedSize(GTI.getIndexedType());
5270 Offset += Size*CI->getSExtValue();
5273 // Found our variable index.
5278 // If there are no variable indices, we must have a constant offset, just
5279 // evaluate it the general way.
5280 if (i == e) return 0;
5282 Value *VariableIdx = GEP->getOperand(i);
5283 // Determine the scale factor of the variable element. For example, this is
5284 // 4 if the variable index is into an array of i32.
5285 uint64_t VariableScale = TD.getTypePaddedSize(GTI.getIndexedType());
5287 // Verify that there are no other variable indices. If so, emit the hard way.
5288 for (++i, ++GTI; i != e; ++i, ++GTI) {
5289 ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i));
5292 // Compute the aggregate offset of constant indices.
5293 if (CI->isZero()) continue;
5295 // Handle a struct index, which adds its field offset to the pointer.
5296 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
5297 Offset += TD.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
5299 uint64_t Size = TD.getTypePaddedSize(GTI.getIndexedType());
5300 Offset += Size*CI->getSExtValue();
5304 // Okay, we know we have a single variable index, which must be a
5305 // pointer/array/vector index. If there is no offset, life is simple, return
5307 unsigned IntPtrWidth = TD.getPointerSizeInBits();
5309 // Cast to intptrty in case a truncation occurs. If an extension is needed,
5310 // we don't need to bother extending: the extension won't affect where the
5311 // computation crosses zero.
5312 if (VariableIdx->getType()->getPrimitiveSizeInBits() > IntPtrWidth)
5313 VariableIdx = new TruncInst(VariableIdx, TD.getIntPtrType(),
5314 VariableIdx->getNameStart(), &I);
5318 // Otherwise, there is an index. The computation we will do will be modulo
5319 // the pointer size, so get it.
5320 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth);
5322 Offset &= PtrSizeMask;
5323 VariableScale &= PtrSizeMask;
5325 // To do this transformation, any constant index must be a multiple of the
5326 // variable scale factor. For example, we can evaluate "12 + 4*i" as "3 + i",
5327 // but we can't evaluate "10 + 3*i" in terms of i. Check that the offset is a
5328 // multiple of the variable scale.
5329 int64_t NewOffs = Offset / (int64_t)VariableScale;
5330 if (Offset != NewOffs*(int64_t)VariableScale)
5333 // Okay, we can do this evaluation. Start by converting the index to intptr.
5334 const Type *IntPtrTy = TD.getIntPtrType();
5335 if (VariableIdx->getType() != IntPtrTy)
5336 VariableIdx = CastInst::CreateIntegerCast(VariableIdx, IntPtrTy,
5338 VariableIdx->getNameStart(), &I);
5339 Constant *OffsetVal = ConstantInt::get(IntPtrTy, NewOffs);
5340 return BinaryOperator::CreateAdd(VariableIdx, OffsetVal, "offset", &I);
5344 /// FoldGEPICmp - Fold comparisons between a GEP instruction and something
5345 /// else. At this point we know that the GEP is on the LHS of the comparison.
5346 Instruction *InstCombiner::FoldGEPICmp(User *GEPLHS, Value *RHS,
5347 ICmpInst::Predicate Cond,
5349 assert(dyn_castGetElementPtr(GEPLHS) && "LHS is not a getelementptr!");
5351 // Look through bitcasts.
5352 if (BitCastInst *BCI = dyn_cast<BitCastInst>(RHS))
5353 RHS = BCI->getOperand(0);
5355 Value *PtrBase = GEPLHS->getOperand(0);
5356 if (PtrBase == RHS) {
5357 // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
5358 // This transformation (ignoring the base and scales) is valid because we
5359 // know pointers can't overflow. See if we can output an optimized form.
5360 Value *Offset = EvaluateGEPOffsetExpression(GEPLHS, I, *this);
5362 // If not, synthesize the offset the hard way.
5364 Offset = EmitGEPOffset(GEPLHS, I, *this);
5365 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Offset,
5366 Constant::getNullValue(Offset->getType()));
5367 } else if (User *GEPRHS = dyn_castGetElementPtr(RHS)) {
5368 // If the base pointers are different, but the indices are the same, just
5369 // compare the base pointer.
5370 if (PtrBase != GEPRHS->getOperand(0)) {
5371 bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands();
5372 IndicesTheSame &= GEPLHS->getOperand(0)->getType() ==
5373 GEPRHS->getOperand(0)->getType();
5375 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
5376 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
5377 IndicesTheSame = false;
5381 // If all indices are the same, just compare the base pointers.
5383 return new ICmpInst(ICmpInst::getSignedPredicate(Cond),
5384 GEPLHS->getOperand(0), GEPRHS->getOperand(0));
5386 // Otherwise, the base pointers are different and the indices are
5387 // different, bail out.
5391 // If one of the GEPs has all zero indices, recurse.
5392 bool AllZeros = true;
5393 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
5394 if (!isa<Constant>(GEPLHS->getOperand(i)) ||
5395 !cast<Constant>(GEPLHS->getOperand(i))->isNullValue()) {
5400 return FoldGEPICmp(GEPRHS, GEPLHS->getOperand(0),
5401 ICmpInst::getSwappedPredicate(Cond), I);
5403 // If the other GEP has all zero indices, recurse.
5405 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
5406 if (!isa<Constant>(GEPRHS->getOperand(i)) ||
5407 !cast<Constant>(GEPRHS->getOperand(i))->isNullValue()) {
5412 return FoldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I);
5414 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
5415 // If the GEPs only differ by one index, compare it.
5416 unsigned NumDifferences = 0; // Keep track of # differences.
5417 unsigned DiffOperand = 0; // The operand that differs.
5418 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
5419 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
5420 if (GEPLHS->getOperand(i)->getType()->getPrimitiveSizeInBits() !=
5421 GEPRHS->getOperand(i)->getType()->getPrimitiveSizeInBits()) {
5422 // Irreconcilable differences.
5426 if (NumDifferences++) break;
5431 if (NumDifferences == 0) // SAME GEP?
5432 return ReplaceInstUsesWith(I, // No comparison is needed here.
5433 ConstantInt::get(Type::Int1Ty,
5434 ICmpInst::isTrueWhenEqual(Cond)));
5436 else if (NumDifferences == 1) {
5437 Value *LHSV = GEPLHS->getOperand(DiffOperand);
5438 Value *RHSV = GEPRHS->getOperand(DiffOperand);
5439 // Make sure we do a signed comparison here.
5440 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), LHSV, RHSV);
5444 // Only lower this if the icmp is the only user of the GEP or if we expect
5445 // the result to fold to a constant!
5446 if ((isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
5447 (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
5448 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
5449 Value *L = EmitGEPOffset(GEPLHS, I, *this);
5450 Value *R = EmitGEPOffset(GEPRHS, I, *this);
5451 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), L, R);
5457 /// FoldFCmp_IntToFP_Cst - Fold fcmp ([us]itofp x, cst) if possible.
5459 Instruction *InstCombiner::FoldFCmp_IntToFP_Cst(FCmpInst &I,
5462 if (!isa<ConstantFP>(RHSC)) return 0;
5463 const APFloat &RHS = cast<ConstantFP>(RHSC)->getValueAPF();
5465 // Get the width of the mantissa. We don't want to hack on conversions that
5466 // might lose information from the integer, e.g. "i64 -> float"
5467 int MantissaWidth = LHSI->getType()->getFPMantissaWidth();
5468 if (MantissaWidth == -1) return 0; // Unknown.
5470 // Check to see that the input is converted from an integer type that is small
5471 // enough that preserves all bits. TODO: check here for "known" sign bits.
5472 // This would allow us to handle (fptosi (x >>s 62) to float) if x is i64 f.e.
5473 unsigned InputSize = LHSI->getOperand(0)->getType()->getPrimitiveSizeInBits();
5475 // If this is a uitofp instruction, we need an extra bit to hold the sign.
5476 bool LHSUnsigned = isa<UIToFPInst>(LHSI);
5480 // If the conversion would lose info, don't hack on this.
5481 if ((int)InputSize > MantissaWidth)
5484 // Otherwise, we can potentially simplify the comparison. We know that it
5485 // will always come through as an integer value and we know the constant is
5486 // not a NAN (it would have been previously simplified).
5487 assert(!RHS.isNaN() && "NaN comparison not already folded!");
5489 ICmpInst::Predicate Pred;
5490 switch (I.getPredicate()) {
5491 default: assert(0 && "Unexpected predicate!");
5492 case FCmpInst::FCMP_UEQ:
5493 case FCmpInst::FCMP_OEQ:
5494 Pred = ICmpInst::ICMP_EQ;
5496 case FCmpInst::FCMP_UGT:
5497 case FCmpInst::FCMP_OGT:
5498 Pred = LHSUnsigned ? ICmpInst::ICMP_UGT : ICmpInst::ICMP_SGT;
5500 case FCmpInst::FCMP_UGE:
5501 case FCmpInst::FCMP_OGE:
5502 Pred = LHSUnsigned ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_SGE;
5504 case FCmpInst::FCMP_ULT:
5505 case FCmpInst::FCMP_OLT:
5506 Pred = LHSUnsigned ? ICmpInst::ICMP_ULT : ICmpInst::ICMP_SLT;
5508 case FCmpInst::FCMP_ULE:
5509 case FCmpInst::FCMP_OLE:
5510 Pred = LHSUnsigned ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_SLE;
5512 case FCmpInst::FCMP_UNE:
5513 case FCmpInst::FCMP_ONE:
5514 Pred = ICmpInst::ICMP_NE;
5516 case FCmpInst::FCMP_ORD:
5517 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5518 case FCmpInst::FCMP_UNO:
5519 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5522 const IntegerType *IntTy = cast<IntegerType>(LHSI->getOperand(0)->getType());
5524 // Now we know that the APFloat is a normal number, zero or inf.
5526 // See if the FP constant is too large for the integer. For example,
5527 // comparing an i8 to 300.0.
5528 unsigned IntWidth = IntTy->getPrimitiveSizeInBits();
5531 // If the RHS value is > SignedMax, fold the comparison. This handles +INF
5532 // and large values.
5533 APFloat SMax(RHS.getSemantics(), APFloat::fcZero, false);
5534 SMax.convertFromAPInt(APInt::getSignedMaxValue(IntWidth), true,
5535 APFloat::rmNearestTiesToEven);
5536 if (SMax.compare(RHS) == APFloat::cmpLessThan) { // smax < 13123.0
5537 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SLT ||
5538 Pred == ICmpInst::ICMP_SLE)
5539 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5540 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5543 // If the RHS value is > UnsignedMax, fold the comparison. This handles
5544 // +INF and large values.
5545 APFloat UMax(RHS.getSemantics(), APFloat::fcZero, false);
5546 UMax.convertFromAPInt(APInt::getMaxValue(IntWidth), false,
5547 APFloat::rmNearestTiesToEven);
5548 if (UMax.compare(RHS) == APFloat::cmpLessThan) { // umax < 13123.0
5549 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_ULT ||
5550 Pred == ICmpInst::ICMP_ULE)
5551 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5552 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5557 // See if the RHS value is < SignedMin.
5558 APFloat SMin(RHS.getSemantics(), APFloat::fcZero, false);
5559 SMin.convertFromAPInt(APInt::getSignedMinValue(IntWidth), true,
5560 APFloat::rmNearestTiesToEven);
5561 if (SMin.compare(RHS) == APFloat::cmpGreaterThan) { // smin > 12312.0
5562 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SGT ||
5563 Pred == ICmpInst::ICMP_SGE)
5564 return ReplaceInstUsesWith(I,ConstantInt::getTrue());
5565 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5569 // Okay, now we know that the FP constant fits in the range [SMIN, SMAX] or
5570 // [0, UMAX], but it may still be fractional. See if it is fractional by
5571 // casting the FP value to the integer value and back, checking for equality.
5572 // Don't do this for zero, because -0.0 is not fractional.
5573 Constant *RHSInt = ConstantExpr::getFPToSI(RHSC, IntTy);
5574 if (!RHS.isZero() &&
5575 ConstantExpr::getSIToFP(RHSInt, RHSC->getType()) != RHSC) {
5576 // If we had a comparison against a fractional value, we have to adjust the
5577 // compare predicate and sometimes the value. RHSC is rounded towards zero
5580 default: assert(0 && "Unexpected integer comparison!");
5581 case ICmpInst::ICMP_NE: // (float)int != 4.4 --> true
5582 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5583 case ICmpInst::ICMP_EQ: // (float)int == 4.4 --> false
5584 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5585 case ICmpInst::ICMP_ULE:
5586 // (float)int <= 4.4 --> int <= 4
5587 // (float)int <= -4.4 --> false
5588 if (RHS.isNegative())
5589 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5591 case ICmpInst::ICMP_SLE:
5592 // (float)int <= 4.4 --> int <= 4
5593 // (float)int <= -4.4 --> int < -4
5594 if (RHS.isNegative())
5595 Pred = ICmpInst::ICMP_SLT;
5597 case ICmpInst::ICMP_ULT:
5598 // (float)int < -4.4 --> false
5599 // (float)int < 4.4 --> int <= 4
5600 if (RHS.isNegative())
5601 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5602 Pred = ICmpInst::ICMP_ULE;
5604 case ICmpInst::ICMP_SLT:
5605 // (float)int < -4.4 --> int < -4
5606 // (float)int < 4.4 --> int <= 4
5607 if (!RHS.isNegative())
5608 Pred = ICmpInst::ICMP_SLE;
5610 case ICmpInst::ICMP_UGT:
5611 // (float)int > 4.4 --> int > 4
5612 // (float)int > -4.4 --> true
5613 if (RHS.isNegative())
5614 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5616 case ICmpInst::ICMP_SGT:
5617 // (float)int > 4.4 --> int > 4
5618 // (float)int > -4.4 --> int >= -4
5619 if (RHS.isNegative())
5620 Pred = ICmpInst::ICMP_SGE;
5622 case ICmpInst::ICMP_UGE:
5623 // (float)int >= -4.4 --> true
5624 // (float)int >= 4.4 --> int > 4
5625 if (!RHS.isNegative())
5626 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5627 Pred = ICmpInst::ICMP_UGT;
5629 case ICmpInst::ICMP_SGE:
5630 // (float)int >= -4.4 --> int >= -4
5631 // (float)int >= 4.4 --> int > 4
5632 if (!RHS.isNegative())
5633 Pred = ICmpInst::ICMP_SGT;
5638 // Lower this FP comparison into an appropriate integer version of the
5640 return new ICmpInst(Pred, LHSI->getOperand(0), RHSInt);
5643 Instruction *InstCombiner::visitFCmpInst(FCmpInst &I) {
5644 bool Changed = SimplifyCompare(I);
5645 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5647 // Fold trivial predicates.
5648 if (I.getPredicate() == FCmpInst::FCMP_FALSE)
5649 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5650 if (I.getPredicate() == FCmpInst::FCMP_TRUE)
5651 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5653 // Simplify 'fcmp pred X, X'
5655 switch (I.getPredicate()) {
5656 default: assert(0 && "Unknown predicate!");
5657 case FCmpInst::FCMP_UEQ: // True if unordered or equal
5658 case FCmpInst::FCMP_UGE: // True if unordered, greater than, or equal
5659 case FCmpInst::FCMP_ULE: // True if unordered, less than, or equal
5660 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5661 case FCmpInst::FCMP_OGT: // True if ordered and greater than
5662 case FCmpInst::FCMP_OLT: // True if ordered and less than
5663 case FCmpInst::FCMP_ONE: // True if ordered and operands are unequal
5664 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5666 case FCmpInst::FCMP_UNO: // True if unordered: isnan(X) | isnan(Y)
5667 case FCmpInst::FCMP_ULT: // True if unordered or less than
5668 case FCmpInst::FCMP_UGT: // True if unordered or greater than
5669 case FCmpInst::FCMP_UNE: // True if unordered or not equal
5670 // Canonicalize these to be 'fcmp uno %X, 0.0'.
5671 I.setPredicate(FCmpInst::FCMP_UNO);
5672 I.setOperand(1, Constant::getNullValue(Op0->getType()));
5675 case FCmpInst::FCMP_ORD: // True if ordered (no nans)
5676 case FCmpInst::FCMP_OEQ: // True if ordered and equal
5677 case FCmpInst::FCMP_OGE: // True if ordered and greater than or equal
5678 case FCmpInst::FCMP_OLE: // True if ordered and less than or equal
5679 // Canonicalize these to be 'fcmp ord %X, 0.0'.
5680 I.setPredicate(FCmpInst::FCMP_ORD);
5681 I.setOperand(1, Constant::getNullValue(Op0->getType()));
5686 if (isa<UndefValue>(Op1)) // fcmp pred X, undef -> undef
5687 return ReplaceInstUsesWith(I, UndefValue::get(Type::Int1Ty));
5689 // Handle fcmp with constant RHS
5690 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
5691 // If the constant is a nan, see if we can fold the comparison based on it.
5692 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHSC)) {
5693 if (CFP->getValueAPF().isNaN()) {
5694 if (FCmpInst::isOrdered(I.getPredicate())) // True if ordered and...
5695 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5696 assert(FCmpInst::isUnordered(I.getPredicate()) &&
5697 "Comparison must be either ordered or unordered!");
5698 // True if unordered.
5699 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5703 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
5704 switch (LHSI->getOpcode()) {
5705 case Instruction::PHI:
5706 // Only fold fcmp into the PHI if the phi and fcmp are in the same
5707 // block. If in the same block, we're encouraging jump threading. If
5708 // not, we are just pessimizing the code by making an i1 phi.
5709 if (LHSI->getParent() == I.getParent())
5710 if (Instruction *NV = FoldOpIntoPhi(I))
5713 case Instruction::SIToFP:
5714 case Instruction::UIToFP:
5715 if (Instruction *NV = FoldFCmp_IntToFP_Cst(I, LHSI, RHSC))
5718 case Instruction::Select:
5719 // If either operand of the select is a constant, we can fold the
5720 // comparison into the select arms, which will cause one to be
5721 // constant folded and the select turned into a bitwise or.
5722 Value *Op1 = 0, *Op2 = 0;
5723 if (LHSI->hasOneUse()) {
5724 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
5725 // Fold the known value into the constant operand.
5726 Op1 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
5727 // Insert a new FCmp of the other select operand.
5728 Op2 = InsertNewInstBefore(new FCmpInst(I.getPredicate(),
5729 LHSI->getOperand(2), RHSC,
5731 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
5732 // Fold the known value into the constant operand.
5733 Op2 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
5734 // Insert a new FCmp of the other select operand.
5735 Op1 = InsertNewInstBefore(new FCmpInst(I.getPredicate(),
5736 LHSI->getOperand(1), RHSC,
5742 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
5747 return Changed ? &I : 0;
5750 Instruction *InstCombiner::visitICmpInst(ICmpInst &I) {
5751 bool Changed = SimplifyCompare(I);
5752 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5753 const Type *Ty = Op0->getType();
5757 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
5758 I.isTrueWhenEqual()));
5760 if (isa<UndefValue>(Op1)) // X icmp undef -> undef
5761 return ReplaceInstUsesWith(I, UndefValue::get(Type::Int1Ty));
5763 // icmp <global/alloca*/null>, <global/alloca*/null> - Global/Stack value
5764 // addresses never equal each other! We already know that Op0 != Op1.
5765 if ((isa<GlobalValue>(Op0) || isa<AllocaInst>(Op0) ||
5766 isa<ConstantPointerNull>(Op0)) &&
5767 (isa<GlobalValue>(Op1) || isa<AllocaInst>(Op1) ||
5768 isa<ConstantPointerNull>(Op1)))
5769 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
5770 !I.isTrueWhenEqual()));
5772 // icmp's with boolean values can always be turned into bitwise operations
5773 if (Ty == Type::Int1Ty) {
5774 switch (I.getPredicate()) {
5775 default: assert(0 && "Invalid icmp instruction!");
5776 case ICmpInst::ICMP_EQ: { // icmp eq i1 A, B -> ~(A^B)
5777 Instruction *Xor = BinaryOperator::CreateXor(Op0, Op1, I.getName()+"tmp");
5778 InsertNewInstBefore(Xor, I);
5779 return BinaryOperator::CreateNot(Xor);
5781 case ICmpInst::ICMP_NE: // icmp eq i1 A, B -> A^B
5782 return BinaryOperator::CreateXor(Op0, Op1);
5784 case ICmpInst::ICMP_UGT:
5785 std::swap(Op0, Op1); // Change icmp ugt -> icmp ult
5787 case ICmpInst::ICMP_ULT:{ // icmp ult i1 A, B -> ~A & B
5788 Instruction *Not = BinaryOperator::CreateNot(Op0, I.getName()+"tmp");
5789 InsertNewInstBefore(Not, I);
5790 return BinaryOperator::CreateAnd(Not, Op1);
5792 case ICmpInst::ICMP_SGT:
5793 std::swap(Op0, Op1); // Change icmp sgt -> icmp slt
5795 case ICmpInst::ICMP_SLT: { // icmp slt i1 A, B -> A & ~B
5796 Instruction *Not = BinaryOperator::CreateNot(Op1, I.getName()+"tmp");
5797 InsertNewInstBefore(Not, I);
5798 return BinaryOperator::CreateAnd(Not, Op0);
5800 case ICmpInst::ICMP_UGE:
5801 std::swap(Op0, Op1); // Change icmp uge -> icmp ule
5803 case ICmpInst::ICMP_ULE: { // icmp ule i1 A, B -> ~A | B
5804 Instruction *Not = BinaryOperator::CreateNot(Op0, I.getName()+"tmp");
5805 InsertNewInstBefore(Not, I);
5806 return BinaryOperator::CreateOr(Not, Op1);
5808 case ICmpInst::ICMP_SGE:
5809 std::swap(Op0, Op1); // Change icmp sge -> icmp sle
5811 case ICmpInst::ICMP_SLE: { // icmp sle i1 A, B -> A | ~B
5812 Instruction *Not = BinaryOperator::CreateNot(Op1, I.getName()+"tmp");
5813 InsertNewInstBefore(Not, I);
5814 return BinaryOperator::CreateOr(Not, Op0);
5819 // See if we are doing a comparison with a constant.
5820 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
5823 // (icmp ne/eq (sub A B) 0) -> (icmp ne/eq A, B)
5824 if (I.isEquality() && CI->isNullValue() &&
5825 match(Op0, m_Sub(m_Value(A), m_Value(B)))) {
5826 // (icmp cond A B) if cond is equality
5827 return new ICmpInst(I.getPredicate(), A, B);
5830 // If we have an icmp le or icmp ge instruction, turn it into the
5831 // appropriate icmp lt or icmp gt instruction. This allows us to rely on
5832 // them being folded in the code below.
5833 switch (I.getPredicate()) {
5835 case ICmpInst::ICMP_ULE:
5836 if (CI->isMaxValue(false)) // A <=u MAX -> TRUE
5837 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5838 return new ICmpInst(ICmpInst::ICMP_ULT, Op0, AddOne(CI));
5839 case ICmpInst::ICMP_SLE:
5840 if (CI->isMaxValue(true)) // A <=s MAX -> TRUE
5841 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5842 return new ICmpInst(ICmpInst::ICMP_SLT, Op0, AddOne(CI));
5843 case ICmpInst::ICMP_UGE:
5844 if (CI->isMinValue(false)) // A >=u MIN -> TRUE
5845 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5846 return new ICmpInst( ICmpInst::ICMP_UGT, Op0, SubOne(CI));
5847 case ICmpInst::ICMP_SGE:
5848 if (CI->isMinValue(true)) // A >=s MIN -> TRUE
5849 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5850 return new ICmpInst(ICmpInst::ICMP_SGT, Op0, SubOne(CI));
5853 // See if we can fold the comparison based on range information we can get
5854 // by checking whether bits are known to be zero or one in the input.
5855 uint32_t BitWidth = cast<IntegerType>(Ty)->getBitWidth();
5856 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
5858 // If this comparison is a normal comparison, it demands all
5859 // bits, if it is a sign bit comparison, it only demands the sign bit.
5861 bool isSignBit = isSignBitCheck(I.getPredicate(), CI, UnusedBit);
5863 if (SimplifyDemandedBits(I.getOperandUse(0),
5864 isSignBit ? APInt::getSignBit(BitWidth)
5865 : APInt::getAllOnesValue(BitWidth),
5866 KnownZero, KnownOne, 0))
5869 // Given the known and unknown bits, compute a range that the LHS could be
5870 // in. Compute the Min, Max and RHS values based on the known bits. For the
5871 // EQ and NE we use unsigned values.
5872 APInt Min(BitWidth, 0), Max(BitWidth, 0);
5873 if (ICmpInst::isSignedPredicate(I.getPredicate()))
5874 ComputeSignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne, Min, Max);
5876 ComputeUnsignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne,Min,Max);
5878 // If Min and Max are known to be the same, then SimplifyDemandedBits
5879 // figured out that the LHS is a constant. Just constant fold this now so
5880 // that code below can assume that Min != Max.
5882 return ReplaceInstUsesWith(I, ConstantExpr::getICmp(I.getPredicate(),
5883 ConstantInt::get(Min),
5886 // Based on the range information we know about the LHS, see if we can
5887 // simplify this comparison. For example, (x&4) < 8 is always true.
5888 const APInt &RHSVal = CI->getValue();
5889 switch (I.getPredicate()) { // LE/GE have been folded already.
5890 default: assert(0 && "Unknown icmp opcode!");
5891 case ICmpInst::ICMP_EQ:
5892 if (Max.ult(RHSVal) || Min.ugt(RHSVal))
5893 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5895 case ICmpInst::ICMP_NE:
5896 if (Max.ult(RHSVal) || Min.ugt(RHSVal))
5897 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5899 case ICmpInst::ICMP_ULT:
5900 if (Max.ult(RHSVal)) // A <u C -> true iff max(A) < C
5901 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5902 if (Min.uge(RHSVal)) // A <u C -> false iff min(A) >= C
5903 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5904 if (RHSVal == Max) // A <u MAX -> A != MAX
5905 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5906 if (RHSVal == Min+1) // A <u MIN+1 -> A == MIN
5907 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, SubOne(CI));
5909 // (x <u 2147483648) -> (x >s -1) -> true if sign bit clear
5910 if (CI->isMinValue(true))
5911 return new ICmpInst(ICmpInst::ICMP_SGT, Op0,
5912 ConstantInt::getAllOnesValue(Op0->getType()));
5914 case ICmpInst::ICMP_UGT:
5915 if (Min.ugt(RHSVal)) // A >u C -> true iff min(A) > C
5916 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5917 if (Max.ule(RHSVal)) // A >u C -> false iff max(A) <= C
5918 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5920 if (RHSVal == Min) // A >u MIN -> A != MIN
5921 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5922 if (RHSVal == Max-1) // A >u MAX-1 -> A == MAX
5923 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, AddOne(CI));
5925 // (x >u 2147483647) -> (x <s 0) -> true if sign bit set
5926 if (CI->isMaxValue(true))
5927 return new ICmpInst(ICmpInst::ICMP_SLT, Op0,
5928 ConstantInt::getNullValue(Op0->getType()));
5930 case ICmpInst::ICMP_SLT:
5931 if (Max.slt(RHSVal)) // A <s C -> true iff max(A) < C
5932 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5933 if (Min.sge(RHSVal)) // A <s C -> false iff min(A) >= C
5934 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5935 if (RHSVal == Max) // A <s MAX -> A != MAX
5936 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5937 if (RHSVal == Min+1) // A <s MIN+1 -> A == MIN
5938 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, SubOne(CI));
5940 case ICmpInst::ICMP_SGT:
5941 if (Min.sgt(RHSVal)) // A >s C -> true iff min(A) > C
5942 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5943 if (Max.sle(RHSVal)) // A >s C -> false iff max(A) <= C
5944 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5946 if (RHSVal == Min) // A >s MIN -> A != MIN
5947 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5948 if (RHSVal == Max-1) // A >s MAX-1 -> A == MAX
5949 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, AddOne(CI));
5954 // Test if the ICmpInst instruction is used exclusively by a select as
5955 // part of a minimum or maximum operation. If so, refrain from doing
5956 // any other folding. This helps out other analyses which understand
5957 // non-obfuscated minimum and maximum idioms, such as ScalarEvolution
5958 // and CodeGen. And in this case, at least one of the comparison
5959 // operands has at least one user besides the compare (the select),
5960 // which would often largely negate the benefit of folding anyway.
5962 if (SelectInst *SI = dyn_cast<SelectInst>(*I.use_begin()))
5963 if ((SI->getOperand(1) == Op0 && SI->getOperand(2) == Op1) ||
5964 (SI->getOperand(2) == Op0 && SI->getOperand(1) == Op1))
5967 // See if we are doing a comparison between a constant and an instruction that
5968 // can be folded into the comparison.
5969 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
5970 // Since the RHS is a ConstantInt (CI), if the left hand side is an
5971 // instruction, see if that instruction also has constants so that the
5972 // instruction can be folded into the icmp
5973 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
5974 if (Instruction *Res = visitICmpInstWithInstAndIntCst(I, LHSI, CI))
5978 // Handle icmp with constant (but not simple integer constant) RHS
5979 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
5980 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
5981 switch (LHSI->getOpcode()) {
5982 case Instruction::GetElementPtr:
5983 if (RHSC->isNullValue()) {
5984 // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
5985 bool isAllZeros = true;
5986 for (unsigned i = 1, e = LHSI->getNumOperands(); i != e; ++i)
5987 if (!isa<Constant>(LHSI->getOperand(i)) ||
5988 !cast<Constant>(LHSI->getOperand(i))->isNullValue()) {
5993 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
5994 Constant::getNullValue(LHSI->getOperand(0)->getType()));
5998 case Instruction::PHI:
5999 // Only fold icmp into the PHI if the phi and fcmp are in the same
6000 // block. If in the same block, we're encouraging jump threading. If
6001 // not, we are just pessimizing the code by making an i1 phi.
6002 if (LHSI->getParent() == I.getParent())
6003 if (Instruction *NV = FoldOpIntoPhi(I))
6006 case Instruction::Select: {
6007 // If either operand of the select is a constant, we can fold the
6008 // comparison into the select arms, which will cause one to be
6009 // constant folded and the select turned into a bitwise or.
6010 Value *Op1 = 0, *Op2 = 0;
6011 if (LHSI->hasOneUse()) {
6012 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
6013 // Fold the known value into the constant operand.
6014 Op1 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
6015 // Insert a new ICmp of the other select operand.
6016 Op2 = InsertNewInstBefore(new ICmpInst(I.getPredicate(),
6017 LHSI->getOperand(2), RHSC,
6019 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
6020 // Fold the known value into the constant operand.
6021 Op2 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
6022 // Insert a new ICmp of the other select operand.
6023 Op1 = InsertNewInstBefore(new ICmpInst(I.getPredicate(),
6024 LHSI->getOperand(1), RHSC,
6030 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
6033 case Instruction::Malloc:
6034 // If we have (malloc != null), and if the malloc has a single use, we
6035 // can assume it is successful and remove the malloc.
6036 if (LHSI->hasOneUse() && isa<ConstantPointerNull>(RHSC)) {
6037 AddToWorkList(LHSI);
6038 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
6039 !I.isTrueWhenEqual()));
6045 // If we can optimize a 'icmp GEP, P' or 'icmp P, GEP', do so now.
6046 if (User *GEP = dyn_castGetElementPtr(Op0))
6047 if (Instruction *NI = FoldGEPICmp(GEP, Op1, I.getPredicate(), I))
6049 if (User *GEP = dyn_castGetElementPtr(Op1))
6050 if (Instruction *NI = FoldGEPICmp(GEP, Op0,
6051 ICmpInst::getSwappedPredicate(I.getPredicate()), I))
6054 // Test to see if the operands of the icmp are casted versions of other
6055 // values. If the ptr->ptr cast can be stripped off both arguments, we do so
6057 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op0)) {
6058 if (isa<PointerType>(Op0->getType()) &&
6059 (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
6060 // We keep moving the cast from the left operand over to the right
6061 // operand, where it can often be eliminated completely.
6062 Op0 = CI->getOperand(0);
6064 // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast
6065 // so eliminate it as well.
6066 if (BitCastInst *CI2 = dyn_cast<BitCastInst>(Op1))
6067 Op1 = CI2->getOperand(0);
6069 // If Op1 is a constant, we can fold the cast into the constant.
6070 if (Op0->getType() != Op1->getType()) {
6071 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
6072 Op1 = ConstantExpr::getBitCast(Op1C, Op0->getType());
6074 // Otherwise, cast the RHS right before the icmp
6075 Op1 = InsertBitCastBefore(Op1, Op0->getType(), I);
6078 return new ICmpInst(I.getPredicate(), Op0, Op1);
6082 if (isa<CastInst>(Op0)) {
6083 // Handle the special case of: icmp (cast bool to X), <cst>
6084 // This comes up when you have code like
6087 // For generality, we handle any zero-extension of any operand comparison
6088 // with a constant or another cast from the same type.
6089 if (isa<ConstantInt>(Op1) || isa<CastInst>(Op1))
6090 if (Instruction *R = visitICmpInstWithCastAndCast(I))
6094 // See if it's the same type of instruction on the left and right.
6095 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
6096 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
6097 if (Op0I->getOpcode() == Op1I->getOpcode() && Op0I->hasOneUse() &&
6098 Op1I->hasOneUse() && Op0I->getOperand(1) == Op1I->getOperand(1)) {
6099 switch (Op0I->getOpcode()) {
6101 case Instruction::Add:
6102 case Instruction::Sub:
6103 case Instruction::Xor:
6104 if (I.isEquality()) // a+x icmp eq/ne b+x --> a icmp b
6105 return new ICmpInst(I.getPredicate(), Op0I->getOperand(0),
6106 Op1I->getOperand(0));
6107 // icmp u/s (a ^ signbit), (b ^ signbit) --> icmp s/u a, b
6108 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) {
6109 if (CI->getValue().isSignBit()) {
6110 ICmpInst::Predicate Pred = I.isSignedPredicate()
6111 ? I.getUnsignedPredicate()
6112 : I.getSignedPredicate();
6113 return new ICmpInst(Pred, Op0I->getOperand(0),
6114 Op1I->getOperand(0));
6117 if (CI->getValue().isMaxSignedValue()) {
6118 ICmpInst::Predicate Pred = I.isSignedPredicate()
6119 ? I.getUnsignedPredicate()
6120 : I.getSignedPredicate();
6121 Pred = I.getSwappedPredicate(Pred);
6122 return new ICmpInst(Pred, Op0I->getOperand(0),
6123 Op1I->getOperand(0));
6127 case Instruction::Mul:
6128 if (!I.isEquality())
6131 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) {
6132 // a * Cst icmp eq/ne b * Cst --> a & Mask icmp b & Mask
6133 // Mask = -1 >> count-trailing-zeros(Cst).
6134 if (!CI->isZero() && !CI->isOne()) {
6135 const APInt &AP = CI->getValue();
6136 ConstantInt *Mask = ConstantInt::get(
6137 APInt::getLowBitsSet(AP.getBitWidth(),
6139 AP.countTrailingZeros()));
6140 Instruction *And1 = BinaryOperator::CreateAnd(Op0I->getOperand(0),
6142 Instruction *And2 = BinaryOperator::CreateAnd(Op1I->getOperand(0),
6144 InsertNewInstBefore(And1, I);
6145 InsertNewInstBefore(And2, I);
6146 return new ICmpInst(I.getPredicate(), And1, And2);
6155 // ~x < ~y --> y < x
6157 if (match(Op0, m_Not(m_Value(A))) &&
6158 match(Op1, m_Not(m_Value(B))))
6159 return new ICmpInst(I.getPredicate(), B, A);
6162 if (I.isEquality()) {
6163 Value *A, *B, *C, *D;
6165 // -x == -y --> x == y
6166 if (match(Op0, m_Neg(m_Value(A))) &&
6167 match(Op1, m_Neg(m_Value(B))))
6168 return new ICmpInst(I.getPredicate(), A, B);
6170 if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
6171 if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0
6172 Value *OtherVal = A == Op1 ? B : A;
6173 return new ICmpInst(I.getPredicate(), OtherVal,
6174 Constant::getNullValue(A->getType()));
6177 if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) {
6178 // A^c1 == C^c2 --> A == C^(c1^c2)
6179 ConstantInt *C1, *C2;
6180 if (match(B, m_ConstantInt(C1)) &&
6181 match(D, m_ConstantInt(C2)) && Op1->hasOneUse()) {
6182 Constant *NC = ConstantInt::get(C1->getValue() ^ C2->getValue());
6183 Instruction *Xor = BinaryOperator::CreateXor(C, NC, "tmp");
6184 return new ICmpInst(I.getPredicate(), A,
6185 InsertNewInstBefore(Xor, I));
6188 // A^B == A^D -> B == D
6189 if (A == C) return new ICmpInst(I.getPredicate(), B, D);
6190 if (A == D) return new ICmpInst(I.getPredicate(), B, C);
6191 if (B == C) return new ICmpInst(I.getPredicate(), A, D);
6192 if (B == D) return new ICmpInst(I.getPredicate(), A, C);
6196 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
6197 (A == Op0 || B == Op0)) {
6198 // A == (A^B) -> B == 0
6199 Value *OtherVal = A == Op0 ? B : A;
6200 return new ICmpInst(I.getPredicate(), OtherVal,
6201 Constant::getNullValue(A->getType()));
6204 // (A-B) == A -> B == 0
6205 if (match(Op0, m_Sub(m_Specific(Op1), m_Value(B))))
6206 return new ICmpInst(I.getPredicate(), B,
6207 Constant::getNullValue(B->getType()));
6209 // A == (A-B) -> B == 0
6210 if (match(Op1, m_Sub(m_Specific(Op0), m_Value(B))))
6211 return new ICmpInst(I.getPredicate(), B,
6212 Constant::getNullValue(B->getType()));
6214 // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
6215 if (Op0->hasOneUse() && Op1->hasOneUse() &&
6216 match(Op0, m_And(m_Value(A), m_Value(B))) &&
6217 match(Op1, m_And(m_Value(C), m_Value(D)))) {
6218 Value *X = 0, *Y = 0, *Z = 0;
6221 X = B; Y = D; Z = A;
6222 } else if (A == D) {
6223 X = B; Y = C; Z = A;
6224 } else if (B == C) {
6225 X = A; Y = D; Z = B;
6226 } else if (B == D) {
6227 X = A; Y = C; Z = B;
6230 if (X) { // Build (X^Y) & Z
6231 Op1 = InsertNewInstBefore(BinaryOperator::CreateXor(X, Y, "tmp"), I);
6232 Op1 = InsertNewInstBefore(BinaryOperator::CreateAnd(Op1, Z, "tmp"), I);
6233 I.setOperand(0, Op1);
6234 I.setOperand(1, Constant::getNullValue(Op1->getType()));
6239 return Changed ? &I : 0;
6243 /// FoldICmpDivCst - Fold "icmp pred, ([su]div X, DivRHS), CmpRHS" where DivRHS
6244 /// and CmpRHS are both known to be integer constants.
6245 Instruction *InstCombiner::FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
6246 ConstantInt *DivRHS) {
6247 ConstantInt *CmpRHS = cast<ConstantInt>(ICI.getOperand(1));
6248 const APInt &CmpRHSV = CmpRHS->getValue();
6250 // FIXME: If the operand types don't match the type of the divide
6251 // then don't attempt this transform. The code below doesn't have the
6252 // logic to deal with a signed divide and an unsigned compare (and
6253 // vice versa). This is because (x /s C1) <s C2 produces different
6254 // results than (x /s C1) <u C2 or (x /u C1) <s C2 or even
6255 // (x /u C1) <u C2. Simply casting the operands and result won't
6256 // work. :( The if statement below tests that condition and bails
6258 bool DivIsSigned = DivI->getOpcode() == Instruction::SDiv;
6259 if (!ICI.isEquality() && DivIsSigned != ICI.isSignedPredicate())
6261 if (DivRHS->isZero())
6262 return 0; // The ProdOV computation fails on divide by zero.
6263 if (DivIsSigned && DivRHS->isAllOnesValue())
6264 return 0; // The overflow computation also screws up here
6265 if (DivRHS->isOne())
6266 return 0; // Not worth bothering, and eliminates some funny cases
6269 // Compute Prod = CI * DivRHS. We are essentially solving an equation
6270 // of form X/C1=C2. We solve for X by multiplying C1 (DivRHS) and
6271 // C2 (CI). By solving for X we can turn this into a range check
6272 // instead of computing a divide.
6273 ConstantInt *Prod = Multiply(CmpRHS, DivRHS);
6275 // Determine if the product overflows by seeing if the product is
6276 // not equal to the divide. Make sure we do the same kind of divide
6277 // as in the LHS instruction that we're folding.
6278 bool ProdOV = (DivIsSigned ? ConstantExpr::getSDiv(Prod, DivRHS) :
6279 ConstantExpr::getUDiv(Prod, DivRHS)) != CmpRHS;
6281 // Get the ICmp opcode
6282 ICmpInst::Predicate Pred = ICI.getPredicate();
6284 // Figure out the interval that is being checked. For example, a comparison
6285 // like "X /u 5 == 0" is really checking that X is in the interval [0, 5).
6286 // Compute this interval based on the constants involved and the signedness of
6287 // the compare/divide. This computes a half-open interval, keeping track of
6288 // whether either value in the interval overflows. After analysis each
6289 // overflow variable is set to 0 if it's corresponding bound variable is valid
6290 // -1 if overflowed off the bottom end, or +1 if overflowed off the top end.
6291 int LoOverflow = 0, HiOverflow = 0;
6292 ConstantInt *LoBound = 0, *HiBound = 0;
6294 if (!DivIsSigned) { // udiv
6295 // e.g. X/5 op 3 --> [15, 20)
6297 HiOverflow = LoOverflow = ProdOV;
6299 HiOverflow = AddWithOverflow(HiBound, LoBound, DivRHS, false);
6300 } else if (DivRHS->getValue().isStrictlyPositive()) { // Divisor is > 0.
6301 if (CmpRHSV == 0) { // (X / pos) op 0
6302 // Can't overflow. e.g. X/2 op 0 --> [-1, 2)
6303 LoBound = cast<ConstantInt>(ConstantExpr::getNeg(SubOne(DivRHS)));
6305 } else if (CmpRHSV.isStrictlyPositive()) { // (X / pos) op pos
6306 LoBound = Prod; // e.g. X/5 op 3 --> [15, 20)
6307 HiOverflow = LoOverflow = ProdOV;
6309 HiOverflow = AddWithOverflow(HiBound, Prod, DivRHS, true);
6310 } else { // (X / pos) op neg
6311 // e.g. X/5 op -3 --> [-15-4, -15+1) --> [-19, -14)
6312 HiBound = AddOne(Prod);
6313 LoOverflow = HiOverflow = ProdOV ? -1 : 0;
6315 ConstantInt* DivNeg = cast<ConstantInt>(ConstantExpr::getNeg(DivRHS));
6316 LoOverflow = AddWithOverflow(LoBound, HiBound, DivNeg,
6320 } else if (DivRHS->getValue().isNegative()) { // Divisor is < 0.
6321 if (CmpRHSV == 0) { // (X / neg) op 0
6322 // e.g. X/-5 op 0 --> [-4, 5)
6323 LoBound = AddOne(DivRHS);
6324 HiBound = cast<ConstantInt>(ConstantExpr::getNeg(DivRHS));
6325 if (HiBound == DivRHS) { // -INTMIN = INTMIN
6326 HiOverflow = 1; // [INTMIN+1, overflow)
6327 HiBound = 0; // e.g. X/INTMIN = 0 --> X > INTMIN
6329 } else if (CmpRHSV.isStrictlyPositive()) { // (X / neg) op pos
6330 // e.g. X/-5 op 3 --> [-19, -14)
6331 HiBound = AddOne(Prod);
6332 HiOverflow = LoOverflow = ProdOV ? -1 : 0;
6334 LoOverflow = AddWithOverflow(LoBound, HiBound, DivRHS, true) ? -1 : 0;
6335 } else { // (X / neg) op neg
6336 LoBound = Prod; // e.g. X/-5 op -3 --> [15, 20)
6337 LoOverflow = HiOverflow = ProdOV;
6339 HiOverflow = SubWithOverflow(HiBound, Prod, DivRHS, true);
6342 // Dividing by a negative swaps the condition. LT <-> GT
6343 Pred = ICmpInst::getSwappedPredicate(Pred);
6346 Value *X = DivI->getOperand(0);
6348 default: assert(0 && "Unhandled icmp opcode!");
6349 case ICmpInst::ICMP_EQ:
6350 if (LoOverflow && HiOverflow)
6351 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
6352 else if (HiOverflow)
6353 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
6354 ICmpInst::ICMP_UGE, X, LoBound);
6355 else if (LoOverflow)
6356 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
6357 ICmpInst::ICMP_ULT, X, HiBound);
6359 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned, true, ICI);
6360 case ICmpInst::ICMP_NE:
6361 if (LoOverflow && HiOverflow)
6362 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
6363 else if (HiOverflow)
6364 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
6365 ICmpInst::ICMP_ULT, X, LoBound);
6366 else if (LoOverflow)
6367 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
6368 ICmpInst::ICMP_UGE, X, HiBound);
6370 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned, false, ICI);
6371 case ICmpInst::ICMP_ULT:
6372 case ICmpInst::ICMP_SLT:
6373 if (LoOverflow == +1) // Low bound is greater than input range.
6374 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
6375 if (LoOverflow == -1) // Low bound is less than input range.
6376 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
6377 return new ICmpInst(Pred, X, LoBound);
6378 case ICmpInst::ICMP_UGT:
6379 case ICmpInst::ICMP_SGT:
6380 if (HiOverflow == +1) // High bound greater than input range.
6381 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
6382 else if (HiOverflow == -1) // High bound less than input range.
6383 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
6384 if (Pred == ICmpInst::ICMP_UGT)
6385 return new ICmpInst(ICmpInst::ICMP_UGE, X, HiBound);
6387 return new ICmpInst(ICmpInst::ICMP_SGE, X, HiBound);
6392 /// visitICmpInstWithInstAndIntCst - Handle "icmp (instr, intcst)".
6394 Instruction *InstCombiner::visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
6397 const APInt &RHSV = RHS->getValue();
6399 switch (LHSI->getOpcode()) {
6400 case Instruction::Trunc:
6401 if (ICI.isEquality() && LHSI->hasOneUse()) {
6402 // Simplify icmp eq (trunc x to i8), 42 -> icmp eq x, 42|highbits if all
6403 // of the high bits truncated out of x are known.
6404 unsigned DstBits = LHSI->getType()->getPrimitiveSizeInBits(),
6405 SrcBits = LHSI->getOperand(0)->getType()->getPrimitiveSizeInBits();
6406 APInt Mask(APInt::getHighBitsSet(SrcBits, SrcBits-DstBits));
6407 APInt KnownZero(SrcBits, 0), KnownOne(SrcBits, 0);
6408 ComputeMaskedBits(LHSI->getOperand(0), Mask, KnownZero, KnownOne);
6410 // If all the high bits are known, we can do this xform.
6411 if ((KnownZero|KnownOne).countLeadingOnes() >= SrcBits-DstBits) {
6412 // Pull in the high bits from known-ones set.
6413 APInt NewRHS(RHS->getValue());
6414 NewRHS.zext(SrcBits);
6416 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
6417 ConstantInt::get(NewRHS));
6422 case Instruction::Xor: // (icmp pred (xor X, XorCST), CI)
6423 if (ConstantInt *XorCST = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
6424 // If this is a comparison that tests the signbit (X < 0) or (x > -1),
6426 if ((ICI.getPredicate() == ICmpInst::ICMP_SLT && RHSV == 0) ||
6427 (ICI.getPredicate() == ICmpInst::ICMP_SGT && RHSV.isAllOnesValue())) {
6428 Value *CompareVal = LHSI->getOperand(0);
6430 // If the sign bit of the XorCST is not set, there is no change to
6431 // the operation, just stop using the Xor.
6432 if (!XorCST->getValue().isNegative()) {
6433 ICI.setOperand(0, CompareVal);
6434 AddToWorkList(LHSI);
6438 // Was the old condition true if the operand is positive?
6439 bool isTrueIfPositive = ICI.getPredicate() == ICmpInst::ICMP_SGT;
6441 // If so, the new one isn't.
6442 isTrueIfPositive ^= true;
6444 if (isTrueIfPositive)
6445 return new ICmpInst(ICmpInst::ICMP_SGT, CompareVal, SubOne(RHS));
6447 return new ICmpInst(ICmpInst::ICMP_SLT, CompareVal, AddOne(RHS));
6450 if (LHSI->hasOneUse()) {
6451 // (icmp u/s (xor A SignBit), C) -> (icmp s/u A, (xor C SignBit))
6452 if (!ICI.isEquality() && XorCST->getValue().isSignBit()) {
6453 const APInt &SignBit = XorCST->getValue();
6454 ICmpInst::Predicate Pred = ICI.isSignedPredicate()
6455 ? ICI.getUnsignedPredicate()
6456 : ICI.getSignedPredicate();
6457 return new ICmpInst(Pred, LHSI->getOperand(0),
6458 ConstantInt::get(RHSV ^ SignBit));
6461 // (icmp u/s (xor A ~SignBit), C) -> (icmp s/u (xor C ~SignBit), A)
6462 if (!ICI.isEquality() && XorCST->getValue().isMaxSignedValue()) {
6463 const APInt &NotSignBit = XorCST->getValue();
6464 ICmpInst::Predicate Pred = ICI.isSignedPredicate()
6465 ? ICI.getUnsignedPredicate()
6466 : ICI.getSignedPredicate();
6467 Pred = ICI.getSwappedPredicate(Pred);
6468 return new ICmpInst(Pred, LHSI->getOperand(0),
6469 ConstantInt::get(RHSV ^ NotSignBit));
6474 case Instruction::And: // (icmp pred (and X, AndCST), RHS)
6475 if (LHSI->hasOneUse() && isa<ConstantInt>(LHSI->getOperand(1)) &&
6476 LHSI->getOperand(0)->hasOneUse()) {
6477 ConstantInt *AndCST = cast<ConstantInt>(LHSI->getOperand(1));
6479 // If the LHS is an AND of a truncating cast, we can widen the
6480 // and/compare to be the input width without changing the value
6481 // produced, eliminating a cast.
6482 if (TruncInst *Cast = dyn_cast<TruncInst>(LHSI->getOperand(0))) {
6483 // We can do this transformation if either the AND constant does not
6484 // have its sign bit set or if it is an equality comparison.
6485 // Extending a relational comparison when we're checking the sign
6486 // bit would not work.
6487 if (Cast->hasOneUse() &&
6488 (ICI.isEquality() ||
6489 (AndCST->getValue().isNonNegative() && RHSV.isNonNegative()))) {
6491 cast<IntegerType>(Cast->getOperand(0)->getType())->getBitWidth();
6492 APInt NewCST = AndCST->getValue();
6493 NewCST.zext(BitWidth);
6495 NewCI.zext(BitWidth);
6496 Instruction *NewAnd =
6497 BinaryOperator::CreateAnd(Cast->getOperand(0),
6498 ConstantInt::get(NewCST),LHSI->getName());
6499 InsertNewInstBefore(NewAnd, ICI);
6500 return new ICmpInst(ICI.getPredicate(), NewAnd,
6501 ConstantInt::get(NewCI));
6505 // If this is: (X >> C1) & C2 != C3 (where any shift and any compare
6506 // could exist), turn it into (X & (C2 << C1)) != (C3 << C1). This
6507 // happens a LOT in code produced by the C front-end, for bitfield
6509 BinaryOperator *Shift = dyn_cast<BinaryOperator>(LHSI->getOperand(0));
6510 if (Shift && !Shift->isShift())
6514 ShAmt = Shift ? dyn_cast<ConstantInt>(Shift->getOperand(1)) : 0;
6515 const Type *Ty = Shift ? Shift->getType() : 0; // Type of the shift.
6516 const Type *AndTy = AndCST->getType(); // Type of the and.
6518 // We can fold this as long as we can't shift unknown bits
6519 // into the mask. This can only happen with signed shift
6520 // rights, as they sign-extend.
6522 bool CanFold = Shift->isLogicalShift();
6524 // To test for the bad case of the signed shr, see if any
6525 // of the bits shifted in could be tested after the mask.
6526 uint32_t TyBits = Ty->getPrimitiveSizeInBits();
6527 int ShAmtVal = TyBits - ShAmt->getLimitedValue(TyBits);
6529 uint32_t BitWidth = AndTy->getPrimitiveSizeInBits();
6530 if ((APInt::getHighBitsSet(BitWidth, BitWidth-ShAmtVal) &
6531 AndCST->getValue()) == 0)
6537 if (Shift->getOpcode() == Instruction::Shl)
6538 NewCst = ConstantExpr::getLShr(RHS, ShAmt);
6540 NewCst = ConstantExpr::getShl(RHS, ShAmt);
6542 // Check to see if we are shifting out any of the bits being
6544 if (ConstantExpr::get(Shift->getOpcode(), NewCst, ShAmt) != RHS) {
6545 // If we shifted bits out, the fold is not going to work out.
6546 // As a special case, check to see if this means that the
6547 // result is always true or false now.
6548 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
6549 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
6550 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
6551 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
6553 ICI.setOperand(1, NewCst);
6554 Constant *NewAndCST;
6555 if (Shift->getOpcode() == Instruction::Shl)
6556 NewAndCST = ConstantExpr::getLShr(AndCST, ShAmt);
6558 NewAndCST = ConstantExpr::getShl(AndCST, ShAmt);
6559 LHSI->setOperand(1, NewAndCST);
6560 LHSI->setOperand(0, Shift->getOperand(0));
6561 AddToWorkList(Shift); // Shift is dead.
6562 AddUsesToWorkList(ICI);
6568 // Turn ((X >> Y) & C) == 0 into (X & (C << Y)) == 0. The later is
6569 // preferable because it allows the C<<Y expression to be hoisted out
6570 // of a loop if Y is invariant and X is not.
6571 if (Shift && Shift->hasOneUse() && RHSV == 0 &&
6572 ICI.isEquality() && !Shift->isArithmeticShift() &&
6573 isa<Instruction>(Shift->getOperand(0))) {
6576 if (Shift->getOpcode() == Instruction::LShr) {
6577 NS = BinaryOperator::CreateShl(AndCST,
6578 Shift->getOperand(1), "tmp");
6580 // Insert a logical shift.
6581 NS = BinaryOperator::CreateLShr(AndCST,
6582 Shift->getOperand(1), "tmp");
6584 InsertNewInstBefore(cast<Instruction>(NS), ICI);
6586 // Compute X & (C << Y).
6587 Instruction *NewAnd =
6588 BinaryOperator::CreateAnd(Shift->getOperand(0), NS, LHSI->getName());
6589 InsertNewInstBefore(NewAnd, ICI);
6591 ICI.setOperand(0, NewAnd);
6597 case Instruction::Shl: { // (icmp pred (shl X, ShAmt), CI)
6598 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
6601 uint32_t TypeBits = RHSV.getBitWidth();
6603 // Check that the shift amount is in range. If not, don't perform
6604 // undefined shifts. When the shift is visited it will be
6606 if (ShAmt->uge(TypeBits))
6609 if (ICI.isEquality()) {
6610 // If we are comparing against bits always shifted out, the
6611 // comparison cannot succeed.
6613 ConstantExpr::getShl(ConstantExpr::getLShr(RHS, ShAmt), ShAmt);
6614 if (Comp != RHS) {// Comparing against a bit that we know is zero.
6615 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
6616 Constant *Cst = ConstantInt::get(Type::Int1Ty, IsICMP_NE);
6617 return ReplaceInstUsesWith(ICI, Cst);
6620 if (LHSI->hasOneUse()) {
6621 // Otherwise strength reduce the shift into an and.
6622 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
6624 ConstantInt::get(APInt::getLowBitsSet(TypeBits, TypeBits-ShAmtVal));
6627 BinaryOperator::CreateAnd(LHSI->getOperand(0),
6628 Mask, LHSI->getName()+".mask");
6629 Value *And = InsertNewInstBefore(AndI, ICI);
6630 return new ICmpInst(ICI.getPredicate(), And,
6631 ConstantInt::get(RHSV.lshr(ShAmtVal)));
6635 // Otherwise, if this is a comparison of the sign bit, simplify to and/test.
6636 bool TrueIfSigned = false;
6637 if (LHSI->hasOneUse() &&
6638 isSignBitCheck(ICI.getPredicate(), RHS, TrueIfSigned)) {
6639 // (X << 31) <s 0 --> (X&1) != 0
6640 Constant *Mask = ConstantInt::get(APInt(TypeBits, 1) <<
6641 (TypeBits-ShAmt->getZExtValue()-1));
6643 BinaryOperator::CreateAnd(LHSI->getOperand(0),
6644 Mask, LHSI->getName()+".mask");
6645 Value *And = InsertNewInstBefore(AndI, ICI);
6647 return new ICmpInst(TrueIfSigned ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ,
6648 And, Constant::getNullValue(And->getType()));
6653 case Instruction::LShr: // (icmp pred (shr X, ShAmt), CI)
6654 case Instruction::AShr: {
6655 // Only handle equality comparisons of shift-by-constant.
6656 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
6657 if (!ShAmt || !ICI.isEquality()) break;
6659 // Check that the shift amount is in range. If not, don't perform
6660 // undefined shifts. When the shift is visited it will be
6662 uint32_t TypeBits = RHSV.getBitWidth();
6663 if (ShAmt->uge(TypeBits))
6666 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
6668 // If we are comparing against bits always shifted out, the
6669 // comparison cannot succeed.
6670 APInt Comp = RHSV << ShAmtVal;
6671 if (LHSI->getOpcode() == Instruction::LShr)
6672 Comp = Comp.lshr(ShAmtVal);
6674 Comp = Comp.ashr(ShAmtVal);
6676 if (Comp != RHSV) { // Comparing against a bit that we know is zero.
6677 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
6678 Constant *Cst = ConstantInt::get(Type::Int1Ty, IsICMP_NE);
6679 return ReplaceInstUsesWith(ICI, Cst);
6682 // Otherwise, check to see if the bits shifted out are known to be zero.
6683 // If so, we can compare against the unshifted value:
6684 // (X & 4) >> 1 == 2 --> (X & 4) == 4.
6685 if (LHSI->hasOneUse() &&
6686 MaskedValueIsZero(LHSI->getOperand(0),
6687 APInt::getLowBitsSet(Comp.getBitWidth(), ShAmtVal))) {
6688 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
6689 ConstantExpr::getShl(RHS, ShAmt));
6692 if (LHSI->hasOneUse()) {
6693 // Otherwise strength reduce the shift into an and.
6694 APInt Val(APInt::getHighBitsSet(TypeBits, TypeBits - ShAmtVal));
6695 Constant *Mask = ConstantInt::get(Val);
6698 BinaryOperator::CreateAnd(LHSI->getOperand(0),
6699 Mask, LHSI->getName()+".mask");
6700 Value *And = InsertNewInstBefore(AndI, ICI);
6701 return new ICmpInst(ICI.getPredicate(), And,
6702 ConstantExpr::getShl(RHS, ShAmt));
6707 case Instruction::SDiv:
6708 case Instruction::UDiv:
6709 // Fold: icmp pred ([us]div X, C1), C2 -> range test
6710 // Fold this div into the comparison, producing a range check.
6711 // Determine, based on the divide type, what the range is being
6712 // checked. If there is an overflow on the low or high side, remember
6713 // it, otherwise compute the range [low, hi) bounding the new value.
6714 // See: InsertRangeTest above for the kinds of replacements possible.
6715 if (ConstantInt *DivRHS = dyn_cast<ConstantInt>(LHSI->getOperand(1)))
6716 if (Instruction *R = FoldICmpDivCst(ICI, cast<BinaryOperator>(LHSI),
6721 case Instruction::Add:
6722 // Fold: icmp pred (add, X, C1), C2
6724 if (!ICI.isEquality()) {
6725 ConstantInt *LHSC = dyn_cast<ConstantInt>(LHSI->getOperand(1));
6727 const APInt &LHSV = LHSC->getValue();
6729 ConstantRange CR = ICI.makeConstantRange(ICI.getPredicate(), RHSV)
6732 if (ICI.isSignedPredicate()) {
6733 if (CR.getLower().isSignBit()) {
6734 return new ICmpInst(ICmpInst::ICMP_SLT, LHSI->getOperand(0),
6735 ConstantInt::get(CR.getUpper()));
6736 } else if (CR.getUpper().isSignBit()) {
6737 return new ICmpInst(ICmpInst::ICMP_SGE, LHSI->getOperand(0),
6738 ConstantInt::get(CR.getLower()));
6741 if (CR.getLower().isMinValue()) {
6742 return new ICmpInst(ICmpInst::ICMP_ULT, LHSI->getOperand(0),
6743 ConstantInt::get(CR.getUpper()));
6744 } else if (CR.getUpper().isMinValue()) {
6745 return new ICmpInst(ICmpInst::ICMP_UGE, LHSI->getOperand(0),
6746 ConstantInt::get(CR.getLower()));
6753 // Simplify icmp_eq and icmp_ne instructions with integer constant RHS.
6754 if (ICI.isEquality()) {
6755 bool isICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
6757 // If the first operand is (add|sub|and|or|xor|rem) with a constant, and
6758 // the second operand is a constant, simplify a bit.
6759 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(LHSI)) {
6760 switch (BO->getOpcode()) {
6761 case Instruction::SRem:
6762 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
6763 if (RHSV == 0 && isa<ConstantInt>(BO->getOperand(1)) &&BO->hasOneUse()){
6764 const APInt &V = cast<ConstantInt>(BO->getOperand(1))->getValue();
6765 if (V.sgt(APInt(V.getBitWidth(), 1)) && V.isPowerOf2()) {
6766 Instruction *NewRem =
6767 BinaryOperator::CreateURem(BO->getOperand(0), BO->getOperand(1),
6769 InsertNewInstBefore(NewRem, ICI);
6770 return new ICmpInst(ICI.getPredicate(), NewRem,
6771 Constant::getNullValue(BO->getType()));
6775 case Instruction::Add:
6776 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
6777 if (ConstantInt *BOp1C = dyn_cast<ConstantInt>(BO->getOperand(1))) {
6778 if (BO->hasOneUse())
6779 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
6780 Subtract(RHS, BOp1C));
6781 } else if (RHSV == 0) {
6782 // Replace ((add A, B) != 0) with (A != -B) if A or B is
6783 // efficiently invertible, or if the add has just this one use.
6784 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
6786 if (Value *NegVal = dyn_castNegVal(BOp1))
6787 return new ICmpInst(ICI.getPredicate(), BOp0, NegVal);
6788 else if (Value *NegVal = dyn_castNegVal(BOp0))
6789 return new ICmpInst(ICI.getPredicate(), NegVal, BOp1);
6790 else if (BO->hasOneUse()) {
6791 Instruction *Neg = BinaryOperator::CreateNeg(BOp1);
6792 InsertNewInstBefore(Neg, ICI);
6794 return new ICmpInst(ICI.getPredicate(), BOp0, Neg);
6798 case Instruction::Xor:
6799 // For the xor case, we can xor two constants together, eliminating
6800 // the explicit xor.
6801 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1)))
6802 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
6803 ConstantExpr::getXor(RHS, BOC));
6806 case Instruction::Sub:
6807 // Replace (([sub|xor] A, B) != 0) with (A != B)
6809 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
6813 case Instruction::Or:
6814 // If bits are being or'd in that are not present in the constant we
6815 // are comparing against, then the comparison could never succeed!
6816 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1))) {
6817 Constant *NotCI = ConstantExpr::getNot(RHS);
6818 if (!ConstantExpr::getAnd(BOC, NotCI)->isNullValue())
6819 return ReplaceInstUsesWith(ICI, ConstantInt::get(Type::Int1Ty,
6824 case Instruction::And:
6825 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
6826 // If bits are being compared against that are and'd out, then the
6827 // comparison can never succeed!
6828 if ((RHSV & ~BOC->getValue()) != 0)
6829 return ReplaceInstUsesWith(ICI, ConstantInt::get(Type::Int1Ty,
6832 // If we have ((X & C) == C), turn it into ((X & C) != 0).
6833 if (RHS == BOC && RHSV.isPowerOf2())
6834 return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ :
6835 ICmpInst::ICMP_NE, LHSI,
6836 Constant::getNullValue(RHS->getType()));
6838 // Replace (and X, (1 << size(X)-1) != 0) with x s< 0
6839 if (BOC->getValue().isSignBit()) {
6840 Value *X = BO->getOperand(0);
6841 Constant *Zero = Constant::getNullValue(X->getType());
6842 ICmpInst::Predicate pred = isICMP_NE ?
6843 ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
6844 return new ICmpInst(pred, X, Zero);
6847 // ((X & ~7) == 0) --> X < 8
6848 if (RHSV == 0 && isHighOnes(BOC)) {
6849 Value *X = BO->getOperand(0);
6850 Constant *NegX = ConstantExpr::getNeg(BOC);
6851 ICmpInst::Predicate pred = isICMP_NE ?
6852 ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
6853 return new ICmpInst(pred, X, NegX);
6858 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(LHSI)) {
6859 // Handle icmp {eq|ne} <intrinsic>, intcst.
6860 if (II->getIntrinsicID() == Intrinsic::bswap) {
6862 ICI.setOperand(0, II->getOperand(1));
6863 ICI.setOperand(1, ConstantInt::get(RHSV.byteSwap()));
6871 /// visitICmpInstWithCastAndCast - Handle icmp (cast x to y), (cast/cst).
6872 /// We only handle extending casts so far.
6874 Instruction *InstCombiner::visitICmpInstWithCastAndCast(ICmpInst &ICI) {
6875 const CastInst *LHSCI = cast<CastInst>(ICI.getOperand(0));
6876 Value *LHSCIOp = LHSCI->getOperand(0);
6877 const Type *SrcTy = LHSCIOp->getType();
6878 const Type *DestTy = LHSCI->getType();
6881 // Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the
6882 // integer type is the same size as the pointer type.
6883 if (LHSCI->getOpcode() == Instruction::PtrToInt &&
6884 getTargetData().getPointerSizeInBits() ==
6885 cast<IntegerType>(DestTy)->getBitWidth()) {
6887 if (Constant *RHSC = dyn_cast<Constant>(ICI.getOperand(1))) {
6888 RHSOp = ConstantExpr::getIntToPtr(RHSC, SrcTy);
6889 } else if (PtrToIntInst *RHSC = dyn_cast<PtrToIntInst>(ICI.getOperand(1))) {
6890 RHSOp = RHSC->getOperand(0);
6891 // If the pointer types don't match, insert a bitcast.
6892 if (LHSCIOp->getType() != RHSOp->getType())
6893 RHSOp = InsertBitCastBefore(RHSOp, LHSCIOp->getType(), ICI);
6897 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSOp);
6900 // The code below only handles extension cast instructions, so far.
6902 if (LHSCI->getOpcode() != Instruction::ZExt &&
6903 LHSCI->getOpcode() != Instruction::SExt)
6906 bool isSignedExt = LHSCI->getOpcode() == Instruction::SExt;
6907 bool isSignedCmp = ICI.isSignedPredicate();
6909 if (CastInst *CI = dyn_cast<CastInst>(ICI.getOperand(1))) {
6910 // Not an extension from the same type?
6911 RHSCIOp = CI->getOperand(0);
6912 if (RHSCIOp->getType() != LHSCIOp->getType())
6915 // If the signedness of the two casts doesn't agree (i.e. one is a sext
6916 // and the other is a zext), then we can't handle this.
6917 if (CI->getOpcode() != LHSCI->getOpcode())
6920 // Deal with equality cases early.
6921 if (ICI.isEquality())
6922 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
6924 // A signed comparison of sign extended values simplifies into a
6925 // signed comparison.
6926 if (isSignedCmp && isSignedExt)
6927 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
6929 // The other three cases all fold into an unsigned comparison.
6930 return new ICmpInst(ICI.getUnsignedPredicate(), LHSCIOp, RHSCIOp);
6933 // If we aren't dealing with a constant on the RHS, exit early
6934 ConstantInt *CI = dyn_cast<ConstantInt>(ICI.getOperand(1));
6938 // Compute the constant that would happen if we truncated to SrcTy then
6939 // reextended to DestTy.
6940 Constant *Res1 = ConstantExpr::getTrunc(CI, SrcTy);
6941 Constant *Res2 = ConstantExpr::getCast(LHSCI->getOpcode(), Res1, DestTy);
6943 // If the re-extended constant didn't change...
6945 // Make sure that sign of the Cmp and the sign of the Cast are the same.
6946 // For example, we might have:
6947 // %A = sext short %X to uint
6948 // %B = icmp ugt uint %A, 1330
6949 // It is incorrect to transform this into
6950 // %B = icmp ugt short %X, 1330
6951 // because %A may have negative value.
6953 // However, we allow this when the compare is EQ/NE, because they are
6955 if (isSignedExt == isSignedCmp || ICI.isEquality())
6956 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
6960 // The re-extended constant changed so the constant cannot be represented
6961 // in the shorter type. Consequently, we cannot emit a simple comparison.
6963 // First, handle some easy cases. We know the result cannot be equal at this
6964 // point so handle the ICI.isEquality() cases
6965 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
6966 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
6967 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
6968 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
6970 // Evaluate the comparison for LT (we invert for GT below). LE and GE cases
6971 // should have been folded away previously and not enter in here.
6974 // We're performing a signed comparison.
6975 if (cast<ConstantInt>(CI)->getValue().isNegative())
6976 Result = ConstantInt::getFalse(); // X < (small) --> false
6978 Result = ConstantInt::getTrue(); // X < (large) --> true
6980 // We're performing an unsigned comparison.
6982 // We're performing an unsigned comp with a sign extended value.
6983 // This is true if the input is >= 0. [aka >s -1]
6984 Constant *NegOne = ConstantInt::getAllOnesValue(SrcTy);
6985 Result = InsertNewInstBefore(new ICmpInst(ICmpInst::ICMP_SGT, LHSCIOp,
6986 NegOne, ICI.getName()), ICI);
6988 // Unsigned extend & unsigned compare -> always true.
6989 Result = ConstantInt::getTrue();
6993 // Finally, return the value computed.
6994 if (ICI.getPredicate() == ICmpInst::ICMP_ULT ||
6995 ICI.getPredicate() == ICmpInst::ICMP_SLT)
6996 return ReplaceInstUsesWith(ICI, Result);
6998 assert((ICI.getPredicate()==ICmpInst::ICMP_UGT ||
6999 ICI.getPredicate()==ICmpInst::ICMP_SGT) &&
7000 "ICmp should be folded!");
7001 if (Constant *CI = dyn_cast<Constant>(Result))
7002 return ReplaceInstUsesWith(ICI, ConstantExpr::getNot(CI));
7003 return BinaryOperator::CreateNot(Result);
7006 Instruction *InstCombiner::visitShl(BinaryOperator &I) {
7007 return commonShiftTransforms(I);
7010 Instruction *InstCombiner::visitLShr(BinaryOperator &I) {
7011 return commonShiftTransforms(I);
7014 Instruction *InstCombiner::visitAShr(BinaryOperator &I) {
7015 if (Instruction *R = commonShiftTransforms(I))
7018 Value *Op0 = I.getOperand(0);
7020 // ashr int -1, X = -1 (for any arithmetic shift rights of ~0)
7021 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
7022 if (CSI->isAllOnesValue())
7023 return ReplaceInstUsesWith(I, CSI);
7025 // See if we can turn a signed shr into an unsigned shr.
7026 if (!isa<VectorType>(I.getType()) &&
7027 MaskedValueIsZero(Op0,
7028 APInt::getSignBit(I.getType()->getPrimitiveSizeInBits())))
7029 return BinaryOperator::CreateLShr(Op0, I.getOperand(1));
7034 Instruction *InstCombiner::commonShiftTransforms(BinaryOperator &I) {
7035 assert(I.getOperand(1)->getType() == I.getOperand(0)->getType());
7036 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
7038 // shl X, 0 == X and shr X, 0 == X
7039 // shl 0, X == 0 and shr 0, X == 0
7040 if (Op1 == Constant::getNullValue(Op1->getType()) ||
7041 Op0 == Constant::getNullValue(Op0->getType()))
7042 return ReplaceInstUsesWith(I, Op0);
7044 if (isa<UndefValue>(Op0)) {
7045 if (I.getOpcode() == Instruction::AShr) // undef >>s X -> undef
7046 return ReplaceInstUsesWith(I, Op0);
7047 else // undef << X -> 0, undef >>u X -> 0
7048 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
7050 if (isa<UndefValue>(Op1)) {
7051 if (I.getOpcode() == Instruction::AShr) // X >>s undef -> X
7052 return ReplaceInstUsesWith(I, Op0);
7053 else // X << undef, X >>u undef -> 0
7054 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
7057 // Try to fold constant and into select arguments.
7058 if (isa<Constant>(Op0))
7059 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
7060 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
7063 if (ConstantInt *CUI = dyn_cast<ConstantInt>(Op1))
7064 if (Instruction *Res = FoldShiftByConstant(Op0, CUI, I))
7069 Instruction *InstCombiner::FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
7070 BinaryOperator &I) {
7071 bool isLeftShift = I.getOpcode() == Instruction::Shl;
7073 // See if we can simplify any instructions used by the instruction whose sole
7074 // purpose is to compute bits we don't care about.
7075 uint32_t TypeBits = Op0->getType()->getPrimitiveSizeInBits();
7076 if (SimplifyDemandedInstructionBits(I))
7079 // shl uint X, 32 = 0 and shr ubyte Y, 9 = 0, ... just don't eliminate shr
7080 // of a signed value.
7082 if (Op1->uge(TypeBits)) {
7083 if (I.getOpcode() != Instruction::AShr)
7084 return ReplaceInstUsesWith(I, Constant::getNullValue(Op0->getType()));
7086 I.setOperand(1, ConstantInt::get(I.getType(), TypeBits-1));
7091 // ((X*C1) << C2) == (X * (C1 << C2))
7092 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0))
7093 if (BO->getOpcode() == Instruction::Mul && isLeftShift)
7094 if (Constant *BOOp = dyn_cast<Constant>(BO->getOperand(1)))
7095 return BinaryOperator::CreateMul(BO->getOperand(0),
7096 ConstantExpr::getShl(BOOp, Op1));
7098 // Try to fold constant and into select arguments.
7099 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
7100 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
7102 if (isa<PHINode>(Op0))
7103 if (Instruction *NV = FoldOpIntoPhi(I))
7106 // Fold shift2(trunc(shift1(x,c1)), c2) -> trunc(shift2(shift1(x,c1),c2))
7107 if (TruncInst *TI = dyn_cast<TruncInst>(Op0)) {
7108 Instruction *TrOp = dyn_cast<Instruction>(TI->getOperand(0));
7109 // If 'shift2' is an ashr, we would have to get the sign bit into a funny
7110 // place. Don't try to do this transformation in this case. Also, we
7111 // require that the input operand is a shift-by-constant so that we have
7112 // confidence that the shifts will get folded together. We could do this
7113 // xform in more cases, but it is unlikely to be profitable.
7114 if (TrOp && I.isLogicalShift() && TrOp->isShift() &&
7115 isa<ConstantInt>(TrOp->getOperand(1))) {
7116 // Okay, we'll do this xform. Make the shift of shift.
7117 Constant *ShAmt = ConstantExpr::getZExt(Op1, TrOp->getType());
7118 Instruction *NSh = BinaryOperator::Create(I.getOpcode(), TrOp, ShAmt,
7120 InsertNewInstBefore(NSh, I); // (shift2 (shift1 & 0x00FF), c2)
7122 // For logical shifts, the truncation has the effect of making the high
7123 // part of the register be zeros. Emulate this by inserting an AND to
7124 // clear the top bits as needed. This 'and' will usually be zapped by
7125 // other xforms later if dead.
7126 unsigned SrcSize = TrOp->getType()->getPrimitiveSizeInBits();
7127 unsigned DstSize = TI->getType()->getPrimitiveSizeInBits();
7128 APInt MaskV(APInt::getLowBitsSet(SrcSize, DstSize));
7130 // The mask we constructed says what the trunc would do if occurring
7131 // between the shifts. We want to know the effect *after* the second
7132 // shift. We know that it is a logical shift by a constant, so adjust the
7133 // mask as appropriate.
7134 if (I.getOpcode() == Instruction::Shl)
7135 MaskV <<= Op1->getZExtValue();
7137 assert(I.getOpcode() == Instruction::LShr && "Unknown logical shift");
7138 MaskV = MaskV.lshr(Op1->getZExtValue());
7141 Instruction *And = BinaryOperator::CreateAnd(NSh, ConstantInt::get(MaskV),
7143 InsertNewInstBefore(And, I); // shift1 & 0x00FF
7145 // Return the value truncated to the interesting size.
7146 return new TruncInst(And, I.getType());
7150 if (Op0->hasOneUse()) {
7151 if (BinaryOperator *Op0BO = dyn_cast<BinaryOperator>(Op0)) {
7152 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
7155 switch (Op0BO->getOpcode()) {
7157 case Instruction::Add:
7158 case Instruction::And:
7159 case Instruction::Or:
7160 case Instruction::Xor: {
7161 // These operators commute.
7162 // Turn (Y + (X >> C)) << C -> (X + (Y << C)) & (~0 << C)
7163 if (isLeftShift && Op0BO->getOperand(1)->hasOneUse() &&
7164 match(Op0BO->getOperand(1), m_Shr(m_Value(V1), m_Specific(Op1)))){
7165 Instruction *YS = BinaryOperator::CreateShl(
7166 Op0BO->getOperand(0), Op1,
7168 InsertNewInstBefore(YS, I); // (Y << C)
7170 BinaryOperator::Create(Op0BO->getOpcode(), YS, V1,
7171 Op0BO->getOperand(1)->getName());
7172 InsertNewInstBefore(X, I); // (X + (Y << C))
7173 uint32_t Op1Val = Op1->getLimitedValue(TypeBits);
7174 return BinaryOperator::CreateAnd(X, ConstantInt::get(
7175 APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
7178 // Turn (Y + ((X >> C) & CC)) << C -> ((X & (CC << C)) + (Y << C))
7179 Value *Op0BOOp1 = Op0BO->getOperand(1);
7180 if (isLeftShift && Op0BOOp1->hasOneUse() &&
7182 m_And(m_Shr(m_Value(V1), m_Specific(Op1)),
7183 m_ConstantInt(CC))) &&
7184 cast<BinaryOperator>(Op0BOOp1)->getOperand(0)->hasOneUse()) {
7185 Instruction *YS = BinaryOperator::CreateShl(
7186 Op0BO->getOperand(0), Op1,
7188 InsertNewInstBefore(YS, I); // (Y << C)
7190 BinaryOperator::CreateAnd(V1, ConstantExpr::getShl(CC, Op1),
7191 V1->getName()+".mask");
7192 InsertNewInstBefore(XM, I); // X & (CC << C)
7194 return BinaryOperator::Create(Op0BO->getOpcode(), YS, XM);
7199 case Instruction::Sub: {
7200 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
7201 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
7202 match(Op0BO->getOperand(0), m_Shr(m_Value(V1), m_Specific(Op1)))){
7203 Instruction *YS = BinaryOperator::CreateShl(
7204 Op0BO->getOperand(1), Op1,
7206 InsertNewInstBefore(YS, I); // (Y << C)
7208 BinaryOperator::Create(Op0BO->getOpcode(), V1, YS,
7209 Op0BO->getOperand(0)->getName());
7210 InsertNewInstBefore(X, I); // (X + (Y << C))
7211 uint32_t Op1Val = Op1->getLimitedValue(TypeBits);
7212 return BinaryOperator::CreateAnd(X, ConstantInt::get(
7213 APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
7216 // Turn (((X >> C)&CC) + Y) << C -> (X + (Y << C)) & (CC << C)
7217 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
7218 match(Op0BO->getOperand(0),
7219 m_And(m_Shr(m_Value(V1), m_Value(V2)),
7220 m_ConstantInt(CC))) && V2 == Op1 &&
7221 cast<BinaryOperator>(Op0BO->getOperand(0))
7222 ->getOperand(0)->hasOneUse()) {
7223 Instruction *YS = BinaryOperator::CreateShl(
7224 Op0BO->getOperand(1), Op1,
7226 InsertNewInstBefore(YS, I); // (Y << C)
7228 BinaryOperator::CreateAnd(V1, ConstantExpr::getShl(CC, Op1),
7229 V1->getName()+".mask");
7230 InsertNewInstBefore(XM, I); // X & (CC << C)
7232 return BinaryOperator::Create(Op0BO->getOpcode(), XM, YS);
7240 // If the operand is an bitwise operator with a constant RHS, and the
7241 // shift is the only use, we can pull it out of the shift.
7242 if (ConstantInt *Op0C = dyn_cast<ConstantInt>(Op0BO->getOperand(1))) {
7243 bool isValid = true; // Valid only for And, Or, Xor
7244 bool highBitSet = false; // Transform if high bit of constant set?
7246 switch (Op0BO->getOpcode()) {
7247 default: isValid = false; break; // Do not perform transform!
7248 case Instruction::Add:
7249 isValid = isLeftShift;
7251 case Instruction::Or:
7252 case Instruction::Xor:
7255 case Instruction::And:
7260 // If this is a signed shift right, and the high bit is modified
7261 // by the logical operation, do not perform the transformation.
7262 // The highBitSet boolean indicates the value of the high bit of
7263 // the constant which would cause it to be modified for this
7266 if (isValid && I.getOpcode() == Instruction::AShr)
7267 isValid = Op0C->getValue()[TypeBits-1] == highBitSet;
7270 Constant *NewRHS = ConstantExpr::get(I.getOpcode(), Op0C, Op1);
7272 Instruction *NewShift =
7273 BinaryOperator::Create(I.getOpcode(), Op0BO->getOperand(0), Op1);
7274 InsertNewInstBefore(NewShift, I);
7275 NewShift->takeName(Op0BO);
7277 return BinaryOperator::Create(Op0BO->getOpcode(), NewShift,
7284 // Find out if this is a shift of a shift by a constant.
7285 BinaryOperator *ShiftOp = dyn_cast<BinaryOperator>(Op0);
7286 if (ShiftOp && !ShiftOp->isShift())
7289 if (ShiftOp && isa<ConstantInt>(ShiftOp->getOperand(1))) {
7290 ConstantInt *ShiftAmt1C = cast<ConstantInt>(ShiftOp->getOperand(1));
7291 uint32_t ShiftAmt1 = ShiftAmt1C->getLimitedValue(TypeBits);
7292 uint32_t ShiftAmt2 = Op1->getLimitedValue(TypeBits);
7293 assert(ShiftAmt2 != 0 && "Should have been simplified earlier");
7294 if (ShiftAmt1 == 0) return 0; // Will be simplified in the future.
7295 Value *X = ShiftOp->getOperand(0);
7297 uint32_t AmtSum = ShiftAmt1+ShiftAmt2; // Fold into one big shift.
7298 if (AmtSum > TypeBits)
7301 const IntegerType *Ty = cast<IntegerType>(I.getType());
7303 // Check for (X << c1) << c2 and (X >> c1) >> c2
7304 if (I.getOpcode() == ShiftOp->getOpcode()) {
7305 return BinaryOperator::Create(I.getOpcode(), X,
7306 ConstantInt::get(Ty, AmtSum));
7307 } else if (ShiftOp->getOpcode() == Instruction::LShr &&
7308 I.getOpcode() == Instruction::AShr) {
7309 // ((X >>u C1) >>s C2) -> (X >>u (C1+C2)) since C1 != 0.
7310 return BinaryOperator::CreateLShr(X, ConstantInt::get(Ty, AmtSum));
7311 } else if (ShiftOp->getOpcode() == Instruction::AShr &&
7312 I.getOpcode() == Instruction::LShr) {
7313 // ((X >>s C1) >>u C2) -> ((X >>s (C1+C2)) & mask) since C1 != 0.
7314 Instruction *Shift =
7315 BinaryOperator::CreateAShr(X, ConstantInt::get(Ty, AmtSum));
7316 InsertNewInstBefore(Shift, I);
7318 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
7319 return BinaryOperator::CreateAnd(Shift, ConstantInt::get(Mask));
7322 // Okay, if we get here, one shift must be left, and the other shift must be
7323 // right. See if the amounts are equal.
7324 if (ShiftAmt1 == ShiftAmt2) {
7325 // If we have ((X >>? C) << C), turn this into X & (-1 << C).
7326 if (I.getOpcode() == Instruction::Shl) {
7327 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt1));
7328 return BinaryOperator::CreateAnd(X, ConstantInt::get(Mask));
7330 // If we have ((X << C) >>u C), turn this into X & (-1 >>u C).
7331 if (I.getOpcode() == Instruction::LShr) {
7332 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt1));
7333 return BinaryOperator::CreateAnd(X, ConstantInt::get(Mask));
7335 // We can simplify ((X << C) >>s C) into a trunc + sext.
7336 // NOTE: we could do this for any C, but that would make 'unusual' integer
7337 // types. For now, just stick to ones well-supported by the code
7339 const Type *SExtType = 0;
7340 switch (Ty->getBitWidth() - ShiftAmt1) {
7347 SExtType = IntegerType::get(Ty->getBitWidth() - ShiftAmt1);
7352 Instruction *NewTrunc = new TruncInst(X, SExtType, "sext");
7353 InsertNewInstBefore(NewTrunc, I);
7354 return new SExtInst(NewTrunc, Ty);
7356 // Otherwise, we can't handle it yet.
7357 } else if (ShiftAmt1 < ShiftAmt2) {
7358 uint32_t ShiftDiff = ShiftAmt2-ShiftAmt1;
7360 // (X >>? C1) << C2 --> X << (C2-C1) & (-1 << C2)
7361 if (I.getOpcode() == Instruction::Shl) {
7362 assert(ShiftOp->getOpcode() == Instruction::LShr ||
7363 ShiftOp->getOpcode() == Instruction::AShr);
7364 Instruction *Shift =
7365 BinaryOperator::CreateShl(X, ConstantInt::get(Ty, ShiftDiff));
7366 InsertNewInstBefore(Shift, I);
7368 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
7369 return BinaryOperator::CreateAnd(Shift, ConstantInt::get(Mask));
7372 // (X << C1) >>u C2 --> X >>u (C2-C1) & (-1 >> C2)
7373 if (I.getOpcode() == Instruction::LShr) {
7374 assert(ShiftOp->getOpcode() == Instruction::Shl);
7375 Instruction *Shift =
7376 BinaryOperator::CreateLShr(X, ConstantInt::get(Ty, ShiftDiff));
7377 InsertNewInstBefore(Shift, I);
7379 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
7380 return BinaryOperator::CreateAnd(Shift, ConstantInt::get(Mask));
7383 // We can't handle (X << C1) >>s C2, it shifts arbitrary bits in.
7385 assert(ShiftAmt2 < ShiftAmt1);
7386 uint32_t ShiftDiff = ShiftAmt1-ShiftAmt2;
7388 // (X >>? C1) << C2 --> X >>? (C1-C2) & (-1 << C2)
7389 if (I.getOpcode() == Instruction::Shl) {
7390 assert(ShiftOp->getOpcode() == Instruction::LShr ||
7391 ShiftOp->getOpcode() == Instruction::AShr);
7392 Instruction *Shift =
7393 BinaryOperator::Create(ShiftOp->getOpcode(), X,
7394 ConstantInt::get(Ty, ShiftDiff));
7395 InsertNewInstBefore(Shift, I);
7397 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
7398 return BinaryOperator::CreateAnd(Shift, ConstantInt::get(Mask));
7401 // (X << C1) >>u C2 --> X << (C1-C2) & (-1 >> C2)
7402 if (I.getOpcode() == Instruction::LShr) {
7403 assert(ShiftOp->getOpcode() == Instruction::Shl);
7404 Instruction *Shift =
7405 BinaryOperator::CreateShl(X, ConstantInt::get(Ty, ShiftDiff));
7406 InsertNewInstBefore(Shift, I);
7408 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
7409 return BinaryOperator::CreateAnd(Shift, ConstantInt::get(Mask));
7412 // We can't handle (X << C1) >>a C2, it shifts arbitrary bits in.
7419 /// DecomposeSimpleLinearExpr - Analyze 'Val', seeing if it is a simple linear
7420 /// expression. If so, decompose it, returning some value X, such that Val is
7423 static Value *DecomposeSimpleLinearExpr(Value *Val, unsigned &Scale,
7425 assert(Val->getType() == Type::Int32Ty && "Unexpected allocation size type!");
7426 if (ConstantInt *CI = dyn_cast<ConstantInt>(Val)) {
7427 Offset = CI->getZExtValue();
7429 return ConstantInt::get(Type::Int32Ty, 0);
7430 } else if (BinaryOperator *I = dyn_cast<BinaryOperator>(Val)) {
7431 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
7432 if (I->getOpcode() == Instruction::Shl) {
7433 // This is a value scaled by '1 << the shift amt'.
7434 Scale = 1U << RHS->getZExtValue();
7436 return I->getOperand(0);
7437 } else if (I->getOpcode() == Instruction::Mul) {
7438 // This value is scaled by 'RHS'.
7439 Scale = RHS->getZExtValue();
7441 return I->getOperand(0);
7442 } else if (I->getOpcode() == Instruction::Add) {
7443 // We have X+C. Check to see if we really have (X*C2)+C1,
7444 // where C1 is divisible by C2.
7447 DecomposeSimpleLinearExpr(I->getOperand(0), SubScale, Offset);
7448 Offset += RHS->getZExtValue();
7455 // Otherwise, we can't look past this.
7462 /// PromoteCastOfAllocation - If we find a cast of an allocation instruction,
7463 /// try to eliminate the cast by moving the type information into the alloc.
7464 Instruction *InstCombiner::PromoteCastOfAllocation(BitCastInst &CI,
7465 AllocationInst &AI) {
7466 const PointerType *PTy = cast<PointerType>(CI.getType());
7468 // Remove any uses of AI that are dead.
7469 assert(!CI.use_empty() && "Dead instructions should be removed earlier!");
7471 for (Value::use_iterator UI = AI.use_begin(), E = AI.use_end(); UI != E; ) {
7472 Instruction *User = cast<Instruction>(*UI++);
7473 if (isInstructionTriviallyDead(User)) {
7474 while (UI != E && *UI == User)
7475 ++UI; // If this instruction uses AI more than once, don't break UI.
7478 DOUT << "IC: DCE: " << *User;
7479 EraseInstFromFunction(*User);
7483 // Get the type really allocated and the type casted to.
7484 const Type *AllocElTy = AI.getAllocatedType();
7485 const Type *CastElTy = PTy->getElementType();
7486 if (!AllocElTy->isSized() || !CastElTy->isSized()) return 0;
7488 unsigned AllocElTyAlign = TD->getABITypeAlignment(AllocElTy);
7489 unsigned CastElTyAlign = TD->getABITypeAlignment(CastElTy);
7490 if (CastElTyAlign < AllocElTyAlign) return 0;
7492 // If the allocation has multiple uses, only promote it if we are strictly
7493 // increasing the alignment of the resultant allocation. If we keep it the
7494 // same, we open the door to infinite loops of various kinds.
7495 if (!AI.hasOneUse() && CastElTyAlign == AllocElTyAlign) return 0;
7497 uint64_t AllocElTySize = TD->getTypePaddedSize(AllocElTy);
7498 uint64_t CastElTySize = TD->getTypePaddedSize(CastElTy);
7499 if (CastElTySize == 0 || AllocElTySize == 0) return 0;
7501 // See if we can satisfy the modulus by pulling a scale out of the array
7503 unsigned ArraySizeScale;
7505 Value *NumElements = // See if the array size is a decomposable linear expr.
7506 DecomposeSimpleLinearExpr(AI.getOperand(0), ArraySizeScale, ArrayOffset);
7508 // If we can now satisfy the modulus, by using a non-1 scale, we really can
7510 if ((AllocElTySize*ArraySizeScale) % CastElTySize != 0 ||
7511 (AllocElTySize*ArrayOffset ) % CastElTySize != 0) return 0;
7513 unsigned Scale = (AllocElTySize*ArraySizeScale)/CastElTySize;
7518 // If the allocation size is constant, form a constant mul expression
7519 Amt = ConstantInt::get(Type::Int32Ty, Scale);
7520 if (isa<ConstantInt>(NumElements))
7521 Amt = Multiply(cast<ConstantInt>(NumElements), cast<ConstantInt>(Amt));
7522 // otherwise multiply the amount and the number of elements
7523 else if (Scale != 1) {
7524 Instruction *Tmp = BinaryOperator::CreateMul(Amt, NumElements, "tmp");
7525 Amt = InsertNewInstBefore(Tmp, AI);
7529 if (int Offset = (AllocElTySize*ArrayOffset)/CastElTySize) {
7530 Value *Off = ConstantInt::get(Type::Int32Ty, Offset, true);
7531 Instruction *Tmp = BinaryOperator::CreateAdd(Amt, Off, "tmp");
7532 Amt = InsertNewInstBefore(Tmp, AI);
7535 AllocationInst *New;
7536 if (isa<MallocInst>(AI))
7537 New = new MallocInst(CastElTy, Amt, AI.getAlignment());
7539 New = new AllocaInst(CastElTy, Amt, AI.getAlignment());
7540 InsertNewInstBefore(New, AI);
7543 // If the allocation has multiple uses, insert a cast and change all things
7544 // that used it to use the new cast. This will also hack on CI, but it will
7546 if (!AI.hasOneUse()) {
7547 AddUsesToWorkList(AI);
7548 // New is the allocation instruction, pointer typed. AI is the original
7549 // allocation instruction, also pointer typed. Thus, cast to use is BitCast.
7550 CastInst *NewCast = new BitCastInst(New, AI.getType(), "tmpcast");
7551 InsertNewInstBefore(NewCast, AI);
7552 AI.replaceAllUsesWith(NewCast);
7554 return ReplaceInstUsesWith(CI, New);
7557 /// CanEvaluateInDifferentType - Return true if we can take the specified value
7558 /// and return it as type Ty without inserting any new casts and without
7559 /// changing the computed value. This is used by code that tries to decide
7560 /// whether promoting or shrinking integer operations to wider or smaller types
7561 /// will allow us to eliminate a truncate or extend.
7563 /// This is a truncation operation if Ty is smaller than V->getType(), or an
7564 /// extension operation if Ty is larger.
7566 /// If CastOpc is a truncation, then Ty will be a type smaller than V. We
7567 /// should return true if trunc(V) can be computed by computing V in the smaller
7568 /// type. If V is an instruction, then trunc(inst(x,y)) can be computed as
7569 /// inst(trunc(x),trunc(y)), which only makes sense if x and y can be
7570 /// efficiently truncated.
7572 /// If CastOpc is a sext or zext, we are asking if the low bits of the value can
7573 /// bit computed in a larger type, which is then and'd or sext_in_reg'd to get
7574 /// the final result.
7575 bool InstCombiner::CanEvaluateInDifferentType(Value *V, const IntegerType *Ty,
7577 int &NumCastsRemoved){
7578 // We can always evaluate constants in another type.
7579 if (isa<ConstantInt>(V))
7582 Instruction *I = dyn_cast<Instruction>(V);
7583 if (!I) return false;
7585 const IntegerType *OrigTy = cast<IntegerType>(V->getType());
7587 // If this is an extension or truncate, we can often eliminate it.
7588 if (isa<TruncInst>(I) || isa<ZExtInst>(I) || isa<SExtInst>(I)) {
7589 // If this is a cast from the destination type, we can trivially eliminate
7590 // it, and this will remove a cast overall.
7591 if (I->getOperand(0)->getType() == Ty) {
7592 // If the first operand is itself a cast, and is eliminable, do not count
7593 // this as an eliminable cast. We would prefer to eliminate those two
7595 if (!isa<CastInst>(I->getOperand(0)) && I->hasOneUse())
7601 // We can't extend or shrink something that has multiple uses: doing so would
7602 // require duplicating the instruction in general, which isn't profitable.
7603 if (!I->hasOneUse()) return false;
7605 unsigned Opc = I->getOpcode();
7607 case Instruction::Add:
7608 case Instruction::Sub:
7609 case Instruction::Mul:
7610 case Instruction::And:
7611 case Instruction::Or:
7612 case Instruction::Xor:
7613 // These operators can all arbitrarily be extended or truncated.
7614 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
7616 CanEvaluateInDifferentType(I->getOperand(1), Ty, CastOpc,
7619 case Instruction::Shl:
7620 // If we are truncating the result of this SHL, and if it's a shift of a
7621 // constant amount, we can always perform a SHL in a smaller type.
7622 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
7623 uint32_t BitWidth = Ty->getBitWidth();
7624 if (BitWidth < OrigTy->getBitWidth() &&
7625 CI->getLimitedValue(BitWidth) < BitWidth)
7626 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
7630 case Instruction::LShr:
7631 // If this is a truncate of a logical shr, we can truncate it to a smaller
7632 // lshr iff we know that the bits we would otherwise be shifting in are
7634 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
7635 uint32_t OrigBitWidth = OrigTy->getBitWidth();
7636 uint32_t BitWidth = Ty->getBitWidth();
7637 if (BitWidth < OrigBitWidth &&
7638 MaskedValueIsZero(I->getOperand(0),
7639 APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth)) &&
7640 CI->getLimitedValue(BitWidth) < BitWidth) {
7641 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
7646 case Instruction::ZExt:
7647 case Instruction::SExt:
7648 case Instruction::Trunc:
7649 // If this is the same kind of case as our original (e.g. zext+zext), we
7650 // can safely replace it. Note that replacing it does not reduce the number
7651 // of casts in the input.
7655 // sext (zext ty1), ty2 -> zext ty2
7656 if (CastOpc == Instruction::SExt && Opc == Instruction::ZExt)
7659 case Instruction::Select: {
7660 SelectInst *SI = cast<SelectInst>(I);
7661 return CanEvaluateInDifferentType(SI->getTrueValue(), Ty, CastOpc,
7663 CanEvaluateInDifferentType(SI->getFalseValue(), Ty, CastOpc,
7666 case Instruction::PHI: {
7667 // We can change a phi if we can change all operands.
7668 PHINode *PN = cast<PHINode>(I);
7669 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
7670 if (!CanEvaluateInDifferentType(PN->getIncomingValue(i), Ty, CastOpc,
7676 // TODO: Can handle more cases here.
7683 /// EvaluateInDifferentType - Given an expression that
7684 /// CanEvaluateInDifferentType returns true for, actually insert the code to
7685 /// evaluate the expression.
7686 Value *InstCombiner::EvaluateInDifferentType(Value *V, const Type *Ty,
7688 if (Constant *C = dyn_cast<Constant>(V))
7689 return ConstantExpr::getIntegerCast(C, Ty, isSigned /*Sext or ZExt*/);
7691 // Otherwise, it must be an instruction.
7692 Instruction *I = cast<Instruction>(V);
7693 Instruction *Res = 0;
7694 unsigned Opc = I->getOpcode();
7696 case Instruction::Add:
7697 case Instruction::Sub:
7698 case Instruction::Mul:
7699 case Instruction::And:
7700 case Instruction::Or:
7701 case Instruction::Xor:
7702 case Instruction::AShr:
7703 case Instruction::LShr:
7704 case Instruction::Shl: {
7705 Value *LHS = EvaluateInDifferentType(I->getOperand(0), Ty, isSigned);
7706 Value *RHS = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
7707 Res = BinaryOperator::Create((Instruction::BinaryOps)Opc, LHS, RHS);
7710 case Instruction::Trunc:
7711 case Instruction::ZExt:
7712 case Instruction::SExt:
7713 // If the source type of the cast is the type we're trying for then we can
7714 // just return the source. There's no need to insert it because it is not
7716 if (I->getOperand(0)->getType() == Ty)
7717 return I->getOperand(0);
7719 // Otherwise, must be the same type of cast, so just reinsert a new one.
7720 Res = CastInst::Create(cast<CastInst>(I)->getOpcode(), I->getOperand(0),
7723 case Instruction::Select: {
7724 Value *True = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
7725 Value *False = EvaluateInDifferentType(I->getOperand(2), Ty, isSigned);
7726 Res = SelectInst::Create(I->getOperand(0), True, False);
7729 case Instruction::PHI: {
7730 PHINode *OPN = cast<PHINode>(I);
7731 PHINode *NPN = PHINode::Create(Ty);
7732 for (unsigned i = 0, e = OPN->getNumIncomingValues(); i != e; ++i) {
7733 Value *V =EvaluateInDifferentType(OPN->getIncomingValue(i), Ty, isSigned);
7734 NPN->addIncoming(V, OPN->getIncomingBlock(i));
7740 // TODO: Can handle more cases here.
7741 assert(0 && "Unreachable!");
7746 return InsertNewInstBefore(Res, *I);
7749 /// @brief Implement the transforms common to all CastInst visitors.
7750 Instruction *InstCombiner::commonCastTransforms(CastInst &CI) {
7751 Value *Src = CI.getOperand(0);
7753 // Many cases of "cast of a cast" are eliminable. If it's eliminable we just
7754 // eliminate it now.
7755 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
7756 if (Instruction::CastOps opc =
7757 isEliminableCastPair(CSrc, CI.getOpcode(), CI.getType(), TD)) {
7758 // The first cast (CSrc) is eliminable so we need to fix up or replace
7759 // the second cast (CI). CSrc will then have a good chance of being dead.
7760 return CastInst::Create(opc, CSrc->getOperand(0), CI.getType());
7764 // If we are casting a select then fold the cast into the select
7765 if (SelectInst *SI = dyn_cast<SelectInst>(Src))
7766 if (Instruction *NV = FoldOpIntoSelect(CI, SI, this))
7769 // If we are casting a PHI then fold the cast into the PHI
7770 if (isa<PHINode>(Src))
7771 if (Instruction *NV = FoldOpIntoPhi(CI))
7777 /// FindElementAtOffset - Given a type and a constant offset, determine whether
7778 /// or not there is a sequence of GEP indices into the type that will land us at
7779 /// the specified offset. If so, fill them into NewIndices and return the
7780 /// resultant element type, otherwise return null.
7781 static const Type *FindElementAtOffset(const Type *Ty, int64_t Offset,
7782 SmallVectorImpl<Value*> &NewIndices,
7783 const TargetData *TD) {
7784 if (!Ty->isSized()) return 0;
7786 // Start with the index over the outer type. Note that the type size
7787 // might be zero (even if the offset isn't zero) if the indexed type
7788 // is something like [0 x {int, int}]
7789 const Type *IntPtrTy = TD->getIntPtrType();
7790 int64_t FirstIdx = 0;
7791 if (int64_t TySize = TD->getTypePaddedSize(Ty)) {
7792 FirstIdx = Offset/TySize;
7793 Offset -= FirstIdx*TySize;
7795 // Handle hosts where % returns negative instead of values [0..TySize).
7799 assert(Offset >= 0);
7801 assert((uint64_t)Offset < (uint64_t)TySize && "Out of range offset");
7804 NewIndices.push_back(ConstantInt::get(IntPtrTy, FirstIdx));
7806 // Index into the types. If we fail, set OrigBase to null.
7808 // Indexing into tail padding between struct/array elements.
7809 if (uint64_t(Offset*8) >= TD->getTypeSizeInBits(Ty))
7812 if (const StructType *STy = dyn_cast<StructType>(Ty)) {
7813 const StructLayout *SL = TD->getStructLayout(STy);
7814 assert(Offset < (int64_t)SL->getSizeInBytes() &&
7815 "Offset must stay within the indexed type");
7817 unsigned Elt = SL->getElementContainingOffset(Offset);
7818 NewIndices.push_back(ConstantInt::get(Type::Int32Ty, Elt));
7820 Offset -= SL->getElementOffset(Elt);
7821 Ty = STy->getElementType(Elt);
7822 } else if (const ArrayType *AT = dyn_cast<ArrayType>(Ty)) {
7823 uint64_t EltSize = TD->getTypePaddedSize(AT->getElementType());
7824 assert(EltSize && "Cannot index into a zero-sized array");
7825 NewIndices.push_back(ConstantInt::get(IntPtrTy,Offset/EltSize));
7827 Ty = AT->getElementType();
7829 // Otherwise, we can't index into the middle of this atomic type, bail.
7837 /// @brief Implement the transforms for cast of pointer (bitcast/ptrtoint)
7838 Instruction *InstCombiner::commonPointerCastTransforms(CastInst &CI) {
7839 Value *Src = CI.getOperand(0);
7841 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Src)) {
7842 // If casting the result of a getelementptr instruction with no offset, turn
7843 // this into a cast of the original pointer!
7844 if (GEP->hasAllZeroIndices()) {
7845 // Changing the cast operand is usually not a good idea but it is safe
7846 // here because the pointer operand is being replaced with another
7847 // pointer operand so the opcode doesn't need to change.
7849 CI.setOperand(0, GEP->getOperand(0));
7853 // If the GEP has a single use, and the base pointer is a bitcast, and the
7854 // GEP computes a constant offset, see if we can convert these three
7855 // instructions into fewer. This typically happens with unions and other
7856 // non-type-safe code.
7857 if (GEP->hasOneUse() && isa<BitCastInst>(GEP->getOperand(0))) {
7858 if (GEP->hasAllConstantIndices()) {
7859 // We are guaranteed to get a constant from EmitGEPOffset.
7860 ConstantInt *OffsetV = cast<ConstantInt>(EmitGEPOffset(GEP, CI, *this));
7861 int64_t Offset = OffsetV->getSExtValue();
7863 // Get the base pointer input of the bitcast, and the type it points to.
7864 Value *OrigBase = cast<BitCastInst>(GEP->getOperand(0))->getOperand(0);
7865 const Type *GEPIdxTy =
7866 cast<PointerType>(OrigBase->getType())->getElementType();
7867 SmallVector<Value*, 8> NewIndices;
7868 if (FindElementAtOffset(GEPIdxTy, Offset, NewIndices, TD)) {
7869 // If we were able to index down into an element, create the GEP
7870 // and bitcast the result. This eliminates one bitcast, potentially
7872 Instruction *NGEP = GetElementPtrInst::Create(OrigBase,
7874 NewIndices.end(), "");
7875 InsertNewInstBefore(NGEP, CI);
7876 NGEP->takeName(GEP);
7878 if (isa<BitCastInst>(CI))
7879 return new BitCastInst(NGEP, CI.getType());
7880 assert(isa<PtrToIntInst>(CI));
7881 return new PtrToIntInst(NGEP, CI.getType());
7887 return commonCastTransforms(CI);
7891 /// Only the TRUNC, ZEXT, SEXT, and BITCAST can both operand and result as
7892 /// integer types. This function implements the common transforms for all those
7894 /// @brief Implement the transforms common to CastInst with integer operands
7895 Instruction *InstCombiner::commonIntCastTransforms(CastInst &CI) {
7896 if (Instruction *Result = commonCastTransforms(CI))
7899 Value *Src = CI.getOperand(0);
7900 const Type *SrcTy = Src->getType();
7901 const Type *DestTy = CI.getType();
7902 uint32_t SrcBitSize = SrcTy->getPrimitiveSizeInBits();
7903 uint32_t DestBitSize = DestTy->getPrimitiveSizeInBits();
7905 // See if we can simplify any instructions used by the LHS whose sole
7906 // purpose is to compute bits we don't care about.
7907 if (SimplifyDemandedInstructionBits(CI))
7910 // If the source isn't an instruction or has more than one use then we
7911 // can't do anything more.
7912 Instruction *SrcI = dyn_cast<Instruction>(Src);
7913 if (!SrcI || !Src->hasOneUse())
7916 // Attempt to propagate the cast into the instruction for int->int casts.
7917 int NumCastsRemoved = 0;
7918 if (!isa<BitCastInst>(CI) &&
7919 CanEvaluateInDifferentType(SrcI, cast<IntegerType>(DestTy),
7920 CI.getOpcode(), NumCastsRemoved)) {
7921 // If this cast is a truncate, evaluting in a different type always
7922 // eliminates the cast, so it is always a win. If this is a zero-extension,
7923 // we need to do an AND to maintain the clear top-part of the computation,
7924 // so we require that the input have eliminated at least one cast. If this
7925 // is a sign extension, we insert two new casts (to do the extension) so we
7926 // require that two casts have been eliminated.
7927 bool DoXForm = false;
7928 bool JustReplace = false;
7929 switch (CI.getOpcode()) {
7931 // All the others use floating point so we shouldn't actually
7932 // get here because of the check above.
7933 assert(0 && "Unknown cast type");
7934 case Instruction::Trunc:
7937 case Instruction::ZExt: {
7938 DoXForm = NumCastsRemoved >= 1;
7939 if (!DoXForm && 0) {
7940 // If it's unnecessary to issue an AND to clear the high bits, it's
7941 // always profitable to do this xform.
7942 Value *TryRes = EvaluateInDifferentType(SrcI, DestTy, false);
7943 APInt Mask(APInt::getBitsSet(DestBitSize, SrcBitSize, DestBitSize));
7944 if (MaskedValueIsZero(TryRes, Mask))
7945 return ReplaceInstUsesWith(CI, TryRes);
7947 if (Instruction *TryI = dyn_cast<Instruction>(TryRes))
7948 if (TryI->use_empty())
7949 EraseInstFromFunction(*TryI);
7953 case Instruction::SExt: {
7954 DoXForm = NumCastsRemoved >= 2;
7955 if (!DoXForm && !isa<TruncInst>(SrcI) && 0) {
7956 // If we do not have to emit the truncate + sext pair, then it's always
7957 // profitable to do this xform.
7959 // It's not safe to eliminate the trunc + sext pair if one of the
7960 // eliminated cast is a truncate. e.g.
7961 // t2 = trunc i32 t1 to i16
7962 // t3 = sext i16 t2 to i32
7965 Value *TryRes = EvaluateInDifferentType(SrcI, DestTy, true);
7966 unsigned NumSignBits = ComputeNumSignBits(TryRes);
7967 if (NumSignBits > (DestBitSize - SrcBitSize))
7968 return ReplaceInstUsesWith(CI, TryRes);
7970 if (Instruction *TryI = dyn_cast<Instruction>(TryRes))
7971 if (TryI->use_empty())
7972 EraseInstFromFunction(*TryI);
7979 DOUT << "ICE: EvaluateInDifferentType converting expression type to avoid"
7981 Value *Res = EvaluateInDifferentType(SrcI, DestTy,
7982 CI.getOpcode() == Instruction::SExt);
7984 // Just replace this cast with the result.
7985 return ReplaceInstUsesWith(CI, Res);
7987 assert(Res->getType() == DestTy);
7988 switch (CI.getOpcode()) {
7989 default: assert(0 && "Unknown cast type!");
7990 case Instruction::Trunc:
7991 case Instruction::BitCast:
7992 // Just replace this cast with the result.
7993 return ReplaceInstUsesWith(CI, Res);
7994 case Instruction::ZExt: {
7995 assert(SrcBitSize < DestBitSize && "Not a zext?");
7997 // If the high bits are already zero, just replace this cast with the
7999 APInt Mask(APInt::getBitsSet(DestBitSize, SrcBitSize, DestBitSize));
8000 if (MaskedValueIsZero(Res, Mask))
8001 return ReplaceInstUsesWith(CI, Res);
8003 // We need to emit an AND to clear the high bits.
8004 Constant *C = ConstantInt::get(APInt::getLowBitsSet(DestBitSize,
8006 return BinaryOperator::CreateAnd(Res, C);
8008 case Instruction::SExt: {
8009 // If the high bits are already filled with sign bit, just replace this
8010 // cast with the result.
8011 unsigned NumSignBits = ComputeNumSignBits(Res);
8012 if (NumSignBits > (DestBitSize - SrcBitSize))
8013 return ReplaceInstUsesWith(CI, Res);
8015 // We need to emit a cast to truncate, then a cast to sext.
8016 return CastInst::Create(Instruction::SExt,
8017 InsertCastBefore(Instruction::Trunc, Res, Src->getType(),
8024 Value *Op0 = SrcI->getNumOperands() > 0 ? SrcI->getOperand(0) : 0;
8025 Value *Op1 = SrcI->getNumOperands() > 1 ? SrcI->getOperand(1) : 0;
8027 switch (SrcI->getOpcode()) {
8028 case Instruction::Add:
8029 case Instruction::Mul:
8030 case Instruction::And:
8031 case Instruction::Or:
8032 case Instruction::Xor:
8033 // If we are discarding information, rewrite.
8034 if (DestBitSize <= SrcBitSize && DestBitSize != 1) {
8035 // Don't insert two casts if they cannot be eliminated. We allow
8036 // two casts to be inserted if the sizes are the same. This could
8037 // only be converting signedness, which is a noop.
8038 if (DestBitSize == SrcBitSize ||
8039 !ValueRequiresCast(CI.getOpcode(), Op1, DestTy,TD) ||
8040 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
8041 Instruction::CastOps opcode = CI.getOpcode();
8042 Value *Op0c = InsertCastBefore(opcode, Op0, DestTy, *SrcI);
8043 Value *Op1c = InsertCastBefore(opcode, Op1, DestTy, *SrcI);
8044 return BinaryOperator::Create(
8045 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
8049 // cast (xor bool X, true) to int --> xor (cast bool X to int), 1
8050 if (isa<ZExtInst>(CI) && SrcBitSize == 1 &&
8051 SrcI->getOpcode() == Instruction::Xor &&
8052 Op1 == ConstantInt::getTrue() &&
8053 (!Op0->hasOneUse() || !isa<CmpInst>(Op0))) {
8054 Value *New = InsertCastBefore(Instruction::ZExt, Op0, DestTy, CI);
8055 return BinaryOperator::CreateXor(New, ConstantInt::get(CI.getType(), 1));
8058 case Instruction::SDiv:
8059 case Instruction::UDiv:
8060 case Instruction::SRem:
8061 case Instruction::URem:
8062 // If we are just changing the sign, rewrite.
8063 if (DestBitSize == SrcBitSize) {
8064 // Don't insert two casts if they cannot be eliminated. We allow
8065 // two casts to be inserted if the sizes are the same. This could
8066 // only be converting signedness, which is a noop.
8067 if (!ValueRequiresCast(CI.getOpcode(), Op1, DestTy, TD) ||
8068 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
8069 Value *Op0c = InsertCastBefore(Instruction::BitCast,
8070 Op0, DestTy, *SrcI);
8071 Value *Op1c = InsertCastBefore(Instruction::BitCast,
8072 Op1, DestTy, *SrcI);
8073 return BinaryOperator::Create(
8074 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
8079 case Instruction::Shl:
8080 // Allow changing the sign of the source operand. Do not allow
8081 // changing the size of the shift, UNLESS the shift amount is a
8082 // constant. We must not change variable sized shifts to a smaller
8083 // size, because it is undefined to shift more bits out than exist
8085 if (DestBitSize == SrcBitSize ||
8086 (DestBitSize < SrcBitSize && isa<Constant>(Op1))) {
8087 Instruction::CastOps opcode = (DestBitSize == SrcBitSize ?
8088 Instruction::BitCast : Instruction::Trunc);
8089 Value *Op0c = InsertCastBefore(opcode, Op0, DestTy, *SrcI);
8090 Value *Op1c = InsertCastBefore(opcode, Op1, DestTy, *SrcI);
8091 return BinaryOperator::CreateShl(Op0c, Op1c);
8094 case Instruction::AShr:
8095 // If this is a signed shr, and if all bits shifted in are about to be
8096 // truncated off, turn it into an unsigned shr to allow greater
8098 if (DestBitSize < SrcBitSize &&
8099 isa<ConstantInt>(Op1)) {
8100 uint32_t ShiftAmt = cast<ConstantInt>(Op1)->getLimitedValue(SrcBitSize);
8101 if (SrcBitSize > ShiftAmt && SrcBitSize-ShiftAmt >= DestBitSize) {
8102 // Insert the new logical shift right.
8103 return BinaryOperator::CreateLShr(Op0, Op1);
8111 Instruction *InstCombiner::visitTrunc(TruncInst &CI) {
8112 if (Instruction *Result = commonIntCastTransforms(CI))
8115 Value *Src = CI.getOperand(0);
8116 const Type *Ty = CI.getType();
8117 uint32_t DestBitWidth = Ty->getPrimitiveSizeInBits();
8118 uint32_t SrcBitWidth = cast<IntegerType>(Src->getType())->getBitWidth();
8120 if (Instruction *SrcI = dyn_cast<Instruction>(Src)) {
8121 switch (SrcI->getOpcode()) {
8123 case Instruction::LShr:
8124 // We can shrink lshr to something smaller if we know the bits shifted in
8125 // are already zeros.
8126 if (ConstantInt *ShAmtV = dyn_cast<ConstantInt>(SrcI->getOperand(1))) {
8127 uint32_t ShAmt = ShAmtV->getLimitedValue(SrcBitWidth);
8129 // Get a mask for the bits shifting in.
8130 APInt Mask(APInt::getLowBitsSet(SrcBitWidth, ShAmt).shl(DestBitWidth));
8131 Value* SrcIOp0 = SrcI->getOperand(0);
8132 if (SrcI->hasOneUse() && MaskedValueIsZero(SrcIOp0, Mask)) {
8133 if (ShAmt >= DestBitWidth) // All zeros.
8134 return ReplaceInstUsesWith(CI, Constant::getNullValue(Ty));
8136 // Okay, we can shrink this. Truncate the input, then return a new
8138 Value *V1 = InsertCastBefore(Instruction::Trunc, SrcIOp0, Ty, CI);
8139 Value *V2 = InsertCastBefore(Instruction::Trunc, SrcI->getOperand(1),
8141 return BinaryOperator::CreateLShr(V1, V2);
8143 } else { // This is a variable shr.
8145 // Turn 'trunc (lshr X, Y) to bool' into '(X & (1 << Y)) != 0'. This is
8146 // more LLVM instructions, but allows '1 << Y' to be hoisted if
8147 // loop-invariant and CSE'd.
8148 if (CI.getType() == Type::Int1Ty && SrcI->hasOneUse()) {
8149 Value *One = ConstantInt::get(SrcI->getType(), 1);
8151 Value *V = InsertNewInstBefore(
8152 BinaryOperator::CreateShl(One, SrcI->getOperand(1),
8154 V = InsertNewInstBefore(BinaryOperator::CreateAnd(V,
8155 SrcI->getOperand(0),
8157 Value *Zero = Constant::getNullValue(V->getType());
8158 return new ICmpInst(ICmpInst::ICMP_NE, V, Zero);
8168 /// transformZExtICmp - Transform (zext icmp) to bitwise / integer operations
8169 /// in order to eliminate the icmp.
8170 Instruction *InstCombiner::transformZExtICmp(ICmpInst *ICI, Instruction &CI,
8172 // If we are just checking for a icmp eq of a single bit and zext'ing it
8173 // to an integer, then shift the bit to the appropriate place and then
8174 // cast to integer to avoid the comparison.
8175 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(ICI->getOperand(1))) {
8176 const APInt &Op1CV = Op1C->getValue();
8178 // zext (x <s 0) to i32 --> x>>u31 true if signbit set.
8179 // zext (x >s -1) to i32 --> (x>>u31)^1 true if signbit clear.
8180 if ((ICI->getPredicate() == ICmpInst::ICMP_SLT && Op1CV == 0) ||
8181 (ICI->getPredicate() == ICmpInst::ICMP_SGT &&Op1CV.isAllOnesValue())) {
8182 if (!DoXform) return ICI;
8184 Value *In = ICI->getOperand(0);
8185 Value *Sh = ConstantInt::get(In->getType(),
8186 In->getType()->getPrimitiveSizeInBits()-1);
8187 In = InsertNewInstBefore(BinaryOperator::CreateLShr(In, Sh,
8188 In->getName()+".lobit"),
8190 if (In->getType() != CI.getType())
8191 In = CastInst::CreateIntegerCast(In, CI.getType(),
8192 false/*ZExt*/, "tmp", &CI);
8194 if (ICI->getPredicate() == ICmpInst::ICMP_SGT) {
8195 Constant *One = ConstantInt::get(In->getType(), 1);
8196 In = InsertNewInstBefore(BinaryOperator::CreateXor(In, One,
8197 In->getName()+".not"),
8201 return ReplaceInstUsesWith(CI, In);
8206 // zext (X == 0) to i32 --> X^1 iff X has only the low bit set.
8207 // zext (X == 0) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
8208 // zext (X == 1) to i32 --> X iff X has only the low bit set.
8209 // zext (X == 2) to i32 --> X>>1 iff X has only the 2nd bit set.
8210 // zext (X != 0) to i32 --> X iff X has only the low bit set.
8211 // zext (X != 0) to i32 --> X>>1 iff X has only the 2nd bit set.
8212 // zext (X != 1) to i32 --> X^1 iff X has only the low bit set.
8213 // zext (X != 2) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
8214 if ((Op1CV == 0 || Op1CV.isPowerOf2()) &&
8215 // This only works for EQ and NE
8216 ICI->isEquality()) {
8217 // If Op1C some other power of two, convert:
8218 uint32_t BitWidth = Op1C->getType()->getBitWidth();
8219 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
8220 APInt TypeMask(APInt::getAllOnesValue(BitWidth));
8221 ComputeMaskedBits(ICI->getOperand(0), TypeMask, KnownZero, KnownOne);
8223 APInt KnownZeroMask(~KnownZero);
8224 if (KnownZeroMask.isPowerOf2()) { // Exactly 1 possible 1?
8225 if (!DoXform) return ICI;
8227 bool isNE = ICI->getPredicate() == ICmpInst::ICMP_NE;
8228 if (Op1CV != 0 && (Op1CV != KnownZeroMask)) {
8229 // (X&4) == 2 --> false
8230 // (X&4) != 2 --> true
8231 Constant *Res = ConstantInt::get(Type::Int1Ty, isNE);
8232 Res = ConstantExpr::getZExt(Res, CI.getType());
8233 return ReplaceInstUsesWith(CI, Res);
8236 uint32_t ShiftAmt = KnownZeroMask.logBase2();
8237 Value *In = ICI->getOperand(0);
8239 // Perform a logical shr by shiftamt.
8240 // Insert the shift to put the result in the low bit.
8241 In = InsertNewInstBefore(BinaryOperator::CreateLShr(In,
8242 ConstantInt::get(In->getType(), ShiftAmt),
8243 In->getName()+".lobit"), CI);
8246 if ((Op1CV != 0) == isNE) { // Toggle the low bit.
8247 Constant *One = ConstantInt::get(In->getType(), 1);
8248 In = BinaryOperator::CreateXor(In, One, "tmp");
8249 InsertNewInstBefore(cast<Instruction>(In), CI);
8252 if (CI.getType() == In->getType())
8253 return ReplaceInstUsesWith(CI, In);
8255 return CastInst::CreateIntegerCast(In, CI.getType(), false/*ZExt*/);
8263 Instruction *InstCombiner::visitZExt(ZExtInst &CI) {
8264 // If one of the common conversion will work ..
8265 if (Instruction *Result = commonIntCastTransforms(CI))
8268 Value *Src = CI.getOperand(0);
8270 // If this is a cast of a cast
8271 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
8272 // If this is a TRUNC followed by a ZEXT then we are dealing with integral
8273 // types and if the sizes are just right we can convert this into a logical
8274 // 'and' which will be much cheaper than the pair of casts.
8275 if (isa<TruncInst>(CSrc)) {
8276 // Get the sizes of the types involved
8277 Value *A = CSrc->getOperand(0);
8278 uint32_t SrcSize = A->getType()->getPrimitiveSizeInBits();
8279 uint32_t MidSize = CSrc->getType()->getPrimitiveSizeInBits();
8280 uint32_t DstSize = CI.getType()->getPrimitiveSizeInBits();
8281 // If we're actually extending zero bits and the trunc is a no-op
8282 if (MidSize < DstSize && SrcSize == DstSize) {
8283 // Replace both of the casts with an And of the type mask.
8284 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
8285 Constant *AndConst = ConstantInt::get(AndValue);
8287 BinaryOperator::CreateAnd(CSrc->getOperand(0), AndConst);
8288 // Unfortunately, if the type changed, we need to cast it back.
8289 if (And->getType() != CI.getType()) {
8290 And->setName(CSrc->getName()+".mask");
8291 InsertNewInstBefore(And, CI);
8292 And = CastInst::CreateIntegerCast(And, CI.getType(), false/*ZExt*/);
8299 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src))
8300 return transformZExtICmp(ICI, CI);
8302 BinaryOperator *SrcI = dyn_cast<BinaryOperator>(Src);
8303 if (SrcI && SrcI->getOpcode() == Instruction::Or) {
8304 // zext (or icmp, icmp) --> or (zext icmp), (zext icmp) if at least one
8305 // of the (zext icmp) will be transformed.
8306 ICmpInst *LHS = dyn_cast<ICmpInst>(SrcI->getOperand(0));
8307 ICmpInst *RHS = dyn_cast<ICmpInst>(SrcI->getOperand(1));
8308 if (LHS && RHS && LHS->hasOneUse() && RHS->hasOneUse() &&
8309 (transformZExtICmp(LHS, CI, false) ||
8310 transformZExtICmp(RHS, CI, false))) {
8311 Value *LCast = InsertCastBefore(Instruction::ZExt, LHS, CI.getType(), CI);
8312 Value *RCast = InsertCastBefore(Instruction::ZExt, RHS, CI.getType(), CI);
8313 return BinaryOperator::Create(Instruction::Or, LCast, RCast);
8320 Instruction *InstCombiner::visitSExt(SExtInst &CI) {
8321 if (Instruction *I = commonIntCastTransforms(CI))
8324 Value *Src = CI.getOperand(0);
8326 // Canonicalize sign-extend from i1 to a select.
8327 if (Src->getType() == Type::Int1Ty)
8328 return SelectInst::Create(Src,
8329 ConstantInt::getAllOnesValue(CI.getType()),
8330 Constant::getNullValue(CI.getType()));
8332 // See if the value being truncated is already sign extended. If so, just
8333 // eliminate the trunc/sext pair.
8334 if (getOpcode(Src) == Instruction::Trunc) {
8335 Value *Op = cast<User>(Src)->getOperand(0);
8336 unsigned OpBits = cast<IntegerType>(Op->getType())->getBitWidth();
8337 unsigned MidBits = cast<IntegerType>(Src->getType())->getBitWidth();
8338 unsigned DestBits = cast<IntegerType>(CI.getType())->getBitWidth();
8339 unsigned NumSignBits = ComputeNumSignBits(Op);
8341 if (OpBits == DestBits) {
8342 // Op is i32, Mid is i8, and Dest is i32. If Op has more than 24 sign
8343 // bits, it is already ready.
8344 if (NumSignBits > DestBits-MidBits)
8345 return ReplaceInstUsesWith(CI, Op);
8346 } else if (OpBits < DestBits) {
8347 // Op is i32, Mid is i8, and Dest is i64. If Op has more than 24 sign
8348 // bits, just sext from i32.
8349 if (NumSignBits > OpBits-MidBits)
8350 return new SExtInst(Op, CI.getType(), "tmp");
8352 // Op is i64, Mid is i8, and Dest is i32. If Op has more than 56 sign
8353 // bits, just truncate to i32.
8354 if (NumSignBits > OpBits-MidBits)
8355 return new TruncInst(Op, CI.getType(), "tmp");
8359 // If the input is a shl/ashr pair of a same constant, then this is a sign
8360 // extension from a smaller value. If we could trust arbitrary bitwidth
8361 // integers, we could turn this into a truncate to the smaller bit and then
8362 // use a sext for the whole extension. Since we don't, look deeper and check
8363 // for a truncate. If the source and dest are the same type, eliminate the
8364 // trunc and extend and just do shifts. For example, turn:
8365 // %a = trunc i32 %i to i8
8366 // %b = shl i8 %a, 6
8367 // %c = ashr i8 %b, 6
8368 // %d = sext i8 %c to i32
8370 // %a = shl i32 %i, 30
8371 // %d = ashr i32 %a, 30
8373 ConstantInt *BA = 0, *CA = 0;
8374 if (match(Src, m_AShr(m_Shl(m_Value(A), m_ConstantInt(BA)),
8375 m_ConstantInt(CA))) &&
8376 BA == CA && isa<TruncInst>(A)) {
8377 Value *I = cast<TruncInst>(A)->getOperand(0);
8378 if (I->getType() == CI.getType()) {
8379 unsigned MidSize = Src->getType()->getPrimitiveSizeInBits();
8380 unsigned SrcDstSize = CI.getType()->getPrimitiveSizeInBits();
8381 unsigned ShAmt = CA->getZExtValue()+SrcDstSize-MidSize;
8382 Constant *ShAmtV = ConstantInt::get(CI.getType(), ShAmt);
8383 I = InsertNewInstBefore(BinaryOperator::CreateShl(I, ShAmtV,
8385 return BinaryOperator::CreateAShr(I, ShAmtV);
8392 /// FitsInFPType - Return a Constant* for the specified FP constant if it fits
8393 /// in the specified FP type without changing its value.
8394 static Constant *FitsInFPType(ConstantFP *CFP, const fltSemantics &Sem) {
8396 APFloat F = CFP->getValueAPF();
8397 (void)F.convert(Sem, APFloat::rmNearestTiesToEven, &losesInfo);
8399 return ConstantFP::get(F);
8403 /// LookThroughFPExtensions - If this is an fp extension instruction, look
8404 /// through it until we get the source value.
8405 static Value *LookThroughFPExtensions(Value *V) {
8406 if (Instruction *I = dyn_cast<Instruction>(V))
8407 if (I->getOpcode() == Instruction::FPExt)
8408 return LookThroughFPExtensions(I->getOperand(0));
8410 // If this value is a constant, return the constant in the smallest FP type
8411 // that can accurately represent it. This allows us to turn
8412 // (float)((double)X+2.0) into x+2.0f.
8413 if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
8414 if (CFP->getType() == Type::PPC_FP128Ty)
8415 return V; // No constant folding of this.
8416 // See if the value can be truncated to float and then reextended.
8417 if (Value *V = FitsInFPType(CFP, APFloat::IEEEsingle))
8419 if (CFP->getType() == Type::DoubleTy)
8420 return V; // Won't shrink.
8421 if (Value *V = FitsInFPType(CFP, APFloat::IEEEdouble))
8423 // Don't try to shrink to various long double types.
8429 Instruction *InstCombiner::visitFPTrunc(FPTruncInst &CI) {
8430 if (Instruction *I = commonCastTransforms(CI))
8433 // If we have fptrunc(add (fpextend x), (fpextend y)), where x and y are
8434 // smaller than the destination type, we can eliminate the truncate by doing
8435 // the add as the smaller type. This applies to add/sub/mul/div as well as
8436 // many builtins (sqrt, etc).
8437 BinaryOperator *OpI = dyn_cast<BinaryOperator>(CI.getOperand(0));
8438 if (OpI && OpI->hasOneUse()) {
8439 switch (OpI->getOpcode()) {
8441 case Instruction::Add:
8442 case Instruction::Sub:
8443 case Instruction::Mul:
8444 case Instruction::FDiv:
8445 case Instruction::FRem:
8446 const Type *SrcTy = OpI->getType();
8447 Value *LHSTrunc = LookThroughFPExtensions(OpI->getOperand(0));
8448 Value *RHSTrunc = LookThroughFPExtensions(OpI->getOperand(1));
8449 if (LHSTrunc->getType() != SrcTy &&
8450 RHSTrunc->getType() != SrcTy) {
8451 unsigned DstSize = CI.getType()->getPrimitiveSizeInBits();
8452 // If the source types were both smaller than the destination type of
8453 // the cast, do this xform.
8454 if (LHSTrunc->getType()->getPrimitiveSizeInBits() <= DstSize &&
8455 RHSTrunc->getType()->getPrimitiveSizeInBits() <= DstSize) {
8456 LHSTrunc = InsertCastBefore(Instruction::FPExt, LHSTrunc,
8458 RHSTrunc = InsertCastBefore(Instruction::FPExt, RHSTrunc,
8460 return BinaryOperator::Create(OpI->getOpcode(), LHSTrunc, RHSTrunc);
8469 Instruction *InstCombiner::visitFPExt(CastInst &CI) {
8470 return commonCastTransforms(CI);
8473 Instruction *InstCombiner::visitFPToUI(FPToUIInst &FI) {
8474 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0));
8476 return commonCastTransforms(FI);
8478 // fptoui(uitofp(X)) --> X
8479 // fptoui(sitofp(X)) --> X
8480 // This is safe if the intermediate type has enough bits in its mantissa to
8481 // accurately represent all values of X. For example, do not do this with
8482 // i64->float->i64. This is also safe for sitofp case, because any negative
8483 // 'X' value would cause an undefined result for the fptoui.
8484 if ((isa<UIToFPInst>(OpI) || isa<SIToFPInst>(OpI)) &&
8485 OpI->getOperand(0)->getType() == FI.getType() &&
8486 (int)FI.getType()->getPrimitiveSizeInBits() < /*extra bit for sign */
8487 OpI->getType()->getFPMantissaWidth())
8488 return ReplaceInstUsesWith(FI, OpI->getOperand(0));
8490 return commonCastTransforms(FI);
8493 Instruction *InstCombiner::visitFPToSI(FPToSIInst &FI) {
8494 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0));
8496 return commonCastTransforms(FI);
8498 // fptosi(sitofp(X)) --> X
8499 // fptosi(uitofp(X)) --> X
8500 // This is safe if the intermediate type has enough bits in its mantissa to
8501 // accurately represent all values of X. For example, do not do this with
8502 // i64->float->i64. This is also safe for sitofp case, because any negative
8503 // 'X' value would cause an undefined result for the fptoui.
8504 if ((isa<UIToFPInst>(OpI) || isa<SIToFPInst>(OpI)) &&
8505 OpI->getOperand(0)->getType() == FI.getType() &&
8506 (int)FI.getType()->getPrimitiveSizeInBits() <=
8507 OpI->getType()->getFPMantissaWidth())
8508 return ReplaceInstUsesWith(FI, OpI->getOperand(0));
8510 return commonCastTransforms(FI);
8513 Instruction *InstCombiner::visitUIToFP(CastInst &CI) {
8514 return commonCastTransforms(CI);
8517 Instruction *InstCombiner::visitSIToFP(CastInst &CI) {
8518 return commonCastTransforms(CI);
8521 Instruction *InstCombiner::visitPtrToInt(CastInst &CI) {
8522 return commonPointerCastTransforms(CI);
8525 Instruction *InstCombiner::visitIntToPtr(IntToPtrInst &CI) {
8526 if (Instruction *I = commonCastTransforms(CI))
8529 const Type *DestPointee = cast<PointerType>(CI.getType())->getElementType();
8530 if (!DestPointee->isSized()) return 0;
8532 // If this is inttoptr(add (ptrtoint x), cst), try to turn this into a GEP.
8535 if (match(CI.getOperand(0), m_Add(m_Cast<PtrToIntInst>(m_Value(X)),
8536 m_ConstantInt(Cst)))) {
8537 // If the source and destination operands have the same type, see if this
8538 // is a single-index GEP.
8539 if (X->getType() == CI.getType()) {
8540 // Get the size of the pointee type.
8541 uint64_t Size = TD->getTypePaddedSize(DestPointee);
8543 // Convert the constant to intptr type.
8544 APInt Offset = Cst->getValue();
8545 Offset.sextOrTrunc(TD->getPointerSizeInBits());
8547 // If Offset is evenly divisible by Size, we can do this xform.
8548 if (Size && !APIntOps::srem(Offset, APInt(Offset.getBitWidth(), Size))){
8549 Offset = APIntOps::sdiv(Offset, APInt(Offset.getBitWidth(), Size));
8550 return GetElementPtrInst::Create(X, ConstantInt::get(Offset));
8553 // TODO: Could handle other cases, e.g. where add is indexing into field of
8555 } else if (CI.getOperand(0)->hasOneUse() &&
8556 match(CI.getOperand(0), m_Add(m_Value(X), m_ConstantInt(Cst)))) {
8557 // Otherwise, if this is inttoptr(add x, cst), try to turn this into an
8558 // "inttoptr+GEP" instead of "add+intptr".
8560 // Get the size of the pointee type.
8561 uint64_t Size = TD->getTypePaddedSize(DestPointee);
8563 // Convert the constant to intptr type.
8564 APInt Offset = Cst->getValue();
8565 Offset.sextOrTrunc(TD->getPointerSizeInBits());
8567 // If Offset is evenly divisible by Size, we can do this xform.
8568 if (Size && !APIntOps::srem(Offset, APInt(Offset.getBitWidth(), Size))){
8569 Offset = APIntOps::sdiv(Offset, APInt(Offset.getBitWidth(), Size));
8571 Instruction *P = InsertNewInstBefore(new IntToPtrInst(X, CI.getType(),
8573 return GetElementPtrInst::Create(P, ConstantInt::get(Offset), "tmp");
8579 Instruction *InstCombiner::visitBitCast(BitCastInst &CI) {
8580 // If the operands are integer typed then apply the integer transforms,
8581 // otherwise just apply the common ones.
8582 Value *Src = CI.getOperand(0);
8583 const Type *SrcTy = Src->getType();
8584 const Type *DestTy = CI.getType();
8586 if (SrcTy->isInteger() && DestTy->isInteger()) {
8587 if (Instruction *Result = commonIntCastTransforms(CI))
8589 } else if (isa<PointerType>(SrcTy)) {
8590 if (Instruction *I = commonPointerCastTransforms(CI))
8593 if (Instruction *Result = commonCastTransforms(CI))
8598 // Get rid of casts from one type to the same type. These are useless and can
8599 // be replaced by the operand.
8600 if (DestTy == Src->getType())
8601 return ReplaceInstUsesWith(CI, Src);
8603 if (const PointerType *DstPTy = dyn_cast<PointerType>(DestTy)) {
8604 const PointerType *SrcPTy = cast<PointerType>(SrcTy);
8605 const Type *DstElTy = DstPTy->getElementType();
8606 const Type *SrcElTy = SrcPTy->getElementType();
8608 // If the address spaces don't match, don't eliminate the bitcast, which is
8609 // required for changing types.
8610 if (SrcPTy->getAddressSpace() != DstPTy->getAddressSpace())
8613 // If we are casting a malloc or alloca to a pointer to a type of the same
8614 // size, rewrite the allocation instruction to allocate the "right" type.
8615 if (AllocationInst *AI = dyn_cast<AllocationInst>(Src))
8616 if (Instruction *V = PromoteCastOfAllocation(CI, *AI))
8619 // If the source and destination are pointers, and this cast is equivalent
8620 // to a getelementptr X, 0, 0, 0... turn it into the appropriate gep.
8621 // This can enhance SROA and other transforms that want type-safe pointers.
8622 Constant *ZeroUInt = Constant::getNullValue(Type::Int32Ty);
8623 unsigned NumZeros = 0;
8624 while (SrcElTy != DstElTy &&
8625 isa<CompositeType>(SrcElTy) && !isa<PointerType>(SrcElTy) &&
8626 SrcElTy->getNumContainedTypes() /* not "{}" */) {
8627 SrcElTy = cast<CompositeType>(SrcElTy)->getTypeAtIndex(ZeroUInt);
8631 // If we found a path from the src to dest, create the getelementptr now.
8632 if (SrcElTy == DstElTy) {
8633 SmallVector<Value*, 8> Idxs(NumZeros+1, ZeroUInt);
8634 return GetElementPtrInst::Create(Src, Idxs.begin(), Idxs.end(), "",
8635 ((Instruction*) NULL));
8639 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(Src)) {
8640 if (SVI->hasOneUse()) {
8641 // Okay, we have (bitconvert (shuffle ..)). Check to see if this is
8642 // a bitconvert to a vector with the same # elts.
8643 if (isa<VectorType>(DestTy) &&
8644 cast<VectorType>(DestTy)->getNumElements() ==
8645 SVI->getType()->getNumElements() &&
8646 SVI->getType()->getNumElements() ==
8647 cast<VectorType>(SVI->getOperand(0)->getType())->getNumElements()) {
8649 // If either of the operands is a cast from CI.getType(), then
8650 // evaluating the shuffle in the casted destination's type will allow
8651 // us to eliminate at least one cast.
8652 if (((Tmp = dyn_cast<CastInst>(SVI->getOperand(0))) &&
8653 Tmp->getOperand(0)->getType() == DestTy) ||
8654 ((Tmp = dyn_cast<CastInst>(SVI->getOperand(1))) &&
8655 Tmp->getOperand(0)->getType() == DestTy)) {
8656 Value *LHS = InsertCastBefore(Instruction::BitCast,
8657 SVI->getOperand(0), DestTy, CI);
8658 Value *RHS = InsertCastBefore(Instruction::BitCast,
8659 SVI->getOperand(1), DestTy, CI);
8660 // Return a new shuffle vector. Use the same element ID's, as we
8661 // know the vector types match #elts.
8662 return new ShuffleVectorInst(LHS, RHS, SVI->getOperand(2));
8670 /// GetSelectFoldableOperands - We want to turn code that looks like this:
8672 /// %D = select %cond, %C, %A
8674 /// %C = select %cond, %B, 0
8677 /// Assuming that the specified instruction is an operand to the select, return
8678 /// a bitmask indicating which operands of this instruction are foldable if they
8679 /// equal the other incoming value of the select.
8681 static unsigned GetSelectFoldableOperands(Instruction *I) {
8682 switch (I->getOpcode()) {
8683 case Instruction::Add:
8684 case Instruction::Mul:
8685 case Instruction::And:
8686 case Instruction::Or:
8687 case Instruction::Xor:
8688 return 3; // Can fold through either operand.
8689 case Instruction::Sub: // Can only fold on the amount subtracted.
8690 case Instruction::Shl: // Can only fold on the shift amount.
8691 case Instruction::LShr:
8692 case Instruction::AShr:
8695 return 0; // Cannot fold
8699 /// GetSelectFoldableConstant - For the same transformation as the previous
8700 /// function, return the identity constant that goes into the select.
8701 static Constant *GetSelectFoldableConstant(Instruction *I) {
8702 switch (I->getOpcode()) {
8703 default: assert(0 && "This cannot happen!"); abort();
8704 case Instruction::Add:
8705 case Instruction::Sub:
8706 case Instruction::Or:
8707 case Instruction::Xor:
8708 case Instruction::Shl:
8709 case Instruction::LShr:
8710 case Instruction::AShr:
8711 return Constant::getNullValue(I->getType());
8712 case Instruction::And:
8713 return Constant::getAllOnesValue(I->getType());
8714 case Instruction::Mul:
8715 return ConstantInt::get(I->getType(), 1);
8719 /// FoldSelectOpOp - Here we have (select c, TI, FI), and we know that TI and FI
8720 /// have the same opcode and only one use each. Try to simplify this.
8721 Instruction *InstCombiner::FoldSelectOpOp(SelectInst &SI, Instruction *TI,
8723 if (TI->getNumOperands() == 1) {
8724 // If this is a non-volatile load or a cast from the same type,
8727 if (TI->getOperand(0)->getType() != FI->getOperand(0)->getType())
8730 return 0; // unknown unary op.
8733 // Fold this by inserting a select from the input values.
8734 SelectInst *NewSI = SelectInst::Create(SI.getCondition(), TI->getOperand(0),
8735 FI->getOperand(0), SI.getName()+".v");
8736 InsertNewInstBefore(NewSI, SI);
8737 return CastInst::Create(Instruction::CastOps(TI->getOpcode()), NewSI,
8741 // Only handle binary operators here.
8742 if (!isa<BinaryOperator>(TI))
8745 // Figure out if the operations have any operands in common.
8746 Value *MatchOp, *OtherOpT, *OtherOpF;
8748 if (TI->getOperand(0) == FI->getOperand(0)) {
8749 MatchOp = TI->getOperand(0);
8750 OtherOpT = TI->getOperand(1);
8751 OtherOpF = FI->getOperand(1);
8752 MatchIsOpZero = true;
8753 } else if (TI->getOperand(1) == FI->getOperand(1)) {
8754 MatchOp = TI->getOperand(1);
8755 OtherOpT = TI->getOperand(0);
8756 OtherOpF = FI->getOperand(0);
8757 MatchIsOpZero = false;
8758 } else if (!TI->isCommutative()) {
8760 } else if (TI->getOperand(0) == FI->getOperand(1)) {
8761 MatchOp = TI->getOperand(0);
8762 OtherOpT = TI->getOperand(1);
8763 OtherOpF = FI->getOperand(0);
8764 MatchIsOpZero = true;
8765 } else if (TI->getOperand(1) == FI->getOperand(0)) {
8766 MatchOp = TI->getOperand(1);
8767 OtherOpT = TI->getOperand(0);
8768 OtherOpF = FI->getOperand(1);
8769 MatchIsOpZero = true;
8774 // If we reach here, they do have operations in common.
8775 SelectInst *NewSI = SelectInst::Create(SI.getCondition(), OtherOpT,
8776 OtherOpF, SI.getName()+".v");
8777 InsertNewInstBefore(NewSI, SI);
8779 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TI)) {
8781 return BinaryOperator::Create(BO->getOpcode(), MatchOp, NewSI);
8783 return BinaryOperator::Create(BO->getOpcode(), NewSI, MatchOp);
8785 assert(0 && "Shouldn't get here");
8789 /// visitSelectInstWithICmp - Visit a SelectInst that has an
8790 /// ICmpInst as its first operand.
8792 Instruction *InstCombiner::visitSelectInstWithICmp(SelectInst &SI,
8794 bool Changed = false;
8795 ICmpInst::Predicate Pred = ICI->getPredicate();
8796 Value *CmpLHS = ICI->getOperand(0);
8797 Value *CmpRHS = ICI->getOperand(1);
8798 Value *TrueVal = SI.getTrueValue();
8799 Value *FalseVal = SI.getFalseValue();
8801 // Check cases where the comparison is with a constant that
8802 // can be adjusted to fit the min/max idiom. We may edit ICI in
8803 // place here, so make sure the select is the only user.
8804 if (ICI->hasOneUse())
8805 if (ConstantInt *CI = dyn_cast<ConstantInt>(CmpRHS)) {
8808 case ICmpInst::ICMP_ULT:
8809 case ICmpInst::ICMP_SLT: {
8810 // X < MIN ? T : F --> F
8811 if (CI->isMinValue(Pred == ICmpInst::ICMP_SLT))
8812 return ReplaceInstUsesWith(SI, FalseVal);
8813 // X < C ? X : C-1 --> X > C-1 ? C-1 : X
8814 Constant *AdjustedRHS = SubOne(CI);
8815 if ((CmpLHS == TrueVal && AdjustedRHS == FalseVal) ||
8816 (CmpLHS == FalseVal && AdjustedRHS == TrueVal)) {
8817 Pred = ICmpInst::getSwappedPredicate(Pred);
8818 CmpRHS = AdjustedRHS;
8819 std::swap(FalseVal, TrueVal);
8820 ICI->setPredicate(Pred);
8821 ICI->setOperand(1, CmpRHS);
8822 SI.setOperand(1, TrueVal);
8823 SI.setOperand(2, FalseVal);
8828 case ICmpInst::ICMP_UGT:
8829 case ICmpInst::ICMP_SGT: {
8830 // X > MAX ? T : F --> F
8831 if (CI->isMaxValue(Pred == ICmpInst::ICMP_SGT))
8832 return ReplaceInstUsesWith(SI, FalseVal);
8833 // X > C ? X : C+1 --> X < C+1 ? C+1 : X
8834 Constant *AdjustedRHS = AddOne(CI);
8835 if ((CmpLHS == TrueVal && AdjustedRHS == FalseVal) ||
8836 (CmpLHS == FalseVal && AdjustedRHS == TrueVal)) {
8837 Pred = ICmpInst::getSwappedPredicate(Pred);
8838 CmpRHS = AdjustedRHS;
8839 std::swap(FalseVal, TrueVal);
8840 ICI->setPredicate(Pred);
8841 ICI->setOperand(1, CmpRHS);
8842 SI.setOperand(1, TrueVal);
8843 SI.setOperand(2, FalseVal);
8850 // (x <s 0) ? -1 : 0 -> ashr x, 31 -> all ones if signed
8851 // (x >s -1) ? -1 : 0 -> ashr x, 31 -> all ones if not signed
8852 CmpInst::Predicate Pred = CmpInst::BAD_ICMP_PREDICATE;
8853 if (match(TrueVal, m_ConstantInt<-1>()) &&
8854 match(FalseVal, m_ConstantInt<0>()))
8855 Pred = ICI->getPredicate();
8856 else if (match(TrueVal, m_ConstantInt<0>()) &&
8857 match(FalseVal, m_ConstantInt<-1>()))
8858 Pred = CmpInst::getInversePredicate(ICI->getPredicate());
8860 if (Pred != CmpInst::BAD_ICMP_PREDICATE) {
8861 // If we are just checking for a icmp eq of a single bit and zext'ing it
8862 // to an integer, then shift the bit to the appropriate place and then
8863 // cast to integer to avoid the comparison.
8864 const APInt &Op1CV = CI->getValue();
8866 // sext (x <s 0) to i32 --> x>>s31 true if signbit set.
8867 // sext (x >s -1) to i32 --> (x>>s31)^-1 true if signbit clear.
8868 if ((Pred == ICmpInst::ICMP_SLT && Op1CV == 0) ||
8869 (Pred == ICmpInst::ICMP_SGT && Op1CV.isAllOnesValue())) {
8870 Value *In = ICI->getOperand(0);
8871 Value *Sh = ConstantInt::get(In->getType(),
8872 In->getType()->getPrimitiveSizeInBits()-1);
8873 In = InsertNewInstBefore(BinaryOperator::CreateAShr(In, Sh,
8874 In->getName()+".lobit"),
8876 if (In->getType() != SI.getType())
8877 In = CastInst::CreateIntegerCast(In, SI.getType(),
8878 true/*SExt*/, "tmp", ICI);
8880 if (Pred == ICmpInst::ICMP_SGT)
8881 In = InsertNewInstBefore(BinaryOperator::CreateNot(In,
8882 In->getName()+".not"), *ICI);
8884 return ReplaceInstUsesWith(SI, In);
8889 if (CmpLHS == TrueVal && CmpRHS == FalseVal) {
8890 // Transform (X == Y) ? X : Y -> Y
8891 if (Pred == ICmpInst::ICMP_EQ)
8892 return ReplaceInstUsesWith(SI, FalseVal);
8893 // Transform (X != Y) ? X : Y -> X
8894 if (Pred == ICmpInst::ICMP_NE)
8895 return ReplaceInstUsesWith(SI, TrueVal);
8896 /// NOTE: if we wanted to, this is where to detect integer MIN/MAX
8898 } else if (CmpLHS == FalseVal && CmpRHS == TrueVal) {
8899 // Transform (X == Y) ? Y : X -> X
8900 if (Pred == ICmpInst::ICMP_EQ)
8901 return ReplaceInstUsesWith(SI, FalseVal);
8902 // Transform (X != Y) ? Y : X -> Y
8903 if (Pred == ICmpInst::ICMP_NE)
8904 return ReplaceInstUsesWith(SI, TrueVal);
8905 /// NOTE: if we wanted to, this is where to detect integer MIN/MAX
8908 /// NOTE: if we wanted to, this is where to detect integer ABS
8910 return Changed ? &SI : 0;
8913 Instruction *InstCombiner::visitSelectInst(SelectInst &SI) {
8914 Value *CondVal = SI.getCondition();
8915 Value *TrueVal = SI.getTrueValue();
8916 Value *FalseVal = SI.getFalseValue();
8918 // select true, X, Y -> X
8919 // select false, X, Y -> Y
8920 if (ConstantInt *C = dyn_cast<ConstantInt>(CondVal))
8921 return ReplaceInstUsesWith(SI, C->getZExtValue() ? TrueVal : FalseVal);
8923 // select C, X, X -> X
8924 if (TrueVal == FalseVal)
8925 return ReplaceInstUsesWith(SI, TrueVal);
8927 if (isa<UndefValue>(TrueVal)) // select C, undef, X -> X
8928 return ReplaceInstUsesWith(SI, FalseVal);
8929 if (isa<UndefValue>(FalseVal)) // select C, X, undef -> X
8930 return ReplaceInstUsesWith(SI, TrueVal);
8931 if (isa<UndefValue>(CondVal)) { // select undef, X, Y -> X or Y
8932 if (isa<Constant>(TrueVal))
8933 return ReplaceInstUsesWith(SI, TrueVal);
8935 return ReplaceInstUsesWith(SI, FalseVal);
8938 if (SI.getType() == Type::Int1Ty) {
8939 if (ConstantInt *C = dyn_cast<ConstantInt>(TrueVal)) {
8940 if (C->getZExtValue()) {
8941 // Change: A = select B, true, C --> A = or B, C
8942 return BinaryOperator::CreateOr(CondVal, FalseVal);
8944 // Change: A = select B, false, C --> A = and !B, C
8946 InsertNewInstBefore(BinaryOperator::CreateNot(CondVal,
8947 "not."+CondVal->getName()), SI);
8948 return BinaryOperator::CreateAnd(NotCond, FalseVal);
8950 } else if (ConstantInt *C = dyn_cast<ConstantInt>(FalseVal)) {
8951 if (C->getZExtValue() == false) {
8952 // Change: A = select B, C, false --> A = and B, C
8953 return BinaryOperator::CreateAnd(CondVal, TrueVal);
8955 // Change: A = select B, C, true --> A = or !B, C
8957 InsertNewInstBefore(BinaryOperator::CreateNot(CondVal,
8958 "not."+CondVal->getName()), SI);
8959 return BinaryOperator::CreateOr(NotCond, TrueVal);
8963 // select a, b, a -> a&b
8964 // select a, a, b -> a|b
8965 if (CondVal == TrueVal)
8966 return BinaryOperator::CreateOr(CondVal, FalseVal);
8967 else if (CondVal == FalseVal)
8968 return BinaryOperator::CreateAnd(CondVal, TrueVal);
8971 // Selecting between two integer constants?
8972 if (ConstantInt *TrueValC = dyn_cast<ConstantInt>(TrueVal))
8973 if (ConstantInt *FalseValC = dyn_cast<ConstantInt>(FalseVal)) {
8974 // select C, 1, 0 -> zext C to int
8975 if (FalseValC->isZero() && TrueValC->getValue() == 1) {
8976 return CastInst::Create(Instruction::ZExt, CondVal, SI.getType());
8977 } else if (TrueValC->isZero() && FalseValC->getValue() == 1) {
8978 // select C, 0, 1 -> zext !C to int
8980 InsertNewInstBefore(BinaryOperator::CreateNot(CondVal,
8981 "not."+CondVal->getName()), SI);
8982 return CastInst::Create(Instruction::ZExt, NotCond, SI.getType());
8985 if (ICmpInst *IC = dyn_cast<ICmpInst>(SI.getCondition())) {
8987 // (x <s 0) ? -1 : 0 -> ashr x, 31
8988 if (TrueValC->isAllOnesValue() && FalseValC->isZero())
8989 if (ConstantInt *CmpCst = dyn_cast<ConstantInt>(IC->getOperand(1))) {
8990 if (IC->getPredicate() == ICmpInst::ICMP_SLT && CmpCst->isZero()) {
8991 // The comparison constant and the result are not neccessarily the
8992 // same width. Make an all-ones value by inserting a AShr.
8993 Value *X = IC->getOperand(0);
8994 uint32_t Bits = X->getType()->getPrimitiveSizeInBits();
8995 Constant *ShAmt = ConstantInt::get(X->getType(), Bits-1);
8996 Instruction *SRA = BinaryOperator::Create(Instruction::AShr, X,
8998 InsertNewInstBefore(SRA, SI);
9000 // Then cast to the appropriate width.
9001 return CastInst::CreateIntegerCast(SRA, SI.getType(), true);
9006 // If one of the constants is zero (we know they can't both be) and we
9007 // have an icmp instruction with zero, and we have an 'and' with the
9008 // non-constant value, eliminate this whole mess. This corresponds to
9009 // cases like this: ((X & 27) ? 27 : 0)
9010 if (TrueValC->isZero() || FalseValC->isZero())
9011 if (IC->isEquality() && isa<ConstantInt>(IC->getOperand(1)) &&
9012 cast<Constant>(IC->getOperand(1))->isNullValue())
9013 if (Instruction *ICA = dyn_cast<Instruction>(IC->getOperand(0)))
9014 if (ICA->getOpcode() == Instruction::And &&
9015 isa<ConstantInt>(ICA->getOperand(1)) &&
9016 (ICA->getOperand(1) == TrueValC ||
9017 ICA->getOperand(1) == FalseValC) &&
9018 isOneBitSet(cast<ConstantInt>(ICA->getOperand(1)))) {
9019 // Okay, now we know that everything is set up, we just don't
9020 // know whether we have a icmp_ne or icmp_eq and whether the
9021 // true or false val is the zero.
9022 bool ShouldNotVal = !TrueValC->isZero();
9023 ShouldNotVal ^= IC->getPredicate() == ICmpInst::ICMP_NE;
9026 V = InsertNewInstBefore(BinaryOperator::Create(
9027 Instruction::Xor, V, ICA->getOperand(1)), SI);
9028 return ReplaceInstUsesWith(SI, V);
9033 // See if we are selecting two values based on a comparison of the two values.
9034 if (FCmpInst *FCI = dyn_cast<FCmpInst>(CondVal)) {
9035 if (FCI->getOperand(0) == TrueVal && FCI->getOperand(1) == FalseVal) {
9036 // Transform (X == Y) ? X : Y -> Y
9037 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ) {
9038 // This is not safe in general for floating point:
9039 // consider X== -0, Y== +0.
9040 // It becomes safe if either operand is a nonzero constant.
9041 ConstantFP *CFPt, *CFPf;
9042 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) &&
9043 !CFPt->getValueAPF().isZero()) ||
9044 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) &&
9045 !CFPf->getValueAPF().isZero()))
9046 return ReplaceInstUsesWith(SI, FalseVal);
9048 // Transform (X != Y) ? X : Y -> X
9049 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
9050 return ReplaceInstUsesWith(SI, TrueVal);
9051 // NOTE: if we wanted to, this is where to detect MIN/MAX
9053 } else if (FCI->getOperand(0) == FalseVal && FCI->getOperand(1) == TrueVal){
9054 // Transform (X == Y) ? Y : X -> X
9055 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ) {
9056 // This is not safe in general for floating point:
9057 // consider X== -0, Y== +0.
9058 // It becomes safe if either operand is a nonzero constant.
9059 ConstantFP *CFPt, *CFPf;
9060 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) &&
9061 !CFPt->getValueAPF().isZero()) ||
9062 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) &&
9063 !CFPf->getValueAPF().isZero()))
9064 return ReplaceInstUsesWith(SI, FalseVal);
9066 // Transform (X != Y) ? Y : X -> Y
9067 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
9068 return ReplaceInstUsesWith(SI, TrueVal);
9069 // NOTE: if we wanted to, this is where to detect MIN/MAX
9071 // NOTE: if we wanted to, this is where to detect ABS
9074 // See if we are selecting two values based on a comparison of the two values.
9075 if (ICmpInst *ICI = dyn_cast<ICmpInst>(CondVal))
9076 if (Instruction *Result = visitSelectInstWithICmp(SI, ICI))
9079 if (Instruction *TI = dyn_cast<Instruction>(TrueVal))
9080 if (Instruction *FI = dyn_cast<Instruction>(FalseVal))
9081 if (TI->hasOneUse() && FI->hasOneUse()) {
9082 Instruction *AddOp = 0, *SubOp = 0;
9084 // Turn (select C, (op X, Y), (op X, Z)) -> (op X, (select C, Y, Z))
9085 if (TI->getOpcode() == FI->getOpcode())
9086 if (Instruction *IV = FoldSelectOpOp(SI, TI, FI))
9089 // Turn select C, (X+Y), (X-Y) --> (X+(select C, Y, (-Y))). This is
9090 // even legal for FP.
9091 if (TI->getOpcode() == Instruction::Sub &&
9092 FI->getOpcode() == Instruction::Add) {
9093 AddOp = FI; SubOp = TI;
9094 } else if (FI->getOpcode() == Instruction::Sub &&
9095 TI->getOpcode() == Instruction::Add) {
9096 AddOp = TI; SubOp = FI;
9100 Value *OtherAddOp = 0;
9101 if (SubOp->getOperand(0) == AddOp->getOperand(0)) {
9102 OtherAddOp = AddOp->getOperand(1);
9103 } else if (SubOp->getOperand(0) == AddOp->getOperand(1)) {
9104 OtherAddOp = AddOp->getOperand(0);
9108 // So at this point we know we have (Y -> OtherAddOp):
9109 // select C, (add X, Y), (sub X, Z)
9110 Value *NegVal; // Compute -Z
9111 if (Constant *C = dyn_cast<Constant>(SubOp->getOperand(1))) {
9112 NegVal = ConstantExpr::getNeg(C);
9114 NegVal = InsertNewInstBefore(
9115 BinaryOperator::CreateNeg(SubOp->getOperand(1), "tmp"), SI);
9118 Value *NewTrueOp = OtherAddOp;
9119 Value *NewFalseOp = NegVal;
9121 std::swap(NewTrueOp, NewFalseOp);
9122 Instruction *NewSel =
9123 SelectInst::Create(CondVal, NewTrueOp,
9124 NewFalseOp, SI.getName() + ".p");
9126 NewSel = InsertNewInstBefore(NewSel, SI);
9127 return BinaryOperator::CreateAdd(SubOp->getOperand(0), NewSel);
9132 // See if we can fold the select into one of our operands.
9133 if (SI.getType()->isInteger()) {
9134 // See the comment above GetSelectFoldableOperands for a description of the
9135 // transformation we are doing here.
9136 if (Instruction *TVI = dyn_cast<Instruction>(TrueVal))
9137 if (TVI->hasOneUse() && TVI->getNumOperands() == 2 &&
9138 !isa<Constant>(FalseVal))
9139 if (unsigned SFO = GetSelectFoldableOperands(TVI)) {
9140 unsigned OpToFold = 0;
9141 if ((SFO & 1) && FalseVal == TVI->getOperand(0)) {
9143 } else if ((SFO & 2) && FalseVal == TVI->getOperand(1)) {
9148 Constant *C = GetSelectFoldableConstant(TVI);
9149 Instruction *NewSel =
9150 SelectInst::Create(SI.getCondition(),
9151 TVI->getOperand(2-OpToFold), C);
9152 InsertNewInstBefore(NewSel, SI);
9153 NewSel->takeName(TVI);
9154 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TVI))
9155 return BinaryOperator::Create(BO->getOpcode(), FalseVal, NewSel);
9157 assert(0 && "Unknown instruction!!");
9162 if (Instruction *FVI = dyn_cast<Instruction>(FalseVal))
9163 if (FVI->hasOneUse() && FVI->getNumOperands() == 2 &&
9164 !isa<Constant>(TrueVal))
9165 if (unsigned SFO = GetSelectFoldableOperands(FVI)) {
9166 unsigned OpToFold = 0;
9167 if ((SFO & 1) && TrueVal == FVI->getOperand(0)) {
9169 } else if ((SFO & 2) && TrueVal == FVI->getOperand(1)) {
9174 Constant *C = GetSelectFoldableConstant(FVI);
9175 Instruction *NewSel =
9176 SelectInst::Create(SI.getCondition(), C,
9177 FVI->getOperand(2-OpToFold));
9178 InsertNewInstBefore(NewSel, SI);
9179 NewSel->takeName(FVI);
9180 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(FVI))
9181 return BinaryOperator::Create(BO->getOpcode(), TrueVal, NewSel);
9183 assert(0 && "Unknown instruction!!");
9188 if (BinaryOperator::isNot(CondVal)) {
9189 SI.setOperand(0, BinaryOperator::getNotArgument(CondVal));
9190 SI.setOperand(1, FalseVal);
9191 SI.setOperand(2, TrueVal);
9198 /// EnforceKnownAlignment - If the specified pointer points to an object that
9199 /// we control, modify the object's alignment to PrefAlign. This isn't
9200 /// often possible though. If alignment is important, a more reliable approach
9201 /// is to simply align all global variables and allocation instructions to
9202 /// their preferred alignment from the beginning.
9204 static unsigned EnforceKnownAlignment(Value *V,
9205 unsigned Align, unsigned PrefAlign) {
9207 User *U = dyn_cast<User>(V);
9208 if (!U) return Align;
9210 switch (getOpcode(U)) {
9212 case Instruction::BitCast:
9213 return EnforceKnownAlignment(U->getOperand(0), Align, PrefAlign);
9214 case Instruction::GetElementPtr: {
9215 // If all indexes are zero, it is just the alignment of the base pointer.
9216 bool AllZeroOperands = true;
9217 for (User::op_iterator i = U->op_begin() + 1, e = U->op_end(); i != e; ++i)
9218 if (!isa<Constant>(*i) ||
9219 !cast<Constant>(*i)->isNullValue()) {
9220 AllZeroOperands = false;
9224 if (AllZeroOperands) {
9225 // Treat this like a bitcast.
9226 return EnforceKnownAlignment(U->getOperand(0), Align, PrefAlign);
9232 if (GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
9233 // If there is a large requested alignment and we can, bump up the alignment
9235 if (!GV->isDeclaration()) {
9236 GV->setAlignment(PrefAlign);
9239 } else if (AllocationInst *AI = dyn_cast<AllocationInst>(V)) {
9240 // If there is a requested alignment and if this is an alloca, round up. We
9241 // don't do this for malloc, because some systems can't respect the request.
9242 if (isa<AllocaInst>(AI)) {
9243 AI->setAlignment(PrefAlign);
9251 /// GetOrEnforceKnownAlignment - If the specified pointer has an alignment that
9252 /// we can determine, return it, otherwise return 0. If PrefAlign is specified,
9253 /// and it is more than the alignment of the ultimate object, see if we can
9254 /// increase the alignment of the ultimate object, making this check succeed.
9255 unsigned InstCombiner::GetOrEnforceKnownAlignment(Value *V,
9256 unsigned PrefAlign) {
9257 unsigned BitWidth = TD ? TD->getTypeSizeInBits(V->getType()) :
9258 sizeof(PrefAlign) * CHAR_BIT;
9259 APInt Mask = APInt::getAllOnesValue(BitWidth);
9260 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
9261 ComputeMaskedBits(V, Mask, KnownZero, KnownOne);
9262 unsigned TrailZ = KnownZero.countTrailingOnes();
9263 unsigned Align = 1u << std::min(BitWidth - 1, TrailZ);
9265 if (PrefAlign > Align)
9266 Align = EnforceKnownAlignment(V, Align, PrefAlign);
9268 // We don't need to make any adjustment.
9272 Instruction *InstCombiner::SimplifyMemTransfer(MemIntrinsic *MI) {
9273 unsigned DstAlign = GetOrEnforceKnownAlignment(MI->getOperand(1));
9274 unsigned SrcAlign = GetOrEnforceKnownAlignment(MI->getOperand(2));
9275 unsigned MinAlign = std::min(DstAlign, SrcAlign);
9276 unsigned CopyAlign = MI->getAlignment()->getZExtValue();
9278 if (CopyAlign < MinAlign) {
9279 MI->setAlignment(ConstantInt::get(Type::Int32Ty, MinAlign));
9283 // If MemCpyInst length is 1/2/4/8 bytes then replace memcpy with
9285 ConstantInt *MemOpLength = dyn_cast<ConstantInt>(MI->getOperand(3));
9286 if (MemOpLength == 0) return 0;
9288 // Source and destination pointer types are always "i8*" for intrinsic. See
9289 // if the size is something we can handle with a single primitive load/store.
9290 // A single load+store correctly handles overlapping memory in the memmove
9292 unsigned Size = MemOpLength->getZExtValue();
9293 if (Size == 0) return MI; // Delete this mem transfer.
9295 if (Size > 8 || (Size&(Size-1)))
9296 return 0; // If not 1/2/4/8 bytes, exit.
9298 // Use an integer load+store unless we can find something better.
9299 Type *NewPtrTy = PointerType::getUnqual(IntegerType::get(Size<<3));
9301 // Memcpy forces the use of i8* for the source and destination. That means
9302 // that if you're using memcpy to move one double around, you'll get a cast
9303 // from double* to i8*. We'd much rather use a double load+store rather than
9304 // an i64 load+store, here because this improves the odds that the source or
9305 // dest address will be promotable. See if we can find a better type than the
9306 // integer datatype.
9307 if (Value *Op = getBitCastOperand(MI->getOperand(1))) {
9308 const Type *SrcETy = cast<PointerType>(Op->getType())->getElementType();
9309 if (SrcETy->isSized() && TD->getTypeStoreSize(SrcETy) == Size) {
9310 // The SrcETy might be something like {{{double}}} or [1 x double]. Rip
9311 // down through these levels if so.
9312 while (!SrcETy->isSingleValueType()) {
9313 if (const StructType *STy = dyn_cast<StructType>(SrcETy)) {
9314 if (STy->getNumElements() == 1)
9315 SrcETy = STy->getElementType(0);
9318 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(SrcETy)) {
9319 if (ATy->getNumElements() == 1)
9320 SrcETy = ATy->getElementType();
9327 if (SrcETy->isSingleValueType())
9328 NewPtrTy = PointerType::getUnqual(SrcETy);
9333 // If the memcpy/memmove provides better alignment info than we can
9335 SrcAlign = std::max(SrcAlign, CopyAlign);
9336 DstAlign = std::max(DstAlign, CopyAlign);
9338 Value *Src = InsertBitCastBefore(MI->getOperand(2), NewPtrTy, *MI);
9339 Value *Dest = InsertBitCastBefore(MI->getOperand(1), NewPtrTy, *MI);
9340 Instruction *L = new LoadInst(Src, "tmp", false, SrcAlign);
9341 InsertNewInstBefore(L, *MI);
9342 InsertNewInstBefore(new StoreInst(L, Dest, false, DstAlign), *MI);
9344 // Set the size of the copy to 0, it will be deleted on the next iteration.
9345 MI->setOperand(3, Constant::getNullValue(MemOpLength->getType()));
9349 Instruction *InstCombiner::SimplifyMemSet(MemSetInst *MI) {
9350 unsigned Alignment = GetOrEnforceKnownAlignment(MI->getDest());
9351 if (MI->getAlignment()->getZExtValue() < Alignment) {
9352 MI->setAlignment(ConstantInt::get(Type::Int32Ty, Alignment));
9356 // Extract the length and alignment and fill if they are constant.
9357 ConstantInt *LenC = dyn_cast<ConstantInt>(MI->getLength());
9358 ConstantInt *FillC = dyn_cast<ConstantInt>(MI->getValue());
9359 if (!LenC || !FillC || FillC->getType() != Type::Int8Ty)
9361 uint64_t Len = LenC->getZExtValue();
9362 Alignment = MI->getAlignment()->getZExtValue();
9364 // If the length is zero, this is a no-op
9365 if (Len == 0) return MI; // memset(d,c,0,a) -> noop
9367 // memset(s,c,n) -> store s, c (for n=1,2,4,8)
9368 if (Len <= 8 && isPowerOf2_32((uint32_t)Len)) {
9369 const Type *ITy = IntegerType::get(Len*8); // n=1 -> i8.
9371 Value *Dest = MI->getDest();
9372 Dest = InsertBitCastBefore(Dest, PointerType::getUnqual(ITy), *MI);
9374 // Alignment 0 is identity for alignment 1 for memset, but not store.
9375 if (Alignment == 0) Alignment = 1;
9377 // Extract the fill value and store.
9378 uint64_t Fill = FillC->getZExtValue()*0x0101010101010101ULL;
9379 InsertNewInstBefore(new StoreInst(ConstantInt::get(ITy, Fill), Dest, false,
9382 // Set the size of the copy to 0, it will be deleted on the next iteration.
9383 MI->setLength(Constant::getNullValue(LenC->getType()));
9391 /// visitCallInst - CallInst simplification. This mostly only handles folding
9392 /// of intrinsic instructions. For normal calls, it allows visitCallSite to do
9393 /// the heavy lifting.
9395 Instruction *InstCombiner::visitCallInst(CallInst &CI) {
9396 IntrinsicInst *II = dyn_cast<IntrinsicInst>(&CI);
9397 if (!II) return visitCallSite(&CI);
9399 // Intrinsics cannot occur in an invoke, so handle them here instead of in
9401 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(II)) {
9402 bool Changed = false;
9404 // memmove/cpy/set of zero bytes is a noop.
9405 if (Constant *NumBytes = dyn_cast<Constant>(MI->getLength())) {
9406 if (NumBytes->isNullValue()) return EraseInstFromFunction(CI);
9408 if (ConstantInt *CI = dyn_cast<ConstantInt>(NumBytes))
9409 if (CI->getZExtValue() == 1) {
9410 // Replace the instruction with just byte operations. We would
9411 // transform other cases to loads/stores, but we don't know if
9412 // alignment is sufficient.
9416 // If we have a memmove and the source operation is a constant global,
9417 // then the source and dest pointers can't alias, so we can change this
9418 // into a call to memcpy.
9419 if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(MI)) {
9420 if (GlobalVariable *GVSrc = dyn_cast<GlobalVariable>(MMI->getSource()))
9421 if (GVSrc->isConstant()) {
9422 Module *M = CI.getParent()->getParent()->getParent();
9423 Intrinsic::ID MemCpyID = Intrinsic::memcpy;
9425 Tys[0] = CI.getOperand(3)->getType();
9427 Intrinsic::getDeclaration(M, MemCpyID, Tys, 1));
9431 // memmove(x,x,size) -> noop.
9432 if (MMI->getSource() == MMI->getDest())
9433 return EraseInstFromFunction(CI);
9436 // If we can determine a pointer alignment that is bigger than currently
9437 // set, update the alignment.
9438 if (isa<MemCpyInst>(MI) || isa<MemMoveInst>(MI)) {
9439 if (Instruction *I = SimplifyMemTransfer(MI))
9441 } else if (MemSetInst *MSI = dyn_cast<MemSetInst>(MI)) {
9442 if (Instruction *I = SimplifyMemSet(MSI))
9446 if (Changed) return II;
9449 switch (II->getIntrinsicID()) {
9451 case Intrinsic::bswap:
9452 // bswap(bswap(x)) -> x
9453 if (IntrinsicInst *Operand = dyn_cast<IntrinsicInst>(II->getOperand(1)))
9454 if (Operand->getIntrinsicID() == Intrinsic::bswap)
9455 return ReplaceInstUsesWith(CI, Operand->getOperand(1));
9457 case Intrinsic::ppc_altivec_lvx:
9458 case Intrinsic::ppc_altivec_lvxl:
9459 case Intrinsic::x86_sse_loadu_ps:
9460 case Intrinsic::x86_sse2_loadu_pd:
9461 case Intrinsic::x86_sse2_loadu_dq:
9462 // Turn PPC lvx -> load if the pointer is known aligned.
9463 // Turn X86 loadups -> load if the pointer is known aligned.
9464 if (GetOrEnforceKnownAlignment(II->getOperand(1), 16) >= 16) {
9465 Value *Ptr = InsertBitCastBefore(II->getOperand(1),
9466 PointerType::getUnqual(II->getType()),
9468 return new LoadInst(Ptr);
9471 case Intrinsic::ppc_altivec_stvx:
9472 case Intrinsic::ppc_altivec_stvxl:
9473 // Turn stvx -> store if the pointer is known aligned.
9474 if (GetOrEnforceKnownAlignment(II->getOperand(2), 16) >= 16) {
9475 const Type *OpPtrTy =
9476 PointerType::getUnqual(II->getOperand(1)->getType());
9477 Value *Ptr = InsertBitCastBefore(II->getOperand(2), OpPtrTy, CI);
9478 return new StoreInst(II->getOperand(1), Ptr);
9481 case Intrinsic::x86_sse_storeu_ps:
9482 case Intrinsic::x86_sse2_storeu_pd:
9483 case Intrinsic::x86_sse2_storeu_dq:
9484 // Turn X86 storeu -> store if the pointer is known aligned.
9485 if (GetOrEnforceKnownAlignment(II->getOperand(1), 16) >= 16) {
9486 const Type *OpPtrTy =
9487 PointerType::getUnqual(II->getOperand(2)->getType());
9488 Value *Ptr = InsertBitCastBefore(II->getOperand(1), OpPtrTy, CI);
9489 return new StoreInst(II->getOperand(2), Ptr);
9493 case Intrinsic::x86_sse_cvttss2si: {
9494 // These intrinsics only demands the 0th element of its input vector. If
9495 // we can simplify the input based on that, do so now.
9497 if (Value *V = SimplifyDemandedVectorElts(II->getOperand(1), 1,
9499 II->setOperand(1, V);
9505 case Intrinsic::ppc_altivec_vperm:
9506 // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant.
9507 if (ConstantVector *Mask = dyn_cast<ConstantVector>(II->getOperand(3))) {
9508 assert(Mask->getNumOperands() == 16 && "Bad type for intrinsic!");
9510 // Check that all of the elements are integer constants or undefs.
9511 bool AllEltsOk = true;
9512 for (unsigned i = 0; i != 16; ++i) {
9513 if (!isa<ConstantInt>(Mask->getOperand(i)) &&
9514 !isa<UndefValue>(Mask->getOperand(i))) {
9521 // Cast the input vectors to byte vectors.
9522 Value *Op0 =InsertBitCastBefore(II->getOperand(1),Mask->getType(),CI);
9523 Value *Op1 =InsertBitCastBefore(II->getOperand(2),Mask->getType(),CI);
9524 Value *Result = UndefValue::get(Op0->getType());
9526 // Only extract each element once.
9527 Value *ExtractedElts[32];
9528 memset(ExtractedElts, 0, sizeof(ExtractedElts));
9530 for (unsigned i = 0; i != 16; ++i) {
9531 if (isa<UndefValue>(Mask->getOperand(i)))
9533 unsigned Idx=cast<ConstantInt>(Mask->getOperand(i))->getZExtValue();
9534 Idx &= 31; // Match the hardware behavior.
9536 if (ExtractedElts[Idx] == 0) {
9538 new ExtractElementInst(Idx < 16 ? Op0 : Op1, Idx&15, "tmp");
9539 InsertNewInstBefore(Elt, CI);
9540 ExtractedElts[Idx] = Elt;
9543 // Insert this value into the result vector.
9544 Result = InsertElementInst::Create(Result, ExtractedElts[Idx],
9546 InsertNewInstBefore(cast<Instruction>(Result), CI);
9548 return CastInst::Create(Instruction::BitCast, Result, CI.getType());
9553 case Intrinsic::stackrestore: {
9554 // If the save is right next to the restore, remove the restore. This can
9555 // happen when variable allocas are DCE'd.
9556 if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getOperand(1))) {
9557 if (SS->getIntrinsicID() == Intrinsic::stacksave) {
9558 BasicBlock::iterator BI = SS;
9560 return EraseInstFromFunction(CI);
9564 // Scan down this block to see if there is another stack restore in the
9565 // same block without an intervening call/alloca.
9566 BasicBlock::iterator BI = II;
9567 TerminatorInst *TI = II->getParent()->getTerminator();
9568 bool CannotRemove = false;
9569 for (++BI; &*BI != TI; ++BI) {
9570 if (isa<AllocaInst>(BI)) {
9571 CannotRemove = true;
9574 if (CallInst *BCI = dyn_cast<CallInst>(BI)) {
9575 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(BCI)) {
9576 // If there is a stackrestore below this one, remove this one.
9577 if (II->getIntrinsicID() == Intrinsic::stackrestore)
9578 return EraseInstFromFunction(CI);
9579 // Otherwise, ignore the intrinsic.
9581 // If we found a non-intrinsic call, we can't remove the stack
9583 CannotRemove = true;
9589 // If the stack restore is in a return/unwind block and if there are no
9590 // allocas or calls between the restore and the return, nuke the restore.
9591 if (!CannotRemove && (isa<ReturnInst>(TI) || isa<UnwindInst>(TI)))
9592 return EraseInstFromFunction(CI);
9597 return visitCallSite(II);
9600 // InvokeInst simplification
9602 Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) {
9603 return visitCallSite(&II);
9606 /// isSafeToEliminateVarargsCast - If this cast does not affect the value
9607 /// passed through the varargs area, we can eliminate the use of the cast.
9608 static bool isSafeToEliminateVarargsCast(const CallSite CS,
9609 const CastInst * const CI,
9610 const TargetData * const TD,
9612 if (!CI->isLosslessCast())
9615 // The size of ByVal arguments is derived from the type, so we
9616 // can't change to a type with a different size. If the size were
9617 // passed explicitly we could avoid this check.
9618 if (!CS.paramHasAttr(ix, Attribute::ByVal))
9622 cast<PointerType>(CI->getOperand(0)->getType())->getElementType();
9623 const Type* DstTy = cast<PointerType>(CI->getType())->getElementType();
9624 if (!SrcTy->isSized() || !DstTy->isSized())
9626 if (TD->getTypePaddedSize(SrcTy) != TD->getTypePaddedSize(DstTy))
9631 // visitCallSite - Improvements for call and invoke instructions.
9633 Instruction *InstCombiner::visitCallSite(CallSite CS) {
9634 bool Changed = false;
9636 // If the callee is a constexpr cast of a function, attempt to move the cast
9637 // to the arguments of the call/invoke.
9638 if (transformConstExprCastCall(CS)) return 0;
9640 Value *Callee = CS.getCalledValue();
9642 if (Function *CalleeF = dyn_cast<Function>(Callee))
9643 if (CalleeF->getCallingConv() != CS.getCallingConv()) {
9644 Instruction *OldCall = CS.getInstruction();
9645 // If the call and callee calling conventions don't match, this call must
9646 // be unreachable, as the call is undefined.
9647 new StoreInst(ConstantInt::getTrue(),
9648 UndefValue::get(PointerType::getUnqual(Type::Int1Ty)),
9650 if (!OldCall->use_empty())
9651 OldCall->replaceAllUsesWith(UndefValue::get(OldCall->getType()));
9652 if (isa<CallInst>(OldCall)) // Not worth removing an invoke here.
9653 return EraseInstFromFunction(*OldCall);
9657 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) {
9658 // This instruction is not reachable, just remove it. We insert a store to
9659 // undef so that we know that this code is not reachable, despite the fact
9660 // that we can't modify the CFG here.
9661 new StoreInst(ConstantInt::getTrue(),
9662 UndefValue::get(PointerType::getUnqual(Type::Int1Ty)),
9663 CS.getInstruction());
9665 if (!CS.getInstruction()->use_empty())
9666 CS.getInstruction()->
9667 replaceAllUsesWith(UndefValue::get(CS.getInstruction()->getType()));
9669 if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) {
9670 // Don't break the CFG, insert a dummy cond branch.
9671 BranchInst::Create(II->getNormalDest(), II->getUnwindDest(),
9672 ConstantInt::getTrue(), II);
9674 return EraseInstFromFunction(*CS.getInstruction());
9677 if (BitCastInst *BC = dyn_cast<BitCastInst>(Callee))
9678 if (IntrinsicInst *In = dyn_cast<IntrinsicInst>(BC->getOperand(0)))
9679 if (In->getIntrinsicID() == Intrinsic::init_trampoline)
9680 return transformCallThroughTrampoline(CS);
9682 const PointerType *PTy = cast<PointerType>(Callee->getType());
9683 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
9684 if (FTy->isVarArg()) {
9685 int ix = FTy->getNumParams() + (isa<InvokeInst>(Callee) ? 3 : 1);
9686 // See if we can optimize any arguments passed through the varargs area of
9688 for (CallSite::arg_iterator I = CS.arg_begin()+FTy->getNumParams(),
9689 E = CS.arg_end(); I != E; ++I, ++ix) {
9690 CastInst *CI = dyn_cast<CastInst>(*I);
9691 if (CI && isSafeToEliminateVarargsCast(CS, CI, TD, ix)) {
9692 *I = CI->getOperand(0);
9698 if (isa<InlineAsm>(Callee) && !CS.doesNotThrow()) {
9699 // Inline asm calls cannot throw - mark them 'nounwind'.
9700 CS.setDoesNotThrow();
9704 return Changed ? CS.getInstruction() : 0;
9707 // transformConstExprCastCall - If the callee is a constexpr cast of a function,
9708 // attempt to move the cast to the arguments of the call/invoke.
9710 bool InstCombiner::transformConstExprCastCall(CallSite CS) {
9711 if (!isa<ConstantExpr>(CS.getCalledValue())) return false;
9712 ConstantExpr *CE = cast<ConstantExpr>(CS.getCalledValue());
9713 if (CE->getOpcode() != Instruction::BitCast ||
9714 !isa<Function>(CE->getOperand(0)))
9716 Function *Callee = cast<Function>(CE->getOperand(0));
9717 Instruction *Caller = CS.getInstruction();
9718 const AttrListPtr &CallerPAL = CS.getAttributes();
9720 // Okay, this is a cast from a function to a different type. Unless doing so
9721 // would cause a type conversion of one of our arguments, change this call to
9722 // be a direct call with arguments casted to the appropriate types.
9724 const FunctionType *FT = Callee->getFunctionType();
9725 const Type *OldRetTy = Caller->getType();
9726 const Type *NewRetTy = FT->getReturnType();
9728 if (isa<StructType>(NewRetTy))
9729 return false; // TODO: Handle multiple return values.
9731 // Check to see if we are changing the return type...
9732 if (OldRetTy != NewRetTy) {
9733 if (Callee->isDeclaration() &&
9734 // Conversion is ok if changing from one pointer type to another or from
9735 // a pointer to an integer of the same size.
9736 !((isa<PointerType>(OldRetTy) || OldRetTy == TD->getIntPtrType()) &&
9737 (isa<PointerType>(NewRetTy) || NewRetTy == TD->getIntPtrType())))
9738 return false; // Cannot transform this return value.
9740 if (!Caller->use_empty() &&
9741 // void -> non-void is handled specially
9742 NewRetTy != Type::VoidTy && !CastInst::isCastable(NewRetTy, OldRetTy))
9743 return false; // Cannot transform this return value.
9745 if (!CallerPAL.isEmpty() && !Caller->use_empty()) {
9746 Attributes RAttrs = CallerPAL.getRetAttributes();
9747 if (RAttrs & Attribute::typeIncompatible(NewRetTy))
9748 return false; // Attribute not compatible with transformed value.
9751 // If the callsite is an invoke instruction, and the return value is used by
9752 // a PHI node in a successor, we cannot change the return type of the call
9753 // because there is no place to put the cast instruction (without breaking
9754 // the critical edge). Bail out in this case.
9755 if (!Caller->use_empty())
9756 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller))
9757 for (Value::use_iterator UI = II->use_begin(), E = II->use_end();
9759 if (PHINode *PN = dyn_cast<PHINode>(*UI))
9760 if (PN->getParent() == II->getNormalDest() ||
9761 PN->getParent() == II->getUnwindDest())
9765 unsigned NumActualArgs = unsigned(CS.arg_end()-CS.arg_begin());
9766 unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs);
9768 CallSite::arg_iterator AI = CS.arg_begin();
9769 for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) {
9770 const Type *ParamTy = FT->getParamType(i);
9771 const Type *ActTy = (*AI)->getType();
9773 if (!CastInst::isCastable(ActTy, ParamTy))
9774 return false; // Cannot transform this parameter value.
9776 if (CallerPAL.getParamAttributes(i + 1)
9777 & Attribute::typeIncompatible(ParamTy))
9778 return false; // Attribute not compatible with transformed value.
9780 // Converting from one pointer type to another or between a pointer and an
9781 // integer of the same size is safe even if we do not have a body.
9782 bool isConvertible = ActTy == ParamTy ||
9783 ((isa<PointerType>(ParamTy) || ParamTy == TD->getIntPtrType()) &&
9784 (isa<PointerType>(ActTy) || ActTy == TD->getIntPtrType()));
9785 if (Callee->isDeclaration() && !isConvertible) return false;
9788 if (FT->getNumParams() < NumActualArgs && !FT->isVarArg() &&
9789 Callee->isDeclaration())
9790 return false; // Do not delete arguments unless we have a function body.
9792 if (FT->getNumParams() < NumActualArgs && FT->isVarArg() &&
9793 !CallerPAL.isEmpty())
9794 // In this case we have more arguments than the new function type, but we
9795 // won't be dropping them. Check that these extra arguments have attributes
9796 // that are compatible with being a vararg call argument.
9797 for (unsigned i = CallerPAL.getNumSlots(); i; --i) {
9798 if (CallerPAL.getSlot(i - 1).Index <= FT->getNumParams())
9800 Attributes PAttrs = CallerPAL.getSlot(i - 1).Attrs;
9801 if (PAttrs & Attribute::VarArgsIncompatible)
9805 // Okay, we decided that this is a safe thing to do: go ahead and start
9806 // inserting cast instructions as necessary...
9807 std::vector<Value*> Args;
9808 Args.reserve(NumActualArgs);
9809 SmallVector<AttributeWithIndex, 8> attrVec;
9810 attrVec.reserve(NumCommonArgs);
9812 // Get any return attributes.
9813 Attributes RAttrs = CallerPAL.getRetAttributes();
9815 // If the return value is not being used, the type may not be compatible
9816 // with the existing attributes. Wipe out any problematic attributes.
9817 RAttrs &= ~Attribute::typeIncompatible(NewRetTy);
9819 // Add the new return attributes.
9821 attrVec.push_back(AttributeWithIndex::get(0, RAttrs));
9823 AI = CS.arg_begin();
9824 for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) {
9825 const Type *ParamTy = FT->getParamType(i);
9826 if ((*AI)->getType() == ParamTy) {
9827 Args.push_back(*AI);
9829 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI,
9830 false, ParamTy, false);
9831 CastInst *NewCast = CastInst::Create(opcode, *AI, ParamTy, "tmp");
9832 Args.push_back(InsertNewInstBefore(NewCast, *Caller));
9835 // Add any parameter attributes.
9836 if (Attributes PAttrs = CallerPAL.getParamAttributes(i + 1))
9837 attrVec.push_back(AttributeWithIndex::get(i + 1, PAttrs));
9840 // If the function takes more arguments than the call was taking, add them
9842 for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i)
9843 Args.push_back(Constant::getNullValue(FT->getParamType(i)));
9845 // If we are removing arguments to the function, emit an obnoxious warning...
9846 if (FT->getNumParams() < NumActualArgs) {
9847 if (!FT->isVarArg()) {
9848 cerr << "WARNING: While resolving call to function '"
9849 << Callee->getName() << "' arguments were dropped!\n";
9851 // Add all of the arguments in their promoted form to the arg list...
9852 for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) {
9853 const Type *PTy = getPromotedType((*AI)->getType());
9854 if (PTy != (*AI)->getType()) {
9855 // Must promote to pass through va_arg area!
9856 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI, false,
9858 Instruction *Cast = CastInst::Create(opcode, *AI, PTy, "tmp");
9859 InsertNewInstBefore(Cast, *Caller);
9860 Args.push_back(Cast);
9862 Args.push_back(*AI);
9865 // Add any parameter attributes.
9866 if (Attributes PAttrs = CallerPAL.getParamAttributes(i + 1))
9867 attrVec.push_back(AttributeWithIndex::get(i + 1, PAttrs));
9872 if (Attributes FnAttrs = CallerPAL.getFnAttributes())
9873 attrVec.push_back(AttributeWithIndex::get(~0, FnAttrs));
9875 if (NewRetTy == Type::VoidTy)
9876 Caller->setName(""); // Void type should not have a name.
9878 const AttrListPtr &NewCallerPAL = AttrListPtr::get(attrVec.begin(),attrVec.end());
9881 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
9882 NC = InvokeInst::Create(Callee, II->getNormalDest(), II->getUnwindDest(),
9883 Args.begin(), Args.end(),
9884 Caller->getName(), Caller);
9885 cast<InvokeInst>(NC)->setCallingConv(II->getCallingConv());
9886 cast<InvokeInst>(NC)->setAttributes(NewCallerPAL);
9888 NC = CallInst::Create(Callee, Args.begin(), Args.end(),
9889 Caller->getName(), Caller);
9890 CallInst *CI = cast<CallInst>(Caller);
9891 if (CI->isTailCall())
9892 cast<CallInst>(NC)->setTailCall();
9893 cast<CallInst>(NC)->setCallingConv(CI->getCallingConv());
9894 cast<CallInst>(NC)->setAttributes(NewCallerPAL);
9897 // Insert a cast of the return type as necessary.
9899 if (OldRetTy != NV->getType() && !Caller->use_empty()) {
9900 if (NV->getType() != Type::VoidTy) {
9901 Instruction::CastOps opcode = CastInst::getCastOpcode(NC, false,
9903 NV = NC = CastInst::Create(opcode, NC, OldRetTy, "tmp");
9905 // If this is an invoke instruction, we should insert it after the first
9906 // non-phi, instruction in the normal successor block.
9907 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
9908 BasicBlock::iterator I = II->getNormalDest()->getFirstNonPHI();
9909 InsertNewInstBefore(NC, *I);
9911 // Otherwise, it's a call, just insert cast right after the call instr
9912 InsertNewInstBefore(NC, *Caller);
9914 AddUsersToWorkList(*Caller);
9916 NV = UndefValue::get(Caller->getType());
9920 if (Caller->getType() != Type::VoidTy && !Caller->use_empty())
9921 Caller->replaceAllUsesWith(NV);
9922 Caller->eraseFromParent();
9923 RemoveFromWorkList(Caller);
9927 // transformCallThroughTrampoline - Turn a call to a function created by the
9928 // init_trampoline intrinsic into a direct call to the underlying function.
9930 Instruction *InstCombiner::transformCallThroughTrampoline(CallSite CS) {
9931 Value *Callee = CS.getCalledValue();
9932 const PointerType *PTy = cast<PointerType>(Callee->getType());
9933 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
9934 const AttrListPtr &Attrs = CS.getAttributes();
9936 // If the call already has the 'nest' attribute somewhere then give up -
9937 // otherwise 'nest' would occur twice after splicing in the chain.
9938 if (Attrs.hasAttrSomewhere(Attribute::Nest))
9941 IntrinsicInst *Tramp =
9942 cast<IntrinsicInst>(cast<BitCastInst>(Callee)->getOperand(0));
9944 Function *NestF = cast<Function>(Tramp->getOperand(2)->stripPointerCasts());
9945 const PointerType *NestFPTy = cast<PointerType>(NestF->getType());
9946 const FunctionType *NestFTy = cast<FunctionType>(NestFPTy->getElementType());
9948 const AttrListPtr &NestAttrs = NestF->getAttributes();
9949 if (!NestAttrs.isEmpty()) {
9950 unsigned NestIdx = 1;
9951 const Type *NestTy = 0;
9952 Attributes NestAttr = Attribute::None;
9954 // Look for a parameter marked with the 'nest' attribute.
9955 for (FunctionType::param_iterator I = NestFTy->param_begin(),
9956 E = NestFTy->param_end(); I != E; ++NestIdx, ++I)
9957 if (NestAttrs.paramHasAttr(NestIdx, Attribute::Nest)) {
9958 // Record the parameter type and any other attributes.
9960 NestAttr = NestAttrs.getParamAttributes(NestIdx);
9965 Instruction *Caller = CS.getInstruction();
9966 std::vector<Value*> NewArgs;
9967 NewArgs.reserve(unsigned(CS.arg_end()-CS.arg_begin())+1);
9969 SmallVector<AttributeWithIndex, 8> NewAttrs;
9970 NewAttrs.reserve(Attrs.getNumSlots() + 1);
9972 // Insert the nest argument into the call argument list, which may
9973 // mean appending it. Likewise for attributes.
9975 // Add any result attributes.
9976 if (Attributes Attr = Attrs.getRetAttributes())
9977 NewAttrs.push_back(AttributeWithIndex::get(0, Attr));
9981 CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
9983 if (Idx == NestIdx) {
9984 // Add the chain argument and attributes.
9985 Value *NestVal = Tramp->getOperand(3);
9986 if (NestVal->getType() != NestTy)
9987 NestVal = new BitCastInst(NestVal, NestTy, "nest", Caller);
9988 NewArgs.push_back(NestVal);
9989 NewAttrs.push_back(AttributeWithIndex::get(NestIdx, NestAttr));
9995 // Add the original argument and attributes.
9996 NewArgs.push_back(*I);
9997 if (Attributes Attr = Attrs.getParamAttributes(Idx))
9999 (AttributeWithIndex::get(Idx + (Idx >= NestIdx), Attr));
10005 // Add any function attributes.
10006 if (Attributes Attr = Attrs.getFnAttributes())
10007 NewAttrs.push_back(AttributeWithIndex::get(~0, Attr));
10009 // The trampoline may have been bitcast to a bogus type (FTy).
10010 // Handle this by synthesizing a new function type, equal to FTy
10011 // with the chain parameter inserted.
10013 std::vector<const Type*> NewTypes;
10014 NewTypes.reserve(FTy->getNumParams()+1);
10016 // Insert the chain's type into the list of parameter types, which may
10017 // mean appending it.
10020 FunctionType::param_iterator I = FTy->param_begin(),
10021 E = FTy->param_end();
10024 if (Idx == NestIdx)
10025 // Add the chain's type.
10026 NewTypes.push_back(NestTy);
10031 // Add the original type.
10032 NewTypes.push_back(*I);
10038 // Replace the trampoline call with a direct call. Let the generic
10039 // code sort out any function type mismatches.
10040 FunctionType *NewFTy =
10041 FunctionType::get(FTy->getReturnType(), NewTypes, FTy->isVarArg());
10042 Constant *NewCallee = NestF->getType() == PointerType::getUnqual(NewFTy) ?
10043 NestF : ConstantExpr::getBitCast(NestF, PointerType::getUnqual(NewFTy));
10044 const AttrListPtr &NewPAL = AttrListPtr::get(NewAttrs.begin(),NewAttrs.end());
10046 Instruction *NewCaller;
10047 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
10048 NewCaller = InvokeInst::Create(NewCallee,
10049 II->getNormalDest(), II->getUnwindDest(),
10050 NewArgs.begin(), NewArgs.end(),
10051 Caller->getName(), Caller);
10052 cast<InvokeInst>(NewCaller)->setCallingConv(II->getCallingConv());
10053 cast<InvokeInst>(NewCaller)->setAttributes(NewPAL);
10055 NewCaller = CallInst::Create(NewCallee, NewArgs.begin(), NewArgs.end(),
10056 Caller->getName(), Caller);
10057 if (cast<CallInst>(Caller)->isTailCall())
10058 cast<CallInst>(NewCaller)->setTailCall();
10059 cast<CallInst>(NewCaller)->
10060 setCallingConv(cast<CallInst>(Caller)->getCallingConv());
10061 cast<CallInst>(NewCaller)->setAttributes(NewPAL);
10063 if (Caller->getType() != Type::VoidTy && !Caller->use_empty())
10064 Caller->replaceAllUsesWith(NewCaller);
10065 Caller->eraseFromParent();
10066 RemoveFromWorkList(Caller);
10071 // Replace the trampoline call with a direct call. Since there is no 'nest'
10072 // parameter, there is no need to adjust the argument list. Let the generic
10073 // code sort out any function type mismatches.
10074 Constant *NewCallee =
10075 NestF->getType() == PTy ? NestF : ConstantExpr::getBitCast(NestF, PTy);
10076 CS.setCalledFunction(NewCallee);
10077 return CS.getInstruction();
10080 /// FoldPHIArgBinOpIntoPHI - If we have something like phi [add (a,b), add(c,d)]
10081 /// and if a/b/c/d and the add's all have a single use, turn this into two phi's
10082 /// and a single binop.
10083 Instruction *InstCombiner::FoldPHIArgBinOpIntoPHI(PHINode &PN) {
10084 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
10085 assert(isa<BinaryOperator>(FirstInst) || isa<CmpInst>(FirstInst));
10086 unsigned Opc = FirstInst->getOpcode();
10087 Value *LHSVal = FirstInst->getOperand(0);
10088 Value *RHSVal = FirstInst->getOperand(1);
10090 const Type *LHSType = LHSVal->getType();
10091 const Type *RHSType = RHSVal->getType();
10093 // Scan to see if all operands are the same opcode, all have one use, and all
10094 // kill their operands (i.e. the operands have one use).
10095 for (unsigned i = 1; i != PN.getNumIncomingValues(); ++i) {
10096 Instruction *I = dyn_cast<Instruction>(PN.getIncomingValue(i));
10097 if (!I || I->getOpcode() != Opc || !I->hasOneUse() ||
10098 // Verify type of the LHS matches so we don't fold cmp's of different
10099 // types or GEP's with different index types.
10100 I->getOperand(0)->getType() != LHSType ||
10101 I->getOperand(1)->getType() != RHSType)
10104 // If they are CmpInst instructions, check their predicates
10105 if (Opc == Instruction::ICmp || Opc == Instruction::FCmp)
10106 if (cast<CmpInst>(I)->getPredicate() !=
10107 cast<CmpInst>(FirstInst)->getPredicate())
10110 // Keep track of which operand needs a phi node.
10111 if (I->getOperand(0) != LHSVal) LHSVal = 0;
10112 if (I->getOperand(1) != RHSVal) RHSVal = 0;
10115 // Otherwise, this is safe to transform!
10117 Value *InLHS = FirstInst->getOperand(0);
10118 Value *InRHS = FirstInst->getOperand(1);
10119 PHINode *NewLHS = 0, *NewRHS = 0;
10121 NewLHS = PHINode::Create(LHSType,
10122 FirstInst->getOperand(0)->getName() + ".pn");
10123 NewLHS->reserveOperandSpace(PN.getNumOperands()/2);
10124 NewLHS->addIncoming(InLHS, PN.getIncomingBlock(0));
10125 InsertNewInstBefore(NewLHS, PN);
10130 NewRHS = PHINode::Create(RHSType,
10131 FirstInst->getOperand(1)->getName() + ".pn");
10132 NewRHS->reserveOperandSpace(PN.getNumOperands()/2);
10133 NewRHS->addIncoming(InRHS, PN.getIncomingBlock(0));
10134 InsertNewInstBefore(NewRHS, PN);
10138 // Add all operands to the new PHIs.
10139 if (NewLHS || NewRHS) {
10140 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
10141 Instruction *InInst = cast<Instruction>(PN.getIncomingValue(i));
10143 Value *NewInLHS = InInst->getOperand(0);
10144 NewLHS->addIncoming(NewInLHS, PN.getIncomingBlock(i));
10147 Value *NewInRHS = InInst->getOperand(1);
10148 NewRHS->addIncoming(NewInRHS, PN.getIncomingBlock(i));
10153 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
10154 return BinaryOperator::Create(BinOp->getOpcode(), LHSVal, RHSVal);
10155 CmpInst *CIOp = cast<CmpInst>(FirstInst);
10156 return CmpInst::Create(CIOp->getOpcode(), CIOp->getPredicate(), LHSVal,
10160 Instruction *InstCombiner::FoldPHIArgGEPIntoPHI(PHINode &PN) {
10161 GetElementPtrInst *FirstInst =cast<GetElementPtrInst>(PN.getIncomingValue(0));
10163 SmallVector<Value*, 16> FixedOperands(FirstInst->op_begin(),
10164 FirstInst->op_end());
10166 // Scan to see if all operands are the same opcode, all have one use, and all
10167 // kill their operands (i.e. the operands have one use).
10168 for (unsigned i = 1; i != PN.getNumIncomingValues(); ++i) {
10169 GetElementPtrInst *GEP= dyn_cast<GetElementPtrInst>(PN.getIncomingValue(i));
10170 if (!GEP || !GEP->hasOneUse() || GEP->getType() != FirstInst->getType() ||
10171 GEP->getNumOperands() != FirstInst->getNumOperands())
10174 // Compare the operand lists.
10175 for (unsigned op = 0, e = FirstInst->getNumOperands(); op != e; ++op) {
10176 if (FirstInst->getOperand(op) == GEP->getOperand(op))
10179 // Don't merge two GEPs when two operands differ (introducing phi nodes)
10180 // if one of the PHIs has a constant for the index. The index may be
10181 // substantially cheaper to compute for the constants, so making it a
10182 // variable index could pessimize the path. This also handles the case
10183 // for struct indices, which must always be constant.
10184 if (isa<ConstantInt>(FirstInst->getOperand(op)) ||
10185 isa<ConstantInt>(GEP->getOperand(op)))
10188 if (FirstInst->getOperand(op)->getType() !=GEP->getOperand(op)->getType())
10190 FixedOperands[op] = 0; // Needs a PHI.
10194 // Otherwise, this is safe to transform. Insert PHI nodes for each operand
10195 // that is variable.
10196 SmallVector<PHINode*, 16> OperandPhis(FixedOperands.size());
10198 bool HasAnyPHIs = false;
10199 for (unsigned i = 0, e = FixedOperands.size(); i != e; ++i) {
10200 if (FixedOperands[i]) continue; // operand doesn't need a phi.
10201 Value *FirstOp = FirstInst->getOperand(i);
10202 PHINode *NewPN = PHINode::Create(FirstOp->getType(),
10203 FirstOp->getName()+".pn");
10204 InsertNewInstBefore(NewPN, PN);
10206 NewPN->reserveOperandSpace(e);
10207 NewPN->addIncoming(FirstOp, PN.getIncomingBlock(0));
10208 OperandPhis[i] = NewPN;
10209 FixedOperands[i] = NewPN;
10214 // Add all operands to the new PHIs.
10216 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
10217 GetElementPtrInst *InGEP =cast<GetElementPtrInst>(PN.getIncomingValue(i));
10218 BasicBlock *InBB = PN.getIncomingBlock(i);
10220 for (unsigned op = 0, e = OperandPhis.size(); op != e; ++op)
10221 if (PHINode *OpPhi = OperandPhis[op])
10222 OpPhi->addIncoming(InGEP->getOperand(op), InBB);
10226 Value *Base = FixedOperands[0];
10227 return GetElementPtrInst::Create(Base, FixedOperands.begin()+1,
10228 FixedOperands.end());
10232 /// isSafeToSinkLoad - Return true if we know that it is safe sink the load out
10233 /// of the block that defines it. This means that it must be obvious the value
10234 /// of the load is not changed from the point of the load to the end of the
10235 /// block it is in.
10237 /// Finally, it is safe, but not profitable, to sink a load targetting a
10238 /// non-address-taken alloca. Doing so will cause us to not promote the alloca
10240 static bool isSafeToSinkLoad(LoadInst *L) {
10241 BasicBlock::iterator BBI = L, E = L->getParent()->end();
10243 for (++BBI; BBI != E; ++BBI)
10244 if (BBI->mayWriteToMemory())
10247 // Check for non-address taken alloca. If not address-taken already, it isn't
10248 // profitable to do this xform.
10249 if (AllocaInst *AI = dyn_cast<AllocaInst>(L->getOperand(0))) {
10250 bool isAddressTaken = false;
10251 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
10253 if (isa<LoadInst>(UI)) continue;
10254 if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
10255 // If storing TO the alloca, then the address isn't taken.
10256 if (SI->getOperand(1) == AI) continue;
10258 isAddressTaken = true;
10262 if (!isAddressTaken)
10270 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
10271 // operator and they all are only used by the PHI, PHI together their
10272 // inputs, and do the operation once, to the result of the PHI.
10273 Instruction *InstCombiner::FoldPHIArgOpIntoPHI(PHINode &PN) {
10274 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
10276 // Scan the instruction, looking for input operations that can be folded away.
10277 // If all input operands to the phi are the same instruction (e.g. a cast from
10278 // the same type or "+42") we can pull the operation through the PHI, reducing
10279 // code size and simplifying code.
10280 Constant *ConstantOp = 0;
10281 const Type *CastSrcTy = 0;
10282 bool isVolatile = false;
10283 if (isa<CastInst>(FirstInst)) {
10284 CastSrcTy = FirstInst->getOperand(0)->getType();
10285 } else if (isa<BinaryOperator>(FirstInst) || isa<CmpInst>(FirstInst)) {
10286 // Can fold binop, compare or shift here if the RHS is a constant,
10287 // otherwise call FoldPHIArgBinOpIntoPHI.
10288 ConstantOp = dyn_cast<Constant>(FirstInst->getOperand(1));
10289 if (ConstantOp == 0)
10290 return FoldPHIArgBinOpIntoPHI(PN);
10291 } else if (LoadInst *LI = dyn_cast<LoadInst>(FirstInst)) {
10292 isVolatile = LI->isVolatile();
10293 // We can't sink the load if the loaded value could be modified between the
10294 // load and the PHI.
10295 if (LI->getParent() != PN.getIncomingBlock(0) ||
10296 !isSafeToSinkLoad(LI))
10299 // If the PHI is of volatile loads and the load block has multiple
10300 // successors, sinking it would remove a load of the volatile value from
10301 // the path through the other successor.
10303 LI->getParent()->getTerminator()->getNumSuccessors() != 1)
10306 } else if (isa<GetElementPtrInst>(FirstInst)) {
10307 return FoldPHIArgGEPIntoPHI(PN);
10309 return 0; // Cannot fold this operation.
10312 // Check to see if all arguments are the same operation.
10313 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
10314 if (!isa<Instruction>(PN.getIncomingValue(i))) return 0;
10315 Instruction *I = cast<Instruction>(PN.getIncomingValue(i));
10316 if (!I->hasOneUse() || !I->isSameOperationAs(FirstInst))
10319 if (I->getOperand(0)->getType() != CastSrcTy)
10320 return 0; // Cast operation must match.
10321 } else if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
10322 // We can't sink the load if the loaded value could be modified between
10323 // the load and the PHI.
10324 if (LI->isVolatile() != isVolatile ||
10325 LI->getParent() != PN.getIncomingBlock(i) ||
10326 !isSafeToSinkLoad(LI))
10329 // If the PHI is of volatile loads and the load block has multiple
10330 // successors, sinking it would remove a load of the volatile value from
10331 // the path through the other successor.
10333 LI->getParent()->getTerminator()->getNumSuccessors() != 1)
10337 } else if (I->getOperand(1) != ConstantOp) {
10342 // Okay, they are all the same operation. Create a new PHI node of the
10343 // correct type, and PHI together all of the LHS's of the instructions.
10344 PHINode *NewPN = PHINode::Create(FirstInst->getOperand(0)->getType(),
10345 PN.getName()+".in");
10346 NewPN->reserveOperandSpace(PN.getNumOperands()/2);
10348 Value *InVal = FirstInst->getOperand(0);
10349 NewPN->addIncoming(InVal, PN.getIncomingBlock(0));
10351 // Add all operands to the new PHI.
10352 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
10353 Value *NewInVal = cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
10354 if (NewInVal != InVal)
10356 NewPN->addIncoming(NewInVal, PN.getIncomingBlock(i));
10361 // The new PHI unions all of the same values together. This is really
10362 // common, so we handle it intelligently here for compile-time speed.
10366 InsertNewInstBefore(NewPN, PN);
10370 // Insert and return the new operation.
10371 if (CastInst* FirstCI = dyn_cast<CastInst>(FirstInst))
10372 return CastInst::Create(FirstCI->getOpcode(), PhiVal, PN.getType());
10373 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
10374 return BinaryOperator::Create(BinOp->getOpcode(), PhiVal, ConstantOp);
10375 if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst))
10376 return CmpInst::Create(CIOp->getOpcode(), CIOp->getPredicate(),
10377 PhiVal, ConstantOp);
10378 assert(isa<LoadInst>(FirstInst) && "Unknown operation");
10380 // If this was a volatile load that we are merging, make sure to loop through
10381 // and mark all the input loads as non-volatile. If we don't do this, we will
10382 // insert a new volatile load and the old ones will not be deletable.
10384 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i)
10385 cast<LoadInst>(PN.getIncomingValue(i))->setVolatile(false);
10387 return new LoadInst(PhiVal, "", isVolatile);
10390 /// DeadPHICycle - Return true if this PHI node is only used by a PHI node cycle
10392 static bool DeadPHICycle(PHINode *PN,
10393 SmallPtrSet<PHINode*, 16> &PotentiallyDeadPHIs) {
10394 if (PN->use_empty()) return true;
10395 if (!PN->hasOneUse()) return false;
10397 // Remember this node, and if we find the cycle, return.
10398 if (!PotentiallyDeadPHIs.insert(PN))
10401 // Don't scan crazily complex things.
10402 if (PotentiallyDeadPHIs.size() == 16)
10405 if (PHINode *PU = dyn_cast<PHINode>(PN->use_back()))
10406 return DeadPHICycle(PU, PotentiallyDeadPHIs);
10411 /// PHIsEqualValue - Return true if this phi node is always equal to
10412 /// NonPhiInVal. This happens with mutually cyclic phi nodes like:
10413 /// z = some value; x = phi (y, z); y = phi (x, z)
10414 static bool PHIsEqualValue(PHINode *PN, Value *NonPhiInVal,
10415 SmallPtrSet<PHINode*, 16> &ValueEqualPHIs) {
10416 // See if we already saw this PHI node.
10417 if (!ValueEqualPHIs.insert(PN))
10420 // Don't scan crazily complex things.
10421 if (ValueEqualPHIs.size() == 16)
10424 // Scan the operands to see if they are either phi nodes or are equal to
10426 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
10427 Value *Op = PN->getIncomingValue(i);
10428 if (PHINode *OpPN = dyn_cast<PHINode>(Op)) {
10429 if (!PHIsEqualValue(OpPN, NonPhiInVal, ValueEqualPHIs))
10431 } else if (Op != NonPhiInVal)
10439 // PHINode simplification
10441 Instruction *InstCombiner::visitPHINode(PHINode &PN) {
10442 // If LCSSA is around, don't mess with Phi nodes
10443 if (MustPreserveLCSSA) return 0;
10445 if (Value *V = PN.hasConstantValue())
10446 return ReplaceInstUsesWith(PN, V);
10448 // If all PHI operands are the same operation, pull them through the PHI,
10449 // reducing code size.
10450 if (isa<Instruction>(PN.getIncomingValue(0)) &&
10451 isa<Instruction>(PN.getIncomingValue(1)) &&
10452 cast<Instruction>(PN.getIncomingValue(0))->getOpcode() ==
10453 cast<Instruction>(PN.getIncomingValue(1))->getOpcode() &&
10454 // FIXME: The hasOneUse check will fail for PHIs that use the value more
10455 // than themselves more than once.
10456 PN.getIncomingValue(0)->hasOneUse())
10457 if (Instruction *Result = FoldPHIArgOpIntoPHI(PN))
10460 // If this is a trivial cycle in the PHI node graph, remove it. Basically, if
10461 // this PHI only has a single use (a PHI), and if that PHI only has one use (a
10462 // PHI)... break the cycle.
10463 if (PN.hasOneUse()) {
10464 Instruction *PHIUser = cast<Instruction>(PN.use_back());
10465 if (PHINode *PU = dyn_cast<PHINode>(PHIUser)) {
10466 SmallPtrSet<PHINode*, 16> PotentiallyDeadPHIs;
10467 PotentiallyDeadPHIs.insert(&PN);
10468 if (DeadPHICycle(PU, PotentiallyDeadPHIs))
10469 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
10472 // If this phi has a single use, and if that use just computes a value for
10473 // the next iteration of a loop, delete the phi. This occurs with unused
10474 // induction variables, e.g. "for (int j = 0; ; ++j);". Detecting this
10475 // common case here is good because the only other things that catch this
10476 // are induction variable analysis (sometimes) and ADCE, which is only run
10478 if (PHIUser->hasOneUse() &&
10479 (isa<BinaryOperator>(PHIUser) || isa<GetElementPtrInst>(PHIUser)) &&
10480 PHIUser->use_back() == &PN) {
10481 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
10485 // We sometimes end up with phi cycles that non-obviously end up being the
10486 // same value, for example:
10487 // z = some value; x = phi (y, z); y = phi (x, z)
10488 // where the phi nodes don't necessarily need to be in the same block. Do a
10489 // quick check to see if the PHI node only contains a single non-phi value, if
10490 // so, scan to see if the phi cycle is actually equal to that value.
10492 unsigned InValNo = 0, NumOperandVals = PN.getNumIncomingValues();
10493 // Scan for the first non-phi operand.
10494 while (InValNo != NumOperandVals &&
10495 isa<PHINode>(PN.getIncomingValue(InValNo)))
10498 if (InValNo != NumOperandVals) {
10499 Value *NonPhiInVal = PN.getOperand(InValNo);
10501 // Scan the rest of the operands to see if there are any conflicts, if so
10502 // there is no need to recursively scan other phis.
10503 for (++InValNo; InValNo != NumOperandVals; ++InValNo) {
10504 Value *OpVal = PN.getIncomingValue(InValNo);
10505 if (OpVal != NonPhiInVal && !isa<PHINode>(OpVal))
10509 // If we scanned over all operands, then we have one unique value plus
10510 // phi values. Scan PHI nodes to see if they all merge in each other or
10512 if (InValNo == NumOperandVals) {
10513 SmallPtrSet<PHINode*, 16> ValueEqualPHIs;
10514 if (PHIsEqualValue(&PN, NonPhiInVal, ValueEqualPHIs))
10515 return ReplaceInstUsesWith(PN, NonPhiInVal);
10522 static Value *InsertCastToIntPtrTy(Value *V, const Type *DTy,
10523 Instruction *InsertPoint,
10524 InstCombiner *IC) {
10525 unsigned PtrSize = DTy->getPrimitiveSizeInBits();
10526 unsigned VTySize = V->getType()->getPrimitiveSizeInBits();
10527 // We must cast correctly to the pointer type. Ensure that we
10528 // sign extend the integer value if it is smaller as this is
10529 // used for address computation.
10530 Instruction::CastOps opcode =
10531 (VTySize < PtrSize ? Instruction::SExt :
10532 (VTySize == PtrSize ? Instruction::BitCast : Instruction::Trunc));
10533 return IC->InsertCastBefore(opcode, V, DTy, *InsertPoint);
10537 Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) {
10538 Value *PtrOp = GEP.getOperand(0);
10539 // Is it 'getelementptr %P, i32 0' or 'getelementptr %P'
10540 // If so, eliminate the noop.
10541 if (GEP.getNumOperands() == 1)
10542 return ReplaceInstUsesWith(GEP, PtrOp);
10544 if (isa<UndefValue>(GEP.getOperand(0)))
10545 return ReplaceInstUsesWith(GEP, UndefValue::get(GEP.getType()));
10547 bool HasZeroPointerIndex = false;
10548 if (Constant *C = dyn_cast<Constant>(GEP.getOperand(1)))
10549 HasZeroPointerIndex = C->isNullValue();
10551 if (GEP.getNumOperands() == 2 && HasZeroPointerIndex)
10552 return ReplaceInstUsesWith(GEP, PtrOp);
10554 // Eliminate unneeded casts for indices.
10555 bool MadeChange = false;
10557 gep_type_iterator GTI = gep_type_begin(GEP);
10558 for (User::op_iterator i = GEP.op_begin() + 1, e = GEP.op_end();
10559 i != e; ++i, ++GTI) {
10560 if (isa<SequentialType>(*GTI)) {
10561 if (CastInst *CI = dyn_cast<CastInst>(*i)) {
10562 if (CI->getOpcode() == Instruction::ZExt ||
10563 CI->getOpcode() == Instruction::SExt) {
10564 const Type *SrcTy = CI->getOperand(0)->getType();
10565 // We can eliminate a cast from i32 to i64 iff the target
10566 // is a 32-bit pointer target.
10567 if (SrcTy->getPrimitiveSizeInBits() >= TD->getPointerSizeInBits()) {
10569 *i = CI->getOperand(0);
10573 // If we are using a wider index than needed for this platform, shrink it
10574 // to what we need. If narrower, sign-extend it to what we need.
10575 // If the incoming value needs a cast instruction,
10576 // insert it. This explicit cast can make subsequent optimizations more
10579 if (TD->getTypeSizeInBits(Op->getType()) > TD->getPointerSizeInBits()) {
10580 if (Constant *C = dyn_cast<Constant>(Op)) {
10581 *i = ConstantExpr::getTrunc(C, TD->getIntPtrType());
10584 Op = InsertCastBefore(Instruction::Trunc, Op, TD->getIntPtrType(),
10589 } else if (TD->getTypeSizeInBits(Op->getType()) < TD->getPointerSizeInBits()) {
10590 if (Constant *C = dyn_cast<Constant>(Op)) {
10591 *i = ConstantExpr::getSExt(C, TD->getIntPtrType());
10594 Op = InsertCastBefore(Instruction::SExt, Op, TD->getIntPtrType(),
10602 if (MadeChange) return &GEP;
10604 // Combine Indices - If the source pointer to this getelementptr instruction
10605 // is a getelementptr instruction, combine the indices of the two
10606 // getelementptr instructions into a single instruction.
10608 SmallVector<Value*, 8> SrcGEPOperands;
10609 if (User *Src = dyn_castGetElementPtr(PtrOp))
10610 SrcGEPOperands.append(Src->op_begin(), Src->op_end());
10612 if (!SrcGEPOperands.empty()) {
10613 // Note that if our source is a gep chain itself that we wait for that
10614 // chain to be resolved before we perform this transformation. This
10615 // avoids us creating a TON of code in some cases.
10617 if (isa<GetElementPtrInst>(SrcGEPOperands[0]) &&
10618 cast<Instruction>(SrcGEPOperands[0])->getNumOperands() == 2)
10619 return 0; // Wait until our source is folded to completion.
10621 SmallVector<Value*, 8> Indices;
10623 // Find out whether the last index in the source GEP is a sequential idx.
10624 bool EndsWithSequential = false;
10625 for (gep_type_iterator I = gep_type_begin(*cast<User>(PtrOp)),
10626 E = gep_type_end(*cast<User>(PtrOp)); I != E; ++I)
10627 EndsWithSequential = !isa<StructType>(*I);
10629 // Can we combine the two pointer arithmetics offsets?
10630 if (EndsWithSequential) {
10631 // Replace: gep (gep %P, long B), long A, ...
10632 // With: T = long A+B; gep %P, T, ...
10634 Value *Sum, *SO1 = SrcGEPOperands.back(), *GO1 = GEP.getOperand(1);
10635 if (SO1 == Constant::getNullValue(SO1->getType())) {
10637 } else if (GO1 == Constant::getNullValue(GO1->getType())) {
10640 // If they aren't the same type, convert both to an integer of the
10641 // target's pointer size.
10642 if (SO1->getType() != GO1->getType()) {
10643 if (Constant *SO1C = dyn_cast<Constant>(SO1)) {
10644 SO1 = ConstantExpr::getIntegerCast(SO1C, GO1->getType(), true);
10645 } else if (Constant *GO1C = dyn_cast<Constant>(GO1)) {
10646 GO1 = ConstantExpr::getIntegerCast(GO1C, SO1->getType(), true);
10648 unsigned PS = TD->getPointerSizeInBits();
10649 if (TD->getTypeSizeInBits(SO1->getType()) == PS) {
10650 // Convert GO1 to SO1's type.
10651 GO1 = InsertCastToIntPtrTy(GO1, SO1->getType(), &GEP, this);
10653 } else if (TD->getTypeSizeInBits(GO1->getType()) == PS) {
10654 // Convert SO1 to GO1's type.
10655 SO1 = InsertCastToIntPtrTy(SO1, GO1->getType(), &GEP, this);
10657 const Type *PT = TD->getIntPtrType();
10658 SO1 = InsertCastToIntPtrTy(SO1, PT, &GEP, this);
10659 GO1 = InsertCastToIntPtrTy(GO1, PT, &GEP, this);
10663 if (isa<Constant>(SO1) && isa<Constant>(GO1))
10664 Sum = ConstantExpr::getAdd(cast<Constant>(SO1), cast<Constant>(GO1));
10666 Sum = BinaryOperator::CreateAdd(SO1, GO1, PtrOp->getName()+".sum");
10667 InsertNewInstBefore(cast<Instruction>(Sum), GEP);
10671 // Recycle the GEP we already have if possible.
10672 if (SrcGEPOperands.size() == 2) {
10673 GEP.setOperand(0, SrcGEPOperands[0]);
10674 GEP.setOperand(1, Sum);
10677 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
10678 SrcGEPOperands.end()-1);
10679 Indices.push_back(Sum);
10680 Indices.insert(Indices.end(), GEP.op_begin()+2, GEP.op_end());
10682 } else if (isa<Constant>(*GEP.idx_begin()) &&
10683 cast<Constant>(*GEP.idx_begin())->isNullValue() &&
10684 SrcGEPOperands.size() != 1) {
10685 // Otherwise we can do the fold if the first index of the GEP is a zero
10686 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
10687 SrcGEPOperands.end());
10688 Indices.insert(Indices.end(), GEP.idx_begin()+1, GEP.idx_end());
10691 if (!Indices.empty())
10692 return GetElementPtrInst::Create(SrcGEPOperands[0], Indices.begin(),
10693 Indices.end(), GEP.getName());
10695 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(PtrOp)) {
10696 // GEP of global variable. If all of the indices for this GEP are
10697 // constants, we can promote this to a constexpr instead of an instruction.
10699 // Scan for nonconstants...
10700 SmallVector<Constant*, 8> Indices;
10701 User::op_iterator I = GEP.idx_begin(), E = GEP.idx_end();
10702 for (; I != E && isa<Constant>(*I); ++I)
10703 Indices.push_back(cast<Constant>(*I));
10705 if (I == E) { // If they are all constants...
10706 Constant *CE = ConstantExpr::getGetElementPtr(GV,
10707 &Indices[0],Indices.size());
10709 // Replace all uses of the GEP with the new constexpr...
10710 return ReplaceInstUsesWith(GEP, CE);
10712 } else if (Value *X = getBitCastOperand(PtrOp)) { // Is the operand a cast?
10713 if (!isa<PointerType>(X->getType())) {
10714 // Not interesting. Source pointer must be a cast from pointer.
10715 } else if (HasZeroPointerIndex) {
10716 // transform: GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ...
10717 // into : GEP [10 x i8]* X, i32 0, ...
10719 // This occurs when the program declares an array extern like "int X[];"
10721 const PointerType *CPTy = cast<PointerType>(PtrOp->getType());
10722 const PointerType *XTy = cast<PointerType>(X->getType());
10723 if (const ArrayType *XATy =
10724 dyn_cast<ArrayType>(XTy->getElementType()))
10725 if (const ArrayType *CATy =
10726 dyn_cast<ArrayType>(CPTy->getElementType()))
10727 if (CATy->getElementType() == XATy->getElementType()) {
10728 // At this point, we know that the cast source type is a pointer
10729 // to an array of the same type as the destination pointer
10730 // array. Because the array type is never stepped over (there
10731 // is a leading zero) we can fold the cast into this GEP.
10732 GEP.setOperand(0, X);
10735 } else if (GEP.getNumOperands() == 2) {
10736 // Transform things like:
10737 // %t = getelementptr i32* bitcast ([2 x i32]* %str to i32*), i32 %V
10738 // into: %t1 = getelementptr [2 x i32]* %str, i32 0, i32 %V; bitcast
10739 const Type *SrcElTy = cast<PointerType>(X->getType())->getElementType();
10740 const Type *ResElTy=cast<PointerType>(PtrOp->getType())->getElementType();
10741 if (isa<ArrayType>(SrcElTy) &&
10742 TD->getTypePaddedSize(cast<ArrayType>(SrcElTy)->getElementType()) ==
10743 TD->getTypePaddedSize(ResElTy)) {
10745 Idx[0] = Constant::getNullValue(Type::Int32Ty);
10746 Idx[1] = GEP.getOperand(1);
10747 Value *V = InsertNewInstBefore(
10748 GetElementPtrInst::Create(X, Idx, Idx + 2, GEP.getName()), GEP);
10749 // V and GEP are both pointer types --> BitCast
10750 return new BitCastInst(V, GEP.getType());
10753 // Transform things like:
10754 // getelementptr i8* bitcast ([100 x double]* X to i8*), i32 %tmp
10755 // (where tmp = 8*tmp2) into:
10756 // getelementptr [100 x double]* %arr, i32 0, i32 %tmp2; bitcast
10758 if (isa<ArrayType>(SrcElTy) && ResElTy == Type::Int8Ty) {
10759 uint64_t ArrayEltSize =
10760 TD->getTypePaddedSize(cast<ArrayType>(SrcElTy)->getElementType());
10762 // Check to see if "tmp" is a scale by a multiple of ArrayEltSize. We
10763 // allow either a mul, shift, or constant here.
10765 ConstantInt *Scale = 0;
10766 if (ArrayEltSize == 1) {
10767 NewIdx = GEP.getOperand(1);
10768 Scale = ConstantInt::get(NewIdx->getType(), 1);
10769 } else if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP.getOperand(1))) {
10770 NewIdx = ConstantInt::get(CI->getType(), 1);
10772 } else if (Instruction *Inst =dyn_cast<Instruction>(GEP.getOperand(1))){
10773 if (Inst->getOpcode() == Instruction::Shl &&
10774 isa<ConstantInt>(Inst->getOperand(1))) {
10775 ConstantInt *ShAmt = cast<ConstantInt>(Inst->getOperand(1));
10776 uint32_t ShAmtVal = ShAmt->getLimitedValue(64);
10777 Scale = ConstantInt::get(Inst->getType(), 1ULL << ShAmtVal);
10778 NewIdx = Inst->getOperand(0);
10779 } else if (Inst->getOpcode() == Instruction::Mul &&
10780 isa<ConstantInt>(Inst->getOperand(1))) {
10781 Scale = cast<ConstantInt>(Inst->getOperand(1));
10782 NewIdx = Inst->getOperand(0);
10786 // If the index will be to exactly the right offset with the scale taken
10787 // out, perform the transformation. Note, we don't know whether Scale is
10788 // signed or not. We'll use unsigned version of division/modulo
10789 // operation after making sure Scale doesn't have the sign bit set.
10790 if (Scale && Scale->getSExtValue() >= 0LL &&
10791 Scale->getZExtValue() % ArrayEltSize == 0) {
10792 Scale = ConstantInt::get(Scale->getType(),
10793 Scale->getZExtValue() / ArrayEltSize);
10794 if (Scale->getZExtValue() != 1) {
10795 Constant *C = ConstantExpr::getIntegerCast(Scale, NewIdx->getType(),
10797 Instruction *Sc = BinaryOperator::CreateMul(NewIdx, C, "idxscale");
10798 NewIdx = InsertNewInstBefore(Sc, GEP);
10801 // Insert the new GEP instruction.
10803 Idx[0] = Constant::getNullValue(Type::Int32Ty);
10805 Instruction *NewGEP =
10806 GetElementPtrInst::Create(X, Idx, Idx + 2, GEP.getName());
10807 NewGEP = InsertNewInstBefore(NewGEP, GEP);
10808 // The NewGEP must be pointer typed, so must the old one -> BitCast
10809 return new BitCastInst(NewGEP, GEP.getType());
10815 /// See if we can simplify:
10816 /// X = bitcast A to B*
10817 /// Y = gep X, <...constant indices...>
10818 /// into a gep of the original struct. This is important for SROA and alias
10819 /// analysis of unions. If "A" is also a bitcast, wait for A/X to be merged.
10820 if (BitCastInst *BCI = dyn_cast<BitCastInst>(PtrOp)) {
10821 if (!isa<BitCastInst>(BCI->getOperand(0)) && GEP.hasAllConstantIndices()) {
10822 // Determine how much the GEP moves the pointer. We are guaranteed to get
10823 // a constant back from EmitGEPOffset.
10824 ConstantInt *OffsetV = cast<ConstantInt>(EmitGEPOffset(&GEP, GEP, *this));
10825 int64_t Offset = OffsetV->getSExtValue();
10827 // If this GEP instruction doesn't move the pointer, just replace the GEP
10828 // with a bitcast of the real input to the dest type.
10830 // If the bitcast is of an allocation, and the allocation will be
10831 // converted to match the type of the cast, don't touch this.
10832 if (isa<AllocationInst>(BCI->getOperand(0))) {
10833 // See if the bitcast simplifies, if so, don't nuke this GEP yet.
10834 if (Instruction *I = visitBitCast(*BCI)) {
10837 BCI->getParent()->getInstList().insert(BCI, I);
10838 ReplaceInstUsesWith(*BCI, I);
10843 return new BitCastInst(BCI->getOperand(0), GEP.getType());
10846 // Otherwise, if the offset is non-zero, we need to find out if there is a
10847 // field at Offset in 'A's type. If so, we can pull the cast through the
10849 SmallVector<Value*, 8> NewIndices;
10851 cast<PointerType>(BCI->getOperand(0)->getType())->getElementType();
10852 if (FindElementAtOffset(InTy, Offset, NewIndices, TD)) {
10853 Instruction *NGEP =
10854 GetElementPtrInst::Create(BCI->getOperand(0), NewIndices.begin(),
10856 if (NGEP->getType() == GEP.getType()) return NGEP;
10857 InsertNewInstBefore(NGEP, GEP);
10858 NGEP->takeName(&GEP);
10859 return new BitCastInst(NGEP, GEP.getType());
10867 Instruction *InstCombiner::visitAllocationInst(AllocationInst &AI) {
10868 // Convert: malloc Ty, C - where C is a constant != 1 into: malloc [C x Ty], 1
10869 if (AI.isArrayAllocation()) { // Check C != 1
10870 if (const ConstantInt *C = dyn_cast<ConstantInt>(AI.getArraySize())) {
10871 const Type *NewTy =
10872 ArrayType::get(AI.getAllocatedType(), C->getZExtValue());
10873 AllocationInst *New = 0;
10875 // Create and insert the replacement instruction...
10876 if (isa<MallocInst>(AI))
10877 New = new MallocInst(NewTy, 0, AI.getAlignment(), AI.getName());
10879 assert(isa<AllocaInst>(AI) && "Unknown type of allocation inst!");
10880 New = new AllocaInst(NewTy, 0, AI.getAlignment(), AI.getName());
10883 InsertNewInstBefore(New, AI);
10885 // Scan to the end of the allocation instructions, to skip over a block of
10886 // allocas if possible...
10888 BasicBlock::iterator It = New;
10889 while (isa<AllocationInst>(*It)) ++It;
10891 // Now that I is pointing to the first non-allocation-inst in the block,
10892 // insert our getelementptr instruction...
10894 Value *NullIdx = Constant::getNullValue(Type::Int32Ty);
10898 Value *V = GetElementPtrInst::Create(New, Idx, Idx + 2,
10899 New->getName()+".sub", It);
10901 // Now make everything use the getelementptr instead of the original
10903 return ReplaceInstUsesWith(AI, V);
10904 } else if (isa<UndefValue>(AI.getArraySize())) {
10905 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
10909 if (isa<AllocaInst>(AI) && AI.getAllocatedType()->isSized()) {
10910 // If alloca'ing a zero byte object, replace the alloca with a null pointer.
10911 // Note that we only do this for alloca's, because malloc should allocate and
10912 // return a unique pointer, even for a zero byte allocation.
10913 if (TD->getTypePaddedSize(AI.getAllocatedType()) == 0)
10914 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
10916 // If the alignment is 0 (unspecified), assign it the preferred alignment.
10917 if (AI.getAlignment() == 0)
10918 AI.setAlignment(TD->getPrefTypeAlignment(AI.getAllocatedType()));
10924 Instruction *InstCombiner::visitFreeInst(FreeInst &FI) {
10925 Value *Op = FI.getOperand(0);
10927 // free undef -> unreachable.
10928 if (isa<UndefValue>(Op)) {
10929 // Insert a new store to null because we cannot modify the CFG here.
10930 new StoreInst(ConstantInt::getTrue(),
10931 UndefValue::get(PointerType::getUnqual(Type::Int1Ty)), &FI);
10932 return EraseInstFromFunction(FI);
10935 // If we have 'free null' delete the instruction. This can happen in stl code
10936 // when lots of inlining happens.
10937 if (isa<ConstantPointerNull>(Op))
10938 return EraseInstFromFunction(FI);
10940 // Change free <ty>* (cast <ty2>* X to <ty>*) into free <ty2>* X
10941 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op)) {
10942 FI.setOperand(0, CI->getOperand(0));
10946 // Change free (gep X, 0,0,0,0) into free(X)
10947 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op)) {
10948 if (GEPI->hasAllZeroIndices()) {
10949 AddToWorkList(GEPI);
10950 FI.setOperand(0, GEPI->getOperand(0));
10955 // Change free(malloc) into nothing, if the malloc has a single use.
10956 if (MallocInst *MI = dyn_cast<MallocInst>(Op))
10957 if (MI->hasOneUse()) {
10958 EraseInstFromFunction(FI);
10959 return EraseInstFromFunction(*MI);
10966 /// InstCombineLoadCast - Fold 'load (cast P)' -> cast (load P)' when possible.
10967 static Instruction *InstCombineLoadCast(InstCombiner &IC, LoadInst &LI,
10968 const TargetData *TD) {
10969 User *CI = cast<User>(LI.getOperand(0));
10970 Value *CastOp = CI->getOperand(0);
10972 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(CI)) {
10973 // Instead of loading constant c string, use corresponding integer value
10974 // directly if string length is small enough.
10976 if (GetConstantStringInfo(CE->getOperand(0), Str) && !Str.empty()) {
10977 unsigned len = Str.length();
10978 const Type *Ty = cast<PointerType>(CE->getType())->getElementType();
10979 unsigned numBits = Ty->getPrimitiveSizeInBits();
10980 // Replace LI with immediate integer store.
10981 if ((numBits >> 3) == len + 1) {
10982 APInt StrVal(numBits, 0);
10983 APInt SingleChar(numBits, 0);
10984 if (TD->isLittleEndian()) {
10985 for (signed i = len-1; i >= 0; i--) {
10986 SingleChar = (uint64_t) Str[i];
10987 StrVal = (StrVal << 8) | SingleChar;
10990 for (unsigned i = 0; i < len; i++) {
10991 SingleChar = (uint64_t) Str[i];
10992 StrVal = (StrVal << 8) | SingleChar;
10994 // Append NULL at the end.
10996 StrVal = (StrVal << 8) | SingleChar;
10998 Value *NL = ConstantInt::get(StrVal);
10999 return IC.ReplaceInstUsesWith(LI, NL);
11004 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
11005 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
11006 const Type *SrcPTy = SrcTy->getElementType();
11008 if (DestPTy->isInteger() || isa<PointerType>(DestPTy) ||
11009 isa<VectorType>(DestPTy)) {
11010 // If the source is an array, the code below will not succeed. Check to
11011 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
11013 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
11014 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
11015 if (ASrcTy->getNumElements() != 0) {
11017 Idxs[0] = Idxs[1] = Constant::getNullValue(Type::Int32Ty);
11018 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs, 2);
11019 SrcTy = cast<PointerType>(CastOp->getType());
11020 SrcPTy = SrcTy->getElementType();
11023 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy) ||
11024 isa<VectorType>(SrcPTy)) &&
11025 // Do not allow turning this into a load of an integer, which is then
11026 // casted to a pointer, this pessimizes pointer analysis a lot.
11027 (isa<PointerType>(SrcPTy) == isa<PointerType>(LI.getType())) &&
11028 IC.getTargetData().getTypeSizeInBits(SrcPTy) ==
11029 IC.getTargetData().getTypeSizeInBits(DestPTy)) {
11031 // Okay, we are casting from one integer or pointer type to another of
11032 // the same size. Instead of casting the pointer before the load, cast
11033 // the result of the loaded value.
11034 Value *NewLoad = IC.InsertNewInstBefore(new LoadInst(CastOp,
11036 LI.isVolatile()),LI);
11037 // Now cast the result of the load.
11038 return new BitCastInst(NewLoad, LI.getType());
11045 /// isSafeToLoadUnconditionally - Return true if we know that executing a load
11046 /// from this value cannot trap. If it is not obviously safe to load from the
11047 /// specified pointer, we do a quick local scan of the basic block containing
11048 /// ScanFrom, to determine if the address is already accessed.
11049 static bool isSafeToLoadUnconditionally(Value *V, Instruction *ScanFrom) {
11050 // If it is an alloca it is always safe to load from.
11051 if (isa<AllocaInst>(V)) return true;
11053 // If it is a global variable it is mostly safe to load from.
11054 if (const GlobalValue *GV = dyn_cast<GlobalVariable>(V))
11055 // Don't try to evaluate aliases. External weak GV can be null.
11056 return !isa<GlobalAlias>(GV) && !GV->hasExternalWeakLinkage();
11058 // Otherwise, be a little bit agressive by scanning the local block where we
11059 // want to check to see if the pointer is already being loaded or stored
11060 // from/to. If so, the previous load or store would have already trapped,
11061 // so there is no harm doing an extra load (also, CSE will later eliminate
11062 // the load entirely).
11063 BasicBlock::iterator BBI = ScanFrom, E = ScanFrom->getParent()->begin();
11068 // If we see a free or a call (which might do a free) the pointer could be
11070 if (isa<FreeInst>(BBI) || isa<CallInst>(BBI))
11073 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
11074 if (LI->getOperand(0) == V) return true;
11075 } else if (StoreInst *SI = dyn_cast<StoreInst>(BBI)) {
11076 if (SI->getOperand(1) == V) return true;
11083 Instruction *InstCombiner::visitLoadInst(LoadInst &LI) {
11084 Value *Op = LI.getOperand(0);
11086 // Attempt to improve the alignment.
11087 unsigned KnownAlign = GetOrEnforceKnownAlignment(Op);
11089 (LI.getAlignment() == 0 ? TD->getABITypeAlignment(LI.getType()) :
11090 LI.getAlignment()))
11091 LI.setAlignment(KnownAlign);
11093 // load (cast X) --> cast (load X) iff safe
11094 if (isa<CastInst>(Op))
11095 if (Instruction *Res = InstCombineLoadCast(*this, LI, TD))
11098 // None of the following transforms are legal for volatile loads.
11099 if (LI.isVolatile()) return 0;
11101 // Do really simple store-to-load forwarding and load CSE, to catch cases
11102 // where there are several consequtive memory accesses to the same location,
11103 // separated by a few arithmetic operations.
11104 BasicBlock::iterator BBI = &LI;
11105 if (Value *AvailableVal = FindAvailableLoadedValue(Op, LI.getParent(), BBI,6))
11106 return ReplaceInstUsesWith(LI, AvailableVal);
11108 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op)) {
11109 const Value *GEPI0 = GEPI->getOperand(0);
11110 // TODO: Consider a target hook for valid address spaces for this xform.
11111 if (isa<ConstantPointerNull>(GEPI0) &&
11112 cast<PointerType>(GEPI0->getType())->getAddressSpace() == 0) {
11113 // Insert a new store to null instruction before the load to indicate
11114 // that this code is not reachable. We do this instead of inserting
11115 // an unreachable instruction directly because we cannot modify the
11117 new StoreInst(UndefValue::get(LI.getType()),
11118 Constant::getNullValue(Op->getType()), &LI);
11119 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
11123 if (Constant *C = dyn_cast<Constant>(Op)) {
11124 // load null/undef -> undef
11125 // TODO: Consider a target hook for valid address spaces for this xform.
11126 if (isa<UndefValue>(C) || (C->isNullValue() &&
11127 cast<PointerType>(Op->getType())->getAddressSpace() == 0)) {
11128 // Insert a new store to null instruction before the load to indicate that
11129 // this code is not reachable. We do this instead of inserting an
11130 // unreachable instruction directly because we cannot modify the CFG.
11131 new StoreInst(UndefValue::get(LI.getType()),
11132 Constant::getNullValue(Op->getType()), &LI);
11133 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
11136 // Instcombine load (constant global) into the value loaded.
11137 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Op))
11138 if (GV->isConstant() && !GV->isDeclaration())
11139 return ReplaceInstUsesWith(LI, GV->getInitializer());
11141 // Instcombine load (constantexpr_GEP global, 0, ...) into the value loaded.
11142 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Op)) {
11143 if (CE->getOpcode() == Instruction::GetElementPtr) {
11144 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(CE->getOperand(0)))
11145 if (GV->isConstant() && !GV->isDeclaration())
11147 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE))
11148 return ReplaceInstUsesWith(LI, V);
11149 if (CE->getOperand(0)->isNullValue()) {
11150 // Insert a new store to null instruction before the load to indicate
11151 // that this code is not reachable. We do this instead of inserting
11152 // an unreachable instruction directly because we cannot modify the
11154 new StoreInst(UndefValue::get(LI.getType()),
11155 Constant::getNullValue(Op->getType()), &LI);
11156 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
11159 } else if (CE->isCast()) {
11160 if (Instruction *Res = InstCombineLoadCast(*this, LI, TD))
11166 // If this load comes from anywhere in a constant global, and if the global
11167 // is all undef or zero, we know what it loads.
11168 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Op->getUnderlyingObject())){
11169 if (GV->isConstant() && GV->hasInitializer()) {
11170 if (GV->getInitializer()->isNullValue())
11171 return ReplaceInstUsesWith(LI, Constant::getNullValue(LI.getType()));
11172 else if (isa<UndefValue>(GV->getInitializer()))
11173 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
11177 if (Op->hasOneUse()) {
11178 // Change select and PHI nodes to select values instead of addresses: this
11179 // helps alias analysis out a lot, allows many others simplifications, and
11180 // exposes redundancy in the code.
11182 // Note that we cannot do the transformation unless we know that the
11183 // introduced loads cannot trap! Something like this is valid as long as
11184 // the condition is always false: load (select bool %C, int* null, int* %G),
11185 // but it would not be valid if we transformed it to load from null
11186 // unconditionally.
11188 if (SelectInst *SI = dyn_cast<SelectInst>(Op)) {
11189 // load (select (Cond, &V1, &V2)) --> select(Cond, load &V1, load &V2).
11190 if (isSafeToLoadUnconditionally(SI->getOperand(1), SI) &&
11191 isSafeToLoadUnconditionally(SI->getOperand(2), SI)) {
11192 Value *V1 = InsertNewInstBefore(new LoadInst(SI->getOperand(1),
11193 SI->getOperand(1)->getName()+".val"), LI);
11194 Value *V2 = InsertNewInstBefore(new LoadInst(SI->getOperand(2),
11195 SI->getOperand(2)->getName()+".val"), LI);
11196 return SelectInst::Create(SI->getCondition(), V1, V2);
11199 // load (select (cond, null, P)) -> load P
11200 if (Constant *C = dyn_cast<Constant>(SI->getOperand(1)))
11201 if (C->isNullValue()) {
11202 LI.setOperand(0, SI->getOperand(2));
11206 // load (select (cond, P, null)) -> load P
11207 if (Constant *C = dyn_cast<Constant>(SI->getOperand(2)))
11208 if (C->isNullValue()) {
11209 LI.setOperand(0, SI->getOperand(1));
11217 /// InstCombineStoreToCast - Fold store V, (cast P) -> store (cast V), P
11218 /// when possible. This makes it generally easy to do alias analysis and/or
11219 /// SROA/mem2reg of the memory object.
11220 static Instruction *InstCombineStoreToCast(InstCombiner &IC, StoreInst &SI) {
11221 User *CI = cast<User>(SI.getOperand(1));
11222 Value *CastOp = CI->getOperand(0);
11224 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
11225 const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType());
11226 if (SrcTy == 0) return 0;
11228 const Type *SrcPTy = SrcTy->getElementType();
11230 if (!DestPTy->isInteger() && !isa<PointerType>(DestPTy))
11233 /// NewGEPIndices - If SrcPTy is an aggregate type, we can emit a "noop gep"
11234 /// to its first element. This allows us to handle things like:
11235 /// store i32 xxx, (bitcast {foo*, float}* %P to i32*)
11236 /// on 32-bit hosts.
11237 SmallVector<Value*, 4> NewGEPIndices;
11239 // If the source is an array, the code below will not succeed. Check to
11240 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
11242 if (isa<ArrayType>(SrcPTy) || isa<StructType>(SrcPTy)) {
11243 // Index through pointer.
11244 Constant *Zero = Constant::getNullValue(Type::Int32Ty);
11245 NewGEPIndices.push_back(Zero);
11248 if (const StructType *STy = dyn_cast<StructType>(SrcPTy)) {
11249 if (!STy->getNumElements()) /* Struct can be empty {} */
11251 NewGEPIndices.push_back(Zero);
11252 SrcPTy = STy->getElementType(0);
11253 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(SrcPTy)) {
11254 NewGEPIndices.push_back(Zero);
11255 SrcPTy = ATy->getElementType();
11261 SrcTy = PointerType::get(SrcPTy, SrcTy->getAddressSpace());
11264 if (!SrcPTy->isInteger() && !isa<PointerType>(SrcPTy))
11267 // If the pointers point into different address spaces or if they point to
11268 // values with different sizes, we can't do the transformation.
11269 if (SrcTy->getAddressSpace() !=
11270 cast<PointerType>(CI->getType())->getAddressSpace() ||
11271 IC.getTargetData().getTypeSizeInBits(SrcPTy) !=
11272 IC.getTargetData().getTypeSizeInBits(DestPTy))
11275 // Okay, we are casting from one integer or pointer type to another of
11276 // the same size. Instead of casting the pointer before
11277 // the store, cast the value to be stored.
11279 Value *SIOp0 = SI.getOperand(0);
11280 Instruction::CastOps opcode = Instruction::BitCast;
11281 const Type* CastSrcTy = SIOp0->getType();
11282 const Type* CastDstTy = SrcPTy;
11283 if (isa<PointerType>(CastDstTy)) {
11284 if (CastSrcTy->isInteger())
11285 opcode = Instruction::IntToPtr;
11286 } else if (isa<IntegerType>(CastDstTy)) {
11287 if (isa<PointerType>(SIOp0->getType()))
11288 opcode = Instruction::PtrToInt;
11291 // SIOp0 is a pointer to aggregate and this is a store to the first field,
11292 // emit a GEP to index into its first field.
11293 if (!NewGEPIndices.empty()) {
11294 if (Constant *C = dyn_cast<Constant>(CastOp))
11295 CastOp = ConstantExpr::getGetElementPtr(C, &NewGEPIndices[0],
11296 NewGEPIndices.size());
11298 CastOp = IC.InsertNewInstBefore(
11299 GetElementPtrInst::Create(CastOp, NewGEPIndices.begin(),
11300 NewGEPIndices.end()), SI);
11303 if (Constant *C = dyn_cast<Constant>(SIOp0))
11304 NewCast = ConstantExpr::getCast(opcode, C, CastDstTy);
11306 NewCast = IC.InsertNewInstBefore(
11307 CastInst::Create(opcode, SIOp0, CastDstTy, SIOp0->getName()+".c"),
11309 return new StoreInst(NewCast, CastOp);
11312 /// equivalentAddressValues - Test if A and B will obviously have the same
11313 /// value. This includes recognizing that %t0 and %t1 will have the same
11314 /// value in code like this:
11315 /// %t0 = getelementptr @a, 0, 3
11316 /// store i32 0, i32* %t0
11317 /// %t1 = getelementptr @a, 0, 3
11318 /// %t2 = load i32* %t1
11320 static bool equivalentAddressValues(Value *A, Value *B) {
11321 // Test if the values are trivially equivalent.
11322 if (A == B) return true;
11324 // Test if the values come form identical arithmetic instructions.
11325 if (isa<BinaryOperator>(A) ||
11326 isa<CastInst>(A) ||
11328 isa<GetElementPtrInst>(A))
11329 if (Instruction *BI = dyn_cast<Instruction>(B))
11330 if (cast<Instruction>(A)->isIdenticalTo(BI))
11333 // Otherwise they may not be equivalent.
11337 Instruction *InstCombiner::visitStoreInst(StoreInst &SI) {
11338 Value *Val = SI.getOperand(0);
11339 Value *Ptr = SI.getOperand(1);
11341 if (isa<UndefValue>(Ptr)) { // store X, undef -> noop (even if volatile)
11342 EraseInstFromFunction(SI);
11347 // If the RHS is an alloca with a single use, zapify the store, making the
11349 if (Ptr->hasOneUse() && !SI.isVolatile()) {
11350 if (isa<AllocaInst>(Ptr)) {
11351 EraseInstFromFunction(SI);
11356 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr))
11357 if (isa<AllocaInst>(GEP->getOperand(0)) &&
11358 GEP->getOperand(0)->hasOneUse()) {
11359 EraseInstFromFunction(SI);
11365 // Attempt to improve the alignment.
11366 unsigned KnownAlign = GetOrEnforceKnownAlignment(Ptr);
11368 (SI.getAlignment() == 0 ? TD->getABITypeAlignment(Val->getType()) :
11369 SI.getAlignment()))
11370 SI.setAlignment(KnownAlign);
11372 // Do really simple DSE, to catch cases where there are several consequtive
11373 // stores to the same location, separated by a few arithmetic operations. This
11374 // situation often occurs with bitfield accesses.
11375 BasicBlock::iterator BBI = &SI;
11376 for (unsigned ScanInsts = 6; BBI != SI.getParent()->begin() && ScanInsts;
11380 if (StoreInst *PrevSI = dyn_cast<StoreInst>(BBI)) {
11381 // Prev store isn't volatile, and stores to the same location?
11382 if (!PrevSI->isVolatile() &&equivalentAddressValues(PrevSI->getOperand(1),
11383 SI.getOperand(1))) {
11386 EraseInstFromFunction(*PrevSI);
11392 // If this is a load, we have to stop. However, if the loaded value is from
11393 // the pointer we're loading and is producing the pointer we're storing,
11394 // then *this* store is dead (X = load P; store X -> P).
11395 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
11396 if (LI == Val && equivalentAddressValues(LI->getOperand(0), Ptr) &&
11397 !SI.isVolatile()) {
11398 EraseInstFromFunction(SI);
11402 // Otherwise, this is a load from some other location. Stores before it
11403 // may not be dead.
11407 // Don't skip over loads or things that can modify memory.
11408 if (BBI->mayWriteToMemory() || BBI->mayReadFromMemory())
11413 if (SI.isVolatile()) return 0; // Don't hack volatile stores.
11415 // store X, null -> turns into 'unreachable' in SimplifyCFG
11416 if (isa<ConstantPointerNull>(Ptr)) {
11417 if (!isa<UndefValue>(Val)) {
11418 SI.setOperand(0, UndefValue::get(Val->getType()));
11419 if (Instruction *U = dyn_cast<Instruction>(Val))
11420 AddToWorkList(U); // Dropped a use.
11423 return 0; // Do not modify these!
11426 // store undef, Ptr -> noop
11427 if (isa<UndefValue>(Val)) {
11428 EraseInstFromFunction(SI);
11433 // If the pointer destination is a cast, see if we can fold the cast into the
11435 if (isa<CastInst>(Ptr))
11436 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
11438 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
11440 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
11444 // If this store is the last instruction in the basic block, and if the block
11445 // ends with an unconditional branch, try to move it to the successor block.
11447 if (BranchInst *BI = dyn_cast<BranchInst>(BBI))
11448 if (BI->isUnconditional())
11449 if (SimplifyStoreAtEndOfBlock(SI))
11450 return 0; // xform done!
11455 /// SimplifyStoreAtEndOfBlock - Turn things like:
11456 /// if () { *P = v1; } else { *P = v2 }
11457 /// into a phi node with a store in the successor.
11459 /// Simplify things like:
11460 /// *P = v1; if () { *P = v2; }
11461 /// into a phi node with a store in the successor.
11463 bool InstCombiner::SimplifyStoreAtEndOfBlock(StoreInst &SI) {
11464 BasicBlock *StoreBB = SI.getParent();
11466 // Check to see if the successor block has exactly two incoming edges. If
11467 // so, see if the other predecessor contains a store to the same location.
11468 // if so, insert a PHI node (if needed) and move the stores down.
11469 BasicBlock *DestBB = StoreBB->getTerminator()->getSuccessor(0);
11471 // Determine whether Dest has exactly two predecessors and, if so, compute
11472 // the other predecessor.
11473 pred_iterator PI = pred_begin(DestBB);
11474 BasicBlock *OtherBB = 0;
11475 if (*PI != StoreBB)
11478 if (PI == pred_end(DestBB))
11481 if (*PI != StoreBB) {
11486 if (++PI != pred_end(DestBB))
11489 // Bail out if all the relevant blocks aren't distinct (this can happen,
11490 // for example, if SI is in an infinite loop)
11491 if (StoreBB == DestBB || OtherBB == DestBB)
11494 // Verify that the other block ends in a branch and is not otherwise empty.
11495 BasicBlock::iterator BBI = OtherBB->getTerminator();
11496 BranchInst *OtherBr = dyn_cast<BranchInst>(BBI);
11497 if (!OtherBr || BBI == OtherBB->begin())
11500 // If the other block ends in an unconditional branch, check for the 'if then
11501 // else' case. there is an instruction before the branch.
11502 StoreInst *OtherStore = 0;
11503 if (OtherBr->isUnconditional()) {
11504 // If this isn't a store, or isn't a store to the same location, bail out.
11506 OtherStore = dyn_cast<StoreInst>(BBI);
11507 if (!OtherStore || OtherStore->getOperand(1) != SI.getOperand(1))
11510 // Otherwise, the other block ended with a conditional branch. If one of the
11511 // destinations is StoreBB, then we have the if/then case.
11512 if (OtherBr->getSuccessor(0) != StoreBB &&
11513 OtherBr->getSuccessor(1) != StoreBB)
11516 // Okay, we know that OtherBr now goes to Dest and StoreBB, so this is an
11517 // if/then triangle. See if there is a store to the same ptr as SI that
11518 // lives in OtherBB.
11520 // Check to see if we find the matching store.
11521 if ((OtherStore = dyn_cast<StoreInst>(BBI))) {
11522 if (OtherStore->getOperand(1) != SI.getOperand(1))
11526 // If we find something that may be using or overwriting the stored
11527 // value, or if we run out of instructions, we can't do the xform.
11528 if (BBI->mayReadFromMemory() || BBI->mayWriteToMemory() ||
11529 BBI == OtherBB->begin())
11533 // In order to eliminate the store in OtherBr, we have to
11534 // make sure nothing reads or overwrites the stored value in
11536 for (BasicBlock::iterator I = StoreBB->begin(); &*I != &SI; ++I) {
11537 // FIXME: This should really be AA driven.
11538 if (I->mayReadFromMemory() || I->mayWriteToMemory())
11543 // Insert a PHI node now if we need it.
11544 Value *MergedVal = OtherStore->getOperand(0);
11545 if (MergedVal != SI.getOperand(0)) {
11546 PHINode *PN = PHINode::Create(MergedVal->getType(), "storemerge");
11547 PN->reserveOperandSpace(2);
11548 PN->addIncoming(SI.getOperand(0), SI.getParent());
11549 PN->addIncoming(OtherStore->getOperand(0), OtherBB);
11550 MergedVal = InsertNewInstBefore(PN, DestBB->front());
11553 // Advance to a place where it is safe to insert the new store and
11555 BBI = DestBB->getFirstNonPHI();
11556 InsertNewInstBefore(new StoreInst(MergedVal, SI.getOperand(1),
11557 OtherStore->isVolatile()), *BBI);
11559 // Nuke the old stores.
11560 EraseInstFromFunction(SI);
11561 EraseInstFromFunction(*OtherStore);
11567 Instruction *InstCombiner::visitBranchInst(BranchInst &BI) {
11568 // Change br (not X), label True, label False to: br X, label False, True
11570 BasicBlock *TrueDest;
11571 BasicBlock *FalseDest;
11572 if (match(&BI, m_Br(m_Not(m_Value(X)), TrueDest, FalseDest)) &&
11573 !isa<Constant>(X)) {
11574 // Swap Destinations and condition...
11575 BI.setCondition(X);
11576 BI.setSuccessor(0, FalseDest);
11577 BI.setSuccessor(1, TrueDest);
11581 // Cannonicalize fcmp_one -> fcmp_oeq
11582 FCmpInst::Predicate FPred; Value *Y;
11583 if (match(&BI, m_Br(m_FCmp(FPred, m_Value(X), m_Value(Y)),
11584 TrueDest, FalseDest)))
11585 if ((FPred == FCmpInst::FCMP_ONE || FPred == FCmpInst::FCMP_OLE ||
11586 FPred == FCmpInst::FCMP_OGE) && BI.getCondition()->hasOneUse()) {
11587 FCmpInst *I = cast<FCmpInst>(BI.getCondition());
11588 FCmpInst::Predicate NewPred = FCmpInst::getInversePredicate(FPred);
11589 Instruction *NewSCC = new FCmpInst(NewPred, X, Y, "", I);
11590 NewSCC->takeName(I);
11591 // Swap Destinations and condition...
11592 BI.setCondition(NewSCC);
11593 BI.setSuccessor(0, FalseDest);
11594 BI.setSuccessor(1, TrueDest);
11595 RemoveFromWorkList(I);
11596 I->eraseFromParent();
11597 AddToWorkList(NewSCC);
11601 // Cannonicalize icmp_ne -> icmp_eq
11602 ICmpInst::Predicate IPred;
11603 if (match(&BI, m_Br(m_ICmp(IPred, m_Value(X), m_Value(Y)),
11604 TrueDest, FalseDest)))
11605 if ((IPred == ICmpInst::ICMP_NE || IPred == ICmpInst::ICMP_ULE ||
11606 IPred == ICmpInst::ICMP_SLE || IPred == ICmpInst::ICMP_UGE ||
11607 IPred == ICmpInst::ICMP_SGE) && BI.getCondition()->hasOneUse()) {
11608 ICmpInst *I = cast<ICmpInst>(BI.getCondition());
11609 ICmpInst::Predicate NewPred = ICmpInst::getInversePredicate(IPred);
11610 Instruction *NewSCC = new ICmpInst(NewPred, X, Y, "", I);
11611 NewSCC->takeName(I);
11612 // Swap Destinations and condition...
11613 BI.setCondition(NewSCC);
11614 BI.setSuccessor(0, FalseDest);
11615 BI.setSuccessor(1, TrueDest);
11616 RemoveFromWorkList(I);
11617 I->eraseFromParent();;
11618 AddToWorkList(NewSCC);
11625 Instruction *InstCombiner::visitSwitchInst(SwitchInst &SI) {
11626 Value *Cond = SI.getCondition();
11627 if (Instruction *I = dyn_cast<Instruction>(Cond)) {
11628 if (I->getOpcode() == Instruction::Add)
11629 if (ConstantInt *AddRHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
11630 // change 'switch (X+4) case 1:' into 'switch (X) case -3'
11631 for (unsigned i = 2, e = SI.getNumOperands(); i != e; i += 2)
11632 SI.setOperand(i,ConstantExpr::getSub(cast<Constant>(SI.getOperand(i)),
11634 SI.setOperand(0, I->getOperand(0));
11642 Instruction *InstCombiner::visitExtractValueInst(ExtractValueInst &EV) {
11643 Value *Agg = EV.getAggregateOperand();
11645 if (!EV.hasIndices())
11646 return ReplaceInstUsesWith(EV, Agg);
11648 if (Constant *C = dyn_cast<Constant>(Agg)) {
11649 if (isa<UndefValue>(C))
11650 return ReplaceInstUsesWith(EV, UndefValue::get(EV.getType()));
11652 if (isa<ConstantAggregateZero>(C))
11653 return ReplaceInstUsesWith(EV, Constant::getNullValue(EV.getType()));
11655 if (isa<ConstantArray>(C) || isa<ConstantStruct>(C)) {
11656 // Extract the element indexed by the first index out of the constant
11657 Value *V = C->getOperand(*EV.idx_begin());
11658 if (EV.getNumIndices() > 1)
11659 // Extract the remaining indices out of the constant indexed by the
11661 return ExtractValueInst::Create(V, EV.idx_begin() + 1, EV.idx_end());
11663 return ReplaceInstUsesWith(EV, V);
11665 return 0; // Can't handle other constants
11667 if (InsertValueInst *IV = dyn_cast<InsertValueInst>(Agg)) {
11668 // We're extracting from an insertvalue instruction, compare the indices
11669 const unsigned *exti, *exte, *insi, *inse;
11670 for (exti = EV.idx_begin(), insi = IV->idx_begin(),
11671 exte = EV.idx_end(), inse = IV->idx_end();
11672 exti != exte && insi != inse;
11674 if (*insi != *exti)
11675 // The insert and extract both reference distinctly different elements.
11676 // This means the extract is not influenced by the insert, and we can
11677 // replace the aggregate operand of the extract with the aggregate
11678 // operand of the insert. i.e., replace
11679 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1
11680 // %E = extractvalue { i32, { i32 } } %I, 0
11682 // %E = extractvalue { i32, { i32 } } %A, 0
11683 return ExtractValueInst::Create(IV->getAggregateOperand(),
11684 EV.idx_begin(), EV.idx_end());
11686 if (exti == exte && insi == inse)
11687 // Both iterators are at the end: Index lists are identical. Replace
11688 // %B = insertvalue { i32, { i32 } } %A, i32 42, 1, 0
11689 // %C = extractvalue { i32, { i32 } } %B, 1, 0
11691 return ReplaceInstUsesWith(EV, IV->getInsertedValueOperand());
11692 if (exti == exte) {
11693 // The extract list is a prefix of the insert list. i.e. replace
11694 // %I = insertvalue { i32, { i32 } } %A, i32 42, 1, 0
11695 // %E = extractvalue { i32, { i32 } } %I, 1
11697 // %X = extractvalue { i32, { i32 } } %A, 1
11698 // %E = insertvalue { i32 } %X, i32 42, 0
11699 // by switching the order of the insert and extract (though the
11700 // insertvalue should be left in, since it may have other uses).
11701 Value *NewEV = InsertNewInstBefore(
11702 ExtractValueInst::Create(IV->getAggregateOperand(),
11703 EV.idx_begin(), EV.idx_end()),
11705 return InsertValueInst::Create(NewEV, IV->getInsertedValueOperand(),
11709 // The insert list is a prefix of the extract list
11710 // We can simply remove the common indices from the extract and make it
11711 // operate on the inserted value instead of the insertvalue result.
11713 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1
11714 // %E = extractvalue { i32, { i32 } } %I, 1, 0
11716 // %E extractvalue { i32 } { i32 42 }, 0
11717 return ExtractValueInst::Create(IV->getInsertedValueOperand(),
11720 // Can't simplify extracts from other values. Note that nested extracts are
11721 // already simplified implicitely by the above (extract ( extract (insert) )
11722 // will be translated into extract ( insert ( extract ) ) first and then just
11723 // the value inserted, if appropriate).
11727 /// CheapToScalarize - Return true if the value is cheaper to scalarize than it
11728 /// is to leave as a vector operation.
11729 static bool CheapToScalarize(Value *V, bool isConstant) {
11730 if (isa<ConstantAggregateZero>(V))
11732 if (ConstantVector *C = dyn_cast<ConstantVector>(V)) {
11733 if (isConstant) return true;
11734 // If all elts are the same, we can extract.
11735 Constant *Op0 = C->getOperand(0);
11736 for (unsigned i = 1; i < C->getNumOperands(); ++i)
11737 if (C->getOperand(i) != Op0)
11741 Instruction *I = dyn_cast<Instruction>(V);
11742 if (!I) return false;
11744 // Insert element gets simplified to the inserted element or is deleted if
11745 // this is constant idx extract element and its a constant idx insertelt.
11746 if (I->getOpcode() == Instruction::InsertElement && isConstant &&
11747 isa<ConstantInt>(I->getOperand(2)))
11749 if (I->getOpcode() == Instruction::Load && I->hasOneUse())
11751 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I))
11752 if (BO->hasOneUse() &&
11753 (CheapToScalarize(BO->getOperand(0), isConstant) ||
11754 CheapToScalarize(BO->getOperand(1), isConstant)))
11756 if (CmpInst *CI = dyn_cast<CmpInst>(I))
11757 if (CI->hasOneUse() &&
11758 (CheapToScalarize(CI->getOperand(0), isConstant) ||
11759 CheapToScalarize(CI->getOperand(1), isConstant)))
11765 /// Read and decode a shufflevector mask.
11767 /// It turns undef elements into values that are larger than the number of
11768 /// elements in the input.
11769 static std::vector<unsigned> getShuffleMask(const ShuffleVectorInst *SVI) {
11770 unsigned NElts = SVI->getType()->getNumElements();
11771 if (isa<ConstantAggregateZero>(SVI->getOperand(2)))
11772 return std::vector<unsigned>(NElts, 0);
11773 if (isa<UndefValue>(SVI->getOperand(2)))
11774 return std::vector<unsigned>(NElts, 2*NElts);
11776 std::vector<unsigned> Result;
11777 const ConstantVector *CP = cast<ConstantVector>(SVI->getOperand(2));
11778 for (User::const_op_iterator i = CP->op_begin(), e = CP->op_end(); i!=e; ++i)
11779 if (isa<UndefValue>(*i))
11780 Result.push_back(NElts*2); // undef -> 8
11782 Result.push_back(cast<ConstantInt>(*i)->getZExtValue());
11786 /// FindScalarElement - Given a vector and an element number, see if the scalar
11787 /// value is already around as a register, for example if it were inserted then
11788 /// extracted from the vector.
11789 static Value *FindScalarElement(Value *V, unsigned EltNo) {
11790 assert(isa<VectorType>(V->getType()) && "Not looking at a vector?");
11791 const VectorType *PTy = cast<VectorType>(V->getType());
11792 unsigned Width = PTy->getNumElements();
11793 if (EltNo >= Width) // Out of range access.
11794 return UndefValue::get(PTy->getElementType());
11796 if (isa<UndefValue>(V))
11797 return UndefValue::get(PTy->getElementType());
11798 else if (isa<ConstantAggregateZero>(V))
11799 return Constant::getNullValue(PTy->getElementType());
11800 else if (ConstantVector *CP = dyn_cast<ConstantVector>(V))
11801 return CP->getOperand(EltNo);
11802 else if (InsertElementInst *III = dyn_cast<InsertElementInst>(V)) {
11803 // If this is an insert to a variable element, we don't know what it is.
11804 if (!isa<ConstantInt>(III->getOperand(2)))
11806 unsigned IIElt = cast<ConstantInt>(III->getOperand(2))->getZExtValue();
11808 // If this is an insert to the element we are looking for, return the
11810 if (EltNo == IIElt)
11811 return III->getOperand(1);
11813 // Otherwise, the insertelement doesn't modify the value, recurse on its
11815 return FindScalarElement(III->getOperand(0), EltNo);
11816 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(V)) {
11817 unsigned LHSWidth =
11818 cast<VectorType>(SVI->getOperand(0)->getType())->getNumElements();
11819 unsigned InEl = getShuffleMask(SVI)[EltNo];
11820 if (InEl < LHSWidth)
11821 return FindScalarElement(SVI->getOperand(0), InEl);
11822 else if (InEl < LHSWidth*2)
11823 return FindScalarElement(SVI->getOperand(1), InEl - LHSWidth);
11825 return UndefValue::get(PTy->getElementType());
11828 // Otherwise, we don't know.
11832 Instruction *InstCombiner::visitExtractElementInst(ExtractElementInst &EI) {
11833 // If vector val is undef, replace extract with scalar undef.
11834 if (isa<UndefValue>(EI.getOperand(0)))
11835 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
11837 // If vector val is constant 0, replace extract with scalar 0.
11838 if (isa<ConstantAggregateZero>(EI.getOperand(0)))
11839 return ReplaceInstUsesWith(EI, Constant::getNullValue(EI.getType()));
11841 if (ConstantVector *C = dyn_cast<ConstantVector>(EI.getOperand(0))) {
11842 // If vector val is constant with all elements the same, replace EI with
11843 // that element. When the elements are not identical, we cannot replace yet
11844 // (we do that below, but only when the index is constant).
11845 Constant *op0 = C->getOperand(0);
11846 for (unsigned i = 1; i < C->getNumOperands(); ++i)
11847 if (C->getOperand(i) != op0) {
11852 return ReplaceInstUsesWith(EI, op0);
11855 // If extracting a specified index from the vector, see if we can recursively
11856 // find a previously computed scalar that was inserted into the vector.
11857 if (ConstantInt *IdxC = dyn_cast<ConstantInt>(EI.getOperand(1))) {
11858 unsigned IndexVal = IdxC->getZExtValue();
11859 unsigned VectorWidth =
11860 cast<VectorType>(EI.getOperand(0)->getType())->getNumElements();
11862 // If this is extracting an invalid index, turn this into undef, to avoid
11863 // crashing the code below.
11864 if (IndexVal >= VectorWidth)
11865 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
11867 // This instruction only demands the single element from the input vector.
11868 // If the input vector has a single use, simplify it based on this use
11870 if (EI.getOperand(0)->hasOneUse() && VectorWidth != 1) {
11871 uint64_t UndefElts;
11872 if (Value *V = SimplifyDemandedVectorElts(EI.getOperand(0),
11875 EI.setOperand(0, V);
11880 if (Value *Elt = FindScalarElement(EI.getOperand(0), IndexVal))
11881 return ReplaceInstUsesWith(EI, Elt);
11883 // If the this extractelement is directly using a bitcast from a vector of
11884 // the same number of elements, see if we can find the source element from
11885 // it. In this case, we will end up needing to bitcast the scalars.
11886 if (BitCastInst *BCI = dyn_cast<BitCastInst>(EI.getOperand(0))) {
11887 if (const VectorType *VT =
11888 dyn_cast<VectorType>(BCI->getOperand(0)->getType()))
11889 if (VT->getNumElements() == VectorWidth)
11890 if (Value *Elt = FindScalarElement(BCI->getOperand(0), IndexVal))
11891 return new BitCastInst(Elt, EI.getType());
11895 if (Instruction *I = dyn_cast<Instruction>(EI.getOperand(0))) {
11896 if (I->hasOneUse()) {
11897 // Push extractelement into predecessor operation if legal and
11898 // profitable to do so
11899 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
11900 bool isConstantElt = isa<ConstantInt>(EI.getOperand(1));
11901 if (CheapToScalarize(BO, isConstantElt)) {
11902 ExtractElementInst *newEI0 =
11903 new ExtractElementInst(BO->getOperand(0), EI.getOperand(1),
11904 EI.getName()+".lhs");
11905 ExtractElementInst *newEI1 =
11906 new ExtractElementInst(BO->getOperand(1), EI.getOperand(1),
11907 EI.getName()+".rhs");
11908 InsertNewInstBefore(newEI0, EI);
11909 InsertNewInstBefore(newEI1, EI);
11910 return BinaryOperator::Create(BO->getOpcode(), newEI0, newEI1);
11912 } else if (isa<LoadInst>(I)) {
11914 cast<PointerType>(I->getOperand(0)->getType())->getAddressSpace();
11915 Value *Ptr = InsertBitCastBefore(I->getOperand(0),
11916 PointerType::get(EI.getType(), AS),EI);
11917 GetElementPtrInst *GEP =
11918 GetElementPtrInst::Create(Ptr, EI.getOperand(1), I->getName()+".gep");
11919 InsertNewInstBefore(GEP, EI);
11920 return new LoadInst(GEP);
11923 if (InsertElementInst *IE = dyn_cast<InsertElementInst>(I)) {
11924 // Extracting the inserted element?
11925 if (IE->getOperand(2) == EI.getOperand(1))
11926 return ReplaceInstUsesWith(EI, IE->getOperand(1));
11927 // If the inserted and extracted elements are constants, they must not
11928 // be the same value, extract from the pre-inserted value instead.
11929 if (isa<Constant>(IE->getOperand(2)) &&
11930 isa<Constant>(EI.getOperand(1))) {
11931 AddUsesToWorkList(EI);
11932 EI.setOperand(0, IE->getOperand(0));
11935 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I)) {
11936 // If this is extracting an element from a shufflevector, figure out where
11937 // it came from and extract from the appropriate input element instead.
11938 if (ConstantInt *Elt = dyn_cast<ConstantInt>(EI.getOperand(1))) {
11939 unsigned SrcIdx = getShuffleMask(SVI)[Elt->getZExtValue()];
11941 unsigned LHSWidth =
11942 cast<VectorType>(SVI->getOperand(0)->getType())->getNumElements();
11944 if (SrcIdx < LHSWidth)
11945 Src = SVI->getOperand(0);
11946 else if (SrcIdx < LHSWidth*2) {
11947 SrcIdx -= LHSWidth;
11948 Src = SVI->getOperand(1);
11950 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
11952 return new ExtractElementInst(Src, SrcIdx);
11959 /// CollectSingleShuffleElements - If V is a shuffle of values that ONLY returns
11960 /// elements from either LHS or RHS, return the shuffle mask and true.
11961 /// Otherwise, return false.
11962 static bool CollectSingleShuffleElements(Value *V, Value *LHS, Value *RHS,
11963 std::vector<Constant*> &Mask) {
11964 assert(V->getType() == LHS->getType() && V->getType() == RHS->getType() &&
11965 "Invalid CollectSingleShuffleElements");
11966 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
11968 if (isa<UndefValue>(V)) {
11969 Mask.assign(NumElts, UndefValue::get(Type::Int32Ty));
11971 } else if (V == LHS) {
11972 for (unsigned i = 0; i != NumElts; ++i)
11973 Mask.push_back(ConstantInt::get(Type::Int32Ty, i));
11975 } else if (V == RHS) {
11976 for (unsigned i = 0; i != NumElts; ++i)
11977 Mask.push_back(ConstantInt::get(Type::Int32Ty, i+NumElts));
11979 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
11980 // If this is an insert of an extract from some other vector, include it.
11981 Value *VecOp = IEI->getOperand(0);
11982 Value *ScalarOp = IEI->getOperand(1);
11983 Value *IdxOp = IEI->getOperand(2);
11985 if (!isa<ConstantInt>(IdxOp))
11987 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
11989 if (isa<UndefValue>(ScalarOp)) { // inserting undef into vector.
11990 // Okay, we can handle this if the vector we are insertinting into is
11991 // transitively ok.
11992 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
11993 // If so, update the mask to reflect the inserted undef.
11994 Mask[InsertedIdx] = UndefValue::get(Type::Int32Ty);
11997 } else if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)){
11998 if (isa<ConstantInt>(EI->getOperand(1)) &&
11999 EI->getOperand(0)->getType() == V->getType()) {
12000 unsigned ExtractedIdx =
12001 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
12003 // This must be extracting from either LHS or RHS.
12004 if (EI->getOperand(0) == LHS || EI->getOperand(0) == RHS) {
12005 // Okay, we can handle this if the vector we are insertinting into is
12006 // transitively ok.
12007 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
12008 // If so, update the mask to reflect the inserted value.
12009 if (EI->getOperand(0) == LHS) {
12010 Mask[InsertedIdx % NumElts] =
12011 ConstantInt::get(Type::Int32Ty, ExtractedIdx);
12013 assert(EI->getOperand(0) == RHS);
12014 Mask[InsertedIdx % NumElts] =
12015 ConstantInt::get(Type::Int32Ty, ExtractedIdx+NumElts);
12024 // TODO: Handle shufflevector here!
12029 /// CollectShuffleElements - We are building a shuffle of V, using RHS as the
12030 /// RHS of the shuffle instruction, if it is not null. Return a shuffle mask
12031 /// that computes V and the LHS value of the shuffle.
12032 static Value *CollectShuffleElements(Value *V, std::vector<Constant*> &Mask,
12034 assert(isa<VectorType>(V->getType()) &&
12035 (RHS == 0 || V->getType() == RHS->getType()) &&
12036 "Invalid shuffle!");
12037 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
12039 if (isa<UndefValue>(V)) {
12040 Mask.assign(NumElts, UndefValue::get(Type::Int32Ty));
12042 } else if (isa<ConstantAggregateZero>(V)) {
12043 Mask.assign(NumElts, ConstantInt::get(Type::Int32Ty, 0));
12045 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
12046 // If this is an insert of an extract from some other vector, include it.
12047 Value *VecOp = IEI->getOperand(0);
12048 Value *ScalarOp = IEI->getOperand(1);
12049 Value *IdxOp = IEI->getOperand(2);
12051 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
12052 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
12053 EI->getOperand(0)->getType() == V->getType()) {
12054 unsigned ExtractedIdx =
12055 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
12056 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
12058 // Either the extracted from or inserted into vector must be RHSVec,
12059 // otherwise we'd end up with a shuffle of three inputs.
12060 if (EI->getOperand(0) == RHS || RHS == 0) {
12061 RHS = EI->getOperand(0);
12062 Value *V = CollectShuffleElements(VecOp, Mask, RHS);
12063 Mask[InsertedIdx % NumElts] =
12064 ConstantInt::get(Type::Int32Ty, NumElts+ExtractedIdx);
12068 if (VecOp == RHS) {
12069 Value *V = CollectShuffleElements(EI->getOperand(0), Mask, RHS);
12070 // Everything but the extracted element is replaced with the RHS.
12071 for (unsigned i = 0; i != NumElts; ++i) {
12072 if (i != InsertedIdx)
12073 Mask[i] = ConstantInt::get(Type::Int32Ty, NumElts+i);
12078 // If this insertelement is a chain that comes from exactly these two
12079 // vectors, return the vector and the effective shuffle.
12080 if (CollectSingleShuffleElements(IEI, EI->getOperand(0), RHS, Mask))
12081 return EI->getOperand(0);
12086 // TODO: Handle shufflevector here!
12088 // Otherwise, can't do anything fancy. Return an identity vector.
12089 for (unsigned i = 0; i != NumElts; ++i)
12090 Mask.push_back(ConstantInt::get(Type::Int32Ty, i));
12094 Instruction *InstCombiner::visitInsertElementInst(InsertElementInst &IE) {
12095 Value *VecOp = IE.getOperand(0);
12096 Value *ScalarOp = IE.getOperand(1);
12097 Value *IdxOp = IE.getOperand(2);
12099 // Inserting an undef or into an undefined place, remove this.
12100 if (isa<UndefValue>(ScalarOp) || isa<UndefValue>(IdxOp))
12101 ReplaceInstUsesWith(IE, VecOp);
12103 // If the inserted element was extracted from some other vector, and if the
12104 // indexes are constant, try to turn this into a shufflevector operation.
12105 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
12106 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
12107 EI->getOperand(0)->getType() == IE.getType()) {
12108 unsigned NumVectorElts = IE.getType()->getNumElements();
12109 unsigned ExtractedIdx =
12110 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
12111 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
12113 if (ExtractedIdx >= NumVectorElts) // Out of range extract.
12114 return ReplaceInstUsesWith(IE, VecOp);
12116 if (InsertedIdx >= NumVectorElts) // Out of range insert.
12117 return ReplaceInstUsesWith(IE, UndefValue::get(IE.getType()));
12119 // If we are extracting a value from a vector, then inserting it right
12120 // back into the same place, just use the input vector.
12121 if (EI->getOperand(0) == VecOp && ExtractedIdx == InsertedIdx)
12122 return ReplaceInstUsesWith(IE, VecOp);
12124 // We could theoretically do this for ANY input. However, doing so could
12125 // turn chains of insertelement instructions into a chain of shufflevector
12126 // instructions, and right now we do not merge shufflevectors. As such,
12127 // only do this in a situation where it is clear that there is benefit.
12128 if (isa<UndefValue>(VecOp) || isa<ConstantAggregateZero>(VecOp)) {
12129 // Turn this into shuffle(EIOp0, VecOp, Mask). The result has all of
12130 // the values of VecOp, except then one read from EIOp0.
12131 // Build a new shuffle mask.
12132 std::vector<Constant*> Mask;
12133 if (isa<UndefValue>(VecOp))
12134 Mask.assign(NumVectorElts, UndefValue::get(Type::Int32Ty));
12136 assert(isa<ConstantAggregateZero>(VecOp) && "Unknown thing");
12137 Mask.assign(NumVectorElts, ConstantInt::get(Type::Int32Ty,
12140 Mask[InsertedIdx] = ConstantInt::get(Type::Int32Ty, ExtractedIdx);
12141 return new ShuffleVectorInst(EI->getOperand(0), VecOp,
12142 ConstantVector::get(Mask));
12145 // If this insertelement isn't used by some other insertelement, turn it
12146 // (and any insertelements it points to), into one big shuffle.
12147 if (!IE.hasOneUse() || !isa<InsertElementInst>(IE.use_back())) {
12148 std::vector<Constant*> Mask;
12150 Value *LHS = CollectShuffleElements(&IE, Mask, RHS);
12151 if (RHS == 0) RHS = UndefValue::get(LHS->getType());
12152 // We now have a shuffle of LHS, RHS, Mask.
12153 return new ShuffleVectorInst(LHS, RHS, ConstantVector::get(Mask));
12162 Instruction *InstCombiner::visitShuffleVectorInst(ShuffleVectorInst &SVI) {
12163 Value *LHS = SVI.getOperand(0);
12164 Value *RHS = SVI.getOperand(1);
12165 std::vector<unsigned> Mask = getShuffleMask(&SVI);
12167 bool MadeChange = false;
12169 // Undefined shuffle mask -> undefined value.
12170 if (isa<UndefValue>(SVI.getOperand(2)))
12171 return ReplaceInstUsesWith(SVI, UndefValue::get(SVI.getType()));
12173 uint64_t UndefElts;
12174 unsigned VWidth = cast<VectorType>(SVI.getType())->getNumElements();
12176 if (VWidth != cast<VectorType>(LHS->getType())->getNumElements())
12179 uint64_t AllOnesEltMask = ~0ULL >> (64-VWidth);
12180 if (VWidth <= 64 &&
12181 SimplifyDemandedVectorElts(&SVI, AllOnesEltMask, UndefElts)) {
12182 LHS = SVI.getOperand(0);
12183 RHS = SVI.getOperand(1);
12187 // Canonicalize shuffle(x ,x,mask) -> shuffle(x, undef,mask')
12188 // Canonicalize shuffle(undef,x,mask) -> shuffle(x, undef,mask').
12189 if (LHS == RHS || isa<UndefValue>(LHS)) {
12190 if (isa<UndefValue>(LHS) && LHS == RHS) {
12191 // shuffle(undef,undef,mask) -> undef.
12192 return ReplaceInstUsesWith(SVI, LHS);
12195 // Remap any references to RHS to use LHS.
12196 std::vector<Constant*> Elts;
12197 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
12198 if (Mask[i] >= 2*e)
12199 Elts.push_back(UndefValue::get(Type::Int32Ty));
12201 if ((Mask[i] >= e && isa<UndefValue>(RHS)) ||
12202 (Mask[i] < e && isa<UndefValue>(LHS))) {
12203 Mask[i] = 2*e; // Turn into undef.
12204 Elts.push_back(UndefValue::get(Type::Int32Ty));
12206 Mask[i] = Mask[i] % e; // Force to LHS.
12207 Elts.push_back(ConstantInt::get(Type::Int32Ty, Mask[i]));
12211 SVI.setOperand(0, SVI.getOperand(1));
12212 SVI.setOperand(1, UndefValue::get(RHS->getType()));
12213 SVI.setOperand(2, ConstantVector::get(Elts));
12214 LHS = SVI.getOperand(0);
12215 RHS = SVI.getOperand(1);
12219 // Analyze the shuffle, are the LHS or RHS and identity shuffles?
12220 bool isLHSID = true, isRHSID = true;
12222 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
12223 if (Mask[i] >= e*2) continue; // Ignore undef values.
12224 // Is this an identity shuffle of the LHS value?
12225 isLHSID &= (Mask[i] == i);
12227 // Is this an identity shuffle of the RHS value?
12228 isRHSID &= (Mask[i]-e == i);
12231 // Eliminate identity shuffles.
12232 if (isLHSID) return ReplaceInstUsesWith(SVI, LHS);
12233 if (isRHSID) return ReplaceInstUsesWith(SVI, RHS);
12235 // If the LHS is a shufflevector itself, see if we can combine it with this
12236 // one without producing an unusual shuffle. Here we are really conservative:
12237 // we are absolutely afraid of producing a shuffle mask not in the input
12238 // program, because the code gen may not be smart enough to turn a merged
12239 // shuffle into two specific shuffles: it may produce worse code. As such,
12240 // we only merge two shuffles if the result is one of the two input shuffle
12241 // masks. In this case, merging the shuffles just removes one instruction,
12242 // which we know is safe. This is good for things like turning:
12243 // (splat(splat)) -> splat.
12244 if (ShuffleVectorInst *LHSSVI = dyn_cast<ShuffleVectorInst>(LHS)) {
12245 if (isa<UndefValue>(RHS)) {
12246 std::vector<unsigned> LHSMask = getShuffleMask(LHSSVI);
12248 std::vector<unsigned> NewMask;
12249 for (unsigned i = 0, e = Mask.size(); i != e; ++i)
12250 if (Mask[i] >= 2*e)
12251 NewMask.push_back(2*e);
12253 NewMask.push_back(LHSMask[Mask[i]]);
12255 // If the result mask is equal to the src shuffle or this shuffle mask, do
12256 // the replacement.
12257 if (NewMask == LHSMask || NewMask == Mask) {
12258 unsigned LHSInNElts =
12259 cast<VectorType>(LHSSVI->getOperand(0)->getType())->getNumElements();
12260 std::vector<Constant*> Elts;
12261 for (unsigned i = 0, e = NewMask.size(); i != e; ++i) {
12262 if (NewMask[i] >= LHSInNElts*2) {
12263 Elts.push_back(UndefValue::get(Type::Int32Ty));
12265 Elts.push_back(ConstantInt::get(Type::Int32Ty, NewMask[i]));
12268 return new ShuffleVectorInst(LHSSVI->getOperand(0),
12269 LHSSVI->getOperand(1),
12270 ConstantVector::get(Elts));
12275 return MadeChange ? &SVI : 0;
12281 /// TryToSinkInstruction - Try to move the specified instruction from its
12282 /// current block into the beginning of DestBlock, which can only happen if it's
12283 /// safe to move the instruction past all of the instructions between it and the
12284 /// end of its block.
12285 static bool TryToSinkInstruction(Instruction *I, BasicBlock *DestBlock) {
12286 assert(I->hasOneUse() && "Invariants didn't hold!");
12288 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
12289 if (isa<PHINode>(I) || I->mayWriteToMemory() || isa<TerminatorInst>(I))
12292 // Do not sink alloca instructions out of the entry block.
12293 if (isa<AllocaInst>(I) && I->getParent() ==
12294 &DestBlock->getParent()->getEntryBlock())
12297 // We can only sink load instructions if there is nothing between the load and
12298 // the end of block that could change the value.
12299 if (I->mayReadFromMemory()) {
12300 for (BasicBlock::iterator Scan = I, E = I->getParent()->end();
12302 if (Scan->mayWriteToMemory())
12306 BasicBlock::iterator InsertPos = DestBlock->getFirstNonPHI();
12308 I->moveBefore(InsertPos);
12314 /// AddReachableCodeToWorklist - Walk the function in depth-first order, adding
12315 /// all reachable code to the worklist.
12317 /// This has a couple of tricks to make the code faster and more powerful. In
12318 /// particular, we constant fold and DCE instructions as we go, to avoid adding
12319 /// them to the worklist (this significantly speeds up instcombine on code where
12320 /// many instructions are dead or constant). Additionally, if we find a branch
12321 /// whose condition is a known constant, we only visit the reachable successors.
12323 static void AddReachableCodeToWorklist(BasicBlock *BB,
12324 SmallPtrSet<BasicBlock*, 64> &Visited,
12326 const TargetData *TD) {
12327 SmallVector<BasicBlock*, 256> Worklist;
12328 Worklist.push_back(BB);
12330 while (!Worklist.empty()) {
12331 BB = Worklist.back();
12332 Worklist.pop_back();
12334 // We have now visited this block! If we've already been here, ignore it.
12335 if (!Visited.insert(BB)) continue;
12337 DbgInfoIntrinsic *DBI_Prev = NULL;
12338 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) {
12339 Instruction *Inst = BBI++;
12341 // DCE instruction if trivially dead.
12342 if (isInstructionTriviallyDead(Inst)) {
12344 DOUT << "IC: DCE: " << *Inst;
12345 Inst->eraseFromParent();
12349 // ConstantProp instruction if trivially constant.
12350 if (Constant *C = ConstantFoldInstruction(Inst, TD)) {
12351 DOUT << "IC: ConstFold to: " << *C << " from: " << *Inst;
12352 Inst->replaceAllUsesWith(C);
12354 Inst->eraseFromParent();
12358 // If there are two consecutive llvm.dbg.stoppoint calls then
12359 // it is likely that the optimizer deleted code in between these
12361 DbgInfoIntrinsic *DBI_Next = dyn_cast<DbgInfoIntrinsic>(Inst);
12364 && DBI_Prev->getIntrinsicID() == llvm::Intrinsic::dbg_stoppoint
12365 && DBI_Next->getIntrinsicID() == llvm::Intrinsic::dbg_stoppoint) {
12366 IC.RemoveFromWorkList(DBI_Prev);
12367 DBI_Prev->eraseFromParent();
12369 DBI_Prev = DBI_Next;
12372 IC.AddToWorkList(Inst);
12375 // Recursively visit successors. If this is a branch or switch on a
12376 // constant, only visit the reachable successor.
12377 TerminatorInst *TI = BB->getTerminator();
12378 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
12379 if (BI->isConditional() && isa<ConstantInt>(BI->getCondition())) {
12380 bool CondVal = cast<ConstantInt>(BI->getCondition())->getZExtValue();
12381 BasicBlock *ReachableBB = BI->getSuccessor(!CondVal);
12382 Worklist.push_back(ReachableBB);
12385 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
12386 if (ConstantInt *Cond = dyn_cast<ConstantInt>(SI->getCondition())) {
12387 // See if this is an explicit destination.
12388 for (unsigned i = 1, e = SI->getNumSuccessors(); i != e; ++i)
12389 if (SI->getCaseValue(i) == Cond) {
12390 BasicBlock *ReachableBB = SI->getSuccessor(i);
12391 Worklist.push_back(ReachableBB);
12395 // Otherwise it is the default destination.
12396 Worklist.push_back(SI->getSuccessor(0));
12401 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
12402 Worklist.push_back(TI->getSuccessor(i));
12406 bool InstCombiner::DoOneIteration(Function &F, unsigned Iteration) {
12407 bool Changed = false;
12408 TD = &getAnalysis<TargetData>();
12410 DEBUG(DOUT << "\n\nINSTCOMBINE ITERATION #" << Iteration << " on "
12411 << F.getNameStr() << "\n");
12414 // Do a depth-first traversal of the function, populate the worklist with
12415 // the reachable instructions. Ignore blocks that are not reachable. Keep
12416 // track of which blocks we visit.
12417 SmallPtrSet<BasicBlock*, 64> Visited;
12418 AddReachableCodeToWorklist(F.begin(), Visited, *this, TD);
12420 // Do a quick scan over the function. If we find any blocks that are
12421 // unreachable, remove any instructions inside of them. This prevents
12422 // the instcombine code from having to deal with some bad special cases.
12423 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
12424 if (!Visited.count(BB)) {
12425 Instruction *Term = BB->getTerminator();
12426 while (Term != BB->begin()) { // Remove instrs bottom-up
12427 BasicBlock::iterator I = Term; --I;
12429 DOUT << "IC: DCE: " << *I;
12432 if (!I->use_empty())
12433 I->replaceAllUsesWith(UndefValue::get(I->getType()));
12434 I->eraseFromParent();
12440 while (!Worklist.empty()) {
12441 Instruction *I = RemoveOneFromWorkList();
12442 if (I == 0) continue; // skip null values.
12444 // Check to see if we can DCE the instruction.
12445 if (isInstructionTriviallyDead(I)) {
12446 // Add operands to the worklist.
12447 if (I->getNumOperands() < 4)
12448 AddUsesToWorkList(*I);
12451 DOUT << "IC: DCE: " << *I;
12453 I->eraseFromParent();
12454 RemoveFromWorkList(I);
12459 // Instruction isn't dead, see if we can constant propagate it.
12460 if (Constant *C = ConstantFoldInstruction(I, TD)) {
12461 DOUT << "IC: ConstFold to: " << *C << " from: " << *I;
12463 // Add operands to the worklist.
12464 AddUsesToWorkList(*I);
12465 ReplaceInstUsesWith(*I, C);
12468 I->eraseFromParent();
12469 RemoveFromWorkList(I);
12474 if (TD && I->getType()->getTypeID() == Type::VoidTyID) {
12475 // See if we can constant fold its operands.
12476 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i)
12477 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(i))
12478 if (Constant *NewC = ConstantFoldConstantExpression(CE, TD))
12485 // See if we can trivially sink this instruction to a successor basic block.
12486 if (I->hasOneUse()) {
12487 BasicBlock *BB = I->getParent();
12488 BasicBlock *UserParent = cast<Instruction>(I->use_back())->getParent();
12489 if (UserParent != BB) {
12490 bool UserIsSuccessor = false;
12491 // See if the user is one of our successors.
12492 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI)
12493 if (*SI == UserParent) {
12494 UserIsSuccessor = true;
12498 // If the user is one of our immediate successors, and if that successor
12499 // only has us as a predecessors (we'd have to split the critical edge
12500 // otherwise), we can keep going.
12501 if (UserIsSuccessor && !isa<PHINode>(I->use_back()) &&
12502 next(pred_begin(UserParent)) == pred_end(UserParent))
12503 // Okay, the CFG is simple enough, try to sink this instruction.
12504 Changed |= TryToSinkInstruction(I, UserParent);
12508 // Now that we have an instruction, try combining it to simplify it...
12512 DEBUG(std::ostringstream SS; I->print(SS); OrigI = SS.str(););
12513 if (Instruction *Result = visit(*I)) {
12515 // Should we replace the old instruction with a new one?
12517 DOUT << "IC: Old = " << *I
12518 << " New = " << *Result;
12520 // Everything uses the new instruction now.
12521 I->replaceAllUsesWith(Result);
12523 // Push the new instruction and any users onto the worklist.
12524 AddToWorkList(Result);
12525 AddUsersToWorkList(*Result);
12527 // Move the name to the new instruction first.
12528 Result->takeName(I);
12530 // Insert the new instruction into the basic block...
12531 BasicBlock *InstParent = I->getParent();
12532 BasicBlock::iterator InsertPos = I;
12534 if (!isa<PHINode>(Result)) // If combining a PHI, don't insert
12535 while (isa<PHINode>(InsertPos)) // middle of a block of PHIs.
12538 InstParent->getInstList().insert(InsertPos, Result);
12540 // Make sure that we reprocess all operands now that we reduced their
12542 AddUsesToWorkList(*I);
12544 // Instructions can end up on the worklist more than once. Make sure
12545 // we do not process an instruction that has been deleted.
12546 RemoveFromWorkList(I);
12548 // Erase the old instruction.
12549 InstParent->getInstList().erase(I);
12552 DOUT << "IC: Mod = " << OrigI
12553 << " New = " << *I;
12556 // If the instruction was modified, it's possible that it is now dead.
12557 // if so, remove it.
12558 if (isInstructionTriviallyDead(I)) {
12559 // Make sure we process all operands now that we are reducing their
12561 AddUsesToWorkList(*I);
12563 // Instructions may end up in the worklist more than once. Erase all
12564 // occurrences of this instruction.
12565 RemoveFromWorkList(I);
12566 I->eraseFromParent();
12569 AddUsersToWorkList(*I);
12576 assert(WorklistMap.empty() && "Worklist empty, but map not?");
12578 // Do an explicit clear, this shrinks the map if needed.
12579 WorklistMap.clear();
12584 bool InstCombiner::runOnFunction(Function &F) {
12585 MustPreserveLCSSA = mustPreserveAnalysisID(LCSSAID);
12587 bool EverMadeChange = false;
12589 // Iterate while there is work to do.
12590 unsigned Iteration = 0;
12591 while (DoOneIteration(F, Iteration++))
12592 EverMadeChange = true;
12593 return EverMadeChange;
12596 FunctionPass *llvm::createInstructionCombiningPass() {
12597 return new InstCombiner();